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 // The APCS parameter registers.
76 static const unsigned GPRArgRegs[] = {
77 ARM::R0, ARM::R1, ARM::R2, ARM::R3
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 // RTABI chapter 4.3.4
402 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
403 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
404 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
407 if (Subtarget->isThumb1Only())
408 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
410 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
411 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
412 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
413 if (!Subtarget->isFPOnlySP())
414 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
416 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
419 if (Subtarget->hasNEON()) {
420 addDRTypeForNEON(MVT::v2f32);
421 addDRTypeForNEON(MVT::v8i8);
422 addDRTypeForNEON(MVT::v4i16);
423 addDRTypeForNEON(MVT::v2i32);
424 addDRTypeForNEON(MVT::v1i64);
426 addQRTypeForNEON(MVT::v4f32);
427 addQRTypeForNEON(MVT::v2f64);
428 addQRTypeForNEON(MVT::v16i8);
429 addQRTypeForNEON(MVT::v8i16);
430 addQRTypeForNEON(MVT::v4i32);
431 addQRTypeForNEON(MVT::v2i64);
433 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
434 // neither Neon nor VFP support any arithmetic operations on it.
435 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
436 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
437 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
438 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
439 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
440 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
441 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
442 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
443 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
444 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
445 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
446 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
447 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
448 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
449 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
450 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
451 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
452 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
453 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
454 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
455 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
456 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
457 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
458 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
460 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
462 // Neon does not support some operations on v1i64 and v2i64 types.
463 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
464 // Custom handling for some quad-vector types to detect VMULL.
465 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
466 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
467 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
468 // Custom handling for some vector types to avoid expensive expansions
469 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
470 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
471 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
472 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
473 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
474 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
475 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
476 // a destination type that is wider than the source.
477 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
478 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
480 setTargetDAGCombine(ISD::INTRINSIC_VOID);
481 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
482 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
483 setTargetDAGCombine(ISD::SHL);
484 setTargetDAGCombine(ISD::SRL);
485 setTargetDAGCombine(ISD::SRA);
486 setTargetDAGCombine(ISD::SIGN_EXTEND);
487 setTargetDAGCombine(ISD::ZERO_EXTEND);
488 setTargetDAGCombine(ISD::ANY_EXTEND);
489 setTargetDAGCombine(ISD::SELECT_CC);
490 setTargetDAGCombine(ISD::BUILD_VECTOR);
491 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
492 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
493 setTargetDAGCombine(ISD::STORE);
496 computeRegisterProperties();
498 // ARM does not have f32 extending load.
499 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
501 // ARM does not have i1 sign extending load.
502 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
504 // ARM supports all 4 flavors of integer indexed load / store.
505 if (!Subtarget->isThumb1Only()) {
506 for (unsigned im = (unsigned)ISD::PRE_INC;
507 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
508 setIndexedLoadAction(im, MVT::i1, Legal);
509 setIndexedLoadAction(im, MVT::i8, Legal);
510 setIndexedLoadAction(im, MVT::i16, Legal);
511 setIndexedLoadAction(im, MVT::i32, Legal);
512 setIndexedStoreAction(im, MVT::i1, Legal);
513 setIndexedStoreAction(im, MVT::i8, Legal);
514 setIndexedStoreAction(im, MVT::i16, Legal);
515 setIndexedStoreAction(im, MVT::i32, Legal);
519 // i64 operation support.
520 setOperationAction(ISD::MUL, MVT::i64, Expand);
521 setOperationAction(ISD::MULHU, MVT::i32, Expand);
522 if (Subtarget->isThumb1Only()) {
523 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
524 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
526 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops())
527 setOperationAction(ISD::MULHS, MVT::i32, Expand);
529 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
530 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
531 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
532 setOperationAction(ISD::SRL, MVT::i64, Custom);
533 setOperationAction(ISD::SRA, MVT::i64, Custom);
535 // ARM does not have ROTL.
536 setOperationAction(ISD::ROTL, MVT::i32, Expand);
537 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
538 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
539 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
540 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
542 // Only ARMv6 has BSWAP.
543 if (!Subtarget->hasV6Ops())
544 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
546 // These are expanded into libcalls.
547 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
548 // v7M has a hardware divider
549 setOperationAction(ISD::SDIV, MVT::i32, Expand);
550 setOperationAction(ISD::UDIV, MVT::i32, Expand);
552 setOperationAction(ISD::SREM, MVT::i32, Expand);
553 setOperationAction(ISD::UREM, MVT::i32, Expand);
554 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
555 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
557 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
558 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
559 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
560 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
561 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
563 setOperationAction(ISD::TRAP, MVT::Other, Legal);
565 // Use the default implementation.
566 setOperationAction(ISD::VASTART, MVT::Other, Custom);
567 setOperationAction(ISD::VAARG, MVT::Other, Expand);
568 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
569 setOperationAction(ISD::VAEND, MVT::Other, Expand);
570 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
571 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
572 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
573 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
574 setExceptionPointerRegister(ARM::R0);
575 setExceptionSelectorRegister(ARM::R1);
577 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
578 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
579 // the default expansion.
580 if (Subtarget->hasDataBarrier() ||
581 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
582 // membarrier needs custom lowering; the rest are legal and handled
584 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
586 // Set them all for expansion, which will force libcalls.
587 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
588 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
589 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
590 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
591 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
592 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
593 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
595 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
596 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
600 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
601 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
602 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
603 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
604 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
605 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
606 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
607 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
608 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
609 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
610 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
611 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
612 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i8, Expand);
613 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i16, Expand);
614 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
615 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i8, Expand);
616 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i16, Expand);
617 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
618 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i8, Expand);
619 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i16, Expand);
620 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
621 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i8, Expand);
622 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i16, Expand);
623 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
624 // Since the libcalls include locking, fold in the fences
625 setShouldFoldAtomicFences(true);
627 // 64-bit versions are always libcalls (for now)
628 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
629 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
630 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
631 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
632 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
633 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
634 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
635 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
637 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
639 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
640 if (!Subtarget->hasV6Ops()) {
641 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
642 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
644 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
646 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
647 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
648 // iff target supports vfp2.
649 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
650 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
653 // We want to custom lower some of our intrinsics.
654 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
655 if (Subtarget->isTargetDarwin()) {
656 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
657 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
658 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
661 setOperationAction(ISD::SETCC, MVT::i32, Expand);
662 setOperationAction(ISD::SETCC, MVT::f32, Expand);
663 setOperationAction(ISD::SETCC, MVT::f64, Expand);
664 setOperationAction(ISD::SELECT, MVT::i32, Custom);
665 setOperationAction(ISD::SELECT, MVT::f32, Custom);
666 setOperationAction(ISD::SELECT, MVT::f64, Custom);
667 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
668 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
669 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
671 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
672 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
673 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
674 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
675 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
677 // We don't support sin/cos/fmod/copysign/pow
678 setOperationAction(ISD::FSIN, MVT::f64, Expand);
679 setOperationAction(ISD::FSIN, MVT::f32, Expand);
680 setOperationAction(ISD::FCOS, MVT::f32, Expand);
681 setOperationAction(ISD::FCOS, MVT::f64, Expand);
682 setOperationAction(ISD::FREM, MVT::f64, Expand);
683 setOperationAction(ISD::FREM, MVT::f32, Expand);
684 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
685 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
686 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
688 setOperationAction(ISD::FPOW, MVT::f64, Expand);
689 setOperationAction(ISD::FPOW, MVT::f32, Expand);
691 // Various VFP goodness
692 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
693 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
694 if (Subtarget->hasVFP2()) {
695 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
696 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
697 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
698 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
700 // Special handling for half-precision FP.
701 if (!Subtarget->hasFP16()) {
702 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
703 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
707 // We have target-specific dag combine patterns for the following nodes:
708 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
709 setTargetDAGCombine(ISD::ADD);
710 setTargetDAGCombine(ISD::SUB);
711 setTargetDAGCombine(ISD::MUL);
713 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
714 setTargetDAGCombine(ISD::OR);
715 if (Subtarget->hasNEON())
716 setTargetDAGCombine(ISD::AND);
718 setStackPointerRegisterToSaveRestore(ARM::SP);
720 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
721 setSchedulingPreference(Sched::RegPressure);
723 setSchedulingPreference(Sched::Hybrid);
725 //// temporary - rewrite interface to use type
726 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
728 // On ARM arguments smaller than 4 bytes are extended, so all arguments
729 // are at least 4 bytes aligned.
730 setMinStackArgumentAlignment(4);
732 benefitFromCodePlacementOpt = true;
734 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
737 // FIXME: It might make sense to define the representative register class as the
738 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
739 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
740 // SPR's representative would be DPR_VFP2. This should work well if register
741 // pressure tracking were modified such that a register use would increment the
742 // pressure of the register class's representative and all of it's super
743 // classes' representatives transitively. We have not implemented this because
744 // of the difficulty prior to coalescing of modeling operand register classes
745 // due to the common occurrence of cross class copies and subregister insertions
747 std::pair<const TargetRegisterClass*, uint8_t>
748 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
749 const TargetRegisterClass *RRC = 0;
751 switch (VT.getSimpleVT().SimpleTy) {
753 return TargetLowering::findRepresentativeClass(VT);
754 // Use DPR as representative register class for all floating point
755 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
756 // the cost is 1 for both f32 and f64.
757 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
758 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
759 RRC = ARM::DPRRegisterClass;
760 // When NEON is used for SP, only half of the register file is available
761 // because operations that define both SP and DP results will be constrained
762 // to the VFP2 class (D0-D15). We currently model this constraint prior to
763 // coalescing by double-counting the SP regs. See the FIXME above.
764 if (Subtarget->useNEONForSinglePrecisionFP())
767 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
768 case MVT::v4f32: case MVT::v2f64:
769 RRC = ARM::DPRRegisterClass;
773 RRC = ARM::DPRRegisterClass;
777 RRC = ARM::DPRRegisterClass;
781 return std::make_pair(RRC, Cost);
784 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
787 case ARMISD::Wrapper: return "ARMISD::Wrapper";
788 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
789 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
790 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
791 case ARMISD::CALL: return "ARMISD::CALL";
792 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
793 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
794 case ARMISD::tCALL: return "ARMISD::tCALL";
795 case ARMISD::BRCOND: return "ARMISD::BRCOND";
796 case ARMISD::BR_JT: return "ARMISD::BR_JT";
797 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
798 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
799 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
800 case ARMISD::CMP: return "ARMISD::CMP";
801 case ARMISD::CMPZ: return "ARMISD::CMPZ";
802 case ARMISD::CMPFP: return "ARMISD::CMPFP";
803 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
804 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
805 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
806 case ARMISD::CMOV: return "ARMISD::CMOV";
808 case ARMISD::RBIT: return "ARMISD::RBIT";
810 case ARMISD::FTOSI: return "ARMISD::FTOSI";
811 case ARMISD::FTOUI: return "ARMISD::FTOUI";
812 case ARMISD::SITOF: return "ARMISD::SITOF";
813 case ARMISD::UITOF: return "ARMISD::UITOF";
815 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
816 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
817 case ARMISD::RRX: return "ARMISD::RRX";
819 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
820 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
822 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
823 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
824 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
826 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
828 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
830 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
832 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
833 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
835 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
837 case ARMISD::VCEQ: return "ARMISD::VCEQ";
838 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
839 case ARMISD::VCGE: return "ARMISD::VCGE";
840 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
841 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
842 case ARMISD::VCGEU: return "ARMISD::VCGEU";
843 case ARMISD::VCGT: return "ARMISD::VCGT";
844 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
845 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
846 case ARMISD::VCGTU: return "ARMISD::VCGTU";
847 case ARMISD::VTST: return "ARMISD::VTST";
849 case ARMISD::VSHL: return "ARMISD::VSHL";
850 case ARMISD::VSHRs: return "ARMISD::VSHRs";
851 case ARMISD::VSHRu: return "ARMISD::VSHRu";
852 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
853 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
854 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
855 case ARMISD::VSHRN: return "ARMISD::VSHRN";
856 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
857 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
858 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
859 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
860 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
861 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
862 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
863 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
864 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
865 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
866 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
867 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
868 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
869 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
870 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
871 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
872 case ARMISD::VDUP: return "ARMISD::VDUP";
873 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
874 case ARMISD::VEXT: return "ARMISD::VEXT";
875 case ARMISD::VREV64: return "ARMISD::VREV64";
876 case ARMISD::VREV32: return "ARMISD::VREV32";
877 case ARMISD::VREV16: return "ARMISD::VREV16";
878 case ARMISD::VZIP: return "ARMISD::VZIP";
879 case ARMISD::VUZP: return "ARMISD::VUZP";
880 case ARMISD::VTRN: return "ARMISD::VTRN";
881 case ARMISD::VTBL1: return "ARMISD::VTBL1";
882 case ARMISD::VTBL2: return "ARMISD::VTBL2";
883 case ARMISD::VMULLs: return "ARMISD::VMULLs";
884 case ARMISD::VMULLu: return "ARMISD::VMULLu";
885 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
886 case ARMISD::FMAX: return "ARMISD::FMAX";
887 case ARMISD::FMIN: return "ARMISD::FMIN";
888 case ARMISD::BFI: return "ARMISD::BFI";
889 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
890 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
891 case ARMISD::VBSL: return "ARMISD::VBSL";
892 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
893 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
894 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
895 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
896 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
897 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
898 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
899 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
900 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
901 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
902 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
903 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
904 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
905 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
906 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
907 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
908 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
909 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
910 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
911 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
915 /// getRegClassFor - Return the register class that should be used for the
916 /// specified value type.
917 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
918 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
919 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
920 // load / store 4 to 8 consecutive D registers.
921 if (Subtarget->hasNEON()) {
922 if (VT == MVT::v4i64)
923 return ARM::QQPRRegisterClass;
924 else if (VT == MVT::v8i64)
925 return ARM::QQQQPRRegisterClass;
927 return TargetLowering::getRegClassFor(VT);
930 // Create a fast isel object.
932 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
933 return ARM::createFastISel(funcInfo);
936 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
937 /// be used for loads / stores from the global.
938 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
939 return (Subtarget->isThumb1Only() ? 127 : 4095);
942 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
943 unsigned NumVals = N->getNumValues();
945 return Sched::RegPressure;
947 for (unsigned i = 0; i != NumVals; ++i) {
948 EVT VT = N->getValueType(i);
949 if (VT == MVT::Glue || VT == MVT::Other)
951 if (VT.isFloatingPoint() || VT.isVector())
952 return Sched::Latency;
955 if (!N->isMachineOpcode())
956 return Sched::RegPressure;
958 // Load are scheduled for latency even if there instruction itinerary
960 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
961 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
963 if (TID.getNumDefs() == 0)
964 return Sched::RegPressure;
965 if (!Itins->isEmpty() &&
966 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
967 return Sched::Latency;
969 return Sched::RegPressure;
972 //===----------------------------------------------------------------------===//
974 //===----------------------------------------------------------------------===//
976 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
977 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
979 default: llvm_unreachable("Unknown condition code!");
980 case ISD::SETNE: return ARMCC::NE;
981 case ISD::SETEQ: return ARMCC::EQ;
982 case ISD::SETGT: return ARMCC::GT;
983 case ISD::SETGE: return ARMCC::GE;
984 case ISD::SETLT: return ARMCC::LT;
985 case ISD::SETLE: return ARMCC::LE;
986 case ISD::SETUGT: return ARMCC::HI;
987 case ISD::SETUGE: return ARMCC::HS;
988 case ISD::SETULT: return ARMCC::LO;
989 case ISD::SETULE: return ARMCC::LS;
993 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
994 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
995 ARMCC::CondCodes &CondCode2) {
996 CondCode2 = ARMCC::AL;
998 default: llvm_unreachable("Unknown FP condition!");
1000 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1002 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1004 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1005 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1006 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1007 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1008 case ISD::SETO: CondCode = ARMCC::VC; break;
1009 case ISD::SETUO: CondCode = ARMCC::VS; break;
1010 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1011 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1012 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1014 case ISD::SETULT: CondCode = ARMCC::LT; break;
1016 case ISD::SETULE: CondCode = ARMCC::LE; break;
1018 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1022 //===----------------------------------------------------------------------===//
1023 // Calling Convention Implementation
1024 //===----------------------------------------------------------------------===//
1026 #include "ARMGenCallingConv.inc"
1028 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1029 /// given CallingConvention value.
1030 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1032 bool isVarArg) const {
1035 llvm_unreachable("Unsupported calling convention");
1036 case CallingConv::Fast:
1037 if (Subtarget->hasVFP2() && !isVarArg) {
1038 if (!Subtarget->isAAPCS_ABI())
1039 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1040 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1041 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1044 case CallingConv::C: {
1045 // Use target triple & subtarget features to do actual dispatch.
1046 if (!Subtarget->isAAPCS_ABI())
1047 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1048 else if (Subtarget->hasVFP2() &&
1049 FloatABIType == FloatABI::Hard && !isVarArg)
1050 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1051 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1053 case CallingConv::ARM_AAPCS_VFP:
1054 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1055 case CallingConv::ARM_AAPCS:
1056 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1057 case CallingConv::ARM_APCS:
1058 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1062 /// LowerCallResult - Lower the result values of a call into the
1063 /// appropriate copies out of appropriate physical registers.
1065 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1066 CallingConv::ID CallConv, bool isVarArg,
1067 const SmallVectorImpl<ISD::InputArg> &Ins,
1068 DebugLoc dl, SelectionDAG &DAG,
1069 SmallVectorImpl<SDValue> &InVals) const {
1071 // Assign locations to each value returned by this call.
1072 SmallVector<CCValAssign, 16> RVLocs;
1073 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1074 RVLocs, *DAG.getContext());
1075 CCInfo.AnalyzeCallResult(Ins,
1076 CCAssignFnForNode(CallConv, /* Return*/ true,
1079 // Copy all of the result registers out of their specified physreg.
1080 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1081 CCValAssign VA = RVLocs[i];
1084 if (VA.needsCustom()) {
1085 // Handle f64 or half of a v2f64.
1086 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1088 Chain = Lo.getValue(1);
1089 InFlag = Lo.getValue(2);
1090 VA = RVLocs[++i]; // skip ahead to next loc
1091 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1093 Chain = Hi.getValue(1);
1094 InFlag = Hi.getValue(2);
1095 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1097 if (VA.getLocVT() == MVT::v2f64) {
1098 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1099 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1100 DAG.getConstant(0, MVT::i32));
1102 VA = RVLocs[++i]; // skip ahead to next loc
1103 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1104 Chain = Lo.getValue(1);
1105 InFlag = Lo.getValue(2);
1106 VA = RVLocs[++i]; // skip ahead to next loc
1107 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1108 Chain = Hi.getValue(1);
1109 InFlag = Hi.getValue(2);
1110 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1111 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1112 DAG.getConstant(1, MVT::i32));
1115 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1117 Chain = Val.getValue(1);
1118 InFlag = Val.getValue(2);
1121 switch (VA.getLocInfo()) {
1122 default: llvm_unreachable("Unknown loc info!");
1123 case CCValAssign::Full: break;
1124 case CCValAssign::BCvt:
1125 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1129 InVals.push_back(Val);
1135 /// LowerMemOpCallTo - Store the argument to the stack.
1137 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1138 SDValue StackPtr, SDValue Arg,
1139 DebugLoc dl, SelectionDAG &DAG,
1140 const CCValAssign &VA,
1141 ISD::ArgFlagsTy Flags) const {
1142 unsigned LocMemOffset = VA.getLocMemOffset();
1143 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1144 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1145 return DAG.getStore(Chain, dl, Arg, PtrOff,
1146 MachinePointerInfo::getStack(LocMemOffset),
1150 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1151 SDValue Chain, SDValue &Arg,
1152 RegsToPassVector &RegsToPass,
1153 CCValAssign &VA, CCValAssign &NextVA,
1155 SmallVector<SDValue, 8> &MemOpChains,
1156 ISD::ArgFlagsTy Flags) const {
1158 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1159 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1160 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1162 if (NextVA.isRegLoc())
1163 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1165 assert(NextVA.isMemLoc());
1166 if (StackPtr.getNode() == 0)
1167 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1169 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1175 /// LowerCall - Lowering a call into a callseq_start <-
1176 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1179 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1180 CallingConv::ID CallConv, bool isVarArg,
1182 const SmallVectorImpl<ISD::OutputArg> &Outs,
1183 const SmallVectorImpl<SDValue> &OutVals,
1184 const SmallVectorImpl<ISD::InputArg> &Ins,
1185 DebugLoc dl, SelectionDAG &DAG,
1186 SmallVectorImpl<SDValue> &InVals) const {
1187 MachineFunction &MF = DAG.getMachineFunction();
1188 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1189 bool IsSibCall = false;
1190 // Temporarily disable tail calls so things don't break.
1191 if (!EnableARMTailCalls)
1194 // Check if it's really possible to do a tail call.
1195 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1196 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1197 Outs, OutVals, Ins, DAG);
1198 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1199 // detected sibcalls.
1206 // Analyze operands of the call, assigning locations to each operand.
1207 SmallVector<CCValAssign, 16> ArgLocs;
1208 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1210 CCInfo.setCallOrPrologue(Call);
1211 CCInfo.AnalyzeCallOperands(Outs,
1212 CCAssignFnForNode(CallConv, /* Return*/ false,
1215 // Get a count of how many bytes are to be pushed on the stack.
1216 unsigned NumBytes = CCInfo.getNextStackOffset();
1218 // For tail calls, memory operands are available in our caller's stack.
1222 // Adjust the stack pointer for the new arguments...
1223 // These operations are automatically eliminated by the prolog/epilog pass
1225 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1227 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1229 RegsToPassVector RegsToPass;
1230 SmallVector<SDValue, 8> MemOpChains;
1232 // Walk the register/memloc assignments, inserting copies/loads. In the case
1233 // of tail call optimization, arguments are handled later.
1234 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1236 ++i, ++realArgIdx) {
1237 CCValAssign &VA = ArgLocs[i];
1238 SDValue Arg = OutVals[realArgIdx];
1239 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1240 bool isByVal = Flags.isByVal();
1242 // Promote the value if needed.
1243 switch (VA.getLocInfo()) {
1244 default: llvm_unreachable("Unknown loc info!");
1245 case CCValAssign::Full: break;
1246 case CCValAssign::SExt:
1247 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1249 case CCValAssign::ZExt:
1250 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1252 case CCValAssign::AExt:
1253 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1255 case CCValAssign::BCvt:
1256 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1260 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1261 if (VA.needsCustom()) {
1262 if (VA.getLocVT() == MVT::v2f64) {
1263 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1264 DAG.getConstant(0, MVT::i32));
1265 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1266 DAG.getConstant(1, MVT::i32));
1268 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1269 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1271 VA = ArgLocs[++i]; // skip ahead to next loc
1272 if (VA.isRegLoc()) {
1273 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1274 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1276 assert(VA.isMemLoc());
1278 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1279 dl, DAG, VA, Flags));
1282 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1283 StackPtr, MemOpChains, Flags);
1285 } else if (VA.isRegLoc()) {
1286 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1287 } else if (isByVal) {
1288 assert(VA.isMemLoc());
1289 unsigned offset = 0;
1291 // True if this byval aggregate will be split between registers
1293 if (CCInfo.isFirstByValRegValid()) {
1294 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1296 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1297 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1298 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1299 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1300 MachinePointerInfo(),
1302 MemOpChains.push_back(Load.getValue(1));
1303 RegsToPass.push_back(std::make_pair(j, Load));
1305 offset = ARM::R4 - CCInfo.getFirstByValReg();
1306 CCInfo.clearFirstByValReg();
1309 unsigned LocMemOffset = VA.getLocMemOffset();
1310 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1311 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1313 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1314 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1315 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1317 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1318 Flags.getByValAlign(),
1319 /*isVolatile=*/false,
1320 /*AlwaysInline=*/false,
1321 MachinePointerInfo(0),
1322 MachinePointerInfo(0)));
1324 } else if (!IsSibCall) {
1325 assert(VA.isMemLoc());
1327 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1328 dl, DAG, VA, Flags));
1332 if (!MemOpChains.empty())
1333 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1334 &MemOpChains[0], MemOpChains.size());
1336 // Build a sequence of copy-to-reg nodes chained together with token chain
1337 // and flag operands which copy the outgoing args into the appropriate regs.
1339 // Tail call byval lowering might overwrite argument registers so in case of
1340 // tail call optimization the copies to registers are lowered later.
1342 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1343 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1344 RegsToPass[i].second, InFlag);
1345 InFlag = Chain.getValue(1);
1348 // For tail calls lower the arguments to the 'real' stack slot.
1350 // Force all the incoming stack arguments to be loaded from the stack
1351 // before any new outgoing arguments are stored to the stack, because the
1352 // outgoing stack slots may alias the incoming argument stack slots, and
1353 // the alias isn't otherwise explicit. This is slightly more conservative
1354 // than necessary, because it means that each store effectively depends
1355 // on every argument instead of just those arguments it would clobber.
1357 // Do not flag preceding copytoreg stuff together with the following stuff.
1359 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1360 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1361 RegsToPass[i].second, InFlag);
1362 InFlag = Chain.getValue(1);
1367 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1368 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1369 // node so that legalize doesn't hack it.
1370 bool isDirect = false;
1371 bool isARMFunc = false;
1372 bool isLocalARMFunc = false;
1373 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1375 if (EnableARMLongCalls) {
1376 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1377 && "long-calls with non-static relocation model!");
1378 // Handle a global address or an external symbol. If it's not one of
1379 // those, the target's already in a register, so we don't need to do
1381 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1382 const GlobalValue *GV = G->getGlobal();
1383 // Create a constant pool entry for the callee address
1384 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1385 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1388 // Get the address of the callee into a register
1389 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1390 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1391 Callee = DAG.getLoad(getPointerTy(), dl,
1392 DAG.getEntryNode(), CPAddr,
1393 MachinePointerInfo::getConstantPool(),
1395 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1396 const char *Sym = S->getSymbol();
1398 // Create a constant pool entry for the callee address
1399 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1400 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1401 Sym, ARMPCLabelIndex, 0);
1402 // Get the address of the callee into a register
1403 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1404 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1405 Callee = DAG.getLoad(getPointerTy(), dl,
1406 DAG.getEntryNode(), CPAddr,
1407 MachinePointerInfo::getConstantPool(),
1410 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1411 const GlobalValue *GV = G->getGlobal();
1413 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1414 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1415 getTargetMachine().getRelocationModel() != Reloc::Static;
1416 isARMFunc = !Subtarget->isThumb() || isStub;
1417 // ARM call to a local ARM function is predicable.
1418 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1419 // tBX takes a register source operand.
1420 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1421 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1422 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1425 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1426 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1427 Callee = DAG.getLoad(getPointerTy(), dl,
1428 DAG.getEntryNode(), CPAddr,
1429 MachinePointerInfo::getConstantPool(),
1431 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1432 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1433 getPointerTy(), Callee, PICLabel);
1435 // On ELF targets for PIC code, direct calls should go through the PLT
1436 unsigned OpFlags = 0;
1437 if (Subtarget->isTargetELF() &&
1438 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1439 OpFlags = ARMII::MO_PLT;
1440 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1442 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1444 bool isStub = Subtarget->isTargetDarwin() &&
1445 getTargetMachine().getRelocationModel() != Reloc::Static;
1446 isARMFunc = !Subtarget->isThumb() || isStub;
1447 // tBX takes a register source operand.
1448 const char *Sym = S->getSymbol();
1449 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1450 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1451 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1452 Sym, ARMPCLabelIndex, 4);
1453 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1454 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1455 Callee = DAG.getLoad(getPointerTy(), dl,
1456 DAG.getEntryNode(), CPAddr,
1457 MachinePointerInfo::getConstantPool(),
1459 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1460 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1461 getPointerTy(), Callee, PICLabel);
1463 unsigned OpFlags = 0;
1464 // On ELF targets for PIC code, direct calls should go through the PLT
1465 if (Subtarget->isTargetELF() &&
1466 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1467 OpFlags = ARMII::MO_PLT;
1468 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1472 // FIXME: handle tail calls differently.
1474 if (Subtarget->isThumb()) {
1475 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1476 CallOpc = ARMISD::CALL_NOLINK;
1478 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1480 CallOpc = (isDirect || Subtarget->hasV5TOps())
1481 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1482 : ARMISD::CALL_NOLINK;
1485 std::vector<SDValue> Ops;
1486 Ops.push_back(Chain);
1487 Ops.push_back(Callee);
1489 // Add argument registers to the end of the list so that they are known live
1491 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1492 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1493 RegsToPass[i].second.getValueType()));
1495 if (InFlag.getNode())
1496 Ops.push_back(InFlag);
1498 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1500 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1502 // Returns a chain and a flag for retval copy to use.
1503 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1504 InFlag = Chain.getValue(1);
1506 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1507 DAG.getIntPtrConstant(0, true), InFlag);
1509 InFlag = Chain.getValue(1);
1511 // Handle result values, copying them out of physregs into vregs that we
1513 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1517 /// HandleByVal - Every parameter *after* a byval parameter is passed
1518 /// on the stack. Remember the next parameter register to allocate,
1519 /// and then confiscate the rest of the parameter registers to insure
1522 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1523 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1524 assert((State->getCallOrPrologue() == Prologue ||
1525 State->getCallOrPrologue() == Call) &&
1526 "unhandled ParmContext");
1527 if ((!State->isFirstByValRegValid()) &&
1528 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1529 State->setFirstByValReg(reg);
1530 // At a call site, a byval parameter that is split between
1531 // registers and memory needs its size truncated here. In a
1532 // function prologue, such byval parameters are reassembled in
1533 // memory, and are not truncated.
1534 if (State->getCallOrPrologue() == Call) {
1535 unsigned excess = 4 * (ARM::R4 - reg);
1536 assert(size >= excess && "expected larger existing stack allocation");
1540 // Confiscate any remaining parameter registers to preclude their
1541 // assignment to subsequent parameters.
1542 while (State->AllocateReg(GPRArgRegs, 4))
1546 /// MatchingStackOffset - Return true if the given stack call argument is
1547 /// already available in the same position (relatively) of the caller's
1548 /// incoming argument stack.
1550 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1551 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1552 const ARMInstrInfo *TII) {
1553 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1555 if (Arg.getOpcode() == ISD::CopyFromReg) {
1556 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1557 if (!TargetRegisterInfo::isVirtualRegister(VR))
1559 MachineInstr *Def = MRI->getVRegDef(VR);
1562 if (!Flags.isByVal()) {
1563 if (!TII->isLoadFromStackSlot(Def, FI))
1568 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1569 if (Flags.isByVal())
1570 // ByVal argument is passed in as a pointer but it's now being
1571 // dereferenced. e.g.
1572 // define @foo(%struct.X* %A) {
1573 // tail call @bar(%struct.X* byval %A)
1576 SDValue Ptr = Ld->getBasePtr();
1577 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1580 FI = FINode->getIndex();
1584 assert(FI != INT_MAX);
1585 if (!MFI->isFixedObjectIndex(FI))
1587 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1590 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1591 /// for tail call optimization. Targets which want to do tail call
1592 /// optimization should implement this function.
1594 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1595 CallingConv::ID CalleeCC,
1597 bool isCalleeStructRet,
1598 bool isCallerStructRet,
1599 const SmallVectorImpl<ISD::OutputArg> &Outs,
1600 const SmallVectorImpl<SDValue> &OutVals,
1601 const SmallVectorImpl<ISD::InputArg> &Ins,
1602 SelectionDAG& DAG) const {
1603 const Function *CallerF = DAG.getMachineFunction().getFunction();
1604 CallingConv::ID CallerCC = CallerF->getCallingConv();
1605 bool CCMatch = CallerCC == CalleeCC;
1607 // Look for obvious safe cases to perform tail call optimization that do not
1608 // require ABI changes. This is what gcc calls sibcall.
1610 // Do not sibcall optimize vararg calls unless the call site is not passing
1612 if (isVarArg && !Outs.empty())
1615 // Also avoid sibcall optimization if either caller or callee uses struct
1616 // return semantics.
1617 if (isCalleeStructRet || isCallerStructRet)
1620 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1621 // emitEpilogue is not ready for them.
1622 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1623 // LR. This means if we need to reload LR, it takes an extra instructions,
1624 // which outweighs the value of the tail call; but here we don't know yet
1625 // whether LR is going to be used. Probably the right approach is to
1626 // generate the tail call here and turn it back into CALL/RET in
1627 // emitEpilogue if LR is used.
1629 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1630 // but we need to make sure there are enough registers; the only valid
1631 // registers are the 4 used for parameters. We don't currently do this
1633 if (Subtarget->isThumb1Only())
1636 // If the calling conventions do not match, then we'd better make sure the
1637 // results are returned in the same way as what the caller expects.
1639 SmallVector<CCValAssign, 16> RVLocs1;
1640 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1641 RVLocs1, *DAG.getContext());
1642 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1644 SmallVector<CCValAssign, 16> RVLocs2;
1645 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1646 RVLocs2, *DAG.getContext());
1647 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1649 if (RVLocs1.size() != RVLocs2.size())
1651 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1652 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1654 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1656 if (RVLocs1[i].isRegLoc()) {
1657 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1660 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1666 // If the callee takes no arguments then go on to check the results of the
1668 if (!Outs.empty()) {
1669 // Check if stack adjustment is needed. For now, do not do this if any
1670 // argument is passed on the stack.
1671 SmallVector<CCValAssign, 16> ArgLocs;
1672 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1673 ArgLocs, *DAG.getContext());
1674 CCInfo.AnalyzeCallOperands(Outs,
1675 CCAssignFnForNode(CalleeCC, false, isVarArg));
1676 if (CCInfo.getNextStackOffset()) {
1677 MachineFunction &MF = DAG.getMachineFunction();
1679 // Check if the arguments are already laid out in the right way as
1680 // the caller's fixed stack objects.
1681 MachineFrameInfo *MFI = MF.getFrameInfo();
1682 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1683 const ARMInstrInfo *TII =
1684 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1685 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1687 ++i, ++realArgIdx) {
1688 CCValAssign &VA = ArgLocs[i];
1689 EVT RegVT = VA.getLocVT();
1690 SDValue Arg = OutVals[realArgIdx];
1691 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1692 if (VA.getLocInfo() == CCValAssign::Indirect)
1694 if (VA.needsCustom()) {
1695 // f64 and vector types are split into multiple registers or
1696 // register/stack-slot combinations. The types will not match
1697 // the registers; give up on memory f64 refs until we figure
1698 // out what to do about this.
1701 if (!ArgLocs[++i].isRegLoc())
1703 if (RegVT == MVT::v2f64) {
1704 if (!ArgLocs[++i].isRegLoc())
1706 if (!ArgLocs[++i].isRegLoc())
1709 } else if (!VA.isRegLoc()) {
1710 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1722 ARMTargetLowering::LowerReturn(SDValue Chain,
1723 CallingConv::ID CallConv, bool isVarArg,
1724 const SmallVectorImpl<ISD::OutputArg> &Outs,
1725 const SmallVectorImpl<SDValue> &OutVals,
1726 DebugLoc dl, SelectionDAG &DAG) const {
1728 // CCValAssign - represent the assignment of the return value to a location.
1729 SmallVector<CCValAssign, 16> RVLocs;
1731 // CCState - Info about the registers and stack slots.
1732 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1735 // Analyze outgoing return values.
1736 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1739 // If this is the first return lowered for this function, add
1740 // the regs to the liveout set for the function.
1741 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1742 for (unsigned i = 0; i != RVLocs.size(); ++i)
1743 if (RVLocs[i].isRegLoc())
1744 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1749 // Copy the result values into the output registers.
1750 for (unsigned i = 0, realRVLocIdx = 0;
1752 ++i, ++realRVLocIdx) {
1753 CCValAssign &VA = RVLocs[i];
1754 assert(VA.isRegLoc() && "Can only return in registers!");
1756 SDValue Arg = OutVals[realRVLocIdx];
1758 switch (VA.getLocInfo()) {
1759 default: llvm_unreachable("Unknown loc info!");
1760 case CCValAssign::Full: break;
1761 case CCValAssign::BCvt:
1762 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1766 if (VA.needsCustom()) {
1767 if (VA.getLocVT() == MVT::v2f64) {
1768 // Extract the first half and return it in two registers.
1769 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1770 DAG.getConstant(0, MVT::i32));
1771 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1772 DAG.getVTList(MVT::i32, MVT::i32), Half);
1774 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1775 Flag = Chain.getValue(1);
1776 VA = RVLocs[++i]; // skip ahead to next loc
1777 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1778 HalfGPRs.getValue(1), Flag);
1779 Flag = Chain.getValue(1);
1780 VA = RVLocs[++i]; // skip ahead to next loc
1782 // Extract the 2nd half and fall through to handle it as an f64 value.
1783 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1784 DAG.getConstant(1, MVT::i32));
1786 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1788 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1789 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1790 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1791 Flag = Chain.getValue(1);
1792 VA = RVLocs[++i]; // skip ahead to next loc
1793 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1796 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1798 // Guarantee that all emitted copies are
1799 // stuck together, avoiding something bad.
1800 Flag = Chain.getValue(1);
1805 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1807 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1812 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1813 if (N->getNumValues() != 1)
1815 if (!N->hasNUsesOfValue(1, 0))
1818 unsigned NumCopies = 0;
1820 SDNode *Use = *N->use_begin();
1821 if (Use->getOpcode() == ISD::CopyToReg) {
1822 Copies[NumCopies++] = Use;
1823 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1824 // f64 returned in a pair of GPRs.
1825 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1827 if (UI->getOpcode() != ISD::CopyToReg)
1829 Copies[UI.getUse().getResNo()] = *UI;
1832 } else if (Use->getOpcode() == ISD::BITCAST) {
1833 // f32 returned in a single GPR.
1834 if (!Use->hasNUsesOfValue(1, 0))
1836 Use = *Use->use_begin();
1837 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1839 Copies[NumCopies++] = Use;
1844 if (NumCopies != 1 && NumCopies != 2)
1847 bool HasRet = false;
1848 for (unsigned i = 0; i < NumCopies; ++i) {
1849 SDNode *Copy = Copies[i];
1850 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1852 if (UI->getOpcode() == ISD::CopyToReg) {
1854 if (Use == Copies[0] || Use == Copies[1])
1858 if (UI->getOpcode() != ARMISD::RET_FLAG)
1867 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1868 if (!EnableARMTailCalls)
1871 if (!CI->isTailCall())
1874 return !Subtarget->isThumb1Only();
1877 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1878 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1879 // one of the above mentioned nodes. It has to be wrapped because otherwise
1880 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1881 // be used to form addressing mode. These wrapped nodes will be selected
1883 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1884 EVT PtrVT = Op.getValueType();
1885 // FIXME there is no actual debug info here
1886 DebugLoc dl = Op.getDebugLoc();
1887 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1889 if (CP->isMachineConstantPoolEntry())
1890 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1891 CP->getAlignment());
1893 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1894 CP->getAlignment());
1895 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1898 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1899 return MachineJumpTableInfo::EK_Inline;
1902 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1903 SelectionDAG &DAG) const {
1904 MachineFunction &MF = DAG.getMachineFunction();
1905 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1906 unsigned ARMPCLabelIndex = 0;
1907 DebugLoc DL = Op.getDebugLoc();
1908 EVT PtrVT = getPointerTy();
1909 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1910 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1912 if (RelocM == Reloc::Static) {
1913 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1915 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1916 ARMPCLabelIndex = AFI->createPICLabelUId();
1917 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1918 ARMCP::CPBlockAddress,
1920 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1922 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1923 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1924 MachinePointerInfo::getConstantPool(),
1926 if (RelocM == Reloc::Static)
1928 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1929 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1932 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1934 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1935 SelectionDAG &DAG) const {
1936 DebugLoc dl = GA->getDebugLoc();
1937 EVT PtrVT = getPointerTy();
1938 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1939 MachineFunction &MF = DAG.getMachineFunction();
1940 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1941 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1942 ARMConstantPoolValue *CPV =
1943 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1944 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1945 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1946 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1947 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1948 MachinePointerInfo::getConstantPool(),
1950 SDValue Chain = Argument.getValue(1);
1952 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1953 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1955 // call __tls_get_addr.
1958 Entry.Node = Argument;
1959 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1960 Args.push_back(Entry);
1961 // FIXME: is there useful debug info available here?
1962 std::pair<SDValue, SDValue> CallResult =
1963 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1964 false, false, false, false,
1965 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1966 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1967 return CallResult.first;
1970 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1971 // "local exec" model.
1973 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1974 SelectionDAG &DAG) const {
1975 const GlobalValue *GV = GA->getGlobal();
1976 DebugLoc dl = GA->getDebugLoc();
1978 SDValue Chain = DAG.getEntryNode();
1979 EVT PtrVT = getPointerTy();
1980 // Get the Thread Pointer
1981 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1983 if (GV->isDeclaration()) {
1984 MachineFunction &MF = DAG.getMachineFunction();
1985 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1986 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1987 // Initial exec model.
1988 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1989 ARMConstantPoolValue *CPV =
1990 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1991 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1992 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1993 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1994 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1995 MachinePointerInfo::getConstantPool(),
1997 Chain = Offset.getValue(1);
1999 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2000 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2002 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2003 MachinePointerInfo::getConstantPool(),
2007 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
2008 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2009 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2010 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2011 MachinePointerInfo::getConstantPool(),
2015 // The address of the thread local variable is the add of the thread
2016 // pointer with the offset of the variable.
2017 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2021 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2022 // TODO: implement the "local dynamic" model
2023 assert(Subtarget->isTargetELF() &&
2024 "TLS not implemented for non-ELF targets");
2025 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2026 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2027 // otherwise use the "Local Exec" TLS Model
2028 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2029 return LowerToTLSGeneralDynamicModel(GA, DAG);
2031 return LowerToTLSExecModels(GA, DAG);
2034 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2035 SelectionDAG &DAG) const {
2036 EVT PtrVT = getPointerTy();
2037 DebugLoc dl = Op.getDebugLoc();
2038 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2039 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2040 if (RelocM == Reloc::PIC_) {
2041 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2042 ARMConstantPoolValue *CPV =
2043 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2044 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2045 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2046 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2048 MachinePointerInfo::getConstantPool(),
2050 SDValue Chain = Result.getValue(1);
2051 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2052 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2054 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2055 MachinePointerInfo::getGOT(), false, false, 0);
2059 // If we have T2 ops, we can materialize the address directly via movt/movw
2060 // pair. This is always cheaper.
2061 if (Subtarget->useMovt()) {
2063 // FIXME: Once remat is capable of dealing with instructions with register
2064 // operands, expand this into two nodes.
2065 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2066 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2068 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2069 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2070 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2071 MachinePointerInfo::getConstantPool(),
2076 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2077 SelectionDAG &DAG) const {
2078 EVT PtrVT = getPointerTy();
2079 DebugLoc dl = Op.getDebugLoc();
2080 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2081 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2082 MachineFunction &MF = DAG.getMachineFunction();
2083 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2085 if (Subtarget->useMovt()) {
2087 // FIXME: Once remat is capable of dealing with instructions with register
2088 // operands, expand this into two nodes.
2089 if (RelocM == Reloc::Static)
2090 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2091 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2093 unsigned Wrapper = (RelocM == Reloc::PIC_)
2094 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2095 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2096 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2097 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2098 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2099 MachinePointerInfo::getGOT(), false, false, 0);
2103 unsigned ARMPCLabelIndex = 0;
2105 if (RelocM == Reloc::Static) {
2106 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2108 ARMPCLabelIndex = AFI->createPICLabelUId();
2109 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2110 ARMConstantPoolValue *CPV =
2111 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2112 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2114 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2116 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2117 MachinePointerInfo::getConstantPool(),
2119 SDValue Chain = Result.getValue(1);
2121 if (RelocM == Reloc::PIC_) {
2122 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2123 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2126 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2127 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2133 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2134 SelectionDAG &DAG) const {
2135 assert(Subtarget->isTargetELF() &&
2136 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2137 MachineFunction &MF = DAG.getMachineFunction();
2138 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2139 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2140 EVT PtrVT = getPointerTy();
2141 DebugLoc dl = Op.getDebugLoc();
2142 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2143 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2144 "_GLOBAL_OFFSET_TABLE_",
2145 ARMPCLabelIndex, PCAdj);
2146 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2147 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2148 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2149 MachinePointerInfo::getConstantPool(),
2151 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2152 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2156 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2158 DebugLoc dl = Op.getDebugLoc();
2159 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2160 Op.getOperand(0), Op.getOperand(1));
2164 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2165 DebugLoc dl = Op.getDebugLoc();
2166 SDValue Val = DAG.getConstant(0, MVT::i32);
2167 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2168 Op.getOperand(1), Val);
2172 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2173 DebugLoc dl = Op.getDebugLoc();
2174 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2175 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2179 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2180 const ARMSubtarget *Subtarget) const {
2181 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2182 DebugLoc dl = Op.getDebugLoc();
2184 default: return SDValue(); // Don't custom lower most intrinsics.
2185 case Intrinsic::arm_thread_pointer: {
2186 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2187 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2189 case Intrinsic::eh_sjlj_lsda: {
2190 MachineFunction &MF = DAG.getMachineFunction();
2191 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2192 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2193 EVT PtrVT = getPointerTy();
2194 DebugLoc dl = Op.getDebugLoc();
2195 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2197 unsigned PCAdj = (RelocM != Reloc::PIC_)
2198 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2199 ARMConstantPoolValue *CPV =
2200 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2201 ARMCP::CPLSDA, PCAdj);
2202 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2203 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2205 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2206 MachinePointerInfo::getConstantPool(),
2209 if (RelocM == Reloc::PIC_) {
2210 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2211 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2215 case Intrinsic::arm_neon_vmulls:
2216 case Intrinsic::arm_neon_vmullu: {
2217 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2218 ? ARMISD::VMULLs : ARMISD::VMULLu;
2219 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2220 Op.getOperand(1), Op.getOperand(2));
2225 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2226 const ARMSubtarget *Subtarget) {
2227 DebugLoc dl = Op.getDebugLoc();
2228 if (!Subtarget->hasDataBarrier()) {
2229 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2230 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2232 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2233 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2234 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2235 DAG.getConstant(0, MVT::i32));
2238 SDValue Op5 = Op.getOperand(5);
2239 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2240 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2241 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2242 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2244 ARM_MB::MemBOpt DMBOpt;
2245 if (isDeviceBarrier)
2246 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2248 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2249 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2250 DAG.getConstant(DMBOpt, MVT::i32));
2253 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2254 const ARMSubtarget *Subtarget) {
2255 // ARM pre v5TE and Thumb1 does not have preload instructions.
2256 if (!(Subtarget->isThumb2() ||
2257 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2258 // Just preserve the chain.
2259 return Op.getOperand(0);
2261 DebugLoc dl = Op.getDebugLoc();
2262 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2264 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2265 // ARMv7 with MP extension has PLDW.
2266 return Op.getOperand(0);
2268 if (Subtarget->isThumb())
2270 isRead = ~isRead & 1;
2271 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2273 // Currently there is no intrinsic that matches pli.
2274 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2275 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2276 DAG.getConstant(isData, MVT::i32));
2279 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2280 MachineFunction &MF = DAG.getMachineFunction();
2281 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2283 // vastart just stores the address of the VarArgsFrameIndex slot into the
2284 // memory location argument.
2285 DebugLoc dl = Op.getDebugLoc();
2286 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2287 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2288 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2289 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2290 MachinePointerInfo(SV), false, false, 0);
2294 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2295 SDValue &Root, SelectionDAG &DAG,
2296 DebugLoc dl) const {
2297 MachineFunction &MF = DAG.getMachineFunction();
2298 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2300 TargetRegisterClass *RC;
2301 if (AFI->isThumb1OnlyFunction())
2302 RC = ARM::tGPRRegisterClass;
2304 RC = ARM::GPRRegisterClass;
2306 // Transform the arguments stored in physical registers into virtual ones.
2307 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2308 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2311 if (NextVA.isMemLoc()) {
2312 MachineFrameInfo *MFI = MF.getFrameInfo();
2313 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2315 // Create load node to retrieve arguments from the stack.
2316 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2317 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2318 MachinePointerInfo::getFixedStack(FI),
2321 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2322 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2325 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2329 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2330 unsigned &VARegSize, unsigned &VARegSaveSize)
2333 if (CCInfo.isFirstByValRegValid())
2334 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2336 unsigned int firstUnalloced;
2337 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2338 sizeof(GPRArgRegs) /
2339 sizeof(GPRArgRegs[0]));
2340 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2343 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2344 VARegSize = NumGPRs * 4;
2345 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2348 // The remaining GPRs hold either the beginning of variable-argument
2349 // data, or the beginning of an aggregate passed by value (usuall
2350 // byval). Either way, we allocate stack slots adjacent to the data
2351 // provided by our caller, and store the unallocated registers there.
2352 // If this is a variadic function, the va_list pointer will begin with
2353 // these values; otherwise, this reassembles a (byval) structure that
2354 // was split between registers and memory.
2356 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2357 DebugLoc dl, SDValue &Chain,
2358 unsigned ArgOffset) const {
2359 MachineFunction &MF = DAG.getMachineFunction();
2360 MachineFrameInfo *MFI = MF.getFrameInfo();
2361 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2362 unsigned firstRegToSaveIndex;
2363 if (CCInfo.isFirstByValRegValid())
2364 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2366 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2367 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2370 unsigned VARegSize, VARegSaveSize;
2371 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2372 if (VARegSaveSize) {
2373 // If this function is vararg, store any remaining integer argument regs
2374 // to their spots on the stack so that they may be loaded by deferencing
2375 // the result of va_next.
2376 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2377 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2378 ArgOffset + VARegSaveSize
2381 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2384 SmallVector<SDValue, 4> MemOps;
2385 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2386 TargetRegisterClass *RC;
2387 if (AFI->isThumb1OnlyFunction())
2388 RC = ARM::tGPRRegisterClass;
2390 RC = ARM::GPRRegisterClass;
2392 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2393 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2395 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2396 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2398 MemOps.push_back(Store);
2399 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2400 DAG.getConstant(4, getPointerTy()));
2402 if (!MemOps.empty())
2403 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2404 &MemOps[0], MemOps.size());
2406 // This will point to the next argument passed via stack.
2407 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2411 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2412 CallingConv::ID CallConv, bool isVarArg,
2413 const SmallVectorImpl<ISD::InputArg>
2415 DebugLoc dl, SelectionDAG &DAG,
2416 SmallVectorImpl<SDValue> &InVals)
2418 MachineFunction &MF = DAG.getMachineFunction();
2419 MachineFrameInfo *MFI = MF.getFrameInfo();
2421 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2423 // Assign locations to all of the incoming arguments.
2424 SmallVector<CCValAssign, 16> ArgLocs;
2425 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2427 CCInfo.setCallOrPrologue(Prologue);
2428 CCInfo.AnalyzeFormalArguments(Ins,
2429 CCAssignFnForNode(CallConv, /* Return*/ false,
2432 SmallVector<SDValue, 16> ArgValues;
2433 int lastInsIndex = -1;
2436 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2437 CCValAssign &VA = ArgLocs[i];
2439 // Arguments stored in registers.
2440 if (VA.isRegLoc()) {
2441 EVT RegVT = VA.getLocVT();
2443 if (VA.needsCustom()) {
2444 // f64 and vector types are split up into multiple registers or
2445 // combinations of registers and stack slots.
2446 if (VA.getLocVT() == MVT::v2f64) {
2447 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2449 VA = ArgLocs[++i]; // skip ahead to next loc
2451 if (VA.isMemLoc()) {
2452 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2453 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2454 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2455 MachinePointerInfo::getFixedStack(FI),
2458 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2461 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2462 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2463 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2464 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2465 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2467 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2470 TargetRegisterClass *RC;
2472 if (RegVT == MVT::f32)
2473 RC = ARM::SPRRegisterClass;
2474 else if (RegVT == MVT::f64)
2475 RC = ARM::DPRRegisterClass;
2476 else if (RegVT == MVT::v2f64)
2477 RC = ARM::QPRRegisterClass;
2478 else if (RegVT == MVT::i32)
2479 RC = (AFI->isThumb1OnlyFunction() ?
2480 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2482 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2484 // Transform the arguments in physical registers into virtual ones.
2485 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2486 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2489 // If this is an 8 or 16-bit value, it is really passed promoted
2490 // to 32 bits. Insert an assert[sz]ext to capture this, then
2491 // truncate to the right size.
2492 switch (VA.getLocInfo()) {
2493 default: llvm_unreachable("Unknown loc info!");
2494 case CCValAssign::Full: break;
2495 case CCValAssign::BCvt:
2496 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2498 case CCValAssign::SExt:
2499 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2500 DAG.getValueType(VA.getValVT()));
2501 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2503 case CCValAssign::ZExt:
2504 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2505 DAG.getValueType(VA.getValVT()));
2506 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2510 InVals.push_back(ArgValue);
2512 } else { // VA.isRegLoc()
2515 assert(VA.isMemLoc());
2516 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2518 int index = ArgLocs[i].getValNo();
2520 // Some Ins[] entries become multiple ArgLoc[] entries.
2521 // Process them only once.
2522 if (index != lastInsIndex)
2524 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2525 // FIXME: For now, all byval parameter objects are marked mutable.
2526 // This can be changed with more analysis.
2527 // In case of tail call optimization mark all arguments mutable.
2528 // Since they could be overwritten by lowering of arguments in case of
2530 if (Flags.isByVal()) {
2531 unsigned VARegSize, VARegSaveSize;
2532 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2533 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2534 unsigned Bytes = Flags.getByValSize() - VARegSize;
2535 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2536 int FI = MFI->CreateFixedObject(Bytes,
2537 VA.getLocMemOffset(), false);
2538 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2540 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2541 VA.getLocMemOffset(), true);
2543 // Create load nodes to retrieve arguments from the stack.
2544 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2545 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2546 MachinePointerInfo::getFixedStack(FI),
2549 lastInsIndex = index;
2556 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2561 /// isFloatingPointZero - Return true if this is +0.0.
2562 static bool isFloatingPointZero(SDValue Op) {
2563 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2564 return CFP->getValueAPF().isPosZero();
2565 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2566 // Maybe this has already been legalized into the constant pool?
2567 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2568 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2569 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2570 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2571 return CFP->getValueAPF().isPosZero();
2577 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2578 /// the given operands.
2580 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2581 SDValue &ARMcc, SelectionDAG &DAG,
2582 DebugLoc dl) const {
2583 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2584 unsigned C = RHSC->getZExtValue();
2585 if (!isLegalICmpImmediate(C)) {
2586 // Constant does not fit, try adjusting it by one?
2591 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2592 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2593 RHS = DAG.getConstant(C-1, MVT::i32);
2598 if (C != 0 && isLegalICmpImmediate(C-1)) {
2599 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2600 RHS = DAG.getConstant(C-1, MVT::i32);
2605 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2606 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2607 RHS = DAG.getConstant(C+1, MVT::i32);
2612 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2613 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2614 RHS = DAG.getConstant(C+1, MVT::i32);
2621 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2622 ARMISD::NodeType CompareType;
2625 CompareType = ARMISD::CMP;
2630 CompareType = ARMISD::CMPZ;
2633 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2634 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2637 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2639 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2640 DebugLoc dl) const {
2642 if (!isFloatingPointZero(RHS))
2643 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2645 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2646 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2649 /// duplicateCmp - Glue values can have only one use, so this function
2650 /// duplicates a comparison node.
2652 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2653 unsigned Opc = Cmp.getOpcode();
2654 DebugLoc DL = Cmp.getDebugLoc();
2655 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2656 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2658 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2659 Cmp = Cmp.getOperand(0);
2660 Opc = Cmp.getOpcode();
2661 if (Opc == ARMISD::CMPFP)
2662 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2664 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2665 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2667 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2670 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2671 SDValue Cond = Op.getOperand(0);
2672 SDValue SelectTrue = Op.getOperand(1);
2673 SDValue SelectFalse = Op.getOperand(2);
2674 DebugLoc dl = Op.getDebugLoc();
2678 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2679 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2681 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2682 const ConstantSDNode *CMOVTrue =
2683 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2684 const ConstantSDNode *CMOVFalse =
2685 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2687 if (CMOVTrue && CMOVFalse) {
2688 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2689 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2693 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2695 False = SelectFalse;
2696 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2701 if (True.getNode() && False.getNode()) {
2702 EVT VT = Op.getValueType();
2703 SDValue ARMcc = Cond.getOperand(2);
2704 SDValue CCR = Cond.getOperand(3);
2705 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2706 assert(True.getValueType() == VT);
2707 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2712 return DAG.getSelectCC(dl, Cond,
2713 DAG.getConstant(0, Cond.getValueType()),
2714 SelectTrue, SelectFalse, ISD::SETNE);
2717 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2718 EVT VT = Op.getValueType();
2719 SDValue LHS = Op.getOperand(0);
2720 SDValue RHS = Op.getOperand(1);
2721 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2722 SDValue TrueVal = Op.getOperand(2);
2723 SDValue FalseVal = Op.getOperand(3);
2724 DebugLoc dl = Op.getDebugLoc();
2726 if (LHS.getValueType() == MVT::i32) {
2728 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2729 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2730 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2733 ARMCC::CondCodes CondCode, CondCode2;
2734 FPCCToARMCC(CC, CondCode, CondCode2);
2736 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2737 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2738 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2739 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2741 if (CondCode2 != ARMCC::AL) {
2742 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2743 // FIXME: Needs another CMP because flag can have but one use.
2744 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2745 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2746 Result, TrueVal, ARMcc2, CCR, Cmp2);
2751 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2752 /// to morph to an integer compare sequence.
2753 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2754 const ARMSubtarget *Subtarget) {
2755 SDNode *N = Op.getNode();
2756 if (!N->hasOneUse())
2757 // Otherwise it requires moving the value from fp to integer registers.
2759 if (!N->getNumValues())
2761 EVT VT = Op.getValueType();
2762 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2763 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2764 // vmrs are very slow, e.g. cortex-a8.
2767 if (isFloatingPointZero(Op)) {
2771 return ISD::isNormalLoad(N);
2774 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2775 if (isFloatingPointZero(Op))
2776 return DAG.getConstant(0, MVT::i32);
2778 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2779 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2780 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2781 Ld->isVolatile(), Ld->isNonTemporal(),
2782 Ld->getAlignment());
2784 llvm_unreachable("Unknown VFP cmp argument!");
2787 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2788 SDValue &RetVal1, SDValue &RetVal2) {
2789 if (isFloatingPointZero(Op)) {
2790 RetVal1 = DAG.getConstant(0, MVT::i32);
2791 RetVal2 = DAG.getConstant(0, MVT::i32);
2795 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2796 SDValue Ptr = Ld->getBasePtr();
2797 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2798 Ld->getChain(), Ptr,
2799 Ld->getPointerInfo(),
2800 Ld->isVolatile(), Ld->isNonTemporal(),
2801 Ld->getAlignment());
2803 EVT PtrType = Ptr.getValueType();
2804 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2805 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2806 PtrType, Ptr, DAG.getConstant(4, PtrType));
2807 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2808 Ld->getChain(), NewPtr,
2809 Ld->getPointerInfo().getWithOffset(4),
2810 Ld->isVolatile(), Ld->isNonTemporal(),
2815 llvm_unreachable("Unknown VFP cmp argument!");
2818 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2819 /// f32 and even f64 comparisons to integer ones.
2821 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2822 SDValue Chain = Op.getOperand(0);
2823 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2824 SDValue LHS = Op.getOperand(2);
2825 SDValue RHS = Op.getOperand(3);
2826 SDValue Dest = Op.getOperand(4);
2827 DebugLoc dl = Op.getDebugLoc();
2829 bool SeenZero = false;
2830 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2831 canChangeToInt(RHS, SeenZero, Subtarget) &&
2832 // If one of the operand is zero, it's safe to ignore the NaN case since
2833 // we only care about equality comparisons.
2834 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2835 // If unsafe fp math optimization is enabled and there are no other uses of
2836 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2837 // to an integer comparison.
2838 if (CC == ISD::SETOEQ)
2840 else if (CC == ISD::SETUNE)
2844 if (LHS.getValueType() == MVT::f32) {
2845 LHS = bitcastf32Toi32(LHS, DAG);
2846 RHS = bitcastf32Toi32(RHS, DAG);
2847 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2848 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2849 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2850 Chain, Dest, ARMcc, CCR, Cmp);
2855 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2856 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2857 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2858 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2859 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2860 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2861 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2867 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2868 SDValue Chain = Op.getOperand(0);
2869 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2870 SDValue LHS = Op.getOperand(2);
2871 SDValue RHS = Op.getOperand(3);
2872 SDValue Dest = Op.getOperand(4);
2873 DebugLoc dl = Op.getDebugLoc();
2875 if (LHS.getValueType() == MVT::i32) {
2877 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2878 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2879 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2880 Chain, Dest, ARMcc, CCR, Cmp);
2883 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2886 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2887 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2888 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2889 if (Result.getNode())
2893 ARMCC::CondCodes CondCode, CondCode2;
2894 FPCCToARMCC(CC, CondCode, CondCode2);
2896 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2897 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2898 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2899 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2900 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2901 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2902 if (CondCode2 != ARMCC::AL) {
2903 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2904 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2905 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2910 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2911 SDValue Chain = Op.getOperand(0);
2912 SDValue Table = Op.getOperand(1);
2913 SDValue Index = Op.getOperand(2);
2914 DebugLoc dl = Op.getDebugLoc();
2916 EVT PTy = getPointerTy();
2917 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2918 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2919 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2920 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2921 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2922 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2923 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2924 if (Subtarget->isThumb2()) {
2925 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2926 // which does another jump to the destination. This also makes it easier
2927 // to translate it to TBB / TBH later.
2928 // FIXME: This might not work if the function is extremely large.
2929 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2930 Addr, Op.getOperand(2), JTI, UId);
2932 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2933 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2934 MachinePointerInfo::getJumpTable(),
2936 Chain = Addr.getValue(1);
2937 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2938 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2940 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2941 MachinePointerInfo::getJumpTable(), false, false, 0);
2942 Chain = Addr.getValue(1);
2943 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2947 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2948 DebugLoc dl = Op.getDebugLoc();
2951 switch (Op.getOpcode()) {
2953 assert(0 && "Invalid opcode!");
2954 case ISD::FP_TO_SINT:
2955 Opc = ARMISD::FTOSI;
2957 case ISD::FP_TO_UINT:
2958 Opc = ARMISD::FTOUI;
2961 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2962 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2965 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2966 EVT VT = Op.getValueType();
2967 DebugLoc dl = Op.getDebugLoc();
2969 EVT OperandVT = Op.getOperand(0).getValueType();
2970 assert(OperandVT == MVT::v4i16 && "Invalid type for custom lowering!");
2971 if (VT != MVT::v4f32)
2972 return DAG.UnrollVectorOp(Op.getNode());
2976 switch (Op.getOpcode()) {
2978 assert(0 && "Invalid opcode!");
2979 case ISD::SINT_TO_FP:
2980 CastOpc = ISD::SIGN_EXTEND;
2981 Opc = ISD::SINT_TO_FP;
2983 case ISD::UINT_TO_FP:
2984 CastOpc = ISD::ZERO_EXTEND;
2985 Opc = ISD::UINT_TO_FP;
2989 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
2990 return DAG.getNode(Opc, dl, VT, Op);
2993 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2994 EVT VT = Op.getValueType();
2996 return LowerVectorINT_TO_FP(Op, DAG);
2998 DebugLoc dl = Op.getDebugLoc();
3001 switch (Op.getOpcode()) {
3003 assert(0 && "Invalid opcode!");
3004 case ISD::SINT_TO_FP:
3005 Opc = ARMISD::SITOF;
3007 case ISD::UINT_TO_FP:
3008 Opc = ARMISD::UITOF;
3012 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3013 return DAG.getNode(Opc, dl, VT, Op);
3016 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3017 // Implement fcopysign with a fabs and a conditional fneg.
3018 SDValue Tmp0 = Op.getOperand(0);
3019 SDValue Tmp1 = Op.getOperand(1);
3020 DebugLoc dl = Op.getDebugLoc();
3021 EVT VT = Op.getValueType();
3022 EVT SrcVT = Tmp1.getValueType();
3023 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3024 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3025 bool UseNEON = !InGPR && Subtarget->hasNEON();
3028 // Use VBSL to copy the sign bit.
3029 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3030 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3031 DAG.getTargetConstant(EncodedVal, MVT::i32));
3032 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3034 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3035 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3036 DAG.getConstant(32, MVT::i32));
3037 else /*if (VT == MVT::f32)*/
3038 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3039 if (SrcVT == MVT::f32) {
3040 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3042 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3043 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3044 DAG.getConstant(32, MVT::i32));
3045 } else if (VT == MVT::f32)
3046 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3047 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3048 DAG.getConstant(32, MVT::i32));
3049 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3050 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3052 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3054 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3055 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3056 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3058 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3059 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3060 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3061 if (VT == MVT::f32) {
3062 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3063 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3064 DAG.getConstant(0, MVT::i32));
3066 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3072 // Bitcast operand 1 to i32.
3073 if (SrcVT == MVT::f64)
3074 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3075 &Tmp1, 1).getValue(1);
3076 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3078 // Or in the signbit with integer operations.
3079 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3080 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3081 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3082 if (VT == MVT::f32) {
3083 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3084 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3085 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3086 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3089 // f64: Or the high part with signbit and then combine two parts.
3090 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3092 SDValue Lo = Tmp0.getValue(0);
3093 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3094 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3095 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3098 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3099 MachineFunction &MF = DAG.getMachineFunction();
3100 MachineFrameInfo *MFI = MF.getFrameInfo();
3101 MFI->setReturnAddressIsTaken(true);
3103 EVT VT = Op.getValueType();
3104 DebugLoc dl = Op.getDebugLoc();
3105 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3107 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3108 SDValue Offset = DAG.getConstant(4, MVT::i32);
3109 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3110 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3111 MachinePointerInfo(), false, false, 0);
3114 // Return LR, which contains the return address. Mark it an implicit live-in.
3115 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3116 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3119 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3120 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3121 MFI->setFrameAddressIsTaken(true);
3123 EVT VT = Op.getValueType();
3124 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3125 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3126 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3127 ? ARM::R7 : ARM::R11;
3128 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3130 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3131 MachinePointerInfo(),
3136 /// ExpandBITCAST - If the target supports VFP, this function is called to
3137 /// expand a bit convert where either the source or destination type is i64 to
3138 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3139 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3140 /// vectors), since the legalizer won't know what to do with that.
3141 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3142 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3143 DebugLoc dl = N->getDebugLoc();
3144 SDValue Op = N->getOperand(0);
3146 // This function is only supposed to be called for i64 types, either as the
3147 // source or destination of the bit convert.
3148 EVT SrcVT = Op.getValueType();
3149 EVT DstVT = N->getValueType(0);
3150 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3151 "ExpandBITCAST called for non-i64 type");
3153 // Turn i64->f64 into VMOVDRR.
3154 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3155 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3156 DAG.getConstant(0, MVT::i32));
3157 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3158 DAG.getConstant(1, MVT::i32));
3159 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3160 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3163 // Turn f64->i64 into VMOVRRD.
3164 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3165 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3166 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3167 // Merge the pieces into a single i64 value.
3168 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3174 /// getZeroVector - Returns a vector of specified type with all zero elements.
3175 /// Zero vectors are used to represent vector negation and in those cases
3176 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3177 /// not support i64 elements, so sometimes the zero vectors will need to be
3178 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3180 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3181 assert(VT.isVector() && "Expected a vector type");
3182 // The canonical modified immediate encoding of a zero vector is....0!
3183 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3184 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3185 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3186 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3189 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3190 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3191 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3192 SelectionDAG &DAG) const {
3193 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3194 EVT VT = Op.getValueType();
3195 unsigned VTBits = VT.getSizeInBits();
3196 DebugLoc dl = Op.getDebugLoc();
3197 SDValue ShOpLo = Op.getOperand(0);
3198 SDValue ShOpHi = Op.getOperand(1);
3199 SDValue ShAmt = Op.getOperand(2);
3201 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3203 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3205 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3206 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3207 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3208 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3209 DAG.getConstant(VTBits, MVT::i32));
3210 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3211 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3212 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3214 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3215 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3217 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3218 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3221 SDValue Ops[2] = { Lo, Hi };
3222 return DAG.getMergeValues(Ops, 2, dl);
3225 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3226 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3227 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3228 SelectionDAG &DAG) const {
3229 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3230 EVT VT = Op.getValueType();
3231 unsigned VTBits = VT.getSizeInBits();
3232 DebugLoc dl = Op.getDebugLoc();
3233 SDValue ShOpLo = Op.getOperand(0);
3234 SDValue ShOpHi = Op.getOperand(1);
3235 SDValue ShAmt = Op.getOperand(2);
3238 assert(Op.getOpcode() == ISD::SHL_PARTS);
3239 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3240 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3241 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3242 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3243 DAG.getConstant(VTBits, MVT::i32));
3244 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3245 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3247 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3248 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3249 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3251 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3252 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3255 SDValue Ops[2] = { Lo, Hi };
3256 return DAG.getMergeValues(Ops, 2, dl);
3259 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3260 SelectionDAG &DAG) const {
3261 // The rounding mode is in bits 23:22 of the FPSCR.
3262 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3263 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3264 // so that the shift + and get folded into a bitfield extract.
3265 DebugLoc dl = Op.getDebugLoc();
3266 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3267 DAG.getConstant(Intrinsic::arm_get_fpscr,
3269 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3270 DAG.getConstant(1U << 22, MVT::i32));
3271 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3272 DAG.getConstant(22, MVT::i32));
3273 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3274 DAG.getConstant(3, MVT::i32));
3277 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3278 const ARMSubtarget *ST) {
3279 EVT VT = N->getValueType(0);
3280 DebugLoc dl = N->getDebugLoc();
3282 if (!ST->hasV6T2Ops())
3285 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3286 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3289 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3290 const ARMSubtarget *ST) {
3291 EVT VT = N->getValueType(0);
3292 DebugLoc dl = N->getDebugLoc();
3297 // Lower vector shifts on NEON to use VSHL.
3298 assert(ST->hasNEON() && "unexpected vector shift");
3300 // Left shifts translate directly to the vshiftu intrinsic.
3301 if (N->getOpcode() == ISD::SHL)
3302 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3303 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3304 N->getOperand(0), N->getOperand(1));
3306 assert((N->getOpcode() == ISD::SRA ||
3307 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3309 // NEON uses the same intrinsics for both left and right shifts. For
3310 // right shifts, the shift amounts are negative, so negate the vector of
3312 EVT ShiftVT = N->getOperand(1).getValueType();
3313 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3314 getZeroVector(ShiftVT, DAG, dl),
3316 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3317 Intrinsic::arm_neon_vshifts :
3318 Intrinsic::arm_neon_vshiftu);
3319 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3320 DAG.getConstant(vshiftInt, MVT::i32),
3321 N->getOperand(0), NegatedCount);
3324 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3325 const ARMSubtarget *ST) {
3326 EVT VT = N->getValueType(0);
3327 DebugLoc dl = N->getDebugLoc();
3329 // We can get here for a node like i32 = ISD::SHL i32, i64
3333 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3334 "Unknown shift to lower!");
3336 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3337 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3338 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3341 // If we are in thumb mode, we don't have RRX.
3342 if (ST->isThumb1Only()) return SDValue();
3344 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3345 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3346 DAG.getConstant(0, MVT::i32));
3347 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3348 DAG.getConstant(1, MVT::i32));
3350 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3351 // captures the result into a carry flag.
3352 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3353 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3355 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3356 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3358 // Merge the pieces into a single i64 value.
3359 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3362 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3363 SDValue TmpOp0, TmpOp1;
3364 bool Invert = false;
3368 SDValue Op0 = Op.getOperand(0);
3369 SDValue Op1 = Op.getOperand(1);
3370 SDValue CC = Op.getOperand(2);
3371 EVT VT = Op.getValueType();
3372 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3373 DebugLoc dl = Op.getDebugLoc();
3375 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3376 switch (SetCCOpcode) {
3377 default: llvm_unreachable("Illegal FP comparison"); break;
3379 case ISD::SETNE: Invert = true; // Fallthrough
3381 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3383 case ISD::SETLT: Swap = true; // Fallthrough
3385 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3387 case ISD::SETLE: Swap = true; // Fallthrough
3389 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3390 case ISD::SETUGE: Swap = true; // Fallthrough
3391 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3392 case ISD::SETUGT: Swap = true; // Fallthrough
3393 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3394 case ISD::SETUEQ: Invert = true; // Fallthrough
3396 // Expand this to (OLT | OGT).
3400 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3401 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3403 case ISD::SETUO: Invert = true; // Fallthrough
3405 // Expand this to (OLT | OGE).
3409 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3410 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3414 // Integer comparisons.
3415 switch (SetCCOpcode) {
3416 default: llvm_unreachable("Illegal integer comparison"); break;
3417 case ISD::SETNE: Invert = true;
3418 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3419 case ISD::SETLT: Swap = true;
3420 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3421 case ISD::SETLE: Swap = true;
3422 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3423 case ISD::SETULT: Swap = true;
3424 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3425 case ISD::SETULE: Swap = true;
3426 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3429 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3430 if (Opc == ARMISD::VCEQ) {
3433 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3435 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3438 // Ignore bitconvert.
3439 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3440 AndOp = AndOp.getOperand(0);
3442 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3444 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3445 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3452 std::swap(Op0, Op1);
3454 // If one of the operands is a constant vector zero, attempt to fold the
3455 // comparison to a specialized compare-against-zero form.
3457 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3459 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3460 if (Opc == ARMISD::VCGE)
3461 Opc = ARMISD::VCLEZ;
3462 else if (Opc == ARMISD::VCGT)
3463 Opc = ARMISD::VCLTZ;
3468 if (SingleOp.getNode()) {
3471 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3473 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3475 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3477 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3479 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3481 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3484 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3488 Result = DAG.getNOT(dl, Result, VT);
3493 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3494 /// valid vector constant for a NEON instruction with a "modified immediate"
3495 /// operand (e.g., VMOV). If so, return the encoded value.
3496 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3497 unsigned SplatBitSize, SelectionDAG &DAG,
3498 EVT &VT, bool is128Bits, NEONModImmType type) {
3499 unsigned OpCmode, Imm;
3501 // SplatBitSize is set to the smallest size that splats the vector, so a
3502 // zero vector will always have SplatBitSize == 8. However, NEON modified
3503 // immediate instructions others than VMOV do not support the 8-bit encoding
3504 // of a zero vector, and the default encoding of zero is supposed to be the
3509 switch (SplatBitSize) {
3511 if (type != VMOVModImm)
3513 // Any 1-byte value is OK. Op=0, Cmode=1110.
3514 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3517 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3521 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3522 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3523 if ((SplatBits & ~0xff) == 0) {
3524 // Value = 0x00nn: Op=x, Cmode=100x.
3529 if ((SplatBits & ~0xff00) == 0) {
3530 // Value = 0xnn00: Op=x, Cmode=101x.
3532 Imm = SplatBits >> 8;
3538 // NEON's 32-bit VMOV supports splat values where:
3539 // * only one byte is nonzero, or
3540 // * the least significant byte is 0xff and the second byte is nonzero, or
3541 // * the least significant 2 bytes are 0xff and the third is nonzero.
3542 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3543 if ((SplatBits & ~0xff) == 0) {
3544 // Value = 0x000000nn: Op=x, Cmode=000x.
3549 if ((SplatBits & ~0xff00) == 0) {
3550 // Value = 0x0000nn00: Op=x, Cmode=001x.
3552 Imm = SplatBits >> 8;
3555 if ((SplatBits & ~0xff0000) == 0) {
3556 // Value = 0x00nn0000: Op=x, Cmode=010x.
3558 Imm = SplatBits >> 16;
3561 if ((SplatBits & ~0xff000000) == 0) {
3562 // Value = 0xnn000000: Op=x, Cmode=011x.
3564 Imm = SplatBits >> 24;
3568 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3569 if (type == OtherModImm) return SDValue();
3571 if ((SplatBits & ~0xffff) == 0 &&
3572 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3573 // Value = 0x0000nnff: Op=x, Cmode=1100.
3575 Imm = SplatBits >> 8;
3580 if ((SplatBits & ~0xffffff) == 0 &&
3581 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3582 // Value = 0x00nnffff: Op=x, Cmode=1101.
3584 Imm = SplatBits >> 16;
3585 SplatBits |= 0xffff;
3589 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3590 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3591 // VMOV.I32. A (very) minor optimization would be to replicate the value
3592 // and fall through here to test for a valid 64-bit splat. But, then the
3593 // caller would also need to check and handle the change in size.
3597 if (type != VMOVModImm)
3599 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3600 uint64_t BitMask = 0xff;
3602 unsigned ImmMask = 1;
3604 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3605 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3608 } else if ((SplatBits & BitMask) != 0) {
3614 // Op=1, Cmode=1110.
3617 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3622 llvm_unreachable("unexpected size for isNEONModifiedImm");
3626 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3627 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3630 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3631 bool &ReverseVEXT, unsigned &Imm) {
3632 unsigned NumElts = VT.getVectorNumElements();
3633 ReverseVEXT = false;
3635 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3641 // If this is a VEXT shuffle, the immediate value is the index of the first
3642 // element. The other shuffle indices must be the successive elements after
3644 unsigned ExpectedElt = Imm;
3645 for (unsigned i = 1; i < NumElts; ++i) {
3646 // Increment the expected index. If it wraps around, it may still be
3647 // a VEXT but the source vectors must be swapped.
3649 if (ExpectedElt == NumElts * 2) {
3654 if (M[i] < 0) continue; // ignore UNDEF indices
3655 if (ExpectedElt != static_cast<unsigned>(M[i]))
3659 // Adjust the index value if the source operands will be swapped.
3666 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3667 /// instruction with the specified blocksize. (The order of the elements
3668 /// within each block of the vector is reversed.)
3669 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3670 unsigned BlockSize) {
3671 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3672 "Only possible block sizes for VREV are: 16, 32, 64");
3674 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3678 unsigned NumElts = VT.getVectorNumElements();
3679 unsigned BlockElts = M[0] + 1;
3680 // If the first shuffle index is UNDEF, be optimistic.
3682 BlockElts = BlockSize / EltSz;
3684 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3687 for (unsigned i = 0; i < NumElts; ++i) {
3688 if (M[i] < 0) continue; // ignore UNDEF indices
3689 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3696 static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3697 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3698 // range, then 0 is placed into the resulting vector. So pretty much any mask
3699 // of 8 elements can work here.
3700 return VT == MVT::v8i8 && M.size() == 8;
3703 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3704 unsigned &WhichResult) {
3705 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3709 unsigned NumElts = VT.getVectorNumElements();
3710 WhichResult = (M[0] == 0 ? 0 : 1);
3711 for (unsigned i = 0; i < NumElts; i += 2) {
3712 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3713 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3719 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3720 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3721 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3722 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3723 unsigned &WhichResult) {
3724 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3728 unsigned NumElts = VT.getVectorNumElements();
3729 WhichResult = (M[0] == 0 ? 0 : 1);
3730 for (unsigned i = 0; i < NumElts; i += 2) {
3731 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3732 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3738 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3739 unsigned &WhichResult) {
3740 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3744 unsigned NumElts = VT.getVectorNumElements();
3745 WhichResult = (M[0] == 0 ? 0 : 1);
3746 for (unsigned i = 0; i != NumElts; ++i) {
3747 if (M[i] < 0) continue; // ignore UNDEF indices
3748 if ((unsigned) M[i] != 2 * i + WhichResult)
3752 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3753 if (VT.is64BitVector() && EltSz == 32)
3759 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3760 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3761 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3762 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3763 unsigned &WhichResult) {
3764 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3768 unsigned Half = VT.getVectorNumElements() / 2;
3769 WhichResult = (M[0] == 0 ? 0 : 1);
3770 for (unsigned j = 0; j != 2; ++j) {
3771 unsigned Idx = WhichResult;
3772 for (unsigned i = 0; i != Half; ++i) {
3773 int MIdx = M[i + j * Half];
3774 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3780 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3781 if (VT.is64BitVector() && EltSz == 32)
3787 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3788 unsigned &WhichResult) {
3789 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3793 unsigned NumElts = VT.getVectorNumElements();
3794 WhichResult = (M[0] == 0 ? 0 : 1);
3795 unsigned Idx = WhichResult * NumElts / 2;
3796 for (unsigned i = 0; i != NumElts; i += 2) {
3797 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3798 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3803 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3804 if (VT.is64BitVector() && EltSz == 32)
3810 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3811 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3812 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3813 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3814 unsigned &WhichResult) {
3815 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3819 unsigned NumElts = VT.getVectorNumElements();
3820 WhichResult = (M[0] == 0 ? 0 : 1);
3821 unsigned Idx = WhichResult * NumElts / 2;
3822 for (unsigned i = 0; i != NumElts; i += 2) {
3823 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3824 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3829 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3830 if (VT.is64BitVector() && EltSz == 32)
3836 // If N is an integer constant that can be moved into a register in one
3837 // instruction, return an SDValue of such a constant (will become a MOV
3838 // instruction). Otherwise return null.
3839 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3840 const ARMSubtarget *ST, DebugLoc dl) {
3842 if (!isa<ConstantSDNode>(N))
3844 Val = cast<ConstantSDNode>(N)->getZExtValue();
3846 if (ST->isThumb1Only()) {
3847 if (Val <= 255 || ~Val <= 255)
3848 return DAG.getConstant(Val, MVT::i32);
3850 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3851 return DAG.getConstant(Val, MVT::i32);
3856 // If this is a case we can't handle, return null and let the default
3857 // expansion code take care of it.
3858 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3859 const ARMSubtarget *ST) const {
3860 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3861 DebugLoc dl = Op.getDebugLoc();
3862 EVT VT = Op.getValueType();
3864 APInt SplatBits, SplatUndef;
3865 unsigned SplatBitSize;
3867 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3868 if (SplatBitSize <= 64) {
3869 // Check if an immediate VMOV works.
3871 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3872 SplatUndef.getZExtValue(), SplatBitSize,
3873 DAG, VmovVT, VT.is128BitVector(),
3875 if (Val.getNode()) {
3876 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3877 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3880 // Try an immediate VMVN.
3881 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3882 ((1LL << SplatBitSize) - 1));
3883 Val = isNEONModifiedImm(NegatedImm,
3884 SplatUndef.getZExtValue(), SplatBitSize,
3885 DAG, VmovVT, VT.is128BitVector(),
3887 if (Val.getNode()) {
3888 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3889 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3894 // Scan through the operands to see if only one value is used.
3895 unsigned NumElts = VT.getVectorNumElements();
3896 bool isOnlyLowElement = true;
3897 bool usesOnlyOneValue = true;
3898 bool isConstant = true;
3900 for (unsigned i = 0; i < NumElts; ++i) {
3901 SDValue V = Op.getOperand(i);
3902 if (V.getOpcode() == ISD::UNDEF)
3905 isOnlyLowElement = false;
3906 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3909 if (!Value.getNode())
3911 else if (V != Value)
3912 usesOnlyOneValue = false;
3915 if (!Value.getNode())
3916 return DAG.getUNDEF(VT);
3918 if (isOnlyLowElement)
3919 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3921 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3923 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3924 // i32 and try again.
3925 if (usesOnlyOneValue && EltSize <= 32) {
3927 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3928 if (VT.getVectorElementType().isFloatingPoint()) {
3929 SmallVector<SDValue, 8> Ops;
3930 for (unsigned i = 0; i < NumElts; ++i)
3931 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3933 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3934 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3935 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3937 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3939 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3941 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3944 // If all elements are constants and the case above didn't get hit, fall back
3945 // to the default expansion, which will generate a load from the constant
3950 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3952 SDValue shuffle = ReconstructShuffle(Op, DAG);
3953 if (shuffle != SDValue())
3957 // Vectors with 32- or 64-bit elements can be built by directly assigning
3958 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3959 // will be legalized.
3960 if (EltSize >= 32) {
3961 // Do the expansion with floating-point types, since that is what the VFP
3962 // registers are defined to use, and since i64 is not legal.
3963 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3964 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3965 SmallVector<SDValue, 8> Ops;
3966 for (unsigned i = 0; i < NumElts; ++i)
3967 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3968 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3969 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3975 // Gather data to see if the operation can be modelled as a
3976 // shuffle in combination with VEXTs.
3977 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
3978 SelectionDAG &DAG) const {
3979 DebugLoc dl = Op.getDebugLoc();
3980 EVT VT = Op.getValueType();
3981 unsigned NumElts = VT.getVectorNumElements();
3983 SmallVector<SDValue, 2> SourceVecs;
3984 SmallVector<unsigned, 2> MinElts;
3985 SmallVector<unsigned, 2> MaxElts;
3987 for (unsigned i = 0; i < NumElts; ++i) {
3988 SDValue V = Op.getOperand(i);
3989 if (V.getOpcode() == ISD::UNDEF)
3991 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3992 // A shuffle can only come from building a vector from various
3993 // elements of other vectors.
3997 // Record this extraction against the appropriate vector if possible...
3998 SDValue SourceVec = V.getOperand(0);
3999 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4000 bool FoundSource = false;
4001 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4002 if (SourceVecs[j] == SourceVec) {
4003 if (MinElts[j] > EltNo)
4005 if (MaxElts[j] < EltNo)
4012 // Or record a new source if not...
4014 SourceVecs.push_back(SourceVec);
4015 MinElts.push_back(EltNo);
4016 MaxElts.push_back(EltNo);
4020 // Currently only do something sane when at most two source vectors
4022 if (SourceVecs.size() > 2)
4025 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4026 int VEXTOffsets[2] = {0, 0};
4028 // This loop extracts the usage patterns of the source vectors
4029 // and prepares appropriate SDValues for a shuffle if possible.
4030 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4031 if (SourceVecs[i].getValueType() == VT) {
4032 // No VEXT necessary
4033 ShuffleSrcs[i] = SourceVecs[i];
4036 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4037 // It probably isn't worth padding out a smaller vector just to
4038 // break it down again in a shuffle.
4042 // Since only 64-bit and 128-bit vectors are legal on ARM and
4043 // we've eliminated the other cases...
4044 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4045 "unexpected vector sizes in ReconstructShuffle");
4047 if (MaxElts[i] - MinElts[i] >= NumElts) {
4048 // Span too large for a VEXT to cope
4052 if (MinElts[i] >= NumElts) {
4053 // The extraction can just take the second half
4054 VEXTOffsets[i] = NumElts;
4055 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4057 DAG.getIntPtrConstant(NumElts));
4058 } else if (MaxElts[i] < NumElts) {
4059 // The extraction can just take the first half
4061 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4063 DAG.getIntPtrConstant(0));
4065 // An actual VEXT is needed
4066 VEXTOffsets[i] = MinElts[i];
4067 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4069 DAG.getIntPtrConstant(0));
4070 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4072 DAG.getIntPtrConstant(NumElts));
4073 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4074 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4078 SmallVector<int, 8> Mask;
4080 for (unsigned i = 0; i < NumElts; ++i) {
4081 SDValue Entry = Op.getOperand(i);
4082 if (Entry.getOpcode() == ISD::UNDEF) {
4087 SDValue ExtractVec = Entry.getOperand(0);
4088 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4089 .getOperand(1))->getSExtValue();
4090 if (ExtractVec == SourceVecs[0]) {
4091 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4093 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4097 // Final check before we try to produce nonsense...
4098 if (isShuffleMaskLegal(Mask, VT))
4099 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4105 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4106 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4107 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4108 /// are assumed to be legal.
4110 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4112 if (VT.getVectorNumElements() == 4 &&
4113 (VT.is128BitVector() || VT.is64BitVector())) {
4114 unsigned PFIndexes[4];
4115 for (unsigned i = 0; i != 4; ++i) {
4119 PFIndexes[i] = M[i];
4122 // Compute the index in the perfect shuffle table.
4123 unsigned PFTableIndex =
4124 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4125 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4126 unsigned Cost = (PFEntry >> 30);
4133 unsigned Imm, WhichResult;
4135 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4136 return (EltSize >= 32 ||
4137 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4138 isVREVMask(M, VT, 64) ||
4139 isVREVMask(M, VT, 32) ||
4140 isVREVMask(M, VT, 16) ||
4141 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4142 isVTBLMask(M, VT) ||
4143 isVTRNMask(M, VT, WhichResult) ||
4144 isVUZPMask(M, VT, WhichResult) ||
4145 isVZIPMask(M, VT, WhichResult) ||
4146 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4147 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4148 isVZIP_v_undef_Mask(M, VT, WhichResult));
4151 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4152 /// the specified operations to build the shuffle.
4153 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4154 SDValue RHS, SelectionDAG &DAG,
4156 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4157 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4158 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4161 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4170 OP_VUZPL, // VUZP, left result
4171 OP_VUZPR, // VUZP, right result
4172 OP_VZIPL, // VZIP, left result
4173 OP_VZIPR, // VZIP, right result
4174 OP_VTRNL, // VTRN, left result
4175 OP_VTRNR // VTRN, right result
4178 if (OpNum == OP_COPY) {
4179 if (LHSID == (1*9+2)*9+3) return LHS;
4180 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4184 SDValue OpLHS, OpRHS;
4185 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4186 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4187 EVT VT = OpLHS.getValueType();
4190 default: llvm_unreachable("Unknown shuffle opcode!");
4192 // VREV divides the vector in half and swaps within the half.
4193 if (VT.getVectorElementType() == MVT::i32 ||
4194 VT.getVectorElementType() == MVT::f32)
4195 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4196 // vrev <4 x i16> -> VREV32
4197 if (VT.getVectorElementType() == MVT::i16)
4198 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4199 // vrev <4 x i8> -> VREV16
4200 assert(VT.getVectorElementType() == MVT::i8);
4201 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4206 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4207 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4211 return DAG.getNode(ARMISD::VEXT, dl, VT,
4213 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4216 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4217 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4220 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4221 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4224 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4225 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4229 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4230 SmallVectorImpl<int> &ShuffleMask,
4231 SelectionDAG &DAG) {
4232 // Check to see if we can use the VTBL instruction.
4233 SDValue V1 = Op.getOperand(0);
4234 SDValue V2 = Op.getOperand(1);
4235 DebugLoc DL = Op.getDebugLoc();
4237 SmallVector<SDValue, 8> VTBLMask;
4238 for (SmallVectorImpl<int>::iterator
4239 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4240 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4242 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4243 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4244 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4247 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4248 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4252 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4253 SDValue V1 = Op.getOperand(0);
4254 SDValue V2 = Op.getOperand(1);
4255 DebugLoc dl = Op.getDebugLoc();
4256 EVT VT = Op.getValueType();
4257 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4258 SmallVector<int, 8> ShuffleMask;
4260 // Convert shuffles that are directly supported on NEON to target-specific
4261 // DAG nodes, instead of keeping them as shuffles and matching them again
4262 // during code selection. This is more efficient and avoids the possibility
4263 // of inconsistencies between legalization and selection.
4264 // FIXME: floating-point vectors should be canonicalized to integer vectors
4265 // of the same time so that they get CSEd properly.
4266 SVN->getMask(ShuffleMask);
4268 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4269 if (EltSize <= 32) {
4270 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4271 int Lane = SVN->getSplatIndex();
4272 // If this is undef splat, generate it via "just" vdup, if possible.
4273 if (Lane == -1) Lane = 0;
4275 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4276 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4278 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4279 DAG.getConstant(Lane, MVT::i32));
4284 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4287 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4288 DAG.getConstant(Imm, MVT::i32));
4291 if (isVREVMask(ShuffleMask, VT, 64))
4292 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4293 if (isVREVMask(ShuffleMask, VT, 32))
4294 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4295 if (isVREVMask(ShuffleMask, VT, 16))
4296 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4298 // Check for Neon shuffles that modify both input vectors in place.
4299 // If both results are used, i.e., if there are two shuffles with the same
4300 // source operands and with masks corresponding to both results of one of
4301 // these operations, DAG memoization will ensure that a single node is
4302 // used for both shuffles.
4303 unsigned WhichResult;
4304 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4305 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4306 V1, V2).getValue(WhichResult);
4307 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4308 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4309 V1, V2).getValue(WhichResult);
4310 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4311 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4312 V1, V2).getValue(WhichResult);
4314 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4315 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4316 V1, V1).getValue(WhichResult);
4317 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4318 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4319 V1, V1).getValue(WhichResult);
4320 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4321 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4322 V1, V1).getValue(WhichResult);
4325 // If the shuffle is not directly supported and it has 4 elements, use
4326 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4327 unsigned NumElts = VT.getVectorNumElements();
4329 unsigned PFIndexes[4];
4330 for (unsigned i = 0; i != 4; ++i) {
4331 if (ShuffleMask[i] < 0)
4334 PFIndexes[i] = ShuffleMask[i];
4337 // Compute the index in the perfect shuffle table.
4338 unsigned PFTableIndex =
4339 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4340 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4341 unsigned Cost = (PFEntry >> 30);
4344 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4347 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4348 if (EltSize >= 32) {
4349 // Do the expansion with floating-point types, since that is what the VFP
4350 // registers are defined to use, and since i64 is not legal.
4351 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4352 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4353 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4354 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4355 SmallVector<SDValue, 8> Ops;
4356 for (unsigned i = 0; i < NumElts; ++i) {
4357 if (ShuffleMask[i] < 0)
4358 Ops.push_back(DAG.getUNDEF(EltVT));
4360 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4361 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4362 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4365 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4366 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4369 if (VT == MVT::v8i8) {
4370 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4371 if (NewOp.getNode())
4378 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4379 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4380 SDValue Lane = Op.getOperand(1);
4381 if (!isa<ConstantSDNode>(Lane))
4384 SDValue Vec = Op.getOperand(0);
4385 if (Op.getValueType() == MVT::i32 &&
4386 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4387 DebugLoc dl = Op.getDebugLoc();
4388 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4394 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4395 // The only time a CONCAT_VECTORS operation can have legal types is when
4396 // two 64-bit vectors are concatenated to a 128-bit vector.
4397 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4398 "unexpected CONCAT_VECTORS");
4399 DebugLoc dl = Op.getDebugLoc();
4400 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4401 SDValue Op0 = Op.getOperand(0);
4402 SDValue Op1 = Op.getOperand(1);
4403 if (Op0.getOpcode() != ISD::UNDEF)
4404 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4405 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4406 DAG.getIntPtrConstant(0));
4407 if (Op1.getOpcode() != ISD::UNDEF)
4408 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4409 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4410 DAG.getIntPtrConstant(1));
4411 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4414 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4415 /// element has been zero/sign-extended, depending on the isSigned parameter,
4416 /// from an integer type half its size.
4417 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4419 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4420 EVT VT = N->getValueType(0);
4421 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4422 SDNode *BVN = N->getOperand(0).getNode();
4423 if (BVN->getValueType(0) != MVT::v4i32 ||
4424 BVN->getOpcode() != ISD::BUILD_VECTOR)
4426 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4427 unsigned HiElt = 1 - LoElt;
4428 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4429 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4430 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4431 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4432 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4435 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4436 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4439 if (Hi0->isNullValue() && Hi1->isNullValue())
4445 if (N->getOpcode() != ISD::BUILD_VECTOR)
4448 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4449 SDNode *Elt = N->getOperand(i).getNode();
4450 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4451 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4452 unsigned HalfSize = EltSize / 2;
4454 int64_t SExtVal = C->getSExtValue();
4455 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4458 if ((C->getZExtValue() >> HalfSize) != 0)
4469 /// isSignExtended - Check if a node is a vector value that is sign-extended
4470 /// or a constant BUILD_VECTOR with sign-extended elements.
4471 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4472 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4474 if (isExtendedBUILD_VECTOR(N, DAG, true))
4479 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4480 /// or a constant BUILD_VECTOR with zero-extended elements.
4481 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4482 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4484 if (isExtendedBUILD_VECTOR(N, DAG, false))
4489 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4490 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4491 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4492 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4493 return N->getOperand(0);
4494 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4495 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4496 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4497 LD->isNonTemporal(), LD->getAlignment());
4498 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4499 // have been legalized as a BITCAST from v4i32.
4500 if (N->getOpcode() == ISD::BITCAST) {
4501 SDNode *BVN = N->getOperand(0).getNode();
4502 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4503 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4504 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4505 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4506 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4508 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4509 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4510 EVT VT = N->getValueType(0);
4511 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4512 unsigned NumElts = VT.getVectorNumElements();
4513 MVT TruncVT = MVT::getIntegerVT(EltSize);
4514 SmallVector<SDValue, 8> Ops;
4515 for (unsigned i = 0; i != NumElts; ++i) {
4516 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4517 const APInt &CInt = C->getAPIntValue();
4518 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4520 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4521 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4524 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4525 unsigned Opcode = N->getOpcode();
4526 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4527 SDNode *N0 = N->getOperand(0).getNode();
4528 SDNode *N1 = N->getOperand(1).getNode();
4529 return N0->hasOneUse() && N1->hasOneUse() &&
4530 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4535 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4536 unsigned Opcode = N->getOpcode();
4537 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4538 SDNode *N0 = N->getOperand(0).getNode();
4539 SDNode *N1 = N->getOperand(1).getNode();
4540 return N0->hasOneUse() && N1->hasOneUse() &&
4541 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4546 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4547 // Multiplications are only custom-lowered for 128-bit vectors so that
4548 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4549 EVT VT = Op.getValueType();
4550 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4551 SDNode *N0 = Op.getOperand(0).getNode();
4552 SDNode *N1 = Op.getOperand(1).getNode();
4553 unsigned NewOpc = 0;
4555 bool isN0SExt = isSignExtended(N0, DAG);
4556 bool isN1SExt = isSignExtended(N1, DAG);
4557 if (isN0SExt && isN1SExt)
4558 NewOpc = ARMISD::VMULLs;
4560 bool isN0ZExt = isZeroExtended(N0, DAG);
4561 bool isN1ZExt = isZeroExtended(N1, DAG);
4562 if (isN0ZExt && isN1ZExt)
4563 NewOpc = ARMISD::VMULLu;
4564 else if (isN1SExt || isN1ZExt) {
4565 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4566 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4567 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4568 NewOpc = ARMISD::VMULLs;
4570 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4571 NewOpc = ARMISD::VMULLu;
4573 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4575 NewOpc = ARMISD::VMULLu;
4581 if (VT == MVT::v2i64)
4582 // Fall through to expand this. It is not legal.
4585 // Other vector multiplications are legal.
4590 // Legalize to a VMULL instruction.
4591 DebugLoc DL = Op.getDebugLoc();
4593 SDValue Op1 = SkipExtension(N1, DAG);
4595 Op0 = SkipExtension(N0, DAG);
4596 assert(Op0.getValueType().is64BitVector() &&
4597 Op1.getValueType().is64BitVector() &&
4598 "unexpected types for extended operands to VMULL");
4599 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4602 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4603 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4610 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4611 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4612 EVT Op1VT = Op1.getValueType();
4613 return DAG.getNode(N0->getOpcode(), DL, VT,
4614 DAG.getNode(NewOpc, DL, VT,
4615 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4616 DAG.getNode(NewOpc, DL, VT,
4617 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4621 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4623 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4624 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4625 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4626 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4627 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4628 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4629 // Get reciprocal estimate.
4630 // float4 recip = vrecpeq_f32(yf);
4631 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4632 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4633 // Because char has a smaller range than uchar, we can actually get away
4634 // without any newton steps. This requires that we use a weird bias
4635 // of 0xb000, however (again, this has been exhaustively tested).
4636 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4637 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4638 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4639 Y = DAG.getConstant(0xb000, MVT::i32);
4640 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4641 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4642 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4643 // Convert back to short.
4644 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4645 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4650 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4652 // Convert to float.
4653 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4654 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4655 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4656 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4657 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4658 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4660 // Use reciprocal estimate and one refinement step.
4661 // float4 recip = vrecpeq_f32(yf);
4662 // recip *= vrecpsq_f32(yf, recip);
4663 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4664 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4665 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4666 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4668 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4669 // Because short has a smaller range than ushort, we can actually get away
4670 // with only a single newton step. This requires that we use a weird bias
4671 // of 89, however (again, this has been exhaustively tested).
4672 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4673 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4674 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4675 N1 = DAG.getConstant(0x89, MVT::i32);
4676 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4677 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4678 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4679 // Convert back to integer and return.
4680 // return vmovn_s32(vcvt_s32_f32(result));
4681 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4682 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4686 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4687 EVT VT = Op.getValueType();
4688 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4689 "unexpected type for custom-lowering ISD::SDIV");
4691 DebugLoc dl = Op.getDebugLoc();
4692 SDValue N0 = Op.getOperand(0);
4693 SDValue N1 = Op.getOperand(1);
4696 if (VT == MVT::v8i8) {
4697 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4698 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4700 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4701 DAG.getIntPtrConstant(4));
4702 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4703 DAG.getIntPtrConstant(4));
4704 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4705 DAG.getIntPtrConstant(0));
4706 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4707 DAG.getIntPtrConstant(0));
4709 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4710 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4712 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4713 N0 = LowerCONCAT_VECTORS(N0, DAG);
4715 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4718 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4721 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4722 EVT VT = Op.getValueType();
4723 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4724 "unexpected type for custom-lowering ISD::UDIV");
4726 DebugLoc dl = Op.getDebugLoc();
4727 SDValue N0 = Op.getOperand(0);
4728 SDValue N1 = Op.getOperand(1);
4731 if (VT == MVT::v8i8) {
4732 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4733 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4735 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4736 DAG.getIntPtrConstant(4));
4737 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4738 DAG.getIntPtrConstant(4));
4739 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4740 DAG.getIntPtrConstant(0));
4741 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4742 DAG.getIntPtrConstant(0));
4744 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4745 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4747 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4748 N0 = LowerCONCAT_VECTORS(N0, DAG);
4750 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4751 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4756 // v4i16 sdiv ... Convert to float.
4757 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4758 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4759 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4760 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4761 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4762 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4764 // Use reciprocal estimate and two refinement steps.
4765 // float4 recip = vrecpeq_f32(yf);
4766 // recip *= vrecpsq_f32(yf, recip);
4767 // recip *= vrecpsq_f32(yf, recip);
4768 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4769 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4770 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4771 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4773 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4774 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4775 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4777 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4778 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4779 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4780 // and that it will never cause us to return an answer too large).
4781 // float4 result = as_float4(as_int4(xf*recip) + 2);
4782 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4783 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4784 N1 = DAG.getConstant(2, MVT::i32);
4785 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4786 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4787 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4788 // Convert back to integer and return.
4789 // return vmovn_u32(vcvt_s32_f32(result));
4790 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4791 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4795 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4796 switch (Op.getOpcode()) {
4797 default: llvm_unreachable("Don't know how to custom lower this!");
4798 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4799 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4800 case ISD::GlobalAddress:
4801 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4802 LowerGlobalAddressELF(Op, DAG);
4803 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4804 case ISD::SELECT: return LowerSELECT(Op, DAG);
4805 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4806 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4807 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4808 case ISD::VASTART: return LowerVASTART(Op, DAG);
4809 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4810 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4811 case ISD::SINT_TO_FP:
4812 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4813 case ISD::FP_TO_SINT:
4814 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4815 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4816 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4817 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4818 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4819 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4820 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4821 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4822 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4824 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4827 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4828 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4829 case ISD::SRL_PARTS:
4830 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4831 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4832 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4833 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4834 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4835 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4836 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4837 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4838 case ISD::MUL: return LowerMUL(Op, DAG);
4839 case ISD::SDIV: return LowerSDIV(Op, DAG);
4840 case ISD::UDIV: return LowerUDIV(Op, DAG);
4845 /// ReplaceNodeResults - Replace the results of node with an illegal result
4846 /// type with new values built out of custom code.
4847 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4848 SmallVectorImpl<SDValue>&Results,
4849 SelectionDAG &DAG) const {
4851 switch (N->getOpcode()) {
4853 llvm_unreachable("Don't know how to custom expand this!");
4856 Res = ExpandBITCAST(N, DAG);
4860 Res = Expand64BitShift(N, DAG, Subtarget);
4864 Results.push_back(Res);
4867 //===----------------------------------------------------------------------===//
4868 // ARM Scheduler Hooks
4869 //===----------------------------------------------------------------------===//
4872 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4873 MachineBasicBlock *BB,
4874 unsigned Size) const {
4875 unsigned dest = MI->getOperand(0).getReg();
4876 unsigned ptr = MI->getOperand(1).getReg();
4877 unsigned oldval = MI->getOperand(2).getReg();
4878 unsigned newval = MI->getOperand(3).getReg();
4879 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4880 DebugLoc dl = MI->getDebugLoc();
4881 bool isThumb2 = Subtarget->isThumb2();
4883 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
4885 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
4886 : ARM::GPRRegisterClass);
4889 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
4890 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
4891 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
4894 unsigned ldrOpc, strOpc;
4896 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4898 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4899 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4902 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4903 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4906 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4907 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4911 MachineFunction *MF = BB->getParent();
4912 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4913 MachineFunction::iterator It = BB;
4914 ++It; // insert the new blocks after the current block
4916 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4917 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4918 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4919 MF->insert(It, loop1MBB);
4920 MF->insert(It, loop2MBB);
4921 MF->insert(It, exitMBB);
4923 // Transfer the remainder of BB and its successor edges to exitMBB.
4924 exitMBB->splice(exitMBB->begin(), BB,
4925 llvm::next(MachineBasicBlock::iterator(MI)),
4927 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4931 // fallthrough --> loop1MBB
4932 BB->addSuccessor(loop1MBB);
4935 // ldrex dest, [ptr]
4939 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4940 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4941 .addReg(dest).addReg(oldval));
4942 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4943 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4944 BB->addSuccessor(loop2MBB);
4945 BB->addSuccessor(exitMBB);
4948 // strex scratch, newval, [ptr]
4952 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4954 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4955 .addReg(scratch).addImm(0));
4956 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4957 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4958 BB->addSuccessor(loop1MBB);
4959 BB->addSuccessor(exitMBB);
4965 MI->eraseFromParent(); // The instruction is gone now.
4971 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4972 unsigned Size, unsigned BinOpcode) const {
4973 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4974 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4976 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4977 MachineFunction *MF = BB->getParent();
4978 MachineFunction::iterator It = BB;
4981 unsigned dest = MI->getOperand(0).getReg();
4982 unsigned ptr = MI->getOperand(1).getReg();
4983 unsigned incr = MI->getOperand(2).getReg();
4984 DebugLoc dl = MI->getDebugLoc();
4986 bool isThumb2 = Subtarget->isThumb2();
4987 unsigned ldrOpc, strOpc;
4989 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4991 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4992 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4995 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4996 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4999 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5000 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5004 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5005 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5006 MF->insert(It, loopMBB);
5007 MF->insert(It, exitMBB);
5009 // Transfer the remainder of BB and its successor edges to exitMBB.
5010 exitMBB->splice(exitMBB->begin(), BB,
5011 llvm::next(MachineBasicBlock::iterator(MI)),
5013 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5015 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5016 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5017 unsigned scratch2 = (!BinOpcode) ? incr :
5018 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5022 // fallthrough --> loopMBB
5023 BB->addSuccessor(loopMBB);
5027 // <binop> scratch2, dest, incr
5028 // strex scratch, scratch2, ptr
5031 // fallthrough --> exitMBB
5033 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5035 // operand order needs to go the other way for NAND
5036 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5037 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5038 addReg(incr).addReg(dest)).addReg(0);
5040 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5041 addReg(dest).addReg(incr)).addReg(0);
5044 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5046 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5047 .addReg(scratch).addImm(0));
5048 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5049 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5051 BB->addSuccessor(loopMBB);
5052 BB->addSuccessor(exitMBB);
5058 MI->eraseFromParent(); // The instruction is gone now.
5064 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5065 MachineBasicBlock *BB,
5068 ARMCC::CondCodes Cond) const {
5069 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5071 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5072 MachineFunction *MF = BB->getParent();
5073 MachineFunction::iterator It = BB;
5076 unsigned dest = MI->getOperand(0).getReg();
5077 unsigned ptr = MI->getOperand(1).getReg();
5078 unsigned incr = MI->getOperand(2).getReg();
5079 unsigned oldval = dest;
5080 DebugLoc dl = MI->getDebugLoc();
5082 bool isThumb2 = Subtarget->isThumb2();
5083 unsigned ldrOpc, strOpc, extendOpc;
5085 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5087 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5088 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5089 extendOpc = isThumb2 ? ARM::t2SXTBr : ARM::SXTBr;
5092 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5093 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5094 extendOpc = isThumb2 ? ARM::t2SXTHr : ARM::SXTHr;
5097 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5098 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5103 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5104 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5105 MF->insert(It, loopMBB);
5106 MF->insert(It, exitMBB);
5108 // Transfer the remainder of BB and its successor edges to exitMBB.
5109 exitMBB->splice(exitMBB->begin(), BB,
5110 llvm::next(MachineBasicBlock::iterator(MI)),
5112 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5114 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5115 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5116 unsigned scratch2 = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5120 // fallthrough --> loopMBB
5121 BB->addSuccessor(loopMBB);
5125 // (sign extend dest, if required)
5127 // cmov.cond scratch2, dest, incr
5128 // strex scratch, scratch2, ptr
5131 // fallthrough --> exitMBB
5133 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5135 // Sign extend the value, if necessary.
5136 if (signExtend && extendOpc) {
5137 oldval = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5138 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval).addReg(dest));
5141 // Build compare and cmov instructions.
5142 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5143 .addReg(oldval).addReg(incr));
5144 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5145 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5147 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5149 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5150 .addReg(scratch).addImm(0));
5151 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5152 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5154 BB->addSuccessor(loopMBB);
5155 BB->addSuccessor(exitMBB);
5161 MI->eraseFromParent(); // The instruction is gone now.
5167 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
5168 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
5169 E = MBB->succ_end(); I != E; ++I)
5172 llvm_unreachable("Expecting a BB with two successors!");
5175 // FIXME: This opcode table should obviously be expressed in the target
5176 // description. We probably just need a "machine opcode" value in the pseudo
5177 // instruction. But the ideal solution maybe to simply remove the "S" version
5178 // of the opcode altogether.
5179 struct AddSubFlagsOpcodePair {
5181 unsigned MachineOpc;
5184 static AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
5185 {ARM::ADCSri, ARM::ADCri},
5186 {ARM::ADCSrr, ARM::ADCrr},
5187 {ARM::ADCSrs, ARM::ADCrs},
5188 {ARM::SBCSri, ARM::SBCri},
5189 {ARM::SBCSrr, ARM::SBCrr},
5190 {ARM::SBCSrs, ARM::SBCrs},
5191 {ARM::RSBSri, ARM::RSBri},
5192 {ARM::RSBSrr, ARM::RSBrr},
5193 {ARM::RSBSrs, ARM::RSBrs},
5194 {ARM::RSCSri, ARM::RSCri},
5195 {ARM::RSCSrs, ARM::RSCrs},
5196 {ARM::t2ADCSri, ARM::t2ADCri},
5197 {ARM::t2ADCSrr, ARM::t2ADCrr},
5198 {ARM::t2ADCSrs, ARM::t2ADCrs},
5199 {ARM::t2SBCSri, ARM::t2SBCri},
5200 {ARM::t2SBCSrr, ARM::t2SBCrr},
5201 {ARM::t2SBCSrs, ARM::t2SBCrs},
5202 {ARM::t2RSBSri, ARM::t2RSBri},
5203 {ARM::t2RSBSrs, ARM::t2RSBrs},
5206 // Convert and Add or Subtract with Carry and Flags to a generic opcode with
5207 // CPSR<def> operand. e.g. ADCS (...) -> ADC (... CPSR<def>).
5209 // FIXME: Somewhere we should assert that CPSR<def> is in the correct
5210 // position to be recognized by the target descrition as the 'S' bit.
5211 bool ARMTargetLowering::RemapAddSubWithFlags(MachineInstr *MI,
5212 MachineBasicBlock *BB) const {
5213 unsigned OldOpc = MI->getOpcode();
5214 unsigned NewOpc = 0;
5216 // This is only called for instructions that need remapping, so iterating over
5217 // the tiny opcode table is not costly.
5218 static const int NPairs =
5219 sizeof(AddSubFlagsOpcodeMap) / sizeof(AddSubFlagsOpcodePair);
5220 for (AddSubFlagsOpcodePair *Pair = &AddSubFlagsOpcodeMap[0],
5221 *End = &AddSubFlagsOpcodeMap[NPairs]; Pair != End; ++Pair) {
5222 if (OldOpc == Pair->PseudoOpc) {
5223 NewOpc = Pair->MachineOpc;
5230 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5231 DebugLoc dl = MI->getDebugLoc();
5232 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
5233 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
5234 MIB.addOperand(MI->getOperand(i));
5235 AddDefaultPred(MIB);
5236 MIB.addReg(ARM::CPSR, RegState::Define); // S bit
5237 MI->eraseFromParent();
5242 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5243 MachineBasicBlock *BB) const {
5244 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5245 DebugLoc dl = MI->getDebugLoc();
5246 bool isThumb2 = Subtarget->isThumb2();
5247 switch (MI->getOpcode()) {
5249 if (RemapAddSubWithFlags(MI, BB))
5253 llvm_unreachable("Unexpected instr type to insert");
5255 case ARM::ATOMIC_LOAD_ADD_I8:
5256 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5257 case ARM::ATOMIC_LOAD_ADD_I16:
5258 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5259 case ARM::ATOMIC_LOAD_ADD_I32:
5260 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5262 case ARM::ATOMIC_LOAD_AND_I8:
5263 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5264 case ARM::ATOMIC_LOAD_AND_I16:
5265 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5266 case ARM::ATOMIC_LOAD_AND_I32:
5267 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5269 case ARM::ATOMIC_LOAD_OR_I8:
5270 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5271 case ARM::ATOMIC_LOAD_OR_I16:
5272 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5273 case ARM::ATOMIC_LOAD_OR_I32:
5274 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5276 case ARM::ATOMIC_LOAD_XOR_I8:
5277 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5278 case ARM::ATOMIC_LOAD_XOR_I16:
5279 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5280 case ARM::ATOMIC_LOAD_XOR_I32:
5281 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5283 case ARM::ATOMIC_LOAD_NAND_I8:
5284 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5285 case ARM::ATOMIC_LOAD_NAND_I16:
5286 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5287 case ARM::ATOMIC_LOAD_NAND_I32:
5288 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5290 case ARM::ATOMIC_LOAD_SUB_I8:
5291 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5292 case ARM::ATOMIC_LOAD_SUB_I16:
5293 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5294 case ARM::ATOMIC_LOAD_SUB_I32:
5295 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5297 case ARM::ATOMIC_LOAD_MIN_I8:
5298 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
5299 case ARM::ATOMIC_LOAD_MIN_I16:
5300 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
5301 case ARM::ATOMIC_LOAD_MIN_I32:
5302 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
5304 case ARM::ATOMIC_LOAD_MAX_I8:
5305 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
5306 case ARM::ATOMIC_LOAD_MAX_I16:
5307 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
5308 case ARM::ATOMIC_LOAD_MAX_I32:
5309 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
5311 case ARM::ATOMIC_LOAD_UMIN_I8:
5312 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
5313 case ARM::ATOMIC_LOAD_UMIN_I16:
5314 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
5315 case ARM::ATOMIC_LOAD_UMIN_I32:
5316 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
5318 case ARM::ATOMIC_LOAD_UMAX_I8:
5319 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
5320 case ARM::ATOMIC_LOAD_UMAX_I16:
5321 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
5322 case ARM::ATOMIC_LOAD_UMAX_I32:
5323 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
5325 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
5326 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
5327 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
5329 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
5330 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
5331 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
5333 case ARM::tMOVCCr_pseudo: {
5334 // To "insert" a SELECT_CC instruction, we actually have to insert the
5335 // diamond control-flow pattern. The incoming instruction knows the
5336 // destination vreg to set, the condition code register to branch on, the
5337 // true/false values to select between, and a branch opcode to use.
5338 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5339 MachineFunction::iterator It = BB;
5345 // cmpTY ccX, r1, r2
5347 // fallthrough --> copy0MBB
5348 MachineBasicBlock *thisMBB = BB;
5349 MachineFunction *F = BB->getParent();
5350 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5351 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5352 F->insert(It, copy0MBB);
5353 F->insert(It, sinkMBB);
5355 // Transfer the remainder of BB and its successor edges to sinkMBB.
5356 sinkMBB->splice(sinkMBB->begin(), BB,
5357 llvm::next(MachineBasicBlock::iterator(MI)),
5359 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5361 BB->addSuccessor(copy0MBB);
5362 BB->addSuccessor(sinkMBB);
5364 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
5365 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
5368 // %FalseValue = ...
5369 // # fallthrough to sinkMBB
5372 // Update machine-CFG edges
5373 BB->addSuccessor(sinkMBB);
5376 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5379 BuildMI(*BB, BB->begin(), dl,
5380 TII->get(ARM::PHI), MI->getOperand(0).getReg())
5381 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5382 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5384 MI->eraseFromParent(); // The pseudo instruction is gone now.
5389 case ARM::BCCZi64: {
5390 // If there is an unconditional branch to the other successor, remove it.
5391 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
5393 // Compare both parts that make up the double comparison separately for
5395 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
5397 unsigned LHS1 = MI->getOperand(1).getReg();
5398 unsigned LHS2 = MI->getOperand(2).getReg();
5400 AddDefaultPred(BuildMI(BB, dl,
5401 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5402 .addReg(LHS1).addImm(0));
5403 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5404 .addReg(LHS2).addImm(0)
5405 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5407 unsigned RHS1 = MI->getOperand(3).getReg();
5408 unsigned RHS2 = MI->getOperand(4).getReg();
5409 AddDefaultPred(BuildMI(BB, dl,
5410 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5411 .addReg(LHS1).addReg(RHS1));
5412 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5413 .addReg(LHS2).addReg(RHS2)
5414 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5417 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
5418 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
5419 if (MI->getOperand(0).getImm() == ARMCC::NE)
5420 std::swap(destMBB, exitMBB);
5422 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5423 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
5424 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
5427 MI->eraseFromParent(); // The pseudo instruction is gone now.
5433 //===----------------------------------------------------------------------===//
5434 // ARM Optimization Hooks
5435 //===----------------------------------------------------------------------===//
5438 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
5439 TargetLowering::DAGCombinerInfo &DCI) {
5440 SelectionDAG &DAG = DCI.DAG;
5441 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5442 EVT VT = N->getValueType(0);
5443 unsigned Opc = N->getOpcode();
5444 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
5445 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
5446 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
5447 ISD::CondCode CC = ISD::SETCC_INVALID;
5450 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
5452 SDValue CCOp = Slct.getOperand(0);
5453 if (CCOp.getOpcode() == ISD::SETCC)
5454 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
5457 bool DoXform = false;
5459 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
5462 if (LHS.getOpcode() == ISD::Constant &&
5463 cast<ConstantSDNode>(LHS)->isNullValue()) {
5465 } else if (CC != ISD::SETCC_INVALID &&
5466 RHS.getOpcode() == ISD::Constant &&
5467 cast<ConstantSDNode>(RHS)->isNullValue()) {
5468 std::swap(LHS, RHS);
5469 SDValue Op0 = Slct.getOperand(0);
5470 EVT OpVT = isSlctCC ? Op0.getValueType() :
5471 Op0.getOperand(0).getValueType();
5472 bool isInt = OpVT.isInteger();
5473 CC = ISD::getSetCCInverse(CC, isInt);
5475 if (!TLI.isCondCodeLegal(CC, OpVT))
5476 return SDValue(); // Inverse operator isn't legal.
5483 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
5485 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
5486 Slct.getOperand(0), Slct.getOperand(1), CC);
5487 SDValue CCOp = Slct.getOperand(0);
5489 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
5490 CCOp.getOperand(0), CCOp.getOperand(1), CC);
5491 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
5492 CCOp, OtherOp, Result);
5497 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
5498 /// operands N0 and N1. This is a helper for PerformADDCombine that is
5499 /// called with the default operands, and if that fails, with commuted
5501 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
5502 TargetLowering::DAGCombinerInfo &DCI) {
5503 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
5504 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
5505 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
5506 if (Result.getNode()) return Result;
5511 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
5513 static SDValue PerformADDCombine(SDNode *N,
5514 TargetLowering::DAGCombinerInfo &DCI) {
5515 SDValue N0 = N->getOperand(0);
5516 SDValue N1 = N->getOperand(1);
5518 // First try with the default operand order.
5519 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
5520 if (Result.getNode())
5523 // If that didn't work, try again with the operands commuted.
5524 return PerformADDCombineWithOperands(N, N1, N0, DCI);
5527 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
5529 static SDValue PerformSUBCombine(SDNode *N,
5530 TargetLowering::DAGCombinerInfo &DCI) {
5531 SDValue N0 = N->getOperand(0);
5532 SDValue N1 = N->getOperand(1);
5534 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
5535 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
5536 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
5537 if (Result.getNode()) return Result;
5543 /// PerformVMULCombine
5544 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
5545 /// special multiplier accumulator forwarding.
5551 static SDValue PerformVMULCombine(SDNode *N,
5552 TargetLowering::DAGCombinerInfo &DCI,
5553 const ARMSubtarget *Subtarget) {
5554 if (!Subtarget->hasVMLxForwarding())
5557 SelectionDAG &DAG = DCI.DAG;
5558 SDValue N0 = N->getOperand(0);
5559 SDValue N1 = N->getOperand(1);
5560 unsigned Opcode = N0.getOpcode();
5561 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5562 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
5563 Opcode = N0.getOpcode();
5564 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5565 Opcode != ISD::FADD && Opcode != ISD::FSUB)
5570 EVT VT = N->getValueType(0);
5571 DebugLoc DL = N->getDebugLoc();
5572 SDValue N00 = N0->getOperand(0);
5573 SDValue N01 = N0->getOperand(1);
5574 return DAG.getNode(Opcode, DL, VT,
5575 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
5576 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
5579 static SDValue PerformMULCombine(SDNode *N,
5580 TargetLowering::DAGCombinerInfo &DCI,
5581 const ARMSubtarget *Subtarget) {
5582 SelectionDAG &DAG = DCI.DAG;
5584 if (Subtarget->isThumb1Only())
5587 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5590 EVT VT = N->getValueType(0);
5591 if (VT.is64BitVector() || VT.is128BitVector())
5592 return PerformVMULCombine(N, DCI, Subtarget);
5596 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
5600 uint64_t MulAmt = C->getZExtValue();
5601 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
5602 ShiftAmt = ShiftAmt & (32 - 1);
5603 SDValue V = N->getOperand(0);
5604 DebugLoc DL = N->getDebugLoc();
5607 MulAmt >>= ShiftAmt;
5608 if (isPowerOf2_32(MulAmt - 1)) {
5609 // (mul x, 2^N + 1) => (add (shl x, N), x)
5610 Res = DAG.getNode(ISD::ADD, DL, VT,
5611 V, DAG.getNode(ISD::SHL, DL, VT,
5612 V, DAG.getConstant(Log2_32(MulAmt-1),
5614 } else if (isPowerOf2_32(MulAmt + 1)) {
5615 // (mul x, 2^N - 1) => (sub (shl x, N), x)
5616 Res = DAG.getNode(ISD::SUB, DL, VT,
5617 DAG.getNode(ISD::SHL, DL, VT,
5618 V, DAG.getConstant(Log2_32(MulAmt+1),
5625 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
5626 DAG.getConstant(ShiftAmt, MVT::i32));
5628 // Do not add new nodes to DAG combiner worklist.
5629 DCI.CombineTo(N, Res, false);
5633 static SDValue PerformANDCombine(SDNode *N,
5634 TargetLowering::DAGCombinerInfo &DCI) {
5636 // Attempt to use immediate-form VBIC
5637 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5638 DebugLoc dl = N->getDebugLoc();
5639 EVT VT = N->getValueType(0);
5640 SelectionDAG &DAG = DCI.DAG;
5642 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5645 APInt SplatBits, SplatUndef;
5646 unsigned SplatBitSize;
5649 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5650 if (SplatBitSize <= 64) {
5652 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
5653 SplatUndef.getZExtValue(), SplatBitSize,
5654 DAG, VbicVT, VT.is128BitVector(),
5656 if (Val.getNode()) {
5658 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
5659 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
5660 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
5668 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
5669 static SDValue PerformORCombine(SDNode *N,
5670 TargetLowering::DAGCombinerInfo &DCI,
5671 const ARMSubtarget *Subtarget) {
5672 // Attempt to use immediate-form VORR
5673 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5674 DebugLoc dl = N->getDebugLoc();
5675 EVT VT = N->getValueType(0);
5676 SelectionDAG &DAG = DCI.DAG;
5678 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5681 APInt SplatBits, SplatUndef;
5682 unsigned SplatBitSize;
5684 if (BVN && Subtarget->hasNEON() &&
5685 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5686 if (SplatBitSize <= 64) {
5688 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5689 SplatUndef.getZExtValue(), SplatBitSize,
5690 DAG, VorrVT, VT.is128BitVector(),
5692 if (Val.getNode()) {
5694 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
5695 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
5696 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
5701 SDValue N0 = N->getOperand(0);
5702 if (N0.getOpcode() != ISD::AND)
5704 SDValue N1 = N->getOperand(1);
5706 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
5707 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
5708 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
5710 unsigned SplatBitSize;
5713 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
5715 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
5716 HasAnyUndefs) && !HasAnyUndefs) {
5717 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
5719 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
5720 HasAnyUndefs) && !HasAnyUndefs &&
5721 SplatBits0 == ~SplatBits1) {
5722 // Canonicalize the vector type to make instruction selection simpler.
5723 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
5724 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
5725 N0->getOperand(1), N0->getOperand(0),
5727 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
5732 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
5735 // BFI is only available on V6T2+
5736 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
5739 DebugLoc DL = N->getDebugLoc();
5740 // 1) or (and A, mask), val => ARMbfi A, val, mask
5741 // iff (val & mask) == val
5743 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5744 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
5745 // && mask == ~mask2
5746 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
5747 // && ~mask == mask2
5748 // (i.e., copy a bitfield value into another bitfield of the same width)
5753 SDValue N00 = N0.getOperand(0);
5755 // The value and the mask need to be constants so we can verify this is
5756 // actually a bitfield set. If the mask is 0xffff, we can do better
5757 // via a movt instruction, so don't use BFI in that case.
5758 SDValue MaskOp = N0.getOperand(1);
5759 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
5762 unsigned Mask = MaskC->getZExtValue();
5766 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
5767 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
5769 unsigned Val = N1C->getZExtValue();
5770 if ((Val & ~Mask) != Val)
5773 if (ARM::isBitFieldInvertedMask(Mask)) {
5774 Val >>= CountTrailingZeros_32(~Mask);
5776 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
5777 DAG.getConstant(Val, MVT::i32),
5778 DAG.getConstant(Mask, MVT::i32));
5780 // Do not add new nodes to DAG combiner worklist.
5781 DCI.CombineTo(N, Res, false);
5784 } else if (N1.getOpcode() == ISD::AND) {
5785 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5786 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5789 unsigned Mask2 = N11C->getZExtValue();
5791 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
5793 if (ARM::isBitFieldInvertedMask(Mask) &&
5795 // The pack halfword instruction works better for masks that fit it,
5796 // so use that when it's available.
5797 if (Subtarget->hasT2ExtractPack() &&
5798 (Mask == 0xffff || Mask == 0xffff0000))
5801 unsigned amt = CountTrailingZeros_32(Mask2);
5802 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
5803 DAG.getConstant(amt, MVT::i32));
5804 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
5805 DAG.getConstant(Mask, MVT::i32));
5806 // Do not add new nodes to DAG combiner worklist.
5807 DCI.CombineTo(N, Res, false);
5809 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
5811 // The pack halfword instruction works better for masks that fit it,
5812 // so use that when it's available.
5813 if (Subtarget->hasT2ExtractPack() &&
5814 (Mask2 == 0xffff || Mask2 == 0xffff0000))
5817 unsigned lsb = CountTrailingZeros_32(Mask);
5818 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
5819 DAG.getConstant(lsb, MVT::i32));
5820 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
5821 DAG.getConstant(Mask2, MVT::i32));
5822 // Do not add new nodes to DAG combiner worklist.
5823 DCI.CombineTo(N, Res, false);
5828 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
5829 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
5830 ARM::isBitFieldInvertedMask(~Mask)) {
5831 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
5832 // where lsb(mask) == #shamt and masked bits of B are known zero.
5833 SDValue ShAmt = N00.getOperand(1);
5834 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5835 unsigned LSB = CountTrailingZeros_32(Mask);
5839 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
5840 DAG.getConstant(~Mask, MVT::i32));
5842 // Do not add new nodes to DAG combiner worklist.
5843 DCI.CombineTo(N, Res, false);
5849 /// PerformBFICombine - (bfi A, (and B, C1), C2) -> (bfi A, B, C2) iff
5851 static SDValue PerformBFICombine(SDNode *N,
5852 TargetLowering::DAGCombinerInfo &DCI) {
5853 SDValue N1 = N->getOperand(1);
5854 if (N1.getOpcode() == ISD::AND) {
5855 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5858 unsigned Mask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
5859 unsigned Mask2 = N11C->getZExtValue();
5860 if ((Mask & Mask2) == Mask2)
5861 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
5862 N->getOperand(0), N1.getOperand(0),
5868 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
5869 /// ARMISD::VMOVRRD.
5870 static SDValue PerformVMOVRRDCombine(SDNode *N,
5871 TargetLowering::DAGCombinerInfo &DCI) {
5872 // vmovrrd(vmovdrr x, y) -> x,y
5873 SDValue InDouble = N->getOperand(0);
5874 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
5875 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
5877 // vmovrrd(load f64) -> (load i32), (load i32)
5878 SDNode *InNode = InDouble.getNode();
5879 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
5880 InNode->getValueType(0) == MVT::f64 &&
5881 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
5882 !cast<LoadSDNode>(InNode)->isVolatile()) {
5883 // TODO: Should this be done for non-FrameIndex operands?
5884 LoadSDNode *LD = cast<LoadSDNode>(InNode);
5886 SelectionDAG &DAG = DCI.DAG;
5887 DebugLoc DL = LD->getDebugLoc();
5888 SDValue BasePtr = LD->getBasePtr();
5889 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
5890 LD->getPointerInfo(), LD->isVolatile(),
5891 LD->isNonTemporal(), LD->getAlignment());
5893 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
5894 DAG.getConstant(4, MVT::i32));
5895 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
5896 LD->getPointerInfo(), LD->isVolatile(),
5897 LD->isNonTemporal(),
5898 std::min(4U, LD->getAlignment() / 2));
5900 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
5901 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
5902 DCI.RemoveFromWorklist(LD);
5910 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
5911 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
5912 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
5913 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
5914 SDValue Op0 = N->getOperand(0);
5915 SDValue Op1 = N->getOperand(1);
5916 if (Op0.getOpcode() == ISD::BITCAST)
5917 Op0 = Op0.getOperand(0);
5918 if (Op1.getOpcode() == ISD::BITCAST)
5919 Op1 = Op1.getOperand(0);
5920 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
5921 Op0.getNode() == Op1.getNode() &&
5922 Op0.getResNo() == 0 && Op1.getResNo() == 1)
5923 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5924 N->getValueType(0), Op0.getOperand(0));
5928 /// PerformSTORECombine - Target-specific dag combine xforms for
5930 static SDValue PerformSTORECombine(SDNode *N,
5931 TargetLowering::DAGCombinerInfo &DCI) {
5932 // Bitcast an i64 store extracted from a vector to f64.
5933 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5934 StoreSDNode *St = cast<StoreSDNode>(N);
5935 SDValue StVal = St->getValue();
5936 if (!ISD::isNormalStore(St) || St->isVolatile())
5939 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
5940 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
5941 SelectionDAG &DAG = DCI.DAG;
5942 DebugLoc DL = St->getDebugLoc();
5943 SDValue BasePtr = St->getBasePtr();
5944 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
5945 StVal.getNode()->getOperand(0), BasePtr,
5946 St->getPointerInfo(), St->isVolatile(),
5947 St->isNonTemporal(), St->getAlignment());
5949 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
5950 DAG.getConstant(4, MVT::i32));
5951 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
5952 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
5953 St->isNonTemporal(),
5954 std::min(4U, St->getAlignment() / 2));
5957 if (StVal.getValueType() != MVT::i64 ||
5958 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5961 SelectionDAG &DAG = DCI.DAG;
5962 DebugLoc dl = StVal.getDebugLoc();
5963 SDValue IntVec = StVal.getOperand(0);
5964 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5965 IntVec.getValueType().getVectorNumElements());
5966 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
5967 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5968 Vec, StVal.getOperand(1));
5969 dl = N->getDebugLoc();
5970 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
5971 // Make the DAGCombiner fold the bitcasts.
5972 DCI.AddToWorklist(Vec.getNode());
5973 DCI.AddToWorklist(ExtElt.getNode());
5974 DCI.AddToWorklist(V.getNode());
5975 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
5976 St->getPointerInfo(), St->isVolatile(),
5977 St->isNonTemporal(), St->getAlignment(),
5981 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
5982 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
5983 /// i64 vector to have f64 elements, since the value can then be loaded
5984 /// directly into a VFP register.
5985 static bool hasNormalLoadOperand(SDNode *N) {
5986 unsigned NumElts = N->getValueType(0).getVectorNumElements();
5987 for (unsigned i = 0; i < NumElts; ++i) {
5988 SDNode *Elt = N->getOperand(i).getNode();
5989 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
5995 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
5996 /// ISD::BUILD_VECTOR.
5997 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
5998 TargetLowering::DAGCombinerInfo &DCI){
5999 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
6000 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
6001 // into a pair of GPRs, which is fine when the value is used as a scalar,
6002 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
6003 SelectionDAG &DAG = DCI.DAG;
6004 if (N->getNumOperands() == 2) {
6005 SDValue RV = PerformVMOVDRRCombine(N, DAG);
6010 // Load i64 elements as f64 values so that type legalization does not split
6011 // them up into i32 values.
6012 EVT VT = N->getValueType(0);
6013 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
6015 DebugLoc dl = N->getDebugLoc();
6016 SmallVector<SDValue, 8> Ops;
6017 unsigned NumElts = VT.getVectorNumElements();
6018 for (unsigned i = 0; i < NumElts; ++i) {
6019 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
6021 // Make the DAGCombiner fold the bitcast.
6022 DCI.AddToWorklist(V.getNode());
6024 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
6025 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
6026 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
6029 /// PerformInsertEltCombine - Target-specific dag combine xforms for
6030 /// ISD::INSERT_VECTOR_ELT.
6031 static SDValue PerformInsertEltCombine(SDNode *N,
6032 TargetLowering::DAGCombinerInfo &DCI) {
6033 // Bitcast an i64 load inserted into a vector to f64.
6034 // Otherwise, the i64 value will be legalized to a pair of i32 values.
6035 EVT VT = N->getValueType(0);
6036 SDNode *Elt = N->getOperand(1).getNode();
6037 if (VT.getVectorElementType() != MVT::i64 ||
6038 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
6041 SelectionDAG &DAG = DCI.DAG;
6042 DebugLoc dl = N->getDebugLoc();
6043 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
6044 VT.getVectorNumElements());
6045 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
6046 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
6047 // Make the DAGCombiner fold the bitcasts.
6048 DCI.AddToWorklist(Vec.getNode());
6049 DCI.AddToWorklist(V.getNode());
6050 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
6051 Vec, V, N->getOperand(2));
6052 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
6055 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
6056 /// ISD::VECTOR_SHUFFLE.
6057 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
6058 // The LLVM shufflevector instruction does not require the shuffle mask
6059 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
6060 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
6061 // operands do not match the mask length, they are extended by concatenating
6062 // them with undef vectors. That is probably the right thing for other
6063 // targets, but for NEON it is better to concatenate two double-register
6064 // size vector operands into a single quad-register size vector. Do that
6065 // transformation here:
6066 // shuffle(concat(v1, undef), concat(v2, undef)) ->
6067 // shuffle(concat(v1, v2), undef)
6068 SDValue Op0 = N->getOperand(0);
6069 SDValue Op1 = N->getOperand(1);
6070 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
6071 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
6072 Op0.getNumOperands() != 2 ||
6073 Op1.getNumOperands() != 2)
6075 SDValue Concat0Op1 = Op0.getOperand(1);
6076 SDValue Concat1Op1 = Op1.getOperand(1);
6077 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
6078 Concat1Op1.getOpcode() != ISD::UNDEF)
6080 // Skip the transformation if any of the types are illegal.
6081 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6082 EVT VT = N->getValueType(0);
6083 if (!TLI.isTypeLegal(VT) ||
6084 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
6085 !TLI.isTypeLegal(Concat1Op1.getValueType()))
6088 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
6089 Op0.getOperand(0), Op1.getOperand(0));
6090 // Translate the shuffle mask.
6091 SmallVector<int, 16> NewMask;
6092 unsigned NumElts = VT.getVectorNumElements();
6093 unsigned HalfElts = NumElts/2;
6094 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
6095 for (unsigned n = 0; n < NumElts; ++n) {
6096 int MaskElt = SVN->getMaskElt(n);
6098 if (MaskElt < (int)HalfElts)
6100 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
6101 NewElt = HalfElts + MaskElt - NumElts;
6102 NewMask.push_back(NewElt);
6104 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
6105 DAG.getUNDEF(VT), NewMask.data());
6108 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
6109 /// NEON load/store intrinsics to merge base address updates.
6110 static SDValue CombineBaseUpdate(SDNode *N,
6111 TargetLowering::DAGCombinerInfo &DCI) {
6112 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6115 SelectionDAG &DAG = DCI.DAG;
6116 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
6117 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
6118 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
6119 SDValue Addr = N->getOperand(AddrOpIdx);
6121 // Search for a use of the address operand that is an increment.
6122 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
6123 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
6125 if (User->getOpcode() != ISD::ADD ||
6126 UI.getUse().getResNo() != Addr.getResNo())
6129 // Check that the add is independent of the load/store. Otherwise, folding
6130 // it would create a cycle.
6131 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
6134 // Find the new opcode for the updating load/store.
6136 bool isLaneOp = false;
6137 unsigned NewOpc = 0;
6138 unsigned NumVecs = 0;
6140 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
6142 default: assert(0 && "unexpected intrinsic for Neon base update");
6143 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
6145 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
6147 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
6149 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
6151 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
6152 NumVecs = 2; isLaneOp = true; break;
6153 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
6154 NumVecs = 3; isLaneOp = true; break;
6155 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
6156 NumVecs = 4; isLaneOp = true; break;
6157 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
6158 NumVecs = 1; isLoad = false; break;
6159 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
6160 NumVecs = 2; isLoad = false; break;
6161 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
6162 NumVecs = 3; isLoad = false; break;
6163 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
6164 NumVecs = 4; isLoad = false; break;
6165 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
6166 NumVecs = 2; isLoad = false; isLaneOp = true; break;
6167 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
6168 NumVecs = 3; isLoad = false; isLaneOp = true; break;
6169 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
6170 NumVecs = 4; isLoad = false; isLaneOp = true; break;
6174 switch (N->getOpcode()) {
6175 default: assert(0 && "unexpected opcode for Neon base update");
6176 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
6177 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
6178 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
6182 // Find the size of memory referenced by the load/store.
6185 VecTy = N->getValueType(0);
6187 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
6188 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
6190 NumBytes /= VecTy.getVectorNumElements();
6192 // If the increment is a constant, it must match the memory ref size.
6193 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
6194 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
6195 uint64_t IncVal = CInc->getZExtValue();
6196 if (IncVal != NumBytes)
6198 } else if (NumBytes >= 3 * 16) {
6199 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
6200 // separate instructions that make it harder to use a non-constant update.
6204 // Create the new updating load/store node.
6206 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
6208 for (n = 0; n < NumResultVecs; ++n)
6210 Tys[n++] = MVT::i32;
6211 Tys[n] = MVT::Other;
6212 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
6213 SmallVector<SDValue, 8> Ops;
6214 Ops.push_back(N->getOperand(0)); // incoming chain
6215 Ops.push_back(N->getOperand(AddrOpIdx));
6217 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
6218 Ops.push_back(N->getOperand(i));
6220 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
6221 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
6222 Ops.data(), Ops.size(),
6223 MemInt->getMemoryVT(),
6224 MemInt->getMemOperand());
6227 std::vector<SDValue> NewResults;
6228 for (unsigned i = 0; i < NumResultVecs; ++i) {
6229 NewResults.push_back(SDValue(UpdN.getNode(), i));
6231 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
6232 DCI.CombineTo(N, NewResults);
6233 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
6240 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
6241 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
6242 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
6244 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
6245 SelectionDAG &DAG = DCI.DAG;
6246 EVT VT = N->getValueType(0);
6247 // vldN-dup instructions only support 64-bit vectors for N > 1.
6248 if (!VT.is64BitVector())
6251 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
6252 SDNode *VLD = N->getOperand(0).getNode();
6253 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
6255 unsigned NumVecs = 0;
6256 unsigned NewOpc = 0;
6257 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
6258 if (IntNo == Intrinsic::arm_neon_vld2lane) {
6260 NewOpc = ARMISD::VLD2DUP;
6261 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
6263 NewOpc = ARMISD::VLD3DUP;
6264 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
6266 NewOpc = ARMISD::VLD4DUP;
6271 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
6272 // numbers match the load.
6273 unsigned VLDLaneNo =
6274 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
6275 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6277 // Ignore uses of the chain result.
6278 if (UI.getUse().getResNo() == NumVecs)
6281 if (User->getOpcode() != ARMISD::VDUPLANE ||
6282 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
6286 // Create the vldN-dup node.
6289 for (n = 0; n < NumVecs; ++n)
6291 Tys[n] = MVT::Other;
6292 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
6293 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
6294 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
6295 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
6296 Ops, 2, VLDMemInt->getMemoryVT(),
6297 VLDMemInt->getMemOperand());
6300 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6302 unsigned ResNo = UI.getUse().getResNo();
6303 // Ignore uses of the chain result.
6304 if (ResNo == NumVecs)
6307 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
6310 // Now the vldN-lane intrinsic is dead except for its chain result.
6311 // Update uses of the chain.
6312 std::vector<SDValue> VLDDupResults;
6313 for (unsigned n = 0; n < NumVecs; ++n)
6314 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
6315 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
6316 DCI.CombineTo(VLD, VLDDupResults);
6321 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
6322 /// ARMISD::VDUPLANE.
6323 static SDValue PerformVDUPLANECombine(SDNode *N,
6324 TargetLowering::DAGCombinerInfo &DCI) {
6325 SDValue Op = N->getOperand(0);
6327 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
6328 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
6329 if (CombineVLDDUP(N, DCI))
6330 return SDValue(N, 0);
6332 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
6333 // redundant. Ignore bit_converts for now; element sizes are checked below.
6334 while (Op.getOpcode() == ISD::BITCAST)
6335 Op = Op.getOperand(0);
6336 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
6339 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
6340 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
6341 // The canonical VMOV for a zero vector uses a 32-bit element size.
6342 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6344 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
6346 EVT VT = N->getValueType(0);
6347 if (EltSize > VT.getVectorElementType().getSizeInBits())
6350 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
6353 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6354 /// operand of a vector shift operation, where all the elements of the
6355 /// build_vector must have the same constant integer value.
6356 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6357 // Ignore bit_converts.
6358 while (Op.getOpcode() == ISD::BITCAST)
6359 Op = Op.getOperand(0);
6360 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6361 APInt SplatBits, SplatUndef;
6362 unsigned SplatBitSize;
6364 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6365 HasAnyUndefs, ElementBits) ||
6366 SplatBitSize > ElementBits)
6368 Cnt = SplatBits.getSExtValue();
6372 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6373 /// operand of a vector shift left operation. That value must be in the range:
6374 /// 0 <= Value < ElementBits for a left shift; or
6375 /// 0 <= Value <= ElementBits for a long left shift.
6376 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6377 assert(VT.isVector() && "vector shift count is not a vector type");
6378 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6379 if (! getVShiftImm(Op, ElementBits, Cnt))
6381 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
6384 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6385 /// operand of a vector shift right operation. For a shift opcode, the value
6386 /// is positive, but for an intrinsic the value count must be negative. The
6387 /// absolute value must be in the range:
6388 /// 1 <= |Value| <= ElementBits for a right shift; or
6389 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6390 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6392 assert(VT.isVector() && "vector shift count is not a vector type");
6393 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6394 if (! getVShiftImm(Op, ElementBits, Cnt))
6398 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
6401 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
6402 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
6403 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
6406 // Don't do anything for most intrinsics.
6409 // Vector shifts: check for immediate versions and lower them.
6410 // Note: This is done during DAG combining instead of DAG legalizing because
6411 // the build_vectors for 64-bit vector element shift counts are generally
6412 // not legal, and it is hard to see their values after they get legalized to
6413 // loads from a constant pool.
6414 case Intrinsic::arm_neon_vshifts:
6415 case Intrinsic::arm_neon_vshiftu:
6416 case Intrinsic::arm_neon_vshiftls:
6417 case Intrinsic::arm_neon_vshiftlu:
6418 case Intrinsic::arm_neon_vshiftn:
6419 case Intrinsic::arm_neon_vrshifts:
6420 case Intrinsic::arm_neon_vrshiftu:
6421 case Intrinsic::arm_neon_vrshiftn:
6422 case Intrinsic::arm_neon_vqshifts:
6423 case Intrinsic::arm_neon_vqshiftu:
6424 case Intrinsic::arm_neon_vqshiftsu:
6425 case Intrinsic::arm_neon_vqshiftns:
6426 case Intrinsic::arm_neon_vqshiftnu:
6427 case Intrinsic::arm_neon_vqshiftnsu:
6428 case Intrinsic::arm_neon_vqrshiftns:
6429 case Intrinsic::arm_neon_vqrshiftnu:
6430 case Intrinsic::arm_neon_vqrshiftnsu: {
6431 EVT VT = N->getOperand(1).getValueType();
6433 unsigned VShiftOpc = 0;
6436 case Intrinsic::arm_neon_vshifts:
6437 case Intrinsic::arm_neon_vshiftu:
6438 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
6439 VShiftOpc = ARMISD::VSHL;
6442 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
6443 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
6444 ARMISD::VSHRs : ARMISD::VSHRu);
6449 case Intrinsic::arm_neon_vshiftls:
6450 case Intrinsic::arm_neon_vshiftlu:
6451 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
6453 llvm_unreachable("invalid shift count for vshll intrinsic");
6455 case Intrinsic::arm_neon_vrshifts:
6456 case Intrinsic::arm_neon_vrshiftu:
6457 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
6461 case Intrinsic::arm_neon_vqshifts:
6462 case Intrinsic::arm_neon_vqshiftu:
6463 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6467 case Intrinsic::arm_neon_vqshiftsu:
6468 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6470 llvm_unreachable("invalid shift count for vqshlu intrinsic");
6472 case Intrinsic::arm_neon_vshiftn:
6473 case Intrinsic::arm_neon_vrshiftn:
6474 case Intrinsic::arm_neon_vqshiftns:
6475 case Intrinsic::arm_neon_vqshiftnu:
6476 case Intrinsic::arm_neon_vqshiftnsu:
6477 case Intrinsic::arm_neon_vqrshiftns:
6478 case Intrinsic::arm_neon_vqrshiftnu:
6479 case Intrinsic::arm_neon_vqrshiftnsu:
6480 // Narrowing shifts require an immediate right shift.
6481 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
6483 llvm_unreachable("invalid shift count for narrowing vector shift "
6487 llvm_unreachable("unhandled vector shift");
6491 case Intrinsic::arm_neon_vshifts:
6492 case Intrinsic::arm_neon_vshiftu:
6493 // Opcode already set above.
6495 case Intrinsic::arm_neon_vshiftls:
6496 case Intrinsic::arm_neon_vshiftlu:
6497 if (Cnt == VT.getVectorElementType().getSizeInBits())
6498 VShiftOpc = ARMISD::VSHLLi;
6500 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
6501 ARMISD::VSHLLs : ARMISD::VSHLLu);
6503 case Intrinsic::arm_neon_vshiftn:
6504 VShiftOpc = ARMISD::VSHRN; break;
6505 case Intrinsic::arm_neon_vrshifts:
6506 VShiftOpc = ARMISD::VRSHRs; break;
6507 case Intrinsic::arm_neon_vrshiftu:
6508 VShiftOpc = ARMISD::VRSHRu; break;
6509 case Intrinsic::arm_neon_vrshiftn:
6510 VShiftOpc = ARMISD::VRSHRN; break;
6511 case Intrinsic::arm_neon_vqshifts:
6512 VShiftOpc = ARMISD::VQSHLs; break;
6513 case Intrinsic::arm_neon_vqshiftu:
6514 VShiftOpc = ARMISD::VQSHLu; break;
6515 case Intrinsic::arm_neon_vqshiftsu:
6516 VShiftOpc = ARMISD::VQSHLsu; break;
6517 case Intrinsic::arm_neon_vqshiftns:
6518 VShiftOpc = ARMISD::VQSHRNs; break;
6519 case Intrinsic::arm_neon_vqshiftnu:
6520 VShiftOpc = ARMISD::VQSHRNu; break;
6521 case Intrinsic::arm_neon_vqshiftnsu:
6522 VShiftOpc = ARMISD::VQSHRNsu; break;
6523 case Intrinsic::arm_neon_vqrshiftns:
6524 VShiftOpc = ARMISD::VQRSHRNs; break;
6525 case Intrinsic::arm_neon_vqrshiftnu:
6526 VShiftOpc = ARMISD::VQRSHRNu; break;
6527 case Intrinsic::arm_neon_vqrshiftnsu:
6528 VShiftOpc = ARMISD::VQRSHRNsu; break;
6531 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6532 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
6535 case Intrinsic::arm_neon_vshiftins: {
6536 EVT VT = N->getOperand(1).getValueType();
6538 unsigned VShiftOpc = 0;
6540 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
6541 VShiftOpc = ARMISD::VSLI;
6542 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
6543 VShiftOpc = ARMISD::VSRI;
6545 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
6548 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6549 N->getOperand(1), N->getOperand(2),
6550 DAG.getConstant(Cnt, MVT::i32));
6553 case Intrinsic::arm_neon_vqrshifts:
6554 case Intrinsic::arm_neon_vqrshiftu:
6555 // No immediate versions of these to check for.
6562 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
6563 /// lowers them. As with the vector shift intrinsics, this is done during DAG
6564 /// combining instead of DAG legalizing because the build_vectors for 64-bit
6565 /// vector element shift counts are generally not legal, and it is hard to see
6566 /// their values after they get legalized to loads from a constant pool.
6567 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
6568 const ARMSubtarget *ST) {
6569 EVT VT = N->getValueType(0);
6571 // Nothing to be done for scalar shifts.
6572 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6573 if (!VT.isVector() || !TLI.isTypeLegal(VT))
6576 assert(ST->hasNEON() && "unexpected vector shift");
6579 switch (N->getOpcode()) {
6580 default: llvm_unreachable("unexpected shift opcode");
6583 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
6584 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
6585 DAG.getConstant(Cnt, MVT::i32));
6590 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
6591 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
6592 ARMISD::VSHRs : ARMISD::VSHRu);
6593 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
6594 DAG.getConstant(Cnt, MVT::i32));
6600 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
6601 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
6602 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
6603 const ARMSubtarget *ST) {
6604 SDValue N0 = N->getOperand(0);
6606 // Check for sign- and zero-extensions of vector extract operations of 8-
6607 // and 16-bit vector elements. NEON supports these directly. They are
6608 // handled during DAG combining because type legalization will promote them
6609 // to 32-bit types and it is messy to recognize the operations after that.
6610 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6611 SDValue Vec = N0.getOperand(0);
6612 SDValue Lane = N0.getOperand(1);
6613 EVT VT = N->getValueType(0);
6614 EVT EltVT = N0.getValueType();
6615 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6617 if (VT == MVT::i32 &&
6618 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
6619 TLI.isTypeLegal(Vec.getValueType()) &&
6620 isa<ConstantSDNode>(Lane)) {
6623 switch (N->getOpcode()) {
6624 default: llvm_unreachable("unexpected opcode");
6625 case ISD::SIGN_EXTEND:
6626 Opc = ARMISD::VGETLANEs;
6628 case ISD::ZERO_EXTEND:
6629 case ISD::ANY_EXTEND:
6630 Opc = ARMISD::VGETLANEu;
6633 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
6640 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
6641 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
6642 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
6643 const ARMSubtarget *ST) {
6644 // If the target supports NEON, try to use vmax/vmin instructions for f32
6645 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
6646 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
6647 // a NaN; only do the transformation when it matches that behavior.
6649 // For now only do this when using NEON for FP operations; if using VFP, it
6650 // is not obvious that the benefit outweighs the cost of switching to the
6652 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
6653 N->getValueType(0) != MVT::f32)
6656 SDValue CondLHS = N->getOperand(0);
6657 SDValue CondRHS = N->getOperand(1);
6658 SDValue LHS = N->getOperand(2);
6659 SDValue RHS = N->getOperand(3);
6660 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
6662 unsigned Opcode = 0;
6664 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
6665 IsReversed = false; // x CC y ? x : y
6666 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
6667 IsReversed = true ; // x CC y ? y : x
6681 // If LHS is NaN, an ordered comparison will be false and the result will
6682 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
6683 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6684 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
6685 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6687 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
6688 // will return -0, so vmin can only be used for unsafe math or if one of
6689 // the operands is known to be nonzero.
6690 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
6692 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6694 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
6703 // If LHS is NaN, an ordered comparison will be false and the result will
6704 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
6705 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6706 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
6707 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6709 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
6710 // will return +0, so vmax can only be used for unsafe math or if one of
6711 // the operands is known to be nonzero.
6712 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
6714 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6716 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
6722 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
6725 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
6726 DAGCombinerInfo &DCI) const {
6727 switch (N->getOpcode()) {
6729 case ISD::ADD: return PerformADDCombine(N, DCI);
6730 case ISD::SUB: return PerformSUBCombine(N, DCI);
6731 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
6732 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
6733 case ISD::AND: return PerformANDCombine(N, DCI);
6734 case ARMISD::BFI: return PerformBFICombine(N, DCI);
6735 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
6736 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
6737 case ISD::STORE: return PerformSTORECombine(N, DCI);
6738 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
6739 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
6740 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
6741 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
6742 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
6745 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
6746 case ISD::SIGN_EXTEND:
6747 case ISD::ZERO_EXTEND:
6748 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
6749 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
6750 case ARMISD::VLD2DUP:
6751 case ARMISD::VLD3DUP:
6752 case ARMISD::VLD4DUP:
6753 return CombineBaseUpdate(N, DCI);
6754 case ISD::INTRINSIC_VOID:
6755 case ISD::INTRINSIC_W_CHAIN:
6756 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
6757 case Intrinsic::arm_neon_vld1:
6758 case Intrinsic::arm_neon_vld2:
6759 case Intrinsic::arm_neon_vld3:
6760 case Intrinsic::arm_neon_vld4:
6761 case Intrinsic::arm_neon_vld2lane:
6762 case Intrinsic::arm_neon_vld3lane:
6763 case Intrinsic::arm_neon_vld4lane:
6764 case Intrinsic::arm_neon_vst1:
6765 case Intrinsic::arm_neon_vst2:
6766 case Intrinsic::arm_neon_vst3:
6767 case Intrinsic::arm_neon_vst4:
6768 case Intrinsic::arm_neon_vst2lane:
6769 case Intrinsic::arm_neon_vst3lane:
6770 case Intrinsic::arm_neon_vst4lane:
6771 return CombineBaseUpdate(N, DCI);
6779 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
6781 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
6784 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
6785 if (!Subtarget->allowsUnalignedMem())
6788 switch (VT.getSimpleVT().SimpleTy) {
6795 // FIXME: VLD1 etc with standard alignment is legal.
6799 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
6804 switch (VT.getSimpleVT().SimpleTy) {
6805 default: return false;
6820 if ((V & (Scale - 1)) != 0)
6823 return V == (V & ((1LL << 5) - 1));
6826 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
6827 const ARMSubtarget *Subtarget) {
6834 switch (VT.getSimpleVT().SimpleTy) {
6835 default: return false;
6840 // + imm12 or - imm8
6842 return V == (V & ((1LL << 8) - 1));
6843 return V == (V & ((1LL << 12) - 1));
6846 // Same as ARM mode. FIXME: NEON?
6847 if (!Subtarget->hasVFP2())
6852 return V == (V & ((1LL << 8) - 1));
6856 /// isLegalAddressImmediate - Return true if the integer value can be used
6857 /// as the offset of the target addressing mode for load / store of the
6859 static bool isLegalAddressImmediate(int64_t V, EVT VT,
6860 const ARMSubtarget *Subtarget) {
6867 if (Subtarget->isThumb1Only())
6868 return isLegalT1AddressImmediate(V, VT);
6869 else if (Subtarget->isThumb2())
6870 return isLegalT2AddressImmediate(V, VT, Subtarget);
6875 switch (VT.getSimpleVT().SimpleTy) {
6876 default: return false;
6881 return V == (V & ((1LL << 12) - 1));
6884 return V == (V & ((1LL << 8) - 1));
6887 if (!Subtarget->hasVFP2()) // FIXME: NEON?
6892 return V == (V & ((1LL << 8) - 1));
6896 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
6898 int Scale = AM.Scale;
6902 switch (VT.getSimpleVT().SimpleTy) {
6903 default: return false;
6912 return Scale == 2 || Scale == 4 || Scale == 8;
6915 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6919 // Note, we allow "void" uses (basically, uses that aren't loads or
6920 // stores), because arm allows folding a scale into many arithmetic
6921 // operations. This should be made more precise and revisited later.
6923 // Allow r << imm, but the imm has to be a multiple of two.
6924 if (Scale & 1) return false;
6925 return isPowerOf2_32(Scale);
6929 /// isLegalAddressingMode - Return true if the addressing mode represented
6930 /// by AM is legal for this target, for a load/store of the specified type.
6931 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6932 const Type *Ty) const {
6933 EVT VT = getValueType(Ty, true);
6934 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
6937 // Can never fold addr of global into load/store.
6942 case 0: // no scale reg, must be "r+i" or "r", or "i".
6945 if (Subtarget->isThumb1Only())
6949 // ARM doesn't support any R+R*scale+imm addr modes.
6956 if (Subtarget->isThumb2())
6957 return isLegalT2ScaledAddressingMode(AM, VT);
6959 int Scale = AM.Scale;
6960 switch (VT.getSimpleVT().SimpleTy) {
6961 default: return false;
6965 if (Scale < 0) Scale = -Scale;
6969 return isPowerOf2_32(Scale & ~1);
6973 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6978 // Note, we allow "void" uses (basically, uses that aren't loads or
6979 // stores), because arm allows folding a scale into many arithmetic
6980 // operations. This should be made more precise and revisited later.
6982 // Allow r << imm, but the imm has to be a multiple of two.
6983 if (Scale & 1) return false;
6984 return isPowerOf2_32(Scale);
6991 /// isLegalICmpImmediate - Return true if the specified immediate is legal
6992 /// icmp immediate, that is the target has icmp instructions which can compare
6993 /// a register against the immediate without having to materialize the
6994 /// immediate into a register.
6995 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
6996 if (!Subtarget->isThumb())
6997 return ARM_AM::getSOImmVal(Imm) != -1;
6998 if (Subtarget->isThumb2())
6999 return ARM_AM::getT2SOImmVal(Imm) != -1;
7000 return Imm >= 0 && Imm <= 255;
7003 /// isLegalAddImmediate - Return true if the specified immediate is legal
7004 /// add immediate, that is the target has add instructions which can add
7005 /// a register with the immediate without having to materialize the
7006 /// immediate into a register.
7007 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
7008 return ARM_AM::getSOImmVal(Imm) != -1;
7011 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
7012 bool isSEXTLoad, SDValue &Base,
7013 SDValue &Offset, bool &isInc,
7014 SelectionDAG &DAG) {
7015 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
7018 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
7020 Base = Ptr->getOperand(0);
7021 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7022 int RHSC = (int)RHS->getZExtValue();
7023 if (RHSC < 0 && RHSC > -256) {
7024 assert(Ptr->getOpcode() == ISD::ADD);
7026 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7030 isInc = (Ptr->getOpcode() == ISD::ADD);
7031 Offset = Ptr->getOperand(1);
7033 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
7035 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7036 int RHSC = (int)RHS->getZExtValue();
7037 if (RHSC < 0 && RHSC > -0x1000) {
7038 assert(Ptr->getOpcode() == ISD::ADD);
7040 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7041 Base = Ptr->getOperand(0);
7046 if (Ptr->getOpcode() == ISD::ADD) {
7048 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
7049 if (ShOpcVal != ARM_AM::no_shift) {
7050 Base = Ptr->getOperand(1);
7051 Offset = Ptr->getOperand(0);
7053 Base = Ptr->getOperand(0);
7054 Offset = Ptr->getOperand(1);
7059 isInc = (Ptr->getOpcode() == ISD::ADD);
7060 Base = Ptr->getOperand(0);
7061 Offset = Ptr->getOperand(1);
7065 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
7069 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
7070 bool isSEXTLoad, SDValue &Base,
7071 SDValue &Offset, bool &isInc,
7072 SelectionDAG &DAG) {
7073 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
7076 Base = Ptr->getOperand(0);
7077 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7078 int RHSC = (int)RHS->getZExtValue();
7079 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
7080 assert(Ptr->getOpcode() == ISD::ADD);
7082 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7084 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
7085 isInc = Ptr->getOpcode() == ISD::ADD;
7086 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
7094 /// getPreIndexedAddressParts - returns true by value, base pointer and
7095 /// offset pointer and addressing mode by reference if the node's address
7096 /// can be legally represented as pre-indexed load / store address.
7098 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7100 ISD::MemIndexedMode &AM,
7101 SelectionDAG &DAG) const {
7102 if (Subtarget->isThumb1Only())
7107 bool isSEXTLoad = false;
7108 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7109 Ptr = LD->getBasePtr();
7110 VT = LD->getMemoryVT();
7111 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7112 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7113 Ptr = ST->getBasePtr();
7114 VT = ST->getMemoryVT();
7119 bool isLegal = false;
7120 if (Subtarget->isThumb2())
7121 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7122 Offset, isInc, DAG);
7124 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7125 Offset, isInc, DAG);
7129 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
7133 /// getPostIndexedAddressParts - returns true by value, base pointer and
7134 /// offset pointer and addressing mode by reference if this node can be
7135 /// combined with a load / store to form a post-indexed load / store.
7136 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7139 ISD::MemIndexedMode &AM,
7140 SelectionDAG &DAG) const {
7141 if (Subtarget->isThumb1Only())
7146 bool isSEXTLoad = false;
7147 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7148 VT = LD->getMemoryVT();
7149 Ptr = LD->getBasePtr();
7150 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7151 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7152 VT = ST->getMemoryVT();
7153 Ptr = ST->getBasePtr();
7158 bool isLegal = false;
7159 if (Subtarget->isThumb2())
7160 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7163 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7169 // Swap base ptr and offset to catch more post-index load / store when
7170 // it's legal. In Thumb2 mode, offset must be an immediate.
7171 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
7172 !Subtarget->isThumb2())
7173 std::swap(Base, Offset);
7175 // Post-indexed load / store update the base pointer.
7180 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
7184 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7188 const SelectionDAG &DAG,
7189 unsigned Depth) const {
7190 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
7191 switch (Op.getOpcode()) {
7193 case ARMISD::CMOV: {
7194 // Bits are known zero/one if known on the LHS and RHS.
7195 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
7196 if (KnownZero == 0 && KnownOne == 0) return;
7198 APInt KnownZeroRHS, KnownOneRHS;
7199 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
7200 KnownZeroRHS, KnownOneRHS, Depth+1);
7201 KnownZero &= KnownZeroRHS;
7202 KnownOne &= KnownOneRHS;
7208 //===----------------------------------------------------------------------===//
7209 // ARM Inline Assembly Support
7210 //===----------------------------------------------------------------------===//
7212 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
7213 // Looking for "rev" which is V6+.
7214 if (!Subtarget->hasV6Ops())
7217 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
7218 std::string AsmStr = IA->getAsmString();
7219 SmallVector<StringRef, 4> AsmPieces;
7220 SplitString(AsmStr, AsmPieces, ";\n");
7222 switch (AsmPieces.size()) {
7223 default: return false;
7225 AsmStr = AsmPieces[0];
7227 SplitString(AsmStr, AsmPieces, " \t,");
7230 if (AsmPieces.size() == 3 &&
7231 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
7232 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
7233 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
7234 if (Ty && Ty->getBitWidth() == 32)
7235 return IntrinsicLowering::LowerToByteSwap(CI);
7243 /// getConstraintType - Given a constraint letter, return the type of
7244 /// constraint it is for this target.
7245 ARMTargetLowering::ConstraintType
7246 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
7247 if (Constraint.size() == 1) {
7248 switch (Constraint[0]) {
7250 case 'l': return C_RegisterClass;
7251 case 'w': return C_RegisterClass;
7254 return TargetLowering::getConstraintType(Constraint);
7257 /// Examine constraint type and operand type and determine a weight value.
7258 /// This object must already have been set up with the operand type
7259 /// and the current alternative constraint selected.
7260 TargetLowering::ConstraintWeight
7261 ARMTargetLowering::getSingleConstraintMatchWeight(
7262 AsmOperandInfo &info, const char *constraint) const {
7263 ConstraintWeight weight = CW_Invalid;
7264 Value *CallOperandVal = info.CallOperandVal;
7265 // If we don't have a value, we can't do a match,
7266 // but allow it at the lowest weight.
7267 if (CallOperandVal == NULL)
7269 const Type *type = CallOperandVal->getType();
7270 // Look at the constraint type.
7271 switch (*constraint) {
7273 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
7276 if (type->isIntegerTy()) {
7277 if (Subtarget->isThumb())
7278 weight = CW_SpecificReg;
7280 weight = CW_Register;
7284 if (type->isFloatingPointTy())
7285 weight = CW_Register;
7291 std::pair<unsigned, const TargetRegisterClass*>
7292 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7294 if (Constraint.size() == 1) {
7295 // GCC ARM Constraint Letters
7296 switch (Constraint[0]) {
7298 if (Subtarget->isThumb())
7299 return std::make_pair(0U, ARM::tGPRRegisterClass);
7301 return std::make_pair(0U, ARM::GPRRegisterClass);
7303 return std::make_pair(0U, ARM::GPRRegisterClass);
7306 return std::make_pair(0U, ARM::SPRRegisterClass);
7307 if (VT.getSizeInBits() == 64)
7308 return std::make_pair(0U, ARM::DPRRegisterClass);
7309 if (VT.getSizeInBits() == 128)
7310 return std::make_pair(0U, ARM::QPRRegisterClass);
7314 if (StringRef("{cc}").equals_lower(Constraint))
7315 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
7317 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7320 std::vector<unsigned> ARMTargetLowering::
7321 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7323 if (Constraint.size() != 1)
7324 return std::vector<unsigned>();
7326 switch (Constraint[0]) { // GCC ARM Constraint Letters
7329 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
7330 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
7333 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
7334 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
7335 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
7336 ARM::R12, ARM::LR, 0);
7339 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
7340 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
7341 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
7342 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
7343 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
7344 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
7345 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
7346 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
7347 if (VT.getSizeInBits() == 64)
7348 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
7349 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
7350 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
7351 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
7352 if (VT.getSizeInBits() == 128)
7353 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
7354 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
7358 return std::vector<unsigned>();
7361 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7362 /// vector. If it is invalid, don't add anything to Ops.
7363 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7365 std::vector<SDValue>&Ops,
7366 SelectionDAG &DAG) const {
7367 SDValue Result(0, 0);
7369 switch (Constraint) {
7371 case 'I': case 'J': case 'K': case 'L':
7372 case 'M': case 'N': case 'O':
7373 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
7377 int64_t CVal64 = C->getSExtValue();
7378 int CVal = (int) CVal64;
7379 // None of these constraints allow values larger than 32 bits. Check
7380 // that the value fits in an int.
7384 switch (Constraint) {
7386 if (Subtarget->isThumb1Only()) {
7387 // This must be a constant between 0 and 255, for ADD
7389 if (CVal >= 0 && CVal <= 255)
7391 } else if (Subtarget->isThumb2()) {
7392 // A constant that can be used as an immediate value in a
7393 // data-processing instruction.
7394 if (ARM_AM::getT2SOImmVal(CVal) != -1)
7397 // A constant that can be used as an immediate value in a
7398 // data-processing instruction.
7399 if (ARM_AM::getSOImmVal(CVal) != -1)
7405 if (Subtarget->isThumb()) { // FIXME thumb2
7406 // This must be a constant between -255 and -1, for negated ADD
7407 // immediates. This can be used in GCC with an "n" modifier that
7408 // prints the negated value, for use with SUB instructions. It is
7409 // not useful otherwise but is implemented for compatibility.
7410 if (CVal >= -255 && CVal <= -1)
7413 // This must be a constant between -4095 and 4095. It is not clear
7414 // what this constraint is intended for. Implemented for
7415 // compatibility with GCC.
7416 if (CVal >= -4095 && CVal <= 4095)
7422 if (Subtarget->isThumb1Only()) {
7423 // A 32-bit value where only one byte has a nonzero value. Exclude
7424 // zero to match GCC. This constraint is used by GCC internally for
7425 // constants that can be loaded with a move/shift combination.
7426 // It is not useful otherwise but is implemented for compatibility.
7427 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
7429 } else if (Subtarget->isThumb2()) {
7430 // A constant whose bitwise inverse can be used as an immediate
7431 // value in a data-processing instruction. This can be used in GCC
7432 // with a "B" modifier that prints the inverted value, for use with
7433 // BIC and MVN instructions. It is not useful otherwise but is
7434 // implemented for compatibility.
7435 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
7438 // A constant whose bitwise inverse can be used as an immediate
7439 // value in a data-processing instruction. This can be used in GCC
7440 // with a "B" modifier that prints the inverted value, for use with
7441 // BIC and MVN instructions. It is not useful otherwise but is
7442 // implemented for compatibility.
7443 if (ARM_AM::getSOImmVal(~CVal) != -1)
7449 if (Subtarget->isThumb1Only()) {
7450 // This must be a constant between -7 and 7,
7451 // for 3-operand ADD/SUB immediate instructions.
7452 if (CVal >= -7 && CVal < 7)
7454 } else if (Subtarget->isThumb2()) {
7455 // A constant whose negation can be used as an immediate value in a
7456 // data-processing instruction. This can be used in GCC with an "n"
7457 // modifier that prints the negated value, for use with SUB
7458 // instructions. It is not useful otherwise but is implemented for
7460 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
7463 // A constant whose negation can be used as an immediate value in a
7464 // data-processing instruction. This can be used in GCC with an "n"
7465 // modifier that prints the negated value, for use with SUB
7466 // instructions. It is not useful otherwise but is implemented for
7468 if (ARM_AM::getSOImmVal(-CVal) != -1)
7474 if (Subtarget->isThumb()) { // FIXME thumb2
7475 // This must be a multiple of 4 between 0 and 1020, for
7476 // ADD sp + immediate.
7477 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
7480 // A power of two or a constant between 0 and 32. This is used in
7481 // GCC for the shift amount on shifted register operands, but it is
7482 // useful in general for any shift amounts.
7483 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
7489 if (Subtarget->isThumb()) { // FIXME thumb2
7490 // This must be a constant between 0 and 31, for shift amounts.
7491 if (CVal >= 0 && CVal <= 31)
7497 if (Subtarget->isThumb()) { // FIXME thumb2
7498 // This must be a multiple of 4 between -508 and 508, for
7499 // ADD/SUB sp = sp + immediate.
7500 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
7505 Result = DAG.getTargetConstant(CVal, Op.getValueType());
7509 if (Result.getNode()) {
7510 Ops.push_back(Result);
7513 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7517 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
7518 // The ARM target isn't yet aware of offsets.
7522 int ARM::getVFPf32Imm(const APFloat &FPImm) {
7523 APInt Imm = FPImm.bitcastToAPInt();
7524 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
7525 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
7526 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
7528 // We can handle 4 bits of mantissa.
7529 // mantissa = (16+UInt(e:f:g:h))/16.
7530 if (Mantissa & 0x7ffff)
7533 if ((Mantissa & 0xf) != Mantissa)
7536 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
7537 if (Exp < -3 || Exp > 4)
7539 Exp = ((Exp+3) & 0x7) ^ 4;
7541 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
7544 int ARM::getVFPf64Imm(const APFloat &FPImm) {
7545 APInt Imm = FPImm.bitcastToAPInt();
7546 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
7547 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
7548 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
7550 // We can handle 4 bits of mantissa.
7551 // mantissa = (16+UInt(e:f:g:h))/16.
7552 if (Mantissa & 0xffffffffffffLL)
7555 if ((Mantissa & 0xf) != Mantissa)
7558 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
7559 if (Exp < -3 || Exp > 4)
7561 Exp = ((Exp+3) & 0x7) ^ 4;
7563 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
7566 bool ARM::isBitFieldInvertedMask(unsigned v) {
7567 if (v == 0xffffffff)
7569 // there can be 1's on either or both "outsides", all the "inside"
7571 unsigned int lsb = 0, msb = 31;
7572 while (v & (1 << msb)) --msb;
7573 while (v & (1 << lsb)) ++lsb;
7574 for (unsigned int i = lsb; i <= msb; ++i) {
7581 /// isFPImmLegal - Returns true if the target can instruction select the
7582 /// specified FP immediate natively. If false, the legalizer will
7583 /// materialize the FP immediate as a load from a constant pool.
7584 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
7585 if (!Subtarget->hasVFP3())
7588 return ARM::getVFPf32Imm(Imm) != -1;
7590 return ARM::getVFPf64Imm(Imm) != -1;
7594 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
7595 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
7596 /// specified in the intrinsic calls.
7597 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
7599 unsigned Intrinsic) const {
7600 switch (Intrinsic) {
7601 case Intrinsic::arm_neon_vld1:
7602 case Intrinsic::arm_neon_vld2:
7603 case Intrinsic::arm_neon_vld3:
7604 case Intrinsic::arm_neon_vld4:
7605 case Intrinsic::arm_neon_vld2lane:
7606 case Intrinsic::arm_neon_vld3lane:
7607 case Intrinsic::arm_neon_vld4lane: {
7608 Info.opc = ISD::INTRINSIC_W_CHAIN;
7609 // Conservatively set memVT to the entire set of vectors loaded.
7610 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
7611 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7612 Info.ptrVal = I.getArgOperand(0);
7614 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7615 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7616 Info.vol = false; // volatile loads with NEON intrinsics not supported
7617 Info.readMem = true;
7618 Info.writeMem = false;
7621 case Intrinsic::arm_neon_vst1:
7622 case Intrinsic::arm_neon_vst2:
7623 case Intrinsic::arm_neon_vst3:
7624 case Intrinsic::arm_neon_vst4:
7625 case Intrinsic::arm_neon_vst2lane:
7626 case Intrinsic::arm_neon_vst3lane:
7627 case Intrinsic::arm_neon_vst4lane: {
7628 Info.opc = ISD::INTRINSIC_VOID;
7629 // Conservatively set memVT to the entire set of vectors stored.
7630 unsigned NumElts = 0;
7631 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
7632 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
7633 if (!ArgTy->isVectorTy())
7635 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
7637 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7638 Info.ptrVal = I.getArgOperand(0);
7640 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7641 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7642 Info.vol = false; // volatile stores with NEON intrinsics not supported
7643 Info.readMem = false;
7644 Info.writeMem = true;