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 if (Subtarget->isThumb1Only())
402 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
404 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
405 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
406 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
407 if (!Subtarget->isFPOnlySP())
408 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
410 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
413 if (Subtarget->hasNEON()) {
414 addDRTypeForNEON(MVT::v2f32);
415 addDRTypeForNEON(MVT::v8i8);
416 addDRTypeForNEON(MVT::v4i16);
417 addDRTypeForNEON(MVT::v2i32);
418 addDRTypeForNEON(MVT::v1i64);
420 addQRTypeForNEON(MVT::v4f32);
421 addQRTypeForNEON(MVT::v2f64);
422 addQRTypeForNEON(MVT::v16i8);
423 addQRTypeForNEON(MVT::v8i16);
424 addQRTypeForNEON(MVT::v4i32);
425 addQRTypeForNEON(MVT::v2i64);
427 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
428 // neither Neon nor VFP support any arithmetic operations on it.
429 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
430 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
431 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
432 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
433 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
434 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
435 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
436 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
437 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
438 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
439 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
440 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
441 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
442 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
443 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
444 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
445 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
446 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
447 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
448 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
449 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
450 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
451 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
452 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
454 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
456 // Neon does not support some operations on v1i64 and v2i64 types.
457 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
458 // Custom handling for some quad-vector types to detect VMULL.
459 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
460 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
461 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
462 // Custom handling for some vector types to avoid expensive expansions
463 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
464 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
465 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
466 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
467 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
468 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
469 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
470 // a destination type that is wider than the source.
471 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
472 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
474 setTargetDAGCombine(ISD::INTRINSIC_VOID);
475 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
476 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
477 setTargetDAGCombine(ISD::SHL);
478 setTargetDAGCombine(ISD::SRL);
479 setTargetDAGCombine(ISD::SRA);
480 setTargetDAGCombine(ISD::SIGN_EXTEND);
481 setTargetDAGCombine(ISD::ZERO_EXTEND);
482 setTargetDAGCombine(ISD::ANY_EXTEND);
483 setTargetDAGCombine(ISD::SELECT_CC);
484 setTargetDAGCombine(ISD::BUILD_VECTOR);
485 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
486 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
487 setTargetDAGCombine(ISD::STORE);
490 computeRegisterProperties();
492 // ARM does not have f32 extending load.
493 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
495 // ARM does not have i1 sign extending load.
496 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
498 // ARM supports all 4 flavors of integer indexed load / store.
499 if (!Subtarget->isThumb1Only()) {
500 for (unsigned im = (unsigned)ISD::PRE_INC;
501 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
502 setIndexedLoadAction(im, MVT::i1, Legal);
503 setIndexedLoadAction(im, MVT::i8, Legal);
504 setIndexedLoadAction(im, MVT::i16, Legal);
505 setIndexedLoadAction(im, MVT::i32, Legal);
506 setIndexedStoreAction(im, MVT::i1, Legal);
507 setIndexedStoreAction(im, MVT::i8, Legal);
508 setIndexedStoreAction(im, MVT::i16, Legal);
509 setIndexedStoreAction(im, MVT::i32, Legal);
513 // i64 operation support.
514 setOperationAction(ISD::MUL, MVT::i64, Expand);
515 setOperationAction(ISD::MULHU, MVT::i32, Expand);
516 if (Subtarget->isThumb1Only()) {
517 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
518 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
520 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops())
521 setOperationAction(ISD::MULHS, MVT::i32, Expand);
523 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
524 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
525 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
526 setOperationAction(ISD::SRL, MVT::i64, Custom);
527 setOperationAction(ISD::SRA, MVT::i64, Custom);
529 // ARM does not have ROTL.
530 setOperationAction(ISD::ROTL, MVT::i32, Expand);
531 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
532 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
533 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
534 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
536 // Only ARMv6 has BSWAP.
537 if (!Subtarget->hasV6Ops())
538 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
540 // These are expanded into libcalls.
541 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
542 // v7M has a hardware divider
543 setOperationAction(ISD::SDIV, MVT::i32, Expand);
544 setOperationAction(ISD::UDIV, MVT::i32, Expand);
546 setOperationAction(ISD::SREM, MVT::i32, Expand);
547 setOperationAction(ISD::UREM, MVT::i32, Expand);
548 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
549 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
551 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
552 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
553 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
554 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
555 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
557 setOperationAction(ISD::TRAP, MVT::Other, Legal);
559 // Use the default implementation.
560 setOperationAction(ISD::VASTART, MVT::Other, Custom);
561 setOperationAction(ISD::VAARG, MVT::Other, Expand);
562 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
563 setOperationAction(ISD::VAEND, MVT::Other, Expand);
564 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
565 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
566 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
567 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
568 setExceptionPointerRegister(ARM::R0);
569 setExceptionSelectorRegister(ARM::R1);
571 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
572 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
573 // the default expansion.
574 if (Subtarget->hasDataBarrier() ||
575 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
576 // membarrier needs custom lowering; the rest are legal and handled
578 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
580 // Set them all for expansion, which will force libcalls.
581 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
582 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
583 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
584 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
585 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
586 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
587 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
588 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
589 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
590 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
591 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
592 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
593 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
595 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
596 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
600 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
601 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
602 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
603 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
604 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
605 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
606 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i8, Expand);
607 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i16, Expand);
608 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
609 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i8, Expand);
610 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i16, Expand);
611 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
612 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i8, Expand);
613 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i16, Expand);
614 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
615 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i8, Expand);
616 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i16, Expand);
617 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
618 // Since the libcalls include locking, fold in the fences
619 setShouldFoldAtomicFences(true);
621 // 64-bit versions are always libcalls (for now)
622 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
623 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
624 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
625 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
626 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
627 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
628 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
629 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
631 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
633 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
634 if (!Subtarget->hasV6Ops()) {
635 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
636 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
638 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
640 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
641 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
642 // iff target supports vfp2.
643 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
644 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
647 // We want to custom lower some of our intrinsics.
648 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
649 if (Subtarget->isTargetDarwin()) {
650 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
651 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
652 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
655 setOperationAction(ISD::SETCC, MVT::i32, Expand);
656 setOperationAction(ISD::SETCC, MVT::f32, Expand);
657 setOperationAction(ISD::SETCC, MVT::f64, Expand);
658 setOperationAction(ISD::SELECT, MVT::i32, Custom);
659 setOperationAction(ISD::SELECT, MVT::f32, Custom);
660 setOperationAction(ISD::SELECT, MVT::f64, Custom);
661 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
662 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
663 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
665 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
666 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
667 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
668 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
669 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
671 // We don't support sin/cos/fmod/copysign/pow
672 setOperationAction(ISD::FSIN, MVT::f64, Expand);
673 setOperationAction(ISD::FSIN, MVT::f32, Expand);
674 setOperationAction(ISD::FCOS, MVT::f32, Expand);
675 setOperationAction(ISD::FCOS, MVT::f64, Expand);
676 setOperationAction(ISD::FREM, MVT::f64, Expand);
677 setOperationAction(ISD::FREM, MVT::f32, Expand);
678 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
679 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
680 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
682 setOperationAction(ISD::FPOW, MVT::f64, Expand);
683 setOperationAction(ISD::FPOW, MVT::f32, Expand);
685 // Various VFP goodness
686 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
687 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
688 if (Subtarget->hasVFP2()) {
689 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
690 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
691 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
692 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
694 // Special handling for half-precision FP.
695 if (!Subtarget->hasFP16()) {
696 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
697 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
701 // We have target-specific dag combine patterns for the following nodes:
702 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
703 setTargetDAGCombine(ISD::ADD);
704 setTargetDAGCombine(ISD::SUB);
705 setTargetDAGCombine(ISD::MUL);
707 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
708 setTargetDAGCombine(ISD::OR);
709 if (Subtarget->hasNEON())
710 setTargetDAGCombine(ISD::AND);
712 setStackPointerRegisterToSaveRestore(ARM::SP);
714 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
715 setSchedulingPreference(Sched::RegPressure);
717 setSchedulingPreference(Sched::Hybrid);
719 //// temporary - rewrite interface to use type
720 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
722 // On ARM arguments smaller than 4 bytes are extended, so all arguments
723 // are at least 4 bytes aligned.
724 setMinStackArgumentAlignment(4);
726 benefitFromCodePlacementOpt = true;
729 // FIXME: It might make sense to define the representative register class as the
730 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
731 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
732 // SPR's representative would be DPR_VFP2. This should work well if register
733 // pressure tracking were modified such that a register use would increment the
734 // pressure of the register class's representative and all of it's super
735 // classes' representatives transitively. We have not implemented this because
736 // of the difficulty prior to coalescing of modeling operand register classes
737 // due to the common occurrence of cross class copies and subregister insertions
739 std::pair<const TargetRegisterClass*, uint8_t>
740 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
741 const TargetRegisterClass *RRC = 0;
743 switch (VT.getSimpleVT().SimpleTy) {
745 return TargetLowering::findRepresentativeClass(VT);
746 // Use DPR as representative register class for all floating point
747 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
748 // the cost is 1 for both f32 and f64.
749 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
750 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
751 RRC = ARM::DPRRegisterClass;
752 // When NEON is used for SP, only half of the register file is available
753 // because operations that define both SP and DP results will be constrained
754 // to the VFP2 class (D0-D15). We currently model this constraint prior to
755 // coalescing by double-counting the SP regs. See the FIXME above.
756 if (Subtarget->useNEONForSinglePrecisionFP())
759 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
760 case MVT::v4f32: case MVT::v2f64:
761 RRC = ARM::DPRRegisterClass;
765 RRC = ARM::DPRRegisterClass;
769 RRC = ARM::DPRRegisterClass;
773 return std::make_pair(RRC, Cost);
776 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
779 case ARMISD::Wrapper: return "ARMISD::Wrapper";
780 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
781 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
782 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
783 case ARMISD::CALL: return "ARMISD::CALL";
784 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
785 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
786 case ARMISD::tCALL: return "ARMISD::tCALL";
787 case ARMISD::BRCOND: return "ARMISD::BRCOND";
788 case ARMISD::BR_JT: return "ARMISD::BR_JT";
789 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
790 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
791 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
792 case ARMISD::CMP: return "ARMISD::CMP";
793 case ARMISD::CMPZ: return "ARMISD::CMPZ";
794 case ARMISD::CMPFP: return "ARMISD::CMPFP";
795 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
796 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
797 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
798 case ARMISD::CMOV: return "ARMISD::CMOV";
800 case ARMISD::RBIT: return "ARMISD::RBIT";
802 case ARMISD::FTOSI: return "ARMISD::FTOSI";
803 case ARMISD::FTOUI: return "ARMISD::FTOUI";
804 case ARMISD::SITOF: return "ARMISD::SITOF";
805 case ARMISD::UITOF: return "ARMISD::UITOF";
807 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
808 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
809 case ARMISD::RRX: return "ARMISD::RRX";
811 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
812 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
814 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
815 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
816 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
818 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
820 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
822 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
824 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
825 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
827 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
829 case ARMISD::VCEQ: return "ARMISD::VCEQ";
830 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
831 case ARMISD::VCGE: return "ARMISD::VCGE";
832 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
833 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
834 case ARMISD::VCGEU: return "ARMISD::VCGEU";
835 case ARMISD::VCGT: return "ARMISD::VCGT";
836 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
837 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
838 case ARMISD::VCGTU: return "ARMISD::VCGTU";
839 case ARMISD::VTST: return "ARMISD::VTST";
841 case ARMISD::VSHL: return "ARMISD::VSHL";
842 case ARMISD::VSHRs: return "ARMISD::VSHRs";
843 case ARMISD::VSHRu: return "ARMISD::VSHRu";
844 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
845 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
846 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
847 case ARMISD::VSHRN: return "ARMISD::VSHRN";
848 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
849 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
850 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
851 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
852 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
853 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
854 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
855 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
856 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
857 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
858 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
859 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
860 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
861 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
862 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
863 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
864 case ARMISD::VDUP: return "ARMISD::VDUP";
865 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
866 case ARMISD::VEXT: return "ARMISD::VEXT";
867 case ARMISD::VREV64: return "ARMISD::VREV64";
868 case ARMISD::VREV32: return "ARMISD::VREV32";
869 case ARMISD::VREV16: return "ARMISD::VREV16";
870 case ARMISD::VZIP: return "ARMISD::VZIP";
871 case ARMISD::VUZP: return "ARMISD::VUZP";
872 case ARMISD::VTRN: return "ARMISD::VTRN";
873 case ARMISD::VTBL1: return "ARMISD::VTBL1";
874 case ARMISD::VTBL2: return "ARMISD::VTBL2";
875 case ARMISD::VMULLs: return "ARMISD::VMULLs";
876 case ARMISD::VMULLu: return "ARMISD::VMULLu";
877 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
878 case ARMISD::FMAX: return "ARMISD::FMAX";
879 case ARMISD::FMIN: return "ARMISD::FMIN";
880 case ARMISD::BFI: return "ARMISD::BFI";
881 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
882 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
883 case ARMISD::VBSL: return "ARMISD::VBSL";
884 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
885 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
886 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
887 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
888 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
889 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
890 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
891 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
892 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
893 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
894 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
895 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
896 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
897 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
898 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
899 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
900 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
901 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
902 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
903 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
907 /// getRegClassFor - Return the register class that should be used for the
908 /// specified value type.
909 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
910 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
911 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
912 // load / store 4 to 8 consecutive D registers.
913 if (Subtarget->hasNEON()) {
914 if (VT == MVT::v4i64)
915 return ARM::QQPRRegisterClass;
916 else if (VT == MVT::v8i64)
917 return ARM::QQQQPRRegisterClass;
919 return TargetLowering::getRegClassFor(VT);
922 // Create a fast isel object.
924 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
925 return ARM::createFastISel(funcInfo);
928 /// getFunctionAlignment - Return the Log2 alignment of this function.
929 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
930 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
933 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
934 /// be used for loads / stores from the global.
935 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
936 return (Subtarget->isThumb1Only() ? 127 : 4095);
939 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
940 unsigned NumVals = N->getNumValues();
942 return Sched::RegPressure;
944 for (unsigned i = 0; i != NumVals; ++i) {
945 EVT VT = N->getValueType(i);
946 if (VT == MVT::Glue || VT == MVT::Other)
948 if (VT.isFloatingPoint() || VT.isVector())
949 return Sched::Latency;
952 if (!N->isMachineOpcode())
953 return Sched::RegPressure;
955 // Load are scheduled for latency even if there instruction itinerary
957 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
958 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
960 if (TID.getNumDefs() == 0)
961 return Sched::RegPressure;
962 if (!Itins->isEmpty() &&
963 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
964 return Sched::Latency;
966 return Sched::RegPressure;
969 //===----------------------------------------------------------------------===//
971 //===----------------------------------------------------------------------===//
973 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
974 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
976 default: llvm_unreachable("Unknown condition code!");
977 case ISD::SETNE: return ARMCC::NE;
978 case ISD::SETEQ: return ARMCC::EQ;
979 case ISD::SETGT: return ARMCC::GT;
980 case ISD::SETGE: return ARMCC::GE;
981 case ISD::SETLT: return ARMCC::LT;
982 case ISD::SETLE: return ARMCC::LE;
983 case ISD::SETUGT: return ARMCC::HI;
984 case ISD::SETUGE: return ARMCC::HS;
985 case ISD::SETULT: return ARMCC::LO;
986 case ISD::SETULE: return ARMCC::LS;
990 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
991 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
992 ARMCC::CondCodes &CondCode2) {
993 CondCode2 = ARMCC::AL;
995 default: llvm_unreachable("Unknown FP condition!");
997 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
999 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1001 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1002 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1003 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1004 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1005 case ISD::SETO: CondCode = ARMCC::VC; break;
1006 case ISD::SETUO: CondCode = ARMCC::VS; break;
1007 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1008 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1009 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1011 case ISD::SETULT: CondCode = ARMCC::LT; break;
1013 case ISD::SETULE: CondCode = ARMCC::LE; break;
1015 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1019 //===----------------------------------------------------------------------===//
1020 // Calling Convention Implementation
1021 //===----------------------------------------------------------------------===//
1023 #include "ARMGenCallingConv.inc"
1025 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1026 /// given CallingConvention value.
1027 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1029 bool isVarArg) const {
1032 llvm_unreachable("Unsupported calling convention");
1033 case CallingConv::Fast:
1034 if (Subtarget->hasVFP2() && !isVarArg) {
1035 if (!Subtarget->isAAPCS_ABI())
1036 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1037 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1038 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1041 case CallingConv::C: {
1042 // Use target triple & subtarget features to do actual dispatch.
1043 if (!Subtarget->isAAPCS_ABI())
1044 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1045 else if (Subtarget->hasVFP2() &&
1046 FloatABIType == FloatABI::Hard && !isVarArg)
1047 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1048 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1050 case CallingConv::ARM_AAPCS_VFP:
1051 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1052 case CallingConv::ARM_AAPCS:
1053 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1054 case CallingConv::ARM_APCS:
1055 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1059 /// LowerCallResult - Lower the result values of a call into the
1060 /// appropriate copies out of appropriate physical registers.
1062 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1063 CallingConv::ID CallConv, bool isVarArg,
1064 const SmallVectorImpl<ISD::InputArg> &Ins,
1065 DebugLoc dl, SelectionDAG &DAG,
1066 SmallVectorImpl<SDValue> &InVals) const {
1068 // Assign locations to each value returned by this call.
1069 SmallVector<CCValAssign, 16> RVLocs;
1070 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1071 RVLocs, *DAG.getContext());
1072 CCInfo.AnalyzeCallResult(Ins,
1073 CCAssignFnForNode(CallConv, /* Return*/ true,
1076 // Copy all of the result registers out of their specified physreg.
1077 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1078 CCValAssign VA = RVLocs[i];
1081 if (VA.needsCustom()) {
1082 // Handle f64 or half of a v2f64.
1083 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1085 Chain = Lo.getValue(1);
1086 InFlag = Lo.getValue(2);
1087 VA = RVLocs[++i]; // skip ahead to next loc
1088 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1090 Chain = Hi.getValue(1);
1091 InFlag = Hi.getValue(2);
1092 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1094 if (VA.getLocVT() == MVT::v2f64) {
1095 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1096 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1097 DAG.getConstant(0, MVT::i32));
1099 VA = RVLocs[++i]; // skip ahead to next loc
1100 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1101 Chain = Lo.getValue(1);
1102 InFlag = Lo.getValue(2);
1103 VA = RVLocs[++i]; // skip ahead to next loc
1104 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1105 Chain = Hi.getValue(1);
1106 InFlag = Hi.getValue(2);
1107 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1108 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1109 DAG.getConstant(1, MVT::i32));
1112 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1114 Chain = Val.getValue(1);
1115 InFlag = Val.getValue(2);
1118 switch (VA.getLocInfo()) {
1119 default: llvm_unreachable("Unknown loc info!");
1120 case CCValAssign::Full: break;
1121 case CCValAssign::BCvt:
1122 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1126 InVals.push_back(Val);
1132 /// LowerMemOpCallTo - Store the argument to the stack.
1134 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1135 SDValue StackPtr, SDValue Arg,
1136 DebugLoc dl, SelectionDAG &DAG,
1137 const CCValAssign &VA,
1138 ISD::ArgFlagsTy Flags) const {
1139 unsigned LocMemOffset = VA.getLocMemOffset();
1140 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1141 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1142 return DAG.getStore(Chain, dl, Arg, PtrOff,
1143 MachinePointerInfo::getStack(LocMemOffset),
1147 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1148 SDValue Chain, SDValue &Arg,
1149 RegsToPassVector &RegsToPass,
1150 CCValAssign &VA, CCValAssign &NextVA,
1152 SmallVector<SDValue, 8> &MemOpChains,
1153 ISD::ArgFlagsTy Flags) const {
1155 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1156 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1157 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1159 if (NextVA.isRegLoc())
1160 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1162 assert(NextVA.isMemLoc());
1163 if (StackPtr.getNode() == 0)
1164 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1166 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1172 /// LowerCall - Lowering a call into a callseq_start <-
1173 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1176 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1177 CallingConv::ID CallConv, bool isVarArg,
1179 const SmallVectorImpl<ISD::OutputArg> &Outs,
1180 const SmallVectorImpl<SDValue> &OutVals,
1181 const SmallVectorImpl<ISD::InputArg> &Ins,
1182 DebugLoc dl, SelectionDAG &DAG,
1183 SmallVectorImpl<SDValue> &InVals) const {
1184 MachineFunction &MF = DAG.getMachineFunction();
1185 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1186 bool IsSibCall = false;
1187 // Temporarily disable tail calls so things don't break.
1188 if (!EnableARMTailCalls)
1191 // Check if it's really possible to do a tail call.
1192 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1193 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1194 Outs, OutVals, Ins, DAG);
1195 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1196 // detected sibcalls.
1203 // Analyze operands of the call, assigning locations to each operand.
1204 SmallVector<CCValAssign, 16> ArgLocs;
1205 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1207 CCInfo.setCallOrPrologue(Call);
1208 CCInfo.AnalyzeCallOperands(Outs,
1209 CCAssignFnForNode(CallConv, /* Return*/ false,
1212 // Get a count of how many bytes are to be pushed on the stack.
1213 unsigned NumBytes = CCInfo.getNextStackOffset();
1215 // For tail calls, memory operands are available in our caller's stack.
1219 // Adjust the stack pointer for the new arguments...
1220 // These operations are automatically eliminated by the prolog/epilog pass
1222 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1224 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1226 RegsToPassVector RegsToPass;
1227 SmallVector<SDValue, 8> MemOpChains;
1229 // Walk the register/memloc assignments, inserting copies/loads. In the case
1230 // of tail call optimization, arguments are handled later.
1231 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1233 ++i, ++realArgIdx) {
1234 CCValAssign &VA = ArgLocs[i];
1235 SDValue Arg = OutVals[realArgIdx];
1236 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1237 bool isByVal = Flags.isByVal();
1239 // Promote the value if needed.
1240 switch (VA.getLocInfo()) {
1241 default: llvm_unreachable("Unknown loc info!");
1242 case CCValAssign::Full: break;
1243 case CCValAssign::SExt:
1244 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1246 case CCValAssign::ZExt:
1247 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1249 case CCValAssign::AExt:
1250 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1252 case CCValAssign::BCvt:
1253 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1257 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1258 if (VA.needsCustom()) {
1259 if (VA.getLocVT() == MVT::v2f64) {
1260 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1261 DAG.getConstant(0, MVT::i32));
1262 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1263 DAG.getConstant(1, MVT::i32));
1265 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1266 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1268 VA = ArgLocs[++i]; // skip ahead to next loc
1269 if (VA.isRegLoc()) {
1270 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1271 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1273 assert(VA.isMemLoc());
1275 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1276 dl, DAG, VA, Flags));
1279 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1280 StackPtr, MemOpChains, Flags);
1282 } else if (VA.isRegLoc()) {
1283 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1284 } else if (isByVal) {
1285 assert(VA.isMemLoc());
1286 unsigned offset = 0;
1288 // True if this byval aggregate will be split between registers
1290 if (CCInfo.isFirstByValRegValid()) {
1291 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1293 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1294 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1295 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1296 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1297 MachinePointerInfo(),
1299 MemOpChains.push_back(Load.getValue(1));
1300 RegsToPass.push_back(std::make_pair(j, Load));
1302 offset = ARM::R4 - CCInfo.getFirstByValReg();
1303 CCInfo.clearFirstByValReg();
1306 unsigned LocMemOffset = VA.getLocMemOffset();
1307 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1308 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1310 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1311 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1312 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1314 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1315 Flags.getByValAlign(),
1316 /*isVolatile=*/false,
1317 /*AlwaysInline=*/false,
1318 MachinePointerInfo(0),
1319 MachinePointerInfo(0)));
1321 } else if (!IsSibCall) {
1322 assert(VA.isMemLoc());
1324 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1325 dl, DAG, VA, Flags));
1329 if (!MemOpChains.empty())
1330 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1331 &MemOpChains[0], MemOpChains.size());
1333 // Build a sequence of copy-to-reg nodes chained together with token chain
1334 // and flag operands which copy the outgoing args into the appropriate regs.
1336 // Tail call byval lowering might overwrite argument registers so in case of
1337 // tail call optimization the copies to registers are lowered later.
1339 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1340 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1341 RegsToPass[i].second, InFlag);
1342 InFlag = Chain.getValue(1);
1345 // For tail calls lower the arguments to the 'real' stack slot.
1347 // Force all the incoming stack arguments to be loaded from the stack
1348 // before any new outgoing arguments are stored to the stack, because the
1349 // outgoing stack slots may alias the incoming argument stack slots, and
1350 // the alias isn't otherwise explicit. This is slightly more conservative
1351 // than necessary, because it means that each store effectively depends
1352 // on every argument instead of just those arguments it would clobber.
1354 // Do not flag preceding copytoreg stuff together with the following stuff.
1356 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1357 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1358 RegsToPass[i].second, InFlag);
1359 InFlag = Chain.getValue(1);
1364 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1365 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1366 // node so that legalize doesn't hack it.
1367 bool isDirect = false;
1368 bool isARMFunc = false;
1369 bool isLocalARMFunc = false;
1370 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1372 if (EnableARMLongCalls) {
1373 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1374 && "long-calls with non-static relocation model!");
1375 // Handle a global address or an external symbol. If it's not one of
1376 // those, the target's already in a register, so we don't need to do
1378 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1379 const GlobalValue *GV = G->getGlobal();
1380 // Create a constant pool entry for the callee address
1381 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1382 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1385 // Get the address of the callee into a register
1386 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1387 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1388 Callee = DAG.getLoad(getPointerTy(), dl,
1389 DAG.getEntryNode(), CPAddr,
1390 MachinePointerInfo::getConstantPool(),
1392 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1393 const char *Sym = S->getSymbol();
1395 // Create a constant pool entry for the callee address
1396 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1397 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1398 Sym, ARMPCLabelIndex, 0);
1399 // Get the address of the callee into a register
1400 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1401 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1402 Callee = DAG.getLoad(getPointerTy(), dl,
1403 DAG.getEntryNode(), CPAddr,
1404 MachinePointerInfo::getConstantPool(),
1407 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1408 const GlobalValue *GV = G->getGlobal();
1410 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1411 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1412 getTargetMachine().getRelocationModel() != Reloc::Static;
1413 isARMFunc = !Subtarget->isThumb() || isStub;
1414 // ARM call to a local ARM function is predicable.
1415 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1416 // tBX takes a register source operand.
1417 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1418 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1419 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1422 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1423 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1424 Callee = DAG.getLoad(getPointerTy(), dl,
1425 DAG.getEntryNode(), CPAddr,
1426 MachinePointerInfo::getConstantPool(),
1428 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1429 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1430 getPointerTy(), Callee, PICLabel);
1432 // On ELF targets for PIC code, direct calls should go through the PLT
1433 unsigned OpFlags = 0;
1434 if (Subtarget->isTargetELF() &&
1435 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1436 OpFlags = ARMII::MO_PLT;
1437 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1439 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1441 bool isStub = Subtarget->isTargetDarwin() &&
1442 getTargetMachine().getRelocationModel() != Reloc::Static;
1443 isARMFunc = !Subtarget->isThumb() || isStub;
1444 // tBX takes a register source operand.
1445 const char *Sym = S->getSymbol();
1446 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1447 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1448 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1449 Sym, ARMPCLabelIndex, 4);
1450 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1451 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1452 Callee = DAG.getLoad(getPointerTy(), dl,
1453 DAG.getEntryNode(), CPAddr,
1454 MachinePointerInfo::getConstantPool(),
1456 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1457 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1458 getPointerTy(), Callee, PICLabel);
1460 unsigned OpFlags = 0;
1461 // On ELF targets for PIC code, direct calls should go through the PLT
1462 if (Subtarget->isTargetELF() &&
1463 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1464 OpFlags = ARMII::MO_PLT;
1465 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1469 // FIXME: handle tail calls differently.
1471 if (Subtarget->isThumb()) {
1472 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1473 CallOpc = ARMISD::CALL_NOLINK;
1475 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1477 CallOpc = (isDirect || Subtarget->hasV5TOps())
1478 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1479 : ARMISD::CALL_NOLINK;
1482 std::vector<SDValue> Ops;
1483 Ops.push_back(Chain);
1484 Ops.push_back(Callee);
1486 // Add argument registers to the end of the list so that they are known live
1488 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1489 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1490 RegsToPass[i].second.getValueType()));
1492 if (InFlag.getNode())
1493 Ops.push_back(InFlag);
1495 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1497 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1499 // Returns a chain and a flag for retval copy to use.
1500 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1501 InFlag = Chain.getValue(1);
1503 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1504 DAG.getIntPtrConstant(0, true), InFlag);
1506 InFlag = Chain.getValue(1);
1508 // Handle result values, copying them out of physregs into vregs that we
1510 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1514 /// HandleByVal - Every parameter *after* a byval parameter is passed
1515 /// on the stack. Remember the next parameter register to allocate,
1516 /// and then confiscate the rest of the parameter registers to insure
1519 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1520 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1521 assert((State->getCallOrPrologue() == Prologue ||
1522 State->getCallOrPrologue() == Call) &&
1523 "unhandled ParmContext");
1524 if ((!State->isFirstByValRegValid()) &&
1525 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1526 State->setFirstByValReg(reg);
1527 // At a call site, a byval parameter that is split between
1528 // registers and memory needs its size truncated here. In a
1529 // function prologue, such byval parameters are reassembled in
1530 // memory, and are not truncated.
1531 if (State->getCallOrPrologue() == Call) {
1532 unsigned excess = 4 * (ARM::R4 - reg);
1533 assert(size >= excess && "expected larger existing stack allocation");
1537 // Confiscate any remaining parameter registers to preclude their
1538 // assignment to subsequent parameters.
1539 while (State->AllocateReg(GPRArgRegs, 4))
1543 /// MatchingStackOffset - Return true if the given stack call argument is
1544 /// already available in the same position (relatively) of the caller's
1545 /// incoming argument stack.
1547 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1548 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1549 const ARMInstrInfo *TII) {
1550 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1552 if (Arg.getOpcode() == ISD::CopyFromReg) {
1553 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1554 if (!TargetRegisterInfo::isVirtualRegister(VR))
1556 MachineInstr *Def = MRI->getVRegDef(VR);
1559 if (!Flags.isByVal()) {
1560 if (!TII->isLoadFromStackSlot(Def, FI))
1565 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1566 if (Flags.isByVal())
1567 // ByVal argument is passed in as a pointer but it's now being
1568 // dereferenced. e.g.
1569 // define @foo(%struct.X* %A) {
1570 // tail call @bar(%struct.X* byval %A)
1573 SDValue Ptr = Ld->getBasePtr();
1574 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1577 FI = FINode->getIndex();
1581 assert(FI != INT_MAX);
1582 if (!MFI->isFixedObjectIndex(FI))
1584 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1587 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1588 /// for tail call optimization. Targets which want to do tail call
1589 /// optimization should implement this function.
1591 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1592 CallingConv::ID CalleeCC,
1594 bool isCalleeStructRet,
1595 bool isCallerStructRet,
1596 const SmallVectorImpl<ISD::OutputArg> &Outs,
1597 const SmallVectorImpl<SDValue> &OutVals,
1598 const SmallVectorImpl<ISD::InputArg> &Ins,
1599 SelectionDAG& DAG) const {
1600 const Function *CallerF = DAG.getMachineFunction().getFunction();
1601 CallingConv::ID CallerCC = CallerF->getCallingConv();
1602 bool CCMatch = CallerCC == CalleeCC;
1604 // Look for obvious safe cases to perform tail call optimization that do not
1605 // require ABI changes. This is what gcc calls sibcall.
1607 // Do not sibcall optimize vararg calls unless the call site is not passing
1609 if (isVarArg && !Outs.empty())
1612 // Also avoid sibcall optimization if either caller or callee uses struct
1613 // return semantics.
1614 if (isCalleeStructRet || isCallerStructRet)
1617 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1618 // emitEpilogue is not ready for them.
1619 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1620 // LR. This means if we need to reload LR, it takes an extra instructions,
1621 // which outweighs the value of the tail call; but here we don't know yet
1622 // whether LR is going to be used. Probably the right approach is to
1623 // generate the tail call here and turn it back into CALL/RET in
1624 // emitEpilogue if LR is used.
1626 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1627 // but we need to make sure there are enough registers; the only valid
1628 // registers are the 4 used for parameters. We don't currently do this
1630 if (Subtarget->isThumb1Only())
1633 // If the calling conventions do not match, then we'd better make sure the
1634 // results are returned in the same way as what the caller expects.
1636 SmallVector<CCValAssign, 16> RVLocs1;
1637 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1638 RVLocs1, *DAG.getContext());
1639 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1641 SmallVector<CCValAssign, 16> RVLocs2;
1642 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1643 RVLocs2, *DAG.getContext());
1644 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1646 if (RVLocs1.size() != RVLocs2.size())
1648 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1649 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1651 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1653 if (RVLocs1[i].isRegLoc()) {
1654 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1657 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1663 // If the callee takes no arguments then go on to check the results of the
1665 if (!Outs.empty()) {
1666 // Check if stack adjustment is needed. For now, do not do this if any
1667 // argument is passed on the stack.
1668 SmallVector<CCValAssign, 16> ArgLocs;
1669 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1670 ArgLocs, *DAG.getContext());
1671 CCInfo.AnalyzeCallOperands(Outs,
1672 CCAssignFnForNode(CalleeCC, false, isVarArg));
1673 if (CCInfo.getNextStackOffset()) {
1674 MachineFunction &MF = DAG.getMachineFunction();
1676 // Check if the arguments are already laid out in the right way as
1677 // the caller's fixed stack objects.
1678 MachineFrameInfo *MFI = MF.getFrameInfo();
1679 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1680 const ARMInstrInfo *TII =
1681 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1682 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1684 ++i, ++realArgIdx) {
1685 CCValAssign &VA = ArgLocs[i];
1686 EVT RegVT = VA.getLocVT();
1687 SDValue Arg = OutVals[realArgIdx];
1688 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1689 if (VA.getLocInfo() == CCValAssign::Indirect)
1691 if (VA.needsCustom()) {
1692 // f64 and vector types are split into multiple registers or
1693 // register/stack-slot combinations. The types will not match
1694 // the registers; give up on memory f64 refs until we figure
1695 // out what to do about this.
1698 if (!ArgLocs[++i].isRegLoc())
1700 if (RegVT == MVT::v2f64) {
1701 if (!ArgLocs[++i].isRegLoc())
1703 if (!ArgLocs[++i].isRegLoc())
1706 } else if (!VA.isRegLoc()) {
1707 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1719 ARMTargetLowering::LowerReturn(SDValue Chain,
1720 CallingConv::ID CallConv, bool isVarArg,
1721 const SmallVectorImpl<ISD::OutputArg> &Outs,
1722 const SmallVectorImpl<SDValue> &OutVals,
1723 DebugLoc dl, SelectionDAG &DAG) const {
1725 // CCValAssign - represent the assignment of the return value to a location.
1726 SmallVector<CCValAssign, 16> RVLocs;
1728 // CCState - Info about the registers and stack slots.
1729 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1732 // Analyze outgoing return values.
1733 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1736 // If this is the first return lowered for this function, add
1737 // the regs to the liveout set for the function.
1738 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1739 for (unsigned i = 0; i != RVLocs.size(); ++i)
1740 if (RVLocs[i].isRegLoc())
1741 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1746 // Copy the result values into the output registers.
1747 for (unsigned i = 0, realRVLocIdx = 0;
1749 ++i, ++realRVLocIdx) {
1750 CCValAssign &VA = RVLocs[i];
1751 assert(VA.isRegLoc() && "Can only return in registers!");
1753 SDValue Arg = OutVals[realRVLocIdx];
1755 switch (VA.getLocInfo()) {
1756 default: llvm_unreachable("Unknown loc info!");
1757 case CCValAssign::Full: break;
1758 case CCValAssign::BCvt:
1759 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1763 if (VA.needsCustom()) {
1764 if (VA.getLocVT() == MVT::v2f64) {
1765 // Extract the first half and return it in two registers.
1766 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1767 DAG.getConstant(0, MVT::i32));
1768 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1769 DAG.getVTList(MVT::i32, MVT::i32), Half);
1771 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1772 Flag = Chain.getValue(1);
1773 VA = RVLocs[++i]; // skip ahead to next loc
1774 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1775 HalfGPRs.getValue(1), Flag);
1776 Flag = Chain.getValue(1);
1777 VA = RVLocs[++i]; // skip ahead to next loc
1779 // Extract the 2nd half and fall through to handle it as an f64 value.
1780 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1781 DAG.getConstant(1, MVT::i32));
1783 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1785 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1786 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1787 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1788 Flag = Chain.getValue(1);
1789 VA = RVLocs[++i]; // skip ahead to next loc
1790 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1793 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1795 // Guarantee that all emitted copies are
1796 // stuck together, avoiding something bad.
1797 Flag = Chain.getValue(1);
1802 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1804 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1809 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1810 if (N->getNumValues() != 1)
1812 if (!N->hasNUsesOfValue(1, 0))
1815 unsigned NumCopies = 0;
1817 SDNode *Use = *N->use_begin();
1818 if (Use->getOpcode() == ISD::CopyToReg) {
1819 Copies[NumCopies++] = Use;
1820 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1821 // f64 returned in a pair of GPRs.
1822 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1824 if (UI->getOpcode() != ISD::CopyToReg)
1826 Copies[UI.getUse().getResNo()] = *UI;
1829 } else if (Use->getOpcode() == ISD::BITCAST) {
1830 // f32 returned in a single GPR.
1831 if (!Use->hasNUsesOfValue(1, 0))
1833 Use = *Use->use_begin();
1834 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1836 Copies[NumCopies++] = Use;
1841 if (NumCopies != 1 && NumCopies != 2)
1844 bool HasRet = false;
1845 for (unsigned i = 0; i < NumCopies; ++i) {
1846 SDNode *Copy = Copies[i];
1847 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1849 if (UI->getOpcode() == ISD::CopyToReg) {
1851 if (Use == Copies[0] || Use == Copies[1])
1855 if (UI->getOpcode() != ARMISD::RET_FLAG)
1864 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1865 if (!EnableARMTailCalls)
1868 if (!CI->isTailCall())
1871 return !Subtarget->isThumb1Only();
1874 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1875 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1876 // one of the above mentioned nodes. It has to be wrapped because otherwise
1877 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1878 // be used to form addressing mode. These wrapped nodes will be selected
1880 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1881 EVT PtrVT = Op.getValueType();
1882 // FIXME there is no actual debug info here
1883 DebugLoc dl = Op.getDebugLoc();
1884 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1886 if (CP->isMachineConstantPoolEntry())
1887 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1888 CP->getAlignment());
1890 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1891 CP->getAlignment());
1892 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1895 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1896 return MachineJumpTableInfo::EK_Inline;
1899 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1900 SelectionDAG &DAG) const {
1901 MachineFunction &MF = DAG.getMachineFunction();
1902 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1903 unsigned ARMPCLabelIndex = 0;
1904 DebugLoc DL = Op.getDebugLoc();
1905 EVT PtrVT = getPointerTy();
1906 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1907 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1909 if (RelocM == Reloc::Static) {
1910 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1912 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1913 ARMPCLabelIndex = AFI->createPICLabelUId();
1914 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1915 ARMCP::CPBlockAddress,
1917 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1919 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1920 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1921 MachinePointerInfo::getConstantPool(),
1923 if (RelocM == Reloc::Static)
1925 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1926 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1929 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1931 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1932 SelectionDAG &DAG) const {
1933 DebugLoc dl = GA->getDebugLoc();
1934 EVT PtrVT = getPointerTy();
1935 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1936 MachineFunction &MF = DAG.getMachineFunction();
1937 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1938 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1939 ARMConstantPoolValue *CPV =
1940 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1941 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1942 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1943 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1944 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1945 MachinePointerInfo::getConstantPool(),
1947 SDValue Chain = Argument.getValue(1);
1949 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1950 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1952 // call __tls_get_addr.
1955 Entry.Node = Argument;
1956 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1957 Args.push_back(Entry);
1958 // FIXME: is there useful debug info available here?
1959 std::pair<SDValue, SDValue> CallResult =
1960 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1961 false, false, false, false,
1962 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1963 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1964 return CallResult.first;
1967 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1968 // "local exec" model.
1970 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1971 SelectionDAG &DAG) const {
1972 const GlobalValue *GV = GA->getGlobal();
1973 DebugLoc dl = GA->getDebugLoc();
1975 SDValue Chain = DAG.getEntryNode();
1976 EVT PtrVT = getPointerTy();
1977 // Get the Thread Pointer
1978 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1980 if (GV->isDeclaration()) {
1981 MachineFunction &MF = DAG.getMachineFunction();
1982 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1983 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1984 // Initial exec model.
1985 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1986 ARMConstantPoolValue *CPV =
1987 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1988 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1989 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1990 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1991 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1992 MachinePointerInfo::getConstantPool(),
1994 Chain = Offset.getValue(1);
1996 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1997 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1999 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2000 MachinePointerInfo::getConstantPool(),
2004 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
2005 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2006 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2007 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2008 MachinePointerInfo::getConstantPool(),
2012 // The address of the thread local variable is the add of the thread
2013 // pointer with the offset of the variable.
2014 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2018 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2019 // TODO: implement the "local dynamic" model
2020 assert(Subtarget->isTargetELF() &&
2021 "TLS not implemented for non-ELF targets");
2022 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2023 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2024 // otherwise use the "Local Exec" TLS Model
2025 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2026 return LowerToTLSGeneralDynamicModel(GA, DAG);
2028 return LowerToTLSExecModels(GA, DAG);
2031 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2032 SelectionDAG &DAG) const {
2033 EVT PtrVT = getPointerTy();
2034 DebugLoc dl = Op.getDebugLoc();
2035 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2036 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2037 if (RelocM == Reloc::PIC_) {
2038 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2039 ARMConstantPoolValue *CPV =
2040 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2041 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2042 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2043 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2045 MachinePointerInfo::getConstantPool(),
2047 SDValue Chain = Result.getValue(1);
2048 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2049 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2051 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2052 MachinePointerInfo::getGOT(), false, false, 0);
2056 // If we have T2 ops, we can materialize the address directly via movt/movw
2057 // pair. This is always cheaper.
2058 if (Subtarget->useMovt()) {
2060 // FIXME: Once remat is capable of dealing with instructions with register
2061 // operands, expand this into two nodes.
2062 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2063 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2065 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2066 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2067 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2068 MachinePointerInfo::getConstantPool(),
2073 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2074 SelectionDAG &DAG) const {
2075 EVT PtrVT = getPointerTy();
2076 DebugLoc dl = Op.getDebugLoc();
2077 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2078 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2079 MachineFunction &MF = DAG.getMachineFunction();
2080 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2082 if (Subtarget->useMovt()) {
2084 // FIXME: Once remat is capable of dealing with instructions with register
2085 // operands, expand this into two nodes.
2086 if (RelocM == Reloc::Static)
2087 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2088 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2090 unsigned Wrapper = (RelocM == Reloc::PIC_)
2091 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2092 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2093 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2094 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2095 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2096 MachinePointerInfo::getGOT(), false, false, 0);
2100 unsigned ARMPCLabelIndex = 0;
2102 if (RelocM == Reloc::Static) {
2103 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2105 ARMPCLabelIndex = AFI->createPICLabelUId();
2106 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2107 ARMConstantPoolValue *CPV =
2108 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2109 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2111 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2113 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2114 MachinePointerInfo::getConstantPool(),
2116 SDValue Chain = Result.getValue(1);
2118 if (RelocM == Reloc::PIC_) {
2119 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2120 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2123 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2124 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2130 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2131 SelectionDAG &DAG) const {
2132 assert(Subtarget->isTargetELF() &&
2133 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2134 MachineFunction &MF = DAG.getMachineFunction();
2135 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2136 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2137 EVT PtrVT = getPointerTy();
2138 DebugLoc dl = Op.getDebugLoc();
2139 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2140 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2141 "_GLOBAL_OFFSET_TABLE_",
2142 ARMPCLabelIndex, PCAdj);
2143 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2144 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2145 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2146 MachinePointerInfo::getConstantPool(),
2148 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2149 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2153 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2155 DebugLoc dl = Op.getDebugLoc();
2156 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2161 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2162 DebugLoc dl = Op.getDebugLoc();
2163 SDValue Val = DAG.getConstant(0, MVT::i32);
2164 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2165 Op.getOperand(1), Val);
2169 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2170 DebugLoc dl = Op.getDebugLoc();
2171 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2172 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2176 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2177 const ARMSubtarget *Subtarget) const {
2178 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2179 DebugLoc dl = Op.getDebugLoc();
2181 default: return SDValue(); // Don't custom lower most intrinsics.
2182 case Intrinsic::arm_thread_pointer: {
2183 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2184 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2186 case Intrinsic::eh_sjlj_lsda: {
2187 MachineFunction &MF = DAG.getMachineFunction();
2188 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2189 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2190 EVT PtrVT = getPointerTy();
2191 DebugLoc dl = Op.getDebugLoc();
2192 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2194 unsigned PCAdj = (RelocM != Reloc::PIC_)
2195 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2196 ARMConstantPoolValue *CPV =
2197 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2198 ARMCP::CPLSDA, PCAdj);
2199 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2200 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2202 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2203 MachinePointerInfo::getConstantPool(),
2206 if (RelocM == Reloc::PIC_) {
2207 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2208 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2212 case Intrinsic::arm_neon_vmulls:
2213 case Intrinsic::arm_neon_vmullu: {
2214 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2215 ? ARMISD::VMULLs : ARMISD::VMULLu;
2216 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2217 Op.getOperand(1), Op.getOperand(2));
2222 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2223 const ARMSubtarget *Subtarget) {
2224 DebugLoc dl = Op.getDebugLoc();
2225 if (!Subtarget->hasDataBarrier()) {
2226 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2227 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2229 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2230 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2231 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2232 DAG.getConstant(0, MVT::i32));
2235 SDValue Op5 = Op.getOperand(5);
2236 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2237 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2238 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2239 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2241 ARM_MB::MemBOpt DMBOpt;
2242 if (isDeviceBarrier)
2243 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2245 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2246 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2247 DAG.getConstant(DMBOpt, MVT::i32));
2250 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2251 const ARMSubtarget *Subtarget) {
2252 // ARM pre v5TE and Thumb1 does not have preload instructions.
2253 if (!(Subtarget->isThumb2() ||
2254 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2255 // Just preserve the chain.
2256 return Op.getOperand(0);
2258 DebugLoc dl = Op.getDebugLoc();
2259 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2261 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2262 // ARMv7 with MP extension has PLDW.
2263 return Op.getOperand(0);
2265 if (Subtarget->isThumb())
2267 isRead = ~isRead & 1;
2268 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2270 // Currently there is no intrinsic that matches pli.
2271 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2272 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2273 DAG.getConstant(isData, MVT::i32));
2276 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2277 MachineFunction &MF = DAG.getMachineFunction();
2278 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2280 // vastart just stores the address of the VarArgsFrameIndex slot into the
2281 // memory location argument.
2282 DebugLoc dl = Op.getDebugLoc();
2283 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2284 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2285 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2286 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2287 MachinePointerInfo(SV), false, false, 0);
2291 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2292 SDValue &Root, SelectionDAG &DAG,
2293 DebugLoc dl) const {
2294 MachineFunction &MF = DAG.getMachineFunction();
2295 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2297 TargetRegisterClass *RC;
2298 if (AFI->isThumb1OnlyFunction())
2299 RC = ARM::tGPRRegisterClass;
2301 RC = ARM::GPRRegisterClass;
2303 // Transform the arguments stored in physical registers into virtual ones.
2304 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2305 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2308 if (NextVA.isMemLoc()) {
2309 MachineFrameInfo *MFI = MF.getFrameInfo();
2310 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2312 // Create load node to retrieve arguments from the stack.
2313 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2314 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2315 MachinePointerInfo::getFixedStack(FI),
2318 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2319 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2322 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2326 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2327 unsigned &VARegSize, unsigned &VARegSaveSize)
2330 if (CCInfo.isFirstByValRegValid())
2331 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2333 unsigned int firstUnalloced;
2334 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2335 sizeof(GPRArgRegs) /
2336 sizeof(GPRArgRegs[0]));
2337 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2340 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2341 VARegSize = NumGPRs * 4;
2342 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2345 // The remaining GPRs hold either the beginning of variable-argument
2346 // data, or the beginning of an aggregate passed by value (usuall
2347 // byval). Either way, we allocate stack slots adjacent to the data
2348 // provided by our caller, and store the unallocated registers there.
2349 // If this is a variadic function, the va_list pointer will begin with
2350 // these values; otherwise, this reassembles a (byval) structure that
2351 // was split between registers and memory.
2353 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2354 DebugLoc dl, SDValue &Chain,
2355 unsigned ArgOffset) const {
2356 MachineFunction &MF = DAG.getMachineFunction();
2357 MachineFrameInfo *MFI = MF.getFrameInfo();
2358 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2359 unsigned firstRegToSaveIndex;
2360 if (CCInfo.isFirstByValRegValid())
2361 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2363 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2364 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2367 unsigned VARegSize, VARegSaveSize;
2368 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2369 if (VARegSaveSize) {
2370 // If this function is vararg, store any remaining integer argument regs
2371 // to their spots on the stack so that they may be loaded by deferencing
2372 // the result of va_next.
2373 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2374 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2375 ArgOffset + VARegSaveSize
2378 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2381 SmallVector<SDValue, 4> MemOps;
2382 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2383 TargetRegisterClass *RC;
2384 if (AFI->isThumb1OnlyFunction())
2385 RC = ARM::tGPRRegisterClass;
2387 RC = ARM::GPRRegisterClass;
2389 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2390 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2392 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2393 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2395 MemOps.push_back(Store);
2396 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2397 DAG.getConstant(4, getPointerTy()));
2399 if (!MemOps.empty())
2400 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2401 &MemOps[0], MemOps.size());
2403 // This will point to the next argument passed via stack.
2404 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2408 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2409 CallingConv::ID CallConv, bool isVarArg,
2410 const SmallVectorImpl<ISD::InputArg>
2412 DebugLoc dl, SelectionDAG &DAG,
2413 SmallVectorImpl<SDValue> &InVals)
2415 MachineFunction &MF = DAG.getMachineFunction();
2416 MachineFrameInfo *MFI = MF.getFrameInfo();
2418 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2420 // Assign locations to all of the incoming arguments.
2421 SmallVector<CCValAssign, 16> ArgLocs;
2422 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2424 CCInfo.setCallOrPrologue(Prologue);
2425 CCInfo.AnalyzeFormalArguments(Ins,
2426 CCAssignFnForNode(CallConv, /* Return*/ false,
2429 SmallVector<SDValue, 16> ArgValues;
2430 int lastInsIndex = -1;
2433 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2434 CCValAssign &VA = ArgLocs[i];
2436 // Arguments stored in registers.
2437 if (VA.isRegLoc()) {
2438 EVT RegVT = VA.getLocVT();
2440 if (VA.needsCustom()) {
2441 // f64 and vector types are split up into multiple registers or
2442 // combinations of registers and stack slots.
2443 if (VA.getLocVT() == MVT::v2f64) {
2444 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2446 VA = ArgLocs[++i]; // skip ahead to next loc
2448 if (VA.isMemLoc()) {
2449 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2450 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2451 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2452 MachinePointerInfo::getFixedStack(FI),
2455 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2458 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2459 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2460 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2461 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2462 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2464 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2467 TargetRegisterClass *RC;
2469 if (RegVT == MVT::f32)
2470 RC = ARM::SPRRegisterClass;
2471 else if (RegVT == MVT::f64)
2472 RC = ARM::DPRRegisterClass;
2473 else if (RegVT == MVT::v2f64)
2474 RC = ARM::QPRRegisterClass;
2475 else if (RegVT == MVT::i32)
2476 RC = (AFI->isThumb1OnlyFunction() ?
2477 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2479 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2481 // Transform the arguments in physical registers into virtual ones.
2482 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2483 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2486 // If this is an 8 or 16-bit value, it is really passed promoted
2487 // to 32 bits. Insert an assert[sz]ext to capture this, then
2488 // truncate to the right size.
2489 switch (VA.getLocInfo()) {
2490 default: llvm_unreachable("Unknown loc info!");
2491 case CCValAssign::Full: break;
2492 case CCValAssign::BCvt:
2493 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2495 case CCValAssign::SExt:
2496 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2497 DAG.getValueType(VA.getValVT()));
2498 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2500 case CCValAssign::ZExt:
2501 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2502 DAG.getValueType(VA.getValVT()));
2503 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2507 InVals.push_back(ArgValue);
2509 } else { // VA.isRegLoc()
2512 assert(VA.isMemLoc());
2513 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2515 int index = ArgLocs[i].getValNo();
2517 // Some Ins[] entries become multiple ArgLoc[] entries.
2518 // Process them only once.
2519 if (index != lastInsIndex)
2521 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2522 // FIXME: For now, all byval parameter objects are marked mutable.
2523 // This can be changed with more analysis.
2524 // In case of tail call optimization mark all arguments mutable.
2525 // Since they could be overwritten by lowering of arguments in case of
2527 if (Flags.isByVal()) {
2528 unsigned VARegSize, VARegSaveSize;
2529 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2530 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2531 unsigned Bytes = Flags.getByValSize() - VARegSize;
2532 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2533 int FI = MFI->CreateFixedObject(Bytes,
2534 VA.getLocMemOffset(), false);
2535 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2537 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2538 VA.getLocMemOffset(), true);
2540 // Create load nodes to retrieve arguments from the stack.
2541 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2542 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2543 MachinePointerInfo::getFixedStack(FI),
2546 lastInsIndex = index;
2553 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2558 /// isFloatingPointZero - Return true if this is +0.0.
2559 static bool isFloatingPointZero(SDValue Op) {
2560 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2561 return CFP->getValueAPF().isPosZero();
2562 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2563 // Maybe this has already been legalized into the constant pool?
2564 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2565 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2566 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2567 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2568 return CFP->getValueAPF().isPosZero();
2574 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2575 /// the given operands.
2577 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2578 SDValue &ARMcc, SelectionDAG &DAG,
2579 DebugLoc dl) const {
2580 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2581 unsigned C = RHSC->getZExtValue();
2582 if (!isLegalICmpImmediate(C)) {
2583 // Constant does not fit, try adjusting it by one?
2588 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2589 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2590 RHS = DAG.getConstant(C-1, MVT::i32);
2595 if (C != 0 && isLegalICmpImmediate(C-1)) {
2596 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2597 RHS = DAG.getConstant(C-1, MVT::i32);
2602 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2603 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2604 RHS = DAG.getConstant(C+1, MVT::i32);
2609 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2610 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2611 RHS = DAG.getConstant(C+1, MVT::i32);
2618 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2619 ARMISD::NodeType CompareType;
2622 CompareType = ARMISD::CMP;
2627 CompareType = ARMISD::CMPZ;
2630 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2631 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2634 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2636 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2637 DebugLoc dl) const {
2639 if (!isFloatingPointZero(RHS))
2640 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2642 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2643 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2646 /// duplicateCmp - Glue values can have only one use, so this function
2647 /// duplicates a comparison node.
2649 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2650 unsigned Opc = Cmp.getOpcode();
2651 DebugLoc DL = Cmp.getDebugLoc();
2652 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2653 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2655 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2656 Cmp = Cmp.getOperand(0);
2657 Opc = Cmp.getOpcode();
2658 if (Opc == ARMISD::CMPFP)
2659 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2661 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2662 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2664 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2667 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2668 SDValue Cond = Op.getOperand(0);
2669 SDValue SelectTrue = Op.getOperand(1);
2670 SDValue SelectFalse = Op.getOperand(2);
2671 DebugLoc dl = Op.getDebugLoc();
2675 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2676 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2678 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2679 const ConstantSDNode *CMOVTrue =
2680 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2681 const ConstantSDNode *CMOVFalse =
2682 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2684 if (CMOVTrue && CMOVFalse) {
2685 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2686 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2690 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2692 False = SelectFalse;
2693 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2698 if (True.getNode() && False.getNode()) {
2699 EVT VT = Cond.getValueType();
2700 SDValue ARMcc = Cond.getOperand(2);
2701 SDValue CCR = Cond.getOperand(3);
2702 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2703 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2708 return DAG.getSelectCC(dl, Cond,
2709 DAG.getConstant(0, Cond.getValueType()),
2710 SelectTrue, SelectFalse, ISD::SETNE);
2713 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2714 EVT VT = Op.getValueType();
2715 SDValue LHS = Op.getOperand(0);
2716 SDValue RHS = Op.getOperand(1);
2717 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2718 SDValue TrueVal = Op.getOperand(2);
2719 SDValue FalseVal = Op.getOperand(3);
2720 DebugLoc dl = Op.getDebugLoc();
2722 if (LHS.getValueType() == MVT::i32) {
2724 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2725 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2726 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2729 ARMCC::CondCodes CondCode, CondCode2;
2730 FPCCToARMCC(CC, CondCode, CondCode2);
2732 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2733 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2734 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2735 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2737 if (CondCode2 != ARMCC::AL) {
2738 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2739 // FIXME: Needs another CMP because flag can have but one use.
2740 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2741 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2742 Result, TrueVal, ARMcc2, CCR, Cmp2);
2747 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2748 /// to morph to an integer compare sequence.
2749 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2750 const ARMSubtarget *Subtarget) {
2751 SDNode *N = Op.getNode();
2752 if (!N->hasOneUse())
2753 // Otherwise it requires moving the value from fp to integer registers.
2755 if (!N->getNumValues())
2757 EVT VT = Op.getValueType();
2758 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2759 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2760 // vmrs are very slow, e.g. cortex-a8.
2763 if (isFloatingPointZero(Op)) {
2767 return ISD::isNormalLoad(N);
2770 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2771 if (isFloatingPointZero(Op))
2772 return DAG.getConstant(0, MVT::i32);
2774 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2775 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2776 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2777 Ld->isVolatile(), Ld->isNonTemporal(),
2778 Ld->getAlignment());
2780 llvm_unreachable("Unknown VFP cmp argument!");
2783 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2784 SDValue &RetVal1, SDValue &RetVal2) {
2785 if (isFloatingPointZero(Op)) {
2786 RetVal1 = DAG.getConstant(0, MVT::i32);
2787 RetVal2 = DAG.getConstant(0, MVT::i32);
2791 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2792 SDValue Ptr = Ld->getBasePtr();
2793 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2794 Ld->getChain(), Ptr,
2795 Ld->getPointerInfo(),
2796 Ld->isVolatile(), Ld->isNonTemporal(),
2797 Ld->getAlignment());
2799 EVT PtrType = Ptr.getValueType();
2800 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2801 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2802 PtrType, Ptr, DAG.getConstant(4, PtrType));
2803 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2804 Ld->getChain(), NewPtr,
2805 Ld->getPointerInfo().getWithOffset(4),
2806 Ld->isVolatile(), Ld->isNonTemporal(),
2811 llvm_unreachable("Unknown VFP cmp argument!");
2814 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2815 /// f32 and even f64 comparisons to integer ones.
2817 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2818 SDValue Chain = Op.getOperand(0);
2819 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2820 SDValue LHS = Op.getOperand(2);
2821 SDValue RHS = Op.getOperand(3);
2822 SDValue Dest = Op.getOperand(4);
2823 DebugLoc dl = Op.getDebugLoc();
2825 bool SeenZero = false;
2826 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2827 canChangeToInt(RHS, SeenZero, Subtarget) &&
2828 // If one of the operand is zero, it's safe to ignore the NaN case since
2829 // we only care about equality comparisons.
2830 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2831 // If unsafe fp math optimization is enabled and there are no other uses of
2832 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2833 // to an integer comparison.
2834 if (CC == ISD::SETOEQ)
2836 else if (CC == ISD::SETUNE)
2840 if (LHS.getValueType() == MVT::f32) {
2841 LHS = bitcastf32Toi32(LHS, DAG);
2842 RHS = bitcastf32Toi32(RHS, DAG);
2843 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2844 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2845 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2846 Chain, Dest, ARMcc, CCR, Cmp);
2851 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2852 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2853 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2854 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2855 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2856 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2857 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2863 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2864 SDValue Chain = Op.getOperand(0);
2865 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2866 SDValue LHS = Op.getOperand(2);
2867 SDValue RHS = Op.getOperand(3);
2868 SDValue Dest = Op.getOperand(4);
2869 DebugLoc dl = Op.getDebugLoc();
2871 if (LHS.getValueType() == MVT::i32) {
2873 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2874 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2875 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2876 Chain, Dest, ARMcc, CCR, Cmp);
2879 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2882 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2883 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2884 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2885 if (Result.getNode())
2889 ARMCC::CondCodes CondCode, CondCode2;
2890 FPCCToARMCC(CC, CondCode, CondCode2);
2892 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2893 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2894 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2895 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2896 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2897 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2898 if (CondCode2 != ARMCC::AL) {
2899 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2900 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2901 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2906 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2907 SDValue Chain = Op.getOperand(0);
2908 SDValue Table = Op.getOperand(1);
2909 SDValue Index = Op.getOperand(2);
2910 DebugLoc dl = Op.getDebugLoc();
2912 EVT PTy = getPointerTy();
2913 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2914 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2915 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2916 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2917 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2918 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2919 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2920 if (Subtarget->isThumb2()) {
2921 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2922 // which does another jump to the destination. This also makes it easier
2923 // to translate it to TBB / TBH later.
2924 // FIXME: This might not work if the function is extremely large.
2925 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2926 Addr, Op.getOperand(2), JTI, UId);
2928 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2929 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2930 MachinePointerInfo::getJumpTable(),
2932 Chain = Addr.getValue(1);
2933 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2934 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2936 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2937 MachinePointerInfo::getJumpTable(), false, false, 0);
2938 Chain = Addr.getValue(1);
2939 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2943 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2944 DebugLoc dl = Op.getDebugLoc();
2947 switch (Op.getOpcode()) {
2949 assert(0 && "Invalid opcode!");
2950 case ISD::FP_TO_SINT:
2951 Opc = ARMISD::FTOSI;
2953 case ISD::FP_TO_UINT:
2954 Opc = ARMISD::FTOUI;
2957 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2958 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2961 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2962 EVT VT = Op.getValueType();
2963 DebugLoc dl = Op.getDebugLoc();
2965 EVT OperandVT = Op.getOperand(0).getValueType();
2966 assert(OperandVT == MVT::v4i16 && "Invalid type for custom lowering!");
2967 if (VT != MVT::v4f32)
2968 return DAG.UnrollVectorOp(Op.getNode());
2972 switch (Op.getOpcode()) {
2974 assert(0 && "Invalid opcode!");
2975 case ISD::SINT_TO_FP:
2976 CastOpc = ISD::SIGN_EXTEND;
2977 Opc = ISD::SINT_TO_FP;
2979 case ISD::UINT_TO_FP:
2980 CastOpc = ISD::ZERO_EXTEND;
2981 Opc = ISD::UINT_TO_FP;
2985 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
2986 return DAG.getNode(Opc, dl, VT, Op);
2989 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2990 EVT VT = Op.getValueType();
2992 return LowerVectorINT_TO_FP(Op, DAG);
2994 DebugLoc dl = Op.getDebugLoc();
2997 switch (Op.getOpcode()) {
2999 assert(0 && "Invalid opcode!");
3000 case ISD::SINT_TO_FP:
3001 Opc = ARMISD::SITOF;
3003 case ISD::UINT_TO_FP:
3004 Opc = ARMISD::UITOF;
3008 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3009 return DAG.getNode(Opc, dl, VT, Op);
3012 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3013 // Implement fcopysign with a fabs and a conditional fneg.
3014 SDValue Tmp0 = Op.getOperand(0);
3015 SDValue Tmp1 = Op.getOperand(1);
3016 DebugLoc dl = Op.getDebugLoc();
3017 EVT VT = Op.getValueType();
3018 EVT SrcVT = Tmp1.getValueType();
3019 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3020 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3021 bool UseNEON = !InGPR && Subtarget->hasNEON();
3024 // Use VBSL to copy the sign bit.
3025 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3026 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3027 DAG.getTargetConstant(EncodedVal, MVT::i32));
3028 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3030 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3031 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3032 DAG.getConstant(32, MVT::i32));
3033 else /*if (VT == MVT::f32)*/
3034 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3035 if (SrcVT == MVT::f32) {
3036 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3038 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3039 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3040 DAG.getConstant(32, MVT::i32));
3041 } else if (VT == MVT::f32)
3042 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3043 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3044 DAG.getConstant(32, MVT::i32));
3045 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3046 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3048 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3050 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3051 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3052 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3054 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3055 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3056 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3057 if (VT == MVT::f32) {
3058 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3059 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3060 DAG.getConstant(0, MVT::i32));
3062 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3068 // Bitcast operand 1 to i32.
3069 if (SrcVT == MVT::f64)
3070 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3071 &Tmp1, 1).getValue(1);
3072 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3074 // Or in the signbit with integer operations.
3075 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3076 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3077 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3078 if (VT == MVT::f32) {
3079 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3080 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3081 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3082 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3085 // f64: Or the high part with signbit and then combine two parts.
3086 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3088 SDValue Lo = Tmp0.getValue(0);
3089 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3090 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3091 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3094 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3095 MachineFunction &MF = DAG.getMachineFunction();
3096 MachineFrameInfo *MFI = MF.getFrameInfo();
3097 MFI->setReturnAddressIsTaken(true);
3099 EVT VT = Op.getValueType();
3100 DebugLoc dl = Op.getDebugLoc();
3101 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3103 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3104 SDValue Offset = DAG.getConstant(4, MVT::i32);
3105 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3106 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3107 MachinePointerInfo(), false, false, 0);
3110 // Return LR, which contains the return address. Mark it an implicit live-in.
3111 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3112 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3115 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3116 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3117 MFI->setFrameAddressIsTaken(true);
3119 EVT VT = Op.getValueType();
3120 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3121 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3122 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3123 ? ARM::R7 : ARM::R11;
3124 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3126 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3127 MachinePointerInfo(),
3132 /// ExpandBITCAST - If the target supports VFP, this function is called to
3133 /// expand a bit convert where either the source or destination type is i64 to
3134 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3135 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3136 /// vectors), since the legalizer won't know what to do with that.
3137 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3138 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3139 DebugLoc dl = N->getDebugLoc();
3140 SDValue Op = N->getOperand(0);
3142 // This function is only supposed to be called for i64 types, either as the
3143 // source or destination of the bit convert.
3144 EVT SrcVT = Op.getValueType();
3145 EVT DstVT = N->getValueType(0);
3146 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3147 "ExpandBITCAST called for non-i64 type");
3149 // Turn i64->f64 into VMOVDRR.
3150 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3151 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3152 DAG.getConstant(0, MVT::i32));
3153 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3154 DAG.getConstant(1, MVT::i32));
3155 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3156 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3159 // Turn f64->i64 into VMOVRRD.
3160 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3161 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3162 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3163 // Merge the pieces into a single i64 value.
3164 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3170 /// getZeroVector - Returns a vector of specified type with all zero elements.
3171 /// Zero vectors are used to represent vector negation and in those cases
3172 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3173 /// not support i64 elements, so sometimes the zero vectors will need to be
3174 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3176 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3177 assert(VT.isVector() && "Expected a vector type");
3178 // The canonical modified immediate encoding of a zero vector is....0!
3179 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3180 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3181 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3182 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3185 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3186 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3187 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3188 SelectionDAG &DAG) const {
3189 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3190 EVT VT = Op.getValueType();
3191 unsigned VTBits = VT.getSizeInBits();
3192 DebugLoc dl = Op.getDebugLoc();
3193 SDValue ShOpLo = Op.getOperand(0);
3194 SDValue ShOpHi = Op.getOperand(1);
3195 SDValue ShAmt = Op.getOperand(2);
3197 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3199 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3201 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3202 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3203 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3204 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3205 DAG.getConstant(VTBits, MVT::i32));
3206 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3207 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3208 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3210 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3211 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3213 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3214 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3217 SDValue Ops[2] = { Lo, Hi };
3218 return DAG.getMergeValues(Ops, 2, dl);
3221 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3222 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3223 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3224 SelectionDAG &DAG) const {
3225 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3226 EVT VT = Op.getValueType();
3227 unsigned VTBits = VT.getSizeInBits();
3228 DebugLoc dl = Op.getDebugLoc();
3229 SDValue ShOpLo = Op.getOperand(0);
3230 SDValue ShOpHi = Op.getOperand(1);
3231 SDValue ShAmt = Op.getOperand(2);
3234 assert(Op.getOpcode() == ISD::SHL_PARTS);
3235 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3236 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3237 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3238 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3239 DAG.getConstant(VTBits, MVT::i32));
3240 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3241 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3243 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3244 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3245 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3247 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3248 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3251 SDValue Ops[2] = { Lo, Hi };
3252 return DAG.getMergeValues(Ops, 2, dl);
3255 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3256 SelectionDAG &DAG) const {
3257 // The rounding mode is in bits 23:22 of the FPSCR.
3258 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3259 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3260 // so that the shift + and get folded into a bitfield extract.
3261 DebugLoc dl = Op.getDebugLoc();
3262 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3263 DAG.getConstant(Intrinsic::arm_get_fpscr,
3265 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3266 DAG.getConstant(1U << 22, MVT::i32));
3267 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3268 DAG.getConstant(22, MVT::i32));
3269 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3270 DAG.getConstant(3, MVT::i32));
3273 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3274 const ARMSubtarget *ST) {
3275 EVT VT = N->getValueType(0);
3276 DebugLoc dl = N->getDebugLoc();
3278 if (!ST->hasV6T2Ops())
3281 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3282 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3285 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3286 const ARMSubtarget *ST) {
3287 EVT VT = N->getValueType(0);
3288 DebugLoc dl = N->getDebugLoc();
3293 // Lower vector shifts on NEON to use VSHL.
3294 assert(ST->hasNEON() && "unexpected vector shift");
3296 // Left shifts translate directly to the vshiftu intrinsic.
3297 if (N->getOpcode() == ISD::SHL)
3298 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3299 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3300 N->getOperand(0), N->getOperand(1));
3302 assert((N->getOpcode() == ISD::SRA ||
3303 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3305 // NEON uses the same intrinsics for both left and right shifts. For
3306 // right shifts, the shift amounts are negative, so negate the vector of
3308 EVT ShiftVT = N->getOperand(1).getValueType();
3309 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3310 getZeroVector(ShiftVT, DAG, dl),
3312 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3313 Intrinsic::arm_neon_vshifts :
3314 Intrinsic::arm_neon_vshiftu);
3315 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3316 DAG.getConstant(vshiftInt, MVT::i32),
3317 N->getOperand(0), NegatedCount);
3320 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3321 const ARMSubtarget *ST) {
3322 EVT VT = N->getValueType(0);
3323 DebugLoc dl = N->getDebugLoc();
3325 // We can get here for a node like i32 = ISD::SHL i32, i64
3329 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3330 "Unknown shift to lower!");
3332 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3333 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3334 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3337 // If we are in thumb mode, we don't have RRX.
3338 if (ST->isThumb1Only()) return SDValue();
3340 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3341 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3342 DAG.getConstant(0, MVT::i32));
3343 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3344 DAG.getConstant(1, MVT::i32));
3346 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3347 // captures the result into a carry flag.
3348 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3349 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3351 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3352 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3354 // Merge the pieces into a single i64 value.
3355 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3358 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3359 SDValue TmpOp0, TmpOp1;
3360 bool Invert = false;
3364 SDValue Op0 = Op.getOperand(0);
3365 SDValue Op1 = Op.getOperand(1);
3366 SDValue CC = Op.getOperand(2);
3367 EVT VT = Op.getValueType();
3368 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3369 DebugLoc dl = Op.getDebugLoc();
3371 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3372 switch (SetCCOpcode) {
3373 default: llvm_unreachable("Illegal FP comparison"); break;
3375 case ISD::SETNE: Invert = true; // Fallthrough
3377 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3379 case ISD::SETLT: Swap = true; // Fallthrough
3381 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3383 case ISD::SETLE: Swap = true; // Fallthrough
3385 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3386 case ISD::SETUGE: Swap = true; // Fallthrough
3387 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3388 case ISD::SETUGT: Swap = true; // Fallthrough
3389 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3390 case ISD::SETUEQ: Invert = true; // Fallthrough
3392 // Expand this to (OLT | OGT).
3396 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3397 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3399 case ISD::SETUO: Invert = true; // Fallthrough
3401 // Expand this to (OLT | OGE).
3405 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3406 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3410 // Integer comparisons.
3411 switch (SetCCOpcode) {
3412 default: llvm_unreachable("Illegal integer comparison"); break;
3413 case ISD::SETNE: Invert = true;
3414 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3415 case ISD::SETLT: Swap = true;
3416 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3417 case ISD::SETLE: Swap = true;
3418 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3419 case ISD::SETULT: Swap = true;
3420 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3421 case ISD::SETULE: Swap = true;
3422 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3425 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3426 if (Opc == ARMISD::VCEQ) {
3429 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3431 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3434 // Ignore bitconvert.
3435 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3436 AndOp = AndOp.getOperand(0);
3438 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3440 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3441 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3448 std::swap(Op0, Op1);
3450 // If one of the operands is a constant vector zero, attempt to fold the
3451 // comparison to a specialized compare-against-zero form.
3453 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3455 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3456 if (Opc == ARMISD::VCGE)
3457 Opc = ARMISD::VCLEZ;
3458 else if (Opc == ARMISD::VCGT)
3459 Opc = ARMISD::VCLTZ;
3464 if (SingleOp.getNode()) {
3467 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3469 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3471 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3473 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3475 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3477 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3480 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3484 Result = DAG.getNOT(dl, Result, VT);
3489 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3490 /// valid vector constant for a NEON instruction with a "modified immediate"
3491 /// operand (e.g., VMOV). If so, return the encoded value.
3492 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3493 unsigned SplatBitSize, SelectionDAG &DAG,
3494 EVT &VT, bool is128Bits, NEONModImmType type) {
3495 unsigned OpCmode, Imm;
3497 // SplatBitSize is set to the smallest size that splats the vector, so a
3498 // zero vector will always have SplatBitSize == 8. However, NEON modified
3499 // immediate instructions others than VMOV do not support the 8-bit encoding
3500 // of a zero vector, and the default encoding of zero is supposed to be the
3505 switch (SplatBitSize) {
3507 if (type != VMOVModImm)
3509 // Any 1-byte value is OK. Op=0, Cmode=1110.
3510 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3513 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3517 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3518 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3519 if ((SplatBits & ~0xff) == 0) {
3520 // Value = 0x00nn: Op=x, Cmode=100x.
3525 if ((SplatBits & ~0xff00) == 0) {
3526 // Value = 0xnn00: Op=x, Cmode=101x.
3528 Imm = SplatBits >> 8;
3534 // NEON's 32-bit VMOV supports splat values where:
3535 // * only one byte is nonzero, or
3536 // * the least significant byte is 0xff and the second byte is nonzero, or
3537 // * the least significant 2 bytes are 0xff and the third is nonzero.
3538 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3539 if ((SplatBits & ~0xff) == 0) {
3540 // Value = 0x000000nn: Op=x, Cmode=000x.
3545 if ((SplatBits & ~0xff00) == 0) {
3546 // Value = 0x0000nn00: Op=x, Cmode=001x.
3548 Imm = SplatBits >> 8;
3551 if ((SplatBits & ~0xff0000) == 0) {
3552 // Value = 0x00nn0000: Op=x, Cmode=010x.
3554 Imm = SplatBits >> 16;
3557 if ((SplatBits & ~0xff000000) == 0) {
3558 // Value = 0xnn000000: Op=x, Cmode=011x.
3560 Imm = SplatBits >> 24;
3564 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3565 if (type == OtherModImm) return SDValue();
3567 if ((SplatBits & ~0xffff) == 0 &&
3568 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3569 // Value = 0x0000nnff: Op=x, Cmode=1100.
3571 Imm = SplatBits >> 8;
3576 if ((SplatBits & ~0xffffff) == 0 &&
3577 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3578 // Value = 0x00nnffff: Op=x, Cmode=1101.
3580 Imm = SplatBits >> 16;
3581 SplatBits |= 0xffff;
3585 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3586 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3587 // VMOV.I32. A (very) minor optimization would be to replicate the value
3588 // and fall through here to test for a valid 64-bit splat. But, then the
3589 // caller would also need to check and handle the change in size.
3593 if (type != VMOVModImm)
3595 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3596 uint64_t BitMask = 0xff;
3598 unsigned ImmMask = 1;
3600 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3601 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3604 } else if ((SplatBits & BitMask) != 0) {
3610 // Op=1, Cmode=1110.
3613 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3618 llvm_unreachable("unexpected size for isNEONModifiedImm");
3622 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3623 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3626 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3627 bool &ReverseVEXT, unsigned &Imm) {
3628 unsigned NumElts = VT.getVectorNumElements();
3629 ReverseVEXT = false;
3631 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3637 // If this is a VEXT shuffle, the immediate value is the index of the first
3638 // element. The other shuffle indices must be the successive elements after
3640 unsigned ExpectedElt = Imm;
3641 for (unsigned i = 1; i < NumElts; ++i) {
3642 // Increment the expected index. If it wraps around, it may still be
3643 // a VEXT but the source vectors must be swapped.
3645 if (ExpectedElt == NumElts * 2) {
3650 if (M[i] < 0) continue; // ignore UNDEF indices
3651 if (ExpectedElt != static_cast<unsigned>(M[i]))
3655 // Adjust the index value if the source operands will be swapped.
3662 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3663 /// instruction with the specified blocksize. (The order of the elements
3664 /// within each block of the vector is reversed.)
3665 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3666 unsigned BlockSize) {
3667 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3668 "Only possible block sizes for VREV are: 16, 32, 64");
3670 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3674 unsigned NumElts = VT.getVectorNumElements();
3675 unsigned BlockElts = M[0] + 1;
3676 // If the first shuffle index is UNDEF, be optimistic.
3678 BlockElts = BlockSize / EltSz;
3680 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3683 for (unsigned i = 0; i < NumElts; ++i) {
3684 if (M[i] < 0) continue; // ignore UNDEF indices
3685 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3692 static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3693 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3694 // range, then 0 is placed into the resulting vector. So pretty much any mask
3695 // of 8 elements can work here.
3696 return VT == MVT::v8i8 && M.size() == 8;
3699 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3700 unsigned &WhichResult) {
3701 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3705 unsigned NumElts = VT.getVectorNumElements();
3706 WhichResult = (M[0] == 0 ? 0 : 1);
3707 for (unsigned i = 0; i < NumElts; i += 2) {
3708 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3709 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3715 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3716 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3717 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3718 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3719 unsigned &WhichResult) {
3720 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3724 unsigned NumElts = VT.getVectorNumElements();
3725 WhichResult = (M[0] == 0 ? 0 : 1);
3726 for (unsigned i = 0; i < NumElts; i += 2) {
3727 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3728 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3734 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3735 unsigned &WhichResult) {
3736 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3740 unsigned NumElts = VT.getVectorNumElements();
3741 WhichResult = (M[0] == 0 ? 0 : 1);
3742 for (unsigned i = 0; i != NumElts; ++i) {
3743 if (M[i] < 0) continue; // ignore UNDEF indices
3744 if ((unsigned) M[i] != 2 * i + WhichResult)
3748 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3749 if (VT.is64BitVector() && EltSz == 32)
3755 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3756 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3757 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3758 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3759 unsigned &WhichResult) {
3760 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3764 unsigned Half = VT.getVectorNumElements() / 2;
3765 WhichResult = (M[0] == 0 ? 0 : 1);
3766 for (unsigned j = 0; j != 2; ++j) {
3767 unsigned Idx = WhichResult;
3768 for (unsigned i = 0; i != Half; ++i) {
3769 int MIdx = M[i + j * Half];
3770 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3776 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3777 if (VT.is64BitVector() && EltSz == 32)
3783 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3784 unsigned &WhichResult) {
3785 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3789 unsigned NumElts = VT.getVectorNumElements();
3790 WhichResult = (M[0] == 0 ? 0 : 1);
3791 unsigned Idx = WhichResult * NumElts / 2;
3792 for (unsigned i = 0; i != NumElts; i += 2) {
3793 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3794 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3799 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3800 if (VT.is64BitVector() && EltSz == 32)
3806 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3807 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3808 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3809 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3810 unsigned &WhichResult) {
3811 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3815 unsigned NumElts = VT.getVectorNumElements();
3816 WhichResult = (M[0] == 0 ? 0 : 1);
3817 unsigned Idx = WhichResult * NumElts / 2;
3818 for (unsigned i = 0; i != NumElts; i += 2) {
3819 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3820 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3825 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3826 if (VT.is64BitVector() && EltSz == 32)
3832 // If N is an integer constant that can be moved into a register in one
3833 // instruction, return an SDValue of such a constant (will become a MOV
3834 // instruction). Otherwise return null.
3835 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3836 const ARMSubtarget *ST, DebugLoc dl) {
3838 if (!isa<ConstantSDNode>(N))
3840 Val = cast<ConstantSDNode>(N)->getZExtValue();
3842 if (ST->isThumb1Only()) {
3843 if (Val <= 255 || ~Val <= 255)
3844 return DAG.getConstant(Val, MVT::i32);
3846 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3847 return DAG.getConstant(Val, MVT::i32);
3852 // If this is a case we can't handle, return null and let the default
3853 // expansion code take care of it.
3854 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3855 const ARMSubtarget *ST) const {
3856 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3857 DebugLoc dl = Op.getDebugLoc();
3858 EVT VT = Op.getValueType();
3860 APInt SplatBits, SplatUndef;
3861 unsigned SplatBitSize;
3863 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3864 if (SplatBitSize <= 64) {
3865 // Check if an immediate VMOV works.
3867 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3868 SplatUndef.getZExtValue(), SplatBitSize,
3869 DAG, VmovVT, VT.is128BitVector(),
3871 if (Val.getNode()) {
3872 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3873 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3876 // Try an immediate VMVN.
3877 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3878 ((1LL << SplatBitSize) - 1));
3879 Val = isNEONModifiedImm(NegatedImm,
3880 SplatUndef.getZExtValue(), SplatBitSize,
3881 DAG, VmovVT, VT.is128BitVector(),
3883 if (Val.getNode()) {
3884 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3885 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3890 // Scan through the operands to see if only one value is used.
3891 unsigned NumElts = VT.getVectorNumElements();
3892 bool isOnlyLowElement = true;
3893 bool usesOnlyOneValue = true;
3894 bool isConstant = true;
3896 for (unsigned i = 0; i < NumElts; ++i) {
3897 SDValue V = Op.getOperand(i);
3898 if (V.getOpcode() == ISD::UNDEF)
3901 isOnlyLowElement = false;
3902 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3905 if (!Value.getNode())
3907 else if (V != Value)
3908 usesOnlyOneValue = false;
3911 if (!Value.getNode())
3912 return DAG.getUNDEF(VT);
3914 if (isOnlyLowElement)
3915 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3917 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3919 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3920 // i32 and try again.
3921 if (usesOnlyOneValue && EltSize <= 32) {
3923 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3924 if (VT.getVectorElementType().isFloatingPoint()) {
3925 SmallVector<SDValue, 8> Ops;
3926 for (unsigned i = 0; i < NumElts; ++i)
3927 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3929 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3930 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3931 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3933 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3935 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3937 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3940 // If all elements are constants and the case above didn't get hit, fall back
3941 // to the default expansion, which will generate a load from the constant
3946 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3948 SDValue shuffle = ReconstructShuffle(Op, DAG);
3949 if (shuffle != SDValue())
3953 // Vectors with 32- or 64-bit elements can be built by directly assigning
3954 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3955 // will be legalized.
3956 if (EltSize >= 32) {
3957 // Do the expansion with floating-point types, since that is what the VFP
3958 // registers are defined to use, and since i64 is not legal.
3959 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3960 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3961 SmallVector<SDValue, 8> Ops;
3962 for (unsigned i = 0; i < NumElts; ++i)
3963 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3964 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3965 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3971 // Gather data to see if the operation can be modelled as a
3972 // shuffle in combination with VEXTs.
3973 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
3974 SelectionDAG &DAG) const {
3975 DebugLoc dl = Op.getDebugLoc();
3976 EVT VT = Op.getValueType();
3977 unsigned NumElts = VT.getVectorNumElements();
3979 SmallVector<SDValue, 2> SourceVecs;
3980 SmallVector<unsigned, 2> MinElts;
3981 SmallVector<unsigned, 2> MaxElts;
3983 for (unsigned i = 0; i < NumElts; ++i) {
3984 SDValue V = Op.getOperand(i);
3985 if (V.getOpcode() == ISD::UNDEF)
3987 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3988 // A shuffle can only come from building a vector from various
3989 // elements of other vectors.
3993 // Record this extraction against the appropriate vector if possible...
3994 SDValue SourceVec = V.getOperand(0);
3995 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3996 bool FoundSource = false;
3997 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3998 if (SourceVecs[j] == SourceVec) {
3999 if (MinElts[j] > EltNo)
4001 if (MaxElts[j] < EltNo)
4008 // Or record a new source if not...
4010 SourceVecs.push_back(SourceVec);
4011 MinElts.push_back(EltNo);
4012 MaxElts.push_back(EltNo);
4016 // Currently only do something sane when at most two source vectors
4018 if (SourceVecs.size() > 2)
4021 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4022 int VEXTOffsets[2] = {0, 0};
4024 // This loop extracts the usage patterns of the source vectors
4025 // and prepares appropriate SDValues for a shuffle if possible.
4026 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4027 if (SourceVecs[i].getValueType() == VT) {
4028 // No VEXT necessary
4029 ShuffleSrcs[i] = SourceVecs[i];
4032 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4033 // It probably isn't worth padding out a smaller vector just to
4034 // break it down again in a shuffle.
4038 // Since only 64-bit and 128-bit vectors are legal on ARM and
4039 // we've eliminated the other cases...
4040 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4041 "unexpected vector sizes in ReconstructShuffle");
4043 if (MaxElts[i] - MinElts[i] >= NumElts) {
4044 // Span too large for a VEXT to cope
4048 if (MinElts[i] >= NumElts) {
4049 // The extraction can just take the second half
4050 VEXTOffsets[i] = NumElts;
4051 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4053 DAG.getIntPtrConstant(NumElts));
4054 } else if (MaxElts[i] < NumElts) {
4055 // The extraction can just take the first half
4057 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4059 DAG.getIntPtrConstant(0));
4061 // An actual VEXT is needed
4062 VEXTOffsets[i] = MinElts[i];
4063 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4065 DAG.getIntPtrConstant(0));
4066 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4068 DAG.getIntPtrConstant(NumElts));
4069 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4070 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4074 SmallVector<int, 8> Mask;
4076 for (unsigned i = 0; i < NumElts; ++i) {
4077 SDValue Entry = Op.getOperand(i);
4078 if (Entry.getOpcode() == ISD::UNDEF) {
4083 SDValue ExtractVec = Entry.getOperand(0);
4084 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4085 .getOperand(1))->getSExtValue();
4086 if (ExtractVec == SourceVecs[0]) {
4087 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4089 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4093 // Final check before we try to produce nonsense...
4094 if (isShuffleMaskLegal(Mask, VT))
4095 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4101 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4102 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4103 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4104 /// are assumed to be legal.
4106 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4108 if (VT.getVectorNumElements() == 4 &&
4109 (VT.is128BitVector() || VT.is64BitVector())) {
4110 unsigned PFIndexes[4];
4111 for (unsigned i = 0; i != 4; ++i) {
4115 PFIndexes[i] = M[i];
4118 // Compute the index in the perfect shuffle table.
4119 unsigned PFTableIndex =
4120 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4121 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4122 unsigned Cost = (PFEntry >> 30);
4129 unsigned Imm, WhichResult;
4131 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4132 return (EltSize >= 32 ||
4133 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4134 isVREVMask(M, VT, 64) ||
4135 isVREVMask(M, VT, 32) ||
4136 isVREVMask(M, VT, 16) ||
4137 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4138 isVTBLMask(M, VT) ||
4139 isVTRNMask(M, VT, WhichResult) ||
4140 isVUZPMask(M, VT, WhichResult) ||
4141 isVZIPMask(M, VT, WhichResult) ||
4142 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4143 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4144 isVZIP_v_undef_Mask(M, VT, WhichResult));
4147 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4148 /// the specified operations to build the shuffle.
4149 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4150 SDValue RHS, SelectionDAG &DAG,
4152 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4153 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4154 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4157 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4166 OP_VUZPL, // VUZP, left result
4167 OP_VUZPR, // VUZP, right result
4168 OP_VZIPL, // VZIP, left result
4169 OP_VZIPR, // VZIP, right result
4170 OP_VTRNL, // VTRN, left result
4171 OP_VTRNR // VTRN, right result
4174 if (OpNum == OP_COPY) {
4175 if (LHSID == (1*9+2)*9+3) return LHS;
4176 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4180 SDValue OpLHS, OpRHS;
4181 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4182 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4183 EVT VT = OpLHS.getValueType();
4186 default: llvm_unreachable("Unknown shuffle opcode!");
4188 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4193 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4194 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4198 return DAG.getNode(ARMISD::VEXT, dl, VT,
4200 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4203 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4204 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4207 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4208 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4211 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4212 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4216 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4217 SmallVectorImpl<int> &ShuffleMask,
4218 SelectionDAG &DAG) {
4219 // Check to see if we can use the VTBL instruction.
4220 SDValue V1 = Op.getOperand(0);
4221 SDValue V2 = Op.getOperand(1);
4222 DebugLoc DL = Op.getDebugLoc();
4224 SmallVector<SDValue, 8> VTBLMask;
4225 for (SmallVectorImpl<int>::iterator
4226 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4227 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4229 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4230 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4231 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4234 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4235 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4239 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4240 SDValue V1 = Op.getOperand(0);
4241 SDValue V2 = Op.getOperand(1);
4242 DebugLoc dl = Op.getDebugLoc();
4243 EVT VT = Op.getValueType();
4244 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4245 SmallVector<int, 8> ShuffleMask;
4247 // Convert shuffles that are directly supported on NEON to target-specific
4248 // DAG nodes, instead of keeping them as shuffles and matching them again
4249 // during code selection. This is more efficient and avoids the possibility
4250 // of inconsistencies between legalization and selection.
4251 // FIXME: floating-point vectors should be canonicalized to integer vectors
4252 // of the same time so that they get CSEd properly.
4253 SVN->getMask(ShuffleMask);
4255 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4256 if (EltSize <= 32) {
4257 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4258 int Lane = SVN->getSplatIndex();
4259 // If this is undef splat, generate it via "just" vdup, if possible.
4260 if (Lane == -1) Lane = 0;
4262 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4263 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4265 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4266 DAG.getConstant(Lane, MVT::i32));
4271 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4274 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4275 DAG.getConstant(Imm, MVT::i32));
4278 if (isVREVMask(ShuffleMask, VT, 64))
4279 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4280 if (isVREVMask(ShuffleMask, VT, 32))
4281 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4282 if (isVREVMask(ShuffleMask, VT, 16))
4283 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4285 // Check for Neon shuffles that modify both input vectors in place.
4286 // If both results are used, i.e., if there are two shuffles with the same
4287 // source operands and with masks corresponding to both results of one of
4288 // these operations, DAG memoization will ensure that a single node is
4289 // used for both shuffles.
4290 unsigned WhichResult;
4291 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4292 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4293 V1, V2).getValue(WhichResult);
4294 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4295 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4296 V1, V2).getValue(WhichResult);
4297 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4298 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4299 V1, V2).getValue(WhichResult);
4301 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4302 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4303 V1, V1).getValue(WhichResult);
4304 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4305 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4306 V1, V1).getValue(WhichResult);
4307 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4308 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4309 V1, V1).getValue(WhichResult);
4312 // If the shuffle is not directly supported and it has 4 elements, use
4313 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4314 unsigned NumElts = VT.getVectorNumElements();
4316 unsigned PFIndexes[4];
4317 for (unsigned i = 0; i != 4; ++i) {
4318 if (ShuffleMask[i] < 0)
4321 PFIndexes[i] = ShuffleMask[i];
4324 // Compute the index in the perfect shuffle table.
4325 unsigned PFTableIndex =
4326 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4327 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4328 unsigned Cost = (PFEntry >> 30);
4331 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4334 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4335 if (EltSize >= 32) {
4336 // Do the expansion with floating-point types, since that is what the VFP
4337 // registers are defined to use, and since i64 is not legal.
4338 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4339 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4340 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4341 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4342 SmallVector<SDValue, 8> Ops;
4343 for (unsigned i = 0; i < NumElts; ++i) {
4344 if (ShuffleMask[i] < 0)
4345 Ops.push_back(DAG.getUNDEF(EltVT));
4347 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4348 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4349 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4352 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4353 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4356 if (VT == MVT::v8i8) {
4357 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4358 if (NewOp.getNode())
4365 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4366 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4367 SDValue Lane = Op.getOperand(1);
4368 if (!isa<ConstantSDNode>(Lane))
4371 SDValue Vec = Op.getOperand(0);
4372 if (Op.getValueType() == MVT::i32 &&
4373 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4374 DebugLoc dl = Op.getDebugLoc();
4375 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4381 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4382 // The only time a CONCAT_VECTORS operation can have legal types is when
4383 // two 64-bit vectors are concatenated to a 128-bit vector.
4384 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4385 "unexpected CONCAT_VECTORS");
4386 DebugLoc dl = Op.getDebugLoc();
4387 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4388 SDValue Op0 = Op.getOperand(0);
4389 SDValue Op1 = Op.getOperand(1);
4390 if (Op0.getOpcode() != ISD::UNDEF)
4391 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4392 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4393 DAG.getIntPtrConstant(0));
4394 if (Op1.getOpcode() != ISD::UNDEF)
4395 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4396 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4397 DAG.getIntPtrConstant(1));
4398 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4401 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4402 /// element has been zero/sign-extended, depending on the isSigned parameter,
4403 /// from an integer type half its size.
4404 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4406 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4407 EVT VT = N->getValueType(0);
4408 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4409 SDNode *BVN = N->getOperand(0).getNode();
4410 if (BVN->getValueType(0) != MVT::v4i32 ||
4411 BVN->getOpcode() != ISD::BUILD_VECTOR)
4413 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4414 unsigned HiElt = 1 - LoElt;
4415 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4416 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4417 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4418 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4419 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4422 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4423 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4426 if (Hi0->isNullValue() && Hi1->isNullValue())
4432 if (N->getOpcode() != ISD::BUILD_VECTOR)
4435 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4436 SDNode *Elt = N->getOperand(i).getNode();
4437 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4438 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4439 unsigned HalfSize = EltSize / 2;
4441 int64_t SExtVal = C->getSExtValue();
4442 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4445 if ((C->getZExtValue() >> HalfSize) != 0)
4456 /// isSignExtended - Check if a node is a vector value that is sign-extended
4457 /// or a constant BUILD_VECTOR with sign-extended elements.
4458 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4459 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4461 if (isExtendedBUILD_VECTOR(N, DAG, true))
4466 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4467 /// or a constant BUILD_VECTOR with zero-extended elements.
4468 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4469 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4471 if (isExtendedBUILD_VECTOR(N, DAG, false))
4476 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4477 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4478 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4479 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4480 return N->getOperand(0);
4481 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4482 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4483 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4484 LD->isNonTemporal(), LD->getAlignment());
4485 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4486 // have been legalized as a BITCAST from v4i32.
4487 if (N->getOpcode() == ISD::BITCAST) {
4488 SDNode *BVN = N->getOperand(0).getNode();
4489 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4490 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4491 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4492 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4493 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4495 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4496 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4497 EVT VT = N->getValueType(0);
4498 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4499 unsigned NumElts = VT.getVectorNumElements();
4500 MVT TruncVT = MVT::getIntegerVT(EltSize);
4501 SmallVector<SDValue, 8> Ops;
4502 for (unsigned i = 0; i != NumElts; ++i) {
4503 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4504 const APInt &CInt = C->getAPIntValue();
4505 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4507 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4508 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4511 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4512 unsigned Opcode = N->getOpcode();
4513 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4514 SDNode *N0 = N->getOperand(0).getNode();
4515 SDNode *N1 = N->getOperand(1).getNode();
4516 return N0->hasOneUse() && N1->hasOneUse() &&
4517 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4522 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4523 unsigned Opcode = N->getOpcode();
4524 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4525 SDNode *N0 = N->getOperand(0).getNode();
4526 SDNode *N1 = N->getOperand(1).getNode();
4527 return N0->hasOneUse() && N1->hasOneUse() &&
4528 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4533 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4534 // Multiplications are only custom-lowered for 128-bit vectors so that
4535 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4536 EVT VT = Op.getValueType();
4537 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4538 SDNode *N0 = Op.getOperand(0).getNode();
4539 SDNode *N1 = Op.getOperand(1).getNode();
4540 unsigned NewOpc = 0;
4542 bool isN0SExt = isSignExtended(N0, DAG);
4543 bool isN1SExt = isSignExtended(N1, DAG);
4544 if (isN0SExt && isN1SExt)
4545 NewOpc = ARMISD::VMULLs;
4547 bool isN0ZExt = isZeroExtended(N0, DAG);
4548 bool isN1ZExt = isZeroExtended(N1, DAG);
4549 if (isN0ZExt && isN1ZExt)
4550 NewOpc = ARMISD::VMULLu;
4551 else if (isN1SExt || isN1ZExt) {
4552 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4553 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4554 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4555 NewOpc = ARMISD::VMULLs;
4557 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4558 NewOpc = ARMISD::VMULLu;
4560 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4562 NewOpc = ARMISD::VMULLu;
4568 if (VT == MVT::v2i64)
4569 // Fall through to expand this. It is not legal.
4572 // Other vector multiplications are legal.
4577 // Legalize to a VMULL instruction.
4578 DebugLoc DL = Op.getDebugLoc();
4580 SDValue Op1 = SkipExtension(N1, DAG);
4582 Op0 = SkipExtension(N0, DAG);
4583 assert(Op0.getValueType().is64BitVector() &&
4584 Op1.getValueType().is64BitVector() &&
4585 "unexpected types for extended operands to VMULL");
4586 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4589 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4590 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4597 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4598 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4599 EVT Op1VT = Op1.getValueType();
4600 return DAG.getNode(N0->getOpcode(), DL, VT,
4601 DAG.getNode(NewOpc, DL, VT,
4602 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4603 DAG.getNode(NewOpc, DL, VT,
4604 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4608 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4610 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4611 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4612 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4613 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4614 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4615 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4616 // Get reciprocal estimate.
4617 // float4 recip = vrecpeq_f32(yf);
4618 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4619 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4620 // Because char has a smaller range than uchar, we can actually get away
4621 // without any newton steps. This requires that we use a weird bias
4622 // of 0xb000, however (again, this has been exhaustively tested).
4623 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4624 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4625 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4626 Y = DAG.getConstant(0xb000, MVT::i32);
4627 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4628 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4629 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4630 // Convert back to short.
4631 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4632 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4637 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4639 // Convert to float.
4640 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4641 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4642 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4643 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4644 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4645 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4647 // Use reciprocal estimate and one refinement step.
4648 // float4 recip = vrecpeq_f32(yf);
4649 // recip *= vrecpsq_f32(yf, recip);
4650 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4651 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4652 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4653 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4655 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4656 // Because short has a smaller range than ushort, we can actually get away
4657 // with only a single newton step. This requires that we use a weird bias
4658 // of 89, however (again, this has been exhaustively tested).
4659 // float4 result = as_float4(as_int4(xf*recip) + 89);
4660 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4661 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4662 N1 = DAG.getConstant(89, MVT::i32);
4663 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4664 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4665 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4666 // Convert back to integer and return.
4667 // return vmovn_s32(vcvt_s32_f32(result));
4668 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4669 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4673 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4674 EVT VT = Op.getValueType();
4675 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4676 "unexpected type for custom-lowering ISD::SDIV");
4678 DebugLoc dl = Op.getDebugLoc();
4679 SDValue N0 = Op.getOperand(0);
4680 SDValue N1 = Op.getOperand(1);
4683 if (VT == MVT::v8i8) {
4684 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4685 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4687 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4688 DAG.getIntPtrConstant(4));
4689 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4690 DAG.getIntPtrConstant(4));
4691 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4692 DAG.getIntPtrConstant(0));
4693 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4694 DAG.getIntPtrConstant(0));
4696 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4697 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4699 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4700 N0 = LowerCONCAT_VECTORS(N0, DAG);
4702 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4705 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4708 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4709 EVT VT = Op.getValueType();
4710 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4711 "unexpected type for custom-lowering ISD::UDIV");
4713 DebugLoc dl = Op.getDebugLoc();
4714 SDValue N0 = Op.getOperand(0);
4715 SDValue N1 = Op.getOperand(1);
4718 if (VT == MVT::v8i8) {
4719 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4720 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4722 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4723 DAG.getIntPtrConstant(4));
4724 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4725 DAG.getIntPtrConstant(4));
4726 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4727 DAG.getIntPtrConstant(0));
4728 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4729 DAG.getIntPtrConstant(0));
4731 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4732 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4734 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4735 N0 = LowerCONCAT_VECTORS(N0, DAG);
4737 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4738 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4743 // v4i16 sdiv ... Convert to float.
4744 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4745 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4746 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4747 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4748 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4749 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4751 // Use reciprocal estimate and two refinement steps.
4752 // float4 recip = vrecpeq_f32(yf);
4753 // recip *= vrecpsq_f32(yf, recip);
4754 // recip *= vrecpsq_f32(yf, recip);
4755 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4756 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4757 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4758 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4760 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4761 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4762 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4764 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4765 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4766 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4767 // and that it will never cause us to return an answer too large).
4768 // float4 result = as_float4(as_int4(xf*recip) + 89);
4769 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4770 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4771 N1 = DAG.getConstant(2, MVT::i32);
4772 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4773 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4774 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4775 // Convert back to integer and return.
4776 // return vmovn_u32(vcvt_s32_f32(result));
4777 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4778 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4782 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4783 switch (Op.getOpcode()) {
4784 default: llvm_unreachable("Don't know how to custom lower this!");
4785 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4786 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4787 case ISD::GlobalAddress:
4788 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4789 LowerGlobalAddressELF(Op, DAG);
4790 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4791 case ISD::SELECT: return LowerSELECT(Op, DAG);
4792 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4793 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4794 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4795 case ISD::VASTART: return LowerVASTART(Op, DAG);
4796 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4797 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4798 case ISD::SINT_TO_FP:
4799 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4800 case ISD::FP_TO_SINT:
4801 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4802 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4803 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4804 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4805 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4806 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4807 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4808 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4809 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4811 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4814 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4815 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4816 case ISD::SRL_PARTS:
4817 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4818 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4819 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4820 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4821 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4822 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4823 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4824 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4825 case ISD::MUL: return LowerMUL(Op, DAG);
4826 case ISD::SDIV: return LowerSDIV(Op, DAG);
4827 case ISD::UDIV: return LowerUDIV(Op, DAG);
4832 /// ReplaceNodeResults - Replace the results of node with an illegal result
4833 /// type with new values built out of custom code.
4834 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4835 SmallVectorImpl<SDValue>&Results,
4836 SelectionDAG &DAG) const {
4838 switch (N->getOpcode()) {
4840 llvm_unreachable("Don't know how to custom expand this!");
4843 Res = ExpandBITCAST(N, DAG);
4847 Res = Expand64BitShift(N, DAG, Subtarget);
4851 Results.push_back(Res);
4854 //===----------------------------------------------------------------------===//
4855 // ARM Scheduler Hooks
4856 //===----------------------------------------------------------------------===//
4859 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4860 MachineBasicBlock *BB,
4861 unsigned Size) const {
4862 unsigned dest = MI->getOperand(0).getReg();
4863 unsigned ptr = MI->getOperand(1).getReg();
4864 unsigned oldval = MI->getOperand(2).getReg();
4865 unsigned newval = MI->getOperand(3).getReg();
4866 unsigned scratch = BB->getParent()->getRegInfo()
4867 .createVirtualRegister(ARM::GPRRegisterClass);
4868 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4869 DebugLoc dl = MI->getDebugLoc();
4870 bool isThumb2 = Subtarget->isThumb2();
4872 unsigned ldrOpc, strOpc;
4874 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4876 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4877 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4880 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4881 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4884 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4885 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4889 MachineFunction *MF = BB->getParent();
4890 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4891 MachineFunction::iterator It = BB;
4892 ++It; // insert the new blocks after the current block
4894 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4895 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4896 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4897 MF->insert(It, loop1MBB);
4898 MF->insert(It, loop2MBB);
4899 MF->insert(It, exitMBB);
4901 // Transfer the remainder of BB and its successor edges to exitMBB.
4902 exitMBB->splice(exitMBB->begin(), BB,
4903 llvm::next(MachineBasicBlock::iterator(MI)),
4905 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4909 // fallthrough --> loop1MBB
4910 BB->addSuccessor(loop1MBB);
4913 // ldrex dest, [ptr]
4917 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4918 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4919 .addReg(dest).addReg(oldval));
4920 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4921 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4922 BB->addSuccessor(loop2MBB);
4923 BB->addSuccessor(exitMBB);
4926 // strex scratch, newval, [ptr]
4930 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4932 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4933 .addReg(scratch).addImm(0));
4934 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4935 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4936 BB->addSuccessor(loop1MBB);
4937 BB->addSuccessor(exitMBB);
4943 MI->eraseFromParent(); // The instruction is gone now.
4949 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4950 unsigned Size, unsigned BinOpcode) const {
4951 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4952 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4954 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4955 MachineFunction *MF = BB->getParent();
4956 MachineFunction::iterator It = BB;
4959 unsigned dest = MI->getOperand(0).getReg();
4960 unsigned ptr = MI->getOperand(1).getReg();
4961 unsigned incr = MI->getOperand(2).getReg();
4962 DebugLoc dl = MI->getDebugLoc();
4964 bool isThumb2 = Subtarget->isThumb2();
4965 unsigned ldrOpc, strOpc;
4967 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4969 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4970 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4973 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4974 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4977 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4978 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4982 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4983 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4984 MF->insert(It, loopMBB);
4985 MF->insert(It, exitMBB);
4987 // Transfer the remainder of BB and its successor edges to exitMBB.
4988 exitMBB->splice(exitMBB->begin(), BB,
4989 llvm::next(MachineBasicBlock::iterator(MI)),
4991 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4993 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4994 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4995 unsigned scratch2 = (!BinOpcode) ? incr :
4996 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5000 // fallthrough --> loopMBB
5001 BB->addSuccessor(loopMBB);
5005 // <binop> scratch2, dest, incr
5006 // strex scratch, scratch2, ptr
5009 // fallthrough --> exitMBB
5011 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5013 // operand order needs to go the other way for NAND
5014 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5015 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5016 addReg(incr).addReg(dest)).addReg(0);
5018 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5019 addReg(dest).addReg(incr)).addReg(0);
5022 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5024 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5025 .addReg(scratch).addImm(0));
5026 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5027 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5029 BB->addSuccessor(loopMBB);
5030 BB->addSuccessor(exitMBB);
5036 MI->eraseFromParent(); // The instruction is gone now.
5042 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5043 MachineBasicBlock *BB,
5046 ARMCC::CondCodes Cond) const {
5047 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5049 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5050 MachineFunction *MF = BB->getParent();
5051 MachineFunction::iterator It = BB;
5054 unsigned dest = MI->getOperand(0).getReg();
5055 unsigned ptr = MI->getOperand(1).getReg();
5056 unsigned incr = MI->getOperand(2).getReg();
5057 unsigned oldval = dest;
5058 DebugLoc dl = MI->getDebugLoc();
5060 bool isThumb2 = Subtarget->isThumb2();
5061 unsigned ldrOpc, strOpc, extendOpc;
5063 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5065 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5066 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5067 extendOpc = isThumb2 ? ARM::t2SXTBr : ARM::SXTBr;
5070 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5071 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5072 extendOpc = isThumb2 ? ARM::t2SXTHr : ARM::SXTHr;
5075 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5076 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5081 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5082 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5083 MF->insert(It, loopMBB);
5084 MF->insert(It, exitMBB);
5086 // Transfer the remainder of BB and its successor edges to exitMBB.
5087 exitMBB->splice(exitMBB->begin(), BB,
5088 llvm::next(MachineBasicBlock::iterator(MI)),
5090 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5092 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5093 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5094 unsigned scratch2 = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5098 // fallthrough --> loopMBB
5099 BB->addSuccessor(loopMBB);
5103 // (sign extend dest, if required)
5105 // cmov.cond scratch2, dest, incr
5106 // strex scratch, scratch2, ptr
5109 // fallthrough --> exitMBB
5111 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5113 // Sign extend the value, if necessary.
5114 if (signExtend && extendOpc) {
5115 oldval = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
5116 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval).addReg(dest));
5119 // Build compare and cmov instructions.
5120 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5121 .addReg(oldval).addReg(incr));
5122 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5123 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5125 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5127 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5128 .addReg(scratch).addImm(0));
5129 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5130 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5132 BB->addSuccessor(loopMBB);
5133 BB->addSuccessor(exitMBB);
5139 MI->eraseFromParent(); // The instruction is gone now.
5145 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
5146 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
5147 E = MBB->succ_end(); I != E; ++I)
5150 llvm_unreachable("Expecting a BB with two successors!");
5153 // FIXME: This opcode table should obviously be expressed in the target
5154 // description. We probably just need a "machine opcode" value in the pseudo
5155 // instruction. But the ideal solution maybe to simply remove the "S" version
5156 // of the opcode altogether.
5157 struct AddSubFlagsOpcodePair {
5159 unsigned MachineOpc;
5162 static AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
5163 {ARM::ADCSri, ARM::ADCri},
5164 {ARM::ADCSrr, ARM::ADCrr},
5165 {ARM::ADCSrs, ARM::ADCrs},
5166 {ARM::SBCSri, ARM::SBCri},
5167 {ARM::SBCSrr, ARM::SBCrr},
5168 {ARM::SBCSrs, ARM::SBCrs},
5169 {ARM::RSBSri, ARM::RSBri},
5170 {ARM::RSBSrr, ARM::RSBrr},
5171 {ARM::RSBSrs, ARM::RSBrs},
5172 {ARM::RSCSri, ARM::RSCri},
5173 {ARM::RSCSrs, ARM::RSCrs},
5174 {ARM::t2ADCSri, ARM::t2ADCri},
5175 {ARM::t2ADCSrr, ARM::t2ADCrr},
5176 {ARM::t2ADCSrs, ARM::t2ADCrs},
5177 {ARM::t2SBCSri, ARM::t2SBCri},
5178 {ARM::t2SBCSrr, ARM::t2SBCrr},
5179 {ARM::t2SBCSrs, ARM::t2SBCrs},
5180 {ARM::t2RSBSri, ARM::t2RSBri},
5181 {ARM::t2RSBSrs, ARM::t2RSBrs},
5184 // Convert and Add or Subtract with Carry and Flags to a generic opcode with
5185 // CPSR<def> operand. e.g. ADCS (...) -> ADC (... CPSR<def>).
5187 // FIXME: Somewhere we should assert that CPSR<def> is in the correct
5188 // position to be recognized by the target descrition as the 'S' bit.
5189 bool ARMTargetLowering::RemapAddSubWithFlags(MachineInstr *MI,
5190 MachineBasicBlock *BB) const {
5191 unsigned OldOpc = MI->getOpcode();
5192 unsigned NewOpc = 0;
5194 // This is only called for instructions that need remapping, so iterating over
5195 // the tiny opcode table is not costly.
5196 static const int NPairs =
5197 sizeof(AddSubFlagsOpcodeMap) / sizeof(AddSubFlagsOpcodePair);
5198 for (AddSubFlagsOpcodePair *Pair = &AddSubFlagsOpcodeMap[0],
5199 *End = &AddSubFlagsOpcodeMap[NPairs]; Pair != End; ++Pair) {
5200 if (OldOpc == Pair->PseudoOpc) {
5201 NewOpc = Pair->MachineOpc;
5208 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5209 DebugLoc dl = MI->getDebugLoc();
5210 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
5211 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
5212 MIB.addOperand(MI->getOperand(i));
5213 AddDefaultPred(MIB);
5214 MIB.addReg(ARM::CPSR, RegState::Define); // S bit
5215 MI->eraseFromParent();
5220 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5221 MachineBasicBlock *BB) const {
5222 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5223 DebugLoc dl = MI->getDebugLoc();
5224 bool isThumb2 = Subtarget->isThumb2();
5225 switch (MI->getOpcode()) {
5227 if (RemapAddSubWithFlags(MI, BB))
5231 llvm_unreachable("Unexpected instr type to insert");
5233 case ARM::ATOMIC_LOAD_ADD_I8:
5234 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5235 case ARM::ATOMIC_LOAD_ADD_I16:
5236 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5237 case ARM::ATOMIC_LOAD_ADD_I32:
5238 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5240 case ARM::ATOMIC_LOAD_AND_I8:
5241 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5242 case ARM::ATOMIC_LOAD_AND_I16:
5243 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5244 case ARM::ATOMIC_LOAD_AND_I32:
5245 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5247 case ARM::ATOMIC_LOAD_OR_I8:
5248 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5249 case ARM::ATOMIC_LOAD_OR_I16:
5250 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5251 case ARM::ATOMIC_LOAD_OR_I32:
5252 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5254 case ARM::ATOMIC_LOAD_XOR_I8:
5255 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5256 case ARM::ATOMIC_LOAD_XOR_I16:
5257 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5258 case ARM::ATOMIC_LOAD_XOR_I32:
5259 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5261 case ARM::ATOMIC_LOAD_NAND_I8:
5262 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5263 case ARM::ATOMIC_LOAD_NAND_I16:
5264 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5265 case ARM::ATOMIC_LOAD_NAND_I32:
5266 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5268 case ARM::ATOMIC_LOAD_SUB_I8:
5269 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5270 case ARM::ATOMIC_LOAD_SUB_I16:
5271 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5272 case ARM::ATOMIC_LOAD_SUB_I32:
5273 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5275 case ARM::ATOMIC_LOAD_MIN_I8:
5276 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
5277 case ARM::ATOMIC_LOAD_MIN_I16:
5278 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
5279 case ARM::ATOMIC_LOAD_MIN_I32:
5280 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
5282 case ARM::ATOMIC_LOAD_MAX_I8:
5283 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
5284 case ARM::ATOMIC_LOAD_MAX_I16:
5285 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
5286 case ARM::ATOMIC_LOAD_MAX_I32:
5287 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
5289 case ARM::ATOMIC_LOAD_UMIN_I8:
5290 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
5291 case ARM::ATOMIC_LOAD_UMIN_I16:
5292 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
5293 case ARM::ATOMIC_LOAD_UMIN_I32:
5294 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
5296 case ARM::ATOMIC_LOAD_UMAX_I8:
5297 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
5298 case ARM::ATOMIC_LOAD_UMAX_I16:
5299 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
5300 case ARM::ATOMIC_LOAD_UMAX_I32:
5301 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
5303 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
5304 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
5305 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
5307 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
5308 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
5309 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
5311 case ARM::tMOVCCr_pseudo: {
5312 // To "insert" a SELECT_CC instruction, we actually have to insert the
5313 // diamond control-flow pattern. The incoming instruction knows the
5314 // destination vreg to set, the condition code register to branch on, the
5315 // true/false values to select between, and a branch opcode to use.
5316 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5317 MachineFunction::iterator It = BB;
5323 // cmpTY ccX, r1, r2
5325 // fallthrough --> copy0MBB
5326 MachineBasicBlock *thisMBB = BB;
5327 MachineFunction *F = BB->getParent();
5328 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5329 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5330 F->insert(It, copy0MBB);
5331 F->insert(It, sinkMBB);
5333 // Transfer the remainder of BB and its successor edges to sinkMBB.
5334 sinkMBB->splice(sinkMBB->begin(), BB,
5335 llvm::next(MachineBasicBlock::iterator(MI)),
5337 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5339 BB->addSuccessor(copy0MBB);
5340 BB->addSuccessor(sinkMBB);
5342 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
5343 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
5346 // %FalseValue = ...
5347 // # fallthrough to sinkMBB
5350 // Update machine-CFG edges
5351 BB->addSuccessor(sinkMBB);
5354 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5357 BuildMI(*BB, BB->begin(), dl,
5358 TII->get(ARM::PHI), MI->getOperand(0).getReg())
5359 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5360 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5362 MI->eraseFromParent(); // The pseudo instruction is gone now.
5367 case ARM::BCCZi64: {
5368 // If there is an unconditional branch to the other successor, remove it.
5369 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
5371 // Compare both parts that make up the double comparison separately for
5373 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
5375 unsigned LHS1 = MI->getOperand(1).getReg();
5376 unsigned LHS2 = MI->getOperand(2).getReg();
5378 AddDefaultPred(BuildMI(BB, dl,
5379 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5380 .addReg(LHS1).addImm(0));
5381 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5382 .addReg(LHS2).addImm(0)
5383 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5385 unsigned RHS1 = MI->getOperand(3).getReg();
5386 unsigned RHS2 = MI->getOperand(4).getReg();
5387 AddDefaultPred(BuildMI(BB, dl,
5388 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5389 .addReg(LHS1).addReg(RHS1));
5390 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5391 .addReg(LHS2).addReg(RHS2)
5392 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5395 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
5396 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
5397 if (MI->getOperand(0).getImm() == ARMCC::NE)
5398 std::swap(destMBB, exitMBB);
5400 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5401 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
5402 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
5405 MI->eraseFromParent(); // The pseudo instruction is gone now.
5411 //===----------------------------------------------------------------------===//
5412 // ARM Optimization Hooks
5413 //===----------------------------------------------------------------------===//
5416 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
5417 TargetLowering::DAGCombinerInfo &DCI) {
5418 SelectionDAG &DAG = DCI.DAG;
5419 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5420 EVT VT = N->getValueType(0);
5421 unsigned Opc = N->getOpcode();
5422 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
5423 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
5424 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
5425 ISD::CondCode CC = ISD::SETCC_INVALID;
5428 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
5430 SDValue CCOp = Slct.getOperand(0);
5431 if (CCOp.getOpcode() == ISD::SETCC)
5432 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
5435 bool DoXform = false;
5437 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
5440 if (LHS.getOpcode() == ISD::Constant &&
5441 cast<ConstantSDNode>(LHS)->isNullValue()) {
5443 } else if (CC != ISD::SETCC_INVALID &&
5444 RHS.getOpcode() == ISD::Constant &&
5445 cast<ConstantSDNode>(RHS)->isNullValue()) {
5446 std::swap(LHS, RHS);
5447 SDValue Op0 = Slct.getOperand(0);
5448 EVT OpVT = isSlctCC ? Op0.getValueType() :
5449 Op0.getOperand(0).getValueType();
5450 bool isInt = OpVT.isInteger();
5451 CC = ISD::getSetCCInverse(CC, isInt);
5453 if (!TLI.isCondCodeLegal(CC, OpVT))
5454 return SDValue(); // Inverse operator isn't legal.
5461 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
5463 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
5464 Slct.getOperand(0), Slct.getOperand(1), CC);
5465 SDValue CCOp = Slct.getOperand(0);
5467 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
5468 CCOp.getOperand(0), CCOp.getOperand(1), CC);
5469 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
5470 CCOp, OtherOp, Result);
5475 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
5476 /// operands N0 and N1. This is a helper for PerformADDCombine that is
5477 /// called with the default operands, and if that fails, with commuted
5479 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
5480 TargetLowering::DAGCombinerInfo &DCI) {
5481 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
5482 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
5483 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
5484 if (Result.getNode()) return Result;
5489 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
5491 static SDValue PerformADDCombine(SDNode *N,
5492 TargetLowering::DAGCombinerInfo &DCI) {
5493 SDValue N0 = N->getOperand(0);
5494 SDValue N1 = N->getOperand(1);
5496 // First try with the default operand order.
5497 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
5498 if (Result.getNode())
5501 // If that didn't work, try again with the operands commuted.
5502 return PerformADDCombineWithOperands(N, N1, N0, DCI);
5505 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
5507 static SDValue PerformSUBCombine(SDNode *N,
5508 TargetLowering::DAGCombinerInfo &DCI) {
5509 SDValue N0 = N->getOperand(0);
5510 SDValue N1 = N->getOperand(1);
5512 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
5513 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
5514 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
5515 if (Result.getNode()) return Result;
5521 /// PerformVMULCombine
5522 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
5523 /// special multiplier accumulator forwarding.
5529 static SDValue PerformVMULCombine(SDNode *N,
5530 TargetLowering::DAGCombinerInfo &DCI,
5531 const ARMSubtarget *Subtarget) {
5532 if (!Subtarget->hasVMLxForwarding())
5535 SelectionDAG &DAG = DCI.DAG;
5536 SDValue N0 = N->getOperand(0);
5537 SDValue N1 = N->getOperand(1);
5538 unsigned Opcode = N0.getOpcode();
5539 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5540 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
5541 Opcode = N0.getOpcode();
5542 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5543 Opcode != ISD::FADD && Opcode != ISD::FSUB)
5548 EVT VT = N->getValueType(0);
5549 DebugLoc DL = N->getDebugLoc();
5550 SDValue N00 = N0->getOperand(0);
5551 SDValue N01 = N0->getOperand(1);
5552 return DAG.getNode(Opcode, DL, VT,
5553 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
5554 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
5557 static SDValue PerformMULCombine(SDNode *N,
5558 TargetLowering::DAGCombinerInfo &DCI,
5559 const ARMSubtarget *Subtarget) {
5560 SelectionDAG &DAG = DCI.DAG;
5562 if (Subtarget->isThumb1Only())
5565 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5568 EVT VT = N->getValueType(0);
5569 if (VT.is64BitVector() || VT.is128BitVector())
5570 return PerformVMULCombine(N, DCI, Subtarget);
5574 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
5578 uint64_t MulAmt = C->getZExtValue();
5579 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
5580 ShiftAmt = ShiftAmt & (32 - 1);
5581 SDValue V = N->getOperand(0);
5582 DebugLoc DL = N->getDebugLoc();
5585 MulAmt >>= ShiftAmt;
5586 if (isPowerOf2_32(MulAmt - 1)) {
5587 // (mul x, 2^N + 1) => (add (shl x, N), x)
5588 Res = DAG.getNode(ISD::ADD, DL, VT,
5589 V, DAG.getNode(ISD::SHL, DL, VT,
5590 V, DAG.getConstant(Log2_32(MulAmt-1),
5592 } else if (isPowerOf2_32(MulAmt + 1)) {
5593 // (mul x, 2^N - 1) => (sub (shl x, N), x)
5594 Res = DAG.getNode(ISD::SUB, DL, VT,
5595 DAG.getNode(ISD::SHL, DL, VT,
5596 V, DAG.getConstant(Log2_32(MulAmt+1),
5603 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
5604 DAG.getConstant(ShiftAmt, MVT::i32));
5606 // Do not add new nodes to DAG combiner worklist.
5607 DCI.CombineTo(N, Res, false);
5611 static SDValue PerformANDCombine(SDNode *N,
5612 TargetLowering::DAGCombinerInfo &DCI) {
5614 // Attempt to use immediate-form VBIC
5615 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5616 DebugLoc dl = N->getDebugLoc();
5617 EVT VT = N->getValueType(0);
5618 SelectionDAG &DAG = DCI.DAG;
5620 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5623 APInt SplatBits, SplatUndef;
5624 unsigned SplatBitSize;
5627 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5628 if (SplatBitSize <= 64) {
5630 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
5631 SplatUndef.getZExtValue(), SplatBitSize,
5632 DAG, VbicVT, VT.is128BitVector(),
5634 if (Val.getNode()) {
5636 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
5637 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
5638 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
5646 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
5647 static SDValue PerformORCombine(SDNode *N,
5648 TargetLowering::DAGCombinerInfo &DCI,
5649 const ARMSubtarget *Subtarget) {
5650 // Attempt to use immediate-form VORR
5651 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5652 DebugLoc dl = N->getDebugLoc();
5653 EVT VT = N->getValueType(0);
5654 SelectionDAG &DAG = DCI.DAG;
5656 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5659 APInt SplatBits, SplatUndef;
5660 unsigned SplatBitSize;
5662 if (BVN && Subtarget->hasNEON() &&
5663 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5664 if (SplatBitSize <= 64) {
5666 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5667 SplatUndef.getZExtValue(), SplatBitSize,
5668 DAG, VorrVT, VT.is128BitVector(),
5670 if (Val.getNode()) {
5672 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
5673 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
5674 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
5679 SDValue N0 = N->getOperand(0);
5680 if (N0.getOpcode() != ISD::AND)
5682 SDValue N1 = N->getOperand(1);
5684 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
5685 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
5686 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
5688 unsigned SplatBitSize;
5691 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
5693 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
5694 HasAnyUndefs) && !HasAnyUndefs) {
5695 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
5697 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
5698 HasAnyUndefs) && !HasAnyUndefs &&
5699 SplatBits0 == ~SplatBits1) {
5700 // Canonicalize the vector type to make instruction selection simpler.
5701 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
5702 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
5703 N0->getOperand(1), N0->getOperand(0),
5705 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
5710 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
5713 // BFI is only available on V6T2+
5714 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
5717 DebugLoc DL = N->getDebugLoc();
5718 // 1) or (and A, mask), val => ARMbfi A, val, mask
5719 // iff (val & mask) == val
5721 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5722 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
5723 // && mask == ~mask2
5724 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
5725 // && ~mask == mask2
5726 // (i.e., copy a bitfield value into another bitfield of the same width)
5731 SDValue N00 = N0.getOperand(0);
5733 // The value and the mask need to be constants so we can verify this is
5734 // actually a bitfield set. If the mask is 0xffff, we can do better
5735 // via a movt instruction, so don't use BFI in that case.
5736 SDValue MaskOp = N0.getOperand(1);
5737 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
5740 unsigned Mask = MaskC->getZExtValue();
5744 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
5745 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
5747 unsigned Val = N1C->getZExtValue();
5748 if ((Val & ~Mask) != Val)
5751 if (ARM::isBitFieldInvertedMask(Mask)) {
5752 Val >>= CountTrailingZeros_32(~Mask);
5754 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
5755 DAG.getConstant(Val, MVT::i32),
5756 DAG.getConstant(Mask, MVT::i32));
5758 // Do not add new nodes to DAG combiner worklist.
5759 DCI.CombineTo(N, Res, false);
5762 } else if (N1.getOpcode() == ISD::AND) {
5763 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5764 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5767 unsigned Mask2 = N11C->getZExtValue();
5769 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
5771 if (ARM::isBitFieldInvertedMask(Mask) &&
5773 // The pack halfword instruction works better for masks that fit it,
5774 // so use that when it's available.
5775 if (Subtarget->hasT2ExtractPack() &&
5776 (Mask == 0xffff || Mask == 0xffff0000))
5779 unsigned amt = CountTrailingZeros_32(Mask2);
5780 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
5781 DAG.getConstant(amt, MVT::i32));
5782 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
5783 DAG.getConstant(Mask, MVT::i32));
5784 // Do not add new nodes to DAG combiner worklist.
5785 DCI.CombineTo(N, Res, false);
5787 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
5789 // The pack halfword instruction works better for masks that fit it,
5790 // so use that when it's available.
5791 if (Subtarget->hasT2ExtractPack() &&
5792 (Mask2 == 0xffff || Mask2 == 0xffff0000))
5795 unsigned lsb = CountTrailingZeros_32(Mask);
5796 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
5797 DAG.getConstant(lsb, MVT::i32));
5798 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
5799 DAG.getConstant(Mask2, MVT::i32));
5800 // Do not add new nodes to DAG combiner worklist.
5801 DCI.CombineTo(N, Res, false);
5806 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
5807 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
5808 ARM::isBitFieldInvertedMask(~Mask)) {
5809 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
5810 // where lsb(mask) == #shamt and masked bits of B are known zero.
5811 SDValue ShAmt = N00.getOperand(1);
5812 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5813 unsigned LSB = CountTrailingZeros_32(Mask);
5817 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
5818 DAG.getConstant(~Mask, MVT::i32));
5820 // Do not add new nodes to DAG combiner worklist.
5821 DCI.CombineTo(N, Res, false);
5827 /// PerformBFICombine - (bfi A, (and B, C1), C2) -> (bfi A, B, C2) iff
5829 static SDValue PerformBFICombine(SDNode *N,
5830 TargetLowering::DAGCombinerInfo &DCI) {
5831 SDValue N1 = N->getOperand(1);
5832 if (N1.getOpcode() == ISD::AND) {
5833 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5836 unsigned Mask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
5837 unsigned Mask2 = N11C->getZExtValue();
5838 if ((Mask & Mask2) == Mask2)
5839 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
5840 N->getOperand(0), N1.getOperand(0),
5846 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
5847 /// ARMISD::VMOVRRD.
5848 static SDValue PerformVMOVRRDCombine(SDNode *N,
5849 TargetLowering::DAGCombinerInfo &DCI) {
5850 // vmovrrd(vmovdrr x, y) -> x,y
5851 SDValue InDouble = N->getOperand(0);
5852 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
5853 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
5855 // vmovrrd(load f64) -> (load i32), (load i32)
5856 SDNode *InNode = InDouble.getNode();
5857 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
5858 InNode->getValueType(0) == MVT::f64 &&
5859 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
5860 !cast<LoadSDNode>(InNode)->isVolatile()) {
5861 // TODO: Should this be done for non-FrameIndex operands?
5862 LoadSDNode *LD = cast<LoadSDNode>(InNode);
5864 SelectionDAG &DAG = DCI.DAG;
5865 DebugLoc DL = LD->getDebugLoc();
5866 SDValue BasePtr = LD->getBasePtr();
5867 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
5868 LD->getPointerInfo(), LD->isVolatile(),
5869 LD->isNonTemporal(), LD->getAlignment());
5871 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
5872 DAG.getConstant(4, MVT::i32));
5873 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
5874 LD->getPointerInfo(), LD->isVolatile(),
5875 LD->isNonTemporal(),
5876 std::min(4U, LD->getAlignment() / 2));
5878 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
5879 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
5880 DCI.RemoveFromWorklist(LD);
5888 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
5889 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
5890 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
5891 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
5892 SDValue Op0 = N->getOperand(0);
5893 SDValue Op1 = N->getOperand(1);
5894 if (Op0.getOpcode() == ISD::BITCAST)
5895 Op0 = Op0.getOperand(0);
5896 if (Op1.getOpcode() == ISD::BITCAST)
5897 Op1 = Op1.getOperand(0);
5898 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
5899 Op0.getNode() == Op1.getNode() &&
5900 Op0.getResNo() == 0 && Op1.getResNo() == 1)
5901 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5902 N->getValueType(0), Op0.getOperand(0));
5906 /// PerformSTORECombine - Target-specific dag combine xforms for
5908 static SDValue PerformSTORECombine(SDNode *N,
5909 TargetLowering::DAGCombinerInfo &DCI) {
5910 // Bitcast an i64 store extracted from a vector to f64.
5911 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5912 StoreSDNode *St = cast<StoreSDNode>(N);
5913 SDValue StVal = St->getValue();
5914 if (!ISD::isNormalStore(St) || St->isVolatile())
5917 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
5918 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
5919 SelectionDAG &DAG = DCI.DAG;
5920 DebugLoc DL = St->getDebugLoc();
5921 SDValue BasePtr = St->getBasePtr();
5922 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
5923 StVal.getNode()->getOperand(0), BasePtr,
5924 St->getPointerInfo(), St->isVolatile(),
5925 St->isNonTemporal(), St->getAlignment());
5927 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
5928 DAG.getConstant(4, MVT::i32));
5929 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
5930 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
5931 St->isNonTemporal(),
5932 std::min(4U, St->getAlignment() / 2));
5935 if (StVal.getValueType() != MVT::i64 ||
5936 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5939 SelectionDAG &DAG = DCI.DAG;
5940 DebugLoc dl = StVal.getDebugLoc();
5941 SDValue IntVec = StVal.getOperand(0);
5942 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5943 IntVec.getValueType().getVectorNumElements());
5944 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
5945 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5946 Vec, StVal.getOperand(1));
5947 dl = N->getDebugLoc();
5948 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
5949 // Make the DAGCombiner fold the bitcasts.
5950 DCI.AddToWorklist(Vec.getNode());
5951 DCI.AddToWorklist(ExtElt.getNode());
5952 DCI.AddToWorklist(V.getNode());
5953 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
5954 St->getPointerInfo(), St->isVolatile(),
5955 St->isNonTemporal(), St->getAlignment(),
5959 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
5960 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
5961 /// i64 vector to have f64 elements, since the value can then be loaded
5962 /// directly into a VFP register.
5963 static bool hasNormalLoadOperand(SDNode *N) {
5964 unsigned NumElts = N->getValueType(0).getVectorNumElements();
5965 for (unsigned i = 0; i < NumElts; ++i) {
5966 SDNode *Elt = N->getOperand(i).getNode();
5967 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
5973 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
5974 /// ISD::BUILD_VECTOR.
5975 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
5976 TargetLowering::DAGCombinerInfo &DCI){
5977 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
5978 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
5979 // into a pair of GPRs, which is fine when the value is used as a scalar,
5980 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
5981 SelectionDAG &DAG = DCI.DAG;
5982 if (N->getNumOperands() == 2) {
5983 SDValue RV = PerformVMOVDRRCombine(N, DAG);
5988 // Load i64 elements as f64 values so that type legalization does not split
5989 // them up into i32 values.
5990 EVT VT = N->getValueType(0);
5991 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
5993 DebugLoc dl = N->getDebugLoc();
5994 SmallVector<SDValue, 8> Ops;
5995 unsigned NumElts = VT.getVectorNumElements();
5996 for (unsigned i = 0; i < NumElts; ++i) {
5997 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
5999 // Make the DAGCombiner fold the bitcast.
6000 DCI.AddToWorklist(V.getNode());
6002 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
6003 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
6004 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
6007 /// PerformInsertEltCombine - Target-specific dag combine xforms for
6008 /// ISD::INSERT_VECTOR_ELT.
6009 static SDValue PerformInsertEltCombine(SDNode *N,
6010 TargetLowering::DAGCombinerInfo &DCI) {
6011 // Bitcast an i64 load inserted into a vector to f64.
6012 // Otherwise, the i64 value will be legalized to a pair of i32 values.
6013 EVT VT = N->getValueType(0);
6014 SDNode *Elt = N->getOperand(1).getNode();
6015 if (VT.getVectorElementType() != MVT::i64 ||
6016 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
6019 SelectionDAG &DAG = DCI.DAG;
6020 DebugLoc dl = N->getDebugLoc();
6021 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
6022 VT.getVectorNumElements());
6023 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
6024 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
6025 // Make the DAGCombiner fold the bitcasts.
6026 DCI.AddToWorklist(Vec.getNode());
6027 DCI.AddToWorklist(V.getNode());
6028 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
6029 Vec, V, N->getOperand(2));
6030 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
6033 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
6034 /// ISD::VECTOR_SHUFFLE.
6035 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
6036 // The LLVM shufflevector instruction does not require the shuffle mask
6037 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
6038 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
6039 // operands do not match the mask length, they are extended by concatenating
6040 // them with undef vectors. That is probably the right thing for other
6041 // targets, but for NEON it is better to concatenate two double-register
6042 // size vector operands into a single quad-register size vector. Do that
6043 // transformation here:
6044 // shuffle(concat(v1, undef), concat(v2, undef)) ->
6045 // shuffle(concat(v1, v2), undef)
6046 SDValue Op0 = N->getOperand(0);
6047 SDValue Op1 = N->getOperand(1);
6048 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
6049 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
6050 Op0.getNumOperands() != 2 ||
6051 Op1.getNumOperands() != 2)
6053 SDValue Concat0Op1 = Op0.getOperand(1);
6054 SDValue Concat1Op1 = Op1.getOperand(1);
6055 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
6056 Concat1Op1.getOpcode() != ISD::UNDEF)
6058 // Skip the transformation if any of the types are illegal.
6059 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6060 EVT VT = N->getValueType(0);
6061 if (!TLI.isTypeLegal(VT) ||
6062 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
6063 !TLI.isTypeLegal(Concat1Op1.getValueType()))
6066 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
6067 Op0.getOperand(0), Op1.getOperand(0));
6068 // Translate the shuffle mask.
6069 SmallVector<int, 16> NewMask;
6070 unsigned NumElts = VT.getVectorNumElements();
6071 unsigned HalfElts = NumElts/2;
6072 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
6073 for (unsigned n = 0; n < NumElts; ++n) {
6074 int MaskElt = SVN->getMaskElt(n);
6076 if (MaskElt < (int)HalfElts)
6078 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
6079 NewElt = HalfElts + MaskElt - NumElts;
6080 NewMask.push_back(NewElt);
6082 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
6083 DAG.getUNDEF(VT), NewMask.data());
6086 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
6087 /// NEON load/store intrinsics to merge base address updates.
6088 static SDValue CombineBaseUpdate(SDNode *N,
6089 TargetLowering::DAGCombinerInfo &DCI) {
6090 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6093 SelectionDAG &DAG = DCI.DAG;
6094 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
6095 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
6096 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
6097 SDValue Addr = N->getOperand(AddrOpIdx);
6099 // Search for a use of the address operand that is an increment.
6100 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
6101 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
6103 if (User->getOpcode() != ISD::ADD ||
6104 UI.getUse().getResNo() != Addr.getResNo())
6107 // Check that the add is independent of the load/store. Otherwise, folding
6108 // it would create a cycle.
6109 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
6112 // Find the new opcode for the updating load/store.
6114 bool isLaneOp = false;
6115 unsigned NewOpc = 0;
6116 unsigned NumVecs = 0;
6118 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
6120 default: assert(0 && "unexpected intrinsic for Neon base update");
6121 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
6123 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
6125 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
6127 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
6129 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
6130 NumVecs = 2; isLaneOp = true; break;
6131 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
6132 NumVecs = 3; isLaneOp = true; break;
6133 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
6134 NumVecs = 4; isLaneOp = true; break;
6135 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
6136 NumVecs = 1; isLoad = false; break;
6137 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
6138 NumVecs = 2; isLoad = false; break;
6139 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
6140 NumVecs = 3; isLoad = false; break;
6141 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
6142 NumVecs = 4; isLoad = false; break;
6143 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
6144 NumVecs = 2; isLoad = false; isLaneOp = true; break;
6145 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
6146 NumVecs = 3; isLoad = false; isLaneOp = true; break;
6147 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
6148 NumVecs = 4; isLoad = false; isLaneOp = true; break;
6152 switch (N->getOpcode()) {
6153 default: assert(0 && "unexpected opcode for Neon base update");
6154 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
6155 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
6156 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
6160 // Find the size of memory referenced by the load/store.
6163 VecTy = N->getValueType(0);
6165 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
6166 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
6168 NumBytes /= VecTy.getVectorNumElements();
6170 // If the increment is a constant, it must match the memory ref size.
6171 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
6172 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
6173 uint64_t IncVal = CInc->getZExtValue();
6174 if (IncVal != NumBytes)
6176 } else if (NumBytes >= 3 * 16) {
6177 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
6178 // separate instructions that make it harder to use a non-constant update.
6182 // Create the new updating load/store node.
6184 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
6186 for (n = 0; n < NumResultVecs; ++n)
6188 Tys[n++] = MVT::i32;
6189 Tys[n] = MVT::Other;
6190 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
6191 SmallVector<SDValue, 8> Ops;
6192 Ops.push_back(N->getOperand(0)); // incoming chain
6193 Ops.push_back(N->getOperand(AddrOpIdx));
6195 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
6196 Ops.push_back(N->getOperand(i));
6198 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
6199 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
6200 Ops.data(), Ops.size(),
6201 MemInt->getMemoryVT(),
6202 MemInt->getMemOperand());
6205 std::vector<SDValue> NewResults;
6206 for (unsigned i = 0; i < NumResultVecs; ++i) {
6207 NewResults.push_back(SDValue(UpdN.getNode(), i));
6209 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
6210 DCI.CombineTo(N, NewResults);
6211 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
6218 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
6219 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
6220 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
6222 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
6223 SelectionDAG &DAG = DCI.DAG;
6224 EVT VT = N->getValueType(0);
6225 // vldN-dup instructions only support 64-bit vectors for N > 1.
6226 if (!VT.is64BitVector())
6229 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
6230 SDNode *VLD = N->getOperand(0).getNode();
6231 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
6233 unsigned NumVecs = 0;
6234 unsigned NewOpc = 0;
6235 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
6236 if (IntNo == Intrinsic::arm_neon_vld2lane) {
6238 NewOpc = ARMISD::VLD2DUP;
6239 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
6241 NewOpc = ARMISD::VLD3DUP;
6242 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
6244 NewOpc = ARMISD::VLD4DUP;
6249 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
6250 // numbers match the load.
6251 unsigned VLDLaneNo =
6252 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
6253 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6255 // Ignore uses of the chain result.
6256 if (UI.getUse().getResNo() == NumVecs)
6259 if (User->getOpcode() != ARMISD::VDUPLANE ||
6260 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
6264 // Create the vldN-dup node.
6267 for (n = 0; n < NumVecs; ++n)
6269 Tys[n] = MVT::Other;
6270 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
6271 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
6272 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
6273 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
6274 Ops, 2, VLDMemInt->getMemoryVT(),
6275 VLDMemInt->getMemOperand());
6278 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6280 unsigned ResNo = UI.getUse().getResNo();
6281 // Ignore uses of the chain result.
6282 if (ResNo == NumVecs)
6285 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
6288 // Now the vldN-lane intrinsic is dead except for its chain result.
6289 // Update uses of the chain.
6290 std::vector<SDValue> VLDDupResults;
6291 for (unsigned n = 0; n < NumVecs; ++n)
6292 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
6293 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
6294 DCI.CombineTo(VLD, VLDDupResults);
6299 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
6300 /// ARMISD::VDUPLANE.
6301 static SDValue PerformVDUPLANECombine(SDNode *N,
6302 TargetLowering::DAGCombinerInfo &DCI) {
6303 SDValue Op = N->getOperand(0);
6305 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
6306 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
6307 if (CombineVLDDUP(N, DCI))
6308 return SDValue(N, 0);
6310 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
6311 // redundant. Ignore bit_converts for now; element sizes are checked below.
6312 while (Op.getOpcode() == ISD::BITCAST)
6313 Op = Op.getOperand(0);
6314 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
6317 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
6318 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
6319 // The canonical VMOV for a zero vector uses a 32-bit element size.
6320 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6322 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
6324 EVT VT = N->getValueType(0);
6325 if (EltSize > VT.getVectorElementType().getSizeInBits())
6328 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
6331 /// getVShiftImm - Check if this is a valid build_vector for the immediate
6332 /// operand of a vector shift operation, where all the elements of the
6333 /// build_vector must have the same constant integer value.
6334 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6335 // Ignore bit_converts.
6336 while (Op.getOpcode() == ISD::BITCAST)
6337 Op = Op.getOperand(0);
6338 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6339 APInt SplatBits, SplatUndef;
6340 unsigned SplatBitSize;
6342 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6343 HasAnyUndefs, ElementBits) ||
6344 SplatBitSize > ElementBits)
6346 Cnt = SplatBits.getSExtValue();
6350 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6351 /// operand of a vector shift left operation. That value must be in the range:
6352 /// 0 <= Value < ElementBits for a left shift; or
6353 /// 0 <= Value <= ElementBits for a long left shift.
6354 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6355 assert(VT.isVector() && "vector shift count is not a vector type");
6356 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6357 if (! getVShiftImm(Op, ElementBits, Cnt))
6359 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
6362 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6363 /// operand of a vector shift right operation. For a shift opcode, the value
6364 /// is positive, but for an intrinsic the value count must be negative. The
6365 /// absolute value must be in the range:
6366 /// 1 <= |Value| <= ElementBits for a right shift; or
6367 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6368 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6370 assert(VT.isVector() && "vector shift count is not a vector type");
6371 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6372 if (! getVShiftImm(Op, ElementBits, Cnt))
6376 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
6379 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
6380 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
6381 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
6384 // Don't do anything for most intrinsics.
6387 // Vector shifts: check for immediate versions and lower them.
6388 // Note: This is done during DAG combining instead of DAG legalizing because
6389 // the build_vectors for 64-bit vector element shift counts are generally
6390 // not legal, and it is hard to see their values after they get legalized to
6391 // loads from a constant pool.
6392 case Intrinsic::arm_neon_vshifts:
6393 case Intrinsic::arm_neon_vshiftu:
6394 case Intrinsic::arm_neon_vshiftls:
6395 case Intrinsic::arm_neon_vshiftlu:
6396 case Intrinsic::arm_neon_vshiftn:
6397 case Intrinsic::arm_neon_vrshifts:
6398 case Intrinsic::arm_neon_vrshiftu:
6399 case Intrinsic::arm_neon_vrshiftn:
6400 case Intrinsic::arm_neon_vqshifts:
6401 case Intrinsic::arm_neon_vqshiftu:
6402 case Intrinsic::arm_neon_vqshiftsu:
6403 case Intrinsic::arm_neon_vqshiftns:
6404 case Intrinsic::arm_neon_vqshiftnu:
6405 case Intrinsic::arm_neon_vqshiftnsu:
6406 case Intrinsic::arm_neon_vqrshiftns:
6407 case Intrinsic::arm_neon_vqrshiftnu:
6408 case Intrinsic::arm_neon_vqrshiftnsu: {
6409 EVT VT = N->getOperand(1).getValueType();
6411 unsigned VShiftOpc = 0;
6414 case Intrinsic::arm_neon_vshifts:
6415 case Intrinsic::arm_neon_vshiftu:
6416 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
6417 VShiftOpc = ARMISD::VSHL;
6420 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
6421 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
6422 ARMISD::VSHRs : ARMISD::VSHRu);
6427 case Intrinsic::arm_neon_vshiftls:
6428 case Intrinsic::arm_neon_vshiftlu:
6429 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
6431 llvm_unreachable("invalid shift count for vshll intrinsic");
6433 case Intrinsic::arm_neon_vrshifts:
6434 case Intrinsic::arm_neon_vrshiftu:
6435 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
6439 case Intrinsic::arm_neon_vqshifts:
6440 case Intrinsic::arm_neon_vqshiftu:
6441 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6445 case Intrinsic::arm_neon_vqshiftsu:
6446 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6448 llvm_unreachable("invalid shift count for vqshlu intrinsic");
6450 case Intrinsic::arm_neon_vshiftn:
6451 case Intrinsic::arm_neon_vrshiftn:
6452 case Intrinsic::arm_neon_vqshiftns:
6453 case Intrinsic::arm_neon_vqshiftnu:
6454 case Intrinsic::arm_neon_vqshiftnsu:
6455 case Intrinsic::arm_neon_vqrshiftns:
6456 case Intrinsic::arm_neon_vqrshiftnu:
6457 case Intrinsic::arm_neon_vqrshiftnsu:
6458 // Narrowing shifts require an immediate right shift.
6459 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
6461 llvm_unreachable("invalid shift count for narrowing vector shift "
6465 llvm_unreachable("unhandled vector shift");
6469 case Intrinsic::arm_neon_vshifts:
6470 case Intrinsic::arm_neon_vshiftu:
6471 // Opcode already set above.
6473 case Intrinsic::arm_neon_vshiftls:
6474 case Intrinsic::arm_neon_vshiftlu:
6475 if (Cnt == VT.getVectorElementType().getSizeInBits())
6476 VShiftOpc = ARMISD::VSHLLi;
6478 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
6479 ARMISD::VSHLLs : ARMISD::VSHLLu);
6481 case Intrinsic::arm_neon_vshiftn:
6482 VShiftOpc = ARMISD::VSHRN; break;
6483 case Intrinsic::arm_neon_vrshifts:
6484 VShiftOpc = ARMISD::VRSHRs; break;
6485 case Intrinsic::arm_neon_vrshiftu:
6486 VShiftOpc = ARMISD::VRSHRu; break;
6487 case Intrinsic::arm_neon_vrshiftn:
6488 VShiftOpc = ARMISD::VRSHRN; break;
6489 case Intrinsic::arm_neon_vqshifts:
6490 VShiftOpc = ARMISD::VQSHLs; break;
6491 case Intrinsic::arm_neon_vqshiftu:
6492 VShiftOpc = ARMISD::VQSHLu; break;
6493 case Intrinsic::arm_neon_vqshiftsu:
6494 VShiftOpc = ARMISD::VQSHLsu; break;
6495 case Intrinsic::arm_neon_vqshiftns:
6496 VShiftOpc = ARMISD::VQSHRNs; break;
6497 case Intrinsic::arm_neon_vqshiftnu:
6498 VShiftOpc = ARMISD::VQSHRNu; break;
6499 case Intrinsic::arm_neon_vqshiftnsu:
6500 VShiftOpc = ARMISD::VQSHRNsu; break;
6501 case Intrinsic::arm_neon_vqrshiftns:
6502 VShiftOpc = ARMISD::VQRSHRNs; break;
6503 case Intrinsic::arm_neon_vqrshiftnu:
6504 VShiftOpc = ARMISD::VQRSHRNu; break;
6505 case Intrinsic::arm_neon_vqrshiftnsu:
6506 VShiftOpc = ARMISD::VQRSHRNsu; break;
6509 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6510 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
6513 case Intrinsic::arm_neon_vshiftins: {
6514 EVT VT = N->getOperand(1).getValueType();
6516 unsigned VShiftOpc = 0;
6518 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
6519 VShiftOpc = ARMISD::VSLI;
6520 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
6521 VShiftOpc = ARMISD::VSRI;
6523 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
6526 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6527 N->getOperand(1), N->getOperand(2),
6528 DAG.getConstant(Cnt, MVT::i32));
6531 case Intrinsic::arm_neon_vqrshifts:
6532 case Intrinsic::arm_neon_vqrshiftu:
6533 // No immediate versions of these to check for.
6540 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
6541 /// lowers them. As with the vector shift intrinsics, this is done during DAG
6542 /// combining instead of DAG legalizing because the build_vectors for 64-bit
6543 /// vector element shift counts are generally not legal, and it is hard to see
6544 /// their values after they get legalized to loads from a constant pool.
6545 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
6546 const ARMSubtarget *ST) {
6547 EVT VT = N->getValueType(0);
6549 // Nothing to be done for scalar shifts.
6550 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6551 if (!VT.isVector() || !TLI.isTypeLegal(VT))
6554 assert(ST->hasNEON() && "unexpected vector shift");
6557 switch (N->getOpcode()) {
6558 default: llvm_unreachable("unexpected shift opcode");
6561 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
6562 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
6563 DAG.getConstant(Cnt, MVT::i32));
6568 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
6569 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
6570 ARMISD::VSHRs : ARMISD::VSHRu);
6571 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
6572 DAG.getConstant(Cnt, MVT::i32));
6578 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
6579 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
6580 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
6581 const ARMSubtarget *ST) {
6582 SDValue N0 = N->getOperand(0);
6584 // Check for sign- and zero-extensions of vector extract operations of 8-
6585 // and 16-bit vector elements. NEON supports these directly. They are
6586 // handled during DAG combining because type legalization will promote them
6587 // to 32-bit types and it is messy to recognize the operations after that.
6588 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6589 SDValue Vec = N0.getOperand(0);
6590 SDValue Lane = N0.getOperand(1);
6591 EVT VT = N->getValueType(0);
6592 EVT EltVT = N0.getValueType();
6593 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6595 if (VT == MVT::i32 &&
6596 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
6597 TLI.isTypeLegal(Vec.getValueType()) &&
6598 isa<ConstantSDNode>(Lane)) {
6601 switch (N->getOpcode()) {
6602 default: llvm_unreachable("unexpected opcode");
6603 case ISD::SIGN_EXTEND:
6604 Opc = ARMISD::VGETLANEs;
6606 case ISD::ZERO_EXTEND:
6607 case ISD::ANY_EXTEND:
6608 Opc = ARMISD::VGETLANEu;
6611 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
6618 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
6619 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
6620 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
6621 const ARMSubtarget *ST) {
6622 // If the target supports NEON, try to use vmax/vmin instructions for f32
6623 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
6624 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
6625 // a NaN; only do the transformation when it matches that behavior.
6627 // For now only do this when using NEON for FP operations; if using VFP, it
6628 // is not obvious that the benefit outweighs the cost of switching to the
6630 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
6631 N->getValueType(0) != MVT::f32)
6634 SDValue CondLHS = N->getOperand(0);
6635 SDValue CondRHS = N->getOperand(1);
6636 SDValue LHS = N->getOperand(2);
6637 SDValue RHS = N->getOperand(3);
6638 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
6640 unsigned Opcode = 0;
6642 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
6643 IsReversed = false; // x CC y ? x : y
6644 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
6645 IsReversed = true ; // x CC y ? y : x
6659 // If LHS is NaN, an ordered comparison will be false and the result will
6660 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
6661 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6662 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
6663 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6665 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
6666 // will return -0, so vmin can only be used for unsafe math or if one of
6667 // the operands is known to be nonzero.
6668 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
6670 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6672 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
6681 // If LHS is NaN, an ordered comparison will be false and the result will
6682 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
6683 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6684 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
6685 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6687 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
6688 // will return +0, so vmax can only be used for unsafe math or if one of
6689 // the operands is known to be nonzero.
6690 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
6692 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6694 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
6700 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
6703 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
6704 DAGCombinerInfo &DCI) const {
6705 switch (N->getOpcode()) {
6707 case ISD::ADD: return PerformADDCombine(N, DCI);
6708 case ISD::SUB: return PerformSUBCombine(N, DCI);
6709 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
6710 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
6711 case ISD::AND: return PerformANDCombine(N, DCI);
6712 case ARMISD::BFI: return PerformBFICombine(N, DCI);
6713 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
6714 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
6715 case ISD::STORE: return PerformSTORECombine(N, DCI);
6716 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
6717 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
6718 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
6719 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
6720 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
6723 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
6724 case ISD::SIGN_EXTEND:
6725 case ISD::ZERO_EXTEND:
6726 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
6727 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
6728 case ARMISD::VLD2DUP:
6729 case ARMISD::VLD3DUP:
6730 case ARMISD::VLD4DUP:
6731 return CombineBaseUpdate(N, DCI);
6732 case ISD::INTRINSIC_VOID:
6733 case ISD::INTRINSIC_W_CHAIN:
6734 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
6735 case Intrinsic::arm_neon_vld1:
6736 case Intrinsic::arm_neon_vld2:
6737 case Intrinsic::arm_neon_vld3:
6738 case Intrinsic::arm_neon_vld4:
6739 case Intrinsic::arm_neon_vld2lane:
6740 case Intrinsic::arm_neon_vld3lane:
6741 case Intrinsic::arm_neon_vld4lane:
6742 case Intrinsic::arm_neon_vst1:
6743 case Intrinsic::arm_neon_vst2:
6744 case Intrinsic::arm_neon_vst3:
6745 case Intrinsic::arm_neon_vst4:
6746 case Intrinsic::arm_neon_vst2lane:
6747 case Intrinsic::arm_neon_vst3lane:
6748 case Intrinsic::arm_neon_vst4lane:
6749 return CombineBaseUpdate(N, DCI);
6757 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
6759 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
6762 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
6763 if (!Subtarget->allowsUnalignedMem())
6766 switch (VT.getSimpleVT().SimpleTy) {
6773 // FIXME: VLD1 etc with standard alignment is legal.
6777 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
6782 switch (VT.getSimpleVT().SimpleTy) {
6783 default: return false;
6798 if ((V & (Scale - 1)) != 0)
6801 return V == (V & ((1LL << 5) - 1));
6804 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
6805 const ARMSubtarget *Subtarget) {
6812 switch (VT.getSimpleVT().SimpleTy) {
6813 default: return false;
6818 // + imm12 or - imm8
6820 return V == (V & ((1LL << 8) - 1));
6821 return V == (V & ((1LL << 12) - 1));
6824 // Same as ARM mode. FIXME: NEON?
6825 if (!Subtarget->hasVFP2())
6830 return V == (V & ((1LL << 8) - 1));
6834 /// isLegalAddressImmediate - Return true if the integer value can be used
6835 /// as the offset of the target addressing mode for load / store of the
6837 static bool isLegalAddressImmediate(int64_t V, EVT VT,
6838 const ARMSubtarget *Subtarget) {
6845 if (Subtarget->isThumb1Only())
6846 return isLegalT1AddressImmediate(V, VT);
6847 else if (Subtarget->isThumb2())
6848 return isLegalT2AddressImmediate(V, VT, Subtarget);
6853 switch (VT.getSimpleVT().SimpleTy) {
6854 default: return false;
6859 return V == (V & ((1LL << 12) - 1));
6862 return V == (V & ((1LL << 8) - 1));
6865 if (!Subtarget->hasVFP2()) // FIXME: NEON?
6870 return V == (V & ((1LL << 8) - 1));
6874 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
6876 int Scale = AM.Scale;
6880 switch (VT.getSimpleVT().SimpleTy) {
6881 default: return false;
6890 return Scale == 2 || Scale == 4 || Scale == 8;
6893 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6897 // Note, we allow "void" uses (basically, uses that aren't loads or
6898 // stores), because arm allows folding a scale into many arithmetic
6899 // operations. This should be made more precise and revisited later.
6901 // Allow r << imm, but the imm has to be a multiple of two.
6902 if (Scale & 1) return false;
6903 return isPowerOf2_32(Scale);
6907 /// isLegalAddressingMode - Return true if the addressing mode represented
6908 /// by AM is legal for this target, for a load/store of the specified type.
6909 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6910 const Type *Ty) const {
6911 EVT VT = getValueType(Ty, true);
6912 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
6915 // Can never fold addr of global into load/store.
6920 case 0: // no scale reg, must be "r+i" or "r", or "i".
6923 if (Subtarget->isThumb1Only())
6927 // ARM doesn't support any R+R*scale+imm addr modes.
6934 if (Subtarget->isThumb2())
6935 return isLegalT2ScaledAddressingMode(AM, VT);
6937 int Scale = AM.Scale;
6938 switch (VT.getSimpleVT().SimpleTy) {
6939 default: return false;
6943 if (Scale < 0) Scale = -Scale;
6947 return isPowerOf2_32(Scale & ~1);
6951 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6956 // Note, we allow "void" uses (basically, uses that aren't loads or
6957 // stores), because arm allows folding a scale into many arithmetic
6958 // operations. This should be made more precise and revisited later.
6960 // Allow r << imm, but the imm has to be a multiple of two.
6961 if (Scale & 1) return false;
6962 return isPowerOf2_32(Scale);
6969 /// isLegalICmpImmediate - Return true if the specified immediate is legal
6970 /// icmp immediate, that is the target has icmp instructions which can compare
6971 /// a register against the immediate without having to materialize the
6972 /// immediate into a register.
6973 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
6974 if (!Subtarget->isThumb())
6975 return ARM_AM::getSOImmVal(Imm) != -1;
6976 if (Subtarget->isThumb2())
6977 return ARM_AM::getT2SOImmVal(Imm) != -1;
6978 return Imm >= 0 && Imm <= 255;
6981 /// isLegalAddImmediate - Return true if the specified immediate is legal
6982 /// add immediate, that is the target has add instructions which can add
6983 /// a register with the immediate without having to materialize the
6984 /// immediate into a register.
6985 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
6986 return ARM_AM::getSOImmVal(Imm) != -1;
6989 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
6990 bool isSEXTLoad, SDValue &Base,
6991 SDValue &Offset, bool &isInc,
6992 SelectionDAG &DAG) {
6993 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6996 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
6998 Base = Ptr->getOperand(0);
6999 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7000 int RHSC = (int)RHS->getZExtValue();
7001 if (RHSC < 0 && RHSC > -256) {
7002 assert(Ptr->getOpcode() == ISD::ADD);
7004 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7008 isInc = (Ptr->getOpcode() == ISD::ADD);
7009 Offset = Ptr->getOperand(1);
7011 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
7013 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7014 int RHSC = (int)RHS->getZExtValue();
7015 if (RHSC < 0 && RHSC > -0x1000) {
7016 assert(Ptr->getOpcode() == ISD::ADD);
7018 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7019 Base = Ptr->getOperand(0);
7024 if (Ptr->getOpcode() == ISD::ADD) {
7026 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
7027 if (ShOpcVal != ARM_AM::no_shift) {
7028 Base = Ptr->getOperand(1);
7029 Offset = Ptr->getOperand(0);
7031 Base = Ptr->getOperand(0);
7032 Offset = Ptr->getOperand(1);
7037 isInc = (Ptr->getOpcode() == ISD::ADD);
7038 Base = Ptr->getOperand(0);
7039 Offset = Ptr->getOperand(1);
7043 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
7047 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
7048 bool isSEXTLoad, SDValue &Base,
7049 SDValue &Offset, bool &isInc,
7050 SelectionDAG &DAG) {
7051 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
7054 Base = Ptr->getOperand(0);
7055 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7056 int RHSC = (int)RHS->getZExtValue();
7057 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
7058 assert(Ptr->getOpcode() == ISD::ADD);
7060 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7062 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
7063 isInc = Ptr->getOpcode() == ISD::ADD;
7064 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
7072 /// getPreIndexedAddressParts - returns true by value, base pointer and
7073 /// offset pointer and addressing mode by reference if the node's address
7074 /// can be legally represented as pre-indexed load / store address.
7076 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7078 ISD::MemIndexedMode &AM,
7079 SelectionDAG &DAG) const {
7080 if (Subtarget->isThumb1Only())
7085 bool isSEXTLoad = false;
7086 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7087 Ptr = LD->getBasePtr();
7088 VT = LD->getMemoryVT();
7089 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7090 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7091 Ptr = ST->getBasePtr();
7092 VT = ST->getMemoryVT();
7097 bool isLegal = false;
7098 if (Subtarget->isThumb2())
7099 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7100 Offset, isInc, DAG);
7102 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7103 Offset, isInc, DAG);
7107 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
7111 /// getPostIndexedAddressParts - returns true by value, base pointer and
7112 /// offset pointer and addressing mode by reference if this node can be
7113 /// combined with a load / store to form a post-indexed load / store.
7114 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7117 ISD::MemIndexedMode &AM,
7118 SelectionDAG &DAG) const {
7119 if (Subtarget->isThumb1Only())
7124 bool isSEXTLoad = false;
7125 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7126 VT = LD->getMemoryVT();
7127 Ptr = LD->getBasePtr();
7128 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7129 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7130 VT = ST->getMemoryVT();
7131 Ptr = ST->getBasePtr();
7136 bool isLegal = false;
7137 if (Subtarget->isThumb2())
7138 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7141 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7147 // Swap base ptr and offset to catch more post-index load / store when
7148 // it's legal. In Thumb2 mode, offset must be an immediate.
7149 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
7150 !Subtarget->isThumb2())
7151 std::swap(Base, Offset);
7153 // Post-indexed load / store update the base pointer.
7158 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
7162 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7166 const SelectionDAG &DAG,
7167 unsigned Depth) const {
7168 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
7169 switch (Op.getOpcode()) {
7171 case ARMISD::CMOV: {
7172 // Bits are known zero/one if known on the LHS and RHS.
7173 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
7174 if (KnownZero == 0 && KnownOne == 0) return;
7176 APInt KnownZeroRHS, KnownOneRHS;
7177 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
7178 KnownZeroRHS, KnownOneRHS, Depth+1);
7179 KnownZero &= KnownZeroRHS;
7180 KnownOne &= KnownOneRHS;
7186 //===----------------------------------------------------------------------===//
7187 // ARM Inline Assembly Support
7188 //===----------------------------------------------------------------------===//
7190 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
7191 // Looking for "rev" which is V6+.
7192 if (!Subtarget->hasV6Ops())
7195 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
7196 std::string AsmStr = IA->getAsmString();
7197 SmallVector<StringRef, 4> AsmPieces;
7198 SplitString(AsmStr, AsmPieces, ";\n");
7200 switch (AsmPieces.size()) {
7201 default: return false;
7203 AsmStr = AsmPieces[0];
7205 SplitString(AsmStr, AsmPieces, " \t,");
7208 if (AsmPieces.size() == 3 &&
7209 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
7210 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
7211 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
7212 if (Ty && Ty->getBitWidth() == 32)
7213 return IntrinsicLowering::LowerToByteSwap(CI);
7221 /// getConstraintType - Given a constraint letter, return the type of
7222 /// constraint it is for this target.
7223 ARMTargetLowering::ConstraintType
7224 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
7225 if (Constraint.size() == 1) {
7226 switch (Constraint[0]) {
7228 case 'l': return C_RegisterClass;
7229 case 'w': return C_RegisterClass;
7232 return TargetLowering::getConstraintType(Constraint);
7235 /// Examine constraint type and operand type and determine a weight value.
7236 /// This object must already have been set up with the operand type
7237 /// and the current alternative constraint selected.
7238 TargetLowering::ConstraintWeight
7239 ARMTargetLowering::getSingleConstraintMatchWeight(
7240 AsmOperandInfo &info, const char *constraint) const {
7241 ConstraintWeight weight = CW_Invalid;
7242 Value *CallOperandVal = info.CallOperandVal;
7243 // If we don't have a value, we can't do a match,
7244 // but allow it at the lowest weight.
7245 if (CallOperandVal == NULL)
7247 const Type *type = CallOperandVal->getType();
7248 // Look at the constraint type.
7249 switch (*constraint) {
7251 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
7254 if (type->isIntegerTy()) {
7255 if (Subtarget->isThumb())
7256 weight = CW_SpecificReg;
7258 weight = CW_Register;
7262 if (type->isFloatingPointTy())
7263 weight = CW_Register;
7269 std::pair<unsigned, const TargetRegisterClass*>
7270 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7272 if (Constraint.size() == 1) {
7273 // GCC ARM Constraint Letters
7274 switch (Constraint[0]) {
7276 if (Subtarget->isThumb())
7277 return std::make_pair(0U, ARM::tGPRRegisterClass);
7279 return std::make_pair(0U, ARM::GPRRegisterClass);
7281 return std::make_pair(0U, ARM::GPRRegisterClass);
7284 return std::make_pair(0U, ARM::SPRRegisterClass);
7285 if (VT.getSizeInBits() == 64)
7286 return std::make_pair(0U, ARM::DPRRegisterClass);
7287 if (VT.getSizeInBits() == 128)
7288 return std::make_pair(0U, ARM::QPRRegisterClass);
7292 if (StringRef("{cc}").equals_lower(Constraint))
7293 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
7295 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7298 std::vector<unsigned> ARMTargetLowering::
7299 getRegClassForInlineAsmConstraint(const std::string &Constraint,
7301 if (Constraint.size() != 1)
7302 return std::vector<unsigned>();
7304 switch (Constraint[0]) { // GCC ARM Constraint Letters
7307 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
7308 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
7311 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
7312 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
7313 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
7314 ARM::R12, ARM::LR, 0);
7317 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
7318 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
7319 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
7320 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
7321 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
7322 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
7323 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
7324 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
7325 if (VT.getSizeInBits() == 64)
7326 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
7327 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
7328 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
7329 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
7330 if (VT.getSizeInBits() == 128)
7331 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
7332 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
7336 return std::vector<unsigned>();
7339 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7340 /// vector. If it is invalid, don't add anything to Ops.
7341 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7343 std::vector<SDValue>&Ops,
7344 SelectionDAG &DAG) const {
7345 SDValue Result(0, 0);
7347 switch (Constraint) {
7349 case 'I': case 'J': case 'K': case 'L':
7350 case 'M': case 'N': case 'O':
7351 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
7355 int64_t CVal64 = C->getSExtValue();
7356 int CVal = (int) CVal64;
7357 // None of these constraints allow values larger than 32 bits. Check
7358 // that the value fits in an int.
7362 switch (Constraint) {
7364 if (Subtarget->isThumb1Only()) {
7365 // This must be a constant between 0 and 255, for ADD
7367 if (CVal >= 0 && CVal <= 255)
7369 } else if (Subtarget->isThumb2()) {
7370 // A constant that can be used as an immediate value in a
7371 // data-processing instruction.
7372 if (ARM_AM::getT2SOImmVal(CVal) != -1)
7375 // A constant that can be used as an immediate value in a
7376 // data-processing instruction.
7377 if (ARM_AM::getSOImmVal(CVal) != -1)
7383 if (Subtarget->isThumb()) { // FIXME thumb2
7384 // This must be a constant between -255 and -1, for negated ADD
7385 // immediates. This can be used in GCC with an "n" modifier that
7386 // prints the negated value, for use with SUB instructions. It is
7387 // not useful otherwise but is implemented for compatibility.
7388 if (CVal >= -255 && CVal <= -1)
7391 // This must be a constant between -4095 and 4095. It is not clear
7392 // what this constraint is intended for. Implemented for
7393 // compatibility with GCC.
7394 if (CVal >= -4095 && CVal <= 4095)
7400 if (Subtarget->isThumb1Only()) {
7401 // A 32-bit value where only one byte has a nonzero value. Exclude
7402 // zero to match GCC. This constraint is used by GCC internally for
7403 // constants that can be loaded with a move/shift combination.
7404 // It is not useful otherwise but is implemented for compatibility.
7405 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
7407 } else if (Subtarget->isThumb2()) {
7408 // A constant whose bitwise inverse can be used as an immediate
7409 // value in a data-processing instruction. This can be used in GCC
7410 // with a "B" modifier that prints the inverted value, for use with
7411 // BIC and MVN instructions. It is not useful otherwise but is
7412 // implemented for compatibility.
7413 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
7416 // A constant whose bitwise inverse can be used as an immediate
7417 // value in a data-processing instruction. This can be used in GCC
7418 // with a "B" modifier that prints the inverted value, for use with
7419 // BIC and MVN instructions. It is not useful otherwise but is
7420 // implemented for compatibility.
7421 if (ARM_AM::getSOImmVal(~CVal) != -1)
7427 if (Subtarget->isThumb1Only()) {
7428 // This must be a constant between -7 and 7,
7429 // for 3-operand ADD/SUB immediate instructions.
7430 if (CVal >= -7 && CVal < 7)
7432 } else if (Subtarget->isThumb2()) {
7433 // A constant whose negation can be used as an immediate value in a
7434 // data-processing instruction. This can be used in GCC with an "n"
7435 // modifier that prints the negated value, for use with SUB
7436 // instructions. It is not useful otherwise but is implemented for
7438 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
7441 // A constant whose negation can be used as an immediate value in a
7442 // data-processing instruction. This can be used in GCC with an "n"
7443 // modifier that prints the negated value, for use with SUB
7444 // instructions. It is not useful otherwise but is implemented for
7446 if (ARM_AM::getSOImmVal(-CVal) != -1)
7452 if (Subtarget->isThumb()) { // FIXME thumb2
7453 // This must be a multiple of 4 between 0 and 1020, for
7454 // ADD sp + immediate.
7455 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
7458 // A power of two or a constant between 0 and 32. This is used in
7459 // GCC for the shift amount on shifted register operands, but it is
7460 // useful in general for any shift amounts.
7461 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
7467 if (Subtarget->isThumb()) { // FIXME thumb2
7468 // This must be a constant between 0 and 31, for shift amounts.
7469 if (CVal >= 0 && CVal <= 31)
7475 if (Subtarget->isThumb()) { // FIXME thumb2
7476 // This must be a multiple of 4 between -508 and 508, for
7477 // ADD/SUB sp = sp + immediate.
7478 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
7483 Result = DAG.getTargetConstant(CVal, Op.getValueType());
7487 if (Result.getNode()) {
7488 Ops.push_back(Result);
7491 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7495 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
7496 // The ARM target isn't yet aware of offsets.
7500 int ARM::getVFPf32Imm(const APFloat &FPImm) {
7501 APInt Imm = FPImm.bitcastToAPInt();
7502 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
7503 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
7504 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
7506 // We can handle 4 bits of mantissa.
7507 // mantissa = (16+UInt(e:f:g:h))/16.
7508 if (Mantissa & 0x7ffff)
7511 if ((Mantissa & 0xf) != Mantissa)
7514 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
7515 if (Exp < -3 || Exp > 4)
7517 Exp = ((Exp+3) & 0x7) ^ 4;
7519 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
7522 int ARM::getVFPf64Imm(const APFloat &FPImm) {
7523 APInt Imm = FPImm.bitcastToAPInt();
7524 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
7525 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
7526 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
7528 // We can handle 4 bits of mantissa.
7529 // mantissa = (16+UInt(e:f:g:h))/16.
7530 if (Mantissa & 0xffffffffffffLL)
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 bool ARM::isBitFieldInvertedMask(unsigned v) {
7545 if (v == 0xffffffff)
7547 // there can be 1's on either or both "outsides", all the "inside"
7549 unsigned int lsb = 0, msb = 31;
7550 while (v & (1 << msb)) --msb;
7551 while (v & (1 << lsb)) ++lsb;
7552 for (unsigned int i = lsb; i <= msb; ++i) {
7559 /// isFPImmLegal - Returns true if the target can instruction select the
7560 /// specified FP immediate natively. If false, the legalizer will
7561 /// materialize the FP immediate as a load from a constant pool.
7562 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
7563 if (!Subtarget->hasVFP3())
7566 return ARM::getVFPf32Imm(Imm) != -1;
7568 return ARM::getVFPf64Imm(Imm) != -1;
7572 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
7573 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
7574 /// specified in the intrinsic calls.
7575 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
7577 unsigned Intrinsic) const {
7578 switch (Intrinsic) {
7579 case Intrinsic::arm_neon_vld1:
7580 case Intrinsic::arm_neon_vld2:
7581 case Intrinsic::arm_neon_vld3:
7582 case Intrinsic::arm_neon_vld4:
7583 case Intrinsic::arm_neon_vld2lane:
7584 case Intrinsic::arm_neon_vld3lane:
7585 case Intrinsic::arm_neon_vld4lane: {
7586 Info.opc = ISD::INTRINSIC_W_CHAIN;
7587 // Conservatively set memVT to the entire set of vectors loaded.
7588 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
7589 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7590 Info.ptrVal = I.getArgOperand(0);
7592 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7593 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7594 Info.vol = false; // volatile loads with NEON intrinsics not supported
7595 Info.readMem = true;
7596 Info.writeMem = false;
7599 case Intrinsic::arm_neon_vst1:
7600 case Intrinsic::arm_neon_vst2:
7601 case Intrinsic::arm_neon_vst3:
7602 case Intrinsic::arm_neon_vst4:
7603 case Intrinsic::arm_neon_vst2lane:
7604 case Intrinsic::arm_neon_vst3lane:
7605 case Intrinsic::arm_neon_vst4lane: {
7606 Info.opc = ISD::INTRINSIC_VOID;
7607 // Conservatively set memVT to the entire set of vectors stored.
7608 unsigned NumElts = 0;
7609 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
7610 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
7611 if (!ArgTy->isVectorTy())
7613 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
7615 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7616 Info.ptrVal = I.getArgOperand(0);
7618 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7619 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7620 Info.vol = false; // volatile stores with NEON intrinsics not supported
7621 Info.readMem = false;
7622 Info.writeMem = true;