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";
782 case ARMISD::RBIT: return "ARMISD::RBIT";
784 case ARMISD::FTOSI: return "ARMISD::FTOSI";
785 case ARMISD::FTOUI: return "ARMISD::FTOUI";
786 case ARMISD::SITOF: return "ARMISD::SITOF";
787 case ARMISD::UITOF: return "ARMISD::UITOF";
789 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
790 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
791 case ARMISD::RRX: return "ARMISD::RRX";
793 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
794 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
796 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
797 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
798 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
800 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
802 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
804 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
806 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
807 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
809 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
811 case ARMISD::VCEQ: return "ARMISD::VCEQ";
812 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
813 case ARMISD::VCGE: return "ARMISD::VCGE";
814 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
815 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
816 case ARMISD::VCGEU: return "ARMISD::VCGEU";
817 case ARMISD::VCGT: return "ARMISD::VCGT";
818 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
819 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
820 case ARMISD::VCGTU: return "ARMISD::VCGTU";
821 case ARMISD::VTST: return "ARMISD::VTST";
823 case ARMISD::VSHL: return "ARMISD::VSHL";
824 case ARMISD::VSHRs: return "ARMISD::VSHRs";
825 case ARMISD::VSHRu: return "ARMISD::VSHRu";
826 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
827 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
828 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
829 case ARMISD::VSHRN: return "ARMISD::VSHRN";
830 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
831 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
832 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
833 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
834 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
835 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
836 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
837 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
838 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
839 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
840 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
841 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
842 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
843 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
844 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
845 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
846 case ARMISD::VDUP: return "ARMISD::VDUP";
847 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
848 case ARMISD::VEXT: return "ARMISD::VEXT";
849 case ARMISD::VREV64: return "ARMISD::VREV64";
850 case ARMISD::VREV32: return "ARMISD::VREV32";
851 case ARMISD::VREV16: return "ARMISD::VREV16";
852 case ARMISD::VZIP: return "ARMISD::VZIP";
853 case ARMISD::VUZP: return "ARMISD::VUZP";
854 case ARMISD::VTRN: return "ARMISD::VTRN";
855 case ARMISD::VMULLs: return "ARMISD::VMULLs";
856 case ARMISD::VMULLu: return "ARMISD::VMULLu";
857 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
858 case ARMISD::FMAX: return "ARMISD::FMAX";
859 case ARMISD::FMIN: return "ARMISD::FMIN";
860 case ARMISD::BFI: return "ARMISD::BFI";
861 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
862 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
863 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
864 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
865 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
866 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
867 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
868 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
869 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
870 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
871 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
872 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
873 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
874 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
875 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
876 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
877 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
878 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
879 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
880 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
881 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
882 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
886 /// getRegClassFor - Return the register class that should be used for the
887 /// specified value type.
888 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
889 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
890 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
891 // load / store 4 to 8 consecutive D registers.
892 if (Subtarget->hasNEON()) {
893 if (VT == MVT::v4i64)
894 return ARM::QQPRRegisterClass;
895 else if (VT == MVT::v8i64)
896 return ARM::QQQQPRRegisterClass;
898 return TargetLowering::getRegClassFor(VT);
901 // Create a fast isel object.
903 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
904 return ARM::createFastISel(funcInfo);
907 /// getFunctionAlignment - Return the Log2 alignment of this function.
908 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
909 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
912 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
913 /// be used for loads / stores from the global.
914 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
915 return (Subtarget->isThumb1Only() ? 127 : 4095);
918 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
919 unsigned NumVals = N->getNumValues();
921 return Sched::RegPressure;
923 for (unsigned i = 0; i != NumVals; ++i) {
924 EVT VT = N->getValueType(i);
925 if (VT == MVT::Glue || VT == MVT::Other)
927 if (VT.isFloatingPoint() || VT.isVector())
928 return Sched::Latency;
931 if (!N->isMachineOpcode())
932 return Sched::RegPressure;
934 // Load are scheduled for latency even if there instruction itinerary
936 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
937 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
939 if (TID.getNumDefs() == 0)
940 return Sched::RegPressure;
941 if (!Itins->isEmpty() &&
942 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
943 return Sched::Latency;
945 return Sched::RegPressure;
948 //===----------------------------------------------------------------------===//
950 //===----------------------------------------------------------------------===//
952 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
953 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
955 default: llvm_unreachable("Unknown condition code!");
956 case ISD::SETNE: return ARMCC::NE;
957 case ISD::SETEQ: return ARMCC::EQ;
958 case ISD::SETGT: return ARMCC::GT;
959 case ISD::SETGE: return ARMCC::GE;
960 case ISD::SETLT: return ARMCC::LT;
961 case ISD::SETLE: return ARMCC::LE;
962 case ISD::SETUGT: return ARMCC::HI;
963 case ISD::SETUGE: return ARMCC::HS;
964 case ISD::SETULT: return ARMCC::LO;
965 case ISD::SETULE: return ARMCC::LS;
969 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
970 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
971 ARMCC::CondCodes &CondCode2) {
972 CondCode2 = ARMCC::AL;
974 default: llvm_unreachable("Unknown FP condition!");
976 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
978 case ISD::SETOGT: CondCode = ARMCC::GT; break;
980 case ISD::SETOGE: CondCode = ARMCC::GE; break;
981 case ISD::SETOLT: CondCode = ARMCC::MI; break;
982 case ISD::SETOLE: CondCode = ARMCC::LS; break;
983 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
984 case ISD::SETO: CondCode = ARMCC::VC; break;
985 case ISD::SETUO: CondCode = ARMCC::VS; break;
986 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
987 case ISD::SETUGT: CondCode = ARMCC::HI; break;
988 case ISD::SETUGE: CondCode = ARMCC::PL; break;
990 case ISD::SETULT: CondCode = ARMCC::LT; break;
992 case ISD::SETULE: CondCode = ARMCC::LE; break;
994 case ISD::SETUNE: CondCode = ARMCC::NE; break;
998 //===----------------------------------------------------------------------===//
999 // Calling Convention Implementation
1000 //===----------------------------------------------------------------------===//
1002 #include "ARMGenCallingConv.inc"
1004 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1005 /// given CallingConvention value.
1006 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1008 bool isVarArg) const {
1011 llvm_unreachable("Unsupported calling convention");
1012 case CallingConv::Fast:
1013 if (Subtarget->hasVFP2() && !isVarArg) {
1014 if (!Subtarget->isAAPCS_ABI())
1015 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1016 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1017 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1020 case CallingConv::C: {
1021 // Use target triple & subtarget features to do actual dispatch.
1022 if (!Subtarget->isAAPCS_ABI())
1023 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1024 else if (Subtarget->hasVFP2() &&
1025 FloatABIType == FloatABI::Hard && !isVarArg)
1026 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1027 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1029 case CallingConv::ARM_AAPCS_VFP:
1030 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1031 case CallingConv::ARM_AAPCS:
1032 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1033 case CallingConv::ARM_APCS:
1034 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1038 /// LowerCallResult - Lower the result values of a call into the
1039 /// appropriate copies out of appropriate physical registers.
1041 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1042 CallingConv::ID CallConv, bool isVarArg,
1043 const SmallVectorImpl<ISD::InputArg> &Ins,
1044 DebugLoc dl, SelectionDAG &DAG,
1045 SmallVectorImpl<SDValue> &InVals) const {
1047 // Assign locations to each value returned by this call.
1048 SmallVector<CCValAssign, 16> RVLocs;
1049 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1050 RVLocs, *DAG.getContext());
1051 CCInfo.AnalyzeCallResult(Ins,
1052 CCAssignFnForNode(CallConv, /* Return*/ true,
1055 // Copy all of the result registers out of their specified physreg.
1056 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1057 CCValAssign VA = RVLocs[i];
1060 if (VA.needsCustom()) {
1061 // Handle f64 or half of a v2f64.
1062 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1064 Chain = Lo.getValue(1);
1065 InFlag = Lo.getValue(2);
1066 VA = RVLocs[++i]; // skip ahead to next loc
1067 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1069 Chain = Hi.getValue(1);
1070 InFlag = Hi.getValue(2);
1071 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1073 if (VA.getLocVT() == MVT::v2f64) {
1074 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1075 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1076 DAG.getConstant(0, MVT::i32));
1078 VA = RVLocs[++i]; // skip ahead to next loc
1079 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1080 Chain = Lo.getValue(1);
1081 InFlag = Lo.getValue(2);
1082 VA = RVLocs[++i]; // skip ahead to next loc
1083 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1084 Chain = Hi.getValue(1);
1085 InFlag = Hi.getValue(2);
1086 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1087 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1088 DAG.getConstant(1, MVT::i32));
1091 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1093 Chain = Val.getValue(1);
1094 InFlag = Val.getValue(2);
1097 switch (VA.getLocInfo()) {
1098 default: llvm_unreachable("Unknown loc info!");
1099 case CCValAssign::Full: break;
1100 case CCValAssign::BCvt:
1101 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1105 InVals.push_back(Val);
1111 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1112 /// by "Src" to address "Dst" of size "Size". Alignment information is
1113 /// specified by the specific parameter attribute. The copy will be passed as
1114 /// a byval function parameter.
1115 /// Sometimes what we are copying is the end of a larger object, the part that
1116 /// does not fit in registers.
1118 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1119 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1121 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1122 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1123 /*isVolatile=*/false, /*AlwaysInline=*/false,
1124 MachinePointerInfo(0), MachinePointerInfo(0));
1127 /// LowerMemOpCallTo - Store the argument to the stack.
1129 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1130 SDValue StackPtr, SDValue Arg,
1131 DebugLoc dl, SelectionDAG &DAG,
1132 const CCValAssign &VA,
1133 ISD::ArgFlagsTy Flags) const {
1134 unsigned LocMemOffset = VA.getLocMemOffset();
1135 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1136 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1137 if (Flags.isByVal())
1138 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1140 return DAG.getStore(Chain, dl, Arg, PtrOff,
1141 MachinePointerInfo::getStack(LocMemOffset),
1145 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1146 SDValue Chain, SDValue &Arg,
1147 RegsToPassVector &RegsToPass,
1148 CCValAssign &VA, CCValAssign &NextVA,
1150 SmallVector<SDValue, 8> &MemOpChains,
1151 ISD::ArgFlagsTy Flags) const {
1153 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1154 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1155 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1157 if (NextVA.isRegLoc())
1158 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1160 assert(NextVA.isMemLoc());
1161 if (StackPtr.getNode() == 0)
1162 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1164 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1170 /// LowerCall - Lowering a call into a callseq_start <-
1171 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1174 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1175 CallingConv::ID CallConv, bool isVarArg,
1177 const SmallVectorImpl<ISD::OutputArg> &Outs,
1178 const SmallVectorImpl<SDValue> &OutVals,
1179 const SmallVectorImpl<ISD::InputArg> &Ins,
1180 DebugLoc dl, SelectionDAG &DAG,
1181 SmallVectorImpl<SDValue> &InVals) const {
1182 MachineFunction &MF = DAG.getMachineFunction();
1183 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1184 bool IsSibCall = false;
1185 // Temporarily disable tail calls so things don't break.
1186 if (!EnableARMTailCalls)
1189 // Check if it's really possible to do a tail call.
1190 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1191 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1192 Outs, OutVals, Ins, DAG);
1193 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1194 // detected sibcalls.
1201 // Analyze operands of the call, assigning locations to each operand.
1202 SmallVector<CCValAssign, 16> ArgLocs;
1203 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1205 CCInfo.AnalyzeCallOperands(Outs,
1206 CCAssignFnForNode(CallConv, /* Return*/ false,
1209 // Get a count of how many bytes are to be pushed on the stack.
1210 unsigned NumBytes = CCInfo.getNextStackOffset();
1212 // For tail calls, memory operands are available in our caller's stack.
1216 // Adjust the stack pointer for the new arguments...
1217 // These operations are automatically eliminated by the prolog/epilog pass
1219 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1221 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1223 RegsToPassVector RegsToPass;
1224 SmallVector<SDValue, 8> MemOpChains;
1226 // Walk the register/memloc assignments, inserting copies/loads. In the case
1227 // of tail call optimization, arguments are handled later.
1228 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1230 ++i, ++realArgIdx) {
1231 CCValAssign &VA = ArgLocs[i];
1232 SDValue Arg = OutVals[realArgIdx];
1233 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1234 bool isByVal = Flags.isByVal();
1236 // Promote the value if needed.
1237 switch (VA.getLocInfo()) {
1238 default: llvm_unreachable("Unknown loc info!");
1239 case CCValAssign::Full: break;
1240 case CCValAssign::SExt:
1241 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1243 case CCValAssign::ZExt:
1244 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1246 case CCValAssign::AExt:
1247 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1249 case CCValAssign::BCvt:
1250 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1254 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1255 if (VA.needsCustom()) {
1256 if (VA.getLocVT() == MVT::v2f64) {
1257 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1258 DAG.getConstant(0, MVT::i32));
1259 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1260 DAG.getConstant(1, MVT::i32));
1262 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1263 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1265 VA = ArgLocs[++i]; // skip ahead to next loc
1266 if (VA.isRegLoc()) {
1267 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1268 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1270 assert(VA.isMemLoc());
1272 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1273 dl, DAG, VA, Flags));
1276 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1277 StackPtr, MemOpChains, Flags);
1279 } else if (VA.isRegLoc()) {
1280 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1281 } else if (!IsSibCall || isByVal) {
1282 assert(VA.isMemLoc());
1284 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1285 dl, DAG, VA, Flags));
1289 if (!MemOpChains.empty())
1290 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1291 &MemOpChains[0], MemOpChains.size());
1293 // Build a sequence of copy-to-reg nodes chained together with token chain
1294 // and flag operands which copy the outgoing args into the appropriate regs.
1296 // Tail call byval lowering might overwrite argument registers so in case of
1297 // tail call optimization the copies to registers are lowered later.
1299 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1300 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1301 RegsToPass[i].second, InFlag);
1302 InFlag = Chain.getValue(1);
1305 // For tail calls lower the arguments to the 'real' stack slot.
1307 // Force all the incoming stack arguments to be loaded from the stack
1308 // before any new outgoing arguments are stored to the stack, because the
1309 // outgoing stack slots may alias the incoming argument stack slots, and
1310 // the alias isn't otherwise explicit. This is slightly more conservative
1311 // than necessary, because it means that each store effectively depends
1312 // on every argument instead of just those arguments it would clobber.
1314 // Do not flag preceeding copytoreg stuff together with the following stuff.
1316 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1317 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1318 RegsToPass[i].second, InFlag);
1319 InFlag = Chain.getValue(1);
1324 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1325 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1326 // node so that legalize doesn't hack it.
1327 bool isDirect = false;
1328 bool isARMFunc = false;
1329 bool isLocalARMFunc = false;
1330 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1332 if (EnableARMLongCalls) {
1333 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1334 && "long-calls with non-static relocation model!");
1335 // Handle a global address or an external symbol. If it's not one of
1336 // those, the target's already in a register, so we don't need to do
1338 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1339 const GlobalValue *GV = G->getGlobal();
1340 // Create a constant pool entry for the callee address
1341 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1342 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1345 // Get the address of the callee into a register
1346 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1347 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1348 Callee = DAG.getLoad(getPointerTy(), dl,
1349 DAG.getEntryNode(), CPAddr,
1350 MachinePointerInfo::getConstantPool(),
1352 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1353 const char *Sym = S->getSymbol();
1355 // Create a constant pool entry for the callee address
1356 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1357 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1358 Sym, ARMPCLabelIndex, 0);
1359 // Get the address of the callee into a register
1360 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1361 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1362 Callee = DAG.getLoad(getPointerTy(), dl,
1363 DAG.getEntryNode(), CPAddr,
1364 MachinePointerInfo::getConstantPool(),
1367 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1368 const GlobalValue *GV = G->getGlobal();
1370 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1371 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1372 getTargetMachine().getRelocationModel() != Reloc::Static;
1373 isARMFunc = !Subtarget->isThumb() || isStub;
1374 // ARM call to a local ARM function is predicable.
1375 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1376 // tBX takes a register source operand.
1377 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1378 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1379 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1382 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1383 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1384 Callee = DAG.getLoad(getPointerTy(), dl,
1385 DAG.getEntryNode(), CPAddr,
1386 MachinePointerInfo::getConstantPool(),
1388 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1389 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1390 getPointerTy(), Callee, PICLabel);
1392 // On ELF targets for PIC code, direct calls should go through the PLT
1393 unsigned OpFlags = 0;
1394 if (Subtarget->isTargetELF() &&
1395 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1396 OpFlags = ARMII::MO_PLT;
1397 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1399 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1401 bool isStub = Subtarget->isTargetDarwin() &&
1402 getTargetMachine().getRelocationModel() != Reloc::Static;
1403 isARMFunc = !Subtarget->isThumb() || isStub;
1404 // tBX takes a register source operand.
1405 const char *Sym = S->getSymbol();
1406 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1407 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1408 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1409 Sym, ARMPCLabelIndex, 4);
1410 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1411 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1412 Callee = DAG.getLoad(getPointerTy(), dl,
1413 DAG.getEntryNode(), CPAddr,
1414 MachinePointerInfo::getConstantPool(),
1416 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1417 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1418 getPointerTy(), Callee, PICLabel);
1420 unsigned OpFlags = 0;
1421 // On ELF targets for PIC code, direct calls should go through the PLT
1422 if (Subtarget->isTargetELF() &&
1423 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1424 OpFlags = ARMII::MO_PLT;
1425 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1429 // FIXME: handle tail calls differently.
1431 if (Subtarget->isThumb()) {
1432 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1433 CallOpc = ARMISD::CALL_NOLINK;
1435 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1437 CallOpc = (isDirect || Subtarget->hasV5TOps())
1438 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1439 : ARMISD::CALL_NOLINK;
1442 std::vector<SDValue> Ops;
1443 Ops.push_back(Chain);
1444 Ops.push_back(Callee);
1446 // Add argument registers to the end of the list so that they are known live
1448 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1449 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1450 RegsToPass[i].second.getValueType()));
1452 if (InFlag.getNode())
1453 Ops.push_back(InFlag);
1455 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1457 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1459 // Returns a chain and a flag for retval copy to use.
1460 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1461 InFlag = Chain.getValue(1);
1463 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1464 DAG.getIntPtrConstant(0, true), InFlag);
1466 InFlag = Chain.getValue(1);
1468 // Handle result values, copying them out of physregs into vregs that we
1470 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1474 /// HandleByVal - Every parameter *after* a byval parameter is passed
1475 /// on the stack. Confiscate all the parameter registers to insure
1478 llvm::ARMTargetLowering::HandleByVal(CCState *State) const {
1479 static const unsigned RegList1[] = {
1480 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1482 do {} while (State->AllocateReg(RegList1, 4));
1485 /// MatchingStackOffset - Return true if the given stack call argument is
1486 /// already available in the same position (relatively) of the caller's
1487 /// incoming argument stack.
1489 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1490 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1491 const ARMInstrInfo *TII) {
1492 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1494 if (Arg.getOpcode() == ISD::CopyFromReg) {
1495 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1496 if (!TargetRegisterInfo::isVirtualRegister(VR))
1498 MachineInstr *Def = MRI->getVRegDef(VR);
1501 if (!Flags.isByVal()) {
1502 if (!TII->isLoadFromStackSlot(Def, FI))
1507 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1508 if (Flags.isByVal())
1509 // ByVal argument is passed in as a pointer but it's now being
1510 // dereferenced. e.g.
1511 // define @foo(%struct.X* %A) {
1512 // tail call @bar(%struct.X* byval %A)
1515 SDValue Ptr = Ld->getBasePtr();
1516 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1519 FI = FINode->getIndex();
1523 assert(FI != INT_MAX);
1524 if (!MFI->isFixedObjectIndex(FI))
1526 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1529 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1530 /// for tail call optimization. Targets which want to do tail call
1531 /// optimization should implement this function.
1533 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1534 CallingConv::ID CalleeCC,
1536 bool isCalleeStructRet,
1537 bool isCallerStructRet,
1538 const SmallVectorImpl<ISD::OutputArg> &Outs,
1539 const SmallVectorImpl<SDValue> &OutVals,
1540 const SmallVectorImpl<ISD::InputArg> &Ins,
1541 SelectionDAG& DAG) const {
1542 const Function *CallerF = DAG.getMachineFunction().getFunction();
1543 CallingConv::ID CallerCC = CallerF->getCallingConv();
1544 bool CCMatch = CallerCC == CalleeCC;
1546 // Look for obvious safe cases to perform tail call optimization that do not
1547 // require ABI changes. This is what gcc calls sibcall.
1549 // Do not sibcall optimize vararg calls unless the call site is not passing
1551 if (isVarArg && !Outs.empty())
1554 // Also avoid sibcall optimization if either caller or callee uses struct
1555 // return semantics.
1556 if (isCalleeStructRet || isCallerStructRet)
1559 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1560 // emitEpilogue is not ready for them.
1561 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1562 // LR. This means if we need to reload LR, it takes an extra instructions,
1563 // which outweighs the value of the tail call; but here we don't know yet
1564 // whether LR is going to be used. Probably the right approach is to
1565 // generate the tail call here and turn it back into CALL/RET in
1566 // emitEpilogue if LR is used.
1568 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1569 // but we need to make sure there are enough registers; the only valid
1570 // registers are the 4 used for parameters. We don't currently do this
1572 if (Subtarget->isThumb1Only())
1575 // If the calling conventions do not match, then we'd better make sure the
1576 // results are returned in the same way as what the caller expects.
1578 SmallVector<CCValAssign, 16> RVLocs1;
1579 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1580 RVLocs1, *DAG.getContext());
1581 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1583 SmallVector<CCValAssign, 16> RVLocs2;
1584 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1585 RVLocs2, *DAG.getContext());
1586 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1588 if (RVLocs1.size() != RVLocs2.size())
1590 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1591 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1593 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1595 if (RVLocs1[i].isRegLoc()) {
1596 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1599 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1605 // If the callee takes no arguments then go on to check the results of the
1607 if (!Outs.empty()) {
1608 // Check if stack adjustment is needed. For now, do not do this if any
1609 // argument is passed on the stack.
1610 SmallVector<CCValAssign, 16> ArgLocs;
1611 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1612 ArgLocs, *DAG.getContext());
1613 CCInfo.AnalyzeCallOperands(Outs,
1614 CCAssignFnForNode(CalleeCC, false, isVarArg));
1615 if (CCInfo.getNextStackOffset()) {
1616 MachineFunction &MF = DAG.getMachineFunction();
1618 // Check if the arguments are already laid out in the right way as
1619 // the caller's fixed stack objects.
1620 MachineFrameInfo *MFI = MF.getFrameInfo();
1621 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1622 const ARMInstrInfo *TII =
1623 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1624 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1626 ++i, ++realArgIdx) {
1627 CCValAssign &VA = ArgLocs[i];
1628 EVT RegVT = VA.getLocVT();
1629 SDValue Arg = OutVals[realArgIdx];
1630 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1631 if (VA.getLocInfo() == CCValAssign::Indirect)
1633 if (VA.needsCustom()) {
1634 // f64 and vector types are split into multiple registers or
1635 // register/stack-slot combinations. The types will not match
1636 // the registers; give up on memory f64 refs until we figure
1637 // out what to do about this.
1640 if (!ArgLocs[++i].isRegLoc())
1642 if (RegVT == MVT::v2f64) {
1643 if (!ArgLocs[++i].isRegLoc())
1645 if (!ArgLocs[++i].isRegLoc())
1648 } else if (!VA.isRegLoc()) {
1649 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1661 ARMTargetLowering::LowerReturn(SDValue Chain,
1662 CallingConv::ID CallConv, bool isVarArg,
1663 const SmallVectorImpl<ISD::OutputArg> &Outs,
1664 const SmallVectorImpl<SDValue> &OutVals,
1665 DebugLoc dl, SelectionDAG &DAG) const {
1667 // CCValAssign - represent the assignment of the return value to a location.
1668 SmallVector<CCValAssign, 16> RVLocs;
1670 // CCState - Info about the registers and stack slots.
1671 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1674 // Analyze outgoing return values.
1675 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1678 // If this is the first return lowered for this function, add
1679 // the regs to the liveout set for the function.
1680 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1681 for (unsigned i = 0; i != RVLocs.size(); ++i)
1682 if (RVLocs[i].isRegLoc())
1683 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1688 // Copy the result values into the output registers.
1689 for (unsigned i = 0, realRVLocIdx = 0;
1691 ++i, ++realRVLocIdx) {
1692 CCValAssign &VA = RVLocs[i];
1693 assert(VA.isRegLoc() && "Can only return in registers!");
1695 SDValue Arg = OutVals[realRVLocIdx];
1697 switch (VA.getLocInfo()) {
1698 default: llvm_unreachable("Unknown loc info!");
1699 case CCValAssign::Full: break;
1700 case CCValAssign::BCvt:
1701 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1705 if (VA.needsCustom()) {
1706 if (VA.getLocVT() == MVT::v2f64) {
1707 // Extract the first half and return it in two registers.
1708 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1709 DAG.getConstant(0, MVT::i32));
1710 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1711 DAG.getVTList(MVT::i32, MVT::i32), Half);
1713 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1714 Flag = Chain.getValue(1);
1715 VA = RVLocs[++i]; // skip ahead to next loc
1716 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1717 HalfGPRs.getValue(1), Flag);
1718 Flag = Chain.getValue(1);
1719 VA = RVLocs[++i]; // skip ahead to next loc
1721 // Extract the 2nd half and fall through to handle it as an f64 value.
1722 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1723 DAG.getConstant(1, MVT::i32));
1725 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1727 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1728 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1729 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1730 Flag = Chain.getValue(1);
1731 VA = RVLocs[++i]; // skip ahead to next loc
1732 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1735 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1737 // Guarantee that all emitted copies are
1738 // stuck together, avoiding something bad.
1739 Flag = Chain.getValue(1);
1744 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1746 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1751 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1752 if (N->getNumValues() != 1)
1754 if (!N->hasNUsesOfValue(1, 0))
1757 unsigned NumCopies = 0;
1759 SDNode *Use = *N->use_begin();
1760 if (Use->getOpcode() == ISD::CopyToReg) {
1761 Copies[NumCopies++] = Use;
1762 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1763 // f64 returned in a pair of GPRs.
1764 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1766 if (UI->getOpcode() != ISD::CopyToReg)
1768 Copies[UI.getUse().getResNo()] = *UI;
1771 } else if (Use->getOpcode() == ISD::BITCAST) {
1772 // f32 returned in a single GPR.
1773 if (!Use->hasNUsesOfValue(1, 0))
1775 Use = *Use->use_begin();
1776 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1778 Copies[NumCopies++] = Use;
1783 if (NumCopies != 1 && NumCopies != 2)
1786 bool HasRet = false;
1787 for (unsigned i = 0; i < NumCopies; ++i) {
1788 SDNode *Copy = Copies[i];
1789 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1791 if (UI->getOpcode() == ISD::CopyToReg) {
1793 if (Use == Copies[0] || Use == Copies[1])
1797 if (UI->getOpcode() != ARMISD::RET_FLAG)
1806 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1807 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1808 // one of the above mentioned nodes. It has to be wrapped because otherwise
1809 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1810 // be used to form addressing mode. These wrapped nodes will be selected
1812 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1813 EVT PtrVT = Op.getValueType();
1814 // FIXME there is no actual debug info here
1815 DebugLoc dl = Op.getDebugLoc();
1816 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1818 if (CP->isMachineConstantPoolEntry())
1819 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1820 CP->getAlignment());
1822 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1823 CP->getAlignment());
1824 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1827 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1828 return MachineJumpTableInfo::EK_Inline;
1831 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1832 SelectionDAG &DAG) const {
1833 MachineFunction &MF = DAG.getMachineFunction();
1834 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1835 unsigned ARMPCLabelIndex = 0;
1836 DebugLoc DL = Op.getDebugLoc();
1837 EVT PtrVT = getPointerTy();
1838 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1839 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1841 if (RelocM == Reloc::Static) {
1842 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1844 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1845 ARMPCLabelIndex = AFI->createPICLabelUId();
1846 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1847 ARMCP::CPBlockAddress,
1849 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1851 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1852 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1853 MachinePointerInfo::getConstantPool(),
1855 if (RelocM == Reloc::Static)
1857 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1858 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1861 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1863 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1864 SelectionDAG &DAG) const {
1865 DebugLoc dl = GA->getDebugLoc();
1866 EVT PtrVT = getPointerTy();
1867 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1868 MachineFunction &MF = DAG.getMachineFunction();
1869 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1870 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1871 ARMConstantPoolValue *CPV =
1872 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1873 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1874 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1875 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1876 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1877 MachinePointerInfo::getConstantPool(),
1879 SDValue Chain = Argument.getValue(1);
1881 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1882 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1884 // call __tls_get_addr.
1887 Entry.Node = Argument;
1888 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1889 Args.push_back(Entry);
1890 // FIXME: is there useful debug info available here?
1891 std::pair<SDValue, SDValue> CallResult =
1892 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1893 false, false, false, false,
1894 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1895 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1896 return CallResult.first;
1899 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1900 // "local exec" model.
1902 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1903 SelectionDAG &DAG) const {
1904 const GlobalValue *GV = GA->getGlobal();
1905 DebugLoc dl = GA->getDebugLoc();
1907 SDValue Chain = DAG.getEntryNode();
1908 EVT PtrVT = getPointerTy();
1909 // Get the Thread Pointer
1910 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1912 if (GV->isDeclaration()) {
1913 MachineFunction &MF = DAG.getMachineFunction();
1914 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1915 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1916 // Initial exec model.
1917 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1918 ARMConstantPoolValue *CPV =
1919 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1920 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1921 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1922 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1923 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1924 MachinePointerInfo::getConstantPool(),
1926 Chain = Offset.getValue(1);
1928 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1929 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1931 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1932 MachinePointerInfo::getConstantPool(),
1936 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
1937 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1938 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1939 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1940 MachinePointerInfo::getConstantPool(),
1944 // The address of the thread local variable is the add of the thread
1945 // pointer with the offset of the variable.
1946 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1950 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1951 // TODO: implement the "local dynamic" model
1952 assert(Subtarget->isTargetELF() &&
1953 "TLS not implemented for non-ELF targets");
1954 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1955 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1956 // otherwise use the "Local Exec" TLS Model
1957 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1958 return LowerToTLSGeneralDynamicModel(GA, DAG);
1960 return LowerToTLSExecModels(GA, DAG);
1963 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1964 SelectionDAG &DAG) const {
1965 EVT PtrVT = getPointerTy();
1966 DebugLoc dl = Op.getDebugLoc();
1967 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1968 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1969 if (RelocM == Reloc::PIC_) {
1970 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1971 ARMConstantPoolValue *CPV =
1972 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
1973 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1974 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1975 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1977 MachinePointerInfo::getConstantPool(),
1979 SDValue Chain = Result.getValue(1);
1980 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1981 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1983 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1984 MachinePointerInfo::getGOT(), false, false, 0);
1988 // If we have T2 ops, we can materialize the address directly via movt/movw
1989 // pair. This is always cheaper.
1990 if (Subtarget->useMovt()) {
1992 // FIXME: Once remat is capable of dealing with instructions with register
1993 // operands, expand this into two nodes.
1994 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1995 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1997 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1998 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1999 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2000 MachinePointerInfo::getConstantPool(),
2005 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2006 SelectionDAG &DAG) const {
2007 EVT PtrVT = getPointerTy();
2008 DebugLoc dl = Op.getDebugLoc();
2009 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2010 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2011 MachineFunction &MF = DAG.getMachineFunction();
2012 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2014 if (Subtarget->useMovt()) {
2016 // FIXME: Once remat is capable of dealing with instructions with register
2017 // operands, expand this into two nodes.
2018 if (RelocM == Reloc::Static)
2019 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2020 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2022 unsigned Wrapper = (RelocM == Reloc::PIC_)
2023 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2024 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2025 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2026 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2027 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2028 MachinePointerInfo::getGOT(), false, false, 0);
2032 unsigned ARMPCLabelIndex = 0;
2034 if (RelocM == Reloc::Static) {
2035 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2037 ARMPCLabelIndex = AFI->createPICLabelUId();
2038 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2039 ARMConstantPoolValue *CPV =
2040 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2041 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2043 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2045 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2046 MachinePointerInfo::getConstantPool(),
2048 SDValue Chain = Result.getValue(1);
2050 if (RelocM == Reloc::PIC_) {
2051 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2052 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2055 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2056 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2062 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2063 SelectionDAG &DAG) const {
2064 assert(Subtarget->isTargetELF() &&
2065 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2066 MachineFunction &MF = DAG.getMachineFunction();
2067 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2068 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2069 EVT PtrVT = getPointerTy();
2070 DebugLoc dl = Op.getDebugLoc();
2071 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2072 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2073 "_GLOBAL_OFFSET_TABLE_",
2074 ARMPCLabelIndex, PCAdj);
2075 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2076 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2077 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2078 MachinePointerInfo::getConstantPool(),
2080 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2081 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2085 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2087 DebugLoc dl = Op.getDebugLoc();
2088 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2089 Op.getOperand(0), Op.getOperand(1));
2093 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2094 DebugLoc dl = Op.getDebugLoc();
2095 SDValue Val = DAG.getConstant(0, MVT::i32);
2096 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2097 Op.getOperand(1), Val);
2101 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2102 DebugLoc dl = Op.getDebugLoc();
2103 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2104 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2108 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2109 const ARMSubtarget *Subtarget) const {
2110 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2111 DebugLoc dl = Op.getDebugLoc();
2113 default: return SDValue(); // Don't custom lower most intrinsics.
2114 case Intrinsic::arm_thread_pointer: {
2115 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2116 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2118 case Intrinsic::eh_sjlj_lsda: {
2119 MachineFunction &MF = DAG.getMachineFunction();
2120 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2121 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2122 EVT PtrVT = getPointerTy();
2123 DebugLoc dl = Op.getDebugLoc();
2124 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2126 unsigned PCAdj = (RelocM != Reloc::PIC_)
2127 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2128 ARMConstantPoolValue *CPV =
2129 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2130 ARMCP::CPLSDA, PCAdj);
2131 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2132 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2134 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2135 MachinePointerInfo::getConstantPool(),
2138 if (RelocM == Reloc::PIC_) {
2139 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2140 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2147 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2148 const ARMSubtarget *Subtarget) {
2149 DebugLoc dl = Op.getDebugLoc();
2150 if (!Subtarget->hasDataBarrier()) {
2151 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2152 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2154 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2155 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2156 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2157 DAG.getConstant(0, MVT::i32));
2160 SDValue Op5 = Op.getOperand(5);
2161 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2162 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2163 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2164 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2166 ARM_MB::MemBOpt DMBOpt;
2167 if (isDeviceBarrier)
2168 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2170 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2171 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2172 DAG.getConstant(DMBOpt, MVT::i32));
2175 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2176 const ARMSubtarget *Subtarget) {
2177 // ARM pre v5TE and Thumb1 does not have preload instructions.
2178 if (!(Subtarget->isThumb2() ||
2179 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2180 // Just preserve the chain.
2181 return Op.getOperand(0);
2183 DebugLoc dl = Op.getDebugLoc();
2184 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2186 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2187 // ARMv7 with MP extension has PLDW.
2188 return Op.getOperand(0);
2190 if (Subtarget->isThumb())
2192 isRead = ~isRead & 1;
2193 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2195 // Currently there is no intrinsic that matches pli.
2196 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2197 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2198 DAG.getConstant(isData, MVT::i32));
2201 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2202 MachineFunction &MF = DAG.getMachineFunction();
2203 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2205 // vastart just stores the address of the VarArgsFrameIndex slot into the
2206 // memory location argument.
2207 DebugLoc dl = Op.getDebugLoc();
2208 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2209 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2210 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2211 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2212 MachinePointerInfo(SV), false, false, 0);
2216 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2217 SDValue &Root, SelectionDAG &DAG,
2218 DebugLoc dl) const {
2219 MachineFunction &MF = DAG.getMachineFunction();
2220 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2222 TargetRegisterClass *RC;
2223 if (AFI->isThumb1OnlyFunction())
2224 RC = ARM::tGPRRegisterClass;
2226 RC = ARM::GPRRegisterClass;
2228 // Transform the arguments stored in physical registers into virtual ones.
2229 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2230 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2233 if (NextVA.isMemLoc()) {
2234 MachineFrameInfo *MFI = MF.getFrameInfo();
2235 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2237 // Create load node to retrieve arguments from the stack.
2238 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2239 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2240 MachinePointerInfo::getFixedStack(FI),
2243 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2244 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2247 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2251 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2252 CallingConv::ID CallConv, bool isVarArg,
2253 const SmallVectorImpl<ISD::InputArg>
2255 DebugLoc dl, SelectionDAG &DAG,
2256 SmallVectorImpl<SDValue> &InVals)
2259 MachineFunction &MF = DAG.getMachineFunction();
2260 MachineFrameInfo *MFI = MF.getFrameInfo();
2262 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2264 // Assign locations to all of the incoming arguments.
2265 SmallVector<CCValAssign, 16> ArgLocs;
2266 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2268 CCInfo.AnalyzeFormalArguments(Ins,
2269 CCAssignFnForNode(CallConv, /* Return*/ false,
2272 SmallVector<SDValue, 16> ArgValues;
2273 int lastInsIndex = -1;
2276 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2277 CCValAssign &VA = ArgLocs[i];
2279 // Arguments stored in registers.
2280 if (VA.isRegLoc()) {
2281 EVT RegVT = VA.getLocVT();
2283 if (VA.needsCustom()) {
2284 // f64 and vector types are split up into multiple registers or
2285 // combinations of registers and stack slots.
2286 if (VA.getLocVT() == MVT::v2f64) {
2287 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2289 VA = ArgLocs[++i]; // skip ahead to next loc
2291 if (VA.isMemLoc()) {
2292 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2293 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2294 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2295 MachinePointerInfo::getFixedStack(FI),
2298 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2301 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2302 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2303 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2304 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2305 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2307 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2310 TargetRegisterClass *RC;
2312 if (RegVT == MVT::f32)
2313 RC = ARM::SPRRegisterClass;
2314 else if (RegVT == MVT::f64)
2315 RC = ARM::DPRRegisterClass;
2316 else if (RegVT == MVT::v2f64)
2317 RC = ARM::QPRRegisterClass;
2318 else if (RegVT == MVT::i32)
2319 RC = (AFI->isThumb1OnlyFunction() ?
2320 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2322 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2324 // Transform the arguments in physical registers into virtual ones.
2325 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2326 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2329 // If this is an 8 or 16-bit value, it is really passed promoted
2330 // to 32 bits. Insert an assert[sz]ext to capture this, then
2331 // truncate to the right size.
2332 switch (VA.getLocInfo()) {
2333 default: llvm_unreachable("Unknown loc info!");
2334 case CCValAssign::Full: break;
2335 case CCValAssign::BCvt:
2336 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2338 case CCValAssign::SExt:
2339 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2340 DAG.getValueType(VA.getValVT()));
2341 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2343 case CCValAssign::ZExt:
2344 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2345 DAG.getValueType(VA.getValVT()));
2346 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2350 InVals.push_back(ArgValue);
2352 } else { // VA.isRegLoc()
2355 assert(VA.isMemLoc());
2356 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2358 int index = ArgLocs[i].getValNo();
2360 // Some Ins[] entries become multiple ArgLoc[] entries.
2361 // Process them only once.
2362 if (index != lastInsIndex)
2364 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2365 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2366 // changed with more analysis.
2367 // In case of tail call optimization mark all arguments mutable. Since they
2368 // could be overwritten by lowering of arguments in case of a tail call.
2369 if (Flags.isByVal()) {
2370 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2371 VA.getLocMemOffset(), false);
2372 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2374 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2375 VA.getLocMemOffset(), true);
2377 // Create load nodes to retrieve arguments from the stack.
2378 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2379 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2380 MachinePointerInfo::getFixedStack(FI),
2383 lastInsIndex = index;
2390 static const unsigned GPRArgRegs[] = {
2391 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2394 unsigned NumGPRs = CCInfo.getFirstUnallocated
2395 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2397 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2398 unsigned VARegSize = (4 - NumGPRs) * 4;
2399 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2400 unsigned ArgOffset = CCInfo.getNextStackOffset();
2401 if (VARegSaveSize) {
2402 // If this function is vararg, store any remaining integer argument regs
2403 // to their spots on the stack so that they may be loaded by deferencing
2404 // the result of va_next.
2405 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2406 AFI->setVarArgsFrameIndex(
2407 MFI->CreateFixedObject(VARegSaveSize,
2408 ArgOffset + VARegSaveSize - VARegSize,
2410 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2413 SmallVector<SDValue, 4> MemOps;
2414 for (; NumGPRs < 4; ++NumGPRs) {
2415 TargetRegisterClass *RC;
2416 if (AFI->isThumb1OnlyFunction())
2417 RC = ARM::tGPRRegisterClass;
2419 RC = ARM::GPRRegisterClass;
2421 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2422 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2424 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2425 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2427 MemOps.push_back(Store);
2428 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2429 DAG.getConstant(4, getPointerTy()));
2431 if (!MemOps.empty())
2432 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2433 &MemOps[0], MemOps.size());
2435 // This will point to the next argument passed via stack.
2436 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2442 /// isFloatingPointZero - Return true if this is +0.0.
2443 static bool isFloatingPointZero(SDValue Op) {
2444 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2445 return CFP->getValueAPF().isPosZero();
2446 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2447 // Maybe this has already been legalized into the constant pool?
2448 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2449 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2450 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2451 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2452 return CFP->getValueAPF().isPosZero();
2458 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2459 /// the given operands.
2461 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2462 SDValue &ARMcc, SelectionDAG &DAG,
2463 DebugLoc dl) const {
2464 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2465 unsigned C = RHSC->getZExtValue();
2466 if (!isLegalICmpImmediate(C)) {
2467 // Constant does not fit, try adjusting it by one?
2472 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2473 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2474 RHS = DAG.getConstant(C-1, MVT::i32);
2479 if (C != 0 && isLegalICmpImmediate(C-1)) {
2480 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2481 RHS = DAG.getConstant(C-1, MVT::i32);
2486 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2487 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2488 RHS = DAG.getConstant(C+1, MVT::i32);
2493 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2494 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2495 RHS = DAG.getConstant(C+1, MVT::i32);
2502 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2503 ARMISD::NodeType CompareType;
2506 CompareType = ARMISD::CMP;
2511 CompareType = ARMISD::CMPZ;
2514 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2515 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2518 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2520 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2521 DebugLoc dl) const {
2523 if (!isFloatingPointZero(RHS))
2524 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2526 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2527 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2530 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2531 SDValue Cond = Op.getOperand(0);
2532 SDValue SelectTrue = Op.getOperand(1);
2533 SDValue SelectFalse = Op.getOperand(2);
2534 DebugLoc dl = Op.getDebugLoc();
2538 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2539 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2541 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2542 const ConstantSDNode *CMOVTrue =
2543 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2544 const ConstantSDNode *CMOVFalse =
2545 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2547 if (CMOVTrue && CMOVFalse) {
2548 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2549 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2553 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2555 False = SelectFalse;
2556 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2561 if (True.getNode() && False.getNode()) {
2562 EVT VT = Cond.getValueType();
2563 SDValue ARMcc = Cond.getOperand(2);
2564 SDValue CCR = Cond.getOperand(3);
2565 SDValue Cmp = Cond.getOperand(4);
2566 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2571 return DAG.getSelectCC(dl, Cond,
2572 DAG.getConstant(0, Cond.getValueType()),
2573 SelectTrue, SelectFalse, ISD::SETNE);
2576 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2577 EVT VT = Op.getValueType();
2578 SDValue LHS = Op.getOperand(0);
2579 SDValue RHS = Op.getOperand(1);
2580 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2581 SDValue TrueVal = Op.getOperand(2);
2582 SDValue FalseVal = Op.getOperand(3);
2583 DebugLoc dl = Op.getDebugLoc();
2585 if (LHS.getValueType() == MVT::i32) {
2587 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2588 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2589 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2592 ARMCC::CondCodes CondCode, CondCode2;
2593 FPCCToARMCC(CC, CondCode, CondCode2);
2595 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2596 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2597 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2598 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2600 if (CondCode2 != ARMCC::AL) {
2601 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2602 // FIXME: Needs another CMP because flag can have but one use.
2603 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2604 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2605 Result, TrueVal, ARMcc2, CCR, Cmp2);
2610 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2611 /// to morph to an integer compare sequence.
2612 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2613 const ARMSubtarget *Subtarget) {
2614 SDNode *N = Op.getNode();
2615 if (!N->hasOneUse())
2616 // Otherwise it requires moving the value from fp to integer registers.
2618 if (!N->getNumValues())
2620 EVT VT = Op.getValueType();
2621 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2622 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2623 // vmrs are very slow, e.g. cortex-a8.
2626 if (isFloatingPointZero(Op)) {
2630 return ISD::isNormalLoad(N);
2633 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2634 if (isFloatingPointZero(Op))
2635 return DAG.getConstant(0, MVT::i32);
2637 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2638 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2639 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2640 Ld->isVolatile(), Ld->isNonTemporal(),
2641 Ld->getAlignment());
2643 llvm_unreachable("Unknown VFP cmp argument!");
2646 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2647 SDValue &RetVal1, SDValue &RetVal2) {
2648 if (isFloatingPointZero(Op)) {
2649 RetVal1 = DAG.getConstant(0, MVT::i32);
2650 RetVal2 = DAG.getConstant(0, MVT::i32);
2654 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2655 SDValue Ptr = Ld->getBasePtr();
2656 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2657 Ld->getChain(), Ptr,
2658 Ld->getPointerInfo(),
2659 Ld->isVolatile(), Ld->isNonTemporal(),
2660 Ld->getAlignment());
2662 EVT PtrType = Ptr.getValueType();
2663 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2664 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2665 PtrType, Ptr, DAG.getConstant(4, PtrType));
2666 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2667 Ld->getChain(), NewPtr,
2668 Ld->getPointerInfo().getWithOffset(4),
2669 Ld->isVolatile(), Ld->isNonTemporal(),
2674 llvm_unreachable("Unknown VFP cmp argument!");
2677 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2678 /// f32 and even f64 comparisons to integer ones.
2680 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2681 SDValue Chain = Op.getOperand(0);
2682 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2683 SDValue LHS = Op.getOperand(2);
2684 SDValue RHS = Op.getOperand(3);
2685 SDValue Dest = Op.getOperand(4);
2686 DebugLoc dl = Op.getDebugLoc();
2688 bool SeenZero = false;
2689 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2690 canChangeToInt(RHS, SeenZero, Subtarget) &&
2691 // If one of the operand is zero, it's safe to ignore the NaN case since
2692 // we only care about equality comparisons.
2693 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2694 // If unsafe fp math optimization is enabled and there are no other uses of
2695 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2696 // to an integer comparison.
2697 if (CC == ISD::SETOEQ)
2699 else if (CC == ISD::SETUNE)
2703 if (LHS.getValueType() == MVT::f32) {
2704 LHS = bitcastf32Toi32(LHS, DAG);
2705 RHS = bitcastf32Toi32(RHS, DAG);
2706 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2707 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2708 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2709 Chain, Dest, ARMcc, CCR, Cmp);
2714 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2715 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2716 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2717 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2718 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2719 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2720 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2726 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2727 SDValue Chain = Op.getOperand(0);
2728 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2729 SDValue LHS = Op.getOperand(2);
2730 SDValue RHS = Op.getOperand(3);
2731 SDValue Dest = Op.getOperand(4);
2732 DebugLoc dl = Op.getDebugLoc();
2734 if (LHS.getValueType() == MVT::i32) {
2736 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2737 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2738 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2739 Chain, Dest, ARMcc, CCR, Cmp);
2742 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2745 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2746 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2747 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2748 if (Result.getNode())
2752 ARMCC::CondCodes CondCode, CondCode2;
2753 FPCCToARMCC(CC, CondCode, CondCode2);
2755 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2756 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2757 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2758 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2759 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2760 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2761 if (CondCode2 != ARMCC::AL) {
2762 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2763 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2764 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2769 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2770 SDValue Chain = Op.getOperand(0);
2771 SDValue Table = Op.getOperand(1);
2772 SDValue Index = Op.getOperand(2);
2773 DebugLoc dl = Op.getDebugLoc();
2775 EVT PTy = getPointerTy();
2776 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2777 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2778 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2779 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2780 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2781 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2782 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2783 if (Subtarget->isThumb2()) {
2784 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2785 // which does another jump to the destination. This also makes it easier
2786 // to translate it to TBB / TBH later.
2787 // FIXME: This might not work if the function is extremely large.
2788 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2789 Addr, Op.getOperand(2), JTI, UId);
2791 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2792 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2793 MachinePointerInfo::getJumpTable(),
2795 Chain = Addr.getValue(1);
2796 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2797 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2799 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2800 MachinePointerInfo::getJumpTable(), false, false, 0);
2801 Chain = Addr.getValue(1);
2802 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2806 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2807 DebugLoc dl = Op.getDebugLoc();
2810 switch (Op.getOpcode()) {
2812 assert(0 && "Invalid opcode!");
2813 case ISD::FP_TO_SINT:
2814 Opc = ARMISD::FTOSI;
2816 case ISD::FP_TO_UINT:
2817 Opc = ARMISD::FTOUI;
2820 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2821 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2824 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2825 EVT VT = Op.getValueType();
2826 DebugLoc dl = Op.getDebugLoc();
2829 switch (Op.getOpcode()) {
2831 assert(0 && "Invalid opcode!");
2832 case ISD::SINT_TO_FP:
2833 Opc = ARMISD::SITOF;
2835 case ISD::UINT_TO_FP:
2836 Opc = ARMISD::UITOF;
2840 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
2841 return DAG.getNode(Opc, dl, VT, Op);
2844 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2845 // Implement fcopysign with a fabs and a conditional fneg.
2846 SDValue Tmp0 = Op.getOperand(0);
2847 SDValue Tmp1 = Op.getOperand(1);
2848 DebugLoc dl = Op.getDebugLoc();
2849 EVT VT = Op.getValueType();
2850 EVT SrcVT = Tmp1.getValueType();
2851 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
2852 Tmp0.getOpcode() == ARMISD::VMOVDRR;
2853 bool UseNEON = !InGPR && Subtarget->hasNEON();
2856 // Use VBSL to copy the sign bit.
2857 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
2858 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
2859 DAG.getTargetConstant(EncodedVal, MVT::i32));
2860 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
2862 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
2863 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
2864 DAG.getConstant(32, MVT::i32));
2865 else /*if (VT == MVT::f32)*/
2866 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
2867 if (SrcVT == MVT::f32) {
2868 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
2870 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
2871 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
2872 DAG.getConstant(32, MVT::i32));
2874 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
2875 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
2877 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
2879 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
2880 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
2881 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
2883 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
2884 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
2885 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
2886 if (VT == MVT::f32) {
2887 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
2888 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
2889 DAG.getConstant(0, MVT::i32));
2891 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
2897 // Bitcast operand 1 to i32.
2898 if (SrcVT == MVT::f64)
2899 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
2900 &Tmp1, 1).getValue(1);
2901 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
2903 // Or in the signbit with integer operations.
2904 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
2905 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
2906 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
2907 if (VT == MVT::f32) {
2908 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
2909 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
2910 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
2911 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
2914 // f64: Or the high part with signbit and then combine two parts.
2915 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
2917 SDValue Lo = Tmp0.getValue(0);
2918 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
2919 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
2920 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2923 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2924 MachineFunction &MF = DAG.getMachineFunction();
2925 MachineFrameInfo *MFI = MF.getFrameInfo();
2926 MFI->setReturnAddressIsTaken(true);
2928 EVT VT = Op.getValueType();
2929 DebugLoc dl = Op.getDebugLoc();
2930 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2932 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2933 SDValue Offset = DAG.getConstant(4, MVT::i32);
2934 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2935 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2936 MachinePointerInfo(), false, false, 0);
2939 // Return LR, which contains the return address. Mark it an implicit live-in.
2940 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2941 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2944 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2945 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2946 MFI->setFrameAddressIsTaken(true);
2948 EVT VT = Op.getValueType();
2949 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2950 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2951 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2952 ? ARM::R7 : ARM::R11;
2953 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2955 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2956 MachinePointerInfo(),
2961 /// ExpandBITCAST - If the target supports VFP, this function is called to
2962 /// expand a bit convert where either the source or destination type is i64 to
2963 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2964 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2965 /// vectors), since the legalizer won't know what to do with that.
2966 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
2967 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2968 DebugLoc dl = N->getDebugLoc();
2969 SDValue Op = N->getOperand(0);
2971 // This function is only supposed to be called for i64 types, either as the
2972 // source or destination of the bit convert.
2973 EVT SrcVT = Op.getValueType();
2974 EVT DstVT = N->getValueType(0);
2975 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2976 "ExpandBITCAST called for non-i64 type");
2978 // Turn i64->f64 into VMOVDRR.
2979 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2980 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2981 DAG.getConstant(0, MVT::i32));
2982 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2983 DAG.getConstant(1, MVT::i32));
2984 return DAG.getNode(ISD::BITCAST, dl, DstVT,
2985 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2988 // Turn f64->i64 into VMOVRRD.
2989 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2990 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2991 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2992 // Merge the pieces into a single i64 value.
2993 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2999 /// getZeroVector - Returns a vector of specified type with all zero elements.
3000 /// Zero vectors are used to represent vector negation and in those cases
3001 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3002 /// not support i64 elements, so sometimes the zero vectors will need to be
3003 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3005 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3006 assert(VT.isVector() && "Expected a vector type");
3007 // The canonical modified immediate encoding of a zero vector is....0!
3008 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3009 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3010 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3011 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3014 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3015 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3016 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3017 SelectionDAG &DAG) const {
3018 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3019 EVT VT = Op.getValueType();
3020 unsigned VTBits = VT.getSizeInBits();
3021 DebugLoc dl = Op.getDebugLoc();
3022 SDValue ShOpLo = Op.getOperand(0);
3023 SDValue ShOpHi = Op.getOperand(1);
3024 SDValue ShAmt = Op.getOperand(2);
3026 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3028 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3030 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3031 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3032 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3033 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3034 DAG.getConstant(VTBits, MVT::i32));
3035 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3036 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3037 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3039 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3040 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3042 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3043 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3046 SDValue Ops[2] = { Lo, Hi };
3047 return DAG.getMergeValues(Ops, 2, dl);
3050 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3051 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3052 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3053 SelectionDAG &DAG) const {
3054 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3055 EVT VT = Op.getValueType();
3056 unsigned VTBits = VT.getSizeInBits();
3057 DebugLoc dl = Op.getDebugLoc();
3058 SDValue ShOpLo = Op.getOperand(0);
3059 SDValue ShOpHi = Op.getOperand(1);
3060 SDValue ShAmt = Op.getOperand(2);
3063 assert(Op.getOpcode() == ISD::SHL_PARTS);
3064 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3065 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3066 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3067 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3068 DAG.getConstant(VTBits, MVT::i32));
3069 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3070 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3072 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3073 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3074 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3076 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3077 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3080 SDValue Ops[2] = { Lo, Hi };
3081 return DAG.getMergeValues(Ops, 2, dl);
3084 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3085 SelectionDAG &DAG) const {
3086 // The rounding mode is in bits 23:22 of the FPSCR.
3087 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3088 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3089 // so that the shift + and get folded into a bitfield extract.
3090 DebugLoc dl = Op.getDebugLoc();
3091 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3092 DAG.getConstant(Intrinsic::arm_get_fpscr,
3094 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3095 DAG.getConstant(1U << 22, MVT::i32));
3096 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3097 DAG.getConstant(22, MVT::i32));
3098 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3099 DAG.getConstant(3, MVT::i32));
3102 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3103 const ARMSubtarget *ST) {
3104 EVT VT = N->getValueType(0);
3105 DebugLoc dl = N->getDebugLoc();
3107 if (!ST->hasV6T2Ops())
3110 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3111 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3114 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3115 const ARMSubtarget *ST) {
3116 EVT VT = N->getValueType(0);
3117 DebugLoc dl = N->getDebugLoc();
3122 // Lower vector shifts on NEON to use VSHL.
3123 assert(ST->hasNEON() && "unexpected vector shift");
3125 // Left shifts translate directly to the vshiftu intrinsic.
3126 if (N->getOpcode() == ISD::SHL)
3127 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3128 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3129 N->getOperand(0), N->getOperand(1));
3131 assert((N->getOpcode() == ISD::SRA ||
3132 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3134 // NEON uses the same intrinsics for both left and right shifts. For
3135 // right shifts, the shift amounts are negative, so negate the vector of
3137 EVT ShiftVT = N->getOperand(1).getValueType();
3138 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3139 getZeroVector(ShiftVT, DAG, dl),
3141 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3142 Intrinsic::arm_neon_vshifts :
3143 Intrinsic::arm_neon_vshiftu);
3144 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3145 DAG.getConstant(vshiftInt, MVT::i32),
3146 N->getOperand(0), NegatedCount);
3149 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3150 const ARMSubtarget *ST) {
3151 EVT VT = N->getValueType(0);
3152 DebugLoc dl = N->getDebugLoc();
3154 // We can get here for a node like i32 = ISD::SHL i32, i64
3158 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3159 "Unknown shift to lower!");
3161 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3162 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3163 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3166 // If we are in thumb mode, we don't have RRX.
3167 if (ST->isThumb1Only()) return SDValue();
3169 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3170 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3171 DAG.getConstant(0, MVT::i32));
3172 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3173 DAG.getConstant(1, MVT::i32));
3175 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3176 // captures the result into a carry flag.
3177 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3178 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3180 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3181 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3183 // Merge the pieces into a single i64 value.
3184 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3187 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3188 SDValue TmpOp0, TmpOp1;
3189 bool Invert = false;
3193 SDValue Op0 = Op.getOperand(0);
3194 SDValue Op1 = Op.getOperand(1);
3195 SDValue CC = Op.getOperand(2);
3196 EVT VT = Op.getValueType();
3197 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3198 DebugLoc dl = Op.getDebugLoc();
3200 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3201 switch (SetCCOpcode) {
3202 default: llvm_unreachable("Illegal FP comparison"); break;
3204 case ISD::SETNE: Invert = true; // Fallthrough
3206 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3208 case ISD::SETLT: Swap = true; // Fallthrough
3210 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3212 case ISD::SETLE: Swap = true; // Fallthrough
3214 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3215 case ISD::SETUGE: Swap = true; // Fallthrough
3216 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3217 case ISD::SETUGT: Swap = true; // Fallthrough
3218 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3219 case ISD::SETUEQ: Invert = true; // Fallthrough
3221 // Expand this to (OLT | OGT).
3225 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3226 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3228 case ISD::SETUO: Invert = true; // Fallthrough
3230 // Expand this to (OLT | OGE).
3234 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3235 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3239 // Integer comparisons.
3240 switch (SetCCOpcode) {
3241 default: llvm_unreachable("Illegal integer comparison"); break;
3242 case ISD::SETNE: Invert = true;
3243 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3244 case ISD::SETLT: Swap = true;
3245 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3246 case ISD::SETLE: Swap = true;
3247 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3248 case ISD::SETULT: Swap = true;
3249 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3250 case ISD::SETULE: Swap = true;
3251 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3254 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3255 if (Opc == ARMISD::VCEQ) {
3258 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3260 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3263 // Ignore bitconvert.
3264 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3265 AndOp = AndOp.getOperand(0);
3267 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3269 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3270 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3277 std::swap(Op0, Op1);
3279 // If one of the operands is a constant vector zero, attempt to fold the
3280 // comparison to a specialized compare-against-zero form.
3282 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3284 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3285 if (Opc == ARMISD::VCGE)
3286 Opc = ARMISD::VCLEZ;
3287 else if (Opc == ARMISD::VCGT)
3288 Opc = ARMISD::VCLTZ;
3293 if (SingleOp.getNode()) {
3296 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3298 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3300 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3302 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3304 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3306 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3309 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3313 Result = DAG.getNOT(dl, Result, VT);
3318 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3319 /// valid vector constant for a NEON instruction with a "modified immediate"
3320 /// operand (e.g., VMOV). If so, return the encoded value.
3321 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3322 unsigned SplatBitSize, SelectionDAG &DAG,
3323 EVT &VT, bool is128Bits, NEONModImmType type) {
3324 unsigned OpCmode, Imm;
3326 // SplatBitSize is set to the smallest size that splats the vector, so a
3327 // zero vector will always have SplatBitSize == 8. However, NEON modified
3328 // immediate instructions others than VMOV do not support the 8-bit encoding
3329 // of a zero vector, and the default encoding of zero is supposed to be the
3334 switch (SplatBitSize) {
3336 if (type != VMOVModImm)
3338 // Any 1-byte value is OK. Op=0, Cmode=1110.
3339 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3342 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3346 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3347 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3348 if ((SplatBits & ~0xff) == 0) {
3349 // Value = 0x00nn: Op=x, Cmode=100x.
3354 if ((SplatBits & ~0xff00) == 0) {
3355 // Value = 0xnn00: Op=x, Cmode=101x.
3357 Imm = SplatBits >> 8;
3363 // NEON's 32-bit VMOV supports splat values where:
3364 // * only one byte is nonzero, or
3365 // * the least significant byte is 0xff and the second byte is nonzero, or
3366 // * the least significant 2 bytes are 0xff and the third is nonzero.
3367 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3368 if ((SplatBits & ~0xff) == 0) {
3369 // Value = 0x000000nn: Op=x, Cmode=000x.
3374 if ((SplatBits & ~0xff00) == 0) {
3375 // Value = 0x0000nn00: Op=x, Cmode=001x.
3377 Imm = SplatBits >> 8;
3380 if ((SplatBits & ~0xff0000) == 0) {
3381 // Value = 0x00nn0000: Op=x, Cmode=010x.
3383 Imm = SplatBits >> 16;
3386 if ((SplatBits & ~0xff000000) == 0) {
3387 // Value = 0xnn000000: Op=x, Cmode=011x.
3389 Imm = SplatBits >> 24;
3393 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3394 if (type == OtherModImm) return SDValue();
3396 if ((SplatBits & ~0xffff) == 0 &&
3397 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3398 // Value = 0x0000nnff: Op=x, Cmode=1100.
3400 Imm = SplatBits >> 8;
3405 if ((SplatBits & ~0xffffff) == 0 &&
3406 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3407 // Value = 0x00nnffff: Op=x, Cmode=1101.
3409 Imm = SplatBits >> 16;
3410 SplatBits |= 0xffff;
3414 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3415 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3416 // VMOV.I32. A (very) minor optimization would be to replicate the value
3417 // and fall through here to test for a valid 64-bit splat. But, then the
3418 // caller would also need to check and handle the change in size.
3422 if (type != VMOVModImm)
3424 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3425 uint64_t BitMask = 0xff;
3427 unsigned ImmMask = 1;
3429 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3430 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3433 } else if ((SplatBits & BitMask) != 0) {
3439 // Op=1, Cmode=1110.
3442 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3447 llvm_unreachable("unexpected size for isNEONModifiedImm");
3451 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3452 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3455 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3456 bool &ReverseVEXT, unsigned &Imm) {
3457 unsigned NumElts = VT.getVectorNumElements();
3458 ReverseVEXT = false;
3460 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3466 // If this is a VEXT shuffle, the immediate value is the index of the first
3467 // element. The other shuffle indices must be the successive elements after
3469 unsigned ExpectedElt = Imm;
3470 for (unsigned i = 1; i < NumElts; ++i) {
3471 // Increment the expected index. If it wraps around, it may still be
3472 // a VEXT but the source vectors must be swapped.
3474 if (ExpectedElt == NumElts * 2) {
3479 if (M[i] < 0) continue; // ignore UNDEF indices
3480 if (ExpectedElt != static_cast<unsigned>(M[i]))
3484 // Adjust the index value if the source operands will be swapped.
3491 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3492 /// instruction with the specified blocksize. (The order of the elements
3493 /// within each block of the vector is reversed.)
3494 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3495 unsigned BlockSize) {
3496 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3497 "Only possible block sizes for VREV are: 16, 32, 64");
3499 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3503 unsigned NumElts = VT.getVectorNumElements();
3504 unsigned BlockElts = M[0] + 1;
3505 // If the first shuffle index is UNDEF, be optimistic.
3507 BlockElts = BlockSize / EltSz;
3509 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3512 for (unsigned i = 0; i < NumElts; ++i) {
3513 if (M[i] < 0) continue; // ignore UNDEF indices
3514 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3521 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3522 unsigned &WhichResult) {
3523 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3527 unsigned NumElts = VT.getVectorNumElements();
3528 WhichResult = (M[0] == 0 ? 0 : 1);
3529 for (unsigned i = 0; i < NumElts; i += 2) {
3530 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3531 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3537 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3538 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3539 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3540 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3541 unsigned &WhichResult) {
3542 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3546 unsigned NumElts = VT.getVectorNumElements();
3547 WhichResult = (M[0] == 0 ? 0 : 1);
3548 for (unsigned i = 0; i < NumElts; i += 2) {
3549 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3550 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3556 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3557 unsigned &WhichResult) {
3558 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3562 unsigned NumElts = VT.getVectorNumElements();
3563 WhichResult = (M[0] == 0 ? 0 : 1);
3564 for (unsigned i = 0; i != NumElts; ++i) {
3565 if (M[i] < 0) continue; // ignore UNDEF indices
3566 if ((unsigned) M[i] != 2 * i + WhichResult)
3570 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3571 if (VT.is64BitVector() && EltSz == 32)
3577 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3578 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3579 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3580 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3581 unsigned &WhichResult) {
3582 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3586 unsigned Half = VT.getVectorNumElements() / 2;
3587 WhichResult = (M[0] == 0 ? 0 : 1);
3588 for (unsigned j = 0; j != 2; ++j) {
3589 unsigned Idx = WhichResult;
3590 for (unsigned i = 0; i != Half; ++i) {
3591 int MIdx = M[i + j * Half];
3592 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3598 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3599 if (VT.is64BitVector() && EltSz == 32)
3605 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3606 unsigned &WhichResult) {
3607 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3611 unsigned NumElts = VT.getVectorNumElements();
3612 WhichResult = (M[0] == 0 ? 0 : 1);
3613 unsigned Idx = WhichResult * NumElts / 2;
3614 for (unsigned i = 0; i != NumElts; i += 2) {
3615 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3616 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3621 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3622 if (VT.is64BitVector() && EltSz == 32)
3628 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3629 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3630 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3631 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3632 unsigned &WhichResult) {
3633 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3637 unsigned NumElts = VT.getVectorNumElements();
3638 WhichResult = (M[0] == 0 ? 0 : 1);
3639 unsigned Idx = WhichResult * NumElts / 2;
3640 for (unsigned i = 0; i != NumElts; i += 2) {
3641 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3642 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3647 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3648 if (VT.is64BitVector() && EltSz == 32)
3654 // If N is an integer constant that can be moved into a register in one
3655 // instruction, return an SDValue of such a constant (will become a MOV
3656 // instruction). Otherwise return null.
3657 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3658 const ARMSubtarget *ST, DebugLoc dl) {
3660 if (!isa<ConstantSDNode>(N))
3662 Val = cast<ConstantSDNode>(N)->getZExtValue();
3664 if (ST->isThumb1Only()) {
3665 if (Val <= 255 || ~Val <= 255)
3666 return DAG.getConstant(Val, MVT::i32);
3668 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3669 return DAG.getConstant(Val, MVT::i32);
3674 // If this is a case we can't handle, return null and let the default
3675 // expansion code take care of it.
3676 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3677 const ARMSubtarget *ST) const {
3678 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3679 DebugLoc dl = Op.getDebugLoc();
3680 EVT VT = Op.getValueType();
3682 APInt SplatBits, SplatUndef;
3683 unsigned SplatBitSize;
3685 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3686 if (SplatBitSize <= 64) {
3687 // Check if an immediate VMOV works.
3689 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3690 SplatUndef.getZExtValue(), SplatBitSize,
3691 DAG, VmovVT, VT.is128BitVector(),
3693 if (Val.getNode()) {
3694 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3695 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3698 // Try an immediate VMVN.
3699 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3700 ((1LL << SplatBitSize) - 1));
3701 Val = isNEONModifiedImm(NegatedImm,
3702 SplatUndef.getZExtValue(), SplatBitSize,
3703 DAG, VmovVT, VT.is128BitVector(),
3705 if (Val.getNode()) {
3706 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3707 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3712 // Scan through the operands to see if only one value is used.
3713 unsigned NumElts = VT.getVectorNumElements();
3714 bool isOnlyLowElement = true;
3715 bool usesOnlyOneValue = true;
3716 bool isConstant = true;
3718 for (unsigned i = 0; i < NumElts; ++i) {
3719 SDValue V = Op.getOperand(i);
3720 if (V.getOpcode() == ISD::UNDEF)
3723 isOnlyLowElement = false;
3724 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3727 if (!Value.getNode())
3729 else if (V != Value)
3730 usesOnlyOneValue = false;
3733 if (!Value.getNode())
3734 return DAG.getUNDEF(VT);
3736 if (isOnlyLowElement)
3737 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3739 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3741 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3742 // i32 and try again.
3743 if (usesOnlyOneValue && EltSize <= 32) {
3745 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3746 if (VT.getVectorElementType().isFloatingPoint()) {
3747 SmallVector<SDValue, 8> Ops;
3748 for (unsigned i = 0; i < NumElts; ++i)
3749 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3751 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3752 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3753 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3755 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3757 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3759 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3762 // If all elements are constants and the case above didn't get hit, fall back
3763 // to the default expansion, which will generate a load from the constant
3768 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3770 SDValue shuffle = ReconstructShuffle(Op, DAG);
3771 if (shuffle != SDValue())
3775 // Vectors with 32- or 64-bit elements can be built by directly assigning
3776 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3777 // will be legalized.
3778 if (EltSize >= 32) {
3779 // Do the expansion with floating-point types, since that is what the VFP
3780 // registers are defined to use, and since i64 is not legal.
3781 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3782 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3783 SmallVector<SDValue, 8> Ops;
3784 for (unsigned i = 0; i < NumElts; ++i)
3785 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3786 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3787 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3793 // Gather data to see if the operation can be modelled as a
3794 // shuffle in combination with VEXTs.
3795 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
3796 SelectionDAG &DAG) const {
3797 DebugLoc dl = Op.getDebugLoc();
3798 EVT VT = Op.getValueType();
3799 unsigned NumElts = VT.getVectorNumElements();
3801 SmallVector<SDValue, 2> SourceVecs;
3802 SmallVector<unsigned, 2> MinElts;
3803 SmallVector<unsigned, 2> MaxElts;
3805 for (unsigned i = 0; i < NumElts; ++i) {
3806 SDValue V = Op.getOperand(i);
3807 if (V.getOpcode() == ISD::UNDEF)
3809 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3810 // A shuffle can only come from building a vector from various
3811 // elements of other vectors.
3815 // Record this extraction against the appropriate vector if possible...
3816 SDValue SourceVec = V.getOperand(0);
3817 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3818 bool FoundSource = false;
3819 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3820 if (SourceVecs[j] == SourceVec) {
3821 if (MinElts[j] > EltNo)
3823 if (MaxElts[j] < EltNo)
3830 // Or record a new source if not...
3832 SourceVecs.push_back(SourceVec);
3833 MinElts.push_back(EltNo);
3834 MaxElts.push_back(EltNo);
3838 // Currently only do something sane when at most two source vectors
3840 if (SourceVecs.size() > 2)
3843 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3844 int VEXTOffsets[2] = {0, 0};
3846 // This loop extracts the usage patterns of the source vectors
3847 // and prepares appropriate SDValues for a shuffle if possible.
3848 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3849 if (SourceVecs[i].getValueType() == VT) {
3850 // No VEXT necessary
3851 ShuffleSrcs[i] = SourceVecs[i];
3854 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3855 // It probably isn't worth padding out a smaller vector just to
3856 // break it down again in a shuffle.
3860 // Since only 64-bit and 128-bit vectors are legal on ARM and
3861 // we've eliminated the other cases...
3862 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
3863 "unexpected vector sizes in ReconstructShuffle");
3865 if (MaxElts[i] - MinElts[i] >= NumElts) {
3866 // Span too large for a VEXT to cope
3870 if (MinElts[i] >= NumElts) {
3871 // The extraction can just take the second half
3872 VEXTOffsets[i] = NumElts;
3873 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3875 DAG.getIntPtrConstant(NumElts));
3876 } else if (MaxElts[i] < NumElts) {
3877 // The extraction can just take the first half
3879 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3881 DAG.getIntPtrConstant(0));
3883 // An actual VEXT is needed
3884 VEXTOffsets[i] = MinElts[i];
3885 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3887 DAG.getIntPtrConstant(0));
3888 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3890 DAG.getIntPtrConstant(NumElts));
3891 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
3892 DAG.getConstant(VEXTOffsets[i], MVT::i32));
3896 SmallVector<int, 8> Mask;
3898 for (unsigned i = 0; i < NumElts; ++i) {
3899 SDValue Entry = Op.getOperand(i);
3900 if (Entry.getOpcode() == ISD::UNDEF) {
3905 SDValue ExtractVec = Entry.getOperand(0);
3906 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
3907 .getOperand(1))->getSExtValue();
3908 if (ExtractVec == SourceVecs[0]) {
3909 Mask.push_back(ExtractElt - VEXTOffsets[0]);
3911 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
3915 // Final check before we try to produce nonsense...
3916 if (isShuffleMaskLegal(Mask, VT))
3917 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
3923 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3924 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3925 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3926 /// are assumed to be legal.
3928 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3930 if (VT.getVectorNumElements() == 4 &&
3931 (VT.is128BitVector() || VT.is64BitVector())) {
3932 unsigned PFIndexes[4];
3933 for (unsigned i = 0; i != 4; ++i) {
3937 PFIndexes[i] = M[i];
3940 // Compute the index in the perfect shuffle table.
3941 unsigned PFTableIndex =
3942 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3943 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3944 unsigned Cost = (PFEntry >> 30);
3951 unsigned Imm, WhichResult;
3953 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3954 return (EltSize >= 32 ||
3955 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3956 isVREVMask(M, VT, 64) ||
3957 isVREVMask(M, VT, 32) ||
3958 isVREVMask(M, VT, 16) ||
3959 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3960 isVTRNMask(M, VT, WhichResult) ||
3961 isVUZPMask(M, VT, WhichResult) ||
3962 isVZIPMask(M, VT, WhichResult) ||
3963 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3964 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3965 isVZIP_v_undef_Mask(M, VT, WhichResult));
3968 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3969 /// the specified operations to build the shuffle.
3970 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3971 SDValue RHS, SelectionDAG &DAG,
3973 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3974 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3975 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3978 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3987 OP_VUZPL, // VUZP, left result
3988 OP_VUZPR, // VUZP, right result
3989 OP_VZIPL, // VZIP, left result
3990 OP_VZIPR, // VZIP, right result
3991 OP_VTRNL, // VTRN, left result
3992 OP_VTRNR // VTRN, right result
3995 if (OpNum == OP_COPY) {
3996 if (LHSID == (1*9+2)*9+3) return LHS;
3997 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4001 SDValue OpLHS, OpRHS;
4002 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4003 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4004 EVT VT = OpLHS.getValueType();
4007 default: llvm_unreachable("Unknown shuffle opcode!");
4009 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4014 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4015 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4019 return DAG.getNode(ARMISD::VEXT, dl, VT,
4021 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4024 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4025 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4028 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4029 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4032 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4033 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4037 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4038 SDValue V1 = Op.getOperand(0);
4039 SDValue V2 = Op.getOperand(1);
4040 DebugLoc dl = Op.getDebugLoc();
4041 EVT VT = Op.getValueType();
4042 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4043 SmallVector<int, 8> ShuffleMask;
4045 // Convert shuffles that are directly supported on NEON to target-specific
4046 // DAG nodes, instead of keeping them as shuffles and matching them again
4047 // during code selection. This is more efficient and avoids the possibility
4048 // of inconsistencies between legalization and selection.
4049 // FIXME: floating-point vectors should be canonicalized to integer vectors
4050 // of the same time so that they get CSEd properly.
4051 SVN->getMask(ShuffleMask);
4053 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4054 if (EltSize <= 32) {
4055 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4056 int Lane = SVN->getSplatIndex();
4057 // If this is undef splat, generate it via "just" vdup, if possible.
4058 if (Lane == -1) Lane = 0;
4060 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4061 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4063 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4064 DAG.getConstant(Lane, MVT::i32));
4069 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4072 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4073 DAG.getConstant(Imm, MVT::i32));
4076 if (isVREVMask(ShuffleMask, VT, 64))
4077 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4078 if (isVREVMask(ShuffleMask, VT, 32))
4079 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4080 if (isVREVMask(ShuffleMask, VT, 16))
4081 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4083 // Check for Neon shuffles that modify both input vectors in place.
4084 // If both results are used, i.e., if there are two shuffles with the same
4085 // source operands and with masks corresponding to both results of one of
4086 // these operations, DAG memoization will ensure that a single node is
4087 // used for both shuffles.
4088 unsigned WhichResult;
4089 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4090 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4091 V1, V2).getValue(WhichResult);
4092 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4093 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4094 V1, V2).getValue(WhichResult);
4095 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4096 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4097 V1, V2).getValue(WhichResult);
4099 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4100 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4101 V1, V1).getValue(WhichResult);
4102 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4103 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4104 V1, V1).getValue(WhichResult);
4105 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4106 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4107 V1, V1).getValue(WhichResult);
4110 // If the shuffle is not directly supported and it has 4 elements, use
4111 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4112 unsigned NumElts = VT.getVectorNumElements();
4114 unsigned PFIndexes[4];
4115 for (unsigned i = 0; i != 4; ++i) {
4116 if (ShuffleMask[i] < 0)
4119 PFIndexes[i] = ShuffleMask[i];
4122 // Compute the index in the perfect shuffle table.
4123 unsigned PFTableIndex =
4124 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4125 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4126 unsigned Cost = (PFEntry >> 30);
4129 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4132 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4133 if (EltSize >= 32) {
4134 // Do the expansion with floating-point types, since that is what the VFP
4135 // registers are defined to use, and since i64 is not legal.
4136 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4137 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4138 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4139 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4140 SmallVector<SDValue, 8> Ops;
4141 for (unsigned i = 0; i < NumElts; ++i) {
4142 if (ShuffleMask[i] < 0)
4143 Ops.push_back(DAG.getUNDEF(EltVT));
4145 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4146 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4147 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4150 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4151 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4157 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4158 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4159 SDValue Lane = Op.getOperand(1);
4160 if (!isa<ConstantSDNode>(Lane))
4163 SDValue Vec = Op.getOperand(0);
4164 if (Op.getValueType() == MVT::i32 &&
4165 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4166 DebugLoc dl = Op.getDebugLoc();
4167 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4173 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4174 // The only time a CONCAT_VECTORS operation can have legal types is when
4175 // two 64-bit vectors are concatenated to a 128-bit vector.
4176 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4177 "unexpected CONCAT_VECTORS");
4178 DebugLoc dl = Op.getDebugLoc();
4179 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4180 SDValue Op0 = Op.getOperand(0);
4181 SDValue Op1 = Op.getOperand(1);
4182 if (Op0.getOpcode() != ISD::UNDEF)
4183 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4184 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4185 DAG.getIntPtrConstant(0));
4186 if (Op1.getOpcode() != ISD::UNDEF)
4187 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4188 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4189 DAG.getIntPtrConstant(1));
4190 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4193 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4194 /// element has been zero/sign-extended, depending on the isSigned parameter,
4195 /// from an integer type half its size.
4196 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4198 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4199 EVT VT = N->getValueType(0);
4200 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4201 SDNode *BVN = N->getOperand(0).getNode();
4202 if (BVN->getValueType(0) != MVT::v4i32 ||
4203 BVN->getOpcode() != ISD::BUILD_VECTOR)
4205 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4206 unsigned HiElt = 1 - LoElt;
4207 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4208 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4209 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4210 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4211 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4214 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4215 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4218 if (Hi0->isNullValue() && Hi1->isNullValue())
4224 if (N->getOpcode() != ISD::BUILD_VECTOR)
4227 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4228 SDNode *Elt = N->getOperand(i).getNode();
4229 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4230 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4231 unsigned HalfSize = EltSize / 2;
4233 int64_t SExtVal = C->getSExtValue();
4234 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4237 if ((C->getZExtValue() >> HalfSize) != 0)
4248 /// isSignExtended - Check if a node is a vector value that is sign-extended
4249 /// or a constant BUILD_VECTOR with sign-extended elements.
4250 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4251 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4253 if (isExtendedBUILD_VECTOR(N, DAG, true))
4258 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4259 /// or a constant BUILD_VECTOR with zero-extended elements.
4260 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4261 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4263 if (isExtendedBUILD_VECTOR(N, DAG, false))
4268 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4269 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4270 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4271 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4272 return N->getOperand(0);
4273 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4274 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4275 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4276 LD->isNonTemporal(), LD->getAlignment());
4277 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4278 // have been legalized as a BITCAST from v4i32.
4279 if (N->getOpcode() == ISD::BITCAST) {
4280 SDNode *BVN = N->getOperand(0).getNode();
4281 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4282 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4283 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4284 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4285 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4287 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4288 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4289 EVT VT = N->getValueType(0);
4290 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4291 unsigned NumElts = VT.getVectorNumElements();
4292 MVT TruncVT = MVT::getIntegerVT(EltSize);
4293 SmallVector<SDValue, 8> Ops;
4294 for (unsigned i = 0; i != NumElts; ++i) {
4295 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4296 const APInt &CInt = C->getAPIntValue();
4297 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4299 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4300 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4303 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4304 // Multiplications are only custom-lowered for 128-bit vectors so that
4305 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4306 EVT VT = Op.getValueType();
4307 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4308 SDNode *N0 = Op.getOperand(0).getNode();
4309 SDNode *N1 = Op.getOperand(1).getNode();
4310 unsigned NewOpc = 0;
4311 if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
4312 NewOpc = ARMISD::VMULLs;
4313 else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
4314 NewOpc = ARMISD::VMULLu;
4315 else if (VT == MVT::v2i64)
4316 // Fall through to expand this. It is not legal.
4319 // Other vector multiplications are legal.
4322 // Legalize to a VMULL instruction.
4323 DebugLoc DL = Op.getDebugLoc();
4324 SDValue Op0 = SkipExtension(N0, DAG);
4325 SDValue Op1 = SkipExtension(N1, DAG);
4327 assert(Op0.getValueType().is64BitVector() &&
4328 Op1.getValueType().is64BitVector() &&
4329 "unexpected types for extended operands to VMULL");
4330 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4334 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4336 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4337 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4338 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4339 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4340 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4341 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4342 // Get reciprocal estimate.
4343 // float4 recip = vrecpeq_f32(yf);
4344 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4345 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4346 // Because char has a smaller range than uchar, we can actually get away
4347 // without any newton steps. This requires that we use a weird bias
4348 // of 0xb000, however (again, this has been exhaustively tested).
4349 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4350 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4351 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4352 Y = DAG.getConstant(0xb000, MVT::i32);
4353 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4354 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4355 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4356 // Convert back to short.
4357 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4358 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4363 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4365 // Convert to float.
4366 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4367 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4368 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4369 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4370 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4371 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4373 // Use reciprocal estimate and one refinement step.
4374 // float4 recip = vrecpeq_f32(yf);
4375 // recip *= vrecpsq_f32(yf, recip);
4376 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4377 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4378 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4379 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4381 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4382 // Because short has a smaller range than ushort, we can actually get away
4383 // with only a single newton step. This requires that we use a weird bias
4384 // of 89, however (again, this has been exhaustively tested).
4385 // float4 result = as_float4(as_int4(xf*recip) + 89);
4386 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4387 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4388 N1 = DAG.getConstant(89, MVT::i32);
4389 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4390 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4391 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4392 // Convert back to integer and return.
4393 // return vmovn_s32(vcvt_s32_f32(result));
4394 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4395 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4399 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4400 EVT VT = Op.getValueType();
4401 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4402 "unexpected type for custom-lowering ISD::SDIV");
4404 DebugLoc dl = Op.getDebugLoc();
4405 SDValue N0 = Op.getOperand(0);
4406 SDValue N1 = Op.getOperand(1);
4409 if (VT == MVT::v8i8) {
4410 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4411 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4413 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4414 DAG.getIntPtrConstant(4));
4415 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4416 DAG.getIntPtrConstant(4));
4417 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4418 DAG.getIntPtrConstant(0));
4419 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4420 DAG.getIntPtrConstant(0));
4422 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4423 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4425 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4426 N0 = LowerCONCAT_VECTORS(N0, DAG);
4428 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4431 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4434 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4435 EVT VT = Op.getValueType();
4436 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4437 "unexpected type for custom-lowering ISD::UDIV");
4439 DebugLoc dl = Op.getDebugLoc();
4440 SDValue N0 = Op.getOperand(0);
4441 SDValue N1 = Op.getOperand(1);
4444 if (VT == MVT::v8i8) {
4445 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4446 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4448 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4449 DAG.getIntPtrConstant(4));
4450 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4451 DAG.getIntPtrConstant(4));
4452 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4453 DAG.getIntPtrConstant(0));
4454 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4455 DAG.getIntPtrConstant(0));
4457 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4458 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4460 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4461 N0 = LowerCONCAT_VECTORS(N0, DAG);
4463 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4464 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4469 // v4i16 sdiv ... Convert to float.
4470 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4471 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4472 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4473 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4474 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4475 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4477 // Use reciprocal estimate and two refinement steps.
4478 // float4 recip = vrecpeq_f32(yf);
4479 // recip *= vrecpsq_f32(yf, recip);
4480 // recip *= vrecpsq_f32(yf, recip);
4481 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4482 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4483 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4484 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4486 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4487 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4488 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4490 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4491 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4492 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4493 // and that it will never cause us to return an answer too large).
4494 // float4 result = as_float4(as_int4(xf*recip) + 89);
4495 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4496 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4497 N1 = DAG.getConstant(2, MVT::i32);
4498 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4499 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4500 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4501 // Convert back to integer and return.
4502 // return vmovn_u32(vcvt_s32_f32(result));
4503 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4504 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4508 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4509 switch (Op.getOpcode()) {
4510 default: llvm_unreachable("Don't know how to custom lower this!");
4511 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4512 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4513 case ISD::GlobalAddress:
4514 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4515 LowerGlobalAddressELF(Op, DAG);
4516 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4517 case ISD::SELECT: return LowerSELECT(Op, DAG);
4518 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4519 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4520 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4521 case ISD::VASTART: return LowerVASTART(Op, DAG);
4522 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4523 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4524 case ISD::SINT_TO_FP:
4525 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4526 case ISD::FP_TO_SINT:
4527 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4528 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4529 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4530 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4531 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4532 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4533 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4534 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4535 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4537 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4540 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4541 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4542 case ISD::SRL_PARTS:
4543 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4544 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4545 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4546 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4547 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4548 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4549 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4550 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4551 case ISD::MUL: return LowerMUL(Op, DAG);
4552 case ISD::SDIV: return LowerSDIV(Op, DAG);
4553 case ISD::UDIV: return LowerUDIV(Op, DAG);
4558 /// ReplaceNodeResults - Replace the results of node with an illegal result
4559 /// type with new values built out of custom code.
4560 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4561 SmallVectorImpl<SDValue>&Results,
4562 SelectionDAG &DAG) const {
4564 switch (N->getOpcode()) {
4566 llvm_unreachable("Don't know how to custom expand this!");
4569 Res = ExpandBITCAST(N, DAG);
4573 Res = Expand64BitShift(N, DAG, Subtarget);
4577 Results.push_back(Res);
4580 //===----------------------------------------------------------------------===//
4581 // ARM Scheduler Hooks
4582 //===----------------------------------------------------------------------===//
4585 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4586 MachineBasicBlock *BB,
4587 unsigned Size) const {
4588 unsigned dest = MI->getOperand(0).getReg();
4589 unsigned ptr = MI->getOperand(1).getReg();
4590 unsigned oldval = MI->getOperand(2).getReg();
4591 unsigned newval = MI->getOperand(3).getReg();
4592 unsigned scratch = BB->getParent()->getRegInfo()
4593 .createVirtualRegister(ARM::GPRRegisterClass);
4594 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4595 DebugLoc dl = MI->getDebugLoc();
4596 bool isThumb2 = Subtarget->isThumb2();
4598 unsigned ldrOpc, strOpc;
4600 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4602 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4603 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4606 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4607 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4610 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4611 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4615 MachineFunction *MF = BB->getParent();
4616 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4617 MachineFunction::iterator It = BB;
4618 ++It; // insert the new blocks after the current block
4620 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4621 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4622 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4623 MF->insert(It, loop1MBB);
4624 MF->insert(It, loop2MBB);
4625 MF->insert(It, exitMBB);
4627 // Transfer the remainder of BB and its successor edges to exitMBB.
4628 exitMBB->splice(exitMBB->begin(), BB,
4629 llvm::next(MachineBasicBlock::iterator(MI)),
4631 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4635 // fallthrough --> loop1MBB
4636 BB->addSuccessor(loop1MBB);
4639 // ldrex dest, [ptr]
4643 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4644 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4645 .addReg(dest).addReg(oldval));
4646 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4647 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4648 BB->addSuccessor(loop2MBB);
4649 BB->addSuccessor(exitMBB);
4652 // strex scratch, newval, [ptr]
4656 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4658 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4659 .addReg(scratch).addImm(0));
4660 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4661 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4662 BB->addSuccessor(loop1MBB);
4663 BB->addSuccessor(exitMBB);
4669 MI->eraseFromParent(); // The instruction is gone now.
4675 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4676 unsigned Size, unsigned BinOpcode) const {
4677 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4678 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4680 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4681 MachineFunction *MF = BB->getParent();
4682 MachineFunction::iterator It = BB;
4685 unsigned dest = MI->getOperand(0).getReg();
4686 unsigned ptr = MI->getOperand(1).getReg();
4687 unsigned incr = MI->getOperand(2).getReg();
4688 DebugLoc dl = MI->getDebugLoc();
4690 bool isThumb2 = Subtarget->isThumb2();
4691 unsigned ldrOpc, strOpc;
4693 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4695 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4696 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4699 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4700 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4703 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4704 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4708 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4709 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4710 MF->insert(It, loopMBB);
4711 MF->insert(It, exitMBB);
4713 // Transfer the remainder of BB and its successor edges to exitMBB.
4714 exitMBB->splice(exitMBB->begin(), BB,
4715 llvm::next(MachineBasicBlock::iterator(MI)),
4717 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4719 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4720 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4721 unsigned scratch2 = (!BinOpcode) ? incr :
4722 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4726 // fallthrough --> loopMBB
4727 BB->addSuccessor(loopMBB);
4731 // <binop> scratch2, dest, incr
4732 // strex scratch, scratch2, ptr
4735 // fallthrough --> exitMBB
4737 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4739 // operand order needs to go the other way for NAND
4740 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4741 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4742 addReg(incr).addReg(dest)).addReg(0);
4744 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4745 addReg(dest).addReg(incr)).addReg(0);
4748 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4750 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4751 .addReg(scratch).addImm(0));
4752 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4753 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4755 BB->addSuccessor(loopMBB);
4756 BB->addSuccessor(exitMBB);
4762 MI->eraseFromParent(); // The instruction is gone now.
4768 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4769 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4770 E = MBB->succ_end(); I != E; ++I)
4773 llvm_unreachable("Expecting a BB with two successors!");
4777 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4778 MachineBasicBlock *BB) const {
4779 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4780 DebugLoc dl = MI->getDebugLoc();
4781 bool isThumb2 = Subtarget->isThumb2();
4782 switch (MI->getOpcode()) {
4785 llvm_unreachable("Unexpected instr type to insert");
4787 case ARM::ATOMIC_LOAD_ADD_I8:
4788 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4789 case ARM::ATOMIC_LOAD_ADD_I16:
4790 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4791 case ARM::ATOMIC_LOAD_ADD_I32:
4792 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4794 case ARM::ATOMIC_LOAD_AND_I8:
4795 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4796 case ARM::ATOMIC_LOAD_AND_I16:
4797 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4798 case ARM::ATOMIC_LOAD_AND_I32:
4799 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4801 case ARM::ATOMIC_LOAD_OR_I8:
4802 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4803 case ARM::ATOMIC_LOAD_OR_I16:
4804 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4805 case ARM::ATOMIC_LOAD_OR_I32:
4806 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4808 case ARM::ATOMIC_LOAD_XOR_I8:
4809 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4810 case ARM::ATOMIC_LOAD_XOR_I16:
4811 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4812 case ARM::ATOMIC_LOAD_XOR_I32:
4813 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4815 case ARM::ATOMIC_LOAD_NAND_I8:
4816 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4817 case ARM::ATOMIC_LOAD_NAND_I16:
4818 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4819 case ARM::ATOMIC_LOAD_NAND_I32:
4820 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4822 case ARM::ATOMIC_LOAD_SUB_I8:
4823 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4824 case ARM::ATOMIC_LOAD_SUB_I16:
4825 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4826 case ARM::ATOMIC_LOAD_SUB_I32:
4827 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4829 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4830 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4831 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4833 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4834 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4835 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4837 case ARM::tMOVCCr_pseudo: {
4838 // To "insert" a SELECT_CC instruction, we actually have to insert the
4839 // diamond control-flow pattern. The incoming instruction knows the
4840 // destination vreg to set, the condition code register to branch on, the
4841 // true/false values to select between, and a branch opcode to use.
4842 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4843 MachineFunction::iterator It = BB;
4849 // cmpTY ccX, r1, r2
4851 // fallthrough --> copy0MBB
4852 MachineBasicBlock *thisMBB = BB;
4853 MachineFunction *F = BB->getParent();
4854 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4855 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4856 F->insert(It, copy0MBB);
4857 F->insert(It, sinkMBB);
4859 // Transfer the remainder of BB and its successor edges to sinkMBB.
4860 sinkMBB->splice(sinkMBB->begin(), BB,
4861 llvm::next(MachineBasicBlock::iterator(MI)),
4863 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4865 BB->addSuccessor(copy0MBB);
4866 BB->addSuccessor(sinkMBB);
4868 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4869 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4872 // %FalseValue = ...
4873 // # fallthrough to sinkMBB
4876 // Update machine-CFG edges
4877 BB->addSuccessor(sinkMBB);
4880 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4883 BuildMI(*BB, BB->begin(), dl,
4884 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4885 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4886 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4888 MI->eraseFromParent(); // The pseudo instruction is gone now.
4893 case ARM::BCCZi64: {
4894 // If there is an unconditional branch to the other successor, remove it.
4895 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
4897 // Compare both parts that make up the double comparison separately for
4899 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4901 unsigned LHS1 = MI->getOperand(1).getReg();
4902 unsigned LHS2 = MI->getOperand(2).getReg();
4904 AddDefaultPred(BuildMI(BB, dl,
4905 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4906 .addReg(LHS1).addImm(0));
4907 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4908 .addReg(LHS2).addImm(0)
4909 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4911 unsigned RHS1 = MI->getOperand(3).getReg();
4912 unsigned RHS2 = MI->getOperand(4).getReg();
4913 AddDefaultPred(BuildMI(BB, dl,
4914 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4915 .addReg(LHS1).addReg(RHS1));
4916 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4917 .addReg(LHS2).addReg(RHS2)
4918 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4921 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4922 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4923 if (MI->getOperand(0).getImm() == ARMCC::NE)
4924 std::swap(destMBB, exitMBB);
4926 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4927 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4928 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4931 MI->eraseFromParent(); // The pseudo instruction is gone now.
4937 //===----------------------------------------------------------------------===//
4938 // ARM Optimization Hooks
4939 //===----------------------------------------------------------------------===//
4942 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4943 TargetLowering::DAGCombinerInfo &DCI) {
4944 SelectionDAG &DAG = DCI.DAG;
4945 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4946 EVT VT = N->getValueType(0);
4947 unsigned Opc = N->getOpcode();
4948 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4949 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4950 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4951 ISD::CondCode CC = ISD::SETCC_INVALID;
4954 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4956 SDValue CCOp = Slct.getOperand(0);
4957 if (CCOp.getOpcode() == ISD::SETCC)
4958 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4961 bool DoXform = false;
4963 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4966 if (LHS.getOpcode() == ISD::Constant &&
4967 cast<ConstantSDNode>(LHS)->isNullValue()) {
4969 } else if (CC != ISD::SETCC_INVALID &&
4970 RHS.getOpcode() == ISD::Constant &&
4971 cast<ConstantSDNode>(RHS)->isNullValue()) {
4972 std::swap(LHS, RHS);
4973 SDValue Op0 = Slct.getOperand(0);
4974 EVT OpVT = isSlctCC ? Op0.getValueType() :
4975 Op0.getOperand(0).getValueType();
4976 bool isInt = OpVT.isInteger();
4977 CC = ISD::getSetCCInverse(CC, isInt);
4979 if (!TLI.isCondCodeLegal(CC, OpVT))
4980 return SDValue(); // Inverse operator isn't legal.
4987 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4989 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4990 Slct.getOperand(0), Slct.getOperand(1), CC);
4991 SDValue CCOp = Slct.getOperand(0);
4993 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4994 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4995 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4996 CCOp, OtherOp, Result);
5001 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
5002 /// operands N0 and N1. This is a helper for PerformADDCombine that is
5003 /// called with the default operands, and if that fails, with commuted
5005 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
5006 TargetLowering::DAGCombinerInfo &DCI) {
5007 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
5008 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
5009 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
5010 if (Result.getNode()) return Result;
5015 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
5017 static SDValue PerformADDCombine(SDNode *N,
5018 TargetLowering::DAGCombinerInfo &DCI) {
5019 SDValue N0 = N->getOperand(0);
5020 SDValue N1 = N->getOperand(1);
5022 // First try with the default operand order.
5023 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
5024 if (Result.getNode())
5027 // If that didn't work, try again with the operands commuted.
5028 return PerformADDCombineWithOperands(N, N1, N0, DCI);
5031 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
5033 static SDValue PerformSUBCombine(SDNode *N,
5034 TargetLowering::DAGCombinerInfo &DCI) {
5035 SDValue N0 = N->getOperand(0);
5036 SDValue N1 = N->getOperand(1);
5038 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
5039 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
5040 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
5041 if (Result.getNode()) return Result;
5047 static SDValue PerformMULCombine(SDNode *N,
5048 TargetLowering::DAGCombinerInfo &DCI,
5049 const ARMSubtarget *Subtarget) {
5050 SelectionDAG &DAG = DCI.DAG;
5052 if (Subtarget->isThumb1Only())
5055 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5058 EVT VT = N->getValueType(0);
5062 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
5066 uint64_t MulAmt = C->getZExtValue();
5067 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
5068 ShiftAmt = ShiftAmt & (32 - 1);
5069 SDValue V = N->getOperand(0);
5070 DebugLoc DL = N->getDebugLoc();
5073 MulAmt >>= ShiftAmt;
5074 if (isPowerOf2_32(MulAmt - 1)) {
5075 // (mul x, 2^N + 1) => (add (shl x, N), x)
5076 Res = DAG.getNode(ISD::ADD, DL, VT,
5077 V, DAG.getNode(ISD::SHL, DL, VT,
5078 V, DAG.getConstant(Log2_32(MulAmt-1),
5080 } else if (isPowerOf2_32(MulAmt + 1)) {
5081 // (mul x, 2^N - 1) => (sub (shl x, N), x)
5082 Res = DAG.getNode(ISD::SUB, DL, VT,
5083 DAG.getNode(ISD::SHL, DL, VT,
5084 V, DAG.getConstant(Log2_32(MulAmt+1),
5091 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
5092 DAG.getConstant(ShiftAmt, MVT::i32));
5094 // Do not add new nodes to DAG combiner worklist.
5095 DCI.CombineTo(N, Res, false);
5099 static SDValue PerformANDCombine(SDNode *N,
5100 TargetLowering::DAGCombinerInfo &DCI) {
5101 // Attempt to use immediate-form VBIC
5102 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5103 DebugLoc dl = N->getDebugLoc();
5104 EVT VT = N->getValueType(0);
5105 SelectionDAG &DAG = DCI.DAG;
5107 APInt SplatBits, SplatUndef;
5108 unsigned SplatBitSize;
5111 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5112 if (SplatBitSize <= 64) {
5114 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
5115 SplatUndef.getZExtValue(), SplatBitSize,
5116 DAG, VbicVT, VT.is128BitVector(),
5118 if (Val.getNode()) {
5120 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
5121 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
5122 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
5130 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
5131 static SDValue PerformORCombine(SDNode *N,
5132 TargetLowering::DAGCombinerInfo &DCI,
5133 const ARMSubtarget *Subtarget) {
5134 // Attempt to use immediate-form VORR
5135 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5136 DebugLoc dl = N->getDebugLoc();
5137 EVT VT = N->getValueType(0);
5138 SelectionDAG &DAG = DCI.DAG;
5140 APInt SplatBits, SplatUndef;
5141 unsigned SplatBitSize;
5143 if (BVN && Subtarget->hasNEON() &&
5144 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5145 if (SplatBitSize <= 64) {
5147 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5148 SplatUndef.getZExtValue(), SplatBitSize,
5149 DAG, VorrVT, VT.is128BitVector(),
5151 if (Val.getNode()) {
5153 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
5154 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
5155 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
5160 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
5163 // BFI is only available on V6T2+
5164 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
5167 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
5168 DebugLoc DL = N->getDebugLoc();
5169 // 1) or (and A, mask), val => ARMbfi A, val, mask
5170 // iff (val & mask) == val
5172 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5173 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
5174 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
5175 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
5176 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
5177 // (i.e., copy a bitfield value into another bitfield of the same width)
5178 if (N0.getOpcode() != ISD::AND)
5184 SDValue N00 = N0.getOperand(0);
5186 // The value and the mask need to be constants so we can verify this is
5187 // actually a bitfield set. If the mask is 0xffff, we can do better
5188 // via a movt instruction, so don't use BFI in that case.
5189 SDValue MaskOp = N0.getOperand(1);
5190 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
5193 unsigned Mask = MaskC->getZExtValue();
5197 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
5198 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
5200 unsigned Val = N1C->getZExtValue();
5201 if ((Val & ~Mask) != Val)
5204 if (ARM::isBitFieldInvertedMask(Mask)) {
5205 Val >>= CountTrailingZeros_32(~Mask);
5207 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
5208 DAG.getConstant(Val, MVT::i32),
5209 DAG.getConstant(Mask, MVT::i32));
5211 // Do not add new nodes to DAG combiner worklist.
5212 DCI.CombineTo(N, Res, false);
5215 } else if (N1.getOpcode() == ISD::AND) {
5216 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5217 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5220 unsigned Mask2 = N11C->getZExtValue();
5222 if (ARM::isBitFieldInvertedMask(Mask) &&
5223 ARM::isBitFieldInvertedMask(~Mask2) &&
5224 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
5225 // The pack halfword instruction works better for masks that fit it,
5226 // so use that when it's available.
5227 if (Subtarget->hasT2ExtractPack() &&
5228 (Mask == 0xffff || Mask == 0xffff0000))
5231 unsigned lsb = CountTrailingZeros_32(Mask2);
5232 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
5233 DAG.getConstant(lsb, MVT::i32));
5234 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
5235 DAG.getConstant(Mask, MVT::i32));
5236 // Do not add new nodes to DAG combiner worklist.
5237 DCI.CombineTo(N, Res, false);
5239 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
5240 ARM::isBitFieldInvertedMask(Mask2) &&
5241 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
5242 // The pack halfword instruction works better for masks that fit it,
5243 // so use that when it's available.
5244 if (Subtarget->hasT2ExtractPack() &&
5245 (Mask2 == 0xffff || Mask2 == 0xffff0000))
5248 unsigned lsb = CountTrailingZeros_32(Mask);
5249 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
5250 DAG.getConstant(lsb, MVT::i32));
5251 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
5252 DAG.getConstant(Mask2, MVT::i32));
5253 // Do not add new nodes to DAG combiner worklist.
5254 DCI.CombineTo(N, Res, false);
5259 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
5260 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
5261 ARM::isBitFieldInvertedMask(~Mask)) {
5262 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
5263 // where lsb(mask) == #shamt and masked bits of B are known zero.
5264 SDValue ShAmt = N00.getOperand(1);
5265 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5266 unsigned LSB = CountTrailingZeros_32(Mask);
5270 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
5271 DAG.getConstant(~Mask, MVT::i32));
5273 // Do not add new nodes to DAG combiner worklist.
5274 DCI.CombineTo(N, Res, false);
5280 /// PerformBFICombine - (bfi A, (and B, C1), C2) -> (bfi A, B, C2) iff
5282 static SDValue PerformBFICombine(SDNode *N,
5283 TargetLowering::DAGCombinerInfo &DCI) {
5284 SDValue N1 = N->getOperand(1);
5285 if (N1.getOpcode() == ISD::AND) {
5286 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5289 unsigned Mask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
5290 unsigned Mask2 = N11C->getZExtValue();
5291 if ((Mask & Mask2) == Mask2)
5292 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
5293 N->getOperand(0), N1.getOperand(0),
5299 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
5300 /// ARMISD::VMOVRRD.
5301 static SDValue PerformVMOVRRDCombine(SDNode *N,
5302 TargetLowering::DAGCombinerInfo &DCI) {
5303 // vmovrrd(vmovdrr x, y) -> x,y
5304 SDValue InDouble = N->getOperand(0);
5305 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
5306 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
5310 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
5311 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
5312 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
5313 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
5314 SDValue Op0 = N->getOperand(0);
5315 SDValue Op1 = N->getOperand(1);
5316 if (Op0.getOpcode() == ISD::BITCAST)
5317 Op0 = Op0.getOperand(0);
5318 if (Op1.getOpcode() == ISD::BITCAST)
5319 Op1 = Op1.getOperand(0);
5320 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
5321 Op0.getNode() == Op1.getNode() &&
5322 Op0.getResNo() == 0 && Op1.getResNo() == 1)
5323 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5324 N->getValueType(0), Op0.getOperand(0));
5328 /// PerformSTORECombine - Target-specific dag combine xforms for
5330 static SDValue PerformSTORECombine(SDNode *N,
5331 TargetLowering::DAGCombinerInfo &DCI) {
5332 // Bitcast an i64 store extracted from a vector to f64.
5333 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5334 StoreSDNode *St = cast<StoreSDNode>(N);
5335 SDValue StVal = St->getValue();
5336 if (!ISD::isNormalStore(St) || St->isVolatile() ||
5337 StVal.getValueType() != MVT::i64 ||
5338 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5341 SelectionDAG &DAG = DCI.DAG;
5342 DebugLoc dl = StVal.getDebugLoc();
5343 SDValue IntVec = StVal.getOperand(0);
5344 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5345 IntVec.getValueType().getVectorNumElements());
5346 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
5347 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5348 Vec, StVal.getOperand(1));
5349 dl = N->getDebugLoc();
5350 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
5351 // Make the DAGCombiner fold the bitcasts.
5352 DCI.AddToWorklist(Vec.getNode());
5353 DCI.AddToWorklist(ExtElt.getNode());
5354 DCI.AddToWorklist(V.getNode());
5355 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
5356 St->getPointerInfo(), St->isVolatile(),
5357 St->isNonTemporal(), St->getAlignment(),
5361 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
5362 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
5363 /// i64 vector to have f64 elements, since the value can then be loaded
5364 /// directly into a VFP register.
5365 static bool hasNormalLoadOperand(SDNode *N) {
5366 unsigned NumElts = N->getValueType(0).getVectorNumElements();
5367 for (unsigned i = 0; i < NumElts; ++i) {
5368 SDNode *Elt = N->getOperand(i).getNode();
5369 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
5375 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
5376 /// ISD::BUILD_VECTOR.
5377 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
5378 TargetLowering::DAGCombinerInfo &DCI){
5379 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
5380 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
5381 // into a pair of GPRs, which is fine when the value is used as a scalar,
5382 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
5383 SelectionDAG &DAG = DCI.DAG;
5384 if (N->getNumOperands() == 2) {
5385 SDValue RV = PerformVMOVDRRCombine(N, DAG);
5390 // Load i64 elements as f64 values so that type legalization does not split
5391 // them up into i32 values.
5392 EVT VT = N->getValueType(0);
5393 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
5395 DebugLoc dl = N->getDebugLoc();
5396 SmallVector<SDValue, 8> Ops;
5397 unsigned NumElts = VT.getVectorNumElements();
5398 for (unsigned i = 0; i < NumElts; ++i) {
5399 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
5401 // Make the DAGCombiner fold the bitcast.
5402 DCI.AddToWorklist(V.getNode());
5404 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
5405 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
5406 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
5409 /// PerformInsertEltCombine - Target-specific dag combine xforms for
5410 /// ISD::INSERT_VECTOR_ELT.
5411 static SDValue PerformInsertEltCombine(SDNode *N,
5412 TargetLowering::DAGCombinerInfo &DCI) {
5413 // Bitcast an i64 load inserted into a vector to f64.
5414 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5415 EVT VT = N->getValueType(0);
5416 SDNode *Elt = N->getOperand(1).getNode();
5417 if (VT.getVectorElementType() != MVT::i64 ||
5418 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
5421 SelectionDAG &DAG = DCI.DAG;
5422 DebugLoc dl = N->getDebugLoc();
5423 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5424 VT.getVectorNumElements());
5425 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
5426 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
5427 // Make the DAGCombiner fold the bitcasts.
5428 DCI.AddToWorklist(Vec.getNode());
5429 DCI.AddToWorklist(V.getNode());
5430 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
5431 Vec, V, N->getOperand(2));
5432 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
5435 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
5436 /// ISD::VECTOR_SHUFFLE.
5437 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
5438 // The LLVM shufflevector instruction does not require the shuffle mask
5439 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
5440 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
5441 // operands do not match the mask length, they are extended by concatenating
5442 // them with undef vectors. That is probably the right thing for other
5443 // targets, but for NEON it is better to concatenate two double-register
5444 // size vector operands into a single quad-register size vector. Do that
5445 // transformation here:
5446 // shuffle(concat(v1, undef), concat(v2, undef)) ->
5447 // shuffle(concat(v1, v2), undef)
5448 SDValue Op0 = N->getOperand(0);
5449 SDValue Op1 = N->getOperand(1);
5450 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
5451 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
5452 Op0.getNumOperands() != 2 ||
5453 Op1.getNumOperands() != 2)
5455 SDValue Concat0Op1 = Op0.getOperand(1);
5456 SDValue Concat1Op1 = Op1.getOperand(1);
5457 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
5458 Concat1Op1.getOpcode() != ISD::UNDEF)
5460 // Skip the transformation if any of the types are illegal.
5461 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5462 EVT VT = N->getValueType(0);
5463 if (!TLI.isTypeLegal(VT) ||
5464 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
5465 !TLI.isTypeLegal(Concat1Op1.getValueType()))
5468 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5469 Op0.getOperand(0), Op1.getOperand(0));
5470 // Translate the shuffle mask.
5471 SmallVector<int, 16> NewMask;
5472 unsigned NumElts = VT.getVectorNumElements();
5473 unsigned HalfElts = NumElts/2;
5474 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
5475 for (unsigned n = 0; n < NumElts; ++n) {
5476 int MaskElt = SVN->getMaskElt(n);
5478 if (MaskElt < (int)HalfElts)
5480 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
5481 NewElt = HalfElts + MaskElt - NumElts;
5482 NewMask.push_back(NewElt);
5484 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
5485 DAG.getUNDEF(VT), NewMask.data());
5488 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
5489 /// NEON load/store intrinsics to merge base address updates.
5490 static SDValue CombineBaseUpdate(SDNode *N,
5491 TargetLowering::DAGCombinerInfo &DCI) {
5492 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5495 SelectionDAG &DAG = DCI.DAG;
5496 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
5497 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
5498 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
5499 SDValue Addr = N->getOperand(AddrOpIdx);
5501 // Search for a use of the address operand that is an increment.
5502 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
5503 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
5505 if (User->getOpcode() != ISD::ADD ||
5506 UI.getUse().getResNo() != Addr.getResNo())
5509 // Check that the add is independent of the load/store. Otherwise, folding
5510 // it would create a cycle.
5511 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
5514 // Find the new opcode for the updating load/store.
5516 bool isLaneOp = false;
5517 unsigned NewOpc = 0;
5518 unsigned NumVecs = 0;
5520 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
5522 default: assert(0 && "unexpected intrinsic for Neon base update");
5523 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
5525 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
5527 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
5529 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
5531 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
5532 NumVecs = 2; isLaneOp = true; break;
5533 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
5534 NumVecs = 3; isLaneOp = true; break;
5535 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
5536 NumVecs = 4; isLaneOp = true; break;
5537 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
5538 NumVecs = 1; isLoad = false; break;
5539 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
5540 NumVecs = 2; isLoad = false; break;
5541 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
5542 NumVecs = 3; isLoad = false; break;
5543 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
5544 NumVecs = 4; isLoad = false; break;
5545 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
5546 NumVecs = 2; isLoad = false; isLaneOp = true; break;
5547 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
5548 NumVecs = 3; isLoad = false; isLaneOp = true; break;
5549 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
5550 NumVecs = 4; isLoad = false; isLaneOp = true; break;
5554 switch (N->getOpcode()) {
5555 default: assert(0 && "unexpected opcode for Neon base update");
5556 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
5557 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
5558 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
5562 // Find the size of memory referenced by the load/store.
5565 VecTy = N->getValueType(0);
5567 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
5568 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
5570 NumBytes /= VecTy.getVectorNumElements();
5572 // If the increment is a constant, it must match the memory ref size.
5573 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
5574 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
5575 uint64_t IncVal = CInc->getZExtValue();
5576 if (IncVal != NumBytes)
5578 } else if (NumBytes >= 3 * 16) {
5579 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
5580 // separate instructions that make it harder to use a non-constant update.
5584 // Create the new updating load/store node.
5586 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
5588 for (n = 0; n < NumResultVecs; ++n)
5590 Tys[n++] = MVT::i32;
5591 Tys[n] = MVT::Other;
5592 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
5593 SmallVector<SDValue, 8> Ops;
5594 Ops.push_back(N->getOperand(0)); // incoming chain
5595 Ops.push_back(N->getOperand(AddrOpIdx));
5597 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
5598 Ops.push_back(N->getOperand(i));
5600 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
5601 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
5602 Ops.data(), Ops.size(),
5603 MemInt->getMemoryVT(),
5604 MemInt->getMemOperand());
5607 std::vector<SDValue> NewResults;
5608 for (unsigned i = 0; i < NumResultVecs; ++i) {
5609 NewResults.push_back(SDValue(UpdN.getNode(), i));
5611 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
5612 DCI.CombineTo(N, NewResults);
5613 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
5620 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
5621 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
5622 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
5624 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
5625 SelectionDAG &DAG = DCI.DAG;
5626 EVT VT = N->getValueType(0);
5627 // vldN-dup instructions only support 64-bit vectors for N > 1.
5628 if (!VT.is64BitVector())
5631 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
5632 SDNode *VLD = N->getOperand(0).getNode();
5633 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
5635 unsigned NumVecs = 0;
5636 unsigned NewOpc = 0;
5637 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
5638 if (IntNo == Intrinsic::arm_neon_vld2lane) {
5640 NewOpc = ARMISD::VLD2DUP;
5641 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
5643 NewOpc = ARMISD::VLD3DUP;
5644 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
5646 NewOpc = ARMISD::VLD4DUP;
5651 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
5652 // numbers match the load.
5653 unsigned VLDLaneNo =
5654 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
5655 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5657 // Ignore uses of the chain result.
5658 if (UI.getUse().getResNo() == NumVecs)
5661 if (User->getOpcode() != ARMISD::VDUPLANE ||
5662 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
5666 // Create the vldN-dup node.
5669 for (n = 0; n < NumVecs; ++n)
5671 Tys[n] = MVT::Other;
5672 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
5673 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
5674 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
5675 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
5676 Ops, 2, VLDMemInt->getMemoryVT(),
5677 VLDMemInt->getMemOperand());
5680 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5682 unsigned ResNo = UI.getUse().getResNo();
5683 // Ignore uses of the chain result.
5684 if (ResNo == NumVecs)
5687 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
5690 // Now the vldN-lane intrinsic is dead except for its chain result.
5691 // Update uses of the chain.
5692 std::vector<SDValue> VLDDupResults;
5693 for (unsigned n = 0; n < NumVecs; ++n)
5694 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
5695 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
5696 DCI.CombineTo(VLD, VLDDupResults);
5701 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
5702 /// ARMISD::VDUPLANE.
5703 static SDValue PerformVDUPLANECombine(SDNode *N,
5704 TargetLowering::DAGCombinerInfo &DCI) {
5705 SDValue Op = N->getOperand(0);
5707 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
5708 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
5709 if (CombineVLDDUP(N, DCI))
5710 return SDValue(N, 0);
5712 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
5713 // redundant. Ignore bit_converts for now; element sizes are checked below.
5714 while (Op.getOpcode() == ISD::BITCAST)
5715 Op = Op.getOperand(0);
5716 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
5719 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
5720 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
5721 // The canonical VMOV for a zero vector uses a 32-bit element size.
5722 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5724 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
5726 EVT VT = N->getValueType(0);
5727 if (EltSize > VT.getVectorElementType().getSizeInBits())
5730 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
5733 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5734 /// operand of a vector shift operation, where all the elements of the
5735 /// build_vector must have the same constant integer value.
5736 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5737 // Ignore bit_converts.
5738 while (Op.getOpcode() == ISD::BITCAST)
5739 Op = Op.getOperand(0);
5740 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5741 APInt SplatBits, SplatUndef;
5742 unsigned SplatBitSize;
5744 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5745 HasAnyUndefs, ElementBits) ||
5746 SplatBitSize > ElementBits)
5748 Cnt = SplatBits.getSExtValue();
5752 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5753 /// operand of a vector shift left operation. That value must be in the range:
5754 /// 0 <= Value < ElementBits for a left shift; or
5755 /// 0 <= Value <= ElementBits for a long left shift.
5756 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5757 assert(VT.isVector() && "vector shift count is not a vector type");
5758 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5759 if (! getVShiftImm(Op, ElementBits, Cnt))
5761 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
5764 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5765 /// operand of a vector shift right operation. For a shift opcode, the value
5766 /// is positive, but for an intrinsic the value count must be negative. The
5767 /// absolute value must be in the range:
5768 /// 1 <= |Value| <= ElementBits for a right shift; or
5769 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5770 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5772 assert(VT.isVector() && "vector shift count is not a vector type");
5773 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5774 if (! getVShiftImm(Op, ElementBits, Cnt))
5778 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
5781 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
5782 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
5783 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5786 // Don't do anything for most intrinsics.
5789 // Vector shifts: check for immediate versions and lower them.
5790 // Note: This is done during DAG combining instead of DAG legalizing because
5791 // the build_vectors for 64-bit vector element shift counts are generally
5792 // not legal, and it is hard to see their values after they get legalized to
5793 // loads from a constant pool.
5794 case Intrinsic::arm_neon_vshifts:
5795 case Intrinsic::arm_neon_vshiftu:
5796 case Intrinsic::arm_neon_vshiftls:
5797 case Intrinsic::arm_neon_vshiftlu:
5798 case Intrinsic::arm_neon_vshiftn:
5799 case Intrinsic::arm_neon_vrshifts:
5800 case Intrinsic::arm_neon_vrshiftu:
5801 case Intrinsic::arm_neon_vrshiftn:
5802 case Intrinsic::arm_neon_vqshifts:
5803 case Intrinsic::arm_neon_vqshiftu:
5804 case Intrinsic::arm_neon_vqshiftsu:
5805 case Intrinsic::arm_neon_vqshiftns:
5806 case Intrinsic::arm_neon_vqshiftnu:
5807 case Intrinsic::arm_neon_vqshiftnsu:
5808 case Intrinsic::arm_neon_vqrshiftns:
5809 case Intrinsic::arm_neon_vqrshiftnu:
5810 case Intrinsic::arm_neon_vqrshiftnsu: {
5811 EVT VT = N->getOperand(1).getValueType();
5813 unsigned VShiftOpc = 0;
5816 case Intrinsic::arm_neon_vshifts:
5817 case Intrinsic::arm_neon_vshiftu:
5818 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
5819 VShiftOpc = ARMISD::VSHL;
5822 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
5823 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
5824 ARMISD::VSHRs : ARMISD::VSHRu);
5829 case Intrinsic::arm_neon_vshiftls:
5830 case Intrinsic::arm_neon_vshiftlu:
5831 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
5833 llvm_unreachable("invalid shift count for vshll intrinsic");
5835 case Intrinsic::arm_neon_vrshifts:
5836 case Intrinsic::arm_neon_vrshiftu:
5837 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
5841 case Intrinsic::arm_neon_vqshifts:
5842 case Intrinsic::arm_neon_vqshiftu:
5843 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5847 case Intrinsic::arm_neon_vqshiftsu:
5848 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5850 llvm_unreachable("invalid shift count for vqshlu intrinsic");
5852 case Intrinsic::arm_neon_vshiftn:
5853 case Intrinsic::arm_neon_vrshiftn:
5854 case Intrinsic::arm_neon_vqshiftns:
5855 case Intrinsic::arm_neon_vqshiftnu:
5856 case Intrinsic::arm_neon_vqshiftnsu:
5857 case Intrinsic::arm_neon_vqrshiftns:
5858 case Intrinsic::arm_neon_vqrshiftnu:
5859 case Intrinsic::arm_neon_vqrshiftnsu:
5860 // Narrowing shifts require an immediate right shift.
5861 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
5863 llvm_unreachable("invalid shift count for narrowing vector shift "
5867 llvm_unreachable("unhandled vector shift");
5871 case Intrinsic::arm_neon_vshifts:
5872 case Intrinsic::arm_neon_vshiftu:
5873 // Opcode already set above.
5875 case Intrinsic::arm_neon_vshiftls:
5876 case Intrinsic::arm_neon_vshiftlu:
5877 if (Cnt == VT.getVectorElementType().getSizeInBits())
5878 VShiftOpc = ARMISD::VSHLLi;
5880 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
5881 ARMISD::VSHLLs : ARMISD::VSHLLu);
5883 case Intrinsic::arm_neon_vshiftn:
5884 VShiftOpc = ARMISD::VSHRN; break;
5885 case Intrinsic::arm_neon_vrshifts:
5886 VShiftOpc = ARMISD::VRSHRs; break;
5887 case Intrinsic::arm_neon_vrshiftu:
5888 VShiftOpc = ARMISD::VRSHRu; break;
5889 case Intrinsic::arm_neon_vrshiftn:
5890 VShiftOpc = ARMISD::VRSHRN; break;
5891 case Intrinsic::arm_neon_vqshifts:
5892 VShiftOpc = ARMISD::VQSHLs; break;
5893 case Intrinsic::arm_neon_vqshiftu:
5894 VShiftOpc = ARMISD::VQSHLu; break;
5895 case Intrinsic::arm_neon_vqshiftsu:
5896 VShiftOpc = ARMISD::VQSHLsu; break;
5897 case Intrinsic::arm_neon_vqshiftns:
5898 VShiftOpc = ARMISD::VQSHRNs; break;
5899 case Intrinsic::arm_neon_vqshiftnu:
5900 VShiftOpc = ARMISD::VQSHRNu; break;
5901 case Intrinsic::arm_neon_vqshiftnsu:
5902 VShiftOpc = ARMISD::VQSHRNsu; break;
5903 case Intrinsic::arm_neon_vqrshiftns:
5904 VShiftOpc = ARMISD::VQRSHRNs; break;
5905 case Intrinsic::arm_neon_vqrshiftnu:
5906 VShiftOpc = ARMISD::VQRSHRNu; break;
5907 case Intrinsic::arm_neon_vqrshiftnsu:
5908 VShiftOpc = ARMISD::VQRSHRNsu; break;
5911 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5912 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
5915 case Intrinsic::arm_neon_vshiftins: {
5916 EVT VT = N->getOperand(1).getValueType();
5918 unsigned VShiftOpc = 0;
5920 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
5921 VShiftOpc = ARMISD::VSLI;
5922 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
5923 VShiftOpc = ARMISD::VSRI;
5925 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
5928 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5929 N->getOperand(1), N->getOperand(2),
5930 DAG.getConstant(Cnt, MVT::i32));
5933 case Intrinsic::arm_neon_vqrshifts:
5934 case Intrinsic::arm_neon_vqrshiftu:
5935 // No immediate versions of these to check for.
5942 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
5943 /// lowers them. As with the vector shift intrinsics, this is done during DAG
5944 /// combining instead of DAG legalizing because the build_vectors for 64-bit
5945 /// vector element shift counts are generally not legal, and it is hard to see
5946 /// their values after they get legalized to loads from a constant pool.
5947 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
5948 const ARMSubtarget *ST) {
5949 EVT VT = N->getValueType(0);
5951 // Nothing to be done for scalar shifts.
5952 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5953 if (!VT.isVector() || !TLI.isTypeLegal(VT))
5956 assert(ST->hasNEON() && "unexpected vector shift");
5959 switch (N->getOpcode()) {
5960 default: llvm_unreachable("unexpected shift opcode");
5963 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
5964 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
5965 DAG.getConstant(Cnt, MVT::i32));
5970 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
5971 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
5972 ARMISD::VSHRs : ARMISD::VSHRu);
5973 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
5974 DAG.getConstant(Cnt, MVT::i32));
5980 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
5981 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
5982 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
5983 const ARMSubtarget *ST) {
5984 SDValue N0 = N->getOperand(0);
5986 // Check for sign- and zero-extensions of vector extract operations of 8-
5987 // and 16-bit vector elements. NEON supports these directly. They are
5988 // handled during DAG combining because type legalization will promote them
5989 // to 32-bit types and it is messy to recognize the operations after that.
5990 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
5991 SDValue Vec = N0.getOperand(0);
5992 SDValue Lane = N0.getOperand(1);
5993 EVT VT = N->getValueType(0);
5994 EVT EltVT = N0.getValueType();
5995 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5997 if (VT == MVT::i32 &&
5998 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
5999 TLI.isTypeLegal(Vec.getValueType()) &&
6000 isa<ConstantSDNode>(Lane)) {
6003 switch (N->getOpcode()) {
6004 default: llvm_unreachable("unexpected opcode");
6005 case ISD::SIGN_EXTEND:
6006 Opc = ARMISD::VGETLANEs;
6008 case ISD::ZERO_EXTEND:
6009 case ISD::ANY_EXTEND:
6010 Opc = ARMISD::VGETLANEu;
6013 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
6020 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
6021 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
6022 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
6023 const ARMSubtarget *ST) {
6024 // If the target supports NEON, try to use vmax/vmin instructions for f32
6025 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
6026 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
6027 // a NaN; only do the transformation when it matches that behavior.
6029 // For now only do this when using NEON for FP operations; if using VFP, it
6030 // is not obvious that the benefit outweighs the cost of switching to the
6032 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
6033 N->getValueType(0) != MVT::f32)
6036 SDValue CondLHS = N->getOperand(0);
6037 SDValue CondRHS = N->getOperand(1);
6038 SDValue LHS = N->getOperand(2);
6039 SDValue RHS = N->getOperand(3);
6040 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
6042 unsigned Opcode = 0;
6044 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
6045 IsReversed = false; // x CC y ? x : y
6046 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
6047 IsReversed = true ; // x CC y ? y : x
6061 // If LHS is NaN, an ordered comparison will be false and the result will
6062 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
6063 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6064 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
6065 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6067 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
6068 // will return -0, so vmin can only be used for unsafe math or if one of
6069 // the operands is known to be nonzero.
6070 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
6072 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6074 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
6083 // If LHS is NaN, an ordered comparison will be false and the result will
6084 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
6085 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6086 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
6087 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6089 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
6090 // will return +0, so vmax can only be used for unsafe math or if one of
6091 // the operands is known to be nonzero.
6092 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
6094 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6096 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
6102 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
6105 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
6106 DAGCombinerInfo &DCI) const {
6107 switch (N->getOpcode()) {
6109 case ISD::ADD: return PerformADDCombine(N, DCI);
6110 case ISD::SUB: return PerformSUBCombine(N, DCI);
6111 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
6112 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
6113 case ISD::AND: return PerformANDCombine(N, DCI);
6114 case ARMISD::BFI: return PerformBFICombine(N, DCI);
6115 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
6116 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
6117 case ISD::STORE: return PerformSTORECombine(N, DCI);
6118 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
6119 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
6120 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
6121 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
6122 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
6125 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
6126 case ISD::SIGN_EXTEND:
6127 case ISD::ZERO_EXTEND:
6128 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
6129 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
6130 case ARMISD::VLD2DUP:
6131 case ARMISD::VLD3DUP:
6132 case ARMISD::VLD4DUP:
6133 return CombineBaseUpdate(N, DCI);
6134 case ISD::INTRINSIC_VOID:
6135 case ISD::INTRINSIC_W_CHAIN:
6136 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
6137 case Intrinsic::arm_neon_vld1:
6138 case Intrinsic::arm_neon_vld2:
6139 case Intrinsic::arm_neon_vld3:
6140 case Intrinsic::arm_neon_vld4:
6141 case Intrinsic::arm_neon_vld2lane:
6142 case Intrinsic::arm_neon_vld3lane:
6143 case Intrinsic::arm_neon_vld4lane:
6144 case Intrinsic::arm_neon_vst1:
6145 case Intrinsic::arm_neon_vst2:
6146 case Intrinsic::arm_neon_vst3:
6147 case Intrinsic::arm_neon_vst4:
6148 case Intrinsic::arm_neon_vst2lane:
6149 case Intrinsic::arm_neon_vst3lane:
6150 case Intrinsic::arm_neon_vst4lane:
6151 return CombineBaseUpdate(N, DCI);
6159 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
6161 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
6164 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
6165 if (!Subtarget->allowsUnalignedMem())
6168 switch (VT.getSimpleVT().SimpleTy) {
6175 // FIXME: VLD1 etc with standard alignment is legal.
6179 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
6184 switch (VT.getSimpleVT().SimpleTy) {
6185 default: return false;
6200 if ((V & (Scale - 1)) != 0)
6203 return V == (V & ((1LL << 5) - 1));
6206 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
6207 const ARMSubtarget *Subtarget) {
6214 switch (VT.getSimpleVT().SimpleTy) {
6215 default: return false;
6220 // + imm12 or - imm8
6222 return V == (V & ((1LL << 8) - 1));
6223 return V == (V & ((1LL << 12) - 1));
6226 // Same as ARM mode. FIXME: NEON?
6227 if (!Subtarget->hasVFP2())
6232 return V == (V & ((1LL << 8) - 1));
6236 /// isLegalAddressImmediate - Return true if the integer value can be used
6237 /// as the offset of the target addressing mode for load / store of the
6239 static bool isLegalAddressImmediate(int64_t V, EVT VT,
6240 const ARMSubtarget *Subtarget) {
6247 if (Subtarget->isThumb1Only())
6248 return isLegalT1AddressImmediate(V, VT);
6249 else if (Subtarget->isThumb2())
6250 return isLegalT2AddressImmediate(V, VT, Subtarget);
6255 switch (VT.getSimpleVT().SimpleTy) {
6256 default: return false;
6261 return V == (V & ((1LL << 12) - 1));
6264 return V == (V & ((1LL << 8) - 1));
6267 if (!Subtarget->hasVFP2()) // FIXME: NEON?
6272 return V == (V & ((1LL << 8) - 1));
6276 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
6278 int Scale = AM.Scale;
6282 switch (VT.getSimpleVT().SimpleTy) {
6283 default: return false;
6292 return Scale == 2 || Scale == 4 || Scale == 8;
6295 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6299 // Note, we allow "void" uses (basically, uses that aren't loads or
6300 // stores), because arm allows folding a scale into many arithmetic
6301 // operations. This should be made more precise and revisited later.
6303 // Allow r << imm, but the imm has to be a multiple of two.
6304 if (Scale & 1) return false;
6305 return isPowerOf2_32(Scale);
6309 /// isLegalAddressingMode - Return true if the addressing mode represented
6310 /// by AM is legal for this target, for a load/store of the specified type.
6311 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6312 const Type *Ty) const {
6313 EVT VT = getValueType(Ty, true);
6314 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
6317 // Can never fold addr of global into load/store.
6322 case 0: // no scale reg, must be "r+i" or "r", or "i".
6325 if (Subtarget->isThumb1Only())
6329 // ARM doesn't support any R+R*scale+imm addr modes.
6336 if (Subtarget->isThumb2())
6337 return isLegalT2ScaledAddressingMode(AM, VT);
6339 int Scale = AM.Scale;
6340 switch (VT.getSimpleVT().SimpleTy) {
6341 default: return false;
6345 if (Scale < 0) Scale = -Scale;
6349 return isPowerOf2_32(Scale & ~1);
6353 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
6358 // Note, we allow "void" uses (basically, uses that aren't loads or
6359 // stores), because arm allows folding a scale into many arithmetic
6360 // operations. This should be made more precise and revisited later.
6362 // Allow r << imm, but the imm has to be a multiple of two.
6363 if (Scale & 1) return false;
6364 return isPowerOf2_32(Scale);
6371 /// isLegalICmpImmediate - Return true if the specified immediate is legal
6372 /// icmp immediate, that is the target has icmp instructions which can compare
6373 /// a register against the immediate without having to materialize the
6374 /// immediate into a register.
6375 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
6376 if (!Subtarget->isThumb())
6377 return ARM_AM::getSOImmVal(Imm) != -1;
6378 if (Subtarget->isThumb2())
6379 return ARM_AM::getT2SOImmVal(Imm) != -1;
6380 return Imm >= 0 && Imm <= 255;
6383 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
6384 bool isSEXTLoad, SDValue &Base,
6385 SDValue &Offset, bool &isInc,
6386 SelectionDAG &DAG) {
6387 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6390 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
6392 Base = Ptr->getOperand(0);
6393 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6394 int RHSC = (int)RHS->getZExtValue();
6395 if (RHSC < 0 && RHSC > -256) {
6396 assert(Ptr->getOpcode() == ISD::ADD);
6398 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6402 isInc = (Ptr->getOpcode() == ISD::ADD);
6403 Offset = Ptr->getOperand(1);
6405 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
6407 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6408 int RHSC = (int)RHS->getZExtValue();
6409 if (RHSC < 0 && RHSC > -0x1000) {
6410 assert(Ptr->getOpcode() == ISD::ADD);
6412 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6413 Base = Ptr->getOperand(0);
6418 if (Ptr->getOpcode() == ISD::ADD) {
6420 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
6421 if (ShOpcVal != ARM_AM::no_shift) {
6422 Base = Ptr->getOperand(1);
6423 Offset = Ptr->getOperand(0);
6425 Base = Ptr->getOperand(0);
6426 Offset = Ptr->getOperand(1);
6431 isInc = (Ptr->getOpcode() == ISD::ADD);
6432 Base = Ptr->getOperand(0);
6433 Offset = Ptr->getOperand(1);
6437 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
6441 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
6442 bool isSEXTLoad, SDValue &Base,
6443 SDValue &Offset, bool &isInc,
6444 SelectionDAG &DAG) {
6445 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6448 Base = Ptr->getOperand(0);
6449 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6450 int RHSC = (int)RHS->getZExtValue();
6451 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
6452 assert(Ptr->getOpcode() == ISD::ADD);
6454 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6456 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
6457 isInc = Ptr->getOpcode() == ISD::ADD;
6458 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
6466 /// getPreIndexedAddressParts - returns true by value, base pointer and
6467 /// offset pointer and addressing mode by reference if the node's address
6468 /// can be legally represented as pre-indexed load / store address.
6470 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
6472 ISD::MemIndexedMode &AM,
6473 SelectionDAG &DAG) const {
6474 if (Subtarget->isThumb1Only())
6479 bool isSEXTLoad = false;
6480 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6481 Ptr = LD->getBasePtr();
6482 VT = LD->getMemoryVT();
6483 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6484 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6485 Ptr = ST->getBasePtr();
6486 VT = ST->getMemoryVT();
6491 bool isLegal = false;
6492 if (Subtarget->isThumb2())
6493 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6494 Offset, isInc, DAG);
6496 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6497 Offset, isInc, DAG);
6501 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
6505 /// getPostIndexedAddressParts - returns true by value, base pointer and
6506 /// offset pointer and addressing mode by reference if this node can be
6507 /// combined with a load / store to form a post-indexed load / store.
6508 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
6511 ISD::MemIndexedMode &AM,
6512 SelectionDAG &DAG) const {
6513 if (Subtarget->isThumb1Only())
6518 bool isSEXTLoad = false;
6519 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6520 VT = LD->getMemoryVT();
6521 Ptr = LD->getBasePtr();
6522 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6523 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6524 VT = ST->getMemoryVT();
6525 Ptr = ST->getBasePtr();
6530 bool isLegal = false;
6531 if (Subtarget->isThumb2())
6532 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6535 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6541 // Swap base ptr and offset to catch more post-index load / store when
6542 // it's legal. In Thumb2 mode, offset must be an immediate.
6543 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
6544 !Subtarget->isThumb2())
6545 std::swap(Base, Offset);
6547 // Post-indexed load / store update the base pointer.
6552 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
6556 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6560 const SelectionDAG &DAG,
6561 unsigned Depth) const {
6562 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
6563 switch (Op.getOpcode()) {
6565 case ARMISD::CMOV: {
6566 // Bits are known zero/one if known on the LHS and RHS.
6567 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
6568 if (KnownZero == 0 && KnownOne == 0) return;
6570 APInt KnownZeroRHS, KnownOneRHS;
6571 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
6572 KnownZeroRHS, KnownOneRHS, Depth+1);
6573 KnownZero &= KnownZeroRHS;
6574 KnownOne &= KnownOneRHS;
6580 //===----------------------------------------------------------------------===//
6581 // ARM Inline Assembly Support
6582 //===----------------------------------------------------------------------===//
6584 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
6585 // Looking for "rev" which is V6+.
6586 if (!Subtarget->hasV6Ops())
6589 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
6590 std::string AsmStr = IA->getAsmString();
6591 SmallVector<StringRef, 4> AsmPieces;
6592 SplitString(AsmStr, AsmPieces, ";\n");
6594 switch (AsmPieces.size()) {
6595 default: return false;
6597 AsmStr = AsmPieces[0];
6599 SplitString(AsmStr, AsmPieces, " \t,");
6602 if (AsmPieces.size() == 3 &&
6603 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
6604 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
6605 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
6606 if (Ty && Ty->getBitWidth() == 32)
6607 return IntrinsicLowering::LowerToByteSwap(CI);
6615 /// getConstraintType - Given a constraint letter, return the type of
6616 /// constraint it is for this target.
6617 ARMTargetLowering::ConstraintType
6618 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
6619 if (Constraint.size() == 1) {
6620 switch (Constraint[0]) {
6622 case 'l': return C_RegisterClass;
6623 case 'w': return C_RegisterClass;
6626 return TargetLowering::getConstraintType(Constraint);
6629 /// Examine constraint type and operand type and determine a weight value.
6630 /// This object must already have been set up with the operand type
6631 /// and the current alternative constraint selected.
6632 TargetLowering::ConstraintWeight
6633 ARMTargetLowering::getSingleConstraintMatchWeight(
6634 AsmOperandInfo &info, const char *constraint) const {
6635 ConstraintWeight weight = CW_Invalid;
6636 Value *CallOperandVal = info.CallOperandVal;
6637 // If we don't have a value, we can't do a match,
6638 // but allow it at the lowest weight.
6639 if (CallOperandVal == NULL)
6641 const Type *type = CallOperandVal->getType();
6642 // Look at the constraint type.
6643 switch (*constraint) {
6645 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6648 if (type->isIntegerTy()) {
6649 if (Subtarget->isThumb())
6650 weight = CW_SpecificReg;
6652 weight = CW_Register;
6656 if (type->isFloatingPointTy())
6657 weight = CW_Register;
6663 std::pair<unsigned, const TargetRegisterClass*>
6664 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6666 if (Constraint.size() == 1) {
6667 // GCC ARM Constraint Letters
6668 switch (Constraint[0]) {
6670 if (Subtarget->isThumb())
6671 return std::make_pair(0U, ARM::tGPRRegisterClass);
6673 return std::make_pair(0U, ARM::GPRRegisterClass);
6675 return std::make_pair(0U, ARM::GPRRegisterClass);
6678 return std::make_pair(0U, ARM::SPRRegisterClass);
6679 if (VT.getSizeInBits() == 64)
6680 return std::make_pair(0U, ARM::DPRRegisterClass);
6681 if (VT.getSizeInBits() == 128)
6682 return std::make_pair(0U, ARM::QPRRegisterClass);
6686 if (StringRef("{cc}").equals_lower(Constraint))
6687 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
6689 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6692 std::vector<unsigned> ARMTargetLowering::
6693 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6695 if (Constraint.size() != 1)
6696 return std::vector<unsigned>();
6698 switch (Constraint[0]) { // GCC ARM Constraint Letters
6701 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6702 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6705 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6706 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6707 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
6708 ARM::R12, ARM::LR, 0);
6711 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
6712 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
6713 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
6714 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
6715 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
6716 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
6717 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
6718 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
6719 if (VT.getSizeInBits() == 64)
6720 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
6721 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
6722 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
6723 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
6724 if (VT.getSizeInBits() == 128)
6725 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
6726 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
6730 return std::vector<unsigned>();
6733 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6734 /// vector. If it is invalid, don't add anything to Ops.
6735 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6737 std::vector<SDValue>&Ops,
6738 SelectionDAG &DAG) const {
6739 SDValue Result(0, 0);
6741 switch (Constraint) {
6743 case 'I': case 'J': case 'K': case 'L':
6744 case 'M': case 'N': case 'O':
6745 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
6749 int64_t CVal64 = C->getSExtValue();
6750 int CVal = (int) CVal64;
6751 // None of these constraints allow values larger than 32 bits. Check
6752 // that the value fits in an int.
6756 switch (Constraint) {
6758 if (Subtarget->isThumb1Only()) {
6759 // This must be a constant between 0 and 255, for ADD
6761 if (CVal >= 0 && CVal <= 255)
6763 } else if (Subtarget->isThumb2()) {
6764 // A constant that can be used as an immediate value in a
6765 // data-processing instruction.
6766 if (ARM_AM::getT2SOImmVal(CVal) != -1)
6769 // A constant that can be used as an immediate value in a
6770 // data-processing instruction.
6771 if (ARM_AM::getSOImmVal(CVal) != -1)
6777 if (Subtarget->isThumb()) { // FIXME thumb2
6778 // This must be a constant between -255 and -1, for negated ADD
6779 // immediates. This can be used in GCC with an "n" modifier that
6780 // prints the negated value, for use with SUB instructions. It is
6781 // not useful otherwise but is implemented for compatibility.
6782 if (CVal >= -255 && CVal <= -1)
6785 // This must be a constant between -4095 and 4095. It is not clear
6786 // what this constraint is intended for. Implemented for
6787 // compatibility with GCC.
6788 if (CVal >= -4095 && CVal <= 4095)
6794 if (Subtarget->isThumb1Only()) {
6795 // A 32-bit value where only one byte has a nonzero value. Exclude
6796 // zero to match GCC. This constraint is used by GCC internally for
6797 // constants that can be loaded with a move/shift combination.
6798 // It is not useful otherwise but is implemented for compatibility.
6799 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
6801 } else if (Subtarget->isThumb2()) {
6802 // A constant whose bitwise inverse can be used as an immediate
6803 // value in a data-processing instruction. This can be used in GCC
6804 // with a "B" modifier that prints the inverted value, for use with
6805 // BIC and MVN instructions. It is not useful otherwise but is
6806 // implemented for compatibility.
6807 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
6810 // A constant whose bitwise inverse can be used as an immediate
6811 // value in a data-processing instruction. This can be used in GCC
6812 // with a "B" modifier that prints the inverted value, for use with
6813 // BIC and MVN instructions. It is not useful otherwise but is
6814 // implemented for compatibility.
6815 if (ARM_AM::getSOImmVal(~CVal) != -1)
6821 if (Subtarget->isThumb1Only()) {
6822 // This must be a constant between -7 and 7,
6823 // for 3-operand ADD/SUB immediate instructions.
6824 if (CVal >= -7 && CVal < 7)
6826 } else if (Subtarget->isThumb2()) {
6827 // A constant whose negation can be used as an immediate value in a
6828 // data-processing instruction. This can be used in GCC with an "n"
6829 // modifier that prints the negated value, for use with SUB
6830 // instructions. It is not useful otherwise but is implemented for
6832 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
6835 // A constant whose negation can be used as an immediate value in a
6836 // data-processing instruction. This can be used in GCC with an "n"
6837 // modifier that prints the negated value, for use with SUB
6838 // instructions. It is not useful otherwise but is implemented for
6840 if (ARM_AM::getSOImmVal(-CVal) != -1)
6846 if (Subtarget->isThumb()) { // FIXME thumb2
6847 // This must be a multiple of 4 between 0 and 1020, for
6848 // ADD sp + immediate.
6849 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
6852 // A power of two or a constant between 0 and 32. This is used in
6853 // GCC for the shift amount on shifted register operands, but it is
6854 // useful in general for any shift amounts.
6855 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
6861 if (Subtarget->isThumb()) { // FIXME thumb2
6862 // This must be a constant between 0 and 31, for shift amounts.
6863 if (CVal >= 0 && CVal <= 31)
6869 if (Subtarget->isThumb()) { // FIXME thumb2
6870 // This must be a multiple of 4 between -508 and 508, for
6871 // ADD/SUB sp = sp + immediate.
6872 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
6877 Result = DAG.getTargetConstant(CVal, Op.getValueType());
6881 if (Result.getNode()) {
6882 Ops.push_back(Result);
6885 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6889 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6890 // The ARM target isn't yet aware of offsets.
6894 int ARM::getVFPf32Imm(const APFloat &FPImm) {
6895 APInt Imm = FPImm.bitcastToAPInt();
6896 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
6897 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
6898 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
6900 // We can handle 4 bits of mantissa.
6901 // mantissa = (16+UInt(e:f:g:h))/16.
6902 if (Mantissa & 0x7ffff)
6905 if ((Mantissa & 0xf) != Mantissa)
6908 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6909 if (Exp < -3 || Exp > 4)
6911 Exp = ((Exp+3) & 0x7) ^ 4;
6913 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6916 int ARM::getVFPf64Imm(const APFloat &FPImm) {
6917 APInt Imm = FPImm.bitcastToAPInt();
6918 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
6919 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
6920 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
6922 // We can handle 4 bits of mantissa.
6923 // mantissa = (16+UInt(e:f:g:h))/16.
6924 if (Mantissa & 0xffffffffffffLL)
6927 if ((Mantissa & 0xf) != Mantissa)
6930 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6931 if (Exp < -3 || Exp > 4)
6933 Exp = ((Exp+3) & 0x7) ^ 4;
6935 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6938 bool ARM::isBitFieldInvertedMask(unsigned v) {
6939 if (v == 0xffffffff)
6941 // there can be 1's on either or both "outsides", all the "inside"
6943 unsigned int lsb = 0, msb = 31;
6944 while (v & (1 << msb)) --msb;
6945 while (v & (1 << lsb)) ++lsb;
6946 for (unsigned int i = lsb; i <= msb; ++i) {
6953 /// isFPImmLegal - Returns true if the target can instruction select the
6954 /// specified FP immediate natively. If false, the legalizer will
6955 /// materialize the FP immediate as a load from a constant pool.
6956 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
6957 if (!Subtarget->hasVFP3())
6960 return ARM::getVFPf32Imm(Imm) != -1;
6962 return ARM::getVFPf64Imm(Imm) != -1;
6966 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6967 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6968 /// specified in the intrinsic calls.
6969 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6971 unsigned Intrinsic) const {
6972 switch (Intrinsic) {
6973 case Intrinsic::arm_neon_vld1:
6974 case Intrinsic::arm_neon_vld2:
6975 case Intrinsic::arm_neon_vld3:
6976 case Intrinsic::arm_neon_vld4:
6977 case Intrinsic::arm_neon_vld2lane:
6978 case Intrinsic::arm_neon_vld3lane:
6979 case Intrinsic::arm_neon_vld4lane: {
6980 Info.opc = ISD::INTRINSIC_W_CHAIN;
6981 // Conservatively set memVT to the entire set of vectors loaded.
6982 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
6983 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6984 Info.ptrVal = I.getArgOperand(0);
6986 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6987 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6988 Info.vol = false; // volatile loads with NEON intrinsics not supported
6989 Info.readMem = true;
6990 Info.writeMem = false;
6993 case Intrinsic::arm_neon_vst1:
6994 case Intrinsic::arm_neon_vst2:
6995 case Intrinsic::arm_neon_vst3:
6996 case Intrinsic::arm_neon_vst4:
6997 case Intrinsic::arm_neon_vst2lane:
6998 case Intrinsic::arm_neon_vst3lane:
6999 case Intrinsic::arm_neon_vst4lane: {
7000 Info.opc = ISD::INTRINSIC_VOID;
7001 // Conservatively set memVT to the entire set of vectors stored.
7002 unsigned NumElts = 0;
7003 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
7004 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
7005 if (!ArgTy->isVectorTy())
7007 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
7009 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7010 Info.ptrVal = I.getArgOperand(0);
7012 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7013 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7014 Info.vol = false; // volatile stores with NEON intrinsics not supported
7015 Info.readMem = false;
7016 Info.writeMem = true;