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 "ARMAddressingModes.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMISelLowering.h"
21 #include "ARMMachineFunctionInfo.h"
22 #include "ARMPerfectShuffle.h"
23 #include "ARMRegisterInfo.h"
24 #include "ARMSubtarget.h"
25 #include "ARMTargetMachine.h"
26 #include "ARMTargetObjectFile.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)"),
75 static cl::opt<std::string>
76 TrapFuncName("arm-trap-func", cl::Hidden,
77 cl::desc("Emit a call to trap function rather than a trap instruction"),
80 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
81 EVT PromotedBitwiseVT) {
82 if (VT != PromotedLdStVT) {
83 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
84 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
85 PromotedLdStVT.getSimpleVT());
87 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
88 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
89 PromotedLdStVT.getSimpleVT());
92 EVT ElemTy = VT.getVectorElementType();
93 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
94 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
95 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
96 if (ElemTy != MVT::i32) {
97 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
98 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
99 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
100 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
102 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
103 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
104 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
105 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
106 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
108 if (VT.isInteger()) {
109 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
110 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
112 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
113 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
114 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
115 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
116 setTruncStoreAction(VT.getSimpleVT(),
117 (MVT::SimpleValueType)InnerVT, Expand);
119 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
121 // Promote all bit-wise operations.
122 if (VT.isInteger() && VT != PromotedBitwiseVT) {
123 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
124 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
125 PromotedBitwiseVT.getSimpleVT());
126 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
127 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
128 PromotedBitwiseVT.getSimpleVT());
129 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
130 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
131 PromotedBitwiseVT.getSimpleVT());
134 // Neon does not support vector divide/remainder operations.
135 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
136 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
137 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
138 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
139 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
143 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
144 addRegisterClass(VT, ARM::DPRRegisterClass);
145 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
148 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
149 addRegisterClass(VT, ARM::QPRRegisterClass);
150 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
153 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
154 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
155 return new TargetLoweringObjectFileMachO();
157 return new ARMElfTargetObjectFile();
160 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
161 : TargetLowering(TM, createTLOF(TM)) {
162 Subtarget = &TM.getSubtarget<ARMSubtarget>();
163 RegInfo = TM.getRegisterInfo();
164 Itins = TM.getInstrItineraryData();
166 if (Subtarget->isTargetDarwin()) {
167 // Uses VFP for Thumb libfuncs if available.
168 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
169 // Single-precision floating-point arithmetic.
170 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
171 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
172 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
173 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
175 // Double-precision floating-point arithmetic.
176 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
177 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
178 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
179 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
181 // Single-precision comparisons.
182 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
183 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
184 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
185 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
186 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
187 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
188 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
189 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
191 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
200 // Double-precision comparisons.
201 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
202 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
203 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
204 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
205 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
206 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
207 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
208 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
210 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
219 // Floating-point to integer conversions.
220 // i64 conversions are done via library routines even when generating VFP
221 // instructions, so use the same ones.
222 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
223 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
224 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
225 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
227 // Conversions between floating types.
228 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
229 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
231 // Integer to floating-point conversions.
232 // i64 conversions are done via library routines even when generating VFP
233 // instructions, so use the same ones.
234 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
235 // e.g., __floatunsidf vs. __floatunssidfvfp.
236 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
237 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
238 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
239 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
243 // These libcalls are not available in 32-bit.
244 setLibcallName(RTLIB::SHL_I128, 0);
245 setLibcallName(RTLIB::SRL_I128, 0);
246 setLibcallName(RTLIB::SRA_I128, 0);
248 if (Subtarget->isAAPCS_ABI()) {
249 // Double-precision floating-point arithmetic helper functions
250 // RTABI chapter 4.1.2, Table 2
251 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
252 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
253 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
254 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
255 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
256 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
257 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
258 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
260 // Double-precision floating-point comparison helper functions
261 // RTABI chapter 4.1.2, Table 3
262 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
263 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
264 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
265 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
266 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
267 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
268 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
269 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
270 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
271 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
272 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
273 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
274 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
275 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
276 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
277 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
278 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
279 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
280 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
281 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
282 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
283 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
284 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
285 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
287 // Single-precision floating-point arithmetic helper functions
288 // RTABI chapter 4.1.2, Table 4
289 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
290 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
291 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
292 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
293 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
298 // Single-precision floating-point comparison helper functions
299 // RTABI chapter 4.1.2, Table 5
300 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
301 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
302 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
303 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
304 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
305 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
306 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
307 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
308 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
309 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
310 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
311 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
312 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
313 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
314 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
315 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
316 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
317 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
318 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
319 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
320 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
321 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
322 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
323 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
325 // Floating-point to integer conversions.
326 // RTABI chapter 4.1.2, Table 6
327 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
328 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
329 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
330 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
331 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
332 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
333 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
334 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
335 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
339 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
340 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
341 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
342 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
344 // Conversions between floating types.
345 // RTABI chapter 4.1.2, Table 7
346 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
347 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
348 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
351 // Integer to floating-point conversions.
352 // RTABI chapter 4.1.2, Table 8
353 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
354 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
355 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
356 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
357 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
358 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
359 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
360 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
361 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
362 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
363 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
364 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
365 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
366 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
367 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
368 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
370 // Long long helper functions
371 // RTABI chapter 4.2, Table 9
372 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
373 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
374 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
375 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
376 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
377 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
378 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
385 // Integer division functions
386 // RTABI chapter 4.3.1
387 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
388 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
389 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
390 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
391 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
392 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
393 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
396 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
397 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
398 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
401 if (HasDivModLibcall) {
402 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
403 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
406 if (Subtarget->isThumb1Only())
407 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
409 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
410 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
411 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
412 if (!Subtarget->isFPOnlySP())
413 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
415 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
418 if (Subtarget->hasNEON()) {
419 addDRTypeForNEON(MVT::v2f32);
420 addDRTypeForNEON(MVT::v8i8);
421 addDRTypeForNEON(MVT::v4i16);
422 addDRTypeForNEON(MVT::v2i32);
423 addDRTypeForNEON(MVT::v1i64);
425 addQRTypeForNEON(MVT::v4f32);
426 addQRTypeForNEON(MVT::v2f64);
427 addQRTypeForNEON(MVT::v16i8);
428 addQRTypeForNEON(MVT::v8i16);
429 addQRTypeForNEON(MVT::v4i32);
430 addQRTypeForNEON(MVT::v2i64);
432 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
433 // neither Neon nor VFP support any arithmetic operations on it.
434 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
435 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
436 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
437 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
438 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
439 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
440 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
441 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
442 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
443 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
444 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
445 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
446 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
447 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
448 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
449 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
450 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
451 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
452 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
453 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
454 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
455 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
456 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
457 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
459 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
461 // Neon does not support some operations on v1i64 and v2i64 types.
462 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
463 // Custom handling for some quad-vector types to detect VMULL.
464 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
465 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
466 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
467 // Custom handling for some vector types to avoid expensive expansions
468 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
469 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
470 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
471 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
472 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
473 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
474 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
475 // a destination type that is wider than the source.
476 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
477 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
479 setTargetDAGCombine(ISD::INTRINSIC_VOID);
480 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
481 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
482 setTargetDAGCombine(ISD::SHL);
483 setTargetDAGCombine(ISD::SRL);
484 setTargetDAGCombine(ISD::SRA);
485 setTargetDAGCombine(ISD::SIGN_EXTEND);
486 setTargetDAGCombine(ISD::ZERO_EXTEND);
487 setTargetDAGCombine(ISD::ANY_EXTEND);
488 setTargetDAGCombine(ISD::SELECT_CC);
489 setTargetDAGCombine(ISD::BUILD_VECTOR);
490 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
491 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
492 setTargetDAGCombine(ISD::STORE);
495 computeRegisterProperties();
497 // ARM does not have f32 extending load.
498 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
500 // ARM does not have i1 sign extending load.
501 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
503 // ARM supports all 4 flavors of integer indexed load / store.
504 if (!Subtarget->isThumb1Only()) {
505 for (unsigned im = (unsigned)ISD::PRE_INC;
506 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
507 setIndexedLoadAction(im, MVT::i1, Legal);
508 setIndexedLoadAction(im, MVT::i8, Legal);
509 setIndexedLoadAction(im, MVT::i16, Legal);
510 setIndexedLoadAction(im, MVT::i32, Legal);
511 setIndexedStoreAction(im, MVT::i1, Legal);
512 setIndexedStoreAction(im, MVT::i8, Legal);
513 setIndexedStoreAction(im, MVT::i16, Legal);
514 setIndexedStoreAction(im, MVT::i32, Legal);
518 // i64 operation support.
519 if (Subtarget->isThumb1Only()) {
520 setOperationAction(ISD::MUL, MVT::i64, Expand);
521 setOperationAction(ISD::MULHU, MVT::i32, Expand);
522 setOperationAction(ISD::MULHS, MVT::i32, Expand);
523 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
524 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
526 setOperationAction(ISD::MUL, MVT::i64, Expand);
527 setOperationAction(ISD::MULHU, MVT::i32, Expand);
528 if (!Subtarget->hasV6Ops())
529 setOperationAction(ISD::MULHS, MVT::i32, Expand);
531 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
532 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
533 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
534 setOperationAction(ISD::SRL, MVT::i64, Custom);
535 setOperationAction(ISD::SRA, MVT::i64, Custom);
537 // ARM does not have ROTL.
538 setOperationAction(ISD::ROTL, MVT::i32, Expand);
539 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
540 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
541 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
542 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
544 // Only ARMv6 has BSWAP.
545 if (!Subtarget->hasV6Ops())
546 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
548 // These are expanded into libcalls.
549 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
550 // v7M has a hardware divider
551 setOperationAction(ISD::SDIV, MVT::i32, Expand);
552 setOperationAction(ISD::UDIV, MVT::i32, Expand);
554 setOperationAction(ISD::SREM, MVT::i32, Expand);
555 setOperationAction(ISD::UREM, MVT::i32, Expand);
556 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
557 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
559 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
560 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
561 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
562 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
563 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
565 if (TrapFuncName.size())
566 setOperationAction(ISD::TRAP, MVT::Other, Custom);
568 setOperationAction(ISD::TRAP, MVT::Other, Legal);
570 // Use the default implementation.
571 setOperationAction(ISD::VASTART, MVT::Other, Custom);
572 setOperationAction(ISD::VAARG, MVT::Other, Expand);
573 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
574 setOperationAction(ISD::VAEND, MVT::Other, Expand);
575 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
576 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
577 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
578 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
579 setExceptionPointerRegister(ARM::R0);
580 setExceptionSelectorRegister(ARM::R1);
582 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
583 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
584 // the default expansion.
585 if (Subtarget->hasDataBarrier() ||
586 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
587 // membarrier needs custom lowering; the rest are legal and handled
589 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
591 // Set them all for expansion, which will force libcalls.
592 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
593 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
594 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
595 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
596 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
597 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
598 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
600 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
601 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
602 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
603 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
604 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
605 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
606 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
607 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
608 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
609 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
610 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
611 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
612 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
613 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
614 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
615 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
616 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
617 // Since the libcalls include locking, fold in the fences
618 setShouldFoldAtomicFences(true);
620 // 64-bit versions are always libcalls (for now)
621 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
622 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
623 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
624 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
625 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
626 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
627 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
628 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
630 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
632 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
633 if (!Subtarget->hasV6Ops()) {
634 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
635 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
637 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
639 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
640 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
641 // iff target supports vfp2.
642 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
643 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
646 // We want to custom lower some of our intrinsics.
647 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
648 if (Subtarget->isTargetDarwin()) {
649 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
650 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
651 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
654 setOperationAction(ISD::SETCC, MVT::i32, Expand);
655 setOperationAction(ISD::SETCC, MVT::f32, Expand);
656 setOperationAction(ISD::SETCC, MVT::f64, Expand);
657 setOperationAction(ISD::SELECT, MVT::i32, Custom);
658 setOperationAction(ISD::SELECT, MVT::f32, Custom);
659 setOperationAction(ISD::SELECT, MVT::f64, Custom);
660 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
661 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
662 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
664 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
665 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
666 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
667 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
668 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
670 // We don't support sin/cos/fmod/copysign/pow
671 setOperationAction(ISD::FSIN, MVT::f64, Expand);
672 setOperationAction(ISD::FSIN, MVT::f32, Expand);
673 setOperationAction(ISD::FCOS, MVT::f32, Expand);
674 setOperationAction(ISD::FCOS, MVT::f64, Expand);
675 setOperationAction(ISD::FREM, MVT::f64, Expand);
676 setOperationAction(ISD::FREM, MVT::f32, Expand);
677 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
678 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
679 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
681 setOperationAction(ISD::FPOW, MVT::f64, Expand);
682 setOperationAction(ISD::FPOW, MVT::f32, Expand);
684 // Various VFP goodness
685 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
686 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
687 if (Subtarget->hasVFP2()) {
688 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
689 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
690 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
691 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
693 // Special handling for half-precision FP.
694 if (!Subtarget->hasFP16()) {
695 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
696 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
700 // We have target-specific dag combine patterns for the following nodes:
701 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
702 setTargetDAGCombine(ISD::ADD);
703 setTargetDAGCombine(ISD::SUB);
704 setTargetDAGCombine(ISD::MUL);
706 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
707 setTargetDAGCombine(ISD::OR);
708 if (Subtarget->hasNEON())
709 setTargetDAGCombine(ISD::AND);
711 setStackPointerRegisterToSaveRestore(ARM::SP);
713 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
714 setSchedulingPreference(Sched::RegPressure);
716 setSchedulingPreference(Sched::Hybrid);
718 //// temporary - rewrite interface to use type
719 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
721 // On ARM arguments smaller than 4 bytes are extended, so all arguments
722 // are at least 4 bytes aligned.
723 setMinStackArgumentAlignment(4);
725 benefitFromCodePlacementOpt = true;
728 // FIXME: It might make sense to define the representative register class as the
729 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
730 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
731 // SPR's representative would be DPR_VFP2. This should work well if register
732 // pressure tracking were modified such that a register use would increment the
733 // pressure of the register class's representative and all of it's super
734 // classes' representatives transitively. We have not implemented this because
735 // of the difficulty prior to coalescing of modeling operand register classes
736 // due to the common occurence of cross class copies and subregister insertions
738 std::pair<const TargetRegisterClass*, uint8_t>
739 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
740 const TargetRegisterClass *RRC = 0;
742 switch (VT.getSimpleVT().SimpleTy) {
744 return TargetLowering::findRepresentativeClass(VT);
745 // Use DPR as representative register class for all floating point
746 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
747 // the cost is 1 for both f32 and f64.
748 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
749 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
750 RRC = ARM::DPRRegisterClass;
751 // When NEON is used for SP, only half of the register file is available
752 // because operations that define both SP and DP results will be constrained
753 // to the VFP2 class (D0-D15). We currently model this constraint prior to
754 // coalescing by double-counting the SP regs. See the FIXME above.
755 if (Subtarget->useNEONForSinglePrecisionFP())
758 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
759 case MVT::v4f32: case MVT::v2f64:
760 RRC = ARM::DPRRegisterClass;
764 RRC = ARM::DPRRegisterClass;
768 RRC = ARM::DPRRegisterClass;
772 return std::make_pair(RRC, Cost);
775 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
778 case ARMISD::Wrapper: return "ARMISD::Wrapper";
779 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
780 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
781 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
782 case ARMISD::CALL: return "ARMISD::CALL";
783 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
784 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
785 case ARMISD::tCALL: return "ARMISD::tCALL";
786 case ARMISD::BRCOND: return "ARMISD::BRCOND";
787 case ARMISD::BR_JT: return "ARMISD::BR_JT";
788 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
789 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
790 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
791 case ARMISD::CMP: return "ARMISD::CMP";
792 case ARMISD::CMPZ: return "ARMISD::CMPZ";
793 case ARMISD::CMPFP: return "ARMISD::CMPFP";
794 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
795 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
796 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
797 case ARMISD::CMOV: return "ARMISD::CMOV";
799 case ARMISD::RBIT: return "ARMISD::RBIT";
801 case ARMISD::FTOSI: return "ARMISD::FTOSI";
802 case ARMISD::FTOUI: return "ARMISD::FTOUI";
803 case ARMISD::SITOF: return "ARMISD::SITOF";
804 case ARMISD::UITOF: return "ARMISD::UITOF";
806 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
807 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
808 case ARMISD::RRX: return "ARMISD::RRX";
810 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
811 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
813 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
814 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
815 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
817 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
819 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
821 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
823 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
824 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
826 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
828 case ARMISD::VCEQ: return "ARMISD::VCEQ";
829 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
830 case ARMISD::VCGE: return "ARMISD::VCGE";
831 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
832 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
833 case ARMISD::VCGEU: return "ARMISD::VCGEU";
834 case ARMISD::VCGT: return "ARMISD::VCGT";
835 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
836 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
837 case ARMISD::VCGTU: return "ARMISD::VCGTU";
838 case ARMISD::VTST: return "ARMISD::VTST";
840 case ARMISD::VSHL: return "ARMISD::VSHL";
841 case ARMISD::VSHRs: return "ARMISD::VSHRs";
842 case ARMISD::VSHRu: return "ARMISD::VSHRu";
843 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
844 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
845 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
846 case ARMISD::VSHRN: return "ARMISD::VSHRN";
847 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
848 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
849 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
850 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
851 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
852 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
853 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
854 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
855 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
856 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
857 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
858 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
859 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
860 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
861 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
862 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
863 case ARMISD::VDUP: return "ARMISD::VDUP";
864 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
865 case ARMISD::VEXT: return "ARMISD::VEXT";
866 case ARMISD::VREV64: return "ARMISD::VREV64";
867 case ARMISD::VREV32: return "ARMISD::VREV32";
868 case ARMISD::VREV16: return "ARMISD::VREV16";
869 case ARMISD::VZIP: return "ARMISD::VZIP";
870 case ARMISD::VUZP: return "ARMISD::VUZP";
871 case ARMISD::VTRN: return "ARMISD::VTRN";
872 case ARMISD::VTBL1: return "ARMISD::VTBL1";
873 case ARMISD::VTBL2: return "ARMISD::VTBL2";
874 case ARMISD::VMULLs: return "ARMISD::VMULLs";
875 case ARMISD::VMULLu: return "ARMISD::VMULLu";
876 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
877 case ARMISD::FMAX: return "ARMISD::FMAX";
878 case ARMISD::FMIN: return "ARMISD::FMIN";
879 case ARMISD::BFI: return "ARMISD::BFI";
880 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
881 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
882 case ARMISD::VBSL: return "ARMISD::VBSL";
883 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
884 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
885 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
886 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
887 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
888 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
889 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
890 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
891 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
892 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
893 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
894 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
895 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
896 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
897 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
898 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
899 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
900 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
901 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
902 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
906 /// getRegClassFor - Return the register class that should be used for the
907 /// specified value type.
908 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
909 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
910 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
911 // load / store 4 to 8 consecutive D registers.
912 if (Subtarget->hasNEON()) {
913 if (VT == MVT::v4i64)
914 return ARM::QQPRRegisterClass;
915 else if (VT == MVT::v8i64)
916 return ARM::QQQQPRRegisterClass;
918 return TargetLowering::getRegClassFor(VT);
921 // Create a fast isel object.
923 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
924 return ARM::createFastISel(funcInfo);
927 /// getFunctionAlignment - Return the Log2 alignment of this function.
928 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
929 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
932 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
933 /// be used for loads / stores from the global.
934 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
935 return (Subtarget->isThumb1Only() ? 127 : 4095);
938 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
939 unsigned NumVals = N->getNumValues();
941 return Sched::RegPressure;
943 for (unsigned i = 0; i != NumVals; ++i) {
944 EVT VT = N->getValueType(i);
945 if (VT == MVT::Glue || VT == MVT::Other)
947 if (VT.isFloatingPoint() || VT.isVector())
948 return Sched::Latency;
951 if (!N->isMachineOpcode())
952 return Sched::RegPressure;
954 // Load are scheduled for latency even if there instruction itinerary
956 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
957 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
959 if (TID.getNumDefs() == 0)
960 return Sched::RegPressure;
961 if (!Itins->isEmpty() &&
962 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
963 return Sched::Latency;
965 return Sched::RegPressure;
968 //===----------------------------------------------------------------------===//
970 //===----------------------------------------------------------------------===//
972 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
973 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
975 default: llvm_unreachable("Unknown condition code!");
976 case ISD::SETNE: return ARMCC::NE;
977 case ISD::SETEQ: return ARMCC::EQ;
978 case ISD::SETGT: return ARMCC::GT;
979 case ISD::SETGE: return ARMCC::GE;
980 case ISD::SETLT: return ARMCC::LT;
981 case ISD::SETLE: return ARMCC::LE;
982 case ISD::SETUGT: return ARMCC::HI;
983 case ISD::SETUGE: return ARMCC::HS;
984 case ISD::SETULT: return ARMCC::LO;
985 case ISD::SETULE: return ARMCC::LS;
989 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
990 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
991 ARMCC::CondCodes &CondCode2) {
992 CondCode2 = ARMCC::AL;
994 default: llvm_unreachable("Unknown FP condition!");
996 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
998 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1000 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1001 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1002 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1003 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1004 case ISD::SETO: CondCode = ARMCC::VC; break;
1005 case ISD::SETUO: CondCode = ARMCC::VS; break;
1006 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1007 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1008 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1010 case ISD::SETULT: CondCode = ARMCC::LT; break;
1012 case ISD::SETULE: CondCode = ARMCC::LE; break;
1014 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1018 //===----------------------------------------------------------------------===//
1019 // Calling Convention Implementation
1020 //===----------------------------------------------------------------------===//
1022 #include "ARMGenCallingConv.inc"
1024 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1025 /// given CallingConvention value.
1026 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1028 bool isVarArg) const {
1031 llvm_unreachable("Unsupported calling convention");
1032 case CallingConv::Fast:
1033 if (Subtarget->hasVFP2() && !isVarArg) {
1034 if (!Subtarget->isAAPCS_ABI())
1035 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1036 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1037 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1040 case CallingConv::C: {
1041 // Use target triple & subtarget features to do actual dispatch.
1042 if (!Subtarget->isAAPCS_ABI())
1043 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1044 else if (Subtarget->hasVFP2() &&
1045 FloatABIType == FloatABI::Hard && !isVarArg)
1046 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1047 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1049 case CallingConv::ARM_AAPCS_VFP:
1050 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1051 case CallingConv::ARM_AAPCS:
1052 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1053 case CallingConv::ARM_APCS:
1054 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1058 /// LowerCallResult - Lower the result values of a call into the
1059 /// appropriate copies out of appropriate physical registers.
1061 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1062 CallingConv::ID CallConv, bool isVarArg,
1063 const SmallVectorImpl<ISD::InputArg> &Ins,
1064 DebugLoc dl, SelectionDAG &DAG,
1065 SmallVectorImpl<SDValue> &InVals) const {
1067 // Assign locations to each value returned by this call.
1068 SmallVector<CCValAssign, 16> RVLocs;
1069 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1070 RVLocs, *DAG.getContext());
1071 CCInfo.AnalyzeCallResult(Ins,
1072 CCAssignFnForNode(CallConv, /* Return*/ true,
1075 // Copy all of the result registers out of their specified physreg.
1076 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1077 CCValAssign VA = RVLocs[i];
1080 if (VA.needsCustom()) {
1081 // Handle f64 or half of a v2f64.
1082 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1084 Chain = Lo.getValue(1);
1085 InFlag = Lo.getValue(2);
1086 VA = RVLocs[++i]; // skip ahead to next loc
1087 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1089 Chain = Hi.getValue(1);
1090 InFlag = Hi.getValue(2);
1091 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1093 if (VA.getLocVT() == MVT::v2f64) {
1094 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1095 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1096 DAG.getConstant(0, MVT::i32));
1098 VA = RVLocs[++i]; // skip ahead to next loc
1099 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1100 Chain = Lo.getValue(1);
1101 InFlag = Lo.getValue(2);
1102 VA = RVLocs[++i]; // skip ahead to next loc
1103 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1104 Chain = Hi.getValue(1);
1105 InFlag = Hi.getValue(2);
1106 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1107 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1108 DAG.getConstant(1, MVT::i32));
1111 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1113 Chain = Val.getValue(1);
1114 InFlag = Val.getValue(2);
1117 switch (VA.getLocInfo()) {
1118 default: llvm_unreachable("Unknown loc info!");
1119 case CCValAssign::Full: break;
1120 case CCValAssign::BCvt:
1121 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1125 InVals.push_back(Val);
1131 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1132 /// by "Src" to address "Dst" of size "Size". Alignment information is
1133 /// specified by the specific parameter attribute. The copy will be passed as
1134 /// a byval function parameter.
1135 /// Sometimes what we are copying is the end of a larger object, the part that
1136 /// does not fit in registers.
1138 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1139 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1141 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1142 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1143 /*isVolatile=*/false, /*AlwaysInline=*/false,
1144 MachinePointerInfo(0), MachinePointerInfo(0));
1147 /// LowerMemOpCallTo - Store the argument to the stack.
1149 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1150 SDValue StackPtr, SDValue Arg,
1151 DebugLoc dl, SelectionDAG &DAG,
1152 const CCValAssign &VA,
1153 ISD::ArgFlagsTy Flags) const {
1154 unsigned LocMemOffset = VA.getLocMemOffset();
1155 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1156 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1157 if (Flags.isByVal())
1158 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1160 return DAG.getStore(Chain, dl, Arg, PtrOff,
1161 MachinePointerInfo::getStack(LocMemOffset),
1165 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1166 SDValue Chain, SDValue &Arg,
1167 RegsToPassVector &RegsToPass,
1168 CCValAssign &VA, CCValAssign &NextVA,
1170 SmallVector<SDValue, 8> &MemOpChains,
1171 ISD::ArgFlagsTy Flags) const {
1173 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1174 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1175 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1177 if (NextVA.isRegLoc())
1178 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1180 assert(NextVA.isMemLoc());
1181 if (StackPtr.getNode() == 0)
1182 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1184 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1190 /// LowerCall - Lowering a call into a callseq_start <-
1191 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1194 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1195 CallingConv::ID CallConv, bool isVarArg,
1197 const SmallVectorImpl<ISD::OutputArg> &Outs,
1198 const SmallVectorImpl<SDValue> &OutVals,
1199 const SmallVectorImpl<ISD::InputArg> &Ins,
1200 DebugLoc dl, SelectionDAG &DAG,
1201 SmallVectorImpl<SDValue> &InVals) const {
1202 MachineFunction &MF = DAG.getMachineFunction();
1203 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1204 bool IsSibCall = false;
1205 // Temporarily disable tail calls so things don't break.
1206 if (!EnableARMTailCalls)
1209 // Check if it's really possible to do a tail call.
1210 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1211 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1212 Outs, OutVals, Ins, DAG);
1213 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1214 // detected sibcalls.
1221 // Analyze operands of the call, assigning locations to each operand.
1222 SmallVector<CCValAssign, 16> ArgLocs;
1223 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1225 CCInfo.AnalyzeCallOperands(Outs,
1226 CCAssignFnForNode(CallConv, /* Return*/ false,
1229 // Get a count of how many bytes are to be pushed on the stack.
1230 unsigned NumBytes = CCInfo.getNextStackOffset();
1232 // For tail calls, memory operands are available in our caller's stack.
1236 // Adjust the stack pointer for the new arguments...
1237 // These operations are automatically eliminated by the prolog/epilog pass
1239 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1241 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1243 RegsToPassVector RegsToPass;
1244 SmallVector<SDValue, 8> MemOpChains;
1246 // Walk the register/memloc assignments, inserting copies/loads. In the case
1247 // of tail call optimization, arguments are handled later.
1248 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1250 ++i, ++realArgIdx) {
1251 CCValAssign &VA = ArgLocs[i];
1252 SDValue Arg = OutVals[realArgIdx];
1253 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1254 bool isByVal = Flags.isByVal();
1256 // Promote the value if needed.
1257 switch (VA.getLocInfo()) {
1258 default: llvm_unreachable("Unknown loc info!");
1259 case CCValAssign::Full: break;
1260 case CCValAssign::SExt:
1261 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1263 case CCValAssign::ZExt:
1264 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1266 case CCValAssign::AExt:
1267 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1269 case CCValAssign::BCvt:
1270 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1274 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1275 if (VA.needsCustom()) {
1276 if (VA.getLocVT() == MVT::v2f64) {
1277 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1278 DAG.getConstant(0, MVT::i32));
1279 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1280 DAG.getConstant(1, MVT::i32));
1282 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1283 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1285 VA = ArgLocs[++i]; // skip ahead to next loc
1286 if (VA.isRegLoc()) {
1287 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1288 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1290 assert(VA.isMemLoc());
1292 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1293 dl, DAG, VA, Flags));
1296 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1297 StackPtr, MemOpChains, Flags);
1299 } else if (VA.isRegLoc()) {
1300 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1301 } else if (!IsSibCall || isByVal) {
1302 assert(VA.isMemLoc());
1304 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1305 dl, DAG, VA, Flags));
1309 if (!MemOpChains.empty())
1310 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1311 &MemOpChains[0], MemOpChains.size());
1313 // Build a sequence of copy-to-reg nodes chained together with token chain
1314 // and flag operands which copy the outgoing args into the appropriate regs.
1316 // Tail call byval lowering might overwrite argument registers so in case of
1317 // tail call optimization the copies to registers are lowered later.
1319 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1320 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1321 RegsToPass[i].second, InFlag);
1322 InFlag = Chain.getValue(1);
1325 // For tail calls lower the arguments to the 'real' stack slot.
1327 // Force all the incoming stack arguments to be loaded from the stack
1328 // before any new outgoing arguments are stored to the stack, because the
1329 // outgoing stack slots may alias the incoming argument stack slots, and
1330 // the alias isn't otherwise explicit. This is slightly more conservative
1331 // than necessary, because it means that each store effectively depends
1332 // on every argument instead of just those arguments it would clobber.
1334 // Do not flag preceeding copytoreg stuff together with the following stuff.
1336 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1337 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1338 RegsToPass[i].second, InFlag);
1339 InFlag = Chain.getValue(1);
1344 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1345 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1346 // node so that legalize doesn't hack it.
1347 bool isDirect = false;
1348 bool isARMFunc = false;
1349 bool isLocalARMFunc = false;
1350 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1352 if (EnableARMLongCalls) {
1353 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1354 && "long-calls with non-static relocation model!");
1355 // Handle a global address or an external symbol. If it's not one of
1356 // those, the target's already in a register, so we don't need to do
1358 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1359 const GlobalValue *GV = G->getGlobal();
1360 // Create a constant pool entry for the callee address
1361 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1362 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1365 // Get the address of the callee into a register
1366 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1367 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1368 Callee = DAG.getLoad(getPointerTy(), dl,
1369 DAG.getEntryNode(), CPAddr,
1370 MachinePointerInfo::getConstantPool(),
1372 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1373 const char *Sym = S->getSymbol();
1375 // Create a constant pool entry for the callee address
1376 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1377 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1378 Sym, ARMPCLabelIndex, 0);
1379 // Get the address of the callee into a register
1380 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1381 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1382 Callee = DAG.getLoad(getPointerTy(), dl,
1383 DAG.getEntryNode(), CPAddr,
1384 MachinePointerInfo::getConstantPool(),
1387 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1388 const GlobalValue *GV = G->getGlobal();
1390 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1391 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1392 getTargetMachine().getRelocationModel() != Reloc::Static;
1393 isARMFunc = !Subtarget->isThumb() || isStub;
1394 // ARM call to a local ARM function is predicable.
1395 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1396 // tBX takes a register source operand.
1397 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1398 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1399 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1402 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1403 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1404 Callee = DAG.getLoad(getPointerTy(), dl,
1405 DAG.getEntryNode(), CPAddr,
1406 MachinePointerInfo::getConstantPool(),
1408 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1409 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1410 getPointerTy(), Callee, PICLabel);
1412 // On ELF targets for PIC code, direct calls should go through the PLT
1413 unsigned OpFlags = 0;
1414 if (Subtarget->isTargetELF() &&
1415 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1416 OpFlags = ARMII::MO_PLT;
1417 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1419 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1421 bool isStub = Subtarget->isTargetDarwin() &&
1422 getTargetMachine().getRelocationModel() != Reloc::Static;
1423 isARMFunc = !Subtarget->isThumb() || isStub;
1424 // tBX takes a register source operand.
1425 const char *Sym = S->getSymbol();
1426 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1427 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1428 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1429 Sym, ARMPCLabelIndex, 4);
1430 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1431 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1432 Callee = DAG.getLoad(getPointerTy(), dl,
1433 DAG.getEntryNode(), CPAddr,
1434 MachinePointerInfo::getConstantPool(),
1436 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1437 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1438 getPointerTy(), Callee, PICLabel);
1440 unsigned OpFlags = 0;
1441 // On ELF targets for PIC code, direct calls should go through the PLT
1442 if (Subtarget->isTargetELF() &&
1443 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1444 OpFlags = ARMII::MO_PLT;
1445 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1449 // FIXME: handle tail calls differently.
1451 if (Subtarget->isThumb()) {
1452 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1453 CallOpc = ARMISD::CALL_NOLINK;
1455 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1457 CallOpc = (isDirect || Subtarget->hasV5TOps())
1458 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1459 : ARMISD::CALL_NOLINK;
1462 std::vector<SDValue> Ops;
1463 Ops.push_back(Chain);
1464 Ops.push_back(Callee);
1466 // Add argument registers to the end of the list so that they are known live
1468 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1469 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1470 RegsToPass[i].second.getValueType()));
1472 if (InFlag.getNode())
1473 Ops.push_back(InFlag);
1475 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1477 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1479 // Returns a chain and a flag for retval copy to use.
1480 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1481 InFlag = Chain.getValue(1);
1483 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1484 DAG.getIntPtrConstant(0, true), InFlag);
1486 InFlag = Chain.getValue(1);
1488 // Handle result values, copying them out of physregs into vregs that we
1490 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1494 /// HandleByVal - Every parameter *after* a byval parameter is passed
1495 /// on the stack. Confiscate all the parameter registers to insure
1498 llvm::ARMTargetLowering::HandleByVal(CCState *State) const {
1499 static const unsigned RegList1[] = {
1500 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1502 do {} while (State->AllocateReg(RegList1, 4));
1505 /// MatchingStackOffset - Return true if the given stack call argument is
1506 /// already available in the same position (relatively) of the caller's
1507 /// incoming argument stack.
1509 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1510 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1511 const ARMInstrInfo *TII) {
1512 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1514 if (Arg.getOpcode() == ISD::CopyFromReg) {
1515 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1516 if (!TargetRegisterInfo::isVirtualRegister(VR))
1518 MachineInstr *Def = MRI->getVRegDef(VR);
1521 if (!Flags.isByVal()) {
1522 if (!TII->isLoadFromStackSlot(Def, FI))
1527 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1528 if (Flags.isByVal())
1529 // ByVal argument is passed in as a pointer but it's now being
1530 // dereferenced. e.g.
1531 // define @foo(%struct.X* %A) {
1532 // tail call @bar(%struct.X* byval %A)
1535 SDValue Ptr = Ld->getBasePtr();
1536 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1539 FI = FINode->getIndex();
1543 assert(FI != INT_MAX);
1544 if (!MFI->isFixedObjectIndex(FI))
1546 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1549 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1550 /// for tail call optimization. Targets which want to do tail call
1551 /// optimization should implement this function.
1553 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1554 CallingConv::ID CalleeCC,
1556 bool isCalleeStructRet,
1557 bool isCallerStructRet,
1558 const SmallVectorImpl<ISD::OutputArg> &Outs,
1559 const SmallVectorImpl<SDValue> &OutVals,
1560 const SmallVectorImpl<ISD::InputArg> &Ins,
1561 SelectionDAG& DAG) const {
1562 const Function *CallerF = DAG.getMachineFunction().getFunction();
1563 CallingConv::ID CallerCC = CallerF->getCallingConv();
1564 bool CCMatch = CallerCC == CalleeCC;
1566 // Look for obvious safe cases to perform tail call optimization that do not
1567 // require ABI changes. This is what gcc calls sibcall.
1569 // Do not sibcall optimize vararg calls unless the call site is not passing
1571 if (isVarArg && !Outs.empty())
1574 // Also avoid sibcall optimization if either caller or callee uses struct
1575 // return semantics.
1576 if (isCalleeStructRet || isCallerStructRet)
1579 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1580 // emitEpilogue is not ready for them.
1581 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1582 // LR. This means if we need to reload LR, it takes an extra instructions,
1583 // which outweighs the value of the tail call; but here we don't know yet
1584 // whether LR is going to be used. Probably the right approach is to
1585 // generate the tail call here and turn it back into CALL/RET in
1586 // emitEpilogue if LR is used.
1588 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1589 // but we need to make sure there are enough registers; the only valid
1590 // registers are the 4 used for parameters. We don't currently do this
1592 if (Subtarget->isThumb1Only())
1595 // If the calling conventions do not match, then we'd better make sure the
1596 // results are returned in the same way as what the caller expects.
1598 SmallVector<CCValAssign, 16> RVLocs1;
1599 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1600 RVLocs1, *DAG.getContext());
1601 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1603 SmallVector<CCValAssign, 16> RVLocs2;
1604 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1605 RVLocs2, *DAG.getContext());
1606 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1608 if (RVLocs1.size() != RVLocs2.size())
1610 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1611 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1613 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1615 if (RVLocs1[i].isRegLoc()) {
1616 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1619 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1625 // If the callee takes no arguments then go on to check the results of the
1627 if (!Outs.empty()) {
1628 // Check if stack adjustment is needed. For now, do not do this if any
1629 // argument is passed on the stack.
1630 SmallVector<CCValAssign, 16> ArgLocs;
1631 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1632 ArgLocs, *DAG.getContext());
1633 CCInfo.AnalyzeCallOperands(Outs,
1634 CCAssignFnForNode(CalleeCC, false, isVarArg));
1635 if (CCInfo.getNextStackOffset()) {
1636 MachineFunction &MF = DAG.getMachineFunction();
1638 // Check if the arguments are already laid out in the right way as
1639 // the caller's fixed stack objects.
1640 MachineFrameInfo *MFI = MF.getFrameInfo();
1641 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1642 const ARMInstrInfo *TII =
1643 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1644 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1646 ++i, ++realArgIdx) {
1647 CCValAssign &VA = ArgLocs[i];
1648 EVT RegVT = VA.getLocVT();
1649 SDValue Arg = OutVals[realArgIdx];
1650 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1651 if (VA.getLocInfo() == CCValAssign::Indirect)
1653 if (VA.needsCustom()) {
1654 // f64 and vector types are split into multiple registers or
1655 // register/stack-slot combinations. The types will not match
1656 // the registers; give up on memory f64 refs until we figure
1657 // out what to do about this.
1660 if (!ArgLocs[++i].isRegLoc())
1662 if (RegVT == MVT::v2f64) {
1663 if (!ArgLocs[++i].isRegLoc())
1665 if (!ArgLocs[++i].isRegLoc())
1668 } else if (!VA.isRegLoc()) {
1669 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1681 ARMTargetLowering::LowerReturn(SDValue Chain,
1682 CallingConv::ID CallConv, bool isVarArg,
1683 const SmallVectorImpl<ISD::OutputArg> &Outs,
1684 const SmallVectorImpl<SDValue> &OutVals,
1685 DebugLoc dl, SelectionDAG &DAG) const {
1687 // CCValAssign - represent the assignment of the return value to a location.
1688 SmallVector<CCValAssign, 16> RVLocs;
1690 // CCState - Info about the registers and stack slots.
1691 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1694 // Analyze outgoing return values.
1695 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1698 // If this is the first return lowered for this function, add
1699 // the regs to the liveout set for the function.
1700 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1701 for (unsigned i = 0; i != RVLocs.size(); ++i)
1702 if (RVLocs[i].isRegLoc())
1703 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1708 // Copy the result values into the output registers.
1709 for (unsigned i = 0, realRVLocIdx = 0;
1711 ++i, ++realRVLocIdx) {
1712 CCValAssign &VA = RVLocs[i];
1713 assert(VA.isRegLoc() && "Can only return in registers!");
1715 SDValue Arg = OutVals[realRVLocIdx];
1717 switch (VA.getLocInfo()) {
1718 default: llvm_unreachable("Unknown loc info!");
1719 case CCValAssign::Full: break;
1720 case CCValAssign::BCvt:
1721 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1725 if (VA.needsCustom()) {
1726 if (VA.getLocVT() == MVT::v2f64) {
1727 // Extract the first half and return it in two registers.
1728 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1729 DAG.getConstant(0, MVT::i32));
1730 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1731 DAG.getVTList(MVT::i32, MVT::i32), Half);
1733 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1734 Flag = Chain.getValue(1);
1735 VA = RVLocs[++i]; // skip ahead to next loc
1736 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1737 HalfGPRs.getValue(1), Flag);
1738 Flag = Chain.getValue(1);
1739 VA = RVLocs[++i]; // skip ahead to next loc
1741 // Extract the 2nd half and fall through to handle it as an f64 value.
1742 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1743 DAG.getConstant(1, MVT::i32));
1745 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1747 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1748 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1749 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1750 Flag = Chain.getValue(1);
1751 VA = RVLocs[++i]; // skip ahead to next loc
1752 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1755 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1757 // Guarantee that all emitted copies are
1758 // stuck together, avoiding something bad.
1759 Flag = Chain.getValue(1);
1764 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1766 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1771 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1772 if (N->getNumValues() != 1)
1774 if (!N->hasNUsesOfValue(1, 0))
1777 unsigned NumCopies = 0;
1779 SDNode *Use = *N->use_begin();
1780 if (Use->getOpcode() == ISD::CopyToReg) {
1781 Copies[NumCopies++] = Use;
1782 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1783 // f64 returned in a pair of GPRs.
1784 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1786 if (UI->getOpcode() != ISD::CopyToReg)
1788 Copies[UI.getUse().getResNo()] = *UI;
1791 } else if (Use->getOpcode() == ISD::BITCAST) {
1792 // f32 returned in a single GPR.
1793 if (!Use->hasNUsesOfValue(1, 0))
1795 Use = *Use->use_begin();
1796 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1798 Copies[NumCopies++] = Use;
1803 if (NumCopies != 1 && NumCopies != 2)
1806 bool HasRet = false;
1807 for (unsigned i = 0; i < NumCopies; ++i) {
1808 SDNode *Copy = Copies[i];
1809 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1811 if (UI->getOpcode() == ISD::CopyToReg) {
1813 if (Use == Copies[0] || Use == Copies[1])
1817 if (UI->getOpcode() != ARMISD::RET_FLAG)
1826 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1827 if (!EnableARMTailCalls)
1830 if (!CI->isTailCall())
1833 return !Subtarget->isThumb1Only();
1836 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1837 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1838 // one of the above mentioned nodes. It has to be wrapped because otherwise
1839 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1840 // be used to form addressing mode. These wrapped nodes will be selected
1842 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1843 EVT PtrVT = Op.getValueType();
1844 // FIXME there is no actual debug info here
1845 DebugLoc dl = Op.getDebugLoc();
1846 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1848 if (CP->isMachineConstantPoolEntry())
1849 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1850 CP->getAlignment());
1852 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1853 CP->getAlignment());
1854 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1857 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1858 return MachineJumpTableInfo::EK_Inline;
1861 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1862 SelectionDAG &DAG) const {
1863 MachineFunction &MF = DAG.getMachineFunction();
1864 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1865 unsigned ARMPCLabelIndex = 0;
1866 DebugLoc DL = Op.getDebugLoc();
1867 EVT PtrVT = getPointerTy();
1868 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1869 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1871 if (RelocM == Reloc::Static) {
1872 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1874 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1875 ARMPCLabelIndex = AFI->createPICLabelUId();
1876 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1877 ARMCP::CPBlockAddress,
1879 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1881 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1882 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1883 MachinePointerInfo::getConstantPool(),
1885 if (RelocM == Reloc::Static)
1887 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1888 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1891 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1893 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1894 SelectionDAG &DAG) const {
1895 DebugLoc dl = GA->getDebugLoc();
1896 EVT PtrVT = getPointerTy();
1897 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1898 MachineFunction &MF = DAG.getMachineFunction();
1899 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1900 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1901 ARMConstantPoolValue *CPV =
1902 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1903 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1904 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1905 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1906 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1907 MachinePointerInfo::getConstantPool(),
1909 SDValue Chain = Argument.getValue(1);
1911 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1912 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1914 // call __tls_get_addr.
1917 Entry.Node = Argument;
1918 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1919 Args.push_back(Entry);
1920 // FIXME: is there useful debug info available here?
1921 std::pair<SDValue, SDValue> CallResult =
1922 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1923 false, false, false, false,
1924 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1925 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1926 return CallResult.first;
1929 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1930 // "local exec" model.
1932 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1933 SelectionDAG &DAG) const {
1934 const GlobalValue *GV = GA->getGlobal();
1935 DebugLoc dl = GA->getDebugLoc();
1937 SDValue Chain = DAG.getEntryNode();
1938 EVT PtrVT = getPointerTy();
1939 // Get the Thread Pointer
1940 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1942 if (GV->isDeclaration()) {
1943 MachineFunction &MF = DAG.getMachineFunction();
1944 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1945 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1946 // Initial exec model.
1947 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1948 ARMConstantPoolValue *CPV =
1949 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1950 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1951 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1952 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1953 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1954 MachinePointerInfo::getConstantPool(),
1956 Chain = Offset.getValue(1);
1958 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1959 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1961 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1962 MachinePointerInfo::getConstantPool(),
1966 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
1967 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1968 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1969 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1970 MachinePointerInfo::getConstantPool(),
1974 // The address of the thread local variable is the add of the thread
1975 // pointer with the offset of the variable.
1976 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1980 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1981 // TODO: implement the "local dynamic" model
1982 assert(Subtarget->isTargetELF() &&
1983 "TLS not implemented for non-ELF targets");
1984 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1985 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1986 // otherwise use the "Local Exec" TLS Model
1987 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1988 return LowerToTLSGeneralDynamicModel(GA, DAG);
1990 return LowerToTLSExecModels(GA, DAG);
1993 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1994 SelectionDAG &DAG) const {
1995 EVT PtrVT = getPointerTy();
1996 DebugLoc dl = Op.getDebugLoc();
1997 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1998 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1999 if (RelocM == Reloc::PIC_) {
2000 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2001 ARMConstantPoolValue *CPV =
2002 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2003 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2004 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2005 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2007 MachinePointerInfo::getConstantPool(),
2009 SDValue Chain = Result.getValue(1);
2010 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2011 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2013 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2014 MachinePointerInfo::getGOT(), false, false, 0);
2018 // If we have T2 ops, we can materialize the address directly via movt/movw
2019 // pair. This is always cheaper.
2020 if (Subtarget->useMovt()) {
2022 // FIXME: Once remat is capable of dealing with instructions with register
2023 // operands, expand this into two nodes.
2024 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2025 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2027 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2028 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2029 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2030 MachinePointerInfo::getConstantPool(),
2035 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2036 SelectionDAG &DAG) const {
2037 EVT PtrVT = getPointerTy();
2038 DebugLoc dl = Op.getDebugLoc();
2039 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2040 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2041 MachineFunction &MF = DAG.getMachineFunction();
2042 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2044 if (Subtarget->useMovt()) {
2046 // FIXME: Once remat is capable of dealing with instructions with register
2047 // operands, expand this into two nodes.
2048 if (RelocM == Reloc::Static)
2049 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2050 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2052 unsigned Wrapper = (RelocM == Reloc::PIC_)
2053 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2054 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2055 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2056 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2057 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2058 MachinePointerInfo::getGOT(), false, false, 0);
2062 unsigned ARMPCLabelIndex = 0;
2064 if (RelocM == Reloc::Static) {
2065 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2067 ARMPCLabelIndex = AFI->createPICLabelUId();
2068 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2069 ARMConstantPoolValue *CPV =
2070 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2071 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2073 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2075 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2076 MachinePointerInfo::getConstantPool(),
2078 SDValue Chain = Result.getValue(1);
2080 if (RelocM == Reloc::PIC_) {
2081 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2082 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2085 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2086 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2092 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2093 SelectionDAG &DAG) const {
2094 assert(Subtarget->isTargetELF() &&
2095 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2096 MachineFunction &MF = DAG.getMachineFunction();
2097 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2098 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2099 EVT PtrVT = getPointerTy();
2100 DebugLoc dl = Op.getDebugLoc();
2101 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2102 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2103 "_GLOBAL_OFFSET_TABLE_",
2104 ARMPCLabelIndex, PCAdj);
2105 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2106 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2107 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2108 MachinePointerInfo::getConstantPool(),
2110 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2111 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2115 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2117 DebugLoc dl = Op.getDebugLoc();
2118 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2123 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2124 DebugLoc dl = Op.getDebugLoc();
2125 SDValue Val = DAG.getConstant(0, MVT::i32);
2126 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2127 Op.getOperand(1), Val);
2131 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2132 DebugLoc dl = Op.getDebugLoc();
2133 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2134 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2138 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2139 const ARMSubtarget *Subtarget) const {
2140 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2141 DebugLoc dl = Op.getDebugLoc();
2143 default: return SDValue(); // Don't custom lower most intrinsics.
2144 case Intrinsic::arm_thread_pointer: {
2145 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2146 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2148 case Intrinsic::eh_sjlj_lsda: {
2149 MachineFunction &MF = DAG.getMachineFunction();
2150 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2151 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2152 EVT PtrVT = getPointerTy();
2153 DebugLoc dl = Op.getDebugLoc();
2154 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2156 unsigned PCAdj = (RelocM != Reloc::PIC_)
2157 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2158 ARMConstantPoolValue *CPV =
2159 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2160 ARMCP::CPLSDA, PCAdj);
2161 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2162 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2164 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2165 MachinePointerInfo::getConstantPool(),
2168 if (RelocM == Reloc::PIC_) {
2169 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2170 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2174 case Intrinsic::arm_neon_vmulls:
2175 case Intrinsic::arm_neon_vmullu: {
2176 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2177 ? ARMISD::VMULLs : ARMISD::VMULLu;
2178 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2179 Op.getOperand(1), Op.getOperand(2));
2184 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2185 const ARMSubtarget *Subtarget) {
2186 DebugLoc dl = Op.getDebugLoc();
2187 if (!Subtarget->hasDataBarrier()) {
2188 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2189 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2191 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2192 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2193 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2194 DAG.getConstant(0, MVT::i32));
2197 SDValue Op5 = Op.getOperand(5);
2198 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2199 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2200 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2201 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2203 ARM_MB::MemBOpt DMBOpt;
2204 if (isDeviceBarrier)
2205 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2207 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2208 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2209 DAG.getConstant(DMBOpt, MVT::i32));
2212 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2213 const ARMSubtarget *Subtarget) {
2214 // ARM pre v5TE and Thumb1 does not have preload instructions.
2215 if (!(Subtarget->isThumb2() ||
2216 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2217 // Just preserve the chain.
2218 return Op.getOperand(0);
2220 DebugLoc dl = Op.getDebugLoc();
2221 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2223 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2224 // ARMv7 with MP extension has PLDW.
2225 return Op.getOperand(0);
2227 if (Subtarget->isThumb())
2229 isRead = ~isRead & 1;
2230 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2232 // Currently there is no intrinsic that matches pli.
2233 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2234 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2235 DAG.getConstant(isData, MVT::i32));
2238 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2239 MachineFunction &MF = DAG.getMachineFunction();
2240 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2242 // vastart just stores the address of the VarArgsFrameIndex slot into the
2243 // memory location argument.
2244 DebugLoc dl = Op.getDebugLoc();
2245 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2246 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2247 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2248 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2249 MachinePointerInfo(SV), false, false, 0);
2253 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2254 SDValue &Root, SelectionDAG &DAG,
2255 DebugLoc dl) const {
2256 MachineFunction &MF = DAG.getMachineFunction();
2257 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2259 TargetRegisterClass *RC;
2260 if (AFI->isThumb1OnlyFunction())
2261 RC = ARM::tGPRRegisterClass;
2263 RC = ARM::GPRRegisterClass;
2265 // Transform the arguments stored in physical registers into virtual ones.
2266 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2267 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2270 if (NextVA.isMemLoc()) {
2271 MachineFrameInfo *MFI = MF.getFrameInfo();
2272 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2274 // Create load node to retrieve arguments from the stack.
2275 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2276 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2277 MachinePointerInfo::getFixedStack(FI),
2280 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2281 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2284 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2288 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2289 CallingConv::ID CallConv, bool isVarArg,
2290 const SmallVectorImpl<ISD::InputArg>
2292 DebugLoc dl, SelectionDAG &DAG,
2293 SmallVectorImpl<SDValue> &InVals)
2296 MachineFunction &MF = DAG.getMachineFunction();
2297 MachineFrameInfo *MFI = MF.getFrameInfo();
2299 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2301 // Assign locations to all of the incoming arguments.
2302 SmallVector<CCValAssign, 16> ArgLocs;
2303 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2305 CCInfo.AnalyzeFormalArguments(Ins,
2306 CCAssignFnForNode(CallConv, /* Return*/ false,
2309 SmallVector<SDValue, 16> ArgValues;
2310 int lastInsIndex = -1;
2313 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2314 CCValAssign &VA = ArgLocs[i];
2316 // Arguments stored in registers.
2317 if (VA.isRegLoc()) {
2318 EVT RegVT = VA.getLocVT();
2320 if (VA.needsCustom()) {
2321 // f64 and vector types are split up into multiple registers or
2322 // combinations of registers and stack slots.
2323 if (VA.getLocVT() == MVT::v2f64) {
2324 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2326 VA = ArgLocs[++i]; // skip ahead to next loc
2328 if (VA.isMemLoc()) {
2329 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2330 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2331 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2332 MachinePointerInfo::getFixedStack(FI),
2335 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2338 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2339 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2340 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2341 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2342 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2344 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2347 TargetRegisterClass *RC;
2349 if (RegVT == MVT::f32)
2350 RC = ARM::SPRRegisterClass;
2351 else if (RegVT == MVT::f64)
2352 RC = ARM::DPRRegisterClass;
2353 else if (RegVT == MVT::v2f64)
2354 RC = ARM::QPRRegisterClass;
2355 else if (RegVT == MVT::i32)
2356 RC = (AFI->isThumb1OnlyFunction() ?
2357 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2359 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2361 // Transform the arguments in physical registers into virtual ones.
2362 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2363 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2366 // If this is an 8 or 16-bit value, it is really passed promoted
2367 // to 32 bits. Insert an assert[sz]ext to capture this, then
2368 // truncate to the right size.
2369 switch (VA.getLocInfo()) {
2370 default: llvm_unreachable("Unknown loc info!");
2371 case CCValAssign::Full: break;
2372 case CCValAssign::BCvt:
2373 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2375 case CCValAssign::SExt:
2376 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2377 DAG.getValueType(VA.getValVT()));
2378 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2380 case CCValAssign::ZExt:
2381 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2382 DAG.getValueType(VA.getValVT()));
2383 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2387 InVals.push_back(ArgValue);
2389 } else { // VA.isRegLoc()
2392 assert(VA.isMemLoc());
2393 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2395 int index = ArgLocs[i].getValNo();
2397 // Some Ins[] entries become multiple ArgLoc[] entries.
2398 // Process them only once.
2399 if (index != lastInsIndex)
2401 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2402 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2403 // changed with more analysis.
2404 // In case of tail call optimization mark all arguments mutable. Since they
2405 // could be overwritten by lowering of arguments in case of a tail call.
2406 if (Flags.isByVal()) {
2407 unsigned Bytes = Flags.getByValSize();
2408 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2409 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), false);
2410 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2412 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2413 VA.getLocMemOffset(), true);
2415 // Create load nodes to retrieve arguments from the stack.
2416 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2417 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2418 MachinePointerInfo::getFixedStack(FI),
2421 lastInsIndex = index;
2428 static const unsigned GPRArgRegs[] = {
2429 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2432 unsigned NumGPRs = CCInfo.getFirstUnallocated
2433 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2435 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2436 unsigned VARegSize = (4 - NumGPRs) * 4;
2437 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2438 unsigned ArgOffset = CCInfo.getNextStackOffset();
2439 if (VARegSaveSize) {
2440 // If this function is vararg, store any remaining integer argument regs
2441 // to their spots on the stack so that they may be loaded by deferencing
2442 // the result of va_next.
2443 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2444 AFI->setVarArgsFrameIndex(
2445 MFI->CreateFixedObject(VARegSaveSize,
2446 ArgOffset + VARegSaveSize - VARegSize,
2448 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2451 SmallVector<SDValue, 4> MemOps;
2452 for (; NumGPRs < 4; ++NumGPRs) {
2453 TargetRegisterClass *RC;
2454 if (AFI->isThumb1OnlyFunction())
2455 RC = ARM::tGPRRegisterClass;
2457 RC = ARM::GPRRegisterClass;
2459 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2460 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2462 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2463 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2465 MemOps.push_back(Store);
2466 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2467 DAG.getConstant(4, getPointerTy()));
2469 if (!MemOps.empty())
2470 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2471 &MemOps[0], MemOps.size());
2473 // This will point to the next argument passed via stack.
2474 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2480 /// isFloatingPointZero - Return true if this is +0.0.
2481 static bool isFloatingPointZero(SDValue Op) {
2482 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2483 return CFP->getValueAPF().isPosZero();
2484 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2485 // Maybe this has already been legalized into the constant pool?
2486 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2487 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2488 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2489 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2490 return CFP->getValueAPF().isPosZero();
2496 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2497 /// the given operands.
2499 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2500 SDValue &ARMcc, SelectionDAG &DAG,
2501 DebugLoc dl) const {
2502 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2503 unsigned C = RHSC->getZExtValue();
2504 if (!isLegalICmpImmediate(C)) {
2505 // Constant does not fit, try adjusting it by one?
2510 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2511 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2512 RHS = DAG.getConstant(C-1, MVT::i32);
2517 if (C != 0 && isLegalICmpImmediate(C-1)) {
2518 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2519 RHS = DAG.getConstant(C-1, MVT::i32);
2524 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2525 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2526 RHS = DAG.getConstant(C+1, MVT::i32);
2531 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2532 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2533 RHS = DAG.getConstant(C+1, MVT::i32);
2540 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2541 ARMISD::NodeType CompareType;
2544 CompareType = ARMISD::CMP;
2549 CompareType = ARMISD::CMPZ;
2552 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2553 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2556 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2558 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2559 DebugLoc dl) const {
2561 if (!isFloatingPointZero(RHS))
2562 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2564 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2565 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2568 /// duplicateCmp - Glue values can have only one use, so this function
2569 /// duplicates a comparison node.
2571 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2572 unsigned Opc = Cmp.getOpcode();
2573 DebugLoc DL = Cmp.getDebugLoc();
2574 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2575 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2577 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2578 Cmp = Cmp.getOperand(0);
2579 Opc = Cmp.getOpcode();
2580 if (Opc == ARMISD::CMPFP)
2581 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2583 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2584 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2586 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2589 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2590 SDValue Cond = Op.getOperand(0);
2591 SDValue SelectTrue = Op.getOperand(1);
2592 SDValue SelectFalse = Op.getOperand(2);
2593 DebugLoc dl = Op.getDebugLoc();
2597 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2598 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2600 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2601 const ConstantSDNode *CMOVTrue =
2602 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2603 const ConstantSDNode *CMOVFalse =
2604 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2606 if (CMOVTrue && CMOVFalse) {
2607 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2608 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2612 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2614 False = SelectFalse;
2615 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2620 if (True.getNode() && False.getNode()) {
2621 EVT VT = Cond.getValueType();
2622 SDValue ARMcc = Cond.getOperand(2);
2623 SDValue CCR = Cond.getOperand(3);
2624 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2625 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2630 return DAG.getSelectCC(dl, Cond,
2631 DAG.getConstant(0, Cond.getValueType()),
2632 SelectTrue, SelectFalse, ISD::SETNE);
2635 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2636 EVT VT = Op.getValueType();
2637 SDValue LHS = Op.getOperand(0);
2638 SDValue RHS = Op.getOperand(1);
2639 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2640 SDValue TrueVal = Op.getOperand(2);
2641 SDValue FalseVal = Op.getOperand(3);
2642 DebugLoc dl = Op.getDebugLoc();
2644 if (LHS.getValueType() == MVT::i32) {
2646 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2647 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2648 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2651 ARMCC::CondCodes CondCode, CondCode2;
2652 FPCCToARMCC(CC, CondCode, CondCode2);
2654 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2655 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2656 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2657 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2659 if (CondCode2 != ARMCC::AL) {
2660 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2661 // FIXME: Needs another CMP because flag can have but one use.
2662 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2663 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2664 Result, TrueVal, ARMcc2, CCR, Cmp2);
2669 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2670 /// to morph to an integer compare sequence.
2671 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2672 const ARMSubtarget *Subtarget) {
2673 SDNode *N = Op.getNode();
2674 if (!N->hasOneUse())
2675 // Otherwise it requires moving the value from fp to integer registers.
2677 if (!N->getNumValues())
2679 EVT VT = Op.getValueType();
2680 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2681 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2682 // vmrs are very slow, e.g. cortex-a8.
2685 if (isFloatingPointZero(Op)) {
2689 return ISD::isNormalLoad(N);
2692 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2693 if (isFloatingPointZero(Op))
2694 return DAG.getConstant(0, MVT::i32);
2696 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2697 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2698 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2699 Ld->isVolatile(), Ld->isNonTemporal(),
2700 Ld->getAlignment());
2702 llvm_unreachable("Unknown VFP cmp argument!");
2705 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2706 SDValue &RetVal1, SDValue &RetVal2) {
2707 if (isFloatingPointZero(Op)) {
2708 RetVal1 = DAG.getConstant(0, MVT::i32);
2709 RetVal2 = DAG.getConstant(0, MVT::i32);
2713 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2714 SDValue Ptr = Ld->getBasePtr();
2715 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2716 Ld->getChain(), Ptr,
2717 Ld->getPointerInfo(),
2718 Ld->isVolatile(), Ld->isNonTemporal(),
2719 Ld->getAlignment());
2721 EVT PtrType = Ptr.getValueType();
2722 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2723 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2724 PtrType, Ptr, DAG.getConstant(4, PtrType));
2725 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2726 Ld->getChain(), NewPtr,
2727 Ld->getPointerInfo().getWithOffset(4),
2728 Ld->isVolatile(), Ld->isNonTemporal(),
2733 llvm_unreachable("Unknown VFP cmp argument!");
2736 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2737 /// f32 and even f64 comparisons to integer ones.
2739 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2740 SDValue Chain = Op.getOperand(0);
2741 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2742 SDValue LHS = Op.getOperand(2);
2743 SDValue RHS = Op.getOperand(3);
2744 SDValue Dest = Op.getOperand(4);
2745 DebugLoc dl = Op.getDebugLoc();
2747 bool SeenZero = false;
2748 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2749 canChangeToInt(RHS, SeenZero, Subtarget) &&
2750 // If one of the operand is zero, it's safe to ignore the NaN case since
2751 // we only care about equality comparisons.
2752 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2753 // If unsafe fp math optimization is enabled and there are no other uses of
2754 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2755 // to an integer comparison.
2756 if (CC == ISD::SETOEQ)
2758 else if (CC == ISD::SETUNE)
2762 if (LHS.getValueType() == MVT::f32) {
2763 LHS = bitcastf32Toi32(LHS, DAG);
2764 RHS = bitcastf32Toi32(RHS, DAG);
2765 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2766 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2767 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2768 Chain, Dest, ARMcc, CCR, Cmp);
2773 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2774 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2775 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2776 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2777 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2778 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2779 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2785 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2786 SDValue Chain = Op.getOperand(0);
2787 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2788 SDValue LHS = Op.getOperand(2);
2789 SDValue RHS = Op.getOperand(3);
2790 SDValue Dest = Op.getOperand(4);
2791 DebugLoc dl = Op.getDebugLoc();
2793 if (LHS.getValueType() == MVT::i32) {
2795 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2796 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2797 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2798 Chain, Dest, ARMcc, CCR, Cmp);
2801 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2804 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2805 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2806 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2807 if (Result.getNode())
2811 ARMCC::CondCodes CondCode, CondCode2;
2812 FPCCToARMCC(CC, CondCode, CondCode2);
2814 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2815 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2816 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2817 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2818 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2819 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2820 if (CondCode2 != ARMCC::AL) {
2821 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2822 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2823 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2828 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2829 SDValue Chain = Op.getOperand(0);
2830 SDValue Table = Op.getOperand(1);
2831 SDValue Index = Op.getOperand(2);
2832 DebugLoc dl = Op.getDebugLoc();
2834 EVT PTy = getPointerTy();
2835 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2836 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2837 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2838 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2839 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2840 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2841 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2842 if (Subtarget->isThumb2()) {
2843 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2844 // which does another jump to the destination. This also makes it easier
2845 // to translate it to TBB / TBH later.
2846 // FIXME: This might not work if the function is extremely large.
2847 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2848 Addr, Op.getOperand(2), JTI, UId);
2850 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2851 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2852 MachinePointerInfo::getJumpTable(),
2854 Chain = Addr.getValue(1);
2855 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2856 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2858 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2859 MachinePointerInfo::getJumpTable(), false, false, 0);
2860 Chain = Addr.getValue(1);
2861 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2865 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2866 DebugLoc dl = Op.getDebugLoc();
2869 switch (Op.getOpcode()) {
2871 assert(0 && "Invalid opcode!");
2872 case ISD::FP_TO_SINT:
2873 Opc = ARMISD::FTOSI;
2875 case ISD::FP_TO_UINT:
2876 Opc = ARMISD::FTOUI;
2879 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2880 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2883 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2884 EVT VT = Op.getValueType();
2885 DebugLoc dl = Op.getDebugLoc();
2887 EVT OperandVT = Op.getOperand(0).getValueType();
2888 assert(OperandVT == MVT::v4i16 && "Invalid type for custom lowering!");
2889 if (VT != MVT::v4f32)
2890 return DAG.UnrollVectorOp(Op.getNode());
2894 switch (Op.getOpcode()) {
2896 assert(0 && "Invalid opcode!");
2897 case ISD::SINT_TO_FP:
2898 CastOpc = ISD::SIGN_EXTEND;
2899 Opc = ISD::SINT_TO_FP;
2901 case ISD::UINT_TO_FP:
2902 CastOpc = ISD::ZERO_EXTEND;
2903 Opc = ISD::UINT_TO_FP;
2907 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
2908 return DAG.getNode(Opc, dl, VT, Op);
2911 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2912 EVT VT = Op.getValueType();
2914 return LowerVectorINT_TO_FP(Op, DAG);
2916 DebugLoc dl = Op.getDebugLoc();
2919 switch (Op.getOpcode()) {
2921 assert(0 && "Invalid opcode!");
2922 case ISD::SINT_TO_FP:
2923 Opc = ARMISD::SITOF;
2925 case ISD::UINT_TO_FP:
2926 Opc = ARMISD::UITOF;
2930 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
2931 return DAG.getNode(Opc, dl, VT, Op);
2934 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2935 // Implement fcopysign with a fabs and a conditional fneg.
2936 SDValue Tmp0 = Op.getOperand(0);
2937 SDValue Tmp1 = Op.getOperand(1);
2938 DebugLoc dl = Op.getDebugLoc();
2939 EVT VT = Op.getValueType();
2940 EVT SrcVT = Tmp1.getValueType();
2941 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
2942 Tmp0.getOpcode() == ARMISD::VMOVDRR;
2943 bool UseNEON = !InGPR && Subtarget->hasNEON();
2946 // Use VBSL to copy the sign bit.
2947 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
2948 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
2949 DAG.getTargetConstant(EncodedVal, MVT::i32));
2950 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
2952 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
2953 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
2954 DAG.getConstant(32, MVT::i32));
2955 else /*if (VT == MVT::f32)*/
2956 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
2957 if (SrcVT == MVT::f32) {
2958 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
2960 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
2961 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
2962 DAG.getConstant(32, MVT::i32));
2964 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
2965 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
2967 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
2969 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
2970 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
2971 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
2973 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
2974 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
2975 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
2976 if (VT == MVT::f32) {
2977 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
2978 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
2979 DAG.getConstant(0, MVT::i32));
2981 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
2987 // Bitcast operand 1 to i32.
2988 if (SrcVT == MVT::f64)
2989 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
2990 &Tmp1, 1).getValue(1);
2991 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
2993 // Or in the signbit with integer operations.
2994 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
2995 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
2996 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
2997 if (VT == MVT::f32) {
2998 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
2999 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3000 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3001 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3004 // f64: Or the high part with signbit and then combine two parts.
3005 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3007 SDValue Lo = Tmp0.getValue(0);
3008 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3009 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3010 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3013 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3014 MachineFunction &MF = DAG.getMachineFunction();
3015 MachineFrameInfo *MFI = MF.getFrameInfo();
3016 MFI->setReturnAddressIsTaken(true);
3018 EVT VT = Op.getValueType();
3019 DebugLoc dl = Op.getDebugLoc();
3020 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3022 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3023 SDValue Offset = DAG.getConstant(4, MVT::i32);
3024 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3025 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3026 MachinePointerInfo(), false, false, 0);
3029 // Return LR, which contains the return address. Mark it an implicit live-in.
3030 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3031 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3034 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3035 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3036 MFI->setFrameAddressIsTaken(true);
3038 EVT VT = Op.getValueType();
3039 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3040 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3041 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3042 ? ARM::R7 : ARM::R11;
3043 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3045 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3046 MachinePointerInfo(),
3051 /// ExpandBITCAST - If the target supports VFP, this function is called to
3052 /// expand a bit convert where either the source or destination type is i64 to
3053 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3054 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3055 /// vectors), since the legalizer won't know what to do with that.
3056 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3057 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3058 DebugLoc dl = N->getDebugLoc();
3059 SDValue Op = N->getOperand(0);
3061 // This function is only supposed to be called for i64 types, either as the
3062 // source or destination of the bit convert.
3063 EVT SrcVT = Op.getValueType();
3064 EVT DstVT = N->getValueType(0);
3065 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3066 "ExpandBITCAST called for non-i64 type");
3068 // Turn i64->f64 into VMOVDRR.
3069 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3070 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3071 DAG.getConstant(0, MVT::i32));
3072 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3073 DAG.getConstant(1, MVT::i32));
3074 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3075 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3078 // Turn f64->i64 into VMOVRRD.
3079 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3080 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3081 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3082 // Merge the pieces into a single i64 value.
3083 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3089 /// getZeroVector - Returns a vector of specified type with all zero elements.
3090 /// Zero vectors are used to represent vector negation and in those cases
3091 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3092 /// not support i64 elements, so sometimes the zero vectors will need to be
3093 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3095 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3096 assert(VT.isVector() && "Expected a vector type");
3097 // The canonical modified immediate encoding of a zero vector is....0!
3098 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3099 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3100 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3101 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3104 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3105 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3106 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3107 SelectionDAG &DAG) const {
3108 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3109 EVT VT = Op.getValueType();
3110 unsigned VTBits = VT.getSizeInBits();
3111 DebugLoc dl = Op.getDebugLoc();
3112 SDValue ShOpLo = Op.getOperand(0);
3113 SDValue ShOpHi = Op.getOperand(1);
3114 SDValue ShAmt = Op.getOperand(2);
3116 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3118 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3120 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3121 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3122 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3123 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3124 DAG.getConstant(VTBits, MVT::i32));
3125 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3126 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3127 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3129 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3130 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3132 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3133 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3136 SDValue Ops[2] = { Lo, Hi };
3137 return DAG.getMergeValues(Ops, 2, dl);
3140 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3141 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3142 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3143 SelectionDAG &DAG) const {
3144 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3145 EVT VT = Op.getValueType();
3146 unsigned VTBits = VT.getSizeInBits();
3147 DebugLoc dl = Op.getDebugLoc();
3148 SDValue ShOpLo = Op.getOperand(0);
3149 SDValue ShOpHi = Op.getOperand(1);
3150 SDValue ShAmt = Op.getOperand(2);
3153 assert(Op.getOpcode() == ISD::SHL_PARTS);
3154 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3155 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3156 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3157 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3158 DAG.getConstant(VTBits, MVT::i32));
3159 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3160 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3162 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3163 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3164 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3166 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3167 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3170 SDValue Ops[2] = { Lo, Hi };
3171 return DAG.getMergeValues(Ops, 2, dl);
3174 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3175 SelectionDAG &DAG) const {
3176 // The rounding mode is in bits 23:22 of the FPSCR.
3177 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3178 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3179 // so that the shift + and get folded into a bitfield extract.
3180 DebugLoc dl = Op.getDebugLoc();
3181 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3182 DAG.getConstant(Intrinsic::arm_get_fpscr,
3184 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3185 DAG.getConstant(1U << 22, MVT::i32));
3186 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3187 DAG.getConstant(22, MVT::i32));
3188 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3189 DAG.getConstant(3, MVT::i32));
3192 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3193 const ARMSubtarget *ST) {
3194 EVT VT = N->getValueType(0);
3195 DebugLoc dl = N->getDebugLoc();
3197 if (!ST->hasV6T2Ops())
3200 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3201 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3204 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3205 const ARMSubtarget *ST) {
3206 EVT VT = N->getValueType(0);
3207 DebugLoc dl = N->getDebugLoc();
3212 // Lower vector shifts on NEON to use VSHL.
3213 assert(ST->hasNEON() && "unexpected vector shift");
3215 // Left shifts translate directly to the vshiftu intrinsic.
3216 if (N->getOpcode() == ISD::SHL)
3217 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3218 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3219 N->getOperand(0), N->getOperand(1));
3221 assert((N->getOpcode() == ISD::SRA ||
3222 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3224 // NEON uses the same intrinsics for both left and right shifts. For
3225 // right shifts, the shift amounts are negative, so negate the vector of
3227 EVT ShiftVT = N->getOperand(1).getValueType();
3228 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3229 getZeroVector(ShiftVT, DAG, dl),
3231 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3232 Intrinsic::arm_neon_vshifts :
3233 Intrinsic::arm_neon_vshiftu);
3234 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3235 DAG.getConstant(vshiftInt, MVT::i32),
3236 N->getOperand(0), NegatedCount);
3239 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3240 const ARMSubtarget *ST) {
3241 EVT VT = N->getValueType(0);
3242 DebugLoc dl = N->getDebugLoc();
3244 // We can get here for a node like i32 = ISD::SHL i32, i64
3248 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3249 "Unknown shift to lower!");
3251 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3252 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3253 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3256 // If we are in thumb mode, we don't have RRX.
3257 if (ST->isThumb1Only()) return SDValue();
3259 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3260 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3261 DAG.getConstant(0, MVT::i32));
3262 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3263 DAG.getConstant(1, MVT::i32));
3265 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3266 // captures the result into a carry flag.
3267 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3268 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3270 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3271 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3273 // Merge the pieces into a single i64 value.
3274 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3277 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3278 SDValue TmpOp0, TmpOp1;
3279 bool Invert = false;
3283 SDValue Op0 = Op.getOperand(0);
3284 SDValue Op1 = Op.getOperand(1);
3285 SDValue CC = Op.getOperand(2);
3286 EVT VT = Op.getValueType();
3287 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3288 DebugLoc dl = Op.getDebugLoc();
3290 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3291 switch (SetCCOpcode) {
3292 default: llvm_unreachable("Illegal FP comparison"); break;
3294 case ISD::SETNE: Invert = true; // Fallthrough
3296 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3298 case ISD::SETLT: Swap = true; // Fallthrough
3300 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3302 case ISD::SETLE: Swap = true; // Fallthrough
3304 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3305 case ISD::SETUGE: Swap = true; // Fallthrough
3306 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3307 case ISD::SETUGT: Swap = true; // Fallthrough
3308 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3309 case ISD::SETUEQ: Invert = true; // Fallthrough
3311 // Expand this to (OLT | OGT).
3315 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3316 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3318 case ISD::SETUO: Invert = true; // Fallthrough
3320 // Expand this to (OLT | OGE).
3324 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3325 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3329 // Integer comparisons.
3330 switch (SetCCOpcode) {
3331 default: llvm_unreachable("Illegal integer comparison"); break;
3332 case ISD::SETNE: Invert = true;
3333 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3334 case ISD::SETLT: Swap = true;
3335 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3336 case ISD::SETLE: Swap = true;
3337 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3338 case ISD::SETULT: Swap = true;
3339 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3340 case ISD::SETULE: Swap = true;
3341 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3344 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3345 if (Opc == ARMISD::VCEQ) {
3348 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3350 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3353 // Ignore bitconvert.
3354 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3355 AndOp = AndOp.getOperand(0);
3357 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3359 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3360 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3367 std::swap(Op0, Op1);
3369 // If one of the operands is a constant vector zero, attempt to fold the
3370 // comparison to a specialized compare-against-zero form.
3372 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3374 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3375 if (Opc == ARMISD::VCGE)
3376 Opc = ARMISD::VCLEZ;
3377 else if (Opc == ARMISD::VCGT)
3378 Opc = ARMISD::VCLTZ;
3383 if (SingleOp.getNode()) {
3386 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3388 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3390 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3392 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3394 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3396 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3399 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3403 Result = DAG.getNOT(dl, Result, VT);
3408 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3409 /// valid vector constant for a NEON instruction with a "modified immediate"
3410 /// operand (e.g., VMOV). If so, return the encoded value.
3411 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3412 unsigned SplatBitSize, SelectionDAG &DAG,
3413 EVT &VT, bool is128Bits, NEONModImmType type) {
3414 unsigned OpCmode, Imm;
3416 // SplatBitSize is set to the smallest size that splats the vector, so a
3417 // zero vector will always have SplatBitSize == 8. However, NEON modified
3418 // immediate instructions others than VMOV do not support the 8-bit encoding
3419 // of a zero vector, and the default encoding of zero is supposed to be the
3424 switch (SplatBitSize) {
3426 if (type != VMOVModImm)
3428 // Any 1-byte value is OK. Op=0, Cmode=1110.
3429 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3432 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3436 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3437 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3438 if ((SplatBits & ~0xff) == 0) {
3439 // Value = 0x00nn: Op=x, Cmode=100x.
3444 if ((SplatBits & ~0xff00) == 0) {
3445 // Value = 0xnn00: Op=x, Cmode=101x.
3447 Imm = SplatBits >> 8;
3453 // NEON's 32-bit VMOV supports splat values where:
3454 // * only one byte is nonzero, or
3455 // * the least significant byte is 0xff and the second byte is nonzero, or
3456 // * the least significant 2 bytes are 0xff and the third is nonzero.
3457 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3458 if ((SplatBits & ~0xff) == 0) {
3459 // Value = 0x000000nn: Op=x, Cmode=000x.
3464 if ((SplatBits & ~0xff00) == 0) {
3465 // Value = 0x0000nn00: Op=x, Cmode=001x.
3467 Imm = SplatBits >> 8;
3470 if ((SplatBits & ~0xff0000) == 0) {
3471 // Value = 0x00nn0000: Op=x, Cmode=010x.
3473 Imm = SplatBits >> 16;
3476 if ((SplatBits & ~0xff000000) == 0) {
3477 // Value = 0xnn000000: Op=x, Cmode=011x.
3479 Imm = SplatBits >> 24;
3483 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3484 if (type == OtherModImm) return SDValue();
3486 if ((SplatBits & ~0xffff) == 0 &&
3487 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3488 // Value = 0x0000nnff: Op=x, Cmode=1100.
3490 Imm = SplatBits >> 8;
3495 if ((SplatBits & ~0xffffff) == 0 &&
3496 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3497 // Value = 0x00nnffff: Op=x, Cmode=1101.
3499 Imm = SplatBits >> 16;
3500 SplatBits |= 0xffff;
3504 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3505 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3506 // VMOV.I32. A (very) minor optimization would be to replicate the value
3507 // and fall through here to test for a valid 64-bit splat. But, then the
3508 // caller would also need to check and handle the change in size.
3512 if (type != VMOVModImm)
3514 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3515 uint64_t BitMask = 0xff;
3517 unsigned ImmMask = 1;
3519 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3520 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3523 } else if ((SplatBits & BitMask) != 0) {
3529 // Op=1, Cmode=1110.
3532 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3537 llvm_unreachable("unexpected size for isNEONModifiedImm");
3541 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3542 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3545 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3546 bool &ReverseVEXT, unsigned &Imm) {
3547 unsigned NumElts = VT.getVectorNumElements();
3548 ReverseVEXT = false;
3550 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3556 // If this is a VEXT shuffle, the immediate value is the index of the first
3557 // element. The other shuffle indices must be the successive elements after
3559 unsigned ExpectedElt = Imm;
3560 for (unsigned i = 1; i < NumElts; ++i) {
3561 // Increment the expected index. If it wraps around, it may still be
3562 // a VEXT but the source vectors must be swapped.
3564 if (ExpectedElt == NumElts * 2) {
3569 if (M[i] < 0) continue; // ignore UNDEF indices
3570 if (ExpectedElt != static_cast<unsigned>(M[i]))
3574 // Adjust the index value if the source operands will be swapped.
3581 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3582 /// instruction with the specified blocksize. (The order of the elements
3583 /// within each block of the vector is reversed.)
3584 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3585 unsigned BlockSize) {
3586 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3587 "Only possible block sizes for VREV are: 16, 32, 64");
3589 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3593 unsigned NumElts = VT.getVectorNumElements();
3594 unsigned BlockElts = M[0] + 1;
3595 // If the first shuffle index is UNDEF, be optimistic.
3597 BlockElts = BlockSize / EltSz;
3599 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3602 for (unsigned i = 0; i < NumElts; ++i) {
3603 if (M[i] < 0) continue; // ignore UNDEF indices
3604 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3611 static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3612 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3613 // range, then 0 is placed into the resulting vector. So pretty much any mask
3614 // of 8 elements can work here.
3615 return VT == MVT::v8i8 && M.size() == 8;
3618 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3619 unsigned &WhichResult) {
3620 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3624 unsigned NumElts = VT.getVectorNumElements();
3625 WhichResult = (M[0] == 0 ? 0 : 1);
3626 for (unsigned i = 0; i < NumElts; i += 2) {
3627 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3628 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3634 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3635 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3636 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3637 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3638 unsigned &WhichResult) {
3639 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3643 unsigned NumElts = VT.getVectorNumElements();
3644 WhichResult = (M[0] == 0 ? 0 : 1);
3645 for (unsigned i = 0; i < NumElts; i += 2) {
3646 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3647 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3653 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3654 unsigned &WhichResult) {
3655 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3659 unsigned NumElts = VT.getVectorNumElements();
3660 WhichResult = (M[0] == 0 ? 0 : 1);
3661 for (unsigned i = 0; i != NumElts; ++i) {
3662 if (M[i] < 0) continue; // ignore UNDEF indices
3663 if ((unsigned) M[i] != 2 * i + WhichResult)
3667 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3668 if (VT.is64BitVector() && EltSz == 32)
3674 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3675 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3676 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3677 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3678 unsigned &WhichResult) {
3679 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3683 unsigned Half = VT.getVectorNumElements() / 2;
3684 WhichResult = (M[0] == 0 ? 0 : 1);
3685 for (unsigned j = 0; j != 2; ++j) {
3686 unsigned Idx = WhichResult;
3687 for (unsigned i = 0; i != Half; ++i) {
3688 int MIdx = M[i + j * Half];
3689 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3695 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3696 if (VT.is64BitVector() && EltSz == 32)
3702 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3703 unsigned &WhichResult) {
3704 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3708 unsigned NumElts = VT.getVectorNumElements();
3709 WhichResult = (M[0] == 0 ? 0 : 1);
3710 unsigned Idx = WhichResult * NumElts / 2;
3711 for (unsigned i = 0; i != NumElts; i += 2) {
3712 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3713 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3718 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3719 if (VT.is64BitVector() && EltSz == 32)
3725 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3726 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3727 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3728 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3729 unsigned &WhichResult) {
3730 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3734 unsigned NumElts = VT.getVectorNumElements();
3735 WhichResult = (M[0] == 0 ? 0 : 1);
3736 unsigned Idx = WhichResult * NumElts / 2;
3737 for (unsigned i = 0; i != NumElts; i += 2) {
3738 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3739 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3744 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3745 if (VT.is64BitVector() && EltSz == 32)
3751 // If N is an integer constant that can be moved into a register in one
3752 // instruction, return an SDValue of such a constant (will become a MOV
3753 // instruction). Otherwise return null.
3754 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3755 const ARMSubtarget *ST, DebugLoc dl) {
3757 if (!isa<ConstantSDNode>(N))
3759 Val = cast<ConstantSDNode>(N)->getZExtValue();
3761 if (ST->isThumb1Only()) {
3762 if (Val <= 255 || ~Val <= 255)
3763 return DAG.getConstant(Val, MVT::i32);
3765 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3766 return DAG.getConstant(Val, MVT::i32);
3771 // If this is a case we can't handle, return null and let the default
3772 // expansion code take care of it.
3773 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3774 const ARMSubtarget *ST) const {
3775 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3776 DebugLoc dl = Op.getDebugLoc();
3777 EVT VT = Op.getValueType();
3779 APInt SplatBits, SplatUndef;
3780 unsigned SplatBitSize;
3782 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3783 if (SplatBitSize <= 64) {
3784 // Check if an immediate VMOV works.
3786 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3787 SplatUndef.getZExtValue(), SplatBitSize,
3788 DAG, VmovVT, VT.is128BitVector(),
3790 if (Val.getNode()) {
3791 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3792 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3795 // Try an immediate VMVN.
3796 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3797 ((1LL << SplatBitSize) - 1));
3798 Val = isNEONModifiedImm(NegatedImm,
3799 SplatUndef.getZExtValue(), SplatBitSize,
3800 DAG, VmovVT, VT.is128BitVector(),
3802 if (Val.getNode()) {
3803 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3804 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3809 // Scan through the operands to see if only one value is used.
3810 unsigned NumElts = VT.getVectorNumElements();
3811 bool isOnlyLowElement = true;
3812 bool usesOnlyOneValue = true;
3813 bool isConstant = true;
3815 for (unsigned i = 0; i < NumElts; ++i) {
3816 SDValue V = Op.getOperand(i);
3817 if (V.getOpcode() == ISD::UNDEF)
3820 isOnlyLowElement = false;
3821 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3824 if (!Value.getNode())
3826 else if (V != Value)
3827 usesOnlyOneValue = false;
3830 if (!Value.getNode())
3831 return DAG.getUNDEF(VT);
3833 if (isOnlyLowElement)
3834 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3836 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3838 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3839 // i32 and try again.
3840 if (usesOnlyOneValue && EltSize <= 32) {
3842 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3843 if (VT.getVectorElementType().isFloatingPoint()) {
3844 SmallVector<SDValue, 8> Ops;
3845 for (unsigned i = 0; i < NumElts; ++i)
3846 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3848 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3849 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3850 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3852 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3854 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3856 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3859 // If all elements are constants and the case above didn't get hit, fall back
3860 // to the default expansion, which will generate a load from the constant
3865 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3867 SDValue shuffle = ReconstructShuffle(Op, DAG);
3868 if (shuffle != SDValue())
3872 // Vectors with 32- or 64-bit elements can be built by directly assigning
3873 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3874 // will be legalized.
3875 if (EltSize >= 32) {
3876 // Do the expansion with floating-point types, since that is what the VFP
3877 // registers are defined to use, and since i64 is not legal.
3878 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3879 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3880 SmallVector<SDValue, 8> Ops;
3881 for (unsigned i = 0; i < NumElts; ++i)
3882 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3883 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3884 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3890 // Gather data to see if the operation can be modelled as a
3891 // shuffle in combination with VEXTs.
3892 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
3893 SelectionDAG &DAG) const {
3894 DebugLoc dl = Op.getDebugLoc();
3895 EVT VT = Op.getValueType();
3896 unsigned NumElts = VT.getVectorNumElements();
3898 SmallVector<SDValue, 2> SourceVecs;
3899 SmallVector<unsigned, 2> MinElts;
3900 SmallVector<unsigned, 2> MaxElts;
3902 for (unsigned i = 0; i < NumElts; ++i) {
3903 SDValue V = Op.getOperand(i);
3904 if (V.getOpcode() == ISD::UNDEF)
3906 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3907 // A shuffle can only come from building a vector from various
3908 // elements of other vectors.
3912 // Record this extraction against the appropriate vector if possible...
3913 SDValue SourceVec = V.getOperand(0);
3914 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3915 bool FoundSource = false;
3916 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3917 if (SourceVecs[j] == SourceVec) {
3918 if (MinElts[j] > EltNo)
3920 if (MaxElts[j] < EltNo)
3927 // Or record a new source if not...
3929 SourceVecs.push_back(SourceVec);
3930 MinElts.push_back(EltNo);
3931 MaxElts.push_back(EltNo);
3935 // Currently only do something sane when at most two source vectors
3937 if (SourceVecs.size() > 2)
3940 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3941 int VEXTOffsets[2] = {0, 0};
3943 // This loop extracts the usage patterns of the source vectors
3944 // and prepares appropriate SDValues for a shuffle if possible.
3945 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3946 if (SourceVecs[i].getValueType() == VT) {
3947 // No VEXT necessary
3948 ShuffleSrcs[i] = SourceVecs[i];
3951 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3952 // It probably isn't worth padding out a smaller vector just to
3953 // break it down again in a shuffle.
3957 // Since only 64-bit and 128-bit vectors are legal on ARM and
3958 // we've eliminated the other cases...
3959 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
3960 "unexpected vector sizes in ReconstructShuffle");
3962 if (MaxElts[i] - MinElts[i] >= NumElts) {
3963 // Span too large for a VEXT to cope
3967 if (MinElts[i] >= NumElts) {
3968 // The extraction can just take the second half
3969 VEXTOffsets[i] = NumElts;
3970 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3972 DAG.getIntPtrConstant(NumElts));
3973 } else if (MaxElts[i] < NumElts) {
3974 // The extraction can just take the first half
3976 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3978 DAG.getIntPtrConstant(0));
3980 // An actual VEXT is needed
3981 VEXTOffsets[i] = MinElts[i];
3982 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3984 DAG.getIntPtrConstant(0));
3985 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3987 DAG.getIntPtrConstant(NumElts));
3988 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
3989 DAG.getConstant(VEXTOffsets[i], MVT::i32));
3993 SmallVector<int, 8> Mask;
3995 for (unsigned i = 0; i < NumElts; ++i) {
3996 SDValue Entry = Op.getOperand(i);
3997 if (Entry.getOpcode() == ISD::UNDEF) {
4002 SDValue ExtractVec = Entry.getOperand(0);
4003 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4004 .getOperand(1))->getSExtValue();
4005 if (ExtractVec == SourceVecs[0]) {
4006 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4008 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4012 // Final check before we try to produce nonsense...
4013 if (isShuffleMaskLegal(Mask, VT))
4014 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4020 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4021 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4022 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4023 /// are assumed to be legal.
4025 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4027 if (VT.getVectorNumElements() == 4 &&
4028 (VT.is128BitVector() || VT.is64BitVector())) {
4029 unsigned PFIndexes[4];
4030 for (unsigned i = 0; i != 4; ++i) {
4034 PFIndexes[i] = M[i];
4037 // Compute the index in the perfect shuffle table.
4038 unsigned PFTableIndex =
4039 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4040 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4041 unsigned Cost = (PFEntry >> 30);
4048 unsigned Imm, WhichResult;
4050 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4051 return (EltSize >= 32 ||
4052 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4053 isVREVMask(M, VT, 64) ||
4054 isVREVMask(M, VT, 32) ||
4055 isVREVMask(M, VT, 16) ||
4056 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4057 isVTBLMask(M, VT) ||
4058 isVTRNMask(M, VT, WhichResult) ||
4059 isVUZPMask(M, VT, WhichResult) ||
4060 isVZIPMask(M, VT, WhichResult) ||
4061 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4062 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4063 isVZIP_v_undef_Mask(M, VT, WhichResult));
4066 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4067 /// the specified operations to build the shuffle.
4068 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4069 SDValue RHS, SelectionDAG &DAG,
4071 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4072 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4073 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4076 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4085 OP_VUZPL, // VUZP, left result
4086 OP_VUZPR, // VUZP, right result
4087 OP_VZIPL, // VZIP, left result
4088 OP_VZIPR, // VZIP, right result
4089 OP_VTRNL, // VTRN, left result
4090 OP_VTRNR // VTRN, right result
4093 if (OpNum == OP_COPY) {
4094 if (LHSID == (1*9+2)*9+3) return LHS;
4095 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4099 SDValue OpLHS, OpRHS;
4100 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4101 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4102 EVT VT = OpLHS.getValueType();
4105 default: llvm_unreachable("Unknown shuffle opcode!");
4107 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4112 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4113 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4117 return DAG.getNode(ARMISD::VEXT, dl, VT,
4119 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4122 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4123 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4126 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4127 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4130 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4131 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4135 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4136 SmallVectorImpl<int> &ShuffleMask,
4137 SelectionDAG &DAG) {
4138 // Check to see if we can use the VTBL instruction.
4139 SDValue V1 = Op.getOperand(0);
4140 SDValue V2 = Op.getOperand(1);
4141 DebugLoc DL = Op.getDebugLoc();
4143 SmallVector<SDValue, 8> VTBLMask;
4144 for (SmallVectorImpl<int>::iterator
4145 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4146 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4148 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4149 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4150 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4153 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4154 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4158 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4159 SDValue V1 = Op.getOperand(0);
4160 SDValue V2 = Op.getOperand(1);
4161 DebugLoc dl = Op.getDebugLoc();
4162 EVT VT = Op.getValueType();
4163 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4164 SmallVector<int, 8> ShuffleMask;
4166 // Convert shuffles that are directly supported on NEON to target-specific
4167 // DAG nodes, instead of keeping them as shuffles and matching them again
4168 // during code selection. This is more efficient and avoids the possibility
4169 // of inconsistencies between legalization and selection.
4170 // FIXME: floating-point vectors should be canonicalized to integer vectors
4171 // of the same time so that they get CSEd properly.
4172 SVN->getMask(ShuffleMask);
4174 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4175 if (EltSize <= 32) {
4176 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4177 int Lane = SVN->getSplatIndex();
4178 // If this is undef splat, generate it via "just" vdup, if possible.
4179 if (Lane == -1) Lane = 0;
4181 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4182 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4184 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4185 DAG.getConstant(Lane, MVT::i32));
4190 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4193 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4194 DAG.getConstant(Imm, MVT::i32));
4197 if (isVREVMask(ShuffleMask, VT, 64))
4198 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4199 if (isVREVMask(ShuffleMask, VT, 32))
4200 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4201 if (isVREVMask(ShuffleMask, VT, 16))
4202 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4204 // Check for Neon shuffles that modify both input vectors in place.
4205 // If both results are used, i.e., if there are two shuffles with the same
4206 // source operands and with masks corresponding to both results of one of
4207 // these operations, DAG memoization will ensure that a single node is
4208 // used for both shuffles.
4209 unsigned WhichResult;
4210 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4211 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4212 V1, V2).getValue(WhichResult);
4213 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4214 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4215 V1, V2).getValue(WhichResult);
4216 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4217 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4218 V1, V2).getValue(WhichResult);
4220 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4221 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4222 V1, V1).getValue(WhichResult);
4223 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4224 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4225 V1, V1).getValue(WhichResult);
4226 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4227 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4228 V1, V1).getValue(WhichResult);
4231 // If the shuffle is not directly supported and it has 4 elements, use
4232 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4233 unsigned NumElts = VT.getVectorNumElements();
4235 unsigned PFIndexes[4];
4236 for (unsigned i = 0; i != 4; ++i) {
4237 if (ShuffleMask[i] < 0)
4240 PFIndexes[i] = ShuffleMask[i];
4243 // Compute the index in the perfect shuffle table.
4244 unsigned PFTableIndex =
4245 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4246 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4247 unsigned Cost = (PFEntry >> 30);
4250 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4253 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4254 if (EltSize >= 32) {
4255 // Do the expansion with floating-point types, since that is what the VFP
4256 // registers are defined to use, and since i64 is not legal.
4257 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4258 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4259 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4260 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4261 SmallVector<SDValue, 8> Ops;
4262 for (unsigned i = 0; i < NumElts; ++i) {
4263 if (ShuffleMask[i] < 0)
4264 Ops.push_back(DAG.getUNDEF(EltVT));
4266 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4267 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4268 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4271 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4272 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4275 if (VT == MVT::v8i8) {
4276 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4277 if (NewOp.getNode())
4284 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4285 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4286 SDValue Lane = Op.getOperand(1);
4287 if (!isa<ConstantSDNode>(Lane))
4290 SDValue Vec = Op.getOperand(0);
4291 if (Op.getValueType() == MVT::i32 &&
4292 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4293 DebugLoc dl = Op.getDebugLoc();
4294 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4300 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4301 // The only time a CONCAT_VECTORS operation can have legal types is when
4302 // two 64-bit vectors are concatenated to a 128-bit vector.
4303 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4304 "unexpected CONCAT_VECTORS");
4305 DebugLoc dl = Op.getDebugLoc();
4306 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4307 SDValue Op0 = Op.getOperand(0);
4308 SDValue Op1 = Op.getOperand(1);
4309 if (Op0.getOpcode() != ISD::UNDEF)
4310 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4311 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4312 DAG.getIntPtrConstant(0));
4313 if (Op1.getOpcode() != ISD::UNDEF)
4314 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4315 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4316 DAG.getIntPtrConstant(1));
4317 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4320 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4321 /// element has been zero/sign-extended, depending on the isSigned parameter,
4322 /// from an integer type half its size.
4323 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4325 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4326 EVT VT = N->getValueType(0);
4327 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4328 SDNode *BVN = N->getOperand(0).getNode();
4329 if (BVN->getValueType(0) != MVT::v4i32 ||
4330 BVN->getOpcode() != ISD::BUILD_VECTOR)
4332 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4333 unsigned HiElt = 1 - LoElt;
4334 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4335 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4336 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4337 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4338 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4341 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4342 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4345 if (Hi0->isNullValue() && Hi1->isNullValue())
4351 if (N->getOpcode() != ISD::BUILD_VECTOR)
4354 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4355 SDNode *Elt = N->getOperand(i).getNode();
4356 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4357 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4358 unsigned HalfSize = EltSize / 2;
4360 int64_t SExtVal = C->getSExtValue();
4361 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4364 if ((C->getZExtValue() >> HalfSize) != 0)
4375 /// isSignExtended - Check if a node is a vector value that is sign-extended
4376 /// or a constant BUILD_VECTOR with sign-extended elements.
4377 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4378 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4380 if (isExtendedBUILD_VECTOR(N, DAG, true))
4385 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4386 /// or a constant BUILD_VECTOR with zero-extended elements.
4387 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4388 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4390 if (isExtendedBUILD_VECTOR(N, DAG, false))
4395 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4396 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4397 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4398 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4399 return N->getOperand(0);
4400 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4401 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4402 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4403 LD->isNonTemporal(), LD->getAlignment());
4404 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4405 // have been legalized as a BITCAST from v4i32.
4406 if (N->getOpcode() == ISD::BITCAST) {
4407 SDNode *BVN = N->getOperand(0).getNode();
4408 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4409 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4410 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4411 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4412 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4414 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4415 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4416 EVT VT = N->getValueType(0);
4417 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4418 unsigned NumElts = VT.getVectorNumElements();
4419 MVT TruncVT = MVT::getIntegerVT(EltSize);
4420 SmallVector<SDValue, 8> Ops;
4421 for (unsigned i = 0; i != NumElts; ++i) {
4422 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4423 const APInt &CInt = C->getAPIntValue();
4424 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4426 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4427 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4430 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4431 unsigned Opcode = N->getOpcode();
4432 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4433 SDNode *N0 = N->getOperand(0).getNode();
4434 SDNode *N1 = N->getOperand(1).getNode();
4435 return N0->hasOneUse() && N1->hasOneUse() &&
4436 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4441 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4442 unsigned Opcode = N->getOpcode();
4443 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4444 SDNode *N0 = N->getOperand(0).getNode();
4445 SDNode *N1 = N->getOperand(1).getNode();
4446 return N0->hasOneUse() && N1->hasOneUse() &&
4447 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4452 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4453 // Multiplications are only custom-lowered for 128-bit vectors so that
4454 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4455 EVT VT = Op.getValueType();
4456 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4457 SDNode *N0 = Op.getOperand(0).getNode();
4458 SDNode *N1 = Op.getOperand(1).getNode();
4459 unsigned NewOpc = 0;
4461 bool isN0SExt = isSignExtended(N0, DAG);
4462 bool isN1SExt = isSignExtended(N1, DAG);
4463 if (isN0SExt && isN1SExt)
4464 NewOpc = ARMISD::VMULLs;
4466 bool isN0ZExt = isZeroExtended(N0, DAG);
4467 bool isN1ZExt = isZeroExtended(N1, DAG);
4468 if (isN0ZExt && isN1ZExt)
4469 NewOpc = ARMISD::VMULLu;
4470 else if (isN1SExt || isN1ZExt) {
4471 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4472 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4473 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4474 NewOpc = ARMISD::VMULLs;
4476 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4477 NewOpc = ARMISD::VMULLu;
4479 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4481 NewOpc = ARMISD::VMULLu;
4487 if (VT == MVT::v2i64)
4488 // Fall through to expand this. It is not legal.
4491 // Other vector multiplications are legal.
4496 // Legalize to a VMULL instruction.
4497 DebugLoc DL = Op.getDebugLoc();
4499 SDValue Op1 = SkipExtension(N1, DAG);
4501 Op0 = SkipExtension(N0, DAG);
4502 assert(Op0.getValueType().is64BitVector() &&
4503 Op1.getValueType().is64BitVector() &&
4504 "unexpected types for extended operands to VMULL");
4505 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4508 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4509 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4516 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4517 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4518 EVT Op1VT = Op1.getValueType();
4519 return DAG.getNode(N0->getOpcode(), DL, VT,
4520 DAG.getNode(NewOpc, DL, VT,
4521 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4522 DAG.getNode(NewOpc, DL, VT,
4523 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4527 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4529 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4530 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4531 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4532 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4533 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4534 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4535 // Get reciprocal estimate.
4536 // float4 recip = vrecpeq_f32(yf);
4537 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4538 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4539 // Because char has a smaller range than uchar, we can actually get away
4540 // without any newton steps. This requires that we use a weird bias
4541 // of 0xb000, however (again, this has been exhaustively tested).
4542 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4543 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4544 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4545 Y = DAG.getConstant(0xb000, MVT::i32);
4546 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4547 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4548 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4549 // Convert back to short.
4550 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4551 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4556 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4558 // Convert to float.
4559 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4560 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4561 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4562 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4563 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4564 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4566 // Use reciprocal estimate and one refinement step.
4567 // float4 recip = vrecpeq_f32(yf);
4568 // recip *= vrecpsq_f32(yf, recip);
4569 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4570 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4571 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4572 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4574 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4575 // Because short has a smaller range than ushort, we can actually get away
4576 // with only a single newton step. This requires that we use a weird bias
4577 // of 89, however (again, this has been exhaustively tested).
4578 // float4 result = as_float4(as_int4(xf*recip) + 89);
4579 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4580 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4581 N1 = DAG.getConstant(89, MVT::i32);
4582 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4583 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4584 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4585 // Convert back to integer and return.
4586 // return vmovn_s32(vcvt_s32_f32(result));
4587 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4588 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4592 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4593 EVT VT = Op.getValueType();
4594 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4595 "unexpected type for custom-lowering ISD::SDIV");
4597 DebugLoc dl = Op.getDebugLoc();
4598 SDValue N0 = Op.getOperand(0);
4599 SDValue N1 = Op.getOperand(1);
4602 if (VT == MVT::v8i8) {
4603 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4604 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4606 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4607 DAG.getIntPtrConstant(4));
4608 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4609 DAG.getIntPtrConstant(4));
4610 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4611 DAG.getIntPtrConstant(0));
4612 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4613 DAG.getIntPtrConstant(0));
4615 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4616 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4618 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4619 N0 = LowerCONCAT_VECTORS(N0, DAG);
4621 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4624 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4627 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4628 EVT VT = Op.getValueType();
4629 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4630 "unexpected type for custom-lowering ISD::UDIV");
4632 DebugLoc dl = Op.getDebugLoc();
4633 SDValue N0 = Op.getOperand(0);
4634 SDValue N1 = Op.getOperand(1);
4637 if (VT == MVT::v8i8) {
4638 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4639 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4641 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4642 DAG.getIntPtrConstant(4));
4643 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4644 DAG.getIntPtrConstant(4));
4645 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4646 DAG.getIntPtrConstant(0));
4647 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4648 DAG.getIntPtrConstant(0));
4650 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4651 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4653 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4654 N0 = LowerCONCAT_VECTORS(N0, DAG);
4656 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4657 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4662 // v4i16 sdiv ... Convert to float.
4663 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4664 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4665 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4666 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4667 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4668 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4670 // Use reciprocal estimate and two refinement steps.
4671 // float4 recip = vrecpeq_f32(yf);
4672 // recip *= vrecpsq_f32(yf, recip);
4673 // recip *= vrecpsq_f32(yf, recip);
4674 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4675 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4676 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4677 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4679 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4680 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4681 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4683 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4684 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4685 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4686 // and that it will never cause us to return an answer too large).
4687 // float4 result = as_float4(as_int4(xf*recip) + 89);
4688 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4689 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4690 N1 = DAG.getConstant(2, MVT::i32);
4691 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4692 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4693 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4694 // Convert back to integer and return.
4695 // return vmovn_u32(vcvt_s32_f32(result));
4696 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4697 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4701 static SDValue LowerTrap(SDValue Op, SelectionDAG &DAG) {
4702 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4703 TargetLowering::ArgListTy Args;
4704 std::pair<SDValue, SDValue> CallResult =
4705 TLI.LowerCallTo(Op.getOperand(0), Type::getVoidTy(*DAG.getContext()),
4706 false, false, false, false, 0, CallingConv::C,
4707 /*isTailCall=*/false,
4708 /*isReturnValueUsed=*/true,
4709 DAG.getExternalSymbol(TrapFuncName.c_str(), TLI.getPointerTy()),
4710 Args, DAG, Op.getDebugLoc());
4711 return CallResult.second;
4714 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4715 switch (Op.getOpcode()) {
4716 default: llvm_unreachable("Don't know how to custom lower this!");
4717 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4718 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4719 case ISD::GlobalAddress:
4720 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4721 LowerGlobalAddressELF(Op, DAG);
4722 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4723 case ISD::SELECT: return LowerSELECT(Op, DAG);
4724 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4725 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4726 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4727 case ISD::VASTART: return LowerVASTART(Op, DAG);
4728 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4729 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4730 case ISD::SINT_TO_FP:
4731 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4732 case ISD::FP_TO_SINT:
4733 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4734 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4735 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4736 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4737 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4738 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4739 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4740 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4741 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4743 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4746 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4747 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4748 case ISD::SRL_PARTS:
4749 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4750 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4751 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4752 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4753 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4754 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4755 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4756 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4757 case ISD::MUL: return LowerMUL(Op, DAG);
4758 case ISD::SDIV: return LowerSDIV(Op, DAG);
4759 case ISD::UDIV: return LowerUDIV(Op, DAG);
4760 case ISD::TRAP: return LowerTrap(Op, DAG);
4765 /// ReplaceNodeResults - Replace the results of node with an illegal result
4766 /// type with new values built out of custom code.
4767 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4768 SmallVectorImpl<SDValue>&Results,
4769 SelectionDAG &DAG) const {
4771 switch (N->getOpcode()) {
4773 llvm_unreachable("Don't know how to custom expand this!");
4776 Res = ExpandBITCAST(N, DAG);
4780 Res = Expand64BitShift(N, DAG, Subtarget);
4784 Results.push_back(Res);
4787 //===----------------------------------------------------------------------===//
4788 // ARM Scheduler Hooks
4789 //===----------------------------------------------------------------------===//
4792 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4793 MachineBasicBlock *BB,
4794 unsigned Size) const {
4795 unsigned dest = MI->getOperand(0).getReg();
4796 unsigned ptr = MI->getOperand(1).getReg();
4797 unsigned oldval = MI->getOperand(2).getReg();
4798 unsigned newval = MI->getOperand(3).getReg();
4799 unsigned scratch = BB->getParent()->getRegInfo()
4800 .createVirtualRegister(ARM::GPRRegisterClass);
4801 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4802 DebugLoc dl = MI->getDebugLoc();
4803 bool isThumb2 = Subtarget->isThumb2();
4805 unsigned ldrOpc, strOpc;
4807 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4809 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4810 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4813 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4814 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4817 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4818 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4822 MachineFunction *MF = BB->getParent();
4823 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4824 MachineFunction::iterator It = BB;
4825 ++It; // insert the new blocks after the current block
4827 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4828 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4829 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4830 MF->insert(It, loop1MBB);
4831 MF->insert(It, loop2MBB);
4832 MF->insert(It, exitMBB);
4834 // Transfer the remainder of BB and its successor edges to exitMBB.
4835 exitMBB->splice(exitMBB->begin(), BB,
4836 llvm::next(MachineBasicBlock::iterator(MI)),
4838 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4842 // fallthrough --> loop1MBB
4843 BB->addSuccessor(loop1MBB);
4846 // ldrex dest, [ptr]
4850 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4851 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4852 .addReg(dest).addReg(oldval));
4853 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4854 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4855 BB->addSuccessor(loop2MBB);
4856 BB->addSuccessor(exitMBB);
4859 // strex scratch, newval, [ptr]
4863 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4865 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4866 .addReg(scratch).addImm(0));
4867 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4868 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4869 BB->addSuccessor(loop1MBB);
4870 BB->addSuccessor(exitMBB);
4876 MI->eraseFromParent(); // The instruction is gone now.
4882 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4883 unsigned Size, unsigned BinOpcode) const {
4884 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4885 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4887 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4888 MachineFunction *MF = BB->getParent();
4889 MachineFunction::iterator It = BB;
4892 unsigned dest = MI->getOperand(0).getReg();
4893 unsigned ptr = MI->getOperand(1).getReg();
4894 unsigned incr = MI->getOperand(2).getReg();
4895 DebugLoc dl = MI->getDebugLoc();
4897 bool isThumb2 = Subtarget->isThumb2();
4898 unsigned ldrOpc, strOpc;
4900 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4902 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4903 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4906 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4907 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4910 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4911 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4915 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4916 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4917 MF->insert(It, loopMBB);
4918 MF->insert(It, exitMBB);
4920 // Transfer the remainder of BB and its successor edges to exitMBB.
4921 exitMBB->splice(exitMBB->begin(), BB,
4922 llvm::next(MachineBasicBlock::iterator(MI)),
4924 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4926 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4927 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4928 unsigned scratch2 = (!BinOpcode) ? incr :
4929 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4933 // fallthrough --> loopMBB
4934 BB->addSuccessor(loopMBB);
4938 // <binop> scratch2, dest, incr
4939 // strex scratch, scratch2, ptr
4942 // fallthrough --> exitMBB
4944 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4946 // operand order needs to go the other way for NAND
4947 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4948 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4949 addReg(incr).addReg(dest)).addReg(0);
4951 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4952 addReg(dest).addReg(incr)).addReg(0);
4955 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4957 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4958 .addReg(scratch).addImm(0));
4959 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4960 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4962 BB->addSuccessor(loopMBB);
4963 BB->addSuccessor(exitMBB);
4969 MI->eraseFromParent(); // The instruction is gone now.
4975 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4976 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4977 E = MBB->succ_end(); I != E; ++I)
4980 llvm_unreachable("Expecting a BB with two successors!");
4984 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4985 MachineBasicBlock *BB) const {
4986 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4987 DebugLoc dl = MI->getDebugLoc();
4988 bool isThumb2 = Subtarget->isThumb2();
4989 switch (MI->getOpcode()) {
4992 llvm_unreachable("Unexpected instr type to insert");
4994 case ARM::ATOMIC_LOAD_ADD_I8:
4995 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4996 case ARM::ATOMIC_LOAD_ADD_I16:
4997 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4998 case ARM::ATOMIC_LOAD_ADD_I32:
4999 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5001 case ARM::ATOMIC_LOAD_AND_I8:
5002 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5003 case ARM::ATOMIC_LOAD_AND_I16:
5004 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5005 case ARM::ATOMIC_LOAD_AND_I32:
5006 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5008 case ARM::ATOMIC_LOAD_OR_I8:
5009 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5010 case ARM::ATOMIC_LOAD_OR_I16:
5011 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5012 case ARM::ATOMIC_LOAD_OR_I32:
5013 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5015 case ARM::ATOMIC_LOAD_XOR_I8:
5016 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5017 case ARM::ATOMIC_LOAD_XOR_I16:
5018 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5019 case ARM::ATOMIC_LOAD_XOR_I32:
5020 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5022 case ARM::ATOMIC_LOAD_NAND_I8:
5023 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5024 case ARM::ATOMIC_LOAD_NAND_I16:
5025 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5026 case ARM::ATOMIC_LOAD_NAND_I32:
5027 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5029 case ARM::ATOMIC_LOAD_SUB_I8:
5030 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5031 case ARM::ATOMIC_LOAD_SUB_I16:
5032 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5033 case ARM::ATOMIC_LOAD_SUB_I32:
5034 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5036 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
5037 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
5038 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
5040 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
5041 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
5042 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
5055 unsigned OldOpc = MI->getOpcode();
5092 llvm_unreachable("Unknown opcode?");
5095 MachineInstrBuilder MIB =
5096 BuildMI(*BB, MI, MI->getDebugLoc(), TII->get(Opc));
5097 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
5098 MIB.addOperand(MI->getOperand(i));
5099 AddDefaultPred(MIB);
5100 MIB.addReg(ARM::CPSR, RegState::Define); // S bit
5101 MI->eraseFromParent();
5106 case ARM::tMOVCCr_pseudo: {
5107 // To "insert" a SELECT_CC instruction, we actually have to insert the
5108 // diamond control-flow pattern. The incoming instruction knows the
5109 // destination vreg to set, the condition code register to branch on, the
5110 // true/false values to select between, and a branch opcode to use.
5111 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5112 MachineFunction::iterator It = BB;
5118 // cmpTY ccX, r1, r2
5120 // fallthrough --> copy0MBB
5121 MachineBasicBlock *thisMBB = BB;
5122 MachineFunction *F = BB->getParent();
5123 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5124 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5125 F->insert(It, copy0MBB);
5126 F->insert(It, sinkMBB);
5128 // Transfer the remainder of BB and its successor edges to sinkMBB.
5129 sinkMBB->splice(sinkMBB->begin(), BB,
5130 llvm::next(MachineBasicBlock::iterator(MI)),
5132 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5134 BB->addSuccessor(copy0MBB);
5135 BB->addSuccessor(sinkMBB);
5137 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
5138 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
5141 // %FalseValue = ...
5142 // # fallthrough to sinkMBB
5145 // Update machine-CFG edges
5146 BB->addSuccessor(sinkMBB);
5149 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5152 BuildMI(*BB, BB->begin(), dl,
5153 TII->get(ARM::PHI), MI->getOperand(0).getReg())
5154 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5155 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5157 MI->eraseFromParent(); // The pseudo instruction is gone now.
5162 case ARM::BCCZi64: {
5163 // If there is an unconditional branch to the other successor, remove it.
5164 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
5166 // Compare both parts that make up the double comparison separately for
5168 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
5170 unsigned LHS1 = MI->getOperand(1).getReg();
5171 unsigned LHS2 = MI->getOperand(2).getReg();
5173 AddDefaultPred(BuildMI(BB, dl,
5174 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5175 .addReg(LHS1).addImm(0));
5176 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5177 .addReg(LHS2).addImm(0)
5178 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5180 unsigned RHS1 = MI->getOperand(3).getReg();
5181 unsigned RHS2 = MI->getOperand(4).getReg();
5182 AddDefaultPred(BuildMI(BB, dl,
5183 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5184 .addReg(LHS1).addReg(RHS1));
5185 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5186 .addReg(LHS2).addReg(RHS2)
5187 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5190 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
5191 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
5192 if (MI->getOperand(0).getImm() == ARMCC::NE)
5193 std::swap(destMBB, exitMBB);
5195 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5196 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
5197 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
5200 MI->eraseFromParent(); // The pseudo instruction is gone now.
5206 //===----------------------------------------------------------------------===//
5207 // ARM Optimization Hooks
5208 //===----------------------------------------------------------------------===//
5211 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
5212 TargetLowering::DAGCombinerInfo &DCI) {
5213 SelectionDAG &DAG = DCI.DAG;
5214 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5215 EVT VT = N->getValueType(0);
5216 unsigned Opc = N->getOpcode();
5217 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
5218 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
5219 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
5220 ISD::CondCode CC = ISD::SETCC_INVALID;
5223 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
5225 SDValue CCOp = Slct.getOperand(0);
5226 if (CCOp.getOpcode() == ISD::SETCC)
5227 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
5230 bool DoXform = false;
5232 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
5235 if (LHS.getOpcode() == ISD::Constant &&
5236 cast<ConstantSDNode>(LHS)->isNullValue()) {
5238 } else if (CC != ISD::SETCC_INVALID &&
5239 RHS.getOpcode() == ISD::Constant &&
5240 cast<ConstantSDNode>(RHS)->isNullValue()) {
5241 std::swap(LHS, RHS);
5242 SDValue Op0 = Slct.getOperand(0);
5243 EVT OpVT = isSlctCC ? Op0.getValueType() :
5244 Op0.getOperand(0).getValueType();
5245 bool isInt = OpVT.isInteger();
5246 CC = ISD::getSetCCInverse(CC, isInt);
5248 if (!TLI.isCondCodeLegal(CC, OpVT))
5249 return SDValue(); // Inverse operator isn't legal.
5256 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
5258 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
5259 Slct.getOperand(0), Slct.getOperand(1), CC);
5260 SDValue CCOp = Slct.getOperand(0);
5262 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
5263 CCOp.getOperand(0), CCOp.getOperand(1), CC);
5264 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
5265 CCOp, OtherOp, Result);
5270 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
5271 /// operands N0 and N1. This is a helper for PerformADDCombine that is
5272 /// called with the default operands, and if that fails, with commuted
5274 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
5275 TargetLowering::DAGCombinerInfo &DCI) {
5276 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
5277 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
5278 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
5279 if (Result.getNode()) return Result;
5284 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
5286 static SDValue PerformADDCombine(SDNode *N,
5287 TargetLowering::DAGCombinerInfo &DCI) {
5288 SDValue N0 = N->getOperand(0);
5289 SDValue N1 = N->getOperand(1);
5291 // First try with the default operand order.
5292 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
5293 if (Result.getNode())
5296 // If that didn't work, try again with the operands commuted.
5297 return PerformADDCombineWithOperands(N, N1, N0, DCI);
5300 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
5302 static SDValue PerformSUBCombine(SDNode *N,
5303 TargetLowering::DAGCombinerInfo &DCI) {
5304 SDValue N0 = N->getOperand(0);
5305 SDValue N1 = N->getOperand(1);
5307 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
5308 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
5309 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
5310 if (Result.getNode()) return Result;
5316 /// PerformVMULCombine
5317 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
5318 /// special multiplier accumulator forwarding.
5324 static SDValue PerformVMULCombine(SDNode *N,
5325 TargetLowering::DAGCombinerInfo &DCI,
5326 const ARMSubtarget *Subtarget) {
5327 if (!Subtarget->hasVMLxForwarding())
5330 SelectionDAG &DAG = DCI.DAG;
5331 SDValue N0 = N->getOperand(0);
5332 SDValue N1 = N->getOperand(1);
5333 unsigned Opcode = N0.getOpcode();
5334 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5335 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
5336 Opcode = N0.getOpcode();
5337 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5338 Opcode != ISD::FADD && Opcode != ISD::FSUB)
5343 EVT VT = N->getValueType(0);
5344 DebugLoc DL = N->getDebugLoc();
5345 SDValue N00 = N0->getOperand(0);
5346 SDValue N01 = N0->getOperand(1);
5347 return DAG.getNode(Opcode, DL, VT,
5348 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
5349 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
5352 static SDValue PerformMULCombine(SDNode *N,
5353 TargetLowering::DAGCombinerInfo &DCI,
5354 const ARMSubtarget *Subtarget) {
5355 SelectionDAG &DAG = DCI.DAG;
5357 if (Subtarget->isThumb1Only())
5360 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5363 EVT VT = N->getValueType(0);
5364 if (VT.is64BitVector() || VT.is128BitVector())
5365 return PerformVMULCombine(N, DCI, Subtarget);
5369 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
5373 uint64_t MulAmt = C->getZExtValue();
5374 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
5375 ShiftAmt = ShiftAmt & (32 - 1);
5376 SDValue V = N->getOperand(0);
5377 DebugLoc DL = N->getDebugLoc();
5380 MulAmt >>= ShiftAmt;
5381 if (isPowerOf2_32(MulAmt - 1)) {
5382 // (mul x, 2^N + 1) => (add (shl x, N), x)
5383 Res = DAG.getNode(ISD::ADD, DL, VT,
5384 V, DAG.getNode(ISD::SHL, DL, VT,
5385 V, DAG.getConstant(Log2_32(MulAmt-1),
5387 } else if (isPowerOf2_32(MulAmt + 1)) {
5388 // (mul x, 2^N - 1) => (sub (shl x, N), x)
5389 Res = DAG.getNode(ISD::SUB, DL, VT,
5390 DAG.getNode(ISD::SHL, DL, VT,
5391 V, DAG.getConstant(Log2_32(MulAmt+1),
5398 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
5399 DAG.getConstant(ShiftAmt, MVT::i32));
5401 // Do not add new nodes to DAG combiner worklist.
5402 DCI.CombineTo(N, Res, false);
5406 static SDValue PerformANDCombine(SDNode *N,
5407 TargetLowering::DAGCombinerInfo &DCI) {
5409 // Attempt to use immediate-form VBIC
5410 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5411 DebugLoc dl = N->getDebugLoc();
5412 EVT VT = N->getValueType(0);
5413 SelectionDAG &DAG = DCI.DAG;
5415 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5418 APInt SplatBits, SplatUndef;
5419 unsigned SplatBitSize;
5422 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5423 if (SplatBitSize <= 64) {
5425 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
5426 SplatUndef.getZExtValue(), SplatBitSize,
5427 DAG, VbicVT, VT.is128BitVector(),
5429 if (Val.getNode()) {
5431 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
5432 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
5433 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
5441 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
5442 static SDValue PerformORCombine(SDNode *N,
5443 TargetLowering::DAGCombinerInfo &DCI,
5444 const ARMSubtarget *Subtarget) {
5445 // Attempt to use immediate-form VORR
5446 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5447 DebugLoc dl = N->getDebugLoc();
5448 EVT VT = N->getValueType(0);
5449 SelectionDAG &DAG = DCI.DAG;
5451 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5454 APInt SplatBits, SplatUndef;
5455 unsigned SplatBitSize;
5457 if (BVN && Subtarget->hasNEON() &&
5458 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5459 if (SplatBitSize <= 64) {
5461 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5462 SplatUndef.getZExtValue(), SplatBitSize,
5463 DAG, VorrVT, VT.is128BitVector(),
5465 if (Val.getNode()) {
5467 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
5468 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
5469 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
5474 SDValue N0 = N->getOperand(0);
5475 if (N0.getOpcode() != ISD::AND)
5477 SDValue N1 = N->getOperand(1);
5479 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
5480 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
5481 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
5483 unsigned SplatBitSize;
5486 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
5488 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
5489 HasAnyUndefs) && !HasAnyUndefs) {
5490 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
5492 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
5493 HasAnyUndefs) && !HasAnyUndefs &&
5494 SplatBits0 == ~SplatBits1) {
5495 // Canonicalize the vector type to make instruction selection simpler.
5496 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
5497 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
5498 N0->getOperand(1), N0->getOperand(0),
5500 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
5505 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
5508 // BFI is only available on V6T2+
5509 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
5512 DebugLoc DL = N->getDebugLoc();
5513 // 1) or (and A, mask), val => ARMbfi A, val, mask
5514 // iff (val & mask) == val
5516 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5517 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
5518 // && mask == ~mask2
5519 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
5520 // && ~mask == mask2
5521 // (i.e., copy a bitfield value into another bitfield of the same width)
5526 SDValue N00 = N0.getOperand(0);
5528 // The value and the mask need to be constants so we can verify this is
5529 // actually a bitfield set. If the mask is 0xffff, we can do better
5530 // via a movt instruction, so don't use BFI in that case.
5531 SDValue MaskOp = N0.getOperand(1);
5532 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
5535 unsigned Mask = MaskC->getZExtValue();
5539 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
5540 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
5542 unsigned Val = N1C->getZExtValue();
5543 if ((Val & ~Mask) != Val)
5546 if (ARM::isBitFieldInvertedMask(Mask)) {
5547 Val >>= CountTrailingZeros_32(~Mask);
5549 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
5550 DAG.getConstant(Val, MVT::i32),
5551 DAG.getConstant(Mask, MVT::i32));
5553 // Do not add new nodes to DAG combiner worklist.
5554 DCI.CombineTo(N, Res, false);
5557 } else if (N1.getOpcode() == ISD::AND) {
5558 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5559 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5562 unsigned Mask2 = N11C->getZExtValue();
5564 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
5566 if (ARM::isBitFieldInvertedMask(Mask) &&
5568 // The pack halfword instruction works better for masks that fit it,
5569 // so use that when it's available.
5570 if (Subtarget->hasT2ExtractPack() &&
5571 (Mask == 0xffff || Mask == 0xffff0000))
5574 unsigned amt = CountTrailingZeros_32(Mask2);
5575 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
5576 DAG.getConstant(amt, MVT::i32));
5577 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
5578 DAG.getConstant(Mask, MVT::i32));
5579 // Do not add new nodes to DAG combiner worklist.
5580 DCI.CombineTo(N, Res, false);
5582 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
5584 // The pack halfword instruction works better for masks that fit it,
5585 // so use that when it's available.
5586 if (Subtarget->hasT2ExtractPack() &&
5587 (Mask2 == 0xffff || Mask2 == 0xffff0000))
5590 unsigned lsb = CountTrailingZeros_32(Mask);
5591 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
5592 DAG.getConstant(lsb, MVT::i32));
5593 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
5594 DAG.getConstant(Mask2, MVT::i32));
5595 // Do not add new nodes to DAG combiner worklist.
5596 DCI.CombineTo(N, Res, false);
5601 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
5602 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
5603 ARM::isBitFieldInvertedMask(~Mask)) {
5604 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
5605 // where lsb(mask) == #shamt and masked bits of B are known zero.
5606 SDValue ShAmt = N00.getOperand(1);
5607 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5608 unsigned LSB = CountTrailingZeros_32(Mask);
5612 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
5613 DAG.getConstant(~Mask, MVT::i32));
5615 // Do not add new nodes to DAG combiner worklist.
5616 DCI.CombineTo(N, Res, false);
5622 /// PerformBFICombine - (bfi A, (and B, C1), C2) -> (bfi A, B, C2) iff
5624 static SDValue PerformBFICombine(SDNode *N,
5625 TargetLowering::DAGCombinerInfo &DCI) {
5626 SDValue N1 = N->getOperand(1);
5627 if (N1.getOpcode() == ISD::AND) {
5628 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5631 unsigned Mask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
5632 unsigned Mask2 = N11C->getZExtValue();
5633 if ((Mask & Mask2) == Mask2)
5634 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
5635 N->getOperand(0), N1.getOperand(0),
5641 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
5642 /// ARMISD::VMOVRRD.
5643 static SDValue PerformVMOVRRDCombine(SDNode *N,
5644 TargetLowering::DAGCombinerInfo &DCI) {
5645 // vmovrrd(vmovdrr x, y) -> x,y
5646 SDValue InDouble = N->getOperand(0);
5647 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
5648 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
5650 // vmovrrd(load f64) -> (load i32), (load i32)
5651 SDNode *InNode = InDouble.getNode();
5652 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
5653 InNode->getValueType(0) == MVT::f64 &&
5654 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
5655 !cast<LoadSDNode>(InNode)->isVolatile()) {
5656 // TODO: Should this be done for non-FrameIndex operands?
5657 LoadSDNode *LD = cast<LoadSDNode>(InNode);
5659 SelectionDAG &DAG = DCI.DAG;
5660 DebugLoc DL = LD->getDebugLoc();
5661 SDValue BasePtr = LD->getBasePtr();
5662 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
5663 LD->getPointerInfo(), LD->isVolatile(),
5664 LD->isNonTemporal(), LD->getAlignment());
5666 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
5667 DAG.getConstant(4, MVT::i32));
5668 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
5669 LD->getPointerInfo(), LD->isVolatile(),
5670 LD->isNonTemporal(),
5671 std::min(4U, LD->getAlignment() / 2));
5673 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
5674 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
5675 DCI.RemoveFromWorklist(LD);
5683 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
5684 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
5685 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
5686 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
5687 SDValue Op0 = N->getOperand(0);
5688 SDValue Op1 = N->getOperand(1);
5689 if (Op0.getOpcode() == ISD::BITCAST)
5690 Op0 = Op0.getOperand(0);
5691 if (Op1.getOpcode() == ISD::BITCAST)
5692 Op1 = Op1.getOperand(0);
5693 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
5694 Op0.getNode() == Op1.getNode() &&
5695 Op0.getResNo() == 0 && Op1.getResNo() == 1)
5696 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5697 N->getValueType(0), Op0.getOperand(0));
5701 /// PerformSTORECombine - Target-specific dag combine xforms for
5703 static SDValue PerformSTORECombine(SDNode *N,
5704 TargetLowering::DAGCombinerInfo &DCI) {
5705 // Bitcast an i64 store extracted from a vector to f64.
5706 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5707 StoreSDNode *St = cast<StoreSDNode>(N);
5708 SDValue StVal = St->getValue();
5709 if (!ISD::isNormalStore(St) || St->isVolatile() ||
5710 StVal.getValueType() != MVT::i64 ||
5711 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5714 SelectionDAG &DAG = DCI.DAG;
5715 DebugLoc dl = StVal.getDebugLoc();
5716 SDValue IntVec = StVal.getOperand(0);
5717 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5718 IntVec.getValueType().getVectorNumElements());
5719 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
5720 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5721 Vec, StVal.getOperand(1));
5722 dl = N->getDebugLoc();
5723 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
5724 // Make the DAGCombiner fold the bitcasts.
5725 DCI.AddToWorklist(Vec.getNode());
5726 DCI.AddToWorklist(ExtElt.getNode());
5727 DCI.AddToWorklist(V.getNode());
5728 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
5729 St->getPointerInfo(), St->isVolatile(),
5730 St->isNonTemporal(), St->getAlignment(),
5734 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
5735 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
5736 /// i64 vector to have f64 elements, since the value can then be loaded
5737 /// directly into a VFP register.
5738 static bool hasNormalLoadOperand(SDNode *N) {
5739 unsigned NumElts = N->getValueType(0).getVectorNumElements();
5740 for (unsigned i = 0; i < NumElts; ++i) {
5741 SDNode *Elt = N->getOperand(i).getNode();
5742 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
5748 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
5749 /// ISD::BUILD_VECTOR.
5750 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
5751 TargetLowering::DAGCombinerInfo &DCI){
5752 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
5753 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
5754 // into a pair of GPRs, which is fine when the value is used as a scalar,
5755 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
5756 SelectionDAG &DAG = DCI.DAG;
5757 if (N->getNumOperands() == 2) {
5758 SDValue RV = PerformVMOVDRRCombine(N, DAG);
5763 // Load i64 elements as f64 values so that type legalization does not split
5764 // them up into i32 values.
5765 EVT VT = N->getValueType(0);
5766 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
5768 DebugLoc dl = N->getDebugLoc();
5769 SmallVector<SDValue, 8> Ops;
5770 unsigned NumElts = VT.getVectorNumElements();
5771 for (unsigned i = 0; i < NumElts; ++i) {
5772 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
5774 // Make the DAGCombiner fold the bitcast.
5775 DCI.AddToWorklist(V.getNode());
5777 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
5778 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
5779 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
5782 /// PerformInsertEltCombine - Target-specific dag combine xforms for
5783 /// ISD::INSERT_VECTOR_ELT.
5784 static SDValue PerformInsertEltCombine(SDNode *N,
5785 TargetLowering::DAGCombinerInfo &DCI) {
5786 // Bitcast an i64 load inserted into a vector to f64.
5787 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5788 EVT VT = N->getValueType(0);
5789 SDNode *Elt = N->getOperand(1).getNode();
5790 if (VT.getVectorElementType() != MVT::i64 ||
5791 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
5794 SelectionDAG &DAG = DCI.DAG;
5795 DebugLoc dl = N->getDebugLoc();
5796 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5797 VT.getVectorNumElements());
5798 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
5799 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
5800 // Make the DAGCombiner fold the bitcasts.
5801 DCI.AddToWorklist(Vec.getNode());
5802 DCI.AddToWorklist(V.getNode());
5803 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
5804 Vec, V, N->getOperand(2));
5805 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
5808 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
5809 /// ISD::VECTOR_SHUFFLE.
5810 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
5811 // The LLVM shufflevector instruction does not require the shuffle mask
5812 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
5813 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
5814 // operands do not match the mask length, they are extended by concatenating
5815 // them with undef vectors. That is probably the right thing for other
5816 // targets, but for NEON it is better to concatenate two double-register
5817 // size vector operands into a single quad-register size vector. Do that
5818 // transformation here:
5819 // shuffle(concat(v1, undef), concat(v2, undef)) ->
5820 // shuffle(concat(v1, v2), undef)
5821 SDValue Op0 = N->getOperand(0);
5822 SDValue Op1 = N->getOperand(1);
5823 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
5824 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
5825 Op0.getNumOperands() != 2 ||
5826 Op1.getNumOperands() != 2)
5828 SDValue Concat0Op1 = Op0.getOperand(1);
5829 SDValue Concat1Op1 = Op1.getOperand(1);
5830 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
5831 Concat1Op1.getOpcode() != ISD::UNDEF)
5833 // Skip the transformation if any of the types are illegal.
5834 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5835 EVT VT = N->getValueType(0);
5836 if (!TLI.isTypeLegal(VT) ||
5837 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
5838 !TLI.isTypeLegal(Concat1Op1.getValueType()))
5841 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5842 Op0.getOperand(0), Op1.getOperand(0));
5843 // Translate the shuffle mask.
5844 SmallVector<int, 16> NewMask;
5845 unsigned NumElts = VT.getVectorNumElements();
5846 unsigned HalfElts = NumElts/2;
5847 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
5848 for (unsigned n = 0; n < NumElts; ++n) {
5849 int MaskElt = SVN->getMaskElt(n);
5851 if (MaskElt < (int)HalfElts)
5853 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
5854 NewElt = HalfElts + MaskElt - NumElts;
5855 NewMask.push_back(NewElt);
5857 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
5858 DAG.getUNDEF(VT), NewMask.data());
5861 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
5862 /// NEON load/store intrinsics to merge base address updates.
5863 static SDValue CombineBaseUpdate(SDNode *N,
5864 TargetLowering::DAGCombinerInfo &DCI) {
5865 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5868 SelectionDAG &DAG = DCI.DAG;
5869 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
5870 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
5871 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
5872 SDValue Addr = N->getOperand(AddrOpIdx);
5874 // Search for a use of the address operand that is an increment.
5875 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
5876 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
5878 if (User->getOpcode() != ISD::ADD ||
5879 UI.getUse().getResNo() != Addr.getResNo())
5882 // Check that the add is independent of the load/store. Otherwise, folding
5883 // it would create a cycle.
5884 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
5887 // Find the new opcode for the updating load/store.
5889 bool isLaneOp = false;
5890 unsigned NewOpc = 0;
5891 unsigned NumVecs = 0;
5893 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
5895 default: assert(0 && "unexpected intrinsic for Neon base update");
5896 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
5898 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
5900 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
5902 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
5904 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
5905 NumVecs = 2; isLaneOp = true; break;
5906 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
5907 NumVecs = 3; isLaneOp = true; break;
5908 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
5909 NumVecs = 4; isLaneOp = true; break;
5910 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
5911 NumVecs = 1; isLoad = false; break;
5912 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
5913 NumVecs = 2; isLoad = false; break;
5914 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
5915 NumVecs = 3; isLoad = false; break;
5916 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
5917 NumVecs = 4; isLoad = false; break;
5918 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
5919 NumVecs = 2; isLoad = false; isLaneOp = true; break;
5920 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
5921 NumVecs = 3; isLoad = false; isLaneOp = true; break;
5922 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
5923 NumVecs = 4; isLoad = false; isLaneOp = true; break;
5927 switch (N->getOpcode()) {
5928 default: assert(0 && "unexpected opcode for Neon base update");
5929 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
5930 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
5931 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
5935 // Find the size of memory referenced by the load/store.
5938 VecTy = N->getValueType(0);
5940 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
5941 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
5943 NumBytes /= VecTy.getVectorNumElements();
5945 // If the increment is a constant, it must match the memory ref size.
5946 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
5947 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
5948 uint64_t IncVal = CInc->getZExtValue();
5949 if (IncVal != NumBytes)
5951 } else if (NumBytes >= 3 * 16) {
5952 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
5953 // separate instructions that make it harder to use a non-constant update.
5957 // Create the new updating load/store node.
5959 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
5961 for (n = 0; n < NumResultVecs; ++n)
5963 Tys[n++] = MVT::i32;
5964 Tys[n] = MVT::Other;
5965 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
5966 SmallVector<SDValue, 8> Ops;
5967 Ops.push_back(N->getOperand(0)); // incoming chain
5968 Ops.push_back(N->getOperand(AddrOpIdx));
5970 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
5971 Ops.push_back(N->getOperand(i));
5973 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
5974 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
5975 Ops.data(), Ops.size(),
5976 MemInt->getMemoryVT(),
5977 MemInt->getMemOperand());
5980 std::vector<SDValue> NewResults;
5981 for (unsigned i = 0; i < NumResultVecs; ++i) {
5982 NewResults.push_back(SDValue(UpdN.getNode(), i));
5984 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
5985 DCI.CombineTo(N, NewResults);
5986 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
5993 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
5994 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
5995 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
5997 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
5998 SelectionDAG &DAG = DCI.DAG;
5999 EVT VT = N->getValueType(0);
6000 // vldN-dup instructions only support 64-bit vectors for N > 1.
6001 if (!VT.is64BitVector())
6004 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
6005 SDNode *VLD = N->getOperand(0).getNode();
6006 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
6008 unsigned NumVecs = 0;
6009 unsigned NewOpc = 0;
6010 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
6011 if (IntNo == Intrinsic::arm_neon_vld2lane) {
6013 NewOpc = ARMISD::VLD2DUP;
6014 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
6016 NewOpc = ARMISD::VLD3DUP;
6017 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
6019 NewOpc = ARMISD::VLD4DUP;
6024 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
6025 // numbers match the load.
6026 unsigned VLDLaneNo =
6027 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
6028 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6030 // Ignore uses of the chain result.
6031 if (UI.getUse().getResNo() == NumVecs)
6034 if (User->getOpcode() != ARMISD::VDUPLANE ||
6035 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
6039 // Create the vldN-dup node.
6042 for (n = 0; n < NumVecs; ++n)
6044 Tys[n] = MVT::Other;
6045 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
6046 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
6047 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
6048 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
6049 Ops, 2, VLDMemInt->getMemoryVT(),
6050 VLDMemInt->getMemOperand());
6053 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6055 unsigned ResNo = UI.getUse().getResNo();
6056 // Ignore uses of the chain result.
6057 if (ResNo == NumVecs)
6060 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
6063 // Now the vldN-lane intrinsic is dead except for its chain result.
6064 // Update uses of the chain.
6065 std::vector<SDValue> VLDDupResults;
6066 for (unsigned n = 0; n < NumVecs; ++n)
6067 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
6068 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
6069 DCI.CombineTo(VLD, VLDDupResults);
6074 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
6075 /// ARMISD::VDUPLANE.
6076 static SDValue PerformVDUPLANECombine(SDNode *N,
6077 TargetLowering::DAGCombinerInfo &DCI) {
6078 SDValue Op = N->getOperand(0);
6080 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
6081 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
6082 if (CombineVLDDUP(N, DCI))
6083 return SDValue(N, 0);
6085 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
6086 // redundant. Ignore bit_converts for now; element sizes are checked below.
6087 while (Op.getOpcode() == ISD::BITCAST)
6088 Op = Op.getOperand(0);
6089 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
6092 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
6093 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
6094 // The canonical VMOV for a zero vector uses a 32-bit element size.
6095 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6097 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
6099 EVT VT = N->getValueType(0);
6100 if (EltSize > VT.getVectorElementType().getSizeInBits())
6103 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
6106 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6107 /// operand of a vector shift operation, where all the elements of the
6108 /// build_vector must have the same constant integer value.
6109 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6110 // Ignore bit_converts.
6111 while (Op.getOpcode() == ISD::BITCAST)
6112 Op = Op.getOperand(0);
6113 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6114 APInt SplatBits, SplatUndef;
6115 unsigned SplatBitSize;
6117 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6118 HasAnyUndefs, ElementBits) ||
6119 SplatBitSize > ElementBits)
6121 Cnt = SplatBits.getSExtValue();
6125 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6126 /// operand of a vector shift left operation. That value must be in the range:
6127 /// 0 <= Value < ElementBits for a left shift; or
6128 /// 0 <= Value <= ElementBits for a long left shift.
6129 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6130 assert(VT.isVector() && "vector shift count is not a vector type");
6131 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6132 if (! getVShiftImm(Op, ElementBits, Cnt))
6134 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
6137 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6138 /// operand of a vector shift right operation. For a shift opcode, the value
6139 /// is positive, but for an intrinsic the value count must be negative. The
6140 /// absolute value must be in the range:
6141 /// 1 <= |Value| <= ElementBits for a right shift; or
6142 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6143 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6145 assert(VT.isVector() && "vector shift count is not a vector type");
6146 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6147 if (! getVShiftImm(Op, ElementBits, Cnt))
6151 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
6154 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
6155 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
6156 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
6159 // Don't do anything for most intrinsics.
6162 // Vector shifts: check for immediate versions and lower them.
6163 // Note: This is done during DAG combining instead of DAG legalizing because
6164 // the build_vectors for 64-bit vector element shift counts are generally
6165 // not legal, and it is hard to see their values after they get legalized to
6166 // loads from a constant pool.
6167 case Intrinsic::arm_neon_vshifts:
6168 case Intrinsic::arm_neon_vshiftu:
6169 case Intrinsic::arm_neon_vshiftls:
6170 case Intrinsic::arm_neon_vshiftlu:
6171 case Intrinsic::arm_neon_vshiftn:
6172 case Intrinsic::arm_neon_vrshifts:
6173 case Intrinsic::arm_neon_vrshiftu:
6174 case Intrinsic::arm_neon_vrshiftn:
6175 case Intrinsic::arm_neon_vqshifts:
6176 case Intrinsic::arm_neon_vqshiftu:
6177 case Intrinsic::arm_neon_vqshiftsu:
6178 case Intrinsic::arm_neon_vqshiftns:
6179 case Intrinsic::arm_neon_vqshiftnu:
6180 case Intrinsic::arm_neon_vqshiftnsu:
6181 case Intrinsic::arm_neon_vqrshiftns:
6182 case Intrinsic::arm_neon_vqrshiftnu:
6183 case Intrinsic::arm_neon_vqrshiftnsu: {
6184 EVT VT = N->getOperand(1).getValueType();
6186 unsigned VShiftOpc = 0;
6189 case Intrinsic::arm_neon_vshifts:
6190 case Intrinsic::arm_neon_vshiftu:
6191 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
6192 VShiftOpc = ARMISD::VSHL;
6195 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
6196 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
6197 ARMISD::VSHRs : ARMISD::VSHRu);
6202 case Intrinsic::arm_neon_vshiftls:
6203 case Intrinsic::arm_neon_vshiftlu:
6204 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
6206 llvm_unreachable("invalid shift count for vshll intrinsic");
6208 case Intrinsic::arm_neon_vrshifts:
6209 case Intrinsic::arm_neon_vrshiftu:
6210 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
6214 case Intrinsic::arm_neon_vqshifts:
6215 case Intrinsic::arm_neon_vqshiftu:
6216 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6220 case Intrinsic::arm_neon_vqshiftsu:
6221 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6223 llvm_unreachable("invalid shift count for vqshlu intrinsic");
6225 case Intrinsic::arm_neon_vshiftn:
6226 case Intrinsic::arm_neon_vrshiftn:
6227 case Intrinsic::arm_neon_vqshiftns:
6228 case Intrinsic::arm_neon_vqshiftnu:
6229 case Intrinsic::arm_neon_vqshiftnsu:
6230 case Intrinsic::arm_neon_vqrshiftns:
6231 case Intrinsic::arm_neon_vqrshiftnu:
6232 case Intrinsic::arm_neon_vqrshiftnsu:
6233 // Narrowing shifts require an immediate right shift.
6234 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
6236 llvm_unreachable("invalid shift count for narrowing vector shift "
6240 llvm_unreachable("unhandled vector shift");
6244 case Intrinsic::arm_neon_vshifts:
6245 case Intrinsic::arm_neon_vshiftu:
6246 // Opcode already set above.
6248 case Intrinsic::arm_neon_vshiftls:
6249 case Intrinsic::arm_neon_vshiftlu:
6250 if (Cnt == VT.getVectorElementType().getSizeInBits())
6251 VShiftOpc = ARMISD::VSHLLi;
6253 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
6254 ARMISD::VSHLLs : ARMISD::VSHLLu);
6256 case Intrinsic::arm_neon_vshiftn:
6257 VShiftOpc = ARMISD::VSHRN; break;
6258 case Intrinsic::arm_neon_vrshifts:
6259 VShiftOpc = ARMISD::VRSHRs; break;
6260 case Intrinsic::arm_neon_vrshiftu:
6261 VShiftOpc = ARMISD::VRSHRu; break;
6262 case Intrinsic::arm_neon_vrshiftn:
6263 VShiftOpc = ARMISD::VRSHRN; break;
6264 case Intrinsic::arm_neon_vqshifts:
6265 VShiftOpc = ARMISD::VQSHLs; break;
6266 case Intrinsic::arm_neon_vqshiftu:
6267 VShiftOpc = ARMISD::VQSHLu; break;
6268 case Intrinsic::arm_neon_vqshiftsu:
6269 VShiftOpc = ARMISD::VQSHLsu; break;
6270 case Intrinsic::arm_neon_vqshiftns:
6271 VShiftOpc = ARMISD::VQSHRNs; break;
6272 case Intrinsic::arm_neon_vqshiftnu:
6273 VShiftOpc = ARMISD::VQSHRNu; break;
6274 case Intrinsic::arm_neon_vqshiftnsu:
6275 VShiftOpc = ARMISD::VQSHRNsu; break;
6276 case Intrinsic::arm_neon_vqrshiftns:
6277 VShiftOpc = ARMISD::VQRSHRNs; break;
6278 case Intrinsic::arm_neon_vqrshiftnu:
6279 VShiftOpc = ARMISD::VQRSHRNu; break;
6280 case Intrinsic::arm_neon_vqrshiftnsu:
6281 VShiftOpc = ARMISD::VQRSHRNsu; break;
6284 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6285 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
6288 case Intrinsic::arm_neon_vshiftins: {
6289 EVT VT = N->getOperand(1).getValueType();
6291 unsigned VShiftOpc = 0;
6293 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
6294 VShiftOpc = ARMISD::VSLI;
6295 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
6296 VShiftOpc = ARMISD::VSRI;
6298 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
6301 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6302 N->getOperand(1), N->getOperand(2),
6303 DAG.getConstant(Cnt, MVT::i32));
6306 case Intrinsic::arm_neon_vqrshifts:
6307 case Intrinsic::arm_neon_vqrshiftu:
6308 // No immediate versions of these to check for.
6315 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
6316 /// lowers them. As with the vector shift intrinsics, this is done during DAG
6317 /// combining instead of DAG legalizing because the build_vectors for 64-bit
6318 /// vector element shift counts are generally not legal, and it is hard to see
6319 /// their values after they get legalized to loads from a constant pool.
6320 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
6321 const ARMSubtarget *ST) {
6322 EVT VT = N->getValueType(0);
6324 // Nothing to be done for scalar shifts.
6325 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6326 if (!VT.isVector() || !TLI.isTypeLegal(VT))
6329 assert(ST->hasNEON() && "unexpected vector shift");
6332 switch (N->getOpcode()) {
6333 default: llvm_unreachable("unexpected shift opcode");
6336 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
6337 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
6338 DAG.getConstant(Cnt, MVT::i32));
6343 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
6344 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
6345 ARMISD::VSHRs : ARMISD::VSHRu);
6346 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
6347 DAG.getConstant(Cnt, MVT::i32));
6353 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
6354 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
6355 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
6356 const ARMSubtarget *ST) {
6357 SDValue N0 = N->getOperand(0);
6359 // Check for sign- and zero-extensions of vector extract operations of 8-
6360 // and 16-bit vector elements. NEON supports these directly. They are
6361 // handled during DAG combining because type legalization will promote them
6362 // to 32-bit types and it is messy to recognize the operations after that.
6363 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6364 SDValue Vec = N0.getOperand(0);
6365 SDValue Lane = N0.getOperand(1);
6366 EVT VT = N->getValueType(0);
6367 EVT EltVT = N0.getValueType();
6368 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6370 if (VT == MVT::i32 &&
6371 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
6372 TLI.isTypeLegal(Vec.getValueType()) &&
6373 isa<ConstantSDNode>(Lane)) {
6376 switch (N->getOpcode()) {
6377 default: llvm_unreachable("unexpected opcode");
6378 case ISD::SIGN_EXTEND:
6379 Opc = ARMISD::VGETLANEs;
6381 case ISD::ZERO_EXTEND:
6382 case ISD::ANY_EXTEND:
6383 Opc = ARMISD::VGETLANEu;
6386 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
6393 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
6394 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
6395 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
6396 const ARMSubtarget *ST) {
6397 // If the target supports NEON, try to use vmax/vmin instructions for f32
6398 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
6399 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
6400 // a NaN; only do the transformation when it matches that behavior.
6402 // For now only do this when using NEON for FP operations; if using VFP, it
6403 // is not obvious that the benefit outweighs the cost of switching to the
6405 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
6406 N->getValueType(0) != MVT::f32)
6409 SDValue CondLHS = N->getOperand(0);
6410 SDValue CondRHS = N->getOperand(1);
6411 SDValue LHS = N->getOperand(2);
6412 SDValue RHS = N->getOperand(3);
6413 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
6415 unsigned Opcode = 0;
6417 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
6418 IsReversed = false; // x CC y ? x : y
6419 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
6420 IsReversed = true ; // x CC y ? y : x
6434 // If LHS is NaN, an ordered comparison will be false and the result will
6435 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
6436 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6437 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
6438 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6440 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
6441 // will return -0, so vmin can only be used for unsafe math or if one of
6442 // the operands is known to be nonzero.
6443 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
6445 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6447 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
6456 // If LHS is NaN, an ordered comparison will be false and the result will
6457 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
6458 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6459 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
6460 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6462 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
6463 // will return +0, so vmax can only be used for unsafe math or if one of
6464 // the operands is known to be nonzero.
6465 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
6467 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6469 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
6475 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
6478 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
6479 DAGCombinerInfo &DCI) const {
6480 switch (N->getOpcode()) {
6482 case ISD::ADD: return PerformADDCombine(N, DCI);
6483 case ISD::SUB: return PerformSUBCombine(N, DCI);
6484 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
6485 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
6486 case ISD::AND: return PerformANDCombine(N, DCI);
6487 case ARMISD::BFI: return PerformBFICombine(N, DCI);
6488 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
6489 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
6490 case ISD::STORE: return PerformSTORECombine(N, DCI);
6491 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
6492 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
6493 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
6494 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
6495 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
6498 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
6499 case ISD::SIGN_EXTEND:
6500 case ISD::ZERO_EXTEND:
6501 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
6502 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
6503 case ARMISD::VLD2DUP:
6504 case ARMISD::VLD3DUP:
6505 case ARMISD::VLD4DUP:
6506 return CombineBaseUpdate(N, DCI);
6507 case ISD::INTRINSIC_VOID:
6508 case ISD::INTRINSIC_W_CHAIN:
6509 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
6510 case Intrinsic::arm_neon_vld1:
6511 case Intrinsic::arm_neon_vld2:
6512 case Intrinsic::arm_neon_vld3:
6513 case Intrinsic::arm_neon_vld4:
6514 case Intrinsic::arm_neon_vld2lane:
6515 case Intrinsic::arm_neon_vld3lane:
6516 case Intrinsic::arm_neon_vld4lane:
6517 case Intrinsic::arm_neon_vst1:
6518 case Intrinsic::arm_neon_vst2:
6519 case Intrinsic::arm_neon_vst3:
6520 case Intrinsic::arm_neon_vst4:
6521 case Intrinsic::arm_neon_vst2lane:
6522 case Intrinsic::arm_neon_vst3lane:
6523 case Intrinsic::arm_neon_vst4lane:
6524 return CombineBaseUpdate(N, DCI);
6532 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
6534 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
6537 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
6538 if (!Subtarget->allowsUnalignedMem())
6541 switch (VT.getSimpleVT().SimpleTy) {
6548 // FIXME: VLD1 etc with standard alignment is legal.
6552 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
6557 switch (VT.getSimpleVT().SimpleTy) {
6558 default: return false;
6573 if ((V & (Scale - 1)) != 0)
6576 return V == (V & ((1LL << 5) - 1));
6579 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
6580 const ARMSubtarget *Subtarget) {
6587 switch (VT.getSimpleVT().SimpleTy) {
6588 default: return false;
6593 // + imm12 or - imm8
6595 return V == (V & ((1LL << 8) - 1));
6596 return V == (V & ((1LL << 12) - 1));
6599 // Same as ARM mode. FIXME: NEON?
6600 if (!Subtarget->hasVFP2())
6605 return V == (V & ((1LL << 8) - 1));
6609 /// isLegalAddressImmediate - Return true if the integer value can be used
6610 /// as the offset of the target addressing mode for load / store of the
6612 static bool isLegalAddressImmediate(int64_t V, EVT VT,
6613 const ARMSubtarget *Subtarget) {
6620 if (Subtarget->isThumb1Only())
6621 return isLegalT1AddressImmediate(V, VT);
6622 else if (Subtarget->isThumb2())
6623 return isLegalT2AddressImmediate(V, VT, Subtarget);
6628 switch (VT.getSimpleVT().SimpleTy) {
6629 default: return false;
6634 return V == (V & ((1LL << 12) - 1));
6637 return V == (V & ((1LL << 8) - 1));
6640 if (!Subtarget->hasVFP2()) // FIXME: NEON?
6645 return V == (V & ((1LL << 8) - 1));
6649 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
6651 int Scale = AM.Scale;
6655 switch (VT.getSimpleVT().SimpleTy) {
6656 default: return false;
6665 return Scale == 2 || Scale == 4 || Scale == 8;
6668 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6672 // Note, we allow "void" uses (basically, uses that aren't loads or
6673 // stores), because arm allows folding a scale into many arithmetic
6674 // operations. This should be made more precise and revisited later.
6676 // Allow r << imm, but the imm has to be a multiple of two.
6677 if (Scale & 1) return false;
6678 return isPowerOf2_32(Scale);
6682 /// isLegalAddressingMode - Return true if the addressing mode represented
6683 /// by AM is legal for this target, for a load/store of the specified type.
6684 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6685 const Type *Ty) const {
6686 EVT VT = getValueType(Ty, true);
6687 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
6690 // Can never fold addr of global into load/store.
6695 case 0: // no scale reg, must be "r+i" or "r", or "i".
6698 if (Subtarget->isThumb1Only())
6702 // ARM doesn't support any R+R*scale+imm addr modes.
6709 if (Subtarget->isThumb2())
6710 return isLegalT2ScaledAddressingMode(AM, VT);
6712 int Scale = AM.Scale;
6713 switch (VT.getSimpleVT().SimpleTy) {
6714 default: return false;
6718 if (Scale < 0) Scale = -Scale;
6722 return isPowerOf2_32(Scale & ~1);
6726 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6731 // Note, we allow "void" uses (basically, uses that aren't loads or
6732 // stores), because arm allows folding a scale into many arithmetic
6733 // operations. This should be made more precise and revisited later.
6735 // Allow r << imm, but the imm has to be a multiple of two.
6736 if (Scale & 1) return false;
6737 return isPowerOf2_32(Scale);
6744 /// isLegalICmpImmediate - Return true if the specified immediate is legal
6745 /// icmp immediate, that is the target has icmp instructions which can compare
6746 /// a register against the immediate without having to materialize the
6747 /// immediate into a register.
6748 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
6749 if (!Subtarget->isThumb())
6750 return ARM_AM::getSOImmVal(Imm) != -1;
6751 if (Subtarget->isThumb2())
6752 return ARM_AM::getT2SOImmVal(Imm) != -1;
6753 return Imm >= 0 && Imm <= 255;
6756 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
6757 bool isSEXTLoad, SDValue &Base,
6758 SDValue &Offset, bool &isInc,
6759 SelectionDAG &DAG) {
6760 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6763 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
6765 Base = Ptr->getOperand(0);
6766 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6767 int RHSC = (int)RHS->getZExtValue();
6768 if (RHSC < 0 && RHSC > -256) {
6769 assert(Ptr->getOpcode() == ISD::ADD);
6771 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6775 isInc = (Ptr->getOpcode() == ISD::ADD);
6776 Offset = Ptr->getOperand(1);
6778 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
6780 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6781 int RHSC = (int)RHS->getZExtValue();
6782 if (RHSC < 0 && RHSC > -0x1000) {
6783 assert(Ptr->getOpcode() == ISD::ADD);
6785 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6786 Base = Ptr->getOperand(0);
6791 if (Ptr->getOpcode() == ISD::ADD) {
6793 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
6794 if (ShOpcVal != ARM_AM::no_shift) {
6795 Base = Ptr->getOperand(1);
6796 Offset = Ptr->getOperand(0);
6798 Base = Ptr->getOperand(0);
6799 Offset = Ptr->getOperand(1);
6804 isInc = (Ptr->getOpcode() == ISD::ADD);
6805 Base = Ptr->getOperand(0);
6806 Offset = Ptr->getOperand(1);
6810 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
6814 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
6815 bool isSEXTLoad, SDValue &Base,
6816 SDValue &Offset, bool &isInc,
6817 SelectionDAG &DAG) {
6818 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6821 Base = Ptr->getOperand(0);
6822 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6823 int RHSC = (int)RHS->getZExtValue();
6824 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
6825 assert(Ptr->getOpcode() == ISD::ADD);
6827 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6829 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
6830 isInc = Ptr->getOpcode() == ISD::ADD;
6831 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
6839 /// getPreIndexedAddressParts - returns true by value, base pointer and
6840 /// offset pointer and addressing mode by reference if the node's address
6841 /// can be legally represented as pre-indexed load / store address.
6843 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
6845 ISD::MemIndexedMode &AM,
6846 SelectionDAG &DAG) const {
6847 if (Subtarget->isThumb1Only())
6852 bool isSEXTLoad = false;
6853 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6854 Ptr = LD->getBasePtr();
6855 VT = LD->getMemoryVT();
6856 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6857 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6858 Ptr = ST->getBasePtr();
6859 VT = ST->getMemoryVT();
6864 bool isLegal = false;
6865 if (Subtarget->isThumb2())
6866 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6867 Offset, isInc, DAG);
6869 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6870 Offset, isInc, DAG);
6874 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
6878 /// getPostIndexedAddressParts - returns true by value, base pointer and
6879 /// offset pointer and addressing mode by reference if this node can be
6880 /// combined with a load / store to form a post-indexed load / store.
6881 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
6884 ISD::MemIndexedMode &AM,
6885 SelectionDAG &DAG) const {
6886 if (Subtarget->isThumb1Only())
6891 bool isSEXTLoad = false;
6892 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6893 VT = LD->getMemoryVT();
6894 Ptr = LD->getBasePtr();
6895 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6896 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6897 VT = ST->getMemoryVT();
6898 Ptr = ST->getBasePtr();
6903 bool isLegal = false;
6904 if (Subtarget->isThumb2())
6905 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6908 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6914 // Swap base ptr and offset to catch more post-index load / store when
6915 // it's legal. In Thumb2 mode, offset must be an immediate.
6916 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
6917 !Subtarget->isThumb2())
6918 std::swap(Base, Offset);
6920 // Post-indexed load / store update the base pointer.
6925 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
6929 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6933 const SelectionDAG &DAG,
6934 unsigned Depth) const {
6935 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
6936 switch (Op.getOpcode()) {
6938 case ARMISD::CMOV: {
6939 // Bits are known zero/one if known on the LHS and RHS.
6940 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
6941 if (KnownZero == 0 && KnownOne == 0) return;
6943 APInt KnownZeroRHS, KnownOneRHS;
6944 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
6945 KnownZeroRHS, KnownOneRHS, Depth+1);
6946 KnownZero &= KnownZeroRHS;
6947 KnownOne &= KnownOneRHS;
6953 //===----------------------------------------------------------------------===//
6954 // ARM Inline Assembly Support
6955 //===----------------------------------------------------------------------===//
6957 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
6958 // Looking for "rev" which is V6+.
6959 if (!Subtarget->hasV6Ops())
6962 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
6963 std::string AsmStr = IA->getAsmString();
6964 SmallVector<StringRef, 4> AsmPieces;
6965 SplitString(AsmStr, AsmPieces, ";\n");
6967 switch (AsmPieces.size()) {
6968 default: return false;
6970 AsmStr = AsmPieces[0];
6972 SplitString(AsmStr, AsmPieces, " \t,");
6975 if (AsmPieces.size() == 3 &&
6976 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
6977 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
6978 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
6979 if (Ty && Ty->getBitWidth() == 32)
6980 return IntrinsicLowering::LowerToByteSwap(CI);
6988 /// getConstraintType - Given a constraint letter, return the type of
6989 /// constraint it is for this target.
6990 ARMTargetLowering::ConstraintType
6991 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
6992 if (Constraint.size() == 1) {
6993 switch (Constraint[0]) {
6995 case 'l': return C_RegisterClass;
6996 case 'w': return C_RegisterClass;
6999 return TargetLowering::getConstraintType(Constraint);
7002 /// Examine constraint type and operand type and determine a weight value.
7003 /// This object must already have been set up with the operand type
7004 /// and the current alternative constraint selected.
7005 TargetLowering::ConstraintWeight
7006 ARMTargetLowering::getSingleConstraintMatchWeight(
7007 AsmOperandInfo &info, const char *constraint) const {
7008 ConstraintWeight weight = CW_Invalid;
7009 Value *CallOperandVal = info.CallOperandVal;
7010 // If we don't have a value, we can't do a match,
7011 // but allow it at the lowest weight.
7012 if (CallOperandVal == NULL)
7014 const Type *type = CallOperandVal->getType();
7015 // Look at the constraint type.
7016 switch (*constraint) {
7018 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
7021 if (type->isIntegerTy()) {
7022 if (Subtarget->isThumb())
7023 weight = CW_SpecificReg;
7025 weight = CW_Register;
7029 if (type->isFloatingPointTy())
7030 weight = CW_Register;
7036 std::pair<unsigned, const TargetRegisterClass*>
7037 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7039 if (Constraint.size() == 1) {
7040 // GCC ARM Constraint Letters
7041 switch (Constraint[0]) {
7043 if (Subtarget->isThumb())
7044 return std::make_pair(0U, ARM::tGPRRegisterClass);
7046 return std::make_pair(0U, ARM::GPRRegisterClass);
7048 return std::make_pair(0U, ARM::GPRRegisterClass);
7051 return std::make_pair(0U, ARM::SPRRegisterClass);
7052 if (VT.getSizeInBits() == 64)
7053 return std::make_pair(0U, ARM::DPRRegisterClass);
7054 if (VT.getSizeInBits() == 128)
7055 return std::make_pair(0U, ARM::QPRRegisterClass);
7059 if (StringRef("{cc}").equals_lower(Constraint))
7060 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
7062 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7065 std::vector<unsigned> ARMTargetLowering::
7066 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7068 if (Constraint.size() != 1)
7069 return std::vector<unsigned>();
7071 switch (Constraint[0]) { // GCC ARM Constraint Letters
7074 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
7075 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
7078 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
7079 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
7080 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
7081 ARM::R12, ARM::LR, 0);
7084 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
7085 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
7086 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
7087 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
7088 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
7089 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
7090 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
7091 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
7092 if (VT.getSizeInBits() == 64)
7093 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
7094 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
7095 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
7096 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
7097 if (VT.getSizeInBits() == 128)
7098 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
7099 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
7103 return std::vector<unsigned>();
7106 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7107 /// vector. If it is invalid, don't add anything to Ops.
7108 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7110 std::vector<SDValue>&Ops,
7111 SelectionDAG &DAG) const {
7112 SDValue Result(0, 0);
7114 switch (Constraint) {
7116 case 'I': case 'J': case 'K': case 'L':
7117 case 'M': case 'N': case 'O':
7118 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
7122 int64_t CVal64 = C->getSExtValue();
7123 int CVal = (int) CVal64;
7124 // None of these constraints allow values larger than 32 bits. Check
7125 // that the value fits in an int.
7129 switch (Constraint) {
7131 if (Subtarget->isThumb1Only()) {
7132 // This must be a constant between 0 and 255, for ADD
7134 if (CVal >= 0 && CVal <= 255)
7136 } else if (Subtarget->isThumb2()) {
7137 // A constant that can be used as an immediate value in a
7138 // data-processing instruction.
7139 if (ARM_AM::getT2SOImmVal(CVal) != -1)
7142 // A constant that can be used as an immediate value in a
7143 // data-processing instruction.
7144 if (ARM_AM::getSOImmVal(CVal) != -1)
7150 if (Subtarget->isThumb()) { // FIXME thumb2
7151 // This must be a constant between -255 and -1, for negated ADD
7152 // immediates. This can be used in GCC with an "n" modifier that
7153 // prints the negated value, for use with SUB instructions. It is
7154 // not useful otherwise but is implemented for compatibility.
7155 if (CVal >= -255 && CVal <= -1)
7158 // This must be a constant between -4095 and 4095. It is not clear
7159 // what this constraint is intended for. Implemented for
7160 // compatibility with GCC.
7161 if (CVal >= -4095 && CVal <= 4095)
7167 if (Subtarget->isThumb1Only()) {
7168 // A 32-bit value where only one byte has a nonzero value. Exclude
7169 // zero to match GCC. This constraint is used by GCC internally for
7170 // constants that can be loaded with a move/shift combination.
7171 // It is not useful otherwise but is implemented for compatibility.
7172 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
7174 } else if (Subtarget->isThumb2()) {
7175 // A constant whose bitwise inverse can be used as an immediate
7176 // value in a data-processing instruction. This can be used in GCC
7177 // with a "B" modifier that prints the inverted value, for use with
7178 // BIC and MVN instructions. It is not useful otherwise but is
7179 // implemented for compatibility.
7180 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
7183 // A constant whose bitwise inverse can be used as an immediate
7184 // value in a data-processing instruction. This can be used in GCC
7185 // with a "B" modifier that prints the inverted value, for use with
7186 // BIC and MVN instructions. It is not useful otherwise but is
7187 // implemented for compatibility.
7188 if (ARM_AM::getSOImmVal(~CVal) != -1)
7194 if (Subtarget->isThumb1Only()) {
7195 // This must be a constant between -7 and 7,
7196 // for 3-operand ADD/SUB immediate instructions.
7197 if (CVal >= -7 && CVal < 7)
7199 } else if (Subtarget->isThumb2()) {
7200 // A constant whose negation can be used as an immediate value in a
7201 // data-processing instruction. This can be used in GCC with an "n"
7202 // modifier that prints the negated value, for use with SUB
7203 // instructions. It is not useful otherwise but is implemented for
7205 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
7208 // A constant whose negation can be used as an immediate value in a
7209 // data-processing instruction. This can be used in GCC with an "n"
7210 // modifier that prints the negated value, for use with SUB
7211 // instructions. It is not useful otherwise but is implemented for
7213 if (ARM_AM::getSOImmVal(-CVal) != -1)
7219 if (Subtarget->isThumb()) { // FIXME thumb2
7220 // This must be a multiple of 4 between 0 and 1020, for
7221 // ADD sp + immediate.
7222 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
7225 // A power of two or a constant between 0 and 32. This is used in
7226 // GCC for the shift amount on shifted register operands, but it is
7227 // useful in general for any shift amounts.
7228 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
7234 if (Subtarget->isThumb()) { // FIXME thumb2
7235 // This must be a constant between 0 and 31, for shift amounts.
7236 if (CVal >= 0 && CVal <= 31)
7242 if (Subtarget->isThumb()) { // FIXME thumb2
7243 // This must be a multiple of 4 between -508 and 508, for
7244 // ADD/SUB sp = sp + immediate.
7245 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
7250 Result = DAG.getTargetConstant(CVal, Op.getValueType());
7254 if (Result.getNode()) {
7255 Ops.push_back(Result);
7258 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7262 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
7263 // The ARM target isn't yet aware of offsets.
7267 int ARM::getVFPf32Imm(const APFloat &FPImm) {
7268 APInt Imm = FPImm.bitcastToAPInt();
7269 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
7270 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
7271 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
7273 // We can handle 4 bits of mantissa.
7274 // mantissa = (16+UInt(e:f:g:h))/16.
7275 if (Mantissa & 0x7ffff)
7278 if ((Mantissa & 0xf) != Mantissa)
7281 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
7282 if (Exp < -3 || Exp > 4)
7284 Exp = ((Exp+3) & 0x7) ^ 4;
7286 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
7289 int ARM::getVFPf64Imm(const APFloat &FPImm) {
7290 APInt Imm = FPImm.bitcastToAPInt();
7291 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
7292 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
7293 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
7295 // We can handle 4 bits of mantissa.
7296 // mantissa = (16+UInt(e:f:g:h))/16.
7297 if (Mantissa & 0xffffffffffffLL)
7300 if ((Mantissa & 0xf) != Mantissa)
7303 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
7304 if (Exp < -3 || Exp > 4)
7306 Exp = ((Exp+3) & 0x7) ^ 4;
7308 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
7311 bool ARM::isBitFieldInvertedMask(unsigned v) {
7312 if (v == 0xffffffff)
7314 // there can be 1's on either or both "outsides", all the "inside"
7316 unsigned int lsb = 0, msb = 31;
7317 while (v & (1 << msb)) --msb;
7318 while (v & (1 << lsb)) ++lsb;
7319 for (unsigned int i = lsb; i <= msb; ++i) {
7326 /// isFPImmLegal - Returns true if the target can instruction select the
7327 /// specified FP immediate natively. If false, the legalizer will
7328 /// materialize the FP immediate as a load from a constant pool.
7329 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
7330 if (!Subtarget->hasVFP3())
7333 return ARM::getVFPf32Imm(Imm) != -1;
7335 return ARM::getVFPf64Imm(Imm) != -1;
7339 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
7340 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
7341 /// specified in the intrinsic calls.
7342 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
7344 unsigned Intrinsic) const {
7345 switch (Intrinsic) {
7346 case Intrinsic::arm_neon_vld1:
7347 case Intrinsic::arm_neon_vld2:
7348 case Intrinsic::arm_neon_vld3:
7349 case Intrinsic::arm_neon_vld4:
7350 case Intrinsic::arm_neon_vld2lane:
7351 case Intrinsic::arm_neon_vld3lane:
7352 case Intrinsic::arm_neon_vld4lane: {
7353 Info.opc = ISD::INTRINSIC_W_CHAIN;
7354 // Conservatively set memVT to the entire set of vectors loaded.
7355 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
7356 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7357 Info.ptrVal = I.getArgOperand(0);
7359 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7360 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7361 Info.vol = false; // volatile loads with NEON intrinsics not supported
7362 Info.readMem = true;
7363 Info.writeMem = false;
7366 case Intrinsic::arm_neon_vst1:
7367 case Intrinsic::arm_neon_vst2:
7368 case Intrinsic::arm_neon_vst3:
7369 case Intrinsic::arm_neon_vst4:
7370 case Intrinsic::arm_neon_vst2lane:
7371 case Intrinsic::arm_neon_vst3lane:
7372 case Intrinsic::arm_neon_vst4lane: {
7373 Info.opc = ISD::INTRINSIC_VOID;
7374 // Conservatively set memVT to the entire set of vectors stored.
7375 unsigned NumElts = 0;
7376 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
7377 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
7378 if (!ArgTy->isVectorTy())
7380 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
7382 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7383 Info.ptrVal = I.getArgOperand(0);
7385 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7386 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7387 Info.vol = false; // volatile stores with NEON intrinsics not supported
7388 Info.readMem = false;
7389 Info.writeMem = true;