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 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
76 EVT PromotedBitwiseVT) {
77 if (VT != PromotedLdStVT) {
78 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
79 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
80 PromotedLdStVT.getSimpleVT());
82 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
83 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
84 PromotedLdStVT.getSimpleVT());
87 EVT ElemTy = VT.getVectorElementType();
88 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
89 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
90 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
91 if (ElemTy != MVT::i32) {
92 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
94 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
95 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
97 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
98 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
99 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
100 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
101 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
102 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
103 if (VT.isInteger()) {
104 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
105 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
106 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
107 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
108 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
109 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
110 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
111 setTruncStoreAction(VT.getSimpleVT(),
112 (MVT::SimpleValueType)InnerVT, Expand);
114 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
116 // Promote all bit-wise operations.
117 if (VT.isInteger() && VT != PromotedBitwiseVT) {
118 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
119 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
120 PromotedBitwiseVT.getSimpleVT());
121 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
122 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
123 PromotedBitwiseVT.getSimpleVT());
124 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
125 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
126 PromotedBitwiseVT.getSimpleVT());
129 // Neon does not support vector divide/remainder operations.
130 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
131 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
132 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
133 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
134 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
135 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
138 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
139 addRegisterClass(VT, ARM::DPRRegisterClass);
140 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
143 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
144 addRegisterClass(VT, ARM::QPRRegisterClass);
145 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
148 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
149 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
150 return new TargetLoweringObjectFileMachO();
152 return new ARMElfTargetObjectFile();
155 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
156 : TargetLowering(TM, createTLOF(TM)) {
157 Subtarget = &TM.getSubtarget<ARMSubtarget>();
158 RegInfo = TM.getRegisterInfo();
159 Itins = TM.getInstrItineraryData();
161 if (Subtarget->isTargetDarwin()) {
162 // Uses VFP for Thumb libfuncs if available.
163 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
164 // Single-precision floating-point arithmetic.
165 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
166 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
167 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
168 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
170 // Double-precision floating-point arithmetic.
171 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
172 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
173 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
174 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
176 // Single-precision comparisons.
177 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
178 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
179 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
180 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
181 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
182 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
183 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
184 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
186 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
195 // Double-precision comparisons.
196 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
197 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
198 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
199 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
200 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
201 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
202 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
203 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
205 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
214 // Floating-point to integer conversions.
215 // i64 conversions are done via library routines even when generating VFP
216 // instructions, so use the same ones.
217 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
218 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
219 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
220 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
222 // Conversions between floating types.
223 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
224 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
226 // Integer to floating-point conversions.
227 // i64 conversions are done via library routines even when generating VFP
228 // instructions, so use the same ones.
229 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
230 // e.g., __floatunsidf vs. __floatunssidfvfp.
231 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
232 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
233 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
234 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
238 // These libcalls are not available in 32-bit.
239 setLibcallName(RTLIB::SHL_I128, 0);
240 setLibcallName(RTLIB::SRL_I128, 0);
241 setLibcallName(RTLIB::SRA_I128, 0);
243 if (Subtarget->isAAPCS_ABI()) {
244 // Double-precision floating-point arithmetic helper functions
245 // RTABI chapter 4.1.2, Table 2
246 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
247 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
248 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
249 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
250 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
251 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
252 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
253 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
255 // Double-precision floating-point comparison helper functions
256 // RTABI chapter 4.1.2, Table 3
257 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
258 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
259 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
260 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
261 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
262 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
263 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
264 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
265 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
266 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
267 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
268 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
269 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
270 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
271 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
272 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
273 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
275 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
276 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
277 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
278 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
279 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
280 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
282 // Single-precision floating-point arithmetic helper functions
283 // RTABI chapter 4.1.2, Table 4
284 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
285 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
286 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
287 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
288 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
289 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
293 // Single-precision floating-point comparison helper functions
294 // RTABI chapter 4.1.2, Table 5
295 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
296 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
297 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
298 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
299 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
300 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
301 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
302 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
303 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
304 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
305 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
306 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
307 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
308 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
309 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
310 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
311 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
313 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
314 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
315 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
316 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
317 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
318 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
320 // Floating-point to integer conversions.
321 // RTABI chapter 4.1.2, Table 6
322 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
323 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
324 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
325 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
326 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
327 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
328 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
329 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
330 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
339 // Conversions between floating types.
340 // RTABI chapter 4.1.2, Table 7
341 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
342 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
343 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
344 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
346 // Integer to floating-point conversions.
347 // RTABI chapter 4.1.2, Table 8
348 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
349 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
350 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
351 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
352 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
353 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
354 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
355 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
356 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
361 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
362 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
363 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
365 // Long long helper functions
366 // RTABI chapter 4.2, Table 9
367 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
368 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
369 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
370 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
371 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
372 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
373 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
380 // Integer division functions
381 // RTABI chapter 4.3.1
382 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
383 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
384 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
385 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
386 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
387 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
388 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
396 if (Subtarget->isThumb1Only())
397 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
399 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
400 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
401 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
402 if (!Subtarget->isFPOnlySP())
403 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
405 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
408 if (Subtarget->hasNEON()) {
409 addDRTypeForNEON(MVT::v2f32);
410 addDRTypeForNEON(MVT::v8i8);
411 addDRTypeForNEON(MVT::v4i16);
412 addDRTypeForNEON(MVT::v2i32);
413 addDRTypeForNEON(MVT::v1i64);
415 addQRTypeForNEON(MVT::v4f32);
416 addQRTypeForNEON(MVT::v2f64);
417 addQRTypeForNEON(MVT::v16i8);
418 addQRTypeForNEON(MVT::v8i16);
419 addQRTypeForNEON(MVT::v4i32);
420 addQRTypeForNEON(MVT::v2i64);
422 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
423 // neither Neon nor VFP support any arithmetic operations on it.
424 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
425 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
426 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
427 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
428 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
429 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
430 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
431 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
432 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
433 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
434 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
435 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
436 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
437 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
438 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
439 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
440 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
441 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
442 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
443 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
444 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
445 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
446 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
447 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
449 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
451 // Neon does not support some operations on v1i64 and v2i64 types.
452 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
453 // Custom handling for some quad-vector types to detect VMULL.
454 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
455 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
456 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
457 // Custom handling for some vector types to avoid expensive expansions
458 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
459 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
460 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
461 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
462 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
463 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
465 setTargetDAGCombine(ISD::INTRINSIC_VOID);
466 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
467 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
468 setTargetDAGCombine(ISD::SHL);
469 setTargetDAGCombine(ISD::SRL);
470 setTargetDAGCombine(ISD::SRA);
471 setTargetDAGCombine(ISD::SIGN_EXTEND);
472 setTargetDAGCombine(ISD::ZERO_EXTEND);
473 setTargetDAGCombine(ISD::ANY_EXTEND);
474 setTargetDAGCombine(ISD::SELECT_CC);
475 setTargetDAGCombine(ISD::BUILD_VECTOR);
476 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
477 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
478 setTargetDAGCombine(ISD::STORE);
481 computeRegisterProperties();
483 // ARM does not have f32 extending load.
484 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
486 // ARM does not have i1 sign extending load.
487 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
489 // ARM supports all 4 flavors of integer indexed load / store.
490 if (!Subtarget->isThumb1Only()) {
491 for (unsigned im = (unsigned)ISD::PRE_INC;
492 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
493 setIndexedLoadAction(im, MVT::i1, Legal);
494 setIndexedLoadAction(im, MVT::i8, Legal);
495 setIndexedLoadAction(im, MVT::i16, Legal);
496 setIndexedLoadAction(im, MVT::i32, Legal);
497 setIndexedStoreAction(im, MVT::i1, Legal);
498 setIndexedStoreAction(im, MVT::i8, Legal);
499 setIndexedStoreAction(im, MVT::i16, Legal);
500 setIndexedStoreAction(im, MVT::i32, Legal);
504 // i64 operation support.
505 if (Subtarget->isThumb1Only()) {
506 setOperationAction(ISD::MUL, MVT::i64, Expand);
507 setOperationAction(ISD::MULHU, MVT::i32, Expand);
508 setOperationAction(ISD::MULHS, MVT::i32, Expand);
509 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
510 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
512 setOperationAction(ISD::MUL, MVT::i64, Expand);
513 setOperationAction(ISD::MULHU, MVT::i32, Expand);
514 if (!Subtarget->hasV6Ops())
515 setOperationAction(ISD::MULHS, MVT::i32, Expand);
517 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
518 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
519 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
520 setOperationAction(ISD::SRL, MVT::i64, Custom);
521 setOperationAction(ISD::SRA, MVT::i64, Custom);
523 // ARM does not have ROTL.
524 setOperationAction(ISD::ROTL, MVT::i32, Expand);
525 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
526 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
527 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
528 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
530 // Only ARMv6 has BSWAP.
531 if (!Subtarget->hasV6Ops())
532 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
534 // These are expanded into libcalls.
535 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
536 // v7M has a hardware divider
537 setOperationAction(ISD::SDIV, MVT::i32, Expand);
538 setOperationAction(ISD::UDIV, MVT::i32, Expand);
540 setOperationAction(ISD::SREM, MVT::i32, Expand);
541 setOperationAction(ISD::UREM, MVT::i32, Expand);
542 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
543 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
545 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
546 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
547 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
548 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
549 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
551 setOperationAction(ISD::TRAP, MVT::Other, Legal);
553 // Use the default implementation.
554 setOperationAction(ISD::VASTART, MVT::Other, Custom);
555 setOperationAction(ISD::VAARG, MVT::Other, Expand);
556 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
557 setOperationAction(ISD::VAEND, MVT::Other, Expand);
558 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
559 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
560 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
561 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
562 setExceptionPointerRegister(ARM::R0);
563 setExceptionSelectorRegister(ARM::R1);
565 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
566 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
567 // the default expansion.
568 if (Subtarget->hasDataBarrier() ||
569 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
570 // membarrier needs custom lowering; the rest are legal and handled
572 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
574 // Set them all for expansion, which will force libcalls.
575 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
576 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
577 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
578 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
579 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
580 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
581 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
584 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
585 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
586 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
587 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
588 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
589 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
590 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
591 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
592 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
593 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
595 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
596 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
600 // Since the libcalls include locking, fold in the fences
601 setShouldFoldAtomicFences(true);
603 // 64-bit versions are always libcalls (for now)
604 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
605 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
606 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
607 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
608 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
609 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
610 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
611 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
613 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
615 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
616 if (!Subtarget->hasV6Ops()) {
617 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
618 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
620 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
622 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
623 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
624 // iff target supports vfp2.
625 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
626 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
629 // We want to custom lower some of our intrinsics.
630 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
631 if (Subtarget->isTargetDarwin()) {
632 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
633 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
634 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
637 setOperationAction(ISD::SETCC, MVT::i32, Expand);
638 setOperationAction(ISD::SETCC, MVT::f32, Expand);
639 setOperationAction(ISD::SETCC, MVT::f64, Expand);
640 setOperationAction(ISD::SELECT, MVT::i32, Custom);
641 setOperationAction(ISD::SELECT, MVT::f32, Custom);
642 setOperationAction(ISD::SELECT, MVT::f64, Custom);
643 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
644 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
645 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
647 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
648 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
649 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
650 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
651 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
653 // We don't support sin/cos/fmod/copysign/pow
654 setOperationAction(ISD::FSIN, MVT::f64, Expand);
655 setOperationAction(ISD::FSIN, MVT::f32, Expand);
656 setOperationAction(ISD::FCOS, MVT::f32, Expand);
657 setOperationAction(ISD::FCOS, MVT::f64, Expand);
658 setOperationAction(ISD::FREM, MVT::f64, Expand);
659 setOperationAction(ISD::FREM, MVT::f32, Expand);
660 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
661 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
662 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
664 setOperationAction(ISD::FPOW, MVT::f64, Expand);
665 setOperationAction(ISD::FPOW, MVT::f32, Expand);
667 // Various VFP goodness
668 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
669 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
670 if (Subtarget->hasVFP2()) {
671 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
672 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
673 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
674 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
676 // Special handling for half-precision FP.
677 if (!Subtarget->hasFP16()) {
678 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
679 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
683 // We have target-specific dag combine patterns for the following nodes:
684 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
685 setTargetDAGCombine(ISD::ADD);
686 setTargetDAGCombine(ISD::SUB);
687 setTargetDAGCombine(ISD::MUL);
689 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
690 setTargetDAGCombine(ISD::OR);
691 if (Subtarget->hasNEON())
692 setTargetDAGCombine(ISD::AND);
694 setStackPointerRegisterToSaveRestore(ARM::SP);
696 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
697 setSchedulingPreference(Sched::RegPressure);
699 setSchedulingPreference(Sched::Hybrid);
701 //// temporary - rewrite interface to use type
702 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
704 // On ARM arguments smaller than 4 bytes are extended, so all arguments
705 // are at least 4 bytes aligned.
706 setMinStackArgumentAlignment(4);
708 benefitFromCodePlacementOpt = true;
711 // FIXME: It might make sense to define the representative register class as the
712 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
713 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
714 // SPR's representative would be DPR_VFP2. This should work well if register
715 // pressure tracking were modified such that a register use would increment the
716 // pressure of the register class's representative and all of it's super
717 // classes' representatives transitively. We have not implemented this because
718 // of the difficulty prior to coalescing of modeling operand register classes
719 // due to the common occurence of cross class copies and subregister insertions
721 std::pair<const TargetRegisterClass*, uint8_t>
722 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
723 const TargetRegisterClass *RRC = 0;
725 switch (VT.getSimpleVT().SimpleTy) {
727 return TargetLowering::findRepresentativeClass(VT);
728 // Use DPR as representative register class for all floating point
729 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
730 // the cost is 1 for both f32 and f64.
731 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
732 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
733 RRC = ARM::DPRRegisterClass;
734 // When NEON is used for SP, only half of the register file is available
735 // because operations that define both SP and DP results will be constrained
736 // to the VFP2 class (D0-D15). We currently model this constraint prior to
737 // coalescing by double-counting the SP regs. See the FIXME above.
738 if (Subtarget->useNEONForSinglePrecisionFP())
741 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
742 case MVT::v4f32: case MVT::v2f64:
743 RRC = ARM::DPRRegisterClass;
747 RRC = ARM::DPRRegisterClass;
751 RRC = ARM::DPRRegisterClass;
755 return std::make_pair(RRC, Cost);
758 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
761 case ARMISD::Wrapper: return "ARMISD::Wrapper";
762 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
763 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
764 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
765 case ARMISD::CALL: return "ARMISD::CALL";
766 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
767 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
768 case ARMISD::tCALL: return "ARMISD::tCALL";
769 case ARMISD::BRCOND: return "ARMISD::BRCOND";
770 case ARMISD::BR_JT: return "ARMISD::BR_JT";
771 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
772 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
773 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
774 case ARMISD::CMP: return "ARMISD::CMP";
775 case ARMISD::CMPZ: return "ARMISD::CMPZ";
776 case ARMISD::CMPFP: return "ARMISD::CMPFP";
777 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
778 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
779 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
780 case ARMISD::CMOV: return "ARMISD::CMOV";
781 case ARMISD::CNEG: return "ARMISD::CNEG";
783 case ARMISD::RBIT: return "ARMISD::RBIT";
785 case ARMISD::FTOSI: return "ARMISD::FTOSI";
786 case ARMISD::FTOUI: return "ARMISD::FTOUI";
787 case ARMISD::SITOF: return "ARMISD::SITOF";
788 case ARMISD::UITOF: return "ARMISD::UITOF";
790 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
791 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
792 case ARMISD::RRX: return "ARMISD::RRX";
794 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
795 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
797 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
798 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
799 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
801 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
803 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
805 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
807 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
808 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
810 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
812 case ARMISD::VCEQ: return "ARMISD::VCEQ";
813 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
814 case ARMISD::VCGE: return "ARMISD::VCGE";
815 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
816 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
817 case ARMISD::VCGEU: return "ARMISD::VCGEU";
818 case ARMISD::VCGT: return "ARMISD::VCGT";
819 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
820 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
821 case ARMISD::VCGTU: return "ARMISD::VCGTU";
822 case ARMISD::VTST: return "ARMISD::VTST";
824 case ARMISD::VSHL: return "ARMISD::VSHL";
825 case ARMISD::VSHRs: return "ARMISD::VSHRs";
826 case ARMISD::VSHRu: return "ARMISD::VSHRu";
827 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
828 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
829 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
830 case ARMISD::VSHRN: return "ARMISD::VSHRN";
831 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
832 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
833 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
834 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
835 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
836 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
837 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
838 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
839 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
840 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
841 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
842 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
843 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
844 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
845 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
846 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
847 case ARMISD::VDUP: return "ARMISD::VDUP";
848 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
849 case ARMISD::VEXT: return "ARMISD::VEXT";
850 case ARMISD::VREV64: return "ARMISD::VREV64";
851 case ARMISD::VREV32: return "ARMISD::VREV32";
852 case ARMISD::VREV16: return "ARMISD::VREV16";
853 case ARMISD::VZIP: return "ARMISD::VZIP";
854 case ARMISD::VUZP: return "ARMISD::VUZP";
855 case ARMISD::VTRN: return "ARMISD::VTRN";
856 case ARMISD::VMULLs: return "ARMISD::VMULLs";
857 case ARMISD::VMULLu: return "ARMISD::VMULLu";
858 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
859 case ARMISD::FMAX: return "ARMISD::FMAX";
860 case ARMISD::FMIN: return "ARMISD::FMIN";
861 case ARMISD::BFI: return "ARMISD::BFI";
862 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
863 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
864 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
865 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
866 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
867 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
868 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
869 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
870 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
871 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
872 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
873 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
874 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
875 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
876 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
877 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
878 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
879 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
880 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
881 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
882 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
883 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
887 /// getRegClassFor - Return the register class that should be used for the
888 /// specified value type.
889 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
890 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
891 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
892 // load / store 4 to 8 consecutive D registers.
893 if (Subtarget->hasNEON()) {
894 if (VT == MVT::v4i64)
895 return ARM::QQPRRegisterClass;
896 else if (VT == MVT::v8i64)
897 return ARM::QQQQPRRegisterClass;
899 return TargetLowering::getRegClassFor(VT);
902 // Create a fast isel object.
904 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
905 return ARM::createFastISel(funcInfo);
908 /// getFunctionAlignment - Return the Log2 alignment of this function.
909 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
910 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
913 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
914 /// be used for loads / stores from the global.
915 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
916 return (Subtarget->isThumb1Only() ? 127 : 4095);
919 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
920 unsigned NumVals = N->getNumValues();
922 return Sched::RegPressure;
924 for (unsigned i = 0; i != NumVals; ++i) {
925 EVT VT = N->getValueType(i);
926 if (VT == MVT::Glue || VT == MVT::Other)
928 if (VT.isFloatingPoint() || VT.isVector())
929 return Sched::Latency;
932 if (!N->isMachineOpcode())
933 return Sched::RegPressure;
935 // Load are scheduled for latency even if there instruction itinerary
937 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
938 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
940 if (TID.getNumDefs() == 0)
941 return Sched::RegPressure;
942 if (!Itins->isEmpty() &&
943 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
944 return Sched::Latency;
946 return Sched::RegPressure;
949 // FIXME: Move to RegInfo
951 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
952 MachineFunction &MF) const {
953 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering();
955 switch (RC->getID()) {
958 case ARM::tGPRRegClassID:
959 return TFI->hasFP(MF) ? 4 : 5;
960 case ARM::GPRRegClassID: {
961 unsigned FP = TFI->hasFP(MF) ? 1 : 0;
962 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
964 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
965 case ARM::DPRRegClassID:
970 //===----------------------------------------------------------------------===//
972 //===----------------------------------------------------------------------===//
974 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
975 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
977 default: llvm_unreachable("Unknown condition code!");
978 case ISD::SETNE: return ARMCC::NE;
979 case ISD::SETEQ: return ARMCC::EQ;
980 case ISD::SETGT: return ARMCC::GT;
981 case ISD::SETGE: return ARMCC::GE;
982 case ISD::SETLT: return ARMCC::LT;
983 case ISD::SETLE: return ARMCC::LE;
984 case ISD::SETUGT: return ARMCC::HI;
985 case ISD::SETUGE: return ARMCC::HS;
986 case ISD::SETULT: return ARMCC::LO;
987 case ISD::SETULE: return ARMCC::LS;
991 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
992 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
993 ARMCC::CondCodes &CondCode2) {
994 CondCode2 = ARMCC::AL;
996 default: llvm_unreachable("Unknown FP condition!");
998 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1000 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1002 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1003 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1004 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1005 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1006 case ISD::SETO: CondCode = ARMCC::VC; break;
1007 case ISD::SETUO: CondCode = ARMCC::VS; break;
1008 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1009 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1010 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1012 case ISD::SETULT: CondCode = ARMCC::LT; break;
1014 case ISD::SETULE: CondCode = ARMCC::LE; break;
1016 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1020 //===----------------------------------------------------------------------===//
1021 // Calling Convention Implementation
1022 //===----------------------------------------------------------------------===//
1024 #include "ARMGenCallingConv.inc"
1026 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1027 /// given CallingConvention value.
1028 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1030 bool isVarArg) const {
1033 llvm_unreachable("Unsupported calling convention");
1034 case CallingConv::Fast:
1035 if (Subtarget->hasVFP2() && !isVarArg) {
1036 if (!Subtarget->isAAPCS_ABI())
1037 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1038 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1039 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1042 case CallingConv::C: {
1043 // Use target triple & subtarget features to do actual dispatch.
1044 if (!Subtarget->isAAPCS_ABI())
1045 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1046 else if (Subtarget->hasVFP2() &&
1047 FloatABIType == FloatABI::Hard && !isVarArg)
1048 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1049 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1051 case CallingConv::ARM_AAPCS_VFP:
1052 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1053 case CallingConv::ARM_AAPCS:
1054 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1055 case CallingConv::ARM_APCS:
1056 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1060 /// LowerCallResult - Lower the result values of a call into the
1061 /// appropriate copies out of appropriate physical registers.
1063 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1064 CallingConv::ID CallConv, bool isVarArg,
1065 const SmallVectorImpl<ISD::InputArg> &Ins,
1066 DebugLoc dl, SelectionDAG &DAG,
1067 SmallVectorImpl<SDValue> &InVals) const {
1069 // Assign locations to each value returned by this call.
1070 SmallVector<CCValAssign, 16> RVLocs;
1071 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1072 RVLocs, *DAG.getContext());
1073 CCInfo.AnalyzeCallResult(Ins,
1074 CCAssignFnForNode(CallConv, /* Return*/ true,
1077 // Copy all of the result registers out of their specified physreg.
1078 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1079 CCValAssign VA = RVLocs[i];
1082 if (VA.needsCustom()) {
1083 // Handle f64 or half of a v2f64.
1084 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1086 Chain = Lo.getValue(1);
1087 InFlag = Lo.getValue(2);
1088 VA = RVLocs[++i]; // skip ahead to next loc
1089 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1091 Chain = Hi.getValue(1);
1092 InFlag = Hi.getValue(2);
1093 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1095 if (VA.getLocVT() == MVT::v2f64) {
1096 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1097 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1098 DAG.getConstant(0, MVT::i32));
1100 VA = RVLocs[++i]; // skip ahead to next loc
1101 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1102 Chain = Lo.getValue(1);
1103 InFlag = Lo.getValue(2);
1104 VA = RVLocs[++i]; // skip ahead to next loc
1105 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1106 Chain = Hi.getValue(1);
1107 InFlag = Hi.getValue(2);
1108 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1109 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1110 DAG.getConstant(1, MVT::i32));
1113 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1115 Chain = Val.getValue(1);
1116 InFlag = Val.getValue(2);
1119 switch (VA.getLocInfo()) {
1120 default: llvm_unreachable("Unknown loc info!");
1121 case CCValAssign::Full: break;
1122 case CCValAssign::BCvt:
1123 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1127 InVals.push_back(Val);
1133 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1134 /// by "Src" to address "Dst" of size "Size". Alignment information is
1135 /// specified by the specific parameter attribute. The copy will be passed as
1136 /// a byval function parameter.
1137 /// Sometimes what we are copying is the end of a larger object, the part that
1138 /// does not fit in registers.
1140 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1141 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1143 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1144 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1145 /*isVolatile=*/false, /*AlwaysInline=*/false,
1146 MachinePointerInfo(0), MachinePointerInfo(0));
1149 /// LowerMemOpCallTo - Store the argument to the stack.
1151 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1152 SDValue StackPtr, SDValue Arg,
1153 DebugLoc dl, SelectionDAG &DAG,
1154 const CCValAssign &VA,
1155 ISD::ArgFlagsTy Flags) const {
1156 unsigned LocMemOffset = VA.getLocMemOffset();
1157 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1158 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1159 if (Flags.isByVal())
1160 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1162 return DAG.getStore(Chain, dl, Arg, PtrOff,
1163 MachinePointerInfo::getStack(LocMemOffset),
1167 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1168 SDValue Chain, SDValue &Arg,
1169 RegsToPassVector &RegsToPass,
1170 CCValAssign &VA, CCValAssign &NextVA,
1172 SmallVector<SDValue, 8> &MemOpChains,
1173 ISD::ArgFlagsTy Flags) const {
1175 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1176 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1177 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1179 if (NextVA.isRegLoc())
1180 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1182 assert(NextVA.isMemLoc());
1183 if (StackPtr.getNode() == 0)
1184 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1186 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1192 /// LowerCall - Lowering a call into a callseq_start <-
1193 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1196 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1197 CallingConv::ID CallConv, bool isVarArg,
1199 const SmallVectorImpl<ISD::OutputArg> &Outs,
1200 const SmallVectorImpl<SDValue> &OutVals,
1201 const SmallVectorImpl<ISD::InputArg> &Ins,
1202 DebugLoc dl, SelectionDAG &DAG,
1203 SmallVectorImpl<SDValue> &InVals) const {
1204 MachineFunction &MF = DAG.getMachineFunction();
1205 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1206 bool IsSibCall = false;
1207 // Temporarily disable tail calls so things don't break.
1208 if (!EnableARMTailCalls)
1211 // Check if it's really possible to do a tail call.
1212 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1213 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1214 Outs, OutVals, Ins, DAG);
1215 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1216 // detected sibcalls.
1223 // Analyze operands of the call, assigning locations to each operand.
1224 SmallVector<CCValAssign, 16> ArgLocs;
1225 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1227 CCInfo.AnalyzeCallOperands(Outs,
1228 CCAssignFnForNode(CallConv, /* Return*/ false,
1231 // Get a count of how many bytes are to be pushed on the stack.
1232 unsigned NumBytes = CCInfo.getNextStackOffset();
1234 // For tail calls, memory operands are available in our caller's stack.
1238 // Adjust the stack pointer for the new arguments...
1239 // These operations are automatically eliminated by the prolog/epilog pass
1241 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1243 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1245 RegsToPassVector RegsToPass;
1246 SmallVector<SDValue, 8> MemOpChains;
1248 // Walk the register/memloc assignments, inserting copies/loads. In the case
1249 // of tail call optimization, arguments are handled later.
1250 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1252 ++i, ++realArgIdx) {
1253 CCValAssign &VA = ArgLocs[i];
1254 SDValue Arg = OutVals[realArgIdx];
1255 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1257 // Promote the value if needed.
1258 switch (VA.getLocInfo()) {
1259 default: llvm_unreachable("Unknown loc info!");
1260 case CCValAssign::Full: break;
1261 case CCValAssign::SExt:
1262 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1264 case CCValAssign::ZExt:
1265 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1267 case CCValAssign::AExt:
1268 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1270 case CCValAssign::BCvt:
1271 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1275 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1276 if (VA.needsCustom()) {
1277 if (VA.getLocVT() == MVT::v2f64) {
1278 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1279 DAG.getConstant(0, MVT::i32));
1280 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1281 DAG.getConstant(1, MVT::i32));
1283 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1284 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1286 VA = ArgLocs[++i]; // skip ahead to next loc
1287 if (VA.isRegLoc()) {
1288 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1289 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1291 assert(VA.isMemLoc());
1293 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1294 dl, DAG, VA, Flags));
1297 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1298 StackPtr, MemOpChains, Flags);
1300 } else if (VA.isRegLoc()) {
1301 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1302 } else if (!IsSibCall) {
1303 assert(VA.isMemLoc());
1305 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1306 dl, DAG, VA, Flags));
1310 if (!MemOpChains.empty())
1311 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1312 &MemOpChains[0], MemOpChains.size());
1314 // Build a sequence of copy-to-reg nodes chained together with token chain
1315 // and flag operands which copy the outgoing args into the appropriate regs.
1317 // Tail call byval lowering might overwrite argument registers so in case of
1318 // tail call optimization the copies to registers are lowered later.
1320 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1321 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1322 RegsToPass[i].second, InFlag);
1323 InFlag = Chain.getValue(1);
1326 // For tail calls lower the arguments to the 'real' stack slot.
1328 // Force all the incoming stack arguments to be loaded from the stack
1329 // before any new outgoing arguments are stored to the stack, because the
1330 // outgoing stack slots may alias the incoming argument stack slots, and
1331 // the alias isn't otherwise explicit. This is slightly more conservative
1332 // than necessary, because it means that each store effectively depends
1333 // on every argument instead of just those arguments it would clobber.
1335 // Do not flag preceeding copytoreg stuff together with the following stuff.
1337 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1338 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1339 RegsToPass[i].second, InFlag);
1340 InFlag = Chain.getValue(1);
1345 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1346 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1347 // node so that legalize doesn't hack it.
1348 bool isDirect = false;
1349 bool isARMFunc = false;
1350 bool isLocalARMFunc = false;
1351 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1353 if (EnableARMLongCalls) {
1354 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1355 && "long-calls with non-static relocation model!");
1356 // Handle a global address or an external symbol. If it's not one of
1357 // those, the target's already in a register, so we don't need to do
1359 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1360 const GlobalValue *GV = G->getGlobal();
1361 // Create a constant pool entry for the callee address
1362 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1363 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1366 // Get the address of the callee into a register
1367 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1368 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1369 Callee = DAG.getLoad(getPointerTy(), dl,
1370 DAG.getEntryNode(), CPAddr,
1371 MachinePointerInfo::getConstantPool(),
1373 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1374 const char *Sym = S->getSymbol();
1376 // Create a constant pool entry for the callee address
1377 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1378 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1379 Sym, ARMPCLabelIndex, 0);
1380 // Get the address of the callee into a register
1381 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1382 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1383 Callee = DAG.getLoad(getPointerTy(), dl,
1384 DAG.getEntryNode(), CPAddr,
1385 MachinePointerInfo::getConstantPool(),
1388 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1389 const GlobalValue *GV = G->getGlobal();
1391 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1392 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1393 getTargetMachine().getRelocationModel() != Reloc::Static;
1394 isARMFunc = !Subtarget->isThumb() || isStub;
1395 // ARM call to a local ARM function is predicable.
1396 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1397 // tBX takes a register source operand.
1398 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1399 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1400 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1403 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1404 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1405 Callee = DAG.getLoad(getPointerTy(), dl,
1406 DAG.getEntryNode(), CPAddr,
1407 MachinePointerInfo::getConstantPool(),
1409 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1410 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1411 getPointerTy(), Callee, PICLabel);
1413 // On ELF targets for PIC code, direct calls should go through the PLT
1414 unsigned OpFlags = 0;
1415 if (Subtarget->isTargetELF() &&
1416 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1417 OpFlags = ARMII::MO_PLT;
1418 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1420 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1422 bool isStub = Subtarget->isTargetDarwin() &&
1423 getTargetMachine().getRelocationModel() != Reloc::Static;
1424 isARMFunc = !Subtarget->isThumb() || isStub;
1425 // tBX takes a register source operand.
1426 const char *Sym = S->getSymbol();
1427 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1428 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1429 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1430 Sym, ARMPCLabelIndex, 4);
1431 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1432 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1433 Callee = DAG.getLoad(getPointerTy(), dl,
1434 DAG.getEntryNode(), CPAddr,
1435 MachinePointerInfo::getConstantPool(),
1437 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1438 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1439 getPointerTy(), Callee, PICLabel);
1441 unsigned OpFlags = 0;
1442 // On ELF targets for PIC code, direct calls should go through the PLT
1443 if (Subtarget->isTargetELF() &&
1444 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1445 OpFlags = ARMII::MO_PLT;
1446 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1450 // FIXME: handle tail calls differently.
1452 if (Subtarget->isThumb()) {
1453 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1454 CallOpc = ARMISD::CALL_NOLINK;
1456 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1458 CallOpc = (isDirect || Subtarget->hasV5TOps())
1459 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1460 : ARMISD::CALL_NOLINK;
1463 std::vector<SDValue> Ops;
1464 Ops.push_back(Chain);
1465 Ops.push_back(Callee);
1467 // Add argument registers to the end of the list so that they are known live
1469 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1470 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1471 RegsToPass[i].second.getValueType()));
1473 if (InFlag.getNode())
1474 Ops.push_back(InFlag);
1476 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1478 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1480 // Returns a chain and a flag for retval copy to use.
1481 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1482 InFlag = Chain.getValue(1);
1484 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1485 DAG.getIntPtrConstant(0, true), InFlag);
1487 InFlag = Chain.getValue(1);
1489 // Handle result values, copying them out of physregs into vregs that we
1491 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1495 /// MatchingStackOffset - Return true if the given stack call argument is
1496 /// already available in the same position (relatively) of the caller's
1497 /// incoming argument stack.
1499 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1500 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1501 const ARMInstrInfo *TII) {
1502 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1504 if (Arg.getOpcode() == ISD::CopyFromReg) {
1505 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1506 if (!TargetRegisterInfo::isVirtualRegister(VR))
1508 MachineInstr *Def = MRI->getVRegDef(VR);
1511 if (!Flags.isByVal()) {
1512 if (!TII->isLoadFromStackSlot(Def, FI))
1517 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1518 if (Flags.isByVal())
1519 // ByVal argument is passed in as a pointer but it's now being
1520 // dereferenced. e.g.
1521 // define @foo(%struct.X* %A) {
1522 // tail call @bar(%struct.X* byval %A)
1525 SDValue Ptr = Ld->getBasePtr();
1526 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1529 FI = FINode->getIndex();
1533 assert(FI != INT_MAX);
1534 if (!MFI->isFixedObjectIndex(FI))
1536 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1539 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1540 /// for tail call optimization. Targets which want to do tail call
1541 /// optimization should implement this function.
1543 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1544 CallingConv::ID CalleeCC,
1546 bool isCalleeStructRet,
1547 bool isCallerStructRet,
1548 const SmallVectorImpl<ISD::OutputArg> &Outs,
1549 const SmallVectorImpl<SDValue> &OutVals,
1550 const SmallVectorImpl<ISD::InputArg> &Ins,
1551 SelectionDAG& DAG) const {
1552 const Function *CallerF = DAG.getMachineFunction().getFunction();
1553 CallingConv::ID CallerCC = CallerF->getCallingConv();
1554 bool CCMatch = CallerCC == CalleeCC;
1556 // Look for obvious safe cases to perform tail call optimization that do not
1557 // require ABI changes. This is what gcc calls sibcall.
1559 // Do not sibcall optimize vararg calls unless the call site is not passing
1561 if (isVarArg && !Outs.empty())
1564 // Also avoid sibcall optimization if either caller or callee uses struct
1565 // return semantics.
1566 if (isCalleeStructRet || isCallerStructRet)
1569 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1570 // emitEpilogue is not ready for them.
1571 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1572 // LR. This means if we need to reload LR, it takes an extra instructions,
1573 // which outweighs the value of the tail call; but here we don't know yet
1574 // whether LR is going to be used. Probably the right approach is to
1575 // generate the tail call here and turn it back into CALL/RET in
1576 // emitEpilogue if LR is used.
1578 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1579 // but we need to make sure there are enough registers; the only valid
1580 // registers are the 4 used for parameters. We don't currently do this
1582 if (Subtarget->isThumb1Only())
1585 // If the calling conventions do not match, then we'd better make sure the
1586 // results are returned in the same way as what the caller expects.
1588 SmallVector<CCValAssign, 16> RVLocs1;
1589 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1590 RVLocs1, *DAG.getContext());
1591 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1593 SmallVector<CCValAssign, 16> RVLocs2;
1594 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1595 RVLocs2, *DAG.getContext());
1596 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1598 if (RVLocs1.size() != RVLocs2.size())
1600 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1601 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1603 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1605 if (RVLocs1[i].isRegLoc()) {
1606 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1609 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1615 // If the callee takes no arguments then go on to check the results of the
1617 if (!Outs.empty()) {
1618 // Check if stack adjustment is needed. For now, do not do this if any
1619 // argument is passed on the stack.
1620 SmallVector<CCValAssign, 16> ArgLocs;
1621 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1622 ArgLocs, *DAG.getContext());
1623 CCInfo.AnalyzeCallOperands(Outs,
1624 CCAssignFnForNode(CalleeCC, false, isVarArg));
1625 if (CCInfo.getNextStackOffset()) {
1626 MachineFunction &MF = DAG.getMachineFunction();
1628 // Check if the arguments are already laid out in the right way as
1629 // the caller's fixed stack objects.
1630 MachineFrameInfo *MFI = MF.getFrameInfo();
1631 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1632 const ARMInstrInfo *TII =
1633 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1634 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1636 ++i, ++realArgIdx) {
1637 CCValAssign &VA = ArgLocs[i];
1638 EVT RegVT = VA.getLocVT();
1639 SDValue Arg = OutVals[realArgIdx];
1640 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1641 if (VA.getLocInfo() == CCValAssign::Indirect)
1643 if (VA.needsCustom()) {
1644 // f64 and vector types are split into multiple registers or
1645 // register/stack-slot combinations. The types will not match
1646 // the registers; give up on memory f64 refs until we figure
1647 // out what to do about this.
1650 if (!ArgLocs[++i].isRegLoc())
1652 if (RegVT == MVT::v2f64) {
1653 if (!ArgLocs[++i].isRegLoc())
1655 if (!ArgLocs[++i].isRegLoc())
1658 } else if (!VA.isRegLoc()) {
1659 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1671 ARMTargetLowering::LowerReturn(SDValue Chain,
1672 CallingConv::ID CallConv, bool isVarArg,
1673 const SmallVectorImpl<ISD::OutputArg> &Outs,
1674 const SmallVectorImpl<SDValue> &OutVals,
1675 DebugLoc dl, SelectionDAG &DAG) const {
1677 // CCValAssign - represent the assignment of the return value to a location.
1678 SmallVector<CCValAssign, 16> RVLocs;
1680 // CCState - Info about the registers and stack slots.
1681 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1684 // Analyze outgoing return values.
1685 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1688 // If this is the first return lowered for this function, add
1689 // the regs to the liveout set for the function.
1690 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1691 for (unsigned i = 0; i != RVLocs.size(); ++i)
1692 if (RVLocs[i].isRegLoc())
1693 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1698 // Copy the result values into the output registers.
1699 for (unsigned i = 0, realRVLocIdx = 0;
1701 ++i, ++realRVLocIdx) {
1702 CCValAssign &VA = RVLocs[i];
1703 assert(VA.isRegLoc() && "Can only return in registers!");
1705 SDValue Arg = OutVals[realRVLocIdx];
1707 switch (VA.getLocInfo()) {
1708 default: llvm_unreachable("Unknown loc info!");
1709 case CCValAssign::Full: break;
1710 case CCValAssign::BCvt:
1711 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1715 if (VA.needsCustom()) {
1716 if (VA.getLocVT() == MVT::v2f64) {
1717 // Extract the first half and return it in two registers.
1718 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1719 DAG.getConstant(0, MVT::i32));
1720 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1721 DAG.getVTList(MVT::i32, MVT::i32), Half);
1723 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1724 Flag = Chain.getValue(1);
1725 VA = RVLocs[++i]; // skip ahead to next loc
1726 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1727 HalfGPRs.getValue(1), Flag);
1728 Flag = Chain.getValue(1);
1729 VA = RVLocs[++i]; // skip ahead to next loc
1731 // Extract the 2nd half and fall through to handle it as an f64 value.
1732 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1733 DAG.getConstant(1, MVT::i32));
1735 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1737 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1738 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1739 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1740 Flag = Chain.getValue(1);
1741 VA = RVLocs[++i]; // skip ahead to next loc
1742 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1745 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1747 // Guarantee that all emitted copies are
1748 // stuck together, avoiding something bad.
1749 Flag = Chain.getValue(1);
1754 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1756 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1761 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1762 if (N->getNumValues() != 1)
1764 if (!N->hasNUsesOfValue(1, 0))
1767 unsigned NumCopies = 0;
1769 SDNode *Use = *N->use_begin();
1770 if (Use->getOpcode() == ISD::CopyToReg) {
1771 Copies[NumCopies++] = Use;
1772 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1773 // f64 returned in a pair of GPRs.
1774 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1776 if (UI->getOpcode() != ISD::CopyToReg)
1778 Copies[UI.getUse().getResNo()] = *UI;
1781 } else if (Use->getOpcode() == ISD::BITCAST) {
1782 // f32 returned in a single GPR.
1783 if (!Use->hasNUsesOfValue(1, 0))
1785 Use = *Use->use_begin();
1786 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1788 Copies[NumCopies++] = Use;
1793 if (NumCopies != 1 && NumCopies != 2)
1796 bool HasRet = false;
1797 for (unsigned i = 0; i < NumCopies; ++i) {
1798 SDNode *Copy = Copies[i];
1799 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1801 if (UI->getOpcode() == ISD::CopyToReg) {
1803 if (Use == Copies[0] || Use == Copies[1])
1807 if (UI->getOpcode() != ARMISD::RET_FLAG)
1816 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1817 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1818 // one of the above mentioned nodes. It has to be wrapped because otherwise
1819 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1820 // be used to form addressing mode. These wrapped nodes will be selected
1822 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1823 EVT PtrVT = Op.getValueType();
1824 // FIXME there is no actual debug info here
1825 DebugLoc dl = Op.getDebugLoc();
1826 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1828 if (CP->isMachineConstantPoolEntry())
1829 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1830 CP->getAlignment());
1832 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1833 CP->getAlignment());
1834 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1837 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1838 return MachineJumpTableInfo::EK_Inline;
1841 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1842 SelectionDAG &DAG) const {
1843 MachineFunction &MF = DAG.getMachineFunction();
1844 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1845 unsigned ARMPCLabelIndex = 0;
1846 DebugLoc DL = Op.getDebugLoc();
1847 EVT PtrVT = getPointerTy();
1848 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1849 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1851 if (RelocM == Reloc::Static) {
1852 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1854 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1855 ARMPCLabelIndex = AFI->createPICLabelUId();
1856 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1857 ARMCP::CPBlockAddress,
1859 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1861 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1862 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1863 MachinePointerInfo::getConstantPool(),
1865 if (RelocM == Reloc::Static)
1867 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1868 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1871 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1873 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1874 SelectionDAG &DAG) const {
1875 DebugLoc dl = GA->getDebugLoc();
1876 EVT PtrVT = getPointerTy();
1877 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1878 MachineFunction &MF = DAG.getMachineFunction();
1879 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1880 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1881 ARMConstantPoolValue *CPV =
1882 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1883 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1884 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1885 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1886 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1887 MachinePointerInfo::getConstantPool(),
1889 SDValue Chain = Argument.getValue(1);
1891 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1892 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1894 // call __tls_get_addr.
1897 Entry.Node = Argument;
1898 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1899 Args.push_back(Entry);
1900 // FIXME: is there useful debug info available here?
1901 std::pair<SDValue, SDValue> CallResult =
1902 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1903 false, false, false, false,
1904 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1905 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1906 return CallResult.first;
1909 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1910 // "local exec" model.
1912 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1913 SelectionDAG &DAG) const {
1914 const GlobalValue *GV = GA->getGlobal();
1915 DebugLoc dl = GA->getDebugLoc();
1917 SDValue Chain = DAG.getEntryNode();
1918 EVT PtrVT = getPointerTy();
1919 // Get the Thread Pointer
1920 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1922 if (GV->isDeclaration()) {
1923 MachineFunction &MF = DAG.getMachineFunction();
1924 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1925 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1926 // Initial exec model.
1927 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1928 ARMConstantPoolValue *CPV =
1929 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1930 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1931 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1932 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1933 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1934 MachinePointerInfo::getConstantPool(),
1936 Chain = Offset.getValue(1);
1938 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1939 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1941 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1942 MachinePointerInfo::getConstantPool(),
1946 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
1947 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1948 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1949 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1950 MachinePointerInfo::getConstantPool(),
1954 // The address of the thread local variable is the add of the thread
1955 // pointer with the offset of the variable.
1956 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1960 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1961 // TODO: implement the "local dynamic" model
1962 assert(Subtarget->isTargetELF() &&
1963 "TLS not implemented for non-ELF targets");
1964 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1965 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1966 // otherwise use the "Local Exec" TLS Model
1967 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1968 return LowerToTLSGeneralDynamicModel(GA, DAG);
1970 return LowerToTLSExecModels(GA, DAG);
1973 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1974 SelectionDAG &DAG) const {
1975 EVT PtrVT = getPointerTy();
1976 DebugLoc dl = Op.getDebugLoc();
1977 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1978 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1979 if (RelocM == Reloc::PIC_) {
1980 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1981 ARMConstantPoolValue *CPV =
1982 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
1983 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1984 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1985 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1987 MachinePointerInfo::getConstantPool(),
1989 SDValue Chain = Result.getValue(1);
1990 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1991 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1993 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1994 MachinePointerInfo::getGOT(), false, false, 0);
1998 // If we have T2 ops, we can materialize the address directly via movt/movw
1999 // pair. This is always cheaper.
2000 if (Subtarget->useMovt()) {
2002 // FIXME: Once remat is capable of dealing with instructions with register
2003 // operands, expand this into two nodes.
2004 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2005 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2007 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2008 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2009 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2010 MachinePointerInfo::getConstantPool(),
2015 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2016 SelectionDAG &DAG) const {
2017 EVT PtrVT = getPointerTy();
2018 DebugLoc dl = Op.getDebugLoc();
2019 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2020 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2021 MachineFunction &MF = DAG.getMachineFunction();
2022 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2024 if (Subtarget->useMovt()) {
2026 // FIXME: Once remat is capable of dealing with instructions with register
2027 // operands, expand this into two nodes.
2028 if (RelocM == Reloc::Static)
2029 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2030 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2032 unsigned Wrapper = (RelocM == Reloc::PIC_)
2033 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2034 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2035 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2036 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2037 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2038 MachinePointerInfo::getGOT(), false, false, 0);
2042 unsigned ARMPCLabelIndex = 0;
2044 if (RelocM == Reloc::Static) {
2045 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2047 ARMPCLabelIndex = AFI->createPICLabelUId();
2048 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2049 ARMConstantPoolValue *CPV =
2050 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2051 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2053 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2055 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2056 MachinePointerInfo::getConstantPool(),
2058 SDValue Chain = Result.getValue(1);
2060 if (RelocM == Reloc::PIC_) {
2061 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2062 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2065 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2066 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2072 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2073 SelectionDAG &DAG) const {
2074 assert(Subtarget->isTargetELF() &&
2075 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2076 MachineFunction &MF = DAG.getMachineFunction();
2077 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2078 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2079 EVT PtrVT = getPointerTy();
2080 DebugLoc dl = Op.getDebugLoc();
2081 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2082 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2083 "_GLOBAL_OFFSET_TABLE_",
2084 ARMPCLabelIndex, PCAdj);
2085 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2086 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2087 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2088 MachinePointerInfo::getConstantPool(),
2090 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2091 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2095 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2097 DebugLoc dl = Op.getDebugLoc();
2098 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2099 Op.getOperand(0), Op.getOperand(1));
2103 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2104 DebugLoc dl = Op.getDebugLoc();
2105 SDValue Val = DAG.getConstant(0, MVT::i32);
2106 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2107 Op.getOperand(1), Val);
2111 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2112 DebugLoc dl = Op.getDebugLoc();
2113 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2114 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2118 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2119 const ARMSubtarget *Subtarget) const {
2120 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2121 DebugLoc dl = Op.getDebugLoc();
2123 default: return SDValue(); // Don't custom lower most intrinsics.
2124 case Intrinsic::arm_thread_pointer: {
2125 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2126 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2128 case Intrinsic::eh_sjlj_lsda: {
2129 MachineFunction &MF = DAG.getMachineFunction();
2130 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2131 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2132 EVT PtrVT = getPointerTy();
2133 DebugLoc dl = Op.getDebugLoc();
2134 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2136 unsigned PCAdj = (RelocM != Reloc::PIC_)
2137 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2138 ARMConstantPoolValue *CPV =
2139 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2140 ARMCP::CPLSDA, PCAdj);
2141 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2142 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2144 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2145 MachinePointerInfo::getConstantPool(),
2148 if (RelocM == Reloc::PIC_) {
2149 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2150 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2157 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2158 const ARMSubtarget *Subtarget) {
2159 DebugLoc dl = Op.getDebugLoc();
2160 if (!Subtarget->hasDataBarrier()) {
2161 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2162 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2164 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2165 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2166 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2167 DAG.getConstant(0, MVT::i32));
2170 SDValue Op5 = Op.getOperand(5);
2171 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2172 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2173 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2174 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2176 ARM_MB::MemBOpt DMBOpt;
2177 if (isDeviceBarrier)
2178 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2180 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2181 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2182 DAG.getConstant(DMBOpt, MVT::i32));
2185 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2186 const ARMSubtarget *Subtarget) {
2187 // ARM pre v5TE and Thumb1 does not have preload instructions.
2188 if (!(Subtarget->isThumb2() ||
2189 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2190 // Just preserve the chain.
2191 return Op.getOperand(0);
2193 DebugLoc dl = Op.getDebugLoc();
2194 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2196 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2197 // ARMv7 with MP extension has PLDW.
2198 return Op.getOperand(0);
2200 if (Subtarget->isThumb())
2202 isRead = ~isRead & 1;
2203 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2205 // Currently there is no intrinsic that matches pli.
2206 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2207 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2208 DAG.getConstant(isData, MVT::i32));
2211 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2212 MachineFunction &MF = DAG.getMachineFunction();
2213 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2215 // vastart just stores the address of the VarArgsFrameIndex slot into the
2216 // memory location argument.
2217 DebugLoc dl = Op.getDebugLoc();
2218 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2219 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2220 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2221 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2222 MachinePointerInfo(SV), false, false, 0);
2226 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2227 SDValue &Root, SelectionDAG &DAG,
2228 DebugLoc dl) const {
2229 MachineFunction &MF = DAG.getMachineFunction();
2230 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2232 TargetRegisterClass *RC;
2233 if (AFI->isThumb1OnlyFunction())
2234 RC = ARM::tGPRRegisterClass;
2236 RC = ARM::GPRRegisterClass;
2238 // Transform the arguments stored in physical registers into virtual ones.
2239 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2240 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2243 if (NextVA.isMemLoc()) {
2244 MachineFrameInfo *MFI = MF.getFrameInfo();
2245 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2247 // Create load node to retrieve arguments from the stack.
2248 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2249 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2250 MachinePointerInfo::getFixedStack(FI),
2253 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2254 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2257 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2261 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2262 CallingConv::ID CallConv, bool isVarArg,
2263 const SmallVectorImpl<ISD::InputArg>
2265 DebugLoc dl, SelectionDAG &DAG,
2266 SmallVectorImpl<SDValue> &InVals)
2269 MachineFunction &MF = DAG.getMachineFunction();
2270 MachineFrameInfo *MFI = MF.getFrameInfo();
2272 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2274 // Assign locations to all of the incoming arguments.
2275 SmallVector<CCValAssign, 16> ArgLocs;
2276 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2278 CCInfo.AnalyzeFormalArguments(Ins,
2279 CCAssignFnForNode(CallConv, /* Return*/ false,
2282 SmallVector<SDValue, 16> ArgValues;
2284 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2285 CCValAssign &VA = ArgLocs[i];
2287 // Arguments stored in registers.
2288 if (VA.isRegLoc()) {
2289 EVT RegVT = VA.getLocVT();
2292 if (VA.needsCustom()) {
2293 // f64 and vector types are split up into multiple registers or
2294 // combinations of registers and stack slots.
2295 if (VA.getLocVT() == MVT::v2f64) {
2296 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2298 VA = ArgLocs[++i]; // skip ahead to next loc
2300 if (VA.isMemLoc()) {
2301 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2302 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2303 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2304 MachinePointerInfo::getFixedStack(FI),
2307 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2310 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2311 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2312 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2313 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2314 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2316 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2319 TargetRegisterClass *RC;
2321 if (RegVT == MVT::f32)
2322 RC = ARM::SPRRegisterClass;
2323 else if (RegVT == MVT::f64)
2324 RC = ARM::DPRRegisterClass;
2325 else if (RegVT == MVT::v2f64)
2326 RC = ARM::QPRRegisterClass;
2327 else if (RegVT == MVT::i32)
2328 RC = (AFI->isThumb1OnlyFunction() ?
2329 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2331 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2333 // Transform the arguments in physical registers into virtual ones.
2334 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2335 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2338 // If this is an 8 or 16-bit value, it is really passed promoted
2339 // to 32 bits. Insert an assert[sz]ext to capture this, then
2340 // truncate to the right size.
2341 switch (VA.getLocInfo()) {
2342 default: llvm_unreachable("Unknown loc info!");
2343 case CCValAssign::Full: break;
2344 case CCValAssign::BCvt:
2345 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2347 case CCValAssign::SExt:
2348 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2349 DAG.getValueType(VA.getValVT()));
2350 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2352 case CCValAssign::ZExt:
2353 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2354 DAG.getValueType(VA.getValVT()));
2355 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2359 InVals.push_back(ArgValue);
2361 } else { // VA.isRegLoc()
2364 assert(VA.isMemLoc());
2365 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2367 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2368 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2370 // Create load nodes to retrieve arguments from the stack.
2371 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2372 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2373 MachinePointerInfo::getFixedStack(FI),
2380 static const unsigned GPRArgRegs[] = {
2381 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2384 unsigned NumGPRs = CCInfo.getFirstUnallocated
2385 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2387 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2388 unsigned VARegSize = (4 - NumGPRs) * 4;
2389 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2390 unsigned ArgOffset = CCInfo.getNextStackOffset();
2391 if (VARegSaveSize) {
2392 // If this function is vararg, store any remaining integer argument regs
2393 // to their spots on the stack so that they may be loaded by deferencing
2394 // the result of va_next.
2395 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2396 AFI->setVarArgsFrameIndex(
2397 MFI->CreateFixedObject(VARegSaveSize,
2398 ArgOffset + VARegSaveSize - VARegSize,
2400 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2403 SmallVector<SDValue, 4> MemOps;
2404 for (; NumGPRs < 4; ++NumGPRs) {
2405 TargetRegisterClass *RC;
2406 if (AFI->isThumb1OnlyFunction())
2407 RC = ARM::tGPRRegisterClass;
2409 RC = ARM::GPRRegisterClass;
2411 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2412 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2414 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2415 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2417 MemOps.push_back(Store);
2418 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2419 DAG.getConstant(4, getPointerTy()));
2421 if (!MemOps.empty())
2422 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2423 &MemOps[0], MemOps.size());
2425 // This will point to the next argument passed via stack.
2426 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2432 /// isFloatingPointZero - Return true if this is +0.0.
2433 static bool isFloatingPointZero(SDValue Op) {
2434 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2435 return CFP->getValueAPF().isPosZero();
2436 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2437 // Maybe this has already been legalized into the constant pool?
2438 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2439 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2440 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2441 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2442 return CFP->getValueAPF().isPosZero();
2448 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2449 /// the given operands.
2451 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2452 SDValue &ARMcc, SelectionDAG &DAG,
2453 DebugLoc dl) const {
2454 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2455 unsigned C = RHSC->getZExtValue();
2456 if (!isLegalICmpImmediate(C)) {
2457 // Constant does not fit, try adjusting it by one?
2462 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2463 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2464 RHS = DAG.getConstant(C-1, MVT::i32);
2469 if (C != 0 && isLegalICmpImmediate(C-1)) {
2470 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2471 RHS = DAG.getConstant(C-1, MVT::i32);
2476 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2477 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2478 RHS = DAG.getConstant(C+1, MVT::i32);
2483 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2484 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2485 RHS = DAG.getConstant(C+1, MVT::i32);
2492 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2493 ARMISD::NodeType CompareType;
2496 CompareType = ARMISD::CMP;
2501 CompareType = ARMISD::CMPZ;
2504 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2505 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2508 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2510 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2511 DebugLoc dl) const {
2513 if (!isFloatingPointZero(RHS))
2514 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2516 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2517 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2520 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2521 SDValue Cond = Op.getOperand(0);
2522 SDValue SelectTrue = Op.getOperand(1);
2523 SDValue SelectFalse = Op.getOperand(2);
2524 DebugLoc dl = Op.getDebugLoc();
2528 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2529 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2531 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2532 const ConstantSDNode *CMOVTrue =
2533 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2534 const ConstantSDNode *CMOVFalse =
2535 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2537 if (CMOVTrue && CMOVFalse) {
2538 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2539 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2543 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2545 False = SelectFalse;
2546 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2551 if (True.getNode() && False.getNode()) {
2552 EVT VT = Cond.getValueType();
2553 SDValue ARMcc = Cond.getOperand(2);
2554 SDValue CCR = Cond.getOperand(3);
2555 SDValue Cmp = Cond.getOperand(4);
2556 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2561 return DAG.getSelectCC(dl, Cond,
2562 DAG.getConstant(0, Cond.getValueType()),
2563 SelectTrue, SelectFalse, ISD::SETNE);
2566 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2567 EVT VT = Op.getValueType();
2568 SDValue LHS = Op.getOperand(0);
2569 SDValue RHS = Op.getOperand(1);
2570 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2571 SDValue TrueVal = Op.getOperand(2);
2572 SDValue FalseVal = Op.getOperand(3);
2573 DebugLoc dl = Op.getDebugLoc();
2575 if (LHS.getValueType() == MVT::i32) {
2577 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2578 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2579 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2582 ARMCC::CondCodes CondCode, CondCode2;
2583 FPCCToARMCC(CC, CondCode, CondCode2);
2585 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2586 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2587 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2588 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2590 if (CondCode2 != ARMCC::AL) {
2591 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2592 // FIXME: Needs another CMP because flag can have but one use.
2593 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2594 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2595 Result, TrueVal, ARMcc2, CCR, Cmp2);
2600 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2601 /// to morph to an integer compare sequence.
2602 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2603 const ARMSubtarget *Subtarget) {
2604 SDNode *N = Op.getNode();
2605 if (!N->hasOneUse())
2606 // Otherwise it requires moving the value from fp to integer registers.
2608 if (!N->getNumValues())
2610 EVT VT = Op.getValueType();
2611 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2612 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2613 // vmrs are very slow, e.g. cortex-a8.
2616 if (isFloatingPointZero(Op)) {
2620 return ISD::isNormalLoad(N);
2623 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2624 if (isFloatingPointZero(Op))
2625 return DAG.getConstant(0, MVT::i32);
2627 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2628 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2629 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2630 Ld->isVolatile(), Ld->isNonTemporal(),
2631 Ld->getAlignment());
2633 llvm_unreachable("Unknown VFP cmp argument!");
2636 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2637 SDValue &RetVal1, SDValue &RetVal2) {
2638 if (isFloatingPointZero(Op)) {
2639 RetVal1 = DAG.getConstant(0, MVT::i32);
2640 RetVal2 = DAG.getConstant(0, MVT::i32);
2644 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2645 SDValue Ptr = Ld->getBasePtr();
2646 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2647 Ld->getChain(), Ptr,
2648 Ld->getPointerInfo(),
2649 Ld->isVolatile(), Ld->isNonTemporal(),
2650 Ld->getAlignment());
2652 EVT PtrType = Ptr.getValueType();
2653 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2654 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2655 PtrType, Ptr, DAG.getConstant(4, PtrType));
2656 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2657 Ld->getChain(), NewPtr,
2658 Ld->getPointerInfo().getWithOffset(4),
2659 Ld->isVolatile(), Ld->isNonTemporal(),
2664 llvm_unreachable("Unknown VFP cmp argument!");
2667 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2668 /// f32 and even f64 comparisons to integer ones.
2670 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2671 SDValue Chain = Op.getOperand(0);
2672 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2673 SDValue LHS = Op.getOperand(2);
2674 SDValue RHS = Op.getOperand(3);
2675 SDValue Dest = Op.getOperand(4);
2676 DebugLoc dl = Op.getDebugLoc();
2678 bool SeenZero = false;
2679 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2680 canChangeToInt(RHS, SeenZero, Subtarget) &&
2681 // If one of the operand is zero, it's safe to ignore the NaN case since
2682 // we only care about equality comparisons.
2683 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2684 // If unsafe fp math optimization is enabled and there are no othter uses of
2685 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2686 // to an integer comparison.
2687 if (CC == ISD::SETOEQ)
2689 else if (CC == ISD::SETUNE)
2693 if (LHS.getValueType() == MVT::f32) {
2694 LHS = bitcastf32Toi32(LHS, DAG);
2695 RHS = bitcastf32Toi32(RHS, DAG);
2696 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2697 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2698 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2699 Chain, Dest, ARMcc, CCR, Cmp);
2704 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2705 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2706 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2707 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2708 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2709 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2710 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2716 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2717 SDValue Chain = Op.getOperand(0);
2718 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2719 SDValue LHS = Op.getOperand(2);
2720 SDValue RHS = Op.getOperand(3);
2721 SDValue Dest = Op.getOperand(4);
2722 DebugLoc dl = Op.getDebugLoc();
2724 if (LHS.getValueType() == MVT::i32) {
2726 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2727 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2728 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2729 Chain, Dest, ARMcc, CCR, Cmp);
2732 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2735 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2736 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2737 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2738 if (Result.getNode())
2742 ARMCC::CondCodes CondCode, CondCode2;
2743 FPCCToARMCC(CC, CondCode, CondCode2);
2745 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2746 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2747 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2748 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2749 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2750 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2751 if (CondCode2 != ARMCC::AL) {
2752 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2753 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2754 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2759 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2760 SDValue Chain = Op.getOperand(0);
2761 SDValue Table = Op.getOperand(1);
2762 SDValue Index = Op.getOperand(2);
2763 DebugLoc dl = Op.getDebugLoc();
2765 EVT PTy = getPointerTy();
2766 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2767 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2768 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2769 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2770 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2771 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2772 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2773 if (Subtarget->isThumb2()) {
2774 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2775 // which does another jump to the destination. This also makes it easier
2776 // to translate it to TBB / TBH later.
2777 // FIXME: This might not work if the function is extremely large.
2778 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2779 Addr, Op.getOperand(2), JTI, UId);
2781 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2782 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2783 MachinePointerInfo::getJumpTable(),
2785 Chain = Addr.getValue(1);
2786 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2787 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2789 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2790 MachinePointerInfo::getJumpTable(), false, false, 0);
2791 Chain = Addr.getValue(1);
2792 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2796 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2797 DebugLoc dl = Op.getDebugLoc();
2800 switch (Op.getOpcode()) {
2802 assert(0 && "Invalid opcode!");
2803 case ISD::FP_TO_SINT:
2804 Opc = ARMISD::FTOSI;
2806 case ISD::FP_TO_UINT:
2807 Opc = ARMISD::FTOUI;
2810 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2811 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2814 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2815 EVT VT = Op.getValueType();
2816 DebugLoc dl = Op.getDebugLoc();
2819 switch (Op.getOpcode()) {
2821 assert(0 && "Invalid opcode!");
2822 case ISD::SINT_TO_FP:
2823 Opc = ARMISD::SITOF;
2825 case ISD::UINT_TO_FP:
2826 Opc = ARMISD::UITOF;
2830 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
2831 return DAG.getNode(Opc, dl, VT, Op);
2834 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2835 // Implement fcopysign with a fabs and a conditional fneg.
2836 SDValue Tmp0 = Op.getOperand(0);
2837 SDValue Tmp1 = Op.getOperand(1);
2838 DebugLoc dl = Op.getDebugLoc();
2839 EVT VT = Op.getValueType();
2840 EVT SrcVT = Tmp1.getValueType();
2841 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
2842 Tmp0.getOpcode() == ARMISD::VMOVDRR;
2843 bool UseNEON = !InGPR && Subtarget->hasNEON();
2846 // Use VBSL to copy the sign bit.
2847 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
2848 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
2849 DAG.getTargetConstant(EncodedVal, MVT::i32));
2850 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
2852 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
2853 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
2854 DAG.getConstant(32, MVT::i32));
2855 else /*if (VT == MVT::f32)*/
2856 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
2857 if (SrcVT == MVT::f32) {
2858 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
2860 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
2861 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
2862 DAG.getConstant(32, MVT::i32));
2864 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
2865 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
2867 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
2869 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
2870 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
2871 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
2873 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
2874 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
2875 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
2876 if (SrcVT == MVT::f32) {
2877 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
2878 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
2879 DAG.getConstant(0, MVT::i32));
2881 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
2887 // Bitcast operand 1 to i32.
2888 if (SrcVT == MVT::f64)
2889 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
2890 &Tmp1, 1).getValue(1);
2891 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
2893 // Or in the signbit with integer operations.
2894 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
2895 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
2896 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
2897 if (VT == MVT::f32) {
2898 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
2899 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
2900 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
2901 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
2904 // f64: Or the high part with signbit and then combine two parts.
2905 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
2907 SDValue Lo = Tmp0.getValue(0);
2908 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
2909 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
2910 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2913 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2914 MachineFunction &MF = DAG.getMachineFunction();
2915 MachineFrameInfo *MFI = MF.getFrameInfo();
2916 MFI->setReturnAddressIsTaken(true);
2918 EVT VT = Op.getValueType();
2919 DebugLoc dl = Op.getDebugLoc();
2920 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2922 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2923 SDValue Offset = DAG.getConstant(4, MVT::i32);
2924 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2925 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2926 MachinePointerInfo(), false, false, 0);
2929 // Return LR, which contains the return address. Mark it an implicit live-in.
2930 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2931 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2934 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2935 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2936 MFI->setFrameAddressIsTaken(true);
2938 EVT VT = Op.getValueType();
2939 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2940 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2941 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2942 ? ARM::R7 : ARM::R11;
2943 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2945 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2946 MachinePointerInfo(),
2951 /// ExpandBITCAST - If the target supports VFP, this function is called to
2952 /// expand a bit convert where either the source or destination type is i64 to
2953 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2954 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2955 /// vectors), since the legalizer won't know what to do with that.
2956 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
2957 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2958 DebugLoc dl = N->getDebugLoc();
2959 SDValue Op = N->getOperand(0);
2961 // This function is only supposed to be called for i64 types, either as the
2962 // source or destination of the bit convert.
2963 EVT SrcVT = Op.getValueType();
2964 EVT DstVT = N->getValueType(0);
2965 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2966 "ExpandBITCAST called for non-i64 type");
2968 // Turn i64->f64 into VMOVDRR.
2969 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2970 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2971 DAG.getConstant(0, MVT::i32));
2972 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2973 DAG.getConstant(1, MVT::i32));
2974 return DAG.getNode(ISD::BITCAST, dl, DstVT,
2975 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2978 // Turn f64->i64 into VMOVRRD.
2979 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2980 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2981 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2982 // Merge the pieces into a single i64 value.
2983 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2989 /// getZeroVector - Returns a vector of specified type with all zero elements.
2990 /// Zero vectors are used to represent vector negation and in those cases
2991 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2992 /// not support i64 elements, so sometimes the zero vectors will need to be
2993 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2995 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2996 assert(VT.isVector() && "Expected a vector type");
2997 // The canonical modified immediate encoding of a zero vector is....0!
2998 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2999 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3000 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3001 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3004 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3005 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3006 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3007 SelectionDAG &DAG) const {
3008 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3009 EVT VT = Op.getValueType();
3010 unsigned VTBits = VT.getSizeInBits();
3011 DebugLoc dl = Op.getDebugLoc();
3012 SDValue ShOpLo = Op.getOperand(0);
3013 SDValue ShOpHi = Op.getOperand(1);
3014 SDValue ShAmt = Op.getOperand(2);
3016 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3018 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3020 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3021 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3022 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3023 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3024 DAG.getConstant(VTBits, MVT::i32));
3025 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3026 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3027 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3029 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3030 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3032 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3033 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3036 SDValue Ops[2] = { Lo, Hi };
3037 return DAG.getMergeValues(Ops, 2, dl);
3040 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3041 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3042 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3043 SelectionDAG &DAG) const {
3044 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3045 EVT VT = Op.getValueType();
3046 unsigned VTBits = VT.getSizeInBits();
3047 DebugLoc dl = Op.getDebugLoc();
3048 SDValue ShOpLo = Op.getOperand(0);
3049 SDValue ShOpHi = Op.getOperand(1);
3050 SDValue ShAmt = Op.getOperand(2);
3053 assert(Op.getOpcode() == ISD::SHL_PARTS);
3054 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3055 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3056 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3057 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3058 DAG.getConstant(VTBits, MVT::i32));
3059 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3060 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3062 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3063 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3064 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3066 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3067 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3070 SDValue Ops[2] = { Lo, Hi };
3071 return DAG.getMergeValues(Ops, 2, dl);
3074 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3075 SelectionDAG &DAG) const {
3076 // The rounding mode is in bits 23:22 of the FPSCR.
3077 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3078 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3079 // so that the shift + and get folded into a bitfield extract.
3080 DebugLoc dl = Op.getDebugLoc();
3081 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3082 DAG.getConstant(Intrinsic::arm_get_fpscr,
3084 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3085 DAG.getConstant(1U << 22, MVT::i32));
3086 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3087 DAG.getConstant(22, MVT::i32));
3088 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3089 DAG.getConstant(3, MVT::i32));
3092 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3093 const ARMSubtarget *ST) {
3094 EVT VT = N->getValueType(0);
3095 DebugLoc dl = N->getDebugLoc();
3097 if (!ST->hasV6T2Ops())
3100 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3101 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3104 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3105 const ARMSubtarget *ST) {
3106 EVT VT = N->getValueType(0);
3107 DebugLoc dl = N->getDebugLoc();
3112 // Lower vector shifts on NEON to use VSHL.
3113 assert(ST->hasNEON() && "unexpected vector shift");
3115 // Left shifts translate directly to the vshiftu intrinsic.
3116 if (N->getOpcode() == ISD::SHL)
3117 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3118 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3119 N->getOperand(0), N->getOperand(1));
3121 assert((N->getOpcode() == ISD::SRA ||
3122 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3124 // NEON uses the same intrinsics for both left and right shifts. For
3125 // right shifts, the shift amounts are negative, so negate the vector of
3127 EVT ShiftVT = N->getOperand(1).getValueType();
3128 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3129 getZeroVector(ShiftVT, DAG, dl),
3131 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3132 Intrinsic::arm_neon_vshifts :
3133 Intrinsic::arm_neon_vshiftu);
3134 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3135 DAG.getConstant(vshiftInt, MVT::i32),
3136 N->getOperand(0), NegatedCount);
3139 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3140 const ARMSubtarget *ST) {
3141 EVT VT = N->getValueType(0);
3142 DebugLoc dl = N->getDebugLoc();
3144 // We can get here for a node like i32 = ISD::SHL i32, i64
3148 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3149 "Unknown shift to lower!");
3151 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3152 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3153 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3156 // If we are in thumb mode, we don't have RRX.
3157 if (ST->isThumb1Only()) return SDValue();
3159 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3160 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3161 DAG.getConstant(0, MVT::i32));
3162 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3163 DAG.getConstant(1, MVT::i32));
3165 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3166 // captures the result into a carry flag.
3167 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3168 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3170 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3171 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3173 // Merge the pieces into a single i64 value.
3174 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3177 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3178 SDValue TmpOp0, TmpOp1;
3179 bool Invert = false;
3183 SDValue Op0 = Op.getOperand(0);
3184 SDValue Op1 = Op.getOperand(1);
3185 SDValue CC = Op.getOperand(2);
3186 EVT VT = Op.getValueType();
3187 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3188 DebugLoc dl = Op.getDebugLoc();
3190 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3191 switch (SetCCOpcode) {
3192 default: llvm_unreachable("Illegal FP comparison"); break;
3194 case ISD::SETNE: Invert = true; // Fallthrough
3196 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3198 case ISD::SETLT: Swap = true; // Fallthrough
3200 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3202 case ISD::SETLE: Swap = true; // Fallthrough
3204 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3205 case ISD::SETUGE: Swap = true; // Fallthrough
3206 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3207 case ISD::SETUGT: Swap = true; // Fallthrough
3208 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3209 case ISD::SETUEQ: Invert = true; // Fallthrough
3211 // Expand this to (OLT | OGT).
3215 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3216 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3218 case ISD::SETUO: Invert = true; // Fallthrough
3220 // Expand this to (OLT | OGE).
3224 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3225 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3229 // Integer comparisons.
3230 switch (SetCCOpcode) {
3231 default: llvm_unreachable("Illegal integer comparison"); break;
3232 case ISD::SETNE: Invert = true;
3233 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3234 case ISD::SETLT: Swap = true;
3235 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3236 case ISD::SETLE: Swap = true;
3237 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3238 case ISD::SETULT: Swap = true;
3239 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3240 case ISD::SETULE: Swap = true;
3241 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3244 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3245 if (Opc == ARMISD::VCEQ) {
3248 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3250 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3253 // Ignore bitconvert.
3254 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3255 AndOp = AndOp.getOperand(0);
3257 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3259 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3260 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3267 std::swap(Op0, Op1);
3269 // If one of the operands is a constant vector zero, attempt to fold the
3270 // comparison to a specialized compare-against-zero form.
3272 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3274 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3275 if (Opc == ARMISD::VCGE)
3276 Opc = ARMISD::VCLEZ;
3277 else if (Opc == ARMISD::VCGT)
3278 Opc = ARMISD::VCLTZ;
3283 if (SingleOp.getNode()) {
3286 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3288 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3290 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3292 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3294 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3296 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3299 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3303 Result = DAG.getNOT(dl, Result, VT);
3308 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3309 /// valid vector constant for a NEON instruction with a "modified immediate"
3310 /// operand (e.g., VMOV). If so, return the encoded value.
3311 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3312 unsigned SplatBitSize, SelectionDAG &DAG,
3313 EVT &VT, bool is128Bits, NEONModImmType type) {
3314 unsigned OpCmode, Imm;
3316 // SplatBitSize is set to the smallest size that splats the vector, so a
3317 // zero vector will always have SplatBitSize == 8. However, NEON modified
3318 // immediate instructions others than VMOV do not support the 8-bit encoding
3319 // of a zero vector, and the default encoding of zero is supposed to be the
3324 switch (SplatBitSize) {
3326 if (type != VMOVModImm)
3328 // Any 1-byte value is OK. Op=0, Cmode=1110.
3329 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3332 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3336 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3337 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3338 if ((SplatBits & ~0xff) == 0) {
3339 // Value = 0x00nn: Op=x, Cmode=100x.
3344 if ((SplatBits & ~0xff00) == 0) {
3345 // Value = 0xnn00: Op=x, Cmode=101x.
3347 Imm = SplatBits >> 8;
3353 // NEON's 32-bit VMOV supports splat values where:
3354 // * only one byte is nonzero, or
3355 // * the least significant byte is 0xff and the second byte is nonzero, or
3356 // * the least significant 2 bytes are 0xff and the third is nonzero.
3357 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3358 if ((SplatBits & ~0xff) == 0) {
3359 // Value = 0x000000nn: Op=x, Cmode=000x.
3364 if ((SplatBits & ~0xff00) == 0) {
3365 // Value = 0x0000nn00: Op=x, Cmode=001x.
3367 Imm = SplatBits >> 8;
3370 if ((SplatBits & ~0xff0000) == 0) {
3371 // Value = 0x00nn0000: Op=x, Cmode=010x.
3373 Imm = SplatBits >> 16;
3376 if ((SplatBits & ~0xff000000) == 0) {
3377 // Value = 0xnn000000: Op=x, Cmode=011x.
3379 Imm = SplatBits >> 24;
3383 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3384 if (type == OtherModImm) return SDValue();
3386 if ((SplatBits & ~0xffff) == 0 &&
3387 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3388 // Value = 0x0000nnff: Op=x, Cmode=1100.
3390 Imm = SplatBits >> 8;
3395 if ((SplatBits & ~0xffffff) == 0 &&
3396 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3397 // Value = 0x00nnffff: Op=x, Cmode=1101.
3399 Imm = SplatBits >> 16;
3400 SplatBits |= 0xffff;
3404 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3405 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3406 // VMOV.I32. A (very) minor optimization would be to replicate the value
3407 // and fall through here to test for a valid 64-bit splat. But, then the
3408 // caller would also need to check and handle the change in size.
3412 if (type != VMOVModImm)
3414 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3415 uint64_t BitMask = 0xff;
3417 unsigned ImmMask = 1;
3419 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3420 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3423 } else if ((SplatBits & BitMask) != 0) {
3429 // Op=1, Cmode=1110.
3432 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3437 llvm_unreachable("unexpected size for isNEONModifiedImm");
3441 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3442 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3445 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3446 bool &ReverseVEXT, unsigned &Imm) {
3447 unsigned NumElts = VT.getVectorNumElements();
3448 ReverseVEXT = false;
3450 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3456 // If this is a VEXT shuffle, the immediate value is the index of the first
3457 // element. The other shuffle indices must be the successive elements after
3459 unsigned ExpectedElt = Imm;
3460 for (unsigned i = 1; i < NumElts; ++i) {
3461 // Increment the expected index. If it wraps around, it may still be
3462 // a VEXT but the source vectors must be swapped.
3464 if (ExpectedElt == NumElts * 2) {
3469 if (M[i] < 0) continue; // ignore UNDEF indices
3470 if (ExpectedElt != static_cast<unsigned>(M[i]))
3474 // Adjust the index value if the source operands will be swapped.
3481 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3482 /// instruction with the specified blocksize. (The order of the elements
3483 /// within each block of the vector is reversed.)
3484 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3485 unsigned BlockSize) {
3486 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3487 "Only possible block sizes for VREV are: 16, 32, 64");
3489 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3493 unsigned NumElts = VT.getVectorNumElements();
3494 unsigned BlockElts = M[0] + 1;
3495 // If the first shuffle index is UNDEF, be optimistic.
3497 BlockElts = BlockSize / EltSz;
3499 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3502 for (unsigned i = 0; i < NumElts; ++i) {
3503 if (M[i] < 0) continue; // ignore UNDEF indices
3504 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3511 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3512 unsigned &WhichResult) {
3513 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3517 unsigned NumElts = VT.getVectorNumElements();
3518 WhichResult = (M[0] == 0 ? 0 : 1);
3519 for (unsigned i = 0; i < NumElts; i += 2) {
3520 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3521 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3527 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3528 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3529 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3530 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3531 unsigned &WhichResult) {
3532 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3536 unsigned NumElts = VT.getVectorNumElements();
3537 WhichResult = (M[0] == 0 ? 0 : 1);
3538 for (unsigned i = 0; i < NumElts; i += 2) {
3539 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3540 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3546 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3547 unsigned &WhichResult) {
3548 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3552 unsigned NumElts = VT.getVectorNumElements();
3553 WhichResult = (M[0] == 0 ? 0 : 1);
3554 for (unsigned i = 0; i != NumElts; ++i) {
3555 if (M[i] < 0) continue; // ignore UNDEF indices
3556 if ((unsigned) M[i] != 2 * i + WhichResult)
3560 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3561 if (VT.is64BitVector() && EltSz == 32)
3567 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3568 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3569 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3570 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3571 unsigned &WhichResult) {
3572 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3576 unsigned Half = VT.getVectorNumElements() / 2;
3577 WhichResult = (M[0] == 0 ? 0 : 1);
3578 for (unsigned j = 0; j != 2; ++j) {
3579 unsigned Idx = WhichResult;
3580 for (unsigned i = 0; i != Half; ++i) {
3581 int MIdx = M[i + j * Half];
3582 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3588 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3589 if (VT.is64BitVector() && EltSz == 32)
3595 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3596 unsigned &WhichResult) {
3597 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3601 unsigned NumElts = VT.getVectorNumElements();
3602 WhichResult = (M[0] == 0 ? 0 : 1);
3603 unsigned Idx = WhichResult * NumElts / 2;
3604 for (unsigned i = 0; i != NumElts; i += 2) {
3605 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3606 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3611 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3612 if (VT.is64BitVector() && EltSz == 32)
3618 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3619 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3620 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3621 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3622 unsigned &WhichResult) {
3623 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3627 unsigned NumElts = VT.getVectorNumElements();
3628 WhichResult = (M[0] == 0 ? 0 : 1);
3629 unsigned Idx = WhichResult * NumElts / 2;
3630 for (unsigned i = 0; i != NumElts; i += 2) {
3631 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3632 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3637 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3638 if (VT.is64BitVector() && EltSz == 32)
3644 // If N is an integer constant that can be moved into a register in one
3645 // instruction, return an SDValue of such a constant (will become a MOV
3646 // instruction). Otherwise return null.
3647 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3648 const ARMSubtarget *ST, DebugLoc dl) {
3650 if (!isa<ConstantSDNode>(N))
3652 Val = cast<ConstantSDNode>(N)->getZExtValue();
3654 if (ST->isThumb1Only()) {
3655 if (Val <= 255 || ~Val <= 255)
3656 return DAG.getConstant(Val, MVT::i32);
3658 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3659 return DAG.getConstant(Val, MVT::i32);
3664 // If this is a case we can't handle, return null and let the default
3665 // expansion code take care of it.
3666 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3667 const ARMSubtarget *ST) const {
3668 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3669 DebugLoc dl = Op.getDebugLoc();
3670 EVT VT = Op.getValueType();
3672 APInt SplatBits, SplatUndef;
3673 unsigned SplatBitSize;
3675 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3676 if (SplatBitSize <= 64) {
3677 // Check if an immediate VMOV works.
3679 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3680 SplatUndef.getZExtValue(), SplatBitSize,
3681 DAG, VmovVT, VT.is128BitVector(),
3683 if (Val.getNode()) {
3684 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3685 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3688 // Try an immediate VMVN.
3689 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3690 ((1LL << SplatBitSize) - 1));
3691 Val = isNEONModifiedImm(NegatedImm,
3692 SplatUndef.getZExtValue(), SplatBitSize,
3693 DAG, VmovVT, VT.is128BitVector(),
3695 if (Val.getNode()) {
3696 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3697 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3702 // Scan through the operands to see if only one value is used.
3703 unsigned NumElts = VT.getVectorNumElements();
3704 bool isOnlyLowElement = true;
3705 bool usesOnlyOneValue = true;
3706 bool isConstant = true;
3708 for (unsigned i = 0; i < NumElts; ++i) {
3709 SDValue V = Op.getOperand(i);
3710 if (V.getOpcode() == ISD::UNDEF)
3713 isOnlyLowElement = false;
3714 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3717 if (!Value.getNode())
3719 else if (V != Value)
3720 usesOnlyOneValue = false;
3723 if (!Value.getNode())
3724 return DAG.getUNDEF(VT);
3726 if (isOnlyLowElement)
3727 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3729 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3731 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3732 // i32 and try again.
3733 if (usesOnlyOneValue && EltSize <= 32) {
3735 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3736 if (VT.getVectorElementType().isFloatingPoint()) {
3737 SmallVector<SDValue, 8> Ops;
3738 for (unsigned i = 0; i < NumElts; ++i)
3739 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3741 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3742 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3743 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3745 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3747 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3749 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3752 // If all elements are constants and the case above didn't get hit, fall back
3753 // to the default expansion, which will generate a load from the constant
3758 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3760 SDValue shuffle = ReconstructShuffle(Op, DAG);
3761 if (shuffle != SDValue())
3765 // Vectors with 32- or 64-bit elements can be built by directly assigning
3766 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3767 // will be legalized.
3768 if (EltSize >= 32) {
3769 // Do the expansion with floating-point types, since that is what the VFP
3770 // registers are defined to use, and since i64 is not legal.
3771 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3772 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3773 SmallVector<SDValue, 8> Ops;
3774 for (unsigned i = 0; i < NumElts; ++i)
3775 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3776 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3777 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3783 // Gather data to see if the operation can be modelled as a
3784 // shuffle in combination with VEXTs.
3785 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
3786 SelectionDAG &DAG) const {
3787 DebugLoc dl = Op.getDebugLoc();
3788 EVT VT = Op.getValueType();
3789 unsigned NumElts = VT.getVectorNumElements();
3791 SmallVector<SDValue, 2> SourceVecs;
3792 SmallVector<unsigned, 2> MinElts;
3793 SmallVector<unsigned, 2> MaxElts;
3795 for (unsigned i = 0; i < NumElts; ++i) {
3796 SDValue V = Op.getOperand(i);
3797 if (V.getOpcode() == ISD::UNDEF)
3799 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3800 // A shuffle can only come from building a vector from various
3801 // elements of other vectors.
3805 // Record this extraction against the appropriate vector if possible...
3806 SDValue SourceVec = V.getOperand(0);
3807 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3808 bool FoundSource = false;
3809 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3810 if (SourceVecs[j] == SourceVec) {
3811 if (MinElts[j] > EltNo)
3813 if (MaxElts[j] < EltNo)
3820 // Or record a new source if not...
3822 SourceVecs.push_back(SourceVec);
3823 MinElts.push_back(EltNo);
3824 MaxElts.push_back(EltNo);
3828 // Currently only do something sane when at most two source vectors
3830 if (SourceVecs.size() > 2)
3833 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3834 int VEXTOffsets[2] = {0, 0};
3836 // This loop extracts the usage patterns of the source vectors
3837 // and prepares appropriate SDValues for a shuffle if possible.
3838 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3839 if (SourceVecs[i].getValueType() == VT) {
3840 // No VEXT necessary
3841 ShuffleSrcs[i] = SourceVecs[i];
3844 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3845 // It probably isn't worth padding out a smaller vector just to
3846 // break it down again in a shuffle.
3850 // Since only 64-bit and 128-bit vectors are legal on ARM and
3851 // we've eliminated the other cases...
3852 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
3853 "unexpected vector sizes in ReconstructShuffle");
3855 if (MaxElts[i] - MinElts[i] >= NumElts) {
3856 // Span too large for a VEXT to cope
3860 if (MinElts[i] >= NumElts) {
3861 // The extraction can just take the second half
3862 VEXTOffsets[i] = NumElts;
3863 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3865 DAG.getIntPtrConstant(NumElts));
3866 } else if (MaxElts[i] < NumElts) {
3867 // The extraction can just take the first half
3869 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3871 DAG.getIntPtrConstant(0));
3873 // An actual VEXT is needed
3874 VEXTOffsets[i] = MinElts[i];
3875 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3877 DAG.getIntPtrConstant(0));
3878 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3880 DAG.getIntPtrConstant(NumElts));
3881 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
3882 DAG.getConstant(VEXTOffsets[i], MVT::i32));
3886 SmallVector<int, 8> Mask;
3888 for (unsigned i = 0; i < NumElts; ++i) {
3889 SDValue Entry = Op.getOperand(i);
3890 if (Entry.getOpcode() == ISD::UNDEF) {
3895 SDValue ExtractVec = Entry.getOperand(0);
3896 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
3897 .getOperand(1))->getSExtValue();
3898 if (ExtractVec == SourceVecs[0]) {
3899 Mask.push_back(ExtractElt - VEXTOffsets[0]);
3901 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
3905 // Final check before we try to produce nonsense...
3906 if (isShuffleMaskLegal(Mask, VT))
3907 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
3913 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3914 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3915 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3916 /// are assumed to be legal.
3918 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3920 if (VT.getVectorNumElements() == 4 &&
3921 (VT.is128BitVector() || VT.is64BitVector())) {
3922 unsigned PFIndexes[4];
3923 for (unsigned i = 0; i != 4; ++i) {
3927 PFIndexes[i] = M[i];
3930 // Compute the index in the perfect shuffle table.
3931 unsigned PFTableIndex =
3932 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3933 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3934 unsigned Cost = (PFEntry >> 30);
3941 unsigned Imm, WhichResult;
3943 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3944 return (EltSize >= 32 ||
3945 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3946 isVREVMask(M, VT, 64) ||
3947 isVREVMask(M, VT, 32) ||
3948 isVREVMask(M, VT, 16) ||
3949 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3950 isVTRNMask(M, VT, WhichResult) ||
3951 isVUZPMask(M, VT, WhichResult) ||
3952 isVZIPMask(M, VT, WhichResult) ||
3953 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3954 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3955 isVZIP_v_undef_Mask(M, VT, WhichResult));
3958 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3959 /// the specified operations to build the shuffle.
3960 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3961 SDValue RHS, SelectionDAG &DAG,
3963 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3964 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3965 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3968 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3977 OP_VUZPL, // VUZP, left result
3978 OP_VUZPR, // VUZP, right result
3979 OP_VZIPL, // VZIP, left result
3980 OP_VZIPR, // VZIP, right result
3981 OP_VTRNL, // VTRN, left result
3982 OP_VTRNR // VTRN, right result
3985 if (OpNum == OP_COPY) {
3986 if (LHSID == (1*9+2)*9+3) return LHS;
3987 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3991 SDValue OpLHS, OpRHS;
3992 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3993 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3994 EVT VT = OpLHS.getValueType();
3997 default: llvm_unreachable("Unknown shuffle opcode!");
3999 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4004 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4005 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4009 return DAG.getNode(ARMISD::VEXT, dl, VT,
4011 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4014 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4015 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4018 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4019 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4022 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4023 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4027 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4028 SDValue V1 = Op.getOperand(0);
4029 SDValue V2 = Op.getOperand(1);
4030 DebugLoc dl = Op.getDebugLoc();
4031 EVT VT = Op.getValueType();
4032 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4033 SmallVector<int, 8> ShuffleMask;
4035 // Convert shuffles that are directly supported on NEON to target-specific
4036 // DAG nodes, instead of keeping them as shuffles and matching them again
4037 // during code selection. This is more efficient and avoids the possibility
4038 // of inconsistencies between legalization and selection.
4039 // FIXME: floating-point vectors should be canonicalized to integer vectors
4040 // of the same time so that they get CSEd properly.
4041 SVN->getMask(ShuffleMask);
4043 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4044 if (EltSize <= 32) {
4045 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4046 int Lane = SVN->getSplatIndex();
4047 // If this is undef splat, generate it via "just" vdup, if possible.
4048 if (Lane == -1) Lane = 0;
4050 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4051 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4053 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4054 DAG.getConstant(Lane, MVT::i32));
4059 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4062 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4063 DAG.getConstant(Imm, MVT::i32));
4066 if (isVREVMask(ShuffleMask, VT, 64))
4067 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4068 if (isVREVMask(ShuffleMask, VT, 32))
4069 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4070 if (isVREVMask(ShuffleMask, VT, 16))
4071 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4073 // Check for Neon shuffles that modify both input vectors in place.
4074 // If both results are used, i.e., if there are two shuffles with the same
4075 // source operands and with masks corresponding to both results of one of
4076 // these operations, DAG memoization will ensure that a single node is
4077 // used for both shuffles.
4078 unsigned WhichResult;
4079 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4080 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4081 V1, V2).getValue(WhichResult);
4082 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4083 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4084 V1, V2).getValue(WhichResult);
4085 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4086 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4087 V1, V2).getValue(WhichResult);
4089 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4090 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4091 V1, V1).getValue(WhichResult);
4092 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4093 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4094 V1, V1).getValue(WhichResult);
4095 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4096 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4097 V1, V1).getValue(WhichResult);
4100 // If the shuffle is not directly supported and it has 4 elements, use
4101 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4102 unsigned NumElts = VT.getVectorNumElements();
4104 unsigned PFIndexes[4];
4105 for (unsigned i = 0; i != 4; ++i) {
4106 if (ShuffleMask[i] < 0)
4109 PFIndexes[i] = ShuffleMask[i];
4112 // Compute the index in the perfect shuffle table.
4113 unsigned PFTableIndex =
4114 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4115 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4116 unsigned Cost = (PFEntry >> 30);
4119 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4122 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4123 if (EltSize >= 32) {
4124 // Do the expansion with floating-point types, since that is what the VFP
4125 // registers are defined to use, and since i64 is not legal.
4126 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4127 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4128 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4129 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4130 SmallVector<SDValue, 8> Ops;
4131 for (unsigned i = 0; i < NumElts; ++i) {
4132 if (ShuffleMask[i] < 0)
4133 Ops.push_back(DAG.getUNDEF(EltVT));
4135 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4136 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4137 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4140 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4141 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4147 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4148 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4149 SDValue Lane = Op.getOperand(1);
4150 if (!isa<ConstantSDNode>(Lane))
4153 SDValue Vec = Op.getOperand(0);
4154 if (Op.getValueType() == MVT::i32 &&
4155 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4156 DebugLoc dl = Op.getDebugLoc();
4157 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4163 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4164 // The only time a CONCAT_VECTORS operation can have legal types is when
4165 // two 64-bit vectors are concatenated to a 128-bit vector.
4166 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4167 "unexpected CONCAT_VECTORS");
4168 DebugLoc dl = Op.getDebugLoc();
4169 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4170 SDValue Op0 = Op.getOperand(0);
4171 SDValue Op1 = Op.getOperand(1);
4172 if (Op0.getOpcode() != ISD::UNDEF)
4173 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4174 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4175 DAG.getIntPtrConstant(0));
4176 if (Op1.getOpcode() != ISD::UNDEF)
4177 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4178 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4179 DAG.getIntPtrConstant(1));
4180 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4183 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4184 /// element has been zero/sign-extended, depending on the isSigned parameter,
4185 /// from an integer type half its size.
4186 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4188 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4189 EVT VT = N->getValueType(0);
4190 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4191 SDNode *BVN = N->getOperand(0).getNode();
4192 if (BVN->getValueType(0) != MVT::v4i32 ||
4193 BVN->getOpcode() != ISD::BUILD_VECTOR)
4195 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4196 unsigned HiElt = 1 - LoElt;
4197 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4198 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4199 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4200 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4201 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4204 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4205 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4208 if (Hi0->isNullValue() && Hi1->isNullValue())
4214 if (N->getOpcode() != ISD::BUILD_VECTOR)
4217 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4218 SDNode *Elt = N->getOperand(i).getNode();
4219 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4220 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4221 unsigned HalfSize = EltSize / 2;
4223 int64_t SExtVal = C->getSExtValue();
4224 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4227 if ((C->getZExtValue() >> HalfSize) != 0)
4238 /// isSignExtended - Check if a node is a vector value that is sign-extended
4239 /// or a constant BUILD_VECTOR with sign-extended elements.
4240 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4241 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4243 if (isExtendedBUILD_VECTOR(N, DAG, true))
4248 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4249 /// or a constant BUILD_VECTOR with zero-extended elements.
4250 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4251 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4253 if (isExtendedBUILD_VECTOR(N, DAG, false))
4258 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4259 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4260 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4261 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4262 return N->getOperand(0);
4263 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4264 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4265 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4266 LD->isNonTemporal(), LD->getAlignment());
4267 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4268 // have been legalized as a BITCAST from v4i32.
4269 if (N->getOpcode() == ISD::BITCAST) {
4270 SDNode *BVN = N->getOperand(0).getNode();
4271 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4272 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4273 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4274 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4275 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4277 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4278 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4279 EVT VT = N->getValueType(0);
4280 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4281 unsigned NumElts = VT.getVectorNumElements();
4282 MVT TruncVT = MVT::getIntegerVT(EltSize);
4283 SmallVector<SDValue, 8> Ops;
4284 for (unsigned i = 0; i != NumElts; ++i) {
4285 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4286 const APInt &CInt = C->getAPIntValue();
4287 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4289 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4290 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4293 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4294 // Multiplications are only custom-lowered for 128-bit vectors so that
4295 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4296 EVT VT = Op.getValueType();
4297 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4298 SDNode *N0 = Op.getOperand(0).getNode();
4299 SDNode *N1 = Op.getOperand(1).getNode();
4300 unsigned NewOpc = 0;
4301 if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
4302 NewOpc = ARMISD::VMULLs;
4303 else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
4304 NewOpc = ARMISD::VMULLu;
4305 else if (VT == MVT::v2i64)
4306 // Fall through to expand this. It is not legal.
4309 // Other vector multiplications are legal.
4312 // Legalize to a VMULL instruction.
4313 DebugLoc DL = Op.getDebugLoc();
4314 SDValue Op0 = SkipExtension(N0, DAG);
4315 SDValue Op1 = SkipExtension(N1, DAG);
4317 assert(Op0.getValueType().is64BitVector() &&
4318 Op1.getValueType().is64BitVector() &&
4319 "unexpected types for extended operands to VMULL");
4320 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4324 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4326 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4327 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4328 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4329 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4330 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4331 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4332 // Get reciprocal estimate.
4333 // float4 recip = vrecpeq_f32(yf);
4334 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4335 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4336 // Because char has a smaller range than uchar, we can actually get away
4337 // without any newton steps. This requires that we use a weird bias
4338 // of 0xb000, however (again, this has been exhaustively tested).
4339 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4340 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4341 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4342 Y = DAG.getConstant(0xb000, MVT::i32);
4343 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4344 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4345 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4346 // Convert back to short.
4347 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4348 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4353 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4355 // Convert to float.
4356 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4357 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4358 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4359 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4360 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4361 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4363 // Use reciprocal estimate and one refinement step.
4364 // float4 recip = vrecpeq_f32(yf);
4365 // recip *= vrecpsq_f32(yf, recip);
4366 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4367 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4368 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4369 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4371 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4372 // Because short has a smaller range than ushort, we can actually get away
4373 // with only a single newton step. This requires that we use a weird bias
4374 // of 89, however (again, this has been exhaustively tested).
4375 // float4 result = as_float4(as_int4(xf*recip) + 89);
4376 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4377 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4378 N1 = DAG.getConstant(89, MVT::i32);
4379 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4380 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4381 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4382 // Convert back to integer and return.
4383 // return vmovn_s32(vcvt_s32_f32(result));
4384 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4385 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4389 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4390 EVT VT = Op.getValueType();
4391 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4392 "unexpected type for custom-lowering ISD::SDIV");
4394 DebugLoc dl = Op.getDebugLoc();
4395 SDValue N0 = Op.getOperand(0);
4396 SDValue N1 = Op.getOperand(1);
4399 if (VT == MVT::v8i8) {
4400 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4401 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4403 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4404 DAG.getIntPtrConstant(4));
4405 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4406 DAG.getIntPtrConstant(4));
4407 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4408 DAG.getIntPtrConstant(0));
4409 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4410 DAG.getIntPtrConstant(0));
4412 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4413 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4415 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4416 N0 = LowerCONCAT_VECTORS(N0, DAG);
4418 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4421 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4424 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4425 EVT VT = Op.getValueType();
4426 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4427 "unexpected type for custom-lowering ISD::UDIV");
4429 DebugLoc dl = Op.getDebugLoc();
4430 SDValue N0 = Op.getOperand(0);
4431 SDValue N1 = Op.getOperand(1);
4434 if (VT == MVT::v8i8) {
4435 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4436 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4438 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4439 DAG.getIntPtrConstant(4));
4440 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4441 DAG.getIntPtrConstant(4));
4442 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4443 DAG.getIntPtrConstant(0));
4444 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4445 DAG.getIntPtrConstant(0));
4447 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4448 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4450 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4451 N0 = LowerCONCAT_VECTORS(N0, DAG);
4453 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4454 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4459 // v4i16 sdiv ... Convert to float.
4460 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4461 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4462 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4463 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4464 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4465 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4467 // Use reciprocal estimate and two refinement steps.
4468 // float4 recip = vrecpeq_f32(yf);
4469 // recip *= vrecpsq_f32(yf, recip);
4470 // recip *= vrecpsq_f32(yf, recip);
4471 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4472 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4473 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4474 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4476 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4477 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4478 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4480 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4481 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4482 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4483 // and that it will never cause us to return an answer too large).
4484 // float4 result = as_float4(as_int4(xf*recip) + 89);
4485 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4486 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4487 N1 = DAG.getConstant(2, MVT::i32);
4488 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4489 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4490 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4491 // Convert back to integer and return.
4492 // return vmovn_u32(vcvt_s32_f32(result));
4493 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4494 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4498 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4499 switch (Op.getOpcode()) {
4500 default: llvm_unreachable("Don't know how to custom lower this!");
4501 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4502 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4503 case ISD::GlobalAddress:
4504 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4505 LowerGlobalAddressELF(Op, DAG);
4506 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4507 case ISD::SELECT: return LowerSELECT(Op, DAG);
4508 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4509 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4510 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4511 case ISD::VASTART: return LowerVASTART(Op, DAG);
4512 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4513 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4514 case ISD::SINT_TO_FP:
4515 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4516 case ISD::FP_TO_SINT:
4517 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4518 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4519 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4520 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4521 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4522 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4523 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4524 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4525 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4527 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4530 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4531 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4532 case ISD::SRL_PARTS:
4533 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4534 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4535 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4536 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4537 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4538 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4539 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4540 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4541 case ISD::MUL: return LowerMUL(Op, DAG);
4542 case ISD::SDIV: return LowerSDIV(Op, DAG);
4543 case ISD::UDIV: return LowerUDIV(Op, DAG);
4548 /// ReplaceNodeResults - Replace the results of node with an illegal result
4549 /// type with new values built out of custom code.
4550 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4551 SmallVectorImpl<SDValue>&Results,
4552 SelectionDAG &DAG) const {
4554 switch (N->getOpcode()) {
4556 llvm_unreachable("Don't know how to custom expand this!");
4559 Res = ExpandBITCAST(N, DAG);
4563 Res = Expand64BitShift(N, DAG, Subtarget);
4567 Results.push_back(Res);
4570 //===----------------------------------------------------------------------===//
4571 // ARM Scheduler Hooks
4572 //===----------------------------------------------------------------------===//
4575 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4576 MachineBasicBlock *BB,
4577 unsigned Size) const {
4578 unsigned dest = MI->getOperand(0).getReg();
4579 unsigned ptr = MI->getOperand(1).getReg();
4580 unsigned oldval = MI->getOperand(2).getReg();
4581 unsigned newval = MI->getOperand(3).getReg();
4582 unsigned scratch = BB->getParent()->getRegInfo()
4583 .createVirtualRegister(ARM::GPRRegisterClass);
4584 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4585 DebugLoc dl = MI->getDebugLoc();
4586 bool isThumb2 = Subtarget->isThumb2();
4588 unsigned ldrOpc, strOpc;
4590 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4592 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4593 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4596 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4597 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4600 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4601 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4605 MachineFunction *MF = BB->getParent();
4606 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4607 MachineFunction::iterator It = BB;
4608 ++It; // insert the new blocks after the current block
4610 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4611 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4612 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4613 MF->insert(It, loop1MBB);
4614 MF->insert(It, loop2MBB);
4615 MF->insert(It, exitMBB);
4617 // Transfer the remainder of BB and its successor edges to exitMBB.
4618 exitMBB->splice(exitMBB->begin(), BB,
4619 llvm::next(MachineBasicBlock::iterator(MI)),
4621 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4625 // fallthrough --> loop1MBB
4626 BB->addSuccessor(loop1MBB);
4629 // ldrex dest, [ptr]
4633 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4634 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4635 .addReg(dest).addReg(oldval));
4636 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4637 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4638 BB->addSuccessor(loop2MBB);
4639 BB->addSuccessor(exitMBB);
4642 // strex scratch, newval, [ptr]
4646 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4648 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4649 .addReg(scratch).addImm(0));
4650 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4651 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4652 BB->addSuccessor(loop1MBB);
4653 BB->addSuccessor(exitMBB);
4659 MI->eraseFromParent(); // The instruction is gone now.
4665 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4666 unsigned Size, unsigned BinOpcode) const {
4667 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4668 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4670 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4671 MachineFunction *MF = BB->getParent();
4672 MachineFunction::iterator It = BB;
4675 unsigned dest = MI->getOperand(0).getReg();
4676 unsigned ptr = MI->getOperand(1).getReg();
4677 unsigned incr = MI->getOperand(2).getReg();
4678 DebugLoc dl = MI->getDebugLoc();
4680 bool isThumb2 = Subtarget->isThumb2();
4681 unsigned ldrOpc, strOpc;
4683 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4685 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4686 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4689 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4690 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4693 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4694 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4698 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4699 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4700 MF->insert(It, loopMBB);
4701 MF->insert(It, exitMBB);
4703 // Transfer the remainder of BB and its successor edges to exitMBB.
4704 exitMBB->splice(exitMBB->begin(), BB,
4705 llvm::next(MachineBasicBlock::iterator(MI)),
4707 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4709 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4710 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4711 unsigned scratch2 = (!BinOpcode) ? incr :
4712 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4716 // fallthrough --> loopMBB
4717 BB->addSuccessor(loopMBB);
4721 // <binop> scratch2, dest, incr
4722 // strex scratch, scratch2, ptr
4725 // fallthrough --> exitMBB
4727 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4729 // operand order needs to go the other way for NAND
4730 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4731 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4732 addReg(incr).addReg(dest)).addReg(0);
4734 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4735 addReg(dest).addReg(incr)).addReg(0);
4738 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4740 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4741 .addReg(scratch).addImm(0));
4742 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4743 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4745 BB->addSuccessor(loopMBB);
4746 BB->addSuccessor(exitMBB);
4752 MI->eraseFromParent(); // The instruction is gone now.
4758 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4759 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4760 E = MBB->succ_end(); I != E; ++I)
4763 llvm_unreachable("Expecting a BB with two successors!");
4767 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4768 MachineBasicBlock *BB) const {
4769 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4770 DebugLoc dl = MI->getDebugLoc();
4771 bool isThumb2 = Subtarget->isThumb2();
4772 switch (MI->getOpcode()) {
4775 llvm_unreachable("Unexpected instr type to insert");
4777 case ARM::ATOMIC_LOAD_ADD_I8:
4778 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4779 case ARM::ATOMIC_LOAD_ADD_I16:
4780 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4781 case ARM::ATOMIC_LOAD_ADD_I32:
4782 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4784 case ARM::ATOMIC_LOAD_AND_I8:
4785 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4786 case ARM::ATOMIC_LOAD_AND_I16:
4787 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4788 case ARM::ATOMIC_LOAD_AND_I32:
4789 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4791 case ARM::ATOMIC_LOAD_OR_I8:
4792 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4793 case ARM::ATOMIC_LOAD_OR_I16:
4794 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4795 case ARM::ATOMIC_LOAD_OR_I32:
4796 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4798 case ARM::ATOMIC_LOAD_XOR_I8:
4799 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4800 case ARM::ATOMIC_LOAD_XOR_I16:
4801 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4802 case ARM::ATOMIC_LOAD_XOR_I32:
4803 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4805 case ARM::ATOMIC_LOAD_NAND_I8:
4806 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4807 case ARM::ATOMIC_LOAD_NAND_I16:
4808 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4809 case ARM::ATOMIC_LOAD_NAND_I32:
4810 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4812 case ARM::ATOMIC_LOAD_SUB_I8:
4813 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4814 case ARM::ATOMIC_LOAD_SUB_I16:
4815 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4816 case ARM::ATOMIC_LOAD_SUB_I32:
4817 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4819 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4820 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4821 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4823 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4824 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4825 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4827 case ARM::tMOVCCr_pseudo: {
4828 // To "insert" a SELECT_CC instruction, we actually have to insert the
4829 // diamond control-flow pattern. The incoming instruction knows the
4830 // destination vreg to set, the condition code register to branch on, the
4831 // true/false values to select between, and a branch opcode to use.
4832 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4833 MachineFunction::iterator It = BB;
4839 // cmpTY ccX, r1, r2
4841 // fallthrough --> copy0MBB
4842 MachineBasicBlock *thisMBB = BB;
4843 MachineFunction *F = BB->getParent();
4844 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4845 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4846 F->insert(It, copy0MBB);
4847 F->insert(It, sinkMBB);
4849 // Transfer the remainder of BB and its successor edges to sinkMBB.
4850 sinkMBB->splice(sinkMBB->begin(), BB,
4851 llvm::next(MachineBasicBlock::iterator(MI)),
4853 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4855 BB->addSuccessor(copy0MBB);
4856 BB->addSuccessor(sinkMBB);
4858 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4859 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4862 // %FalseValue = ...
4863 // # fallthrough to sinkMBB
4866 // Update machine-CFG edges
4867 BB->addSuccessor(sinkMBB);
4870 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4873 BuildMI(*BB, BB->begin(), dl,
4874 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4875 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4876 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4878 MI->eraseFromParent(); // The pseudo instruction is gone now.
4883 case ARM::BCCZi64: {
4884 // If there is an unconditional branch to the other successor, remove it.
4885 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
4887 // Compare both parts that make up the double comparison separately for
4889 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4891 unsigned LHS1 = MI->getOperand(1).getReg();
4892 unsigned LHS2 = MI->getOperand(2).getReg();
4894 AddDefaultPred(BuildMI(BB, dl,
4895 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4896 .addReg(LHS1).addImm(0));
4897 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4898 .addReg(LHS2).addImm(0)
4899 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4901 unsigned RHS1 = MI->getOperand(3).getReg();
4902 unsigned RHS2 = MI->getOperand(4).getReg();
4903 AddDefaultPred(BuildMI(BB, dl,
4904 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4905 .addReg(LHS1).addReg(RHS1));
4906 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4907 .addReg(LHS2).addReg(RHS2)
4908 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4911 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4912 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4913 if (MI->getOperand(0).getImm() == ARMCC::NE)
4914 std::swap(destMBB, exitMBB);
4916 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4917 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4918 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4921 MI->eraseFromParent(); // The pseudo instruction is gone now.
4927 //===----------------------------------------------------------------------===//
4928 // ARM Optimization Hooks
4929 //===----------------------------------------------------------------------===//
4932 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4933 TargetLowering::DAGCombinerInfo &DCI) {
4934 SelectionDAG &DAG = DCI.DAG;
4935 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4936 EVT VT = N->getValueType(0);
4937 unsigned Opc = N->getOpcode();
4938 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4939 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4940 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4941 ISD::CondCode CC = ISD::SETCC_INVALID;
4944 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4946 SDValue CCOp = Slct.getOperand(0);
4947 if (CCOp.getOpcode() == ISD::SETCC)
4948 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4951 bool DoXform = false;
4953 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4956 if (LHS.getOpcode() == ISD::Constant &&
4957 cast<ConstantSDNode>(LHS)->isNullValue()) {
4959 } else if (CC != ISD::SETCC_INVALID &&
4960 RHS.getOpcode() == ISD::Constant &&
4961 cast<ConstantSDNode>(RHS)->isNullValue()) {
4962 std::swap(LHS, RHS);
4963 SDValue Op0 = Slct.getOperand(0);
4964 EVT OpVT = isSlctCC ? Op0.getValueType() :
4965 Op0.getOperand(0).getValueType();
4966 bool isInt = OpVT.isInteger();
4967 CC = ISD::getSetCCInverse(CC, isInt);
4969 if (!TLI.isCondCodeLegal(CC, OpVT))
4970 return SDValue(); // Inverse operator isn't legal.
4977 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4979 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4980 Slct.getOperand(0), Slct.getOperand(1), CC);
4981 SDValue CCOp = Slct.getOperand(0);
4983 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4984 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4985 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4986 CCOp, OtherOp, Result);
4991 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4992 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4993 /// called with the default operands, and if that fails, with commuted
4995 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4996 TargetLowering::DAGCombinerInfo &DCI) {
4997 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4998 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4999 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
5000 if (Result.getNode()) return Result;
5005 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
5007 static SDValue PerformADDCombine(SDNode *N,
5008 TargetLowering::DAGCombinerInfo &DCI) {
5009 SDValue N0 = N->getOperand(0);
5010 SDValue N1 = N->getOperand(1);
5012 // First try with the default operand order.
5013 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
5014 if (Result.getNode())
5017 // If that didn't work, try again with the operands commuted.
5018 return PerformADDCombineWithOperands(N, N1, N0, DCI);
5021 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
5023 static SDValue PerformSUBCombine(SDNode *N,
5024 TargetLowering::DAGCombinerInfo &DCI) {
5025 SDValue N0 = N->getOperand(0);
5026 SDValue N1 = N->getOperand(1);
5028 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
5029 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
5030 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
5031 if (Result.getNode()) return Result;
5037 static SDValue PerformMULCombine(SDNode *N,
5038 TargetLowering::DAGCombinerInfo &DCI,
5039 const ARMSubtarget *Subtarget) {
5040 SelectionDAG &DAG = DCI.DAG;
5042 if (Subtarget->isThumb1Only())
5045 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5048 EVT VT = N->getValueType(0);
5052 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
5056 uint64_t MulAmt = C->getZExtValue();
5057 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
5058 ShiftAmt = ShiftAmt & (32 - 1);
5059 SDValue V = N->getOperand(0);
5060 DebugLoc DL = N->getDebugLoc();
5063 MulAmt >>= ShiftAmt;
5064 if (isPowerOf2_32(MulAmt - 1)) {
5065 // (mul x, 2^N + 1) => (add (shl x, N), x)
5066 Res = DAG.getNode(ISD::ADD, DL, VT,
5067 V, DAG.getNode(ISD::SHL, DL, VT,
5068 V, DAG.getConstant(Log2_32(MulAmt-1),
5070 } else if (isPowerOf2_32(MulAmt + 1)) {
5071 // (mul x, 2^N - 1) => (sub (shl x, N), x)
5072 Res = DAG.getNode(ISD::SUB, DL, VT,
5073 DAG.getNode(ISD::SHL, DL, VT,
5074 V, DAG.getConstant(Log2_32(MulAmt+1),
5081 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
5082 DAG.getConstant(ShiftAmt, MVT::i32));
5084 // Do not add new nodes to DAG combiner worklist.
5085 DCI.CombineTo(N, Res, false);
5089 static SDValue PerformANDCombine(SDNode *N,
5090 TargetLowering::DAGCombinerInfo &DCI) {
5091 // Attempt to use immediate-form VBIC
5092 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5093 DebugLoc dl = N->getDebugLoc();
5094 EVT VT = N->getValueType(0);
5095 SelectionDAG &DAG = DCI.DAG;
5097 APInt SplatBits, SplatUndef;
5098 unsigned SplatBitSize;
5101 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5102 if (SplatBitSize <= 64) {
5104 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
5105 SplatUndef.getZExtValue(), SplatBitSize,
5106 DAG, VbicVT, VT.is128BitVector(),
5108 if (Val.getNode()) {
5110 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
5111 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
5112 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
5120 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
5121 static SDValue PerformORCombine(SDNode *N,
5122 TargetLowering::DAGCombinerInfo &DCI,
5123 const ARMSubtarget *Subtarget) {
5124 // Attempt to use immediate-form VORR
5125 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5126 DebugLoc dl = N->getDebugLoc();
5127 EVT VT = N->getValueType(0);
5128 SelectionDAG &DAG = DCI.DAG;
5130 APInt SplatBits, SplatUndef;
5131 unsigned SplatBitSize;
5133 if (BVN && Subtarget->hasNEON() &&
5134 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5135 if (SplatBitSize <= 64) {
5137 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5138 SplatUndef.getZExtValue(), SplatBitSize,
5139 DAG, VorrVT, VT.is128BitVector(),
5141 if (Val.getNode()) {
5143 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
5144 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
5145 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
5150 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
5153 // BFI is only available on V6T2+
5154 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
5157 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
5158 DebugLoc DL = N->getDebugLoc();
5159 // 1) or (and A, mask), val => ARMbfi A, val, mask
5160 // iff (val & mask) == val
5162 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5163 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
5164 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
5165 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
5166 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
5167 // (i.e., copy a bitfield value into another bitfield of the same width)
5168 if (N0.getOpcode() != ISD::AND)
5174 SDValue N00 = N0.getOperand(0);
5176 // The value and the mask need to be constants so we can verify this is
5177 // actually a bitfield set. If the mask is 0xffff, we can do better
5178 // via a movt instruction, so don't use BFI in that case.
5179 SDValue MaskOp = N0.getOperand(1);
5180 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
5183 unsigned Mask = MaskC->getZExtValue();
5187 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
5188 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
5190 unsigned Val = N1C->getZExtValue();
5191 if ((Val & ~Mask) != Val)
5194 if (ARM::isBitFieldInvertedMask(Mask)) {
5195 Val >>= CountTrailingZeros_32(~Mask);
5197 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
5198 DAG.getConstant(Val, MVT::i32),
5199 DAG.getConstant(Mask, MVT::i32));
5201 // Do not add new nodes to DAG combiner worklist.
5202 DCI.CombineTo(N, Res, false);
5205 } else if (N1.getOpcode() == ISD::AND) {
5206 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5207 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5210 unsigned Mask2 = N11C->getZExtValue();
5212 if (ARM::isBitFieldInvertedMask(Mask) &&
5213 ARM::isBitFieldInvertedMask(~Mask2) &&
5214 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
5215 // The pack halfword instruction works better for masks that fit it,
5216 // so use that when it's available.
5217 if (Subtarget->hasT2ExtractPack() &&
5218 (Mask == 0xffff || Mask == 0xffff0000))
5221 unsigned lsb = CountTrailingZeros_32(Mask2);
5222 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
5223 DAG.getConstant(lsb, MVT::i32));
5224 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
5225 DAG.getConstant(Mask, MVT::i32));
5226 // Do not add new nodes to DAG combiner worklist.
5227 DCI.CombineTo(N, Res, false);
5229 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
5230 ARM::isBitFieldInvertedMask(Mask2) &&
5231 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
5232 // The pack halfword instruction works better for masks that fit it,
5233 // so use that when it's available.
5234 if (Subtarget->hasT2ExtractPack() &&
5235 (Mask2 == 0xffff || Mask2 == 0xffff0000))
5238 unsigned lsb = CountTrailingZeros_32(Mask);
5239 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
5240 DAG.getConstant(lsb, MVT::i32));
5241 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
5242 DAG.getConstant(Mask2, MVT::i32));
5243 // Do not add new nodes to DAG combiner worklist.
5244 DCI.CombineTo(N, Res, false);
5249 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
5250 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
5251 ARM::isBitFieldInvertedMask(~Mask)) {
5252 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
5253 // where lsb(mask) == #shamt and masked bits of B are known zero.
5254 SDValue ShAmt = N00.getOperand(1);
5255 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5256 unsigned LSB = CountTrailingZeros_32(Mask);
5260 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
5261 DAG.getConstant(~Mask, MVT::i32));
5263 // Do not add new nodes to DAG combiner worklist.
5264 DCI.CombineTo(N, Res, false);
5270 /// PerformBFICombine - (bfi A, (and B, C1), C2) -> (bfi A, B, C2) iff
5272 static SDValue PerformBFICombine(SDNode *N,
5273 TargetLowering::DAGCombinerInfo &DCI) {
5274 SDValue N1 = N->getOperand(1);
5275 if (N1.getOpcode() == ISD::AND) {
5276 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5279 unsigned Mask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
5280 unsigned Mask2 = N11C->getZExtValue();
5281 if ((Mask & Mask2) == Mask2)
5282 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
5283 N->getOperand(0), N1.getOperand(0),
5289 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
5290 /// ARMISD::VMOVRRD.
5291 static SDValue PerformVMOVRRDCombine(SDNode *N,
5292 TargetLowering::DAGCombinerInfo &DCI) {
5293 // vmovrrd(vmovdrr x, y) -> x,y
5294 SDValue InDouble = N->getOperand(0);
5295 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
5296 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
5300 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
5301 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
5302 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
5303 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
5304 SDValue Op0 = N->getOperand(0);
5305 SDValue Op1 = N->getOperand(1);
5306 if (Op0.getOpcode() == ISD::BITCAST)
5307 Op0 = Op0.getOperand(0);
5308 if (Op1.getOpcode() == ISD::BITCAST)
5309 Op1 = Op1.getOperand(0);
5310 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
5311 Op0.getNode() == Op1.getNode() &&
5312 Op0.getResNo() == 0 && Op1.getResNo() == 1)
5313 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5314 N->getValueType(0), Op0.getOperand(0));
5318 /// PerformSTORECombine - Target-specific dag combine xforms for
5320 static SDValue PerformSTORECombine(SDNode *N,
5321 TargetLowering::DAGCombinerInfo &DCI) {
5322 // Bitcast an i64 store extracted from a vector to f64.
5323 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5324 StoreSDNode *St = cast<StoreSDNode>(N);
5325 SDValue StVal = St->getValue();
5326 if (!ISD::isNormalStore(St) || St->isVolatile() ||
5327 StVal.getValueType() != MVT::i64 ||
5328 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5331 SelectionDAG &DAG = DCI.DAG;
5332 DebugLoc dl = StVal.getDebugLoc();
5333 SDValue IntVec = StVal.getOperand(0);
5334 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5335 IntVec.getValueType().getVectorNumElements());
5336 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
5337 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5338 Vec, StVal.getOperand(1));
5339 dl = N->getDebugLoc();
5340 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
5341 // Make the DAGCombiner fold the bitcasts.
5342 DCI.AddToWorklist(Vec.getNode());
5343 DCI.AddToWorklist(ExtElt.getNode());
5344 DCI.AddToWorklist(V.getNode());
5345 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
5346 St->getPointerInfo(), St->isVolatile(),
5347 St->isNonTemporal(), St->getAlignment(),
5351 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
5352 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
5353 /// i64 vector to have f64 elements, since the value can then be loaded
5354 /// directly into a VFP register.
5355 static bool hasNormalLoadOperand(SDNode *N) {
5356 unsigned NumElts = N->getValueType(0).getVectorNumElements();
5357 for (unsigned i = 0; i < NumElts; ++i) {
5358 SDNode *Elt = N->getOperand(i).getNode();
5359 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
5365 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
5366 /// ISD::BUILD_VECTOR.
5367 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
5368 TargetLowering::DAGCombinerInfo &DCI){
5369 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
5370 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
5371 // into a pair of GPRs, which is fine when the value is used as a scalar,
5372 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
5373 SelectionDAG &DAG = DCI.DAG;
5374 if (N->getNumOperands() == 2) {
5375 SDValue RV = PerformVMOVDRRCombine(N, DAG);
5380 // Load i64 elements as f64 values so that type legalization does not split
5381 // them up into i32 values.
5382 EVT VT = N->getValueType(0);
5383 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
5385 DebugLoc dl = N->getDebugLoc();
5386 SmallVector<SDValue, 8> Ops;
5387 unsigned NumElts = VT.getVectorNumElements();
5388 for (unsigned i = 0; i < NumElts; ++i) {
5389 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
5391 // Make the DAGCombiner fold the bitcast.
5392 DCI.AddToWorklist(V.getNode());
5394 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
5395 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
5396 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
5399 /// PerformInsertEltCombine - Target-specific dag combine xforms for
5400 /// ISD::INSERT_VECTOR_ELT.
5401 static SDValue PerformInsertEltCombine(SDNode *N,
5402 TargetLowering::DAGCombinerInfo &DCI) {
5403 // Bitcast an i64 load inserted into a vector to f64.
5404 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5405 EVT VT = N->getValueType(0);
5406 SDNode *Elt = N->getOperand(1).getNode();
5407 if (VT.getVectorElementType() != MVT::i64 ||
5408 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
5411 SelectionDAG &DAG = DCI.DAG;
5412 DebugLoc dl = N->getDebugLoc();
5413 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5414 VT.getVectorNumElements());
5415 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
5416 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
5417 // Make the DAGCombiner fold the bitcasts.
5418 DCI.AddToWorklist(Vec.getNode());
5419 DCI.AddToWorklist(V.getNode());
5420 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
5421 Vec, V, N->getOperand(2));
5422 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
5425 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
5426 /// ISD::VECTOR_SHUFFLE.
5427 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
5428 // The LLVM shufflevector instruction does not require the shuffle mask
5429 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
5430 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
5431 // operands do not match the mask length, they are extended by concatenating
5432 // them with undef vectors. That is probably the right thing for other
5433 // targets, but for NEON it is better to concatenate two double-register
5434 // size vector operands into a single quad-register size vector. Do that
5435 // transformation here:
5436 // shuffle(concat(v1, undef), concat(v2, undef)) ->
5437 // shuffle(concat(v1, v2), undef)
5438 SDValue Op0 = N->getOperand(0);
5439 SDValue Op1 = N->getOperand(1);
5440 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
5441 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
5442 Op0.getNumOperands() != 2 ||
5443 Op1.getNumOperands() != 2)
5445 SDValue Concat0Op1 = Op0.getOperand(1);
5446 SDValue Concat1Op1 = Op1.getOperand(1);
5447 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
5448 Concat1Op1.getOpcode() != ISD::UNDEF)
5450 // Skip the transformation if any of the types are illegal.
5451 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5452 EVT VT = N->getValueType(0);
5453 if (!TLI.isTypeLegal(VT) ||
5454 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
5455 !TLI.isTypeLegal(Concat1Op1.getValueType()))
5458 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5459 Op0.getOperand(0), Op1.getOperand(0));
5460 // Translate the shuffle mask.
5461 SmallVector<int, 16> NewMask;
5462 unsigned NumElts = VT.getVectorNumElements();
5463 unsigned HalfElts = NumElts/2;
5464 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
5465 for (unsigned n = 0; n < NumElts; ++n) {
5466 int MaskElt = SVN->getMaskElt(n);
5468 if (MaskElt < (int)HalfElts)
5470 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
5471 NewElt = HalfElts + MaskElt - NumElts;
5472 NewMask.push_back(NewElt);
5474 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
5475 DAG.getUNDEF(VT), NewMask.data());
5478 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
5479 /// NEON load/store intrinsics to merge base address updates.
5480 static SDValue CombineBaseUpdate(SDNode *N,
5481 TargetLowering::DAGCombinerInfo &DCI) {
5482 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5485 SelectionDAG &DAG = DCI.DAG;
5486 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
5487 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
5488 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
5489 SDValue Addr = N->getOperand(AddrOpIdx);
5491 // Search for a use of the address operand that is an increment.
5492 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
5493 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
5495 if (User->getOpcode() != ISD::ADD ||
5496 UI.getUse().getResNo() != Addr.getResNo())
5499 // Check that the add is independent of the load/store. Otherwise, folding
5500 // it would create a cycle.
5501 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
5504 // Find the new opcode for the updating load/store.
5506 bool isLaneOp = false;
5507 unsigned NewOpc = 0;
5508 unsigned NumVecs = 0;
5510 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
5512 default: assert(0 && "unexpected intrinsic for Neon base update");
5513 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
5515 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
5517 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
5519 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
5521 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
5522 NumVecs = 2; isLaneOp = true; break;
5523 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
5524 NumVecs = 3; isLaneOp = true; break;
5525 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
5526 NumVecs = 4; isLaneOp = true; break;
5527 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
5528 NumVecs = 1; isLoad = false; break;
5529 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
5530 NumVecs = 2; isLoad = false; break;
5531 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
5532 NumVecs = 3; isLoad = false; break;
5533 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
5534 NumVecs = 4; isLoad = false; break;
5535 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
5536 NumVecs = 2; isLoad = false; isLaneOp = true; break;
5537 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
5538 NumVecs = 3; isLoad = false; isLaneOp = true; break;
5539 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
5540 NumVecs = 4; isLoad = false; isLaneOp = true; break;
5544 switch (N->getOpcode()) {
5545 default: assert(0 && "unexpected opcode for Neon base update");
5546 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
5547 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
5548 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
5552 // Find the size of memory referenced by the load/store.
5555 VecTy = N->getValueType(0);
5557 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
5558 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
5560 NumBytes /= VecTy.getVectorNumElements();
5562 // If the increment is a constant, it must match the memory ref size.
5563 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
5564 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
5565 uint64_t IncVal = CInc->getZExtValue();
5566 if (IncVal != NumBytes)
5568 } else if (NumBytes >= 3 * 16) {
5569 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
5570 // separate instructions that make it harder to use a non-constant update.
5574 // Create the new updating load/store node.
5576 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
5578 for (n = 0; n < NumResultVecs; ++n)
5580 Tys[n++] = MVT::i32;
5581 Tys[n] = MVT::Other;
5582 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
5583 SmallVector<SDValue, 8> Ops;
5584 Ops.push_back(N->getOperand(0)); // incoming chain
5585 Ops.push_back(N->getOperand(AddrOpIdx));
5587 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
5588 Ops.push_back(N->getOperand(i));
5590 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
5591 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
5592 Ops.data(), Ops.size(),
5593 MemInt->getMemoryVT(),
5594 MemInt->getMemOperand());
5597 std::vector<SDValue> NewResults;
5598 for (unsigned i = 0; i < NumResultVecs; ++i) {
5599 NewResults.push_back(SDValue(UpdN.getNode(), i));
5601 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
5602 DCI.CombineTo(N, NewResults);
5603 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
5610 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
5611 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
5612 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
5614 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
5615 SelectionDAG &DAG = DCI.DAG;
5616 EVT VT = N->getValueType(0);
5617 // vldN-dup instructions only support 64-bit vectors for N > 1.
5618 if (!VT.is64BitVector())
5621 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
5622 SDNode *VLD = N->getOperand(0).getNode();
5623 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
5625 unsigned NumVecs = 0;
5626 unsigned NewOpc = 0;
5627 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
5628 if (IntNo == Intrinsic::arm_neon_vld2lane) {
5630 NewOpc = ARMISD::VLD2DUP;
5631 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
5633 NewOpc = ARMISD::VLD3DUP;
5634 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
5636 NewOpc = ARMISD::VLD4DUP;
5641 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
5642 // numbers match the load.
5643 unsigned VLDLaneNo =
5644 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
5645 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5647 // Ignore uses of the chain result.
5648 if (UI.getUse().getResNo() == NumVecs)
5651 if (User->getOpcode() != ARMISD::VDUPLANE ||
5652 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
5656 // Create the vldN-dup node.
5659 for (n = 0; n < NumVecs; ++n)
5661 Tys[n] = MVT::Other;
5662 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
5663 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
5664 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
5665 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
5666 Ops, 2, VLDMemInt->getMemoryVT(),
5667 VLDMemInt->getMemOperand());
5670 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5672 unsigned ResNo = UI.getUse().getResNo();
5673 // Ignore uses of the chain result.
5674 if (ResNo == NumVecs)
5677 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
5680 // Now the vldN-lane intrinsic is dead except for its chain result.
5681 // Update uses of the chain.
5682 std::vector<SDValue> VLDDupResults;
5683 for (unsigned n = 0; n < NumVecs; ++n)
5684 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
5685 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
5686 DCI.CombineTo(VLD, VLDDupResults);
5691 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
5692 /// ARMISD::VDUPLANE.
5693 static SDValue PerformVDUPLANECombine(SDNode *N,
5694 TargetLowering::DAGCombinerInfo &DCI) {
5695 SDValue Op = N->getOperand(0);
5697 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
5698 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
5699 if (CombineVLDDUP(N, DCI))
5700 return SDValue(N, 0);
5702 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
5703 // redundant. Ignore bit_converts for now; element sizes are checked below.
5704 while (Op.getOpcode() == ISD::BITCAST)
5705 Op = Op.getOperand(0);
5706 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
5709 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
5710 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
5711 // The canonical VMOV for a zero vector uses a 32-bit element size.
5712 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5714 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
5716 EVT VT = N->getValueType(0);
5717 if (EltSize > VT.getVectorElementType().getSizeInBits())
5720 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
5723 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5724 /// operand of a vector shift operation, where all the elements of the
5725 /// build_vector must have the same constant integer value.
5726 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5727 // Ignore bit_converts.
5728 while (Op.getOpcode() == ISD::BITCAST)
5729 Op = Op.getOperand(0);
5730 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5731 APInt SplatBits, SplatUndef;
5732 unsigned SplatBitSize;
5734 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5735 HasAnyUndefs, ElementBits) ||
5736 SplatBitSize > ElementBits)
5738 Cnt = SplatBits.getSExtValue();
5742 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5743 /// operand of a vector shift left operation. That value must be in the range:
5744 /// 0 <= Value < ElementBits for a left shift; or
5745 /// 0 <= Value <= ElementBits for a long left shift.
5746 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5747 assert(VT.isVector() && "vector shift count is not a vector type");
5748 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5749 if (! getVShiftImm(Op, ElementBits, Cnt))
5751 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
5754 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5755 /// operand of a vector shift right operation. For a shift opcode, the value
5756 /// is positive, but for an intrinsic the value count must be negative. The
5757 /// absolute value must be in the range:
5758 /// 1 <= |Value| <= ElementBits for a right shift; or
5759 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5760 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5762 assert(VT.isVector() && "vector shift count is not a vector type");
5763 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5764 if (! getVShiftImm(Op, ElementBits, Cnt))
5768 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
5771 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
5772 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
5773 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5776 // Don't do anything for most intrinsics.
5779 // Vector shifts: check for immediate versions and lower them.
5780 // Note: This is done during DAG combining instead of DAG legalizing because
5781 // the build_vectors for 64-bit vector element shift counts are generally
5782 // not legal, and it is hard to see their values after they get legalized to
5783 // loads from a constant pool.
5784 case Intrinsic::arm_neon_vshifts:
5785 case Intrinsic::arm_neon_vshiftu:
5786 case Intrinsic::arm_neon_vshiftls:
5787 case Intrinsic::arm_neon_vshiftlu:
5788 case Intrinsic::arm_neon_vshiftn:
5789 case Intrinsic::arm_neon_vrshifts:
5790 case Intrinsic::arm_neon_vrshiftu:
5791 case Intrinsic::arm_neon_vrshiftn:
5792 case Intrinsic::arm_neon_vqshifts:
5793 case Intrinsic::arm_neon_vqshiftu:
5794 case Intrinsic::arm_neon_vqshiftsu:
5795 case Intrinsic::arm_neon_vqshiftns:
5796 case Intrinsic::arm_neon_vqshiftnu:
5797 case Intrinsic::arm_neon_vqshiftnsu:
5798 case Intrinsic::arm_neon_vqrshiftns:
5799 case Intrinsic::arm_neon_vqrshiftnu:
5800 case Intrinsic::arm_neon_vqrshiftnsu: {
5801 EVT VT = N->getOperand(1).getValueType();
5803 unsigned VShiftOpc = 0;
5806 case Intrinsic::arm_neon_vshifts:
5807 case Intrinsic::arm_neon_vshiftu:
5808 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
5809 VShiftOpc = ARMISD::VSHL;
5812 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
5813 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
5814 ARMISD::VSHRs : ARMISD::VSHRu);
5819 case Intrinsic::arm_neon_vshiftls:
5820 case Intrinsic::arm_neon_vshiftlu:
5821 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
5823 llvm_unreachable("invalid shift count for vshll intrinsic");
5825 case Intrinsic::arm_neon_vrshifts:
5826 case Intrinsic::arm_neon_vrshiftu:
5827 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
5831 case Intrinsic::arm_neon_vqshifts:
5832 case Intrinsic::arm_neon_vqshiftu:
5833 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5837 case Intrinsic::arm_neon_vqshiftsu:
5838 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5840 llvm_unreachable("invalid shift count for vqshlu intrinsic");
5842 case Intrinsic::arm_neon_vshiftn:
5843 case Intrinsic::arm_neon_vrshiftn:
5844 case Intrinsic::arm_neon_vqshiftns:
5845 case Intrinsic::arm_neon_vqshiftnu:
5846 case Intrinsic::arm_neon_vqshiftnsu:
5847 case Intrinsic::arm_neon_vqrshiftns:
5848 case Intrinsic::arm_neon_vqrshiftnu:
5849 case Intrinsic::arm_neon_vqrshiftnsu:
5850 // Narrowing shifts require an immediate right shift.
5851 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
5853 llvm_unreachable("invalid shift count for narrowing vector shift "
5857 llvm_unreachable("unhandled vector shift");
5861 case Intrinsic::arm_neon_vshifts:
5862 case Intrinsic::arm_neon_vshiftu:
5863 // Opcode already set above.
5865 case Intrinsic::arm_neon_vshiftls:
5866 case Intrinsic::arm_neon_vshiftlu:
5867 if (Cnt == VT.getVectorElementType().getSizeInBits())
5868 VShiftOpc = ARMISD::VSHLLi;
5870 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
5871 ARMISD::VSHLLs : ARMISD::VSHLLu);
5873 case Intrinsic::arm_neon_vshiftn:
5874 VShiftOpc = ARMISD::VSHRN; break;
5875 case Intrinsic::arm_neon_vrshifts:
5876 VShiftOpc = ARMISD::VRSHRs; break;
5877 case Intrinsic::arm_neon_vrshiftu:
5878 VShiftOpc = ARMISD::VRSHRu; break;
5879 case Intrinsic::arm_neon_vrshiftn:
5880 VShiftOpc = ARMISD::VRSHRN; break;
5881 case Intrinsic::arm_neon_vqshifts:
5882 VShiftOpc = ARMISD::VQSHLs; break;
5883 case Intrinsic::arm_neon_vqshiftu:
5884 VShiftOpc = ARMISD::VQSHLu; break;
5885 case Intrinsic::arm_neon_vqshiftsu:
5886 VShiftOpc = ARMISD::VQSHLsu; break;
5887 case Intrinsic::arm_neon_vqshiftns:
5888 VShiftOpc = ARMISD::VQSHRNs; break;
5889 case Intrinsic::arm_neon_vqshiftnu:
5890 VShiftOpc = ARMISD::VQSHRNu; break;
5891 case Intrinsic::arm_neon_vqshiftnsu:
5892 VShiftOpc = ARMISD::VQSHRNsu; break;
5893 case Intrinsic::arm_neon_vqrshiftns:
5894 VShiftOpc = ARMISD::VQRSHRNs; break;
5895 case Intrinsic::arm_neon_vqrshiftnu:
5896 VShiftOpc = ARMISD::VQRSHRNu; break;
5897 case Intrinsic::arm_neon_vqrshiftnsu:
5898 VShiftOpc = ARMISD::VQRSHRNsu; break;
5901 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5902 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
5905 case Intrinsic::arm_neon_vshiftins: {
5906 EVT VT = N->getOperand(1).getValueType();
5908 unsigned VShiftOpc = 0;
5910 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
5911 VShiftOpc = ARMISD::VSLI;
5912 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
5913 VShiftOpc = ARMISD::VSRI;
5915 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
5918 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5919 N->getOperand(1), N->getOperand(2),
5920 DAG.getConstant(Cnt, MVT::i32));
5923 case Intrinsic::arm_neon_vqrshifts:
5924 case Intrinsic::arm_neon_vqrshiftu:
5925 // No immediate versions of these to check for.
5932 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
5933 /// lowers them. As with the vector shift intrinsics, this is done during DAG
5934 /// combining instead of DAG legalizing because the build_vectors for 64-bit
5935 /// vector element shift counts are generally not legal, and it is hard to see
5936 /// their values after they get legalized to loads from a constant pool.
5937 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
5938 const ARMSubtarget *ST) {
5939 EVT VT = N->getValueType(0);
5941 // Nothing to be done for scalar shifts.
5942 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5943 if (!VT.isVector() || !TLI.isTypeLegal(VT))
5946 assert(ST->hasNEON() && "unexpected vector shift");
5949 switch (N->getOpcode()) {
5950 default: llvm_unreachable("unexpected shift opcode");
5953 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
5954 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
5955 DAG.getConstant(Cnt, MVT::i32));
5960 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
5961 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
5962 ARMISD::VSHRs : ARMISD::VSHRu);
5963 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
5964 DAG.getConstant(Cnt, MVT::i32));
5970 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
5971 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
5972 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
5973 const ARMSubtarget *ST) {
5974 SDValue N0 = N->getOperand(0);
5976 // Check for sign- and zero-extensions of vector extract operations of 8-
5977 // and 16-bit vector elements. NEON supports these directly. They are
5978 // handled during DAG combining because type legalization will promote them
5979 // to 32-bit types and it is messy to recognize the operations after that.
5980 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
5981 SDValue Vec = N0.getOperand(0);
5982 SDValue Lane = N0.getOperand(1);
5983 EVT VT = N->getValueType(0);
5984 EVT EltVT = N0.getValueType();
5985 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5987 if (VT == MVT::i32 &&
5988 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
5989 TLI.isTypeLegal(Vec.getValueType()) &&
5990 isa<ConstantSDNode>(Lane)) {
5993 switch (N->getOpcode()) {
5994 default: llvm_unreachable("unexpected opcode");
5995 case ISD::SIGN_EXTEND:
5996 Opc = ARMISD::VGETLANEs;
5998 case ISD::ZERO_EXTEND:
5999 case ISD::ANY_EXTEND:
6000 Opc = ARMISD::VGETLANEu;
6003 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
6010 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
6011 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
6012 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
6013 const ARMSubtarget *ST) {
6014 // If the target supports NEON, try to use vmax/vmin instructions for f32
6015 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
6016 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
6017 // a NaN; only do the transformation when it matches that behavior.
6019 // For now only do this when using NEON for FP operations; if using VFP, it
6020 // is not obvious that the benefit outweighs the cost of switching to the
6022 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
6023 N->getValueType(0) != MVT::f32)
6026 SDValue CondLHS = N->getOperand(0);
6027 SDValue CondRHS = N->getOperand(1);
6028 SDValue LHS = N->getOperand(2);
6029 SDValue RHS = N->getOperand(3);
6030 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
6032 unsigned Opcode = 0;
6034 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
6035 IsReversed = false; // x CC y ? x : y
6036 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
6037 IsReversed = true ; // x CC y ? y : x
6051 // If LHS is NaN, an ordered comparison will be false and the result will
6052 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
6053 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6054 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
6055 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6057 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
6058 // will return -0, so vmin can only be used for unsafe math or if one of
6059 // the operands is known to be nonzero.
6060 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
6062 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6064 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
6073 // If LHS is NaN, an ordered comparison will be false and the result will
6074 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
6075 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6076 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
6077 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6079 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
6080 // will return +0, so vmax can only be used for unsafe math or if one of
6081 // the operands is known to be nonzero.
6082 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
6084 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6086 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
6092 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
6095 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
6096 DAGCombinerInfo &DCI) const {
6097 switch (N->getOpcode()) {
6099 case ISD::ADD: return PerformADDCombine(N, DCI);
6100 case ISD::SUB: return PerformSUBCombine(N, DCI);
6101 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
6102 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
6103 case ISD::AND: return PerformANDCombine(N, DCI);
6104 case ARMISD::BFI: return PerformBFICombine(N, DCI);
6105 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
6106 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
6107 case ISD::STORE: return PerformSTORECombine(N, DCI);
6108 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
6109 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
6110 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
6111 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
6112 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
6115 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
6116 case ISD::SIGN_EXTEND:
6117 case ISD::ZERO_EXTEND:
6118 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
6119 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
6120 case ARMISD::VLD2DUP:
6121 case ARMISD::VLD3DUP:
6122 case ARMISD::VLD4DUP:
6123 return CombineBaseUpdate(N, DCI);
6124 case ISD::INTRINSIC_VOID:
6125 case ISD::INTRINSIC_W_CHAIN:
6126 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
6127 case Intrinsic::arm_neon_vld1:
6128 case Intrinsic::arm_neon_vld2:
6129 case Intrinsic::arm_neon_vld3:
6130 case Intrinsic::arm_neon_vld4:
6131 case Intrinsic::arm_neon_vld2lane:
6132 case Intrinsic::arm_neon_vld3lane:
6133 case Intrinsic::arm_neon_vld4lane:
6134 case Intrinsic::arm_neon_vst1:
6135 case Intrinsic::arm_neon_vst2:
6136 case Intrinsic::arm_neon_vst3:
6137 case Intrinsic::arm_neon_vst4:
6138 case Intrinsic::arm_neon_vst2lane:
6139 case Intrinsic::arm_neon_vst3lane:
6140 case Intrinsic::arm_neon_vst4lane:
6141 return CombineBaseUpdate(N, DCI);
6149 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
6151 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
6154 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
6155 if (!Subtarget->allowsUnalignedMem())
6158 switch (VT.getSimpleVT().SimpleTy) {
6165 // FIXME: VLD1 etc with standard alignment is legal.
6169 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
6174 switch (VT.getSimpleVT().SimpleTy) {
6175 default: return false;
6190 if ((V & (Scale - 1)) != 0)
6193 return V == (V & ((1LL << 5) - 1));
6196 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
6197 const ARMSubtarget *Subtarget) {
6204 switch (VT.getSimpleVT().SimpleTy) {
6205 default: return false;
6210 // + imm12 or - imm8
6212 return V == (V & ((1LL << 8) - 1));
6213 return V == (V & ((1LL << 12) - 1));
6216 // Same as ARM mode. FIXME: NEON?
6217 if (!Subtarget->hasVFP2())
6222 return V == (V & ((1LL << 8) - 1));
6226 /// isLegalAddressImmediate - Return true if the integer value can be used
6227 /// as the offset of the target addressing mode for load / store of the
6229 static bool isLegalAddressImmediate(int64_t V, EVT VT,
6230 const ARMSubtarget *Subtarget) {
6237 if (Subtarget->isThumb1Only())
6238 return isLegalT1AddressImmediate(V, VT);
6239 else if (Subtarget->isThumb2())
6240 return isLegalT2AddressImmediate(V, VT, Subtarget);
6245 switch (VT.getSimpleVT().SimpleTy) {
6246 default: return false;
6251 return V == (V & ((1LL << 12) - 1));
6254 return V == (V & ((1LL << 8) - 1));
6257 if (!Subtarget->hasVFP2()) // FIXME: NEON?
6262 return V == (V & ((1LL << 8) - 1));
6266 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
6268 int Scale = AM.Scale;
6272 switch (VT.getSimpleVT().SimpleTy) {
6273 default: return false;
6282 return Scale == 2 || Scale == 4 || Scale == 8;
6285 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6289 // Note, we allow "void" uses (basically, uses that aren't loads or
6290 // stores), because arm allows folding a scale into many arithmetic
6291 // operations. This should be made more precise and revisited later.
6293 // Allow r << imm, but the imm has to be a multiple of two.
6294 if (Scale & 1) return false;
6295 return isPowerOf2_32(Scale);
6299 /// isLegalAddressingMode - Return true if the addressing mode represented
6300 /// by AM is legal for this target, for a load/store of the specified type.
6301 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6302 const Type *Ty) const {
6303 EVT VT = getValueType(Ty, true);
6304 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
6307 // Can never fold addr of global into load/store.
6312 case 0: // no scale reg, must be "r+i" or "r", or "i".
6315 if (Subtarget->isThumb1Only())
6319 // ARM doesn't support any R+R*scale+imm addr modes.
6326 if (Subtarget->isThumb2())
6327 return isLegalT2ScaledAddressingMode(AM, VT);
6329 int Scale = AM.Scale;
6330 switch (VT.getSimpleVT().SimpleTy) {
6331 default: return false;
6335 if (Scale < 0) Scale = -Scale;
6339 return isPowerOf2_32(Scale & ~1);
6343 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6348 // Note, we allow "void" uses (basically, uses that aren't loads or
6349 // stores), because arm allows folding a scale into many arithmetic
6350 // operations. This should be made more precise and revisited later.
6352 // Allow r << imm, but the imm has to be a multiple of two.
6353 if (Scale & 1) return false;
6354 return isPowerOf2_32(Scale);
6361 /// isLegalICmpImmediate - Return true if the specified immediate is legal
6362 /// icmp immediate, that is the target has icmp instructions which can compare
6363 /// a register against the immediate without having to materialize the
6364 /// immediate into a register.
6365 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
6366 if (!Subtarget->isThumb())
6367 return ARM_AM::getSOImmVal(Imm) != -1;
6368 if (Subtarget->isThumb2())
6369 return ARM_AM::getT2SOImmVal(Imm) != -1;
6370 return Imm >= 0 && Imm <= 255;
6373 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
6374 bool isSEXTLoad, SDValue &Base,
6375 SDValue &Offset, bool &isInc,
6376 SelectionDAG &DAG) {
6377 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6380 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
6382 Base = Ptr->getOperand(0);
6383 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6384 int RHSC = (int)RHS->getZExtValue();
6385 if (RHSC < 0 && RHSC > -256) {
6386 assert(Ptr->getOpcode() == ISD::ADD);
6388 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6392 isInc = (Ptr->getOpcode() == ISD::ADD);
6393 Offset = Ptr->getOperand(1);
6395 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
6397 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6398 int RHSC = (int)RHS->getZExtValue();
6399 if (RHSC < 0 && RHSC > -0x1000) {
6400 assert(Ptr->getOpcode() == ISD::ADD);
6402 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6403 Base = Ptr->getOperand(0);
6408 if (Ptr->getOpcode() == ISD::ADD) {
6410 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
6411 if (ShOpcVal != ARM_AM::no_shift) {
6412 Base = Ptr->getOperand(1);
6413 Offset = Ptr->getOperand(0);
6415 Base = Ptr->getOperand(0);
6416 Offset = Ptr->getOperand(1);
6421 isInc = (Ptr->getOpcode() == ISD::ADD);
6422 Base = Ptr->getOperand(0);
6423 Offset = Ptr->getOperand(1);
6427 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
6431 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
6432 bool isSEXTLoad, SDValue &Base,
6433 SDValue &Offset, bool &isInc,
6434 SelectionDAG &DAG) {
6435 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6438 Base = Ptr->getOperand(0);
6439 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6440 int RHSC = (int)RHS->getZExtValue();
6441 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
6442 assert(Ptr->getOpcode() == ISD::ADD);
6444 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6446 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
6447 isInc = Ptr->getOpcode() == ISD::ADD;
6448 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
6456 /// getPreIndexedAddressParts - returns true by value, base pointer and
6457 /// offset pointer and addressing mode by reference if the node's address
6458 /// can be legally represented as pre-indexed load / store address.
6460 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
6462 ISD::MemIndexedMode &AM,
6463 SelectionDAG &DAG) const {
6464 if (Subtarget->isThumb1Only())
6469 bool isSEXTLoad = false;
6470 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6471 Ptr = LD->getBasePtr();
6472 VT = LD->getMemoryVT();
6473 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6474 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6475 Ptr = ST->getBasePtr();
6476 VT = ST->getMemoryVT();
6481 bool isLegal = false;
6482 if (Subtarget->isThumb2())
6483 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6484 Offset, isInc, DAG);
6486 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6487 Offset, isInc, DAG);
6491 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
6495 /// getPostIndexedAddressParts - returns true by value, base pointer and
6496 /// offset pointer and addressing mode by reference if this node can be
6497 /// combined with a load / store to form a post-indexed load / store.
6498 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
6501 ISD::MemIndexedMode &AM,
6502 SelectionDAG &DAG) const {
6503 if (Subtarget->isThumb1Only())
6508 bool isSEXTLoad = false;
6509 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6510 VT = LD->getMemoryVT();
6511 Ptr = LD->getBasePtr();
6512 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6513 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6514 VT = ST->getMemoryVT();
6515 Ptr = ST->getBasePtr();
6520 bool isLegal = false;
6521 if (Subtarget->isThumb2())
6522 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6525 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6531 // Swap base ptr and offset to catch more post-index load / store when
6532 // it's legal. In Thumb2 mode, offset must be an immediate.
6533 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
6534 !Subtarget->isThumb2())
6535 std::swap(Base, Offset);
6537 // Post-indexed load / store update the base pointer.
6542 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
6546 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6550 const SelectionDAG &DAG,
6551 unsigned Depth) const {
6552 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
6553 switch (Op.getOpcode()) {
6555 case ARMISD::CMOV: {
6556 // Bits are known zero/one if known on the LHS and RHS.
6557 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
6558 if (KnownZero == 0 && KnownOne == 0) return;
6560 APInt KnownZeroRHS, KnownOneRHS;
6561 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
6562 KnownZeroRHS, KnownOneRHS, Depth+1);
6563 KnownZero &= KnownZeroRHS;
6564 KnownOne &= KnownOneRHS;
6570 //===----------------------------------------------------------------------===//
6571 // ARM Inline Assembly Support
6572 //===----------------------------------------------------------------------===//
6574 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
6575 // Looking for "rev" which is V6+.
6576 if (!Subtarget->hasV6Ops())
6579 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
6580 std::string AsmStr = IA->getAsmString();
6581 SmallVector<StringRef, 4> AsmPieces;
6582 SplitString(AsmStr, AsmPieces, ";\n");
6584 switch (AsmPieces.size()) {
6585 default: return false;
6587 AsmStr = AsmPieces[0];
6589 SplitString(AsmStr, AsmPieces, " \t,");
6592 if (AsmPieces.size() == 3 &&
6593 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
6594 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
6595 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
6596 if (Ty && Ty->getBitWidth() == 32)
6597 return IntrinsicLowering::LowerToByteSwap(CI);
6605 /// getConstraintType - Given a constraint letter, return the type of
6606 /// constraint it is for this target.
6607 ARMTargetLowering::ConstraintType
6608 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
6609 if (Constraint.size() == 1) {
6610 switch (Constraint[0]) {
6612 case 'l': return C_RegisterClass;
6613 case 'w': return C_RegisterClass;
6616 return TargetLowering::getConstraintType(Constraint);
6619 /// Examine constraint type and operand type and determine a weight value.
6620 /// This object must already have been set up with the operand type
6621 /// and the current alternative constraint selected.
6622 TargetLowering::ConstraintWeight
6623 ARMTargetLowering::getSingleConstraintMatchWeight(
6624 AsmOperandInfo &info, const char *constraint) const {
6625 ConstraintWeight weight = CW_Invalid;
6626 Value *CallOperandVal = info.CallOperandVal;
6627 // If we don't have a value, we can't do a match,
6628 // but allow it at the lowest weight.
6629 if (CallOperandVal == NULL)
6631 const Type *type = CallOperandVal->getType();
6632 // Look at the constraint type.
6633 switch (*constraint) {
6635 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6638 if (type->isIntegerTy()) {
6639 if (Subtarget->isThumb())
6640 weight = CW_SpecificReg;
6642 weight = CW_Register;
6646 if (type->isFloatingPointTy())
6647 weight = CW_Register;
6653 std::pair<unsigned, const TargetRegisterClass*>
6654 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6656 if (Constraint.size() == 1) {
6657 // GCC ARM Constraint Letters
6658 switch (Constraint[0]) {
6660 if (Subtarget->isThumb())
6661 return std::make_pair(0U, ARM::tGPRRegisterClass);
6663 return std::make_pair(0U, ARM::GPRRegisterClass);
6665 return std::make_pair(0U, ARM::GPRRegisterClass);
6668 return std::make_pair(0U, ARM::SPRRegisterClass);
6669 if (VT.getSizeInBits() == 64)
6670 return std::make_pair(0U, ARM::DPRRegisterClass);
6671 if (VT.getSizeInBits() == 128)
6672 return std::make_pair(0U, ARM::QPRRegisterClass);
6676 if (StringRef("{cc}").equals_lower(Constraint))
6677 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
6679 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6682 std::vector<unsigned> ARMTargetLowering::
6683 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6685 if (Constraint.size() != 1)
6686 return std::vector<unsigned>();
6688 switch (Constraint[0]) { // GCC ARM Constraint Letters
6691 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6692 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6695 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6696 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6697 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
6698 ARM::R12, ARM::LR, 0);
6701 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
6702 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
6703 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
6704 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
6705 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
6706 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
6707 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
6708 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
6709 if (VT.getSizeInBits() == 64)
6710 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
6711 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
6712 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
6713 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
6714 if (VT.getSizeInBits() == 128)
6715 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
6716 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
6720 return std::vector<unsigned>();
6723 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6724 /// vector. If it is invalid, don't add anything to Ops.
6725 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6727 std::vector<SDValue>&Ops,
6728 SelectionDAG &DAG) const {
6729 SDValue Result(0, 0);
6731 switch (Constraint) {
6733 case 'I': case 'J': case 'K': case 'L':
6734 case 'M': case 'N': case 'O':
6735 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
6739 int64_t CVal64 = C->getSExtValue();
6740 int CVal = (int) CVal64;
6741 // None of these constraints allow values larger than 32 bits. Check
6742 // that the value fits in an int.
6746 switch (Constraint) {
6748 if (Subtarget->isThumb1Only()) {
6749 // This must be a constant between 0 and 255, for ADD
6751 if (CVal >= 0 && CVal <= 255)
6753 } else if (Subtarget->isThumb2()) {
6754 // A constant that can be used as an immediate value in a
6755 // data-processing instruction.
6756 if (ARM_AM::getT2SOImmVal(CVal) != -1)
6759 // A constant that can be used as an immediate value in a
6760 // data-processing instruction.
6761 if (ARM_AM::getSOImmVal(CVal) != -1)
6767 if (Subtarget->isThumb()) { // FIXME thumb2
6768 // This must be a constant between -255 and -1, for negated ADD
6769 // immediates. This can be used in GCC with an "n" modifier that
6770 // prints the negated value, for use with SUB instructions. It is
6771 // not useful otherwise but is implemented for compatibility.
6772 if (CVal >= -255 && CVal <= -1)
6775 // This must be a constant between -4095 and 4095. It is not clear
6776 // what this constraint is intended for. Implemented for
6777 // compatibility with GCC.
6778 if (CVal >= -4095 && CVal <= 4095)
6784 if (Subtarget->isThumb1Only()) {
6785 // A 32-bit value where only one byte has a nonzero value. Exclude
6786 // zero to match GCC. This constraint is used by GCC internally for
6787 // constants that can be loaded with a move/shift combination.
6788 // It is not useful otherwise but is implemented for compatibility.
6789 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
6791 } else if (Subtarget->isThumb2()) {
6792 // A constant whose bitwise inverse can be used as an immediate
6793 // value in a data-processing instruction. This can be used in GCC
6794 // with a "B" modifier that prints the inverted value, for use with
6795 // BIC and MVN instructions. It is not useful otherwise but is
6796 // implemented for compatibility.
6797 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
6800 // A constant whose bitwise inverse can be used as an immediate
6801 // value in a data-processing instruction. This can be used in GCC
6802 // with a "B" modifier that prints the inverted value, for use with
6803 // BIC and MVN instructions. It is not useful otherwise but is
6804 // implemented for compatibility.
6805 if (ARM_AM::getSOImmVal(~CVal) != -1)
6811 if (Subtarget->isThumb1Only()) {
6812 // This must be a constant between -7 and 7,
6813 // for 3-operand ADD/SUB immediate instructions.
6814 if (CVal >= -7 && CVal < 7)
6816 } else if (Subtarget->isThumb2()) {
6817 // A constant whose negation can be used as an immediate value in a
6818 // data-processing instruction. This can be used in GCC with an "n"
6819 // modifier that prints the negated value, for use with SUB
6820 // instructions. It is not useful otherwise but is implemented for
6822 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
6825 // A constant whose negation can be used as an immediate value in a
6826 // data-processing instruction. This can be used in GCC with an "n"
6827 // modifier that prints the negated value, for use with SUB
6828 // instructions. It is not useful otherwise but is implemented for
6830 if (ARM_AM::getSOImmVal(-CVal) != -1)
6836 if (Subtarget->isThumb()) { // FIXME thumb2
6837 // This must be a multiple of 4 between 0 and 1020, for
6838 // ADD sp + immediate.
6839 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
6842 // A power of two or a constant between 0 and 32. This is used in
6843 // GCC for the shift amount on shifted register operands, but it is
6844 // useful in general for any shift amounts.
6845 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
6851 if (Subtarget->isThumb()) { // FIXME thumb2
6852 // This must be a constant between 0 and 31, for shift amounts.
6853 if (CVal >= 0 && CVal <= 31)
6859 if (Subtarget->isThumb()) { // FIXME thumb2
6860 // This must be a multiple of 4 between -508 and 508, for
6861 // ADD/SUB sp = sp + immediate.
6862 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
6867 Result = DAG.getTargetConstant(CVal, Op.getValueType());
6871 if (Result.getNode()) {
6872 Ops.push_back(Result);
6875 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6879 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6880 // The ARM target isn't yet aware of offsets.
6884 int ARM::getVFPf32Imm(const APFloat &FPImm) {
6885 APInt Imm = FPImm.bitcastToAPInt();
6886 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
6887 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
6888 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
6890 // We can handle 4 bits of mantissa.
6891 // mantissa = (16+UInt(e:f:g:h))/16.
6892 if (Mantissa & 0x7ffff)
6895 if ((Mantissa & 0xf) != Mantissa)
6898 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6899 if (Exp < -3 || Exp > 4)
6901 Exp = ((Exp+3) & 0x7) ^ 4;
6903 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6906 int ARM::getVFPf64Imm(const APFloat &FPImm) {
6907 APInt Imm = FPImm.bitcastToAPInt();
6908 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
6909 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
6910 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
6912 // We can handle 4 bits of mantissa.
6913 // mantissa = (16+UInt(e:f:g:h))/16.
6914 if (Mantissa & 0xffffffffffffLL)
6917 if ((Mantissa & 0xf) != Mantissa)
6920 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6921 if (Exp < -3 || Exp > 4)
6923 Exp = ((Exp+3) & 0x7) ^ 4;
6925 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6928 bool ARM::isBitFieldInvertedMask(unsigned v) {
6929 if (v == 0xffffffff)
6931 // there can be 1's on either or both "outsides", all the "inside"
6933 unsigned int lsb = 0, msb = 31;
6934 while (v & (1 << msb)) --msb;
6935 while (v & (1 << lsb)) ++lsb;
6936 for (unsigned int i = lsb; i <= msb; ++i) {
6943 /// isFPImmLegal - Returns true if the target can instruction select the
6944 /// specified FP immediate natively. If false, the legalizer will
6945 /// materialize the FP immediate as a load from a constant pool.
6946 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
6947 if (!Subtarget->hasVFP3())
6950 return ARM::getVFPf32Imm(Imm) != -1;
6952 return ARM::getVFPf64Imm(Imm) != -1;
6956 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6957 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6958 /// specified in the intrinsic calls.
6959 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6961 unsigned Intrinsic) const {
6962 switch (Intrinsic) {
6963 case Intrinsic::arm_neon_vld1:
6964 case Intrinsic::arm_neon_vld2:
6965 case Intrinsic::arm_neon_vld3:
6966 case Intrinsic::arm_neon_vld4:
6967 case Intrinsic::arm_neon_vld2lane:
6968 case Intrinsic::arm_neon_vld3lane:
6969 case Intrinsic::arm_neon_vld4lane: {
6970 Info.opc = ISD::INTRINSIC_W_CHAIN;
6971 // Conservatively set memVT to the entire set of vectors loaded.
6972 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
6973 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6974 Info.ptrVal = I.getArgOperand(0);
6976 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6977 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6978 Info.vol = false; // volatile loads with NEON intrinsics not supported
6979 Info.readMem = true;
6980 Info.writeMem = false;
6983 case Intrinsic::arm_neon_vst1:
6984 case Intrinsic::arm_neon_vst2:
6985 case Intrinsic::arm_neon_vst3:
6986 case Intrinsic::arm_neon_vst4:
6987 case Intrinsic::arm_neon_vst2lane:
6988 case Intrinsic::arm_neon_vst3lane:
6989 case Intrinsic::arm_neon_vst4lane: {
6990 Info.opc = ISD::INTRINSIC_VOID;
6991 // Conservatively set memVT to the entire set of vectors stored.
6992 unsigned NumElts = 0;
6993 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6994 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
6995 if (!ArgTy->isVectorTy())
6997 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
6999 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7000 Info.ptrVal = I.getArgOperand(0);
7002 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7003 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7004 Info.vol = false; // volatile stores with NEON intrinsics not supported
7005 Info.readMem = false;
7006 Info.writeMem = true;