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 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
458 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
460 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
461 setTargetDAGCombine(ISD::SHL);
462 setTargetDAGCombine(ISD::SRL);
463 setTargetDAGCombine(ISD::SRA);
464 setTargetDAGCombine(ISD::SIGN_EXTEND);
465 setTargetDAGCombine(ISD::ZERO_EXTEND);
466 setTargetDAGCombine(ISD::ANY_EXTEND);
467 setTargetDAGCombine(ISD::SELECT_CC);
468 setTargetDAGCombine(ISD::BUILD_VECTOR);
469 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
470 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
471 setTargetDAGCombine(ISD::STORE);
474 computeRegisterProperties();
476 // ARM does not have f32 extending load.
477 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
479 // ARM does not have i1 sign extending load.
480 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
482 // ARM supports all 4 flavors of integer indexed load / store.
483 if (!Subtarget->isThumb1Only()) {
484 for (unsigned im = (unsigned)ISD::PRE_INC;
485 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
486 setIndexedLoadAction(im, MVT::i1, Legal);
487 setIndexedLoadAction(im, MVT::i8, Legal);
488 setIndexedLoadAction(im, MVT::i16, Legal);
489 setIndexedLoadAction(im, MVT::i32, Legal);
490 setIndexedStoreAction(im, MVT::i1, Legal);
491 setIndexedStoreAction(im, MVT::i8, Legal);
492 setIndexedStoreAction(im, MVT::i16, Legal);
493 setIndexedStoreAction(im, MVT::i32, Legal);
497 // i64 operation support.
498 if (Subtarget->isThumb1Only()) {
499 setOperationAction(ISD::MUL, MVT::i64, Expand);
500 setOperationAction(ISD::MULHU, MVT::i32, Expand);
501 setOperationAction(ISD::MULHS, MVT::i32, Expand);
502 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
503 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
505 setOperationAction(ISD::MUL, MVT::i64, Expand);
506 setOperationAction(ISD::MULHU, MVT::i32, Expand);
507 if (!Subtarget->hasV6Ops())
508 setOperationAction(ISD::MULHS, MVT::i32, Expand);
510 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
511 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
512 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
513 setOperationAction(ISD::SRL, MVT::i64, Custom);
514 setOperationAction(ISD::SRA, MVT::i64, Custom);
516 // ARM does not have ROTL.
517 setOperationAction(ISD::ROTL, MVT::i32, Expand);
518 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
519 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
520 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
521 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
523 // Only ARMv6 has BSWAP.
524 if (!Subtarget->hasV6Ops())
525 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
527 // These are expanded into libcalls.
528 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
529 // v7M has a hardware divider
530 setOperationAction(ISD::SDIV, MVT::i32, Expand);
531 setOperationAction(ISD::UDIV, MVT::i32, Expand);
533 setOperationAction(ISD::SREM, MVT::i32, Expand);
534 setOperationAction(ISD::UREM, MVT::i32, Expand);
535 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
536 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
538 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
539 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
540 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
541 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
542 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
544 setOperationAction(ISD::TRAP, MVT::Other, Legal);
546 // Use the default implementation.
547 setOperationAction(ISD::VASTART, MVT::Other, Custom);
548 setOperationAction(ISD::VAARG, MVT::Other, Expand);
549 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
550 setOperationAction(ISD::VAEND, MVT::Other, Expand);
551 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
552 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
553 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
554 // FIXME: Shouldn't need this, since no register is used, but the legalizer
555 // doesn't yet know how to not do that for SjLj.
556 setExceptionSelectorRegister(ARM::R0);
557 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
558 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
559 // the default expansion.
560 if (Subtarget->hasDataBarrier() ||
561 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
562 // membarrier needs custom lowering; the rest are legal and handled
564 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
566 // Set them all for expansion, which will force libcalls.
567 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
568 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
569 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
570 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
571 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
572 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
573 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
584 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
585 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
586 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
587 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
588 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
589 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
590 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
591 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
592 // Since the libcalls include locking, fold in the fences
593 setShouldFoldAtomicFences(true);
595 // 64-bit versions are always libcalls (for now)
596 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
597 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
600 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
601 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
602 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
603 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
605 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
607 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
608 if (!Subtarget->hasV6Ops()) {
609 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
610 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
612 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
614 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
615 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
616 // iff target supports vfp2.
617 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
618 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
621 // We want to custom lower some of our intrinsics.
622 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
623 if (Subtarget->isTargetDarwin()) {
624 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
625 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
626 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
629 setOperationAction(ISD::SETCC, MVT::i32, Expand);
630 setOperationAction(ISD::SETCC, MVT::f32, Expand);
631 setOperationAction(ISD::SETCC, MVT::f64, Expand);
632 setOperationAction(ISD::SELECT, MVT::i32, Custom);
633 setOperationAction(ISD::SELECT, MVT::f32, Custom);
634 setOperationAction(ISD::SELECT, MVT::f64, Custom);
635 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
636 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
637 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
639 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
640 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
641 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
642 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
643 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
645 // We don't support sin/cos/fmod/copysign/pow
646 setOperationAction(ISD::FSIN, MVT::f64, Expand);
647 setOperationAction(ISD::FSIN, MVT::f32, Expand);
648 setOperationAction(ISD::FCOS, MVT::f32, Expand);
649 setOperationAction(ISD::FCOS, MVT::f64, Expand);
650 setOperationAction(ISD::FREM, MVT::f64, Expand);
651 setOperationAction(ISD::FREM, MVT::f32, Expand);
652 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
653 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
654 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
656 setOperationAction(ISD::FPOW, MVT::f64, Expand);
657 setOperationAction(ISD::FPOW, MVT::f32, Expand);
659 // Various VFP goodness
660 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
661 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
662 if (Subtarget->hasVFP2()) {
663 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
664 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
665 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
666 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
668 // Special handling for half-precision FP.
669 if (!Subtarget->hasFP16()) {
670 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
671 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
675 // We have target-specific dag combine patterns for the following nodes:
676 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
677 setTargetDAGCombine(ISD::ADD);
678 setTargetDAGCombine(ISD::SUB);
679 setTargetDAGCombine(ISD::MUL);
681 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
682 setTargetDAGCombine(ISD::OR);
683 if (Subtarget->hasNEON())
684 setTargetDAGCombine(ISD::AND);
686 setStackPointerRegisterToSaveRestore(ARM::SP);
688 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
689 setSchedulingPreference(Sched::RegPressure);
691 setSchedulingPreference(Sched::Hybrid);
693 //// temporary - rewrite interface to use type
694 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
696 // On ARM arguments smaller than 4 bytes are extended, so all arguments
697 // are at least 4 bytes aligned.
698 setMinStackArgumentAlignment(4);
700 benefitFromCodePlacementOpt = true;
703 // FIXME: It might make sense to define the representative register class as the
704 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
705 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
706 // SPR's representative would be DPR_VFP2. This should work well if register
707 // pressure tracking were modified such that a register use would increment the
708 // pressure of the register class's representative and all of it's super
709 // classes' representatives transitively. We have not implemented this because
710 // of the difficulty prior to coalescing of modeling operand register classes
711 // due to the common occurence of cross class copies and subregister insertions
713 std::pair<const TargetRegisterClass*, uint8_t>
714 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
715 const TargetRegisterClass *RRC = 0;
717 switch (VT.getSimpleVT().SimpleTy) {
719 return TargetLowering::findRepresentativeClass(VT);
720 // Use DPR as representative register class for all floating point
721 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
722 // the cost is 1 for both f32 and f64.
723 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
724 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
725 RRC = ARM::DPRRegisterClass;
726 // When NEON is used for SP, only half of the register file is available
727 // because operations that define both SP and DP results will be constrained
728 // to the VFP2 class (D0-D15). We currently model this constraint prior to
729 // coalescing by double-counting the SP regs. See the FIXME above.
730 if (Subtarget->useNEONForSinglePrecisionFP())
733 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
734 case MVT::v4f32: case MVT::v2f64:
735 RRC = ARM::DPRRegisterClass;
739 RRC = ARM::DPRRegisterClass;
743 RRC = ARM::DPRRegisterClass;
747 return std::make_pair(RRC, Cost);
750 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
753 case ARMISD::Wrapper: return "ARMISD::Wrapper";
754 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
755 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
756 case ARMISD::CALL: return "ARMISD::CALL";
757 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
758 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
759 case ARMISD::tCALL: return "ARMISD::tCALL";
760 case ARMISD::BRCOND: return "ARMISD::BRCOND";
761 case ARMISD::BR_JT: return "ARMISD::BR_JT";
762 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
763 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
764 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
765 case ARMISD::CMP: return "ARMISD::CMP";
766 case ARMISD::CMPZ: return "ARMISD::CMPZ";
767 case ARMISD::CMPFP: return "ARMISD::CMPFP";
768 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
769 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
770 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
771 case ARMISD::CMOV: return "ARMISD::CMOV";
772 case ARMISD::CNEG: return "ARMISD::CNEG";
774 case ARMISD::RBIT: return "ARMISD::RBIT";
776 case ARMISD::FTOSI: return "ARMISD::FTOSI";
777 case ARMISD::FTOUI: return "ARMISD::FTOUI";
778 case ARMISD::SITOF: return "ARMISD::SITOF";
779 case ARMISD::UITOF: return "ARMISD::UITOF";
781 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
782 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
783 case ARMISD::RRX: return "ARMISD::RRX";
785 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
786 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
788 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
789 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
790 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
792 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
794 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
796 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
798 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
799 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
801 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
803 case ARMISD::VCEQ: return "ARMISD::VCEQ";
804 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
805 case ARMISD::VCGE: return "ARMISD::VCGE";
806 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
807 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
808 case ARMISD::VCGEU: return "ARMISD::VCGEU";
809 case ARMISD::VCGT: return "ARMISD::VCGT";
810 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
811 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
812 case ARMISD::VCGTU: return "ARMISD::VCGTU";
813 case ARMISD::VTST: return "ARMISD::VTST";
815 case ARMISD::VSHL: return "ARMISD::VSHL";
816 case ARMISD::VSHRs: return "ARMISD::VSHRs";
817 case ARMISD::VSHRu: return "ARMISD::VSHRu";
818 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
819 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
820 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
821 case ARMISD::VSHRN: return "ARMISD::VSHRN";
822 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
823 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
824 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
825 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
826 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
827 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
828 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
829 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
830 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
831 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
832 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
833 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
834 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
835 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
836 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
837 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
838 case ARMISD::VDUP: return "ARMISD::VDUP";
839 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
840 case ARMISD::VEXT: return "ARMISD::VEXT";
841 case ARMISD::VREV64: return "ARMISD::VREV64";
842 case ARMISD::VREV32: return "ARMISD::VREV32";
843 case ARMISD::VREV16: return "ARMISD::VREV16";
844 case ARMISD::VZIP: return "ARMISD::VZIP";
845 case ARMISD::VUZP: return "ARMISD::VUZP";
846 case ARMISD::VTRN: return "ARMISD::VTRN";
847 case ARMISD::VMULLs: return "ARMISD::VMULLs";
848 case ARMISD::VMULLu: return "ARMISD::VMULLu";
849 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
850 case ARMISD::FMAX: return "ARMISD::FMAX";
851 case ARMISD::FMIN: return "ARMISD::FMIN";
852 case ARMISD::BFI: return "ARMISD::BFI";
853 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
854 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
855 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
856 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
857 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
861 /// getRegClassFor - Return the register class that should be used for the
862 /// specified value type.
863 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
864 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
865 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
866 // load / store 4 to 8 consecutive D registers.
867 if (Subtarget->hasNEON()) {
868 if (VT == MVT::v4i64)
869 return ARM::QQPRRegisterClass;
870 else if (VT == MVT::v8i64)
871 return ARM::QQQQPRRegisterClass;
873 return TargetLowering::getRegClassFor(VT);
876 // Create a fast isel object.
878 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
879 return ARM::createFastISel(funcInfo);
882 /// getFunctionAlignment - Return the Log2 alignment of this function.
883 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
884 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
887 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
888 /// be used for loads / stores from the global.
889 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
890 return (Subtarget->isThumb1Only() ? 127 : 4095);
893 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
894 unsigned NumVals = N->getNumValues();
896 return Sched::RegPressure;
898 for (unsigned i = 0; i != NumVals; ++i) {
899 EVT VT = N->getValueType(i);
900 if (VT == MVT::Glue || VT == MVT::Other)
902 if (VT.isFloatingPoint() || VT.isVector())
903 return Sched::Latency;
906 if (!N->isMachineOpcode())
907 return Sched::RegPressure;
909 // Load are scheduled for latency even if there instruction itinerary
911 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
912 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
914 if (TID.getNumDefs() == 0)
915 return Sched::RegPressure;
916 if (!Itins->isEmpty() &&
917 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
918 return Sched::Latency;
920 return Sched::RegPressure;
923 // FIXME: Move to RegInfo
925 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
926 MachineFunction &MF) const {
927 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering();
929 switch (RC->getID()) {
932 case ARM::tGPRRegClassID:
933 return TFI->hasFP(MF) ? 4 : 5;
934 case ARM::GPRRegClassID: {
935 unsigned FP = TFI->hasFP(MF) ? 1 : 0;
936 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
938 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
939 case ARM::DPRRegClassID:
944 //===----------------------------------------------------------------------===//
946 //===----------------------------------------------------------------------===//
948 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
949 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
951 default: llvm_unreachable("Unknown condition code!");
952 case ISD::SETNE: return ARMCC::NE;
953 case ISD::SETEQ: return ARMCC::EQ;
954 case ISD::SETGT: return ARMCC::GT;
955 case ISD::SETGE: return ARMCC::GE;
956 case ISD::SETLT: return ARMCC::LT;
957 case ISD::SETLE: return ARMCC::LE;
958 case ISD::SETUGT: return ARMCC::HI;
959 case ISD::SETUGE: return ARMCC::HS;
960 case ISD::SETULT: return ARMCC::LO;
961 case ISD::SETULE: return ARMCC::LS;
965 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
966 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
967 ARMCC::CondCodes &CondCode2) {
968 CondCode2 = ARMCC::AL;
970 default: llvm_unreachable("Unknown FP condition!");
972 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
974 case ISD::SETOGT: CondCode = ARMCC::GT; break;
976 case ISD::SETOGE: CondCode = ARMCC::GE; break;
977 case ISD::SETOLT: CondCode = ARMCC::MI; break;
978 case ISD::SETOLE: CondCode = ARMCC::LS; break;
979 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
980 case ISD::SETO: CondCode = ARMCC::VC; break;
981 case ISD::SETUO: CondCode = ARMCC::VS; break;
982 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
983 case ISD::SETUGT: CondCode = ARMCC::HI; break;
984 case ISD::SETUGE: CondCode = ARMCC::PL; break;
986 case ISD::SETULT: CondCode = ARMCC::LT; break;
988 case ISD::SETULE: CondCode = ARMCC::LE; break;
990 case ISD::SETUNE: CondCode = ARMCC::NE; break;
994 //===----------------------------------------------------------------------===//
995 // Calling Convention Implementation
996 //===----------------------------------------------------------------------===//
998 #include "ARMGenCallingConv.inc"
1000 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1001 /// given CallingConvention value.
1002 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1004 bool isVarArg) const {
1007 llvm_unreachable("Unsupported calling convention");
1008 case CallingConv::Fast:
1009 if (Subtarget->hasVFP2() && !isVarArg) {
1010 if (!Subtarget->isAAPCS_ABI())
1011 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1012 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1013 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1016 case CallingConv::C: {
1017 // Use target triple & subtarget features to do actual dispatch.
1018 if (!Subtarget->isAAPCS_ABI())
1019 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1020 else if (Subtarget->hasVFP2() &&
1021 FloatABIType == FloatABI::Hard && !isVarArg)
1022 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1023 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1025 case CallingConv::ARM_AAPCS_VFP:
1026 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1027 case CallingConv::ARM_AAPCS:
1028 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1029 case CallingConv::ARM_APCS:
1030 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1034 /// LowerCallResult - Lower the result values of a call into the
1035 /// appropriate copies out of appropriate physical registers.
1037 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1038 CallingConv::ID CallConv, bool isVarArg,
1039 const SmallVectorImpl<ISD::InputArg> &Ins,
1040 DebugLoc dl, SelectionDAG &DAG,
1041 SmallVectorImpl<SDValue> &InVals) const {
1043 // Assign locations to each value returned by this call.
1044 SmallVector<CCValAssign, 16> RVLocs;
1045 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1046 RVLocs, *DAG.getContext());
1047 CCInfo.AnalyzeCallResult(Ins,
1048 CCAssignFnForNode(CallConv, /* Return*/ true,
1051 // Copy all of the result registers out of their specified physreg.
1052 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1053 CCValAssign VA = RVLocs[i];
1056 if (VA.needsCustom()) {
1057 // Handle f64 or half of a v2f64.
1058 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1060 Chain = Lo.getValue(1);
1061 InFlag = Lo.getValue(2);
1062 VA = RVLocs[++i]; // skip ahead to next loc
1063 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1065 Chain = Hi.getValue(1);
1066 InFlag = Hi.getValue(2);
1067 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1069 if (VA.getLocVT() == MVT::v2f64) {
1070 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1071 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1072 DAG.getConstant(0, MVT::i32));
1074 VA = RVLocs[++i]; // skip ahead to next loc
1075 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1076 Chain = Lo.getValue(1);
1077 InFlag = Lo.getValue(2);
1078 VA = RVLocs[++i]; // skip ahead to next loc
1079 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1080 Chain = Hi.getValue(1);
1081 InFlag = Hi.getValue(2);
1082 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1083 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1084 DAG.getConstant(1, MVT::i32));
1087 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1089 Chain = Val.getValue(1);
1090 InFlag = Val.getValue(2);
1093 switch (VA.getLocInfo()) {
1094 default: llvm_unreachable("Unknown loc info!");
1095 case CCValAssign::Full: break;
1096 case CCValAssign::BCvt:
1097 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1101 InVals.push_back(Val);
1107 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1108 /// by "Src" to address "Dst" of size "Size". Alignment information is
1109 /// specified by the specific parameter attribute. The copy will be passed as
1110 /// a byval function parameter.
1111 /// Sometimes what we are copying is the end of a larger object, the part that
1112 /// does not fit in registers.
1114 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1115 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1117 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1118 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1119 /*isVolatile=*/false, /*AlwaysInline=*/false,
1120 MachinePointerInfo(0), MachinePointerInfo(0));
1123 /// LowerMemOpCallTo - Store the argument to the stack.
1125 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1126 SDValue StackPtr, SDValue Arg,
1127 DebugLoc dl, SelectionDAG &DAG,
1128 const CCValAssign &VA,
1129 ISD::ArgFlagsTy Flags) const {
1130 unsigned LocMemOffset = VA.getLocMemOffset();
1131 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1132 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1133 if (Flags.isByVal())
1134 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1136 return DAG.getStore(Chain, dl, Arg, PtrOff,
1137 MachinePointerInfo::getStack(LocMemOffset),
1141 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1142 SDValue Chain, SDValue &Arg,
1143 RegsToPassVector &RegsToPass,
1144 CCValAssign &VA, CCValAssign &NextVA,
1146 SmallVector<SDValue, 8> &MemOpChains,
1147 ISD::ArgFlagsTy Flags) const {
1149 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1150 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1151 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1153 if (NextVA.isRegLoc())
1154 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1156 assert(NextVA.isMemLoc());
1157 if (StackPtr.getNode() == 0)
1158 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1160 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1166 /// LowerCall - Lowering a call into a callseq_start <-
1167 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1170 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1171 CallingConv::ID CallConv, bool isVarArg,
1173 const SmallVectorImpl<ISD::OutputArg> &Outs,
1174 const SmallVectorImpl<SDValue> &OutVals,
1175 const SmallVectorImpl<ISD::InputArg> &Ins,
1176 DebugLoc dl, SelectionDAG &DAG,
1177 SmallVectorImpl<SDValue> &InVals) const {
1178 MachineFunction &MF = DAG.getMachineFunction();
1179 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1180 bool IsSibCall = false;
1181 // Temporarily disable tail calls so things don't break.
1182 if (!EnableARMTailCalls)
1185 // Check if it's really possible to do a tail call.
1186 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1187 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1188 Outs, OutVals, Ins, DAG);
1189 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1190 // detected sibcalls.
1197 // Analyze operands of the call, assigning locations to each operand.
1198 SmallVector<CCValAssign, 16> ArgLocs;
1199 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1201 CCInfo.AnalyzeCallOperands(Outs,
1202 CCAssignFnForNode(CallConv, /* Return*/ false,
1205 // Get a count of how many bytes are to be pushed on the stack.
1206 unsigned NumBytes = CCInfo.getNextStackOffset();
1208 // For tail calls, memory operands are available in our caller's stack.
1212 // Adjust the stack pointer for the new arguments...
1213 // These operations are automatically eliminated by the prolog/epilog pass
1215 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1217 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1219 RegsToPassVector RegsToPass;
1220 SmallVector<SDValue, 8> MemOpChains;
1222 // Walk the register/memloc assignments, inserting copies/loads. In the case
1223 // of tail call optimization, arguments are handled later.
1224 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1226 ++i, ++realArgIdx) {
1227 CCValAssign &VA = ArgLocs[i];
1228 SDValue Arg = OutVals[realArgIdx];
1229 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1231 // Promote the value if needed.
1232 switch (VA.getLocInfo()) {
1233 default: llvm_unreachable("Unknown loc info!");
1234 case CCValAssign::Full: break;
1235 case CCValAssign::SExt:
1236 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1238 case CCValAssign::ZExt:
1239 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1241 case CCValAssign::AExt:
1242 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1244 case CCValAssign::BCvt:
1245 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1249 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1250 if (VA.needsCustom()) {
1251 if (VA.getLocVT() == MVT::v2f64) {
1252 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1253 DAG.getConstant(0, MVT::i32));
1254 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1255 DAG.getConstant(1, MVT::i32));
1257 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1258 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1260 VA = ArgLocs[++i]; // skip ahead to next loc
1261 if (VA.isRegLoc()) {
1262 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1263 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1265 assert(VA.isMemLoc());
1267 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1268 dl, DAG, VA, Flags));
1271 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1272 StackPtr, MemOpChains, Flags);
1274 } else if (VA.isRegLoc()) {
1275 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1276 } else if (!IsSibCall) {
1277 assert(VA.isMemLoc());
1279 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1280 dl, DAG, VA, Flags));
1284 if (!MemOpChains.empty())
1285 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1286 &MemOpChains[0], MemOpChains.size());
1288 // Build a sequence of copy-to-reg nodes chained together with token chain
1289 // and flag operands which copy the outgoing args into the appropriate regs.
1291 // Tail call byval lowering might overwrite argument registers so in case of
1292 // tail call optimization the copies to registers are lowered later.
1294 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1295 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1296 RegsToPass[i].second, InFlag);
1297 InFlag = Chain.getValue(1);
1300 // For tail calls lower the arguments to the 'real' stack slot.
1302 // Force all the incoming stack arguments to be loaded from the stack
1303 // before any new outgoing arguments are stored to the stack, because the
1304 // outgoing stack slots may alias the incoming argument stack slots, and
1305 // the alias isn't otherwise explicit. This is slightly more conservative
1306 // than necessary, because it means that each store effectively depends
1307 // on every argument instead of just those arguments it would clobber.
1309 // Do not flag preceeding copytoreg stuff together with the following stuff.
1311 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1312 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1313 RegsToPass[i].second, InFlag);
1314 InFlag = Chain.getValue(1);
1319 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1320 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1321 // node so that legalize doesn't hack it.
1322 bool isDirect = false;
1323 bool isARMFunc = false;
1324 bool isLocalARMFunc = false;
1325 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1327 if (EnableARMLongCalls) {
1328 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1329 && "long-calls with non-static relocation model!");
1330 // Handle a global address or an external symbol. If it's not one of
1331 // those, the target's already in a register, so we don't need to do
1333 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1334 const GlobalValue *GV = G->getGlobal();
1335 // Create a constant pool entry for the callee address
1336 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1337 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1340 // Get the address of the callee into a register
1341 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1342 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1343 Callee = DAG.getLoad(getPointerTy(), dl,
1344 DAG.getEntryNode(), CPAddr,
1345 MachinePointerInfo::getConstantPool(),
1347 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1348 const char *Sym = S->getSymbol();
1350 // Create a constant pool entry for the callee address
1351 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1352 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1353 Sym, ARMPCLabelIndex, 0);
1354 // Get the address of the callee into a register
1355 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1356 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1357 Callee = DAG.getLoad(getPointerTy(), dl,
1358 DAG.getEntryNode(), CPAddr,
1359 MachinePointerInfo::getConstantPool(),
1362 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1363 const GlobalValue *GV = G->getGlobal();
1365 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1366 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1367 getTargetMachine().getRelocationModel() != Reloc::Static;
1368 isARMFunc = !Subtarget->isThumb() || isStub;
1369 // ARM call to a local ARM function is predicable.
1370 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1371 // tBX takes a register source operand.
1372 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1373 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1374 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1377 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1378 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1379 Callee = DAG.getLoad(getPointerTy(), dl,
1380 DAG.getEntryNode(), CPAddr,
1381 MachinePointerInfo::getConstantPool(),
1383 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1384 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1385 getPointerTy(), Callee, PICLabel);
1387 // On ELF targets for PIC code, direct calls should go through the PLT
1388 unsigned OpFlags = 0;
1389 if (Subtarget->isTargetELF() &&
1390 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1391 OpFlags = ARMII::MO_PLT;
1392 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1394 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1396 bool isStub = Subtarget->isTargetDarwin() &&
1397 getTargetMachine().getRelocationModel() != Reloc::Static;
1398 isARMFunc = !Subtarget->isThumb() || isStub;
1399 // tBX takes a register source operand.
1400 const char *Sym = S->getSymbol();
1401 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1402 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1403 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1404 Sym, ARMPCLabelIndex, 4);
1405 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1406 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1407 Callee = DAG.getLoad(getPointerTy(), dl,
1408 DAG.getEntryNode(), CPAddr,
1409 MachinePointerInfo::getConstantPool(),
1411 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1412 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1413 getPointerTy(), Callee, PICLabel);
1415 unsigned OpFlags = 0;
1416 // On ELF targets for PIC code, direct calls should go through the PLT
1417 if (Subtarget->isTargetELF() &&
1418 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1419 OpFlags = ARMII::MO_PLT;
1420 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1424 // FIXME: handle tail calls differently.
1426 if (Subtarget->isThumb()) {
1427 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1428 CallOpc = ARMISD::CALL_NOLINK;
1430 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1432 CallOpc = (isDirect || Subtarget->hasV5TOps())
1433 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1434 : ARMISD::CALL_NOLINK;
1437 std::vector<SDValue> Ops;
1438 Ops.push_back(Chain);
1439 Ops.push_back(Callee);
1441 // Add argument registers to the end of the list so that they are known live
1443 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1444 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1445 RegsToPass[i].second.getValueType()));
1447 if (InFlag.getNode())
1448 Ops.push_back(InFlag);
1450 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1452 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1454 // Returns a chain and a flag for retval copy to use.
1455 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1456 InFlag = Chain.getValue(1);
1458 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1459 DAG.getIntPtrConstant(0, true), InFlag);
1461 InFlag = Chain.getValue(1);
1463 // Handle result values, copying them out of physregs into vregs that we
1465 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1469 /// MatchingStackOffset - Return true if the given stack call argument is
1470 /// already available in the same position (relatively) of the caller's
1471 /// incoming argument stack.
1473 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1474 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1475 const ARMInstrInfo *TII) {
1476 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1478 if (Arg.getOpcode() == ISD::CopyFromReg) {
1479 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1480 if (!TargetRegisterInfo::isVirtualRegister(VR))
1482 MachineInstr *Def = MRI->getVRegDef(VR);
1485 if (!Flags.isByVal()) {
1486 if (!TII->isLoadFromStackSlot(Def, FI))
1491 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1492 if (Flags.isByVal())
1493 // ByVal argument is passed in as a pointer but it's now being
1494 // dereferenced. e.g.
1495 // define @foo(%struct.X* %A) {
1496 // tail call @bar(%struct.X* byval %A)
1499 SDValue Ptr = Ld->getBasePtr();
1500 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1503 FI = FINode->getIndex();
1507 assert(FI != INT_MAX);
1508 if (!MFI->isFixedObjectIndex(FI))
1510 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1513 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1514 /// for tail call optimization. Targets which want to do tail call
1515 /// optimization should implement this function.
1517 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1518 CallingConv::ID CalleeCC,
1520 bool isCalleeStructRet,
1521 bool isCallerStructRet,
1522 const SmallVectorImpl<ISD::OutputArg> &Outs,
1523 const SmallVectorImpl<SDValue> &OutVals,
1524 const SmallVectorImpl<ISD::InputArg> &Ins,
1525 SelectionDAG& DAG) const {
1526 const Function *CallerF = DAG.getMachineFunction().getFunction();
1527 CallingConv::ID CallerCC = CallerF->getCallingConv();
1528 bool CCMatch = CallerCC == CalleeCC;
1530 // Look for obvious safe cases to perform tail call optimization that do not
1531 // require ABI changes. This is what gcc calls sibcall.
1533 // Do not sibcall optimize vararg calls unless the call site is not passing
1535 if (isVarArg && !Outs.empty())
1538 // Also avoid sibcall optimization if either caller or callee uses struct
1539 // return semantics.
1540 if (isCalleeStructRet || isCallerStructRet)
1543 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1544 // emitEpilogue is not ready for them.
1545 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1546 // LR. This means if we need to reload LR, it takes an extra instructions,
1547 // which outweighs the value of the tail call; but here we don't know yet
1548 // whether LR is going to be used. Probably the right approach is to
1549 // generate the tail call here and turn it back into CALL/RET in
1550 // emitEpilogue if LR is used.
1552 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1553 // but we need to make sure there are enough registers; the only valid
1554 // registers are the 4 used for parameters. We don't currently do this
1556 if (Subtarget->isThumb1Only())
1559 // If the calling conventions do not match, then we'd better make sure the
1560 // results are returned in the same way as what the caller expects.
1562 SmallVector<CCValAssign, 16> RVLocs1;
1563 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1564 RVLocs1, *DAG.getContext());
1565 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1567 SmallVector<CCValAssign, 16> RVLocs2;
1568 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1569 RVLocs2, *DAG.getContext());
1570 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1572 if (RVLocs1.size() != RVLocs2.size())
1574 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1575 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1577 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1579 if (RVLocs1[i].isRegLoc()) {
1580 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1583 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1589 // If the callee takes no arguments then go on to check the results of the
1591 if (!Outs.empty()) {
1592 // Check if stack adjustment is needed. For now, do not do this if any
1593 // argument is passed on the stack.
1594 SmallVector<CCValAssign, 16> ArgLocs;
1595 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1596 ArgLocs, *DAG.getContext());
1597 CCInfo.AnalyzeCallOperands(Outs,
1598 CCAssignFnForNode(CalleeCC, false, isVarArg));
1599 if (CCInfo.getNextStackOffset()) {
1600 MachineFunction &MF = DAG.getMachineFunction();
1602 // Check if the arguments are already laid out in the right way as
1603 // the caller's fixed stack objects.
1604 MachineFrameInfo *MFI = MF.getFrameInfo();
1605 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1606 const ARMInstrInfo *TII =
1607 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1608 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1610 ++i, ++realArgIdx) {
1611 CCValAssign &VA = ArgLocs[i];
1612 EVT RegVT = VA.getLocVT();
1613 SDValue Arg = OutVals[realArgIdx];
1614 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1615 if (VA.getLocInfo() == CCValAssign::Indirect)
1617 if (VA.needsCustom()) {
1618 // f64 and vector types are split into multiple registers or
1619 // register/stack-slot combinations. The types will not match
1620 // the registers; give up on memory f64 refs until we figure
1621 // out what to do about this.
1624 if (!ArgLocs[++i].isRegLoc())
1626 if (RegVT == MVT::v2f64) {
1627 if (!ArgLocs[++i].isRegLoc())
1629 if (!ArgLocs[++i].isRegLoc())
1632 } else if (!VA.isRegLoc()) {
1633 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1645 ARMTargetLowering::LowerReturn(SDValue Chain,
1646 CallingConv::ID CallConv, bool isVarArg,
1647 const SmallVectorImpl<ISD::OutputArg> &Outs,
1648 const SmallVectorImpl<SDValue> &OutVals,
1649 DebugLoc dl, SelectionDAG &DAG) const {
1651 // CCValAssign - represent the assignment of the return value to a location.
1652 SmallVector<CCValAssign, 16> RVLocs;
1654 // CCState - Info about the registers and stack slots.
1655 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1658 // Analyze outgoing return values.
1659 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1662 // If this is the first return lowered for this function, add
1663 // the regs to the liveout set for the function.
1664 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1665 for (unsigned i = 0; i != RVLocs.size(); ++i)
1666 if (RVLocs[i].isRegLoc())
1667 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1672 // Copy the result values into the output registers.
1673 for (unsigned i = 0, realRVLocIdx = 0;
1675 ++i, ++realRVLocIdx) {
1676 CCValAssign &VA = RVLocs[i];
1677 assert(VA.isRegLoc() && "Can only return in registers!");
1679 SDValue Arg = OutVals[realRVLocIdx];
1681 switch (VA.getLocInfo()) {
1682 default: llvm_unreachable("Unknown loc info!");
1683 case CCValAssign::Full: break;
1684 case CCValAssign::BCvt:
1685 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1689 if (VA.needsCustom()) {
1690 if (VA.getLocVT() == MVT::v2f64) {
1691 // Extract the first half and return it in two registers.
1692 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1693 DAG.getConstant(0, MVT::i32));
1694 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1695 DAG.getVTList(MVT::i32, MVT::i32), Half);
1697 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1698 Flag = Chain.getValue(1);
1699 VA = RVLocs[++i]; // skip ahead to next loc
1700 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1701 HalfGPRs.getValue(1), Flag);
1702 Flag = Chain.getValue(1);
1703 VA = RVLocs[++i]; // skip ahead to next loc
1705 // Extract the 2nd half and fall through to handle it as an f64 value.
1706 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1707 DAG.getConstant(1, MVT::i32));
1709 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1711 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1712 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1713 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1714 Flag = Chain.getValue(1);
1715 VA = RVLocs[++i]; // skip ahead to next loc
1716 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1719 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1721 // Guarantee that all emitted copies are
1722 // stuck together, avoiding something bad.
1723 Flag = Chain.getValue(1);
1728 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1730 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1735 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1736 if (N->getNumValues() != 1)
1738 if (!N->hasNUsesOfValue(1, 0))
1741 unsigned NumCopies = 0;
1743 SDNode *Use = *N->use_begin();
1744 if (Use->getOpcode() == ISD::CopyToReg) {
1745 Copies[NumCopies++] = Use;
1746 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1747 // f64 returned in a pair of GPRs.
1748 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1750 if (UI->getOpcode() != ISD::CopyToReg)
1752 Copies[UI.getUse().getResNo()] = *UI;
1755 } else if (Use->getOpcode() == ISD::BITCAST) {
1756 // f32 returned in a single GPR.
1757 if (!Use->hasNUsesOfValue(1, 0))
1759 Use = *Use->use_begin();
1760 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1762 Copies[NumCopies++] = Use;
1767 if (NumCopies != 1 && NumCopies != 2)
1770 bool HasRet = false;
1771 for (unsigned i = 0; i < NumCopies; ++i) {
1772 SDNode *Copy = Copies[i];
1773 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1775 if (UI->getOpcode() == ISD::CopyToReg) {
1777 if (Use == Copies[0] || Use == Copies[1])
1781 if (UI->getOpcode() != ARMISD::RET_FLAG)
1790 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1791 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1792 // one of the above mentioned nodes. It has to be wrapped because otherwise
1793 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1794 // be used to form addressing mode. These wrapped nodes will be selected
1796 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1797 EVT PtrVT = Op.getValueType();
1798 // FIXME there is no actual debug info here
1799 DebugLoc dl = Op.getDebugLoc();
1800 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1802 if (CP->isMachineConstantPoolEntry())
1803 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1804 CP->getAlignment());
1806 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1807 CP->getAlignment());
1808 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1811 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1812 return MachineJumpTableInfo::EK_Inline;
1815 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1816 SelectionDAG &DAG) const {
1817 MachineFunction &MF = DAG.getMachineFunction();
1818 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1819 unsigned ARMPCLabelIndex = 0;
1820 DebugLoc DL = Op.getDebugLoc();
1821 EVT PtrVT = getPointerTy();
1822 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1823 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1825 if (RelocM == Reloc::Static) {
1826 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1828 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1829 ARMPCLabelIndex = AFI->createPICLabelUId();
1830 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1831 ARMCP::CPBlockAddress,
1833 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1835 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1836 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1837 MachinePointerInfo::getConstantPool(),
1839 if (RelocM == Reloc::Static)
1841 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1842 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1845 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1847 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1848 SelectionDAG &DAG) const {
1849 DebugLoc dl = GA->getDebugLoc();
1850 EVT PtrVT = getPointerTy();
1851 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1852 MachineFunction &MF = DAG.getMachineFunction();
1853 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1854 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1855 ARMConstantPoolValue *CPV =
1856 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1857 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1858 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1859 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1860 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1861 MachinePointerInfo::getConstantPool(),
1863 SDValue Chain = Argument.getValue(1);
1865 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1866 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1868 // call __tls_get_addr.
1871 Entry.Node = Argument;
1872 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1873 Args.push_back(Entry);
1874 // FIXME: is there useful debug info available here?
1875 std::pair<SDValue, SDValue> CallResult =
1876 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1877 false, false, false, false,
1878 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1879 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1880 return CallResult.first;
1883 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1884 // "local exec" model.
1886 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1887 SelectionDAG &DAG) const {
1888 const GlobalValue *GV = GA->getGlobal();
1889 DebugLoc dl = GA->getDebugLoc();
1891 SDValue Chain = DAG.getEntryNode();
1892 EVT PtrVT = getPointerTy();
1893 // Get the Thread Pointer
1894 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1896 if (GV->isDeclaration()) {
1897 MachineFunction &MF = DAG.getMachineFunction();
1898 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1899 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1900 // Initial exec model.
1901 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1902 ARMConstantPoolValue *CPV =
1903 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1904 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1905 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1906 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1907 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1908 MachinePointerInfo::getConstantPool(),
1910 Chain = Offset.getValue(1);
1912 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1913 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1915 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1916 MachinePointerInfo::getConstantPool(),
1920 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
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(),
1928 // The address of the thread local variable is the add of the thread
1929 // pointer with the offset of the variable.
1930 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1934 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1935 // TODO: implement the "local dynamic" model
1936 assert(Subtarget->isTargetELF() &&
1937 "TLS not implemented for non-ELF targets");
1938 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1939 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1940 // otherwise use the "Local Exec" TLS Model
1941 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1942 return LowerToTLSGeneralDynamicModel(GA, DAG);
1944 return LowerToTLSExecModels(GA, DAG);
1947 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1948 SelectionDAG &DAG) const {
1949 EVT PtrVT = getPointerTy();
1950 DebugLoc dl = Op.getDebugLoc();
1951 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1952 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1953 if (RelocM == Reloc::PIC_) {
1954 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1955 ARMConstantPoolValue *CPV =
1956 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
1957 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1958 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1959 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1961 MachinePointerInfo::getConstantPool(),
1963 SDValue Chain = Result.getValue(1);
1964 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1965 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1967 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1968 MachinePointerInfo::getGOT(), false, false, 0);
1972 // If we have T2 ops, we can materialize the address directly via movt/movw
1973 // pair. This is always cheaper.
1974 if (Subtarget->useMovt()) {
1976 // FIXME: Once remat is capable of dealing with instructions with register
1977 // operands, expand this into two nodes.
1978 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1979 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1981 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1982 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1983 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1984 MachinePointerInfo::getConstantPool(),
1989 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1990 SelectionDAG &DAG) const {
1991 EVT PtrVT = getPointerTy();
1992 DebugLoc dl = Op.getDebugLoc();
1993 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1994 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1995 MachineFunction &MF = DAG.getMachineFunction();
1996 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1998 if (Subtarget->useMovt()) {
2000 // FIXME: Once remat is capable of dealing with instructions with register
2001 // operands, expand this into two nodes.
2002 if (RelocM != Reloc::PIC_)
2003 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2004 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2006 SDValue Result = DAG.getNode(ARMISD::WrapperPIC, dl, PtrVT,
2007 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2008 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2009 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2010 MachinePointerInfo::getGOT(), false, false, 0);
2014 unsigned ARMPCLabelIndex = 0;
2016 if (RelocM == Reloc::Static) {
2017 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2019 ARMPCLabelIndex = AFI->createPICLabelUId();
2020 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2021 ARMConstantPoolValue *CPV =
2022 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2023 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2025 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2027 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2028 MachinePointerInfo::getConstantPool(),
2030 SDValue Chain = Result.getValue(1);
2032 if (RelocM == Reloc::PIC_) {
2033 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2034 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2037 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2038 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2044 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2045 SelectionDAG &DAG) const {
2046 assert(Subtarget->isTargetELF() &&
2047 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2048 MachineFunction &MF = DAG.getMachineFunction();
2049 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2050 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2051 EVT PtrVT = getPointerTy();
2052 DebugLoc dl = Op.getDebugLoc();
2053 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2054 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2055 "_GLOBAL_OFFSET_TABLE_",
2056 ARMPCLabelIndex, PCAdj);
2057 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2058 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2059 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2060 MachinePointerInfo::getConstantPool(),
2062 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2063 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2067 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2069 DebugLoc dl = Op.getDebugLoc();
2070 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2071 Op.getOperand(0), Op.getOperand(1));
2075 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2076 DebugLoc dl = Op.getDebugLoc();
2077 SDValue Val = DAG.getConstant(0, MVT::i32);
2078 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2079 Op.getOperand(1), Val);
2083 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2084 DebugLoc dl = Op.getDebugLoc();
2085 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2086 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2090 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2091 const ARMSubtarget *Subtarget) const {
2092 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2093 DebugLoc dl = Op.getDebugLoc();
2095 default: return SDValue(); // Don't custom lower most intrinsics.
2096 case Intrinsic::arm_thread_pointer: {
2097 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2098 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2100 case Intrinsic::eh_sjlj_lsda: {
2101 MachineFunction &MF = DAG.getMachineFunction();
2102 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2103 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2104 EVT PtrVT = getPointerTy();
2105 DebugLoc dl = Op.getDebugLoc();
2106 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2108 unsigned PCAdj = (RelocM != Reloc::PIC_)
2109 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2110 ARMConstantPoolValue *CPV =
2111 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2112 ARMCP::CPLSDA, PCAdj);
2113 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2114 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2116 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2117 MachinePointerInfo::getConstantPool(),
2120 if (RelocM == Reloc::PIC_) {
2121 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2122 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2129 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2130 const ARMSubtarget *Subtarget) {
2131 DebugLoc dl = Op.getDebugLoc();
2132 if (!Subtarget->hasDataBarrier()) {
2133 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2134 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2136 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2137 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2138 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2139 DAG.getConstant(0, MVT::i32));
2142 SDValue Op5 = Op.getOperand(5);
2143 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2144 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2145 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2146 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2148 ARM_MB::MemBOpt DMBOpt;
2149 if (isDeviceBarrier)
2150 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2152 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2153 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2154 DAG.getConstant(DMBOpt, MVT::i32));
2157 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2158 const ARMSubtarget *Subtarget) {
2159 // ARM pre v5TE and Thumb1 does not have preload instructions.
2160 if (!(Subtarget->isThumb2() ||
2161 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2162 // Just preserve the chain.
2163 return Op.getOperand(0);
2165 DebugLoc dl = Op.getDebugLoc();
2166 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2168 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2169 // ARMv7 with MP extension has PLDW.
2170 return Op.getOperand(0);
2172 if (Subtarget->isThumb())
2174 isRead = ~isRead & 1;
2175 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2177 // Currently there is no intrinsic that matches pli.
2178 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2179 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2180 DAG.getConstant(isData, MVT::i32));
2183 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2184 MachineFunction &MF = DAG.getMachineFunction();
2185 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2187 // vastart just stores the address of the VarArgsFrameIndex slot into the
2188 // memory location argument.
2189 DebugLoc dl = Op.getDebugLoc();
2190 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2191 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2192 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2193 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2194 MachinePointerInfo(SV), false, false, 0);
2198 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2199 SDValue &Root, SelectionDAG &DAG,
2200 DebugLoc dl) const {
2201 MachineFunction &MF = DAG.getMachineFunction();
2202 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2204 TargetRegisterClass *RC;
2205 if (AFI->isThumb1OnlyFunction())
2206 RC = ARM::tGPRRegisterClass;
2208 RC = ARM::GPRRegisterClass;
2210 // Transform the arguments stored in physical registers into virtual ones.
2211 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2212 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2215 if (NextVA.isMemLoc()) {
2216 MachineFrameInfo *MFI = MF.getFrameInfo();
2217 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2219 // Create load node to retrieve arguments from the stack.
2220 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2221 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2222 MachinePointerInfo::getFixedStack(FI),
2225 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2226 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2229 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2233 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2234 CallingConv::ID CallConv, bool isVarArg,
2235 const SmallVectorImpl<ISD::InputArg>
2237 DebugLoc dl, SelectionDAG &DAG,
2238 SmallVectorImpl<SDValue> &InVals)
2241 MachineFunction &MF = DAG.getMachineFunction();
2242 MachineFrameInfo *MFI = MF.getFrameInfo();
2244 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2246 // Assign locations to all of the incoming arguments.
2247 SmallVector<CCValAssign, 16> ArgLocs;
2248 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2250 CCInfo.AnalyzeFormalArguments(Ins,
2251 CCAssignFnForNode(CallConv, /* Return*/ false,
2254 SmallVector<SDValue, 16> ArgValues;
2256 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2257 CCValAssign &VA = ArgLocs[i];
2259 // Arguments stored in registers.
2260 if (VA.isRegLoc()) {
2261 EVT RegVT = VA.getLocVT();
2264 if (VA.needsCustom()) {
2265 // f64 and vector types are split up into multiple registers or
2266 // combinations of registers and stack slots.
2267 if (VA.getLocVT() == MVT::v2f64) {
2268 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2270 VA = ArgLocs[++i]; // skip ahead to next loc
2272 if (VA.isMemLoc()) {
2273 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2274 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2275 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2276 MachinePointerInfo::getFixedStack(FI),
2279 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2282 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2283 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2284 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2285 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2286 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2288 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2291 TargetRegisterClass *RC;
2293 if (RegVT == MVT::f32)
2294 RC = ARM::SPRRegisterClass;
2295 else if (RegVT == MVT::f64)
2296 RC = ARM::DPRRegisterClass;
2297 else if (RegVT == MVT::v2f64)
2298 RC = ARM::QPRRegisterClass;
2299 else if (RegVT == MVT::i32)
2300 RC = (AFI->isThumb1OnlyFunction() ?
2301 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2303 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2305 // Transform the arguments in physical registers into virtual ones.
2306 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2307 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2310 // If this is an 8 or 16-bit value, it is really passed promoted
2311 // to 32 bits. Insert an assert[sz]ext to capture this, then
2312 // truncate to the right size.
2313 switch (VA.getLocInfo()) {
2314 default: llvm_unreachable("Unknown loc info!");
2315 case CCValAssign::Full: break;
2316 case CCValAssign::BCvt:
2317 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2319 case CCValAssign::SExt:
2320 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2321 DAG.getValueType(VA.getValVT()));
2322 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2324 case CCValAssign::ZExt:
2325 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2326 DAG.getValueType(VA.getValVT()));
2327 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2331 InVals.push_back(ArgValue);
2333 } else { // VA.isRegLoc()
2336 assert(VA.isMemLoc());
2337 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2339 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2340 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2342 // Create load nodes to retrieve arguments from the stack.
2343 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2344 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2345 MachinePointerInfo::getFixedStack(FI),
2352 static const unsigned GPRArgRegs[] = {
2353 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2356 unsigned NumGPRs = CCInfo.getFirstUnallocated
2357 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2359 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2360 unsigned VARegSize = (4 - NumGPRs) * 4;
2361 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2362 unsigned ArgOffset = CCInfo.getNextStackOffset();
2363 if (VARegSaveSize) {
2364 // If this function is vararg, store any remaining integer argument regs
2365 // to their spots on the stack so that they may be loaded by deferencing
2366 // the result of va_next.
2367 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2368 AFI->setVarArgsFrameIndex(
2369 MFI->CreateFixedObject(VARegSaveSize,
2370 ArgOffset + VARegSaveSize - VARegSize,
2372 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2375 SmallVector<SDValue, 4> MemOps;
2376 for (; NumGPRs < 4; ++NumGPRs) {
2377 TargetRegisterClass *RC;
2378 if (AFI->isThumb1OnlyFunction())
2379 RC = ARM::tGPRRegisterClass;
2381 RC = ARM::GPRRegisterClass;
2383 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2384 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2386 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2387 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2389 MemOps.push_back(Store);
2390 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2391 DAG.getConstant(4, getPointerTy()));
2393 if (!MemOps.empty())
2394 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2395 &MemOps[0], MemOps.size());
2397 // This will point to the next argument passed via stack.
2398 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2404 /// isFloatingPointZero - Return true if this is +0.0.
2405 static bool isFloatingPointZero(SDValue Op) {
2406 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2407 return CFP->getValueAPF().isPosZero();
2408 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2409 // Maybe this has already been legalized into the constant pool?
2410 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2411 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2412 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2413 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2414 return CFP->getValueAPF().isPosZero();
2420 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2421 /// the given operands.
2423 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2424 SDValue &ARMcc, SelectionDAG &DAG,
2425 DebugLoc dl) const {
2426 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2427 unsigned C = RHSC->getZExtValue();
2428 if (!isLegalICmpImmediate(C)) {
2429 // Constant does not fit, try adjusting it by one?
2434 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2435 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2436 RHS = DAG.getConstant(C-1, MVT::i32);
2441 if (C != 0 && isLegalICmpImmediate(C-1)) {
2442 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2443 RHS = DAG.getConstant(C-1, MVT::i32);
2448 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2449 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2450 RHS = DAG.getConstant(C+1, MVT::i32);
2455 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2456 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2457 RHS = DAG.getConstant(C+1, MVT::i32);
2464 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2465 ARMISD::NodeType CompareType;
2468 CompareType = ARMISD::CMP;
2473 CompareType = ARMISD::CMPZ;
2476 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2477 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2480 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2482 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2483 DebugLoc dl) const {
2485 if (!isFloatingPointZero(RHS))
2486 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2488 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2489 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2492 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2493 SDValue Cond = Op.getOperand(0);
2494 SDValue SelectTrue = Op.getOperand(1);
2495 SDValue SelectFalse = Op.getOperand(2);
2496 DebugLoc dl = Op.getDebugLoc();
2500 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2501 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2503 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2504 const ConstantSDNode *CMOVTrue =
2505 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2506 const ConstantSDNode *CMOVFalse =
2507 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2509 if (CMOVTrue && CMOVFalse) {
2510 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2511 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2515 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2517 False = SelectFalse;
2518 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2523 if (True.getNode() && False.getNode()) {
2524 EVT VT = Cond.getValueType();
2525 SDValue ARMcc = Cond.getOperand(2);
2526 SDValue CCR = Cond.getOperand(3);
2527 SDValue Cmp = Cond.getOperand(4);
2528 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2533 return DAG.getSelectCC(dl, Cond,
2534 DAG.getConstant(0, Cond.getValueType()),
2535 SelectTrue, SelectFalse, ISD::SETNE);
2538 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2539 EVT VT = Op.getValueType();
2540 SDValue LHS = Op.getOperand(0);
2541 SDValue RHS = Op.getOperand(1);
2542 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2543 SDValue TrueVal = Op.getOperand(2);
2544 SDValue FalseVal = Op.getOperand(3);
2545 DebugLoc dl = Op.getDebugLoc();
2547 if (LHS.getValueType() == MVT::i32) {
2549 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2550 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2551 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2554 ARMCC::CondCodes CondCode, CondCode2;
2555 FPCCToARMCC(CC, CondCode, CondCode2);
2557 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2558 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2559 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2560 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2562 if (CondCode2 != ARMCC::AL) {
2563 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2564 // FIXME: Needs another CMP because flag can have but one use.
2565 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2566 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2567 Result, TrueVal, ARMcc2, CCR, Cmp2);
2572 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2573 /// to morph to an integer compare sequence.
2574 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2575 const ARMSubtarget *Subtarget) {
2576 SDNode *N = Op.getNode();
2577 if (!N->hasOneUse())
2578 // Otherwise it requires moving the value from fp to integer registers.
2580 if (!N->getNumValues())
2582 EVT VT = Op.getValueType();
2583 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2584 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2585 // vmrs are very slow, e.g. cortex-a8.
2588 if (isFloatingPointZero(Op)) {
2592 return ISD::isNormalLoad(N);
2595 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2596 if (isFloatingPointZero(Op))
2597 return DAG.getConstant(0, MVT::i32);
2599 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2600 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2601 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2602 Ld->isVolatile(), Ld->isNonTemporal(),
2603 Ld->getAlignment());
2605 llvm_unreachable("Unknown VFP cmp argument!");
2608 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2609 SDValue &RetVal1, SDValue &RetVal2) {
2610 if (isFloatingPointZero(Op)) {
2611 RetVal1 = DAG.getConstant(0, MVT::i32);
2612 RetVal2 = DAG.getConstant(0, MVT::i32);
2616 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2617 SDValue Ptr = Ld->getBasePtr();
2618 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2619 Ld->getChain(), Ptr,
2620 Ld->getPointerInfo(),
2621 Ld->isVolatile(), Ld->isNonTemporal(),
2622 Ld->getAlignment());
2624 EVT PtrType = Ptr.getValueType();
2625 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2626 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2627 PtrType, Ptr, DAG.getConstant(4, PtrType));
2628 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2629 Ld->getChain(), NewPtr,
2630 Ld->getPointerInfo().getWithOffset(4),
2631 Ld->isVolatile(), Ld->isNonTemporal(),
2636 llvm_unreachable("Unknown VFP cmp argument!");
2639 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2640 /// f32 and even f64 comparisons to integer ones.
2642 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2643 SDValue Chain = Op.getOperand(0);
2644 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2645 SDValue LHS = Op.getOperand(2);
2646 SDValue RHS = Op.getOperand(3);
2647 SDValue Dest = Op.getOperand(4);
2648 DebugLoc dl = Op.getDebugLoc();
2650 bool SeenZero = false;
2651 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2652 canChangeToInt(RHS, SeenZero, Subtarget) &&
2653 // If one of the operand is zero, it's safe to ignore the NaN case since
2654 // we only care about equality comparisons.
2655 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2656 // If unsafe fp math optimization is enabled and there are no othter uses of
2657 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2658 // to an integer comparison.
2659 if (CC == ISD::SETOEQ)
2661 else if (CC == ISD::SETUNE)
2665 if (LHS.getValueType() == MVT::f32) {
2666 LHS = bitcastf32Toi32(LHS, DAG);
2667 RHS = bitcastf32Toi32(RHS, DAG);
2668 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2669 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2670 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2671 Chain, Dest, ARMcc, CCR, Cmp);
2676 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2677 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2678 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2679 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2680 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2681 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2682 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2688 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2689 SDValue Chain = Op.getOperand(0);
2690 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2691 SDValue LHS = Op.getOperand(2);
2692 SDValue RHS = Op.getOperand(3);
2693 SDValue Dest = Op.getOperand(4);
2694 DebugLoc dl = Op.getDebugLoc();
2696 if (LHS.getValueType() == MVT::i32) {
2698 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2699 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2700 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2701 Chain, Dest, ARMcc, CCR, Cmp);
2704 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2707 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2708 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2709 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2710 if (Result.getNode())
2714 ARMCC::CondCodes CondCode, CondCode2;
2715 FPCCToARMCC(CC, CondCode, CondCode2);
2717 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2718 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2719 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2720 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2721 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2722 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2723 if (CondCode2 != ARMCC::AL) {
2724 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2725 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2726 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2731 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2732 SDValue Chain = Op.getOperand(0);
2733 SDValue Table = Op.getOperand(1);
2734 SDValue Index = Op.getOperand(2);
2735 DebugLoc dl = Op.getDebugLoc();
2737 EVT PTy = getPointerTy();
2738 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2739 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2740 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2741 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2742 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2743 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2744 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2745 if (Subtarget->isThumb2()) {
2746 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2747 // which does another jump to the destination. This also makes it easier
2748 // to translate it to TBB / TBH later.
2749 // FIXME: This might not work if the function is extremely large.
2750 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2751 Addr, Op.getOperand(2), JTI, UId);
2753 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2754 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2755 MachinePointerInfo::getJumpTable(),
2757 Chain = Addr.getValue(1);
2758 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2759 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2761 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2762 MachinePointerInfo::getJumpTable(), false, false, 0);
2763 Chain = Addr.getValue(1);
2764 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2768 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2769 DebugLoc dl = Op.getDebugLoc();
2772 switch (Op.getOpcode()) {
2774 assert(0 && "Invalid opcode!");
2775 case ISD::FP_TO_SINT:
2776 Opc = ARMISD::FTOSI;
2778 case ISD::FP_TO_UINT:
2779 Opc = ARMISD::FTOUI;
2782 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2783 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2786 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2787 EVT VT = Op.getValueType();
2788 DebugLoc dl = Op.getDebugLoc();
2791 switch (Op.getOpcode()) {
2793 assert(0 && "Invalid opcode!");
2794 case ISD::SINT_TO_FP:
2795 Opc = ARMISD::SITOF;
2797 case ISD::UINT_TO_FP:
2798 Opc = ARMISD::UITOF;
2802 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
2803 return DAG.getNode(Opc, dl, VT, Op);
2806 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2807 // Implement fcopysign with a fabs and a conditional fneg.
2808 SDValue Tmp0 = Op.getOperand(0);
2809 SDValue Tmp1 = Op.getOperand(1);
2810 DebugLoc dl = Op.getDebugLoc();
2811 EVT VT = Op.getValueType();
2812 EVT SrcVT = Tmp1.getValueType();
2813 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2814 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2815 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2816 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2817 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2818 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2821 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2822 MachineFunction &MF = DAG.getMachineFunction();
2823 MachineFrameInfo *MFI = MF.getFrameInfo();
2824 MFI->setReturnAddressIsTaken(true);
2826 EVT VT = Op.getValueType();
2827 DebugLoc dl = Op.getDebugLoc();
2828 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2830 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2831 SDValue Offset = DAG.getConstant(4, MVT::i32);
2832 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2833 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2834 MachinePointerInfo(), false, false, 0);
2837 // Return LR, which contains the return address. Mark it an implicit live-in.
2838 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2839 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2842 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2843 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2844 MFI->setFrameAddressIsTaken(true);
2846 EVT VT = Op.getValueType();
2847 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2848 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2849 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2850 ? ARM::R7 : ARM::R11;
2851 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2853 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2854 MachinePointerInfo(),
2859 /// ExpandBITCAST - If the target supports VFP, this function is called to
2860 /// expand a bit convert where either the source or destination type is i64 to
2861 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2862 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2863 /// vectors), since the legalizer won't know what to do with that.
2864 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
2865 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2866 DebugLoc dl = N->getDebugLoc();
2867 SDValue Op = N->getOperand(0);
2869 // This function is only supposed to be called for i64 types, either as the
2870 // source or destination of the bit convert.
2871 EVT SrcVT = Op.getValueType();
2872 EVT DstVT = N->getValueType(0);
2873 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2874 "ExpandBITCAST called for non-i64 type");
2876 // Turn i64->f64 into VMOVDRR.
2877 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2878 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2879 DAG.getConstant(0, MVT::i32));
2880 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2881 DAG.getConstant(1, MVT::i32));
2882 return DAG.getNode(ISD::BITCAST, dl, DstVT,
2883 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2886 // Turn f64->i64 into VMOVRRD.
2887 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2888 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2889 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2890 // Merge the pieces into a single i64 value.
2891 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2897 /// getZeroVector - Returns a vector of specified type with all zero elements.
2898 /// Zero vectors are used to represent vector negation and in those cases
2899 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2900 /// not support i64 elements, so sometimes the zero vectors will need to be
2901 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2903 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2904 assert(VT.isVector() && "Expected a vector type");
2905 // The canonical modified immediate encoding of a zero vector is....0!
2906 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2907 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2908 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2909 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
2912 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2913 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2914 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2915 SelectionDAG &DAG) const {
2916 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2917 EVT VT = Op.getValueType();
2918 unsigned VTBits = VT.getSizeInBits();
2919 DebugLoc dl = Op.getDebugLoc();
2920 SDValue ShOpLo = Op.getOperand(0);
2921 SDValue ShOpHi = Op.getOperand(1);
2922 SDValue ShAmt = Op.getOperand(2);
2924 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2926 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2928 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2929 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2930 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2931 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2932 DAG.getConstant(VTBits, MVT::i32));
2933 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2934 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2935 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2937 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2938 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2940 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2941 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2944 SDValue Ops[2] = { Lo, Hi };
2945 return DAG.getMergeValues(Ops, 2, dl);
2948 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2949 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2950 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2951 SelectionDAG &DAG) const {
2952 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2953 EVT VT = Op.getValueType();
2954 unsigned VTBits = VT.getSizeInBits();
2955 DebugLoc dl = Op.getDebugLoc();
2956 SDValue ShOpLo = Op.getOperand(0);
2957 SDValue ShOpHi = Op.getOperand(1);
2958 SDValue ShAmt = Op.getOperand(2);
2961 assert(Op.getOpcode() == ISD::SHL_PARTS);
2962 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2963 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2964 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2965 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2966 DAG.getConstant(VTBits, MVT::i32));
2967 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2968 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2970 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2971 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2972 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2974 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2975 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2978 SDValue Ops[2] = { Lo, Hi };
2979 return DAG.getMergeValues(Ops, 2, dl);
2982 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2983 SelectionDAG &DAG) const {
2984 // The rounding mode is in bits 23:22 of the FPSCR.
2985 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2986 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2987 // so that the shift + and get folded into a bitfield extract.
2988 DebugLoc dl = Op.getDebugLoc();
2989 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2990 DAG.getConstant(Intrinsic::arm_get_fpscr,
2992 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2993 DAG.getConstant(1U << 22, MVT::i32));
2994 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2995 DAG.getConstant(22, MVT::i32));
2996 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2997 DAG.getConstant(3, MVT::i32));
3000 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3001 const ARMSubtarget *ST) {
3002 EVT VT = N->getValueType(0);
3003 DebugLoc dl = N->getDebugLoc();
3005 if (!ST->hasV6T2Ops())
3008 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3009 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3012 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3013 const ARMSubtarget *ST) {
3014 EVT VT = N->getValueType(0);
3015 DebugLoc dl = N->getDebugLoc();
3020 // Lower vector shifts on NEON to use VSHL.
3021 assert(ST->hasNEON() && "unexpected vector shift");
3023 // Left shifts translate directly to the vshiftu intrinsic.
3024 if (N->getOpcode() == ISD::SHL)
3025 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3026 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3027 N->getOperand(0), N->getOperand(1));
3029 assert((N->getOpcode() == ISD::SRA ||
3030 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3032 // NEON uses the same intrinsics for both left and right shifts. For
3033 // right shifts, the shift amounts are negative, so negate the vector of
3035 EVT ShiftVT = N->getOperand(1).getValueType();
3036 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3037 getZeroVector(ShiftVT, DAG, dl),
3039 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3040 Intrinsic::arm_neon_vshifts :
3041 Intrinsic::arm_neon_vshiftu);
3042 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3043 DAG.getConstant(vshiftInt, MVT::i32),
3044 N->getOperand(0), NegatedCount);
3047 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3048 const ARMSubtarget *ST) {
3049 EVT VT = N->getValueType(0);
3050 DebugLoc dl = N->getDebugLoc();
3052 // We can get here for a node like i32 = ISD::SHL i32, i64
3056 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3057 "Unknown shift to lower!");
3059 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3060 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3061 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3064 // If we are in thumb mode, we don't have RRX.
3065 if (ST->isThumb1Only()) return SDValue();
3067 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3068 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3069 DAG.getConstant(0, MVT::i32));
3070 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3071 DAG.getConstant(1, MVT::i32));
3073 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3074 // captures the result into a carry flag.
3075 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3076 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3078 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3079 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3081 // Merge the pieces into a single i64 value.
3082 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3085 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3086 SDValue TmpOp0, TmpOp1;
3087 bool Invert = false;
3091 SDValue Op0 = Op.getOperand(0);
3092 SDValue Op1 = Op.getOperand(1);
3093 SDValue CC = Op.getOperand(2);
3094 EVT VT = Op.getValueType();
3095 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3096 DebugLoc dl = Op.getDebugLoc();
3098 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3099 switch (SetCCOpcode) {
3100 default: llvm_unreachable("Illegal FP comparison"); break;
3102 case ISD::SETNE: Invert = true; // Fallthrough
3104 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3106 case ISD::SETLT: Swap = true; // Fallthrough
3108 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3110 case ISD::SETLE: Swap = true; // Fallthrough
3112 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3113 case ISD::SETUGE: Swap = true; // Fallthrough
3114 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3115 case ISD::SETUGT: Swap = true; // Fallthrough
3116 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3117 case ISD::SETUEQ: Invert = true; // Fallthrough
3119 // Expand this to (OLT | OGT).
3123 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3124 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3126 case ISD::SETUO: Invert = true; // Fallthrough
3128 // Expand this to (OLT | OGE).
3132 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3133 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3137 // Integer comparisons.
3138 switch (SetCCOpcode) {
3139 default: llvm_unreachable("Illegal integer comparison"); break;
3140 case ISD::SETNE: Invert = true;
3141 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3142 case ISD::SETLT: Swap = true;
3143 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3144 case ISD::SETLE: Swap = true;
3145 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3146 case ISD::SETULT: Swap = true;
3147 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3148 case ISD::SETULE: Swap = true;
3149 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3152 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3153 if (Opc == ARMISD::VCEQ) {
3156 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3158 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3161 // Ignore bitconvert.
3162 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3163 AndOp = AndOp.getOperand(0);
3165 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3167 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3168 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3175 std::swap(Op0, Op1);
3177 // If one of the operands is a constant vector zero, attempt to fold the
3178 // comparison to a specialized compare-against-zero form.
3180 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3182 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3183 if (Opc == ARMISD::VCGE)
3184 Opc = ARMISD::VCLEZ;
3185 else if (Opc == ARMISD::VCGT)
3186 Opc = ARMISD::VCLTZ;
3191 if (SingleOp.getNode()) {
3194 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3196 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3198 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3200 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3202 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3204 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3207 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3211 Result = DAG.getNOT(dl, Result, VT);
3216 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3217 /// valid vector constant for a NEON instruction with a "modified immediate"
3218 /// operand (e.g., VMOV). If so, return the encoded value.
3219 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3220 unsigned SplatBitSize, SelectionDAG &DAG,
3221 EVT &VT, bool is128Bits, NEONModImmType type) {
3222 unsigned OpCmode, Imm;
3224 // SplatBitSize is set to the smallest size that splats the vector, so a
3225 // zero vector will always have SplatBitSize == 8. However, NEON modified
3226 // immediate instructions others than VMOV do not support the 8-bit encoding
3227 // of a zero vector, and the default encoding of zero is supposed to be the
3232 switch (SplatBitSize) {
3234 if (type != VMOVModImm)
3236 // Any 1-byte value is OK. Op=0, Cmode=1110.
3237 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3240 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3244 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3245 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3246 if ((SplatBits & ~0xff) == 0) {
3247 // Value = 0x00nn: Op=x, Cmode=100x.
3252 if ((SplatBits & ~0xff00) == 0) {
3253 // Value = 0xnn00: Op=x, Cmode=101x.
3255 Imm = SplatBits >> 8;
3261 // NEON's 32-bit VMOV supports splat values where:
3262 // * only one byte is nonzero, or
3263 // * the least significant byte is 0xff and the second byte is nonzero, or
3264 // * the least significant 2 bytes are 0xff and the third is nonzero.
3265 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3266 if ((SplatBits & ~0xff) == 0) {
3267 // Value = 0x000000nn: Op=x, Cmode=000x.
3272 if ((SplatBits & ~0xff00) == 0) {
3273 // Value = 0x0000nn00: Op=x, Cmode=001x.
3275 Imm = SplatBits >> 8;
3278 if ((SplatBits & ~0xff0000) == 0) {
3279 // Value = 0x00nn0000: Op=x, Cmode=010x.
3281 Imm = SplatBits >> 16;
3284 if ((SplatBits & ~0xff000000) == 0) {
3285 // Value = 0xnn000000: Op=x, Cmode=011x.
3287 Imm = SplatBits >> 24;
3291 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3292 if (type == OtherModImm) return SDValue();
3294 if ((SplatBits & ~0xffff) == 0 &&
3295 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3296 // Value = 0x0000nnff: Op=x, Cmode=1100.
3298 Imm = SplatBits >> 8;
3303 if ((SplatBits & ~0xffffff) == 0 &&
3304 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3305 // Value = 0x00nnffff: Op=x, Cmode=1101.
3307 Imm = SplatBits >> 16;
3308 SplatBits |= 0xffff;
3312 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3313 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3314 // VMOV.I32. A (very) minor optimization would be to replicate the value
3315 // and fall through here to test for a valid 64-bit splat. But, then the
3316 // caller would also need to check and handle the change in size.
3320 if (type != VMOVModImm)
3322 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3323 uint64_t BitMask = 0xff;
3325 unsigned ImmMask = 1;
3327 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3328 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3331 } else if ((SplatBits & BitMask) != 0) {
3337 // Op=1, Cmode=1110.
3340 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3345 llvm_unreachable("unexpected size for isNEONModifiedImm");
3349 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3350 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3353 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3354 bool &ReverseVEXT, unsigned &Imm) {
3355 unsigned NumElts = VT.getVectorNumElements();
3356 ReverseVEXT = false;
3358 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3364 // If this is a VEXT shuffle, the immediate value is the index of the first
3365 // element. The other shuffle indices must be the successive elements after
3367 unsigned ExpectedElt = Imm;
3368 for (unsigned i = 1; i < NumElts; ++i) {
3369 // Increment the expected index. If it wraps around, it may still be
3370 // a VEXT but the source vectors must be swapped.
3372 if (ExpectedElt == NumElts * 2) {
3377 if (M[i] < 0) continue; // ignore UNDEF indices
3378 if (ExpectedElt != static_cast<unsigned>(M[i]))
3382 // Adjust the index value if the source operands will be swapped.
3389 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3390 /// instruction with the specified blocksize. (The order of the elements
3391 /// within each block of the vector is reversed.)
3392 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3393 unsigned BlockSize) {
3394 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3395 "Only possible block sizes for VREV are: 16, 32, 64");
3397 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3401 unsigned NumElts = VT.getVectorNumElements();
3402 unsigned BlockElts = M[0] + 1;
3403 // If the first shuffle index is UNDEF, be optimistic.
3405 BlockElts = BlockSize / EltSz;
3407 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3410 for (unsigned i = 0; i < NumElts; ++i) {
3411 if (M[i] < 0) continue; // ignore UNDEF indices
3412 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3419 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3420 unsigned &WhichResult) {
3421 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3425 unsigned NumElts = VT.getVectorNumElements();
3426 WhichResult = (M[0] == 0 ? 0 : 1);
3427 for (unsigned i = 0; i < NumElts; i += 2) {
3428 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3429 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3435 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3436 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3437 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3438 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3439 unsigned &WhichResult) {
3440 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3444 unsigned NumElts = VT.getVectorNumElements();
3445 WhichResult = (M[0] == 0 ? 0 : 1);
3446 for (unsigned i = 0; i < NumElts; i += 2) {
3447 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3448 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3454 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3455 unsigned &WhichResult) {
3456 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3460 unsigned NumElts = VT.getVectorNumElements();
3461 WhichResult = (M[0] == 0 ? 0 : 1);
3462 for (unsigned i = 0; i != NumElts; ++i) {
3463 if (M[i] < 0) continue; // ignore UNDEF indices
3464 if ((unsigned) M[i] != 2 * i + WhichResult)
3468 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3469 if (VT.is64BitVector() && EltSz == 32)
3475 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3476 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3477 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3478 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3479 unsigned &WhichResult) {
3480 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3484 unsigned Half = VT.getVectorNumElements() / 2;
3485 WhichResult = (M[0] == 0 ? 0 : 1);
3486 for (unsigned j = 0; j != 2; ++j) {
3487 unsigned Idx = WhichResult;
3488 for (unsigned i = 0; i != Half; ++i) {
3489 int MIdx = M[i + j * Half];
3490 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3496 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3497 if (VT.is64BitVector() && EltSz == 32)
3503 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3504 unsigned &WhichResult) {
3505 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3509 unsigned NumElts = VT.getVectorNumElements();
3510 WhichResult = (M[0] == 0 ? 0 : 1);
3511 unsigned Idx = WhichResult * NumElts / 2;
3512 for (unsigned i = 0; i != NumElts; i += 2) {
3513 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3514 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3519 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3520 if (VT.is64BitVector() && EltSz == 32)
3526 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3527 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3528 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3529 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3530 unsigned &WhichResult) {
3531 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3535 unsigned NumElts = VT.getVectorNumElements();
3536 WhichResult = (M[0] == 0 ? 0 : 1);
3537 unsigned Idx = WhichResult * NumElts / 2;
3538 for (unsigned i = 0; i != NumElts; i += 2) {
3539 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3540 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3545 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3546 if (VT.is64BitVector() && EltSz == 32)
3552 // If N is an integer constant that can be moved into a register in one
3553 // instruction, return an SDValue of such a constant (will become a MOV
3554 // instruction). Otherwise return null.
3555 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3556 const ARMSubtarget *ST, DebugLoc dl) {
3558 if (!isa<ConstantSDNode>(N))
3560 Val = cast<ConstantSDNode>(N)->getZExtValue();
3562 if (ST->isThumb1Only()) {
3563 if (Val <= 255 || ~Val <= 255)
3564 return DAG.getConstant(Val, MVT::i32);
3566 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3567 return DAG.getConstant(Val, MVT::i32);
3572 // If this is a case we can't handle, return null and let the default
3573 // expansion code take care of it.
3574 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3575 const ARMSubtarget *ST) const {
3576 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3577 DebugLoc dl = Op.getDebugLoc();
3578 EVT VT = Op.getValueType();
3580 APInt SplatBits, SplatUndef;
3581 unsigned SplatBitSize;
3583 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3584 if (SplatBitSize <= 64) {
3585 // Check if an immediate VMOV works.
3587 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3588 SplatUndef.getZExtValue(), SplatBitSize,
3589 DAG, VmovVT, VT.is128BitVector(),
3591 if (Val.getNode()) {
3592 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3593 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3596 // Try an immediate VMVN.
3597 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3598 ((1LL << SplatBitSize) - 1));
3599 Val = isNEONModifiedImm(NegatedImm,
3600 SplatUndef.getZExtValue(), SplatBitSize,
3601 DAG, VmovVT, VT.is128BitVector(),
3603 if (Val.getNode()) {
3604 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3605 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3610 // Scan through the operands to see if only one value is used.
3611 unsigned NumElts = VT.getVectorNumElements();
3612 bool isOnlyLowElement = true;
3613 bool usesOnlyOneValue = true;
3614 bool isConstant = true;
3616 for (unsigned i = 0; i < NumElts; ++i) {
3617 SDValue V = Op.getOperand(i);
3618 if (V.getOpcode() == ISD::UNDEF)
3621 isOnlyLowElement = false;
3622 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3625 if (!Value.getNode())
3627 else if (V != Value)
3628 usesOnlyOneValue = false;
3631 if (!Value.getNode())
3632 return DAG.getUNDEF(VT);
3634 if (isOnlyLowElement)
3635 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3637 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3639 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3640 // i32 and try again.
3641 if (usesOnlyOneValue && EltSize <= 32) {
3643 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3644 if (VT.getVectorElementType().isFloatingPoint()) {
3645 SmallVector<SDValue, 8> Ops;
3646 for (unsigned i = 0; i < NumElts; ++i)
3647 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3649 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3650 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3651 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3653 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3655 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3657 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3660 // If all elements are constants and the case above didn't get hit, fall back
3661 // to the default expansion, which will generate a load from the constant
3666 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3668 SDValue shuffle = ReconstructShuffle(Op, DAG);
3669 if (shuffle != SDValue())
3673 // Vectors with 32- or 64-bit elements can be built by directly assigning
3674 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3675 // will be legalized.
3676 if (EltSize >= 32) {
3677 // Do the expansion with floating-point types, since that is what the VFP
3678 // registers are defined to use, and since i64 is not legal.
3679 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3680 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3681 SmallVector<SDValue, 8> Ops;
3682 for (unsigned i = 0; i < NumElts; ++i)
3683 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3684 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3685 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3691 // Gather data to see if the operation can be modelled as a
3692 // shuffle in combination with VEXTs.
3693 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
3694 SelectionDAG &DAG) const {
3695 DebugLoc dl = Op.getDebugLoc();
3696 EVT VT = Op.getValueType();
3697 unsigned NumElts = VT.getVectorNumElements();
3699 SmallVector<SDValue, 2> SourceVecs;
3700 SmallVector<unsigned, 2> MinElts;
3701 SmallVector<unsigned, 2> MaxElts;
3703 for (unsigned i = 0; i < NumElts; ++i) {
3704 SDValue V = Op.getOperand(i);
3705 if (V.getOpcode() == ISD::UNDEF)
3707 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3708 // A shuffle can only come from building a vector from various
3709 // elements of other vectors.
3713 // Record this extraction against the appropriate vector if possible...
3714 SDValue SourceVec = V.getOperand(0);
3715 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3716 bool FoundSource = false;
3717 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3718 if (SourceVecs[j] == SourceVec) {
3719 if (MinElts[j] > EltNo)
3721 if (MaxElts[j] < EltNo)
3728 // Or record a new source if not...
3730 SourceVecs.push_back(SourceVec);
3731 MinElts.push_back(EltNo);
3732 MaxElts.push_back(EltNo);
3736 // Currently only do something sane when at most two source vectors
3738 if (SourceVecs.size() > 2)
3741 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3742 int VEXTOffsets[2] = {0, 0};
3744 // This loop extracts the usage patterns of the source vectors
3745 // and prepares appropriate SDValues for a shuffle if possible.
3746 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3747 if (SourceVecs[i].getValueType() == VT) {
3748 // No VEXT necessary
3749 ShuffleSrcs[i] = SourceVecs[i];
3752 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3753 // It probably isn't worth padding out a smaller vector just to
3754 // break it down again in a shuffle.
3758 // Since only 64-bit and 128-bit vectors are legal on ARM and
3759 // we've eliminated the other cases...
3760 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
3761 "unexpected vector sizes in ReconstructShuffle");
3763 if (MaxElts[i] - MinElts[i] >= NumElts) {
3764 // Span too large for a VEXT to cope
3768 if (MinElts[i] >= NumElts) {
3769 // The extraction can just take the second half
3770 VEXTOffsets[i] = NumElts;
3771 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3773 DAG.getIntPtrConstant(NumElts));
3774 } else if (MaxElts[i] < NumElts) {
3775 // The extraction can just take the first half
3777 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3779 DAG.getIntPtrConstant(0));
3781 // An actual VEXT is needed
3782 VEXTOffsets[i] = MinElts[i];
3783 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3785 DAG.getIntPtrConstant(0));
3786 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
3788 DAG.getIntPtrConstant(NumElts));
3789 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
3790 DAG.getConstant(VEXTOffsets[i], MVT::i32));
3794 SmallVector<int, 8> Mask;
3796 for (unsigned i = 0; i < NumElts; ++i) {
3797 SDValue Entry = Op.getOperand(i);
3798 if (Entry.getOpcode() == ISD::UNDEF) {
3803 SDValue ExtractVec = Entry.getOperand(0);
3804 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
3805 .getOperand(1))->getSExtValue();
3806 if (ExtractVec == SourceVecs[0]) {
3807 Mask.push_back(ExtractElt - VEXTOffsets[0]);
3809 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
3813 // Final check before we try to produce nonsense...
3814 if (isShuffleMaskLegal(Mask, VT))
3815 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
3821 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3822 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3823 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3824 /// are assumed to be legal.
3826 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3828 if (VT.getVectorNumElements() == 4 &&
3829 (VT.is128BitVector() || VT.is64BitVector())) {
3830 unsigned PFIndexes[4];
3831 for (unsigned i = 0; i != 4; ++i) {
3835 PFIndexes[i] = M[i];
3838 // Compute the index in the perfect shuffle table.
3839 unsigned PFTableIndex =
3840 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3841 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3842 unsigned Cost = (PFEntry >> 30);
3849 unsigned Imm, WhichResult;
3851 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3852 return (EltSize >= 32 ||
3853 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3854 isVREVMask(M, VT, 64) ||
3855 isVREVMask(M, VT, 32) ||
3856 isVREVMask(M, VT, 16) ||
3857 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3858 isVTRNMask(M, VT, WhichResult) ||
3859 isVUZPMask(M, VT, WhichResult) ||
3860 isVZIPMask(M, VT, WhichResult) ||
3861 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3862 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3863 isVZIP_v_undef_Mask(M, VT, WhichResult));
3866 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3867 /// the specified operations to build the shuffle.
3868 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3869 SDValue RHS, SelectionDAG &DAG,
3871 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3872 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3873 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3876 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3885 OP_VUZPL, // VUZP, left result
3886 OP_VUZPR, // VUZP, right result
3887 OP_VZIPL, // VZIP, left result
3888 OP_VZIPR, // VZIP, right result
3889 OP_VTRNL, // VTRN, left result
3890 OP_VTRNR // VTRN, right result
3893 if (OpNum == OP_COPY) {
3894 if (LHSID == (1*9+2)*9+3) return LHS;
3895 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3899 SDValue OpLHS, OpRHS;
3900 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3901 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3902 EVT VT = OpLHS.getValueType();
3905 default: llvm_unreachable("Unknown shuffle opcode!");
3907 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3912 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3913 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3917 return DAG.getNode(ARMISD::VEXT, dl, VT,
3919 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3922 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3923 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3926 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3927 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3930 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3931 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3935 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3936 SDValue V1 = Op.getOperand(0);
3937 SDValue V2 = Op.getOperand(1);
3938 DebugLoc dl = Op.getDebugLoc();
3939 EVT VT = Op.getValueType();
3940 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3941 SmallVector<int, 8> ShuffleMask;
3943 // Convert shuffles that are directly supported on NEON to target-specific
3944 // DAG nodes, instead of keeping them as shuffles and matching them again
3945 // during code selection. This is more efficient and avoids the possibility
3946 // of inconsistencies between legalization and selection.
3947 // FIXME: floating-point vectors should be canonicalized to integer vectors
3948 // of the same time so that they get CSEd properly.
3949 SVN->getMask(ShuffleMask);
3951 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3952 if (EltSize <= 32) {
3953 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3954 int Lane = SVN->getSplatIndex();
3955 // If this is undef splat, generate it via "just" vdup, if possible.
3956 if (Lane == -1) Lane = 0;
3958 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3959 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3961 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3962 DAG.getConstant(Lane, MVT::i32));
3967 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3970 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3971 DAG.getConstant(Imm, MVT::i32));
3974 if (isVREVMask(ShuffleMask, VT, 64))
3975 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3976 if (isVREVMask(ShuffleMask, VT, 32))
3977 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3978 if (isVREVMask(ShuffleMask, VT, 16))
3979 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3981 // Check for Neon shuffles that modify both input vectors in place.
3982 // If both results are used, i.e., if there are two shuffles with the same
3983 // source operands and with masks corresponding to both results of one of
3984 // these operations, DAG memoization will ensure that a single node is
3985 // used for both shuffles.
3986 unsigned WhichResult;
3987 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3988 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3989 V1, V2).getValue(WhichResult);
3990 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3991 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3992 V1, V2).getValue(WhichResult);
3993 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3994 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3995 V1, V2).getValue(WhichResult);
3997 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3998 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3999 V1, V1).getValue(WhichResult);
4000 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4001 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4002 V1, V1).getValue(WhichResult);
4003 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4004 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4005 V1, V1).getValue(WhichResult);
4008 // If the shuffle is not directly supported and it has 4 elements, use
4009 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4010 unsigned NumElts = VT.getVectorNumElements();
4012 unsigned PFIndexes[4];
4013 for (unsigned i = 0; i != 4; ++i) {
4014 if (ShuffleMask[i] < 0)
4017 PFIndexes[i] = ShuffleMask[i];
4020 // Compute the index in the perfect shuffle table.
4021 unsigned PFTableIndex =
4022 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4023 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4024 unsigned Cost = (PFEntry >> 30);
4027 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4030 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4031 if (EltSize >= 32) {
4032 // Do the expansion with floating-point types, since that is what the VFP
4033 // registers are defined to use, and since i64 is not legal.
4034 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4035 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4036 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4037 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4038 SmallVector<SDValue, 8> Ops;
4039 for (unsigned i = 0; i < NumElts; ++i) {
4040 if (ShuffleMask[i] < 0)
4041 Ops.push_back(DAG.getUNDEF(EltVT));
4043 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4044 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4045 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4048 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4049 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4055 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4056 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4057 SDValue Lane = Op.getOperand(1);
4058 if (!isa<ConstantSDNode>(Lane))
4061 SDValue Vec = Op.getOperand(0);
4062 if (Op.getValueType() == MVT::i32 &&
4063 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4064 DebugLoc dl = Op.getDebugLoc();
4065 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4071 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4072 // The only time a CONCAT_VECTORS operation can have legal types is when
4073 // two 64-bit vectors are concatenated to a 128-bit vector.
4074 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4075 "unexpected CONCAT_VECTORS");
4076 DebugLoc dl = Op.getDebugLoc();
4077 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4078 SDValue Op0 = Op.getOperand(0);
4079 SDValue Op1 = Op.getOperand(1);
4080 if (Op0.getOpcode() != ISD::UNDEF)
4081 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4082 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4083 DAG.getIntPtrConstant(0));
4084 if (Op1.getOpcode() != ISD::UNDEF)
4085 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4086 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4087 DAG.getIntPtrConstant(1));
4088 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4091 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4092 /// element has been zero/sign-extended, depending on the isSigned parameter,
4093 /// from an integer type half its size.
4094 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4096 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4097 EVT VT = N->getValueType(0);
4098 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4099 SDNode *BVN = N->getOperand(0).getNode();
4100 if (BVN->getValueType(0) != MVT::v4i32 ||
4101 BVN->getOpcode() != ISD::BUILD_VECTOR)
4103 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4104 unsigned HiElt = 1 - LoElt;
4105 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4106 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4107 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4108 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4109 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4112 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4113 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4116 if (Hi0->isNullValue() && Hi1->isNullValue())
4122 if (N->getOpcode() != ISD::BUILD_VECTOR)
4125 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4126 SDNode *Elt = N->getOperand(i).getNode();
4127 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4128 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4129 unsigned HalfSize = EltSize / 2;
4131 int64_t SExtVal = C->getSExtValue();
4132 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4135 if ((C->getZExtValue() >> HalfSize) != 0)
4146 /// isSignExtended - Check if a node is a vector value that is sign-extended
4147 /// or a constant BUILD_VECTOR with sign-extended elements.
4148 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4149 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4151 if (isExtendedBUILD_VECTOR(N, DAG, true))
4156 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4157 /// or a constant BUILD_VECTOR with zero-extended elements.
4158 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4159 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4161 if (isExtendedBUILD_VECTOR(N, DAG, false))
4166 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4167 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4168 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4169 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4170 return N->getOperand(0);
4171 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4172 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4173 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4174 LD->isNonTemporal(), LD->getAlignment());
4175 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4176 // have been legalized as a BITCAST from v4i32.
4177 if (N->getOpcode() == ISD::BITCAST) {
4178 SDNode *BVN = N->getOperand(0).getNode();
4179 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4180 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4181 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4182 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4183 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4185 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4186 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4187 EVT VT = N->getValueType(0);
4188 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4189 unsigned NumElts = VT.getVectorNumElements();
4190 MVT TruncVT = MVT::getIntegerVT(EltSize);
4191 SmallVector<SDValue, 8> Ops;
4192 for (unsigned i = 0; i != NumElts; ++i) {
4193 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4194 const APInt &CInt = C->getAPIntValue();
4195 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4197 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4198 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4201 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4202 // Multiplications are only custom-lowered for 128-bit vectors so that
4203 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4204 EVT VT = Op.getValueType();
4205 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4206 SDNode *N0 = Op.getOperand(0).getNode();
4207 SDNode *N1 = Op.getOperand(1).getNode();
4208 unsigned NewOpc = 0;
4209 if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
4210 NewOpc = ARMISD::VMULLs;
4211 else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
4212 NewOpc = ARMISD::VMULLu;
4213 else if (VT == MVT::v2i64)
4214 // Fall through to expand this. It is not legal.
4217 // Other vector multiplications are legal.
4220 // Legalize to a VMULL instruction.
4221 DebugLoc DL = Op.getDebugLoc();
4222 SDValue Op0 = SkipExtension(N0, DAG);
4223 SDValue Op1 = SkipExtension(N1, DAG);
4225 assert(Op0.getValueType().is64BitVector() &&
4226 Op1.getValueType().is64BitVector() &&
4227 "unexpected types for extended operands to VMULL");
4228 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4231 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4232 switch (Op.getOpcode()) {
4233 default: llvm_unreachable("Don't know how to custom lower this!");
4234 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4235 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4236 case ISD::GlobalAddress:
4237 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4238 LowerGlobalAddressELF(Op, DAG);
4239 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4240 case ISD::SELECT: return LowerSELECT(Op, DAG);
4241 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4242 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4243 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4244 case ISD::VASTART: return LowerVASTART(Op, DAG);
4245 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4246 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4247 case ISD::SINT_TO_FP:
4248 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4249 case ISD::FP_TO_SINT:
4250 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4251 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4252 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4253 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4254 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4255 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4256 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4257 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4258 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4260 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4263 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4264 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4265 case ISD::SRL_PARTS:
4266 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4267 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4268 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4269 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4270 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4271 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4272 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4273 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4274 case ISD::MUL: return LowerMUL(Op, DAG);
4279 /// ReplaceNodeResults - Replace the results of node with an illegal result
4280 /// type with new values built out of custom code.
4281 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4282 SmallVectorImpl<SDValue>&Results,
4283 SelectionDAG &DAG) const {
4285 switch (N->getOpcode()) {
4287 llvm_unreachable("Don't know how to custom expand this!");
4290 Res = ExpandBITCAST(N, DAG);
4294 Res = Expand64BitShift(N, DAG, Subtarget);
4298 Results.push_back(Res);
4301 //===----------------------------------------------------------------------===//
4302 // ARM Scheduler Hooks
4303 //===----------------------------------------------------------------------===//
4306 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4307 MachineBasicBlock *BB,
4308 unsigned Size) const {
4309 unsigned dest = MI->getOperand(0).getReg();
4310 unsigned ptr = MI->getOperand(1).getReg();
4311 unsigned oldval = MI->getOperand(2).getReg();
4312 unsigned newval = MI->getOperand(3).getReg();
4313 unsigned scratch = BB->getParent()->getRegInfo()
4314 .createVirtualRegister(ARM::GPRRegisterClass);
4315 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4316 DebugLoc dl = MI->getDebugLoc();
4317 bool isThumb2 = Subtarget->isThumb2();
4319 unsigned ldrOpc, strOpc;
4321 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4323 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4324 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
4327 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4328 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4331 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4332 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4336 MachineFunction *MF = BB->getParent();
4337 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4338 MachineFunction::iterator It = BB;
4339 ++It; // insert the new blocks after the current block
4341 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4342 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4343 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4344 MF->insert(It, loop1MBB);
4345 MF->insert(It, loop2MBB);
4346 MF->insert(It, exitMBB);
4348 // Transfer the remainder of BB and its successor edges to exitMBB.
4349 exitMBB->splice(exitMBB->begin(), BB,
4350 llvm::next(MachineBasicBlock::iterator(MI)),
4352 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4356 // fallthrough --> loop1MBB
4357 BB->addSuccessor(loop1MBB);
4360 // ldrex dest, [ptr]
4364 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4365 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4366 .addReg(dest).addReg(oldval));
4367 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4368 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4369 BB->addSuccessor(loop2MBB);
4370 BB->addSuccessor(exitMBB);
4373 // strex scratch, newval, [ptr]
4377 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4379 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4380 .addReg(scratch).addImm(0));
4381 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4382 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4383 BB->addSuccessor(loop1MBB);
4384 BB->addSuccessor(exitMBB);
4390 MI->eraseFromParent(); // The instruction is gone now.
4396 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4397 unsigned Size, unsigned BinOpcode) const {
4398 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4399 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4401 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4402 MachineFunction *MF = BB->getParent();
4403 MachineFunction::iterator It = BB;
4406 unsigned dest = MI->getOperand(0).getReg();
4407 unsigned ptr = MI->getOperand(1).getReg();
4408 unsigned incr = MI->getOperand(2).getReg();
4409 DebugLoc dl = MI->getDebugLoc();
4411 bool isThumb2 = Subtarget->isThumb2();
4412 unsigned ldrOpc, strOpc;
4414 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4416 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4417 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4420 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4421 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4424 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4425 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4429 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4430 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4431 MF->insert(It, loopMBB);
4432 MF->insert(It, exitMBB);
4434 // Transfer the remainder of BB and its successor edges to exitMBB.
4435 exitMBB->splice(exitMBB->begin(), BB,
4436 llvm::next(MachineBasicBlock::iterator(MI)),
4438 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4440 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4441 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4442 unsigned scratch2 = (!BinOpcode) ? incr :
4443 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4447 // fallthrough --> loopMBB
4448 BB->addSuccessor(loopMBB);
4452 // <binop> scratch2, dest, incr
4453 // strex scratch, scratch2, ptr
4456 // fallthrough --> exitMBB
4458 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4460 // operand order needs to go the other way for NAND
4461 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4462 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4463 addReg(incr).addReg(dest)).addReg(0);
4465 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4466 addReg(dest).addReg(incr)).addReg(0);
4469 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4471 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4472 .addReg(scratch).addImm(0));
4473 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4474 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4476 BB->addSuccessor(loopMBB);
4477 BB->addSuccessor(exitMBB);
4483 MI->eraseFromParent(); // The instruction is gone now.
4489 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4490 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4491 E = MBB->succ_end(); I != E; ++I)
4494 llvm_unreachable("Expecting a BB with two successors!");
4498 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4499 MachineBasicBlock *BB) const {
4500 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4501 DebugLoc dl = MI->getDebugLoc();
4502 bool isThumb2 = Subtarget->isThumb2();
4503 switch (MI->getOpcode()) {
4506 llvm_unreachable("Unexpected instr type to insert");
4508 case ARM::ATOMIC_LOAD_ADD_I8:
4509 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4510 case ARM::ATOMIC_LOAD_ADD_I16:
4511 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4512 case ARM::ATOMIC_LOAD_ADD_I32:
4513 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4515 case ARM::ATOMIC_LOAD_AND_I8:
4516 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4517 case ARM::ATOMIC_LOAD_AND_I16:
4518 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4519 case ARM::ATOMIC_LOAD_AND_I32:
4520 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4522 case ARM::ATOMIC_LOAD_OR_I8:
4523 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4524 case ARM::ATOMIC_LOAD_OR_I16:
4525 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4526 case ARM::ATOMIC_LOAD_OR_I32:
4527 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4529 case ARM::ATOMIC_LOAD_XOR_I8:
4530 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4531 case ARM::ATOMIC_LOAD_XOR_I16:
4532 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4533 case ARM::ATOMIC_LOAD_XOR_I32:
4534 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4536 case ARM::ATOMIC_LOAD_NAND_I8:
4537 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4538 case ARM::ATOMIC_LOAD_NAND_I16:
4539 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4540 case ARM::ATOMIC_LOAD_NAND_I32:
4541 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4543 case ARM::ATOMIC_LOAD_SUB_I8:
4544 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4545 case ARM::ATOMIC_LOAD_SUB_I16:
4546 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4547 case ARM::ATOMIC_LOAD_SUB_I32:
4548 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4550 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4551 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4552 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4554 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4555 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4556 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4558 case ARM::tMOVCCr_pseudo: {
4559 // To "insert" a SELECT_CC instruction, we actually have to insert the
4560 // diamond control-flow pattern. The incoming instruction knows the
4561 // destination vreg to set, the condition code register to branch on, the
4562 // true/false values to select between, and a branch opcode to use.
4563 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4564 MachineFunction::iterator It = BB;
4570 // cmpTY ccX, r1, r2
4572 // fallthrough --> copy0MBB
4573 MachineBasicBlock *thisMBB = BB;
4574 MachineFunction *F = BB->getParent();
4575 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4576 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4577 F->insert(It, copy0MBB);
4578 F->insert(It, sinkMBB);
4580 // Transfer the remainder of BB and its successor edges to sinkMBB.
4581 sinkMBB->splice(sinkMBB->begin(), BB,
4582 llvm::next(MachineBasicBlock::iterator(MI)),
4584 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4586 BB->addSuccessor(copy0MBB);
4587 BB->addSuccessor(sinkMBB);
4589 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4590 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4593 // %FalseValue = ...
4594 // # fallthrough to sinkMBB
4597 // Update machine-CFG edges
4598 BB->addSuccessor(sinkMBB);
4601 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4604 BuildMI(*BB, BB->begin(), dl,
4605 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4606 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4607 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4609 MI->eraseFromParent(); // The pseudo instruction is gone now.
4614 case ARM::BCCZi64: {
4615 // If there is an unconditional branch to the other successor, remove it.
4616 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
4618 // Compare both parts that make up the double comparison separately for
4620 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4622 unsigned LHS1 = MI->getOperand(1).getReg();
4623 unsigned LHS2 = MI->getOperand(2).getReg();
4625 AddDefaultPred(BuildMI(BB, dl,
4626 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4627 .addReg(LHS1).addImm(0));
4628 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4629 .addReg(LHS2).addImm(0)
4630 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4632 unsigned RHS1 = MI->getOperand(3).getReg();
4633 unsigned RHS2 = MI->getOperand(4).getReg();
4634 AddDefaultPred(BuildMI(BB, dl,
4635 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4636 .addReg(LHS1).addReg(RHS1));
4637 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4638 .addReg(LHS2).addReg(RHS2)
4639 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4642 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4643 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4644 if (MI->getOperand(0).getImm() == ARMCC::NE)
4645 std::swap(destMBB, exitMBB);
4647 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4648 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4649 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4652 MI->eraseFromParent(); // The pseudo instruction is gone now.
4658 //===----------------------------------------------------------------------===//
4659 // ARM Optimization Hooks
4660 //===----------------------------------------------------------------------===//
4663 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4664 TargetLowering::DAGCombinerInfo &DCI) {
4665 SelectionDAG &DAG = DCI.DAG;
4666 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4667 EVT VT = N->getValueType(0);
4668 unsigned Opc = N->getOpcode();
4669 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4670 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4671 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4672 ISD::CondCode CC = ISD::SETCC_INVALID;
4675 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4677 SDValue CCOp = Slct.getOperand(0);
4678 if (CCOp.getOpcode() == ISD::SETCC)
4679 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4682 bool DoXform = false;
4684 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4687 if (LHS.getOpcode() == ISD::Constant &&
4688 cast<ConstantSDNode>(LHS)->isNullValue()) {
4690 } else if (CC != ISD::SETCC_INVALID &&
4691 RHS.getOpcode() == ISD::Constant &&
4692 cast<ConstantSDNode>(RHS)->isNullValue()) {
4693 std::swap(LHS, RHS);
4694 SDValue Op0 = Slct.getOperand(0);
4695 EVT OpVT = isSlctCC ? Op0.getValueType() :
4696 Op0.getOperand(0).getValueType();
4697 bool isInt = OpVT.isInteger();
4698 CC = ISD::getSetCCInverse(CC, isInt);
4700 if (!TLI.isCondCodeLegal(CC, OpVT))
4701 return SDValue(); // Inverse operator isn't legal.
4708 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4710 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4711 Slct.getOperand(0), Slct.getOperand(1), CC);
4712 SDValue CCOp = Slct.getOperand(0);
4714 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4715 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4716 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4717 CCOp, OtherOp, Result);
4722 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4723 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4724 /// called with the default operands, and if that fails, with commuted
4726 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4727 TargetLowering::DAGCombinerInfo &DCI) {
4728 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4729 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4730 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4731 if (Result.getNode()) return Result;
4736 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4738 static SDValue PerformADDCombine(SDNode *N,
4739 TargetLowering::DAGCombinerInfo &DCI) {
4740 SDValue N0 = N->getOperand(0);
4741 SDValue N1 = N->getOperand(1);
4743 // First try with the default operand order.
4744 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4745 if (Result.getNode())
4748 // If that didn't work, try again with the operands commuted.
4749 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4752 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4754 static SDValue PerformSUBCombine(SDNode *N,
4755 TargetLowering::DAGCombinerInfo &DCI) {
4756 SDValue N0 = N->getOperand(0);
4757 SDValue N1 = N->getOperand(1);
4759 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4760 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4761 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4762 if (Result.getNode()) return Result;
4768 static SDValue PerformMULCombine(SDNode *N,
4769 TargetLowering::DAGCombinerInfo &DCI,
4770 const ARMSubtarget *Subtarget) {
4771 SelectionDAG &DAG = DCI.DAG;
4773 if (Subtarget->isThumb1Only())
4776 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4779 EVT VT = N->getValueType(0);
4783 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4787 uint64_t MulAmt = C->getZExtValue();
4788 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4789 ShiftAmt = ShiftAmt & (32 - 1);
4790 SDValue V = N->getOperand(0);
4791 DebugLoc DL = N->getDebugLoc();
4794 MulAmt >>= ShiftAmt;
4795 if (isPowerOf2_32(MulAmt - 1)) {
4796 // (mul x, 2^N + 1) => (add (shl x, N), x)
4797 Res = DAG.getNode(ISD::ADD, DL, VT,
4798 V, DAG.getNode(ISD::SHL, DL, VT,
4799 V, DAG.getConstant(Log2_32(MulAmt-1),
4801 } else if (isPowerOf2_32(MulAmt + 1)) {
4802 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4803 Res = DAG.getNode(ISD::SUB, DL, VT,
4804 DAG.getNode(ISD::SHL, DL, VT,
4805 V, DAG.getConstant(Log2_32(MulAmt+1),
4812 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4813 DAG.getConstant(ShiftAmt, MVT::i32));
4815 // Do not add new nodes to DAG combiner worklist.
4816 DCI.CombineTo(N, Res, false);
4820 static SDValue PerformANDCombine(SDNode *N,
4821 TargetLowering::DAGCombinerInfo &DCI) {
4822 // Attempt to use immediate-form VBIC
4823 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
4824 DebugLoc dl = N->getDebugLoc();
4825 EVT VT = N->getValueType(0);
4826 SelectionDAG &DAG = DCI.DAG;
4828 APInt SplatBits, SplatUndef;
4829 unsigned SplatBitSize;
4832 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4833 if (SplatBitSize <= 64) {
4835 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
4836 SplatUndef.getZExtValue(), SplatBitSize,
4837 DAG, VbicVT, VT.is128BitVector(),
4839 if (Val.getNode()) {
4841 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
4842 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
4843 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
4851 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4852 static SDValue PerformORCombine(SDNode *N,
4853 TargetLowering::DAGCombinerInfo &DCI,
4854 const ARMSubtarget *Subtarget) {
4855 // Attempt to use immediate-form VORR
4856 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
4857 DebugLoc dl = N->getDebugLoc();
4858 EVT VT = N->getValueType(0);
4859 SelectionDAG &DAG = DCI.DAG;
4861 APInt SplatBits, SplatUndef;
4862 unsigned SplatBitSize;
4864 if (BVN && Subtarget->hasNEON() &&
4865 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4866 if (SplatBitSize <= 64) {
4868 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4869 SplatUndef.getZExtValue(), SplatBitSize,
4870 DAG, VorrVT, VT.is128BitVector(),
4872 if (Val.getNode()) {
4874 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
4875 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
4876 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
4881 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4884 // BFI is only available on V6T2+
4885 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4888 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4889 DebugLoc DL = N->getDebugLoc();
4890 // 1) or (and A, mask), val => ARMbfi A, val, mask
4891 // iff (val & mask) == val
4893 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4894 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4895 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4896 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4897 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4898 // (i.e., copy a bitfield value into another bitfield of the same width)
4899 if (N0.getOpcode() != ISD::AND)
4905 SDValue N00 = N0.getOperand(0);
4907 // The value and the mask need to be constants so we can verify this is
4908 // actually a bitfield set. If the mask is 0xffff, we can do better
4909 // via a movt instruction, so don't use BFI in that case.
4910 SDValue MaskOp = N0.getOperand(1);
4911 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
4914 unsigned Mask = MaskC->getZExtValue();
4918 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4919 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4921 unsigned Val = N1C->getZExtValue();
4922 if ((Val & ~Mask) != Val)
4925 if (ARM::isBitFieldInvertedMask(Mask)) {
4926 Val >>= CountTrailingZeros_32(~Mask);
4928 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
4929 DAG.getConstant(Val, MVT::i32),
4930 DAG.getConstant(Mask, MVT::i32));
4932 // Do not add new nodes to DAG combiner worklist.
4933 DCI.CombineTo(N, Res, false);
4936 } else if (N1.getOpcode() == ISD::AND) {
4937 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4938 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4941 unsigned Mask2 = N11C->getZExtValue();
4943 if (ARM::isBitFieldInvertedMask(Mask) &&
4944 ARM::isBitFieldInvertedMask(~Mask2) &&
4945 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4946 // The pack halfword instruction works better for masks that fit it,
4947 // so use that when it's available.
4948 if (Subtarget->hasT2ExtractPack() &&
4949 (Mask == 0xffff || Mask == 0xffff0000))
4952 unsigned lsb = CountTrailingZeros_32(Mask2);
4953 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4954 DAG.getConstant(lsb, MVT::i32));
4955 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
4956 DAG.getConstant(Mask, MVT::i32));
4957 // Do not add new nodes to DAG combiner worklist.
4958 DCI.CombineTo(N, Res, false);
4960 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4961 ARM::isBitFieldInvertedMask(Mask2) &&
4962 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4963 // The pack halfword instruction works better for masks that fit it,
4964 // so use that when it's available.
4965 if (Subtarget->hasT2ExtractPack() &&
4966 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4969 unsigned lsb = CountTrailingZeros_32(Mask);
4970 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
4971 DAG.getConstant(lsb, MVT::i32));
4972 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4973 DAG.getConstant(Mask2, MVT::i32));
4974 // Do not add new nodes to DAG combiner worklist.
4975 DCI.CombineTo(N, Res, false);
4980 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
4981 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
4982 ARM::isBitFieldInvertedMask(~Mask)) {
4983 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
4984 // where lsb(mask) == #shamt and masked bits of B are known zero.
4985 SDValue ShAmt = N00.getOperand(1);
4986 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4987 unsigned LSB = CountTrailingZeros_32(Mask);
4991 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
4992 DAG.getConstant(~Mask, MVT::i32));
4994 // Do not add new nodes to DAG combiner worklist.
4995 DCI.CombineTo(N, Res, false);
5001 /// PerformBFICombine - (bfi A, (and B, C1), C2) -> (bfi A, B, C2) iff
5003 static SDValue PerformBFICombine(SDNode *N,
5004 TargetLowering::DAGCombinerInfo &DCI) {
5005 SDValue N1 = N->getOperand(1);
5006 if (N1.getOpcode() == ISD::AND) {
5007 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5010 unsigned Mask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
5011 unsigned Mask2 = N11C->getZExtValue();
5012 if ((Mask & Mask2) == Mask2)
5013 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
5014 N->getOperand(0), N1.getOperand(0),
5020 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
5021 /// ARMISD::VMOVRRD.
5022 static SDValue PerformVMOVRRDCombine(SDNode *N,
5023 TargetLowering::DAGCombinerInfo &DCI) {
5024 // vmovrrd(vmovdrr x, y) -> x,y
5025 SDValue InDouble = N->getOperand(0);
5026 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
5027 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
5031 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
5032 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
5033 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
5034 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
5035 SDValue Op0 = N->getOperand(0);
5036 SDValue Op1 = N->getOperand(1);
5037 if (Op0.getOpcode() == ISD::BITCAST)
5038 Op0 = Op0.getOperand(0);
5039 if (Op1.getOpcode() == ISD::BITCAST)
5040 Op1 = Op1.getOperand(0);
5041 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
5042 Op0.getNode() == Op1.getNode() &&
5043 Op0.getResNo() == 0 && Op1.getResNo() == 1)
5044 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5045 N->getValueType(0), Op0.getOperand(0));
5049 /// PerformSTORECombine - Target-specific dag combine xforms for
5051 static SDValue PerformSTORECombine(SDNode *N,
5052 TargetLowering::DAGCombinerInfo &DCI) {
5053 // Bitcast an i64 store extracted from a vector to f64.
5054 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5055 StoreSDNode *St = cast<StoreSDNode>(N);
5056 SDValue StVal = St->getValue();
5057 if (!ISD::isNormalStore(St) || St->isVolatile() ||
5058 StVal.getValueType() != MVT::i64 ||
5059 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5062 SelectionDAG &DAG = DCI.DAG;
5063 DebugLoc dl = StVal.getDebugLoc();
5064 SDValue IntVec = StVal.getOperand(0);
5065 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5066 IntVec.getValueType().getVectorNumElements());
5067 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
5068 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5069 Vec, StVal.getOperand(1));
5070 dl = N->getDebugLoc();
5071 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
5072 // Make the DAGCombiner fold the bitcasts.
5073 DCI.AddToWorklist(Vec.getNode());
5074 DCI.AddToWorklist(ExtElt.getNode());
5075 DCI.AddToWorklist(V.getNode());
5076 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
5077 St->getPointerInfo(), St->isVolatile(),
5078 St->isNonTemporal(), St->getAlignment(),
5082 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
5083 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
5084 /// i64 vector to have f64 elements, since the value can then be loaded
5085 /// directly into a VFP register.
5086 static bool hasNormalLoadOperand(SDNode *N) {
5087 unsigned NumElts = N->getValueType(0).getVectorNumElements();
5088 for (unsigned i = 0; i < NumElts; ++i) {
5089 SDNode *Elt = N->getOperand(i).getNode();
5090 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
5096 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
5097 /// ISD::BUILD_VECTOR.
5098 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
5099 TargetLowering::DAGCombinerInfo &DCI){
5100 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
5101 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
5102 // into a pair of GPRs, which is fine when the value is used as a scalar,
5103 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
5104 SelectionDAG &DAG = DCI.DAG;
5105 if (N->getNumOperands() == 2) {
5106 SDValue RV = PerformVMOVDRRCombine(N, DAG);
5111 // Load i64 elements as f64 values so that type legalization does not split
5112 // them up into i32 values.
5113 EVT VT = N->getValueType(0);
5114 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
5116 DebugLoc dl = N->getDebugLoc();
5117 SmallVector<SDValue, 8> Ops;
5118 unsigned NumElts = VT.getVectorNumElements();
5119 for (unsigned i = 0; i < NumElts; ++i) {
5120 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
5122 // Make the DAGCombiner fold the bitcast.
5123 DCI.AddToWorklist(V.getNode());
5125 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
5126 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
5127 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
5130 /// PerformInsertEltCombine - Target-specific dag combine xforms for
5131 /// ISD::INSERT_VECTOR_ELT.
5132 static SDValue PerformInsertEltCombine(SDNode *N,
5133 TargetLowering::DAGCombinerInfo &DCI) {
5134 // Bitcast an i64 load inserted into a vector to f64.
5135 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5136 EVT VT = N->getValueType(0);
5137 SDNode *Elt = N->getOperand(1).getNode();
5138 if (VT.getVectorElementType() != MVT::i64 ||
5139 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
5142 SelectionDAG &DAG = DCI.DAG;
5143 DebugLoc dl = N->getDebugLoc();
5144 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5145 VT.getVectorNumElements());
5146 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
5147 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
5148 // Make the DAGCombiner fold the bitcasts.
5149 DCI.AddToWorklist(Vec.getNode());
5150 DCI.AddToWorklist(V.getNode());
5151 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
5152 Vec, V, N->getOperand(2));
5153 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
5156 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
5157 /// ISD::VECTOR_SHUFFLE.
5158 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
5159 // The LLVM shufflevector instruction does not require the shuffle mask
5160 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
5161 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
5162 // operands do not match the mask length, they are extended by concatenating
5163 // them with undef vectors. That is probably the right thing for other
5164 // targets, but for NEON it is better to concatenate two double-register
5165 // size vector operands into a single quad-register size vector. Do that
5166 // transformation here:
5167 // shuffle(concat(v1, undef), concat(v2, undef)) ->
5168 // shuffle(concat(v1, v2), undef)
5169 SDValue Op0 = N->getOperand(0);
5170 SDValue Op1 = N->getOperand(1);
5171 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
5172 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
5173 Op0.getNumOperands() != 2 ||
5174 Op1.getNumOperands() != 2)
5176 SDValue Concat0Op1 = Op0.getOperand(1);
5177 SDValue Concat1Op1 = Op1.getOperand(1);
5178 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
5179 Concat1Op1.getOpcode() != ISD::UNDEF)
5181 // Skip the transformation if any of the types are illegal.
5182 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5183 EVT VT = N->getValueType(0);
5184 if (!TLI.isTypeLegal(VT) ||
5185 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
5186 !TLI.isTypeLegal(Concat1Op1.getValueType()))
5189 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5190 Op0.getOperand(0), Op1.getOperand(0));
5191 // Translate the shuffle mask.
5192 SmallVector<int, 16> NewMask;
5193 unsigned NumElts = VT.getVectorNumElements();
5194 unsigned HalfElts = NumElts/2;
5195 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
5196 for (unsigned n = 0; n < NumElts; ++n) {
5197 int MaskElt = SVN->getMaskElt(n);
5199 if (MaskElt < (int)HalfElts)
5201 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
5202 NewElt = HalfElts + MaskElt - NumElts;
5203 NewMask.push_back(NewElt);
5205 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
5206 DAG.getUNDEF(VT), NewMask.data());
5209 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
5210 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
5211 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
5213 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
5214 SelectionDAG &DAG = DCI.DAG;
5215 EVT VT = N->getValueType(0);
5216 // vldN-dup instructions only support 64-bit vectors for N > 1.
5217 if (!VT.is64BitVector())
5220 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
5221 SDNode *VLD = N->getOperand(0).getNode();
5222 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
5224 unsigned NumVecs = 0;
5225 unsigned NewOpc = 0;
5226 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
5227 if (IntNo == Intrinsic::arm_neon_vld2lane) {
5229 NewOpc = ARMISD::VLD2DUP;
5230 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
5232 NewOpc = ARMISD::VLD3DUP;
5233 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
5235 NewOpc = ARMISD::VLD4DUP;
5240 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
5241 // numbers match the load.
5242 unsigned VLDLaneNo =
5243 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
5244 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5246 // Ignore uses of the chain result.
5247 if (UI.getUse().getResNo() == NumVecs)
5250 if (User->getOpcode() != ARMISD::VDUPLANE ||
5251 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
5255 // Create the vldN-dup node.
5258 for (n = 0; n < NumVecs; ++n)
5260 Tys[n] = MVT::Other;
5261 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
5262 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
5263 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
5264 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
5265 Ops, 2, VLDMemInt->getMemoryVT(),
5266 VLDMemInt->getMemOperand());
5269 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5271 unsigned ResNo = UI.getUse().getResNo();
5272 // Ignore uses of the chain result.
5273 if (ResNo == NumVecs)
5276 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
5279 // Now the vldN-lane intrinsic is dead except for its chain result.
5280 // Update uses of the chain.
5281 std::vector<SDValue> VLDDupResults;
5282 for (unsigned n = 0; n < NumVecs; ++n)
5283 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
5284 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
5285 DCI.CombineTo(VLD, VLDDupResults);
5290 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
5291 /// ARMISD::VDUPLANE.
5292 static SDValue PerformVDUPLANECombine(SDNode *N,
5293 TargetLowering::DAGCombinerInfo &DCI) {
5294 SDValue Op = N->getOperand(0);
5296 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
5297 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
5298 if (CombineVLDDUP(N, DCI))
5299 return SDValue(N, 0);
5301 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
5302 // redundant. Ignore bit_converts for now; element sizes are checked below.
5303 while (Op.getOpcode() == ISD::BITCAST)
5304 Op = Op.getOperand(0);
5305 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
5308 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
5309 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
5310 // The canonical VMOV for a zero vector uses a 32-bit element size.
5311 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5313 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
5315 EVT VT = N->getValueType(0);
5316 if (EltSize > VT.getVectorElementType().getSizeInBits())
5319 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
5322 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5323 /// operand of a vector shift operation, where all the elements of the
5324 /// build_vector must have the same constant integer value.
5325 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5326 // Ignore bit_converts.
5327 while (Op.getOpcode() == ISD::BITCAST)
5328 Op = Op.getOperand(0);
5329 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5330 APInt SplatBits, SplatUndef;
5331 unsigned SplatBitSize;
5333 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5334 HasAnyUndefs, ElementBits) ||
5335 SplatBitSize > ElementBits)
5337 Cnt = SplatBits.getSExtValue();
5341 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5342 /// operand of a vector shift left operation. That value must be in the range:
5343 /// 0 <= Value < ElementBits for a left shift; or
5344 /// 0 <= Value <= ElementBits for a long left shift.
5345 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5346 assert(VT.isVector() && "vector shift count is not a vector type");
5347 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5348 if (! getVShiftImm(Op, ElementBits, Cnt))
5350 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
5353 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5354 /// operand of a vector shift right operation. For a shift opcode, the value
5355 /// is positive, but for an intrinsic the value count must be negative. The
5356 /// absolute value must be in the range:
5357 /// 1 <= |Value| <= ElementBits for a right shift; or
5358 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5359 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5361 assert(VT.isVector() && "vector shift count is not a vector type");
5362 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5363 if (! getVShiftImm(Op, ElementBits, Cnt))
5367 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
5370 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
5371 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
5372 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5375 // Don't do anything for most intrinsics.
5378 // Vector shifts: check for immediate versions and lower them.
5379 // Note: This is done during DAG combining instead of DAG legalizing because
5380 // the build_vectors for 64-bit vector element shift counts are generally
5381 // not legal, and it is hard to see their values after they get legalized to
5382 // loads from a constant pool.
5383 case Intrinsic::arm_neon_vshifts:
5384 case Intrinsic::arm_neon_vshiftu:
5385 case Intrinsic::arm_neon_vshiftls:
5386 case Intrinsic::arm_neon_vshiftlu:
5387 case Intrinsic::arm_neon_vshiftn:
5388 case Intrinsic::arm_neon_vrshifts:
5389 case Intrinsic::arm_neon_vrshiftu:
5390 case Intrinsic::arm_neon_vrshiftn:
5391 case Intrinsic::arm_neon_vqshifts:
5392 case Intrinsic::arm_neon_vqshiftu:
5393 case Intrinsic::arm_neon_vqshiftsu:
5394 case Intrinsic::arm_neon_vqshiftns:
5395 case Intrinsic::arm_neon_vqshiftnu:
5396 case Intrinsic::arm_neon_vqshiftnsu:
5397 case Intrinsic::arm_neon_vqrshiftns:
5398 case Intrinsic::arm_neon_vqrshiftnu:
5399 case Intrinsic::arm_neon_vqrshiftnsu: {
5400 EVT VT = N->getOperand(1).getValueType();
5402 unsigned VShiftOpc = 0;
5405 case Intrinsic::arm_neon_vshifts:
5406 case Intrinsic::arm_neon_vshiftu:
5407 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
5408 VShiftOpc = ARMISD::VSHL;
5411 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
5412 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
5413 ARMISD::VSHRs : ARMISD::VSHRu);
5418 case Intrinsic::arm_neon_vshiftls:
5419 case Intrinsic::arm_neon_vshiftlu:
5420 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
5422 llvm_unreachable("invalid shift count for vshll intrinsic");
5424 case Intrinsic::arm_neon_vrshifts:
5425 case Intrinsic::arm_neon_vrshiftu:
5426 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
5430 case Intrinsic::arm_neon_vqshifts:
5431 case Intrinsic::arm_neon_vqshiftu:
5432 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5436 case Intrinsic::arm_neon_vqshiftsu:
5437 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5439 llvm_unreachable("invalid shift count for vqshlu intrinsic");
5441 case Intrinsic::arm_neon_vshiftn:
5442 case Intrinsic::arm_neon_vrshiftn:
5443 case Intrinsic::arm_neon_vqshiftns:
5444 case Intrinsic::arm_neon_vqshiftnu:
5445 case Intrinsic::arm_neon_vqshiftnsu:
5446 case Intrinsic::arm_neon_vqrshiftns:
5447 case Intrinsic::arm_neon_vqrshiftnu:
5448 case Intrinsic::arm_neon_vqrshiftnsu:
5449 // Narrowing shifts require an immediate right shift.
5450 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
5452 llvm_unreachable("invalid shift count for narrowing vector shift "
5456 llvm_unreachable("unhandled vector shift");
5460 case Intrinsic::arm_neon_vshifts:
5461 case Intrinsic::arm_neon_vshiftu:
5462 // Opcode already set above.
5464 case Intrinsic::arm_neon_vshiftls:
5465 case Intrinsic::arm_neon_vshiftlu:
5466 if (Cnt == VT.getVectorElementType().getSizeInBits())
5467 VShiftOpc = ARMISD::VSHLLi;
5469 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
5470 ARMISD::VSHLLs : ARMISD::VSHLLu);
5472 case Intrinsic::arm_neon_vshiftn:
5473 VShiftOpc = ARMISD::VSHRN; break;
5474 case Intrinsic::arm_neon_vrshifts:
5475 VShiftOpc = ARMISD::VRSHRs; break;
5476 case Intrinsic::arm_neon_vrshiftu:
5477 VShiftOpc = ARMISD::VRSHRu; break;
5478 case Intrinsic::arm_neon_vrshiftn:
5479 VShiftOpc = ARMISD::VRSHRN; break;
5480 case Intrinsic::arm_neon_vqshifts:
5481 VShiftOpc = ARMISD::VQSHLs; break;
5482 case Intrinsic::arm_neon_vqshiftu:
5483 VShiftOpc = ARMISD::VQSHLu; break;
5484 case Intrinsic::arm_neon_vqshiftsu:
5485 VShiftOpc = ARMISD::VQSHLsu; break;
5486 case Intrinsic::arm_neon_vqshiftns:
5487 VShiftOpc = ARMISD::VQSHRNs; break;
5488 case Intrinsic::arm_neon_vqshiftnu:
5489 VShiftOpc = ARMISD::VQSHRNu; break;
5490 case Intrinsic::arm_neon_vqshiftnsu:
5491 VShiftOpc = ARMISD::VQSHRNsu; break;
5492 case Intrinsic::arm_neon_vqrshiftns:
5493 VShiftOpc = ARMISD::VQRSHRNs; break;
5494 case Intrinsic::arm_neon_vqrshiftnu:
5495 VShiftOpc = ARMISD::VQRSHRNu; break;
5496 case Intrinsic::arm_neon_vqrshiftnsu:
5497 VShiftOpc = ARMISD::VQRSHRNsu; break;
5500 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5501 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
5504 case Intrinsic::arm_neon_vshiftins: {
5505 EVT VT = N->getOperand(1).getValueType();
5507 unsigned VShiftOpc = 0;
5509 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
5510 VShiftOpc = ARMISD::VSLI;
5511 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
5512 VShiftOpc = ARMISD::VSRI;
5514 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
5517 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5518 N->getOperand(1), N->getOperand(2),
5519 DAG.getConstant(Cnt, MVT::i32));
5522 case Intrinsic::arm_neon_vqrshifts:
5523 case Intrinsic::arm_neon_vqrshiftu:
5524 // No immediate versions of these to check for.
5531 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
5532 /// lowers them. As with the vector shift intrinsics, this is done during DAG
5533 /// combining instead of DAG legalizing because the build_vectors for 64-bit
5534 /// vector element shift counts are generally not legal, and it is hard to see
5535 /// their values after they get legalized to loads from a constant pool.
5536 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
5537 const ARMSubtarget *ST) {
5538 EVT VT = N->getValueType(0);
5540 // Nothing to be done for scalar shifts.
5541 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5542 if (!VT.isVector() || !TLI.isTypeLegal(VT))
5545 assert(ST->hasNEON() && "unexpected vector shift");
5548 switch (N->getOpcode()) {
5549 default: llvm_unreachable("unexpected shift opcode");
5552 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
5553 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
5554 DAG.getConstant(Cnt, MVT::i32));
5559 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
5560 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
5561 ARMISD::VSHRs : ARMISD::VSHRu);
5562 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
5563 DAG.getConstant(Cnt, MVT::i32));
5569 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
5570 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
5571 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
5572 const ARMSubtarget *ST) {
5573 SDValue N0 = N->getOperand(0);
5575 // Check for sign- and zero-extensions of vector extract operations of 8-
5576 // and 16-bit vector elements. NEON supports these directly. They are
5577 // handled during DAG combining because type legalization will promote them
5578 // to 32-bit types and it is messy to recognize the operations after that.
5579 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
5580 SDValue Vec = N0.getOperand(0);
5581 SDValue Lane = N0.getOperand(1);
5582 EVT VT = N->getValueType(0);
5583 EVT EltVT = N0.getValueType();
5584 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5586 if (VT == MVT::i32 &&
5587 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
5588 TLI.isTypeLegal(Vec.getValueType()) &&
5589 isa<ConstantSDNode>(Lane)) {
5592 switch (N->getOpcode()) {
5593 default: llvm_unreachable("unexpected opcode");
5594 case ISD::SIGN_EXTEND:
5595 Opc = ARMISD::VGETLANEs;
5597 case ISD::ZERO_EXTEND:
5598 case ISD::ANY_EXTEND:
5599 Opc = ARMISD::VGETLANEu;
5602 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
5609 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
5610 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
5611 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
5612 const ARMSubtarget *ST) {
5613 // If the target supports NEON, try to use vmax/vmin instructions for f32
5614 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
5615 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
5616 // a NaN; only do the transformation when it matches that behavior.
5618 // For now only do this when using NEON for FP operations; if using VFP, it
5619 // is not obvious that the benefit outweighs the cost of switching to the
5621 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
5622 N->getValueType(0) != MVT::f32)
5625 SDValue CondLHS = N->getOperand(0);
5626 SDValue CondRHS = N->getOperand(1);
5627 SDValue LHS = N->getOperand(2);
5628 SDValue RHS = N->getOperand(3);
5629 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
5631 unsigned Opcode = 0;
5633 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
5634 IsReversed = false; // x CC y ? x : y
5635 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
5636 IsReversed = true ; // x CC y ? y : x
5650 // If LHS is NaN, an ordered comparison will be false and the result will
5651 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
5652 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5653 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
5654 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5656 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
5657 // will return -0, so vmin can only be used for unsafe math or if one of
5658 // the operands is known to be nonzero.
5659 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
5661 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5663 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
5672 // If LHS is NaN, an ordered comparison will be false and the result will
5673 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
5674 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5675 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
5676 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5678 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
5679 // will return +0, so vmax can only be used for unsafe math or if one of
5680 // the operands is known to be nonzero.
5681 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
5683 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5685 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
5691 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
5694 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
5695 DAGCombinerInfo &DCI) const {
5696 switch (N->getOpcode()) {
5698 case ISD::ADD: return PerformADDCombine(N, DCI);
5699 case ISD::SUB: return PerformSUBCombine(N, DCI);
5700 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
5701 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
5702 case ISD::AND: return PerformANDCombine(N, DCI);
5703 case ARMISD::BFI: return PerformBFICombine(N, DCI);
5704 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
5705 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
5706 case ISD::STORE: return PerformSTORECombine(N, DCI);
5707 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
5708 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
5709 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
5710 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
5711 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
5714 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
5715 case ISD::SIGN_EXTEND:
5716 case ISD::ZERO_EXTEND:
5717 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
5718 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
5723 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
5724 if (!Subtarget->allowsUnalignedMem())
5727 switch (VT.getSimpleVT().SimpleTy) {
5734 // FIXME: VLD1 etc with standard alignment is legal.
5738 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
5743 switch (VT.getSimpleVT().SimpleTy) {
5744 default: return false;
5759 if ((V & (Scale - 1)) != 0)
5762 return V == (V & ((1LL << 5) - 1));
5765 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
5766 const ARMSubtarget *Subtarget) {
5773 switch (VT.getSimpleVT().SimpleTy) {
5774 default: return false;
5779 // + imm12 or - imm8
5781 return V == (V & ((1LL << 8) - 1));
5782 return V == (V & ((1LL << 12) - 1));
5785 // Same as ARM mode. FIXME: NEON?
5786 if (!Subtarget->hasVFP2())
5791 return V == (V & ((1LL << 8) - 1));
5795 /// isLegalAddressImmediate - Return true if the integer value can be used
5796 /// as the offset of the target addressing mode for load / store of the
5798 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5799 const ARMSubtarget *Subtarget) {
5806 if (Subtarget->isThumb1Only())
5807 return isLegalT1AddressImmediate(V, VT);
5808 else if (Subtarget->isThumb2())
5809 return isLegalT2AddressImmediate(V, VT, Subtarget);
5814 switch (VT.getSimpleVT().SimpleTy) {
5815 default: return false;
5820 return V == (V & ((1LL << 12) - 1));
5823 return V == (V & ((1LL << 8) - 1));
5826 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5831 return V == (V & ((1LL << 8) - 1));
5835 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5837 int Scale = AM.Scale;
5841 switch (VT.getSimpleVT().SimpleTy) {
5842 default: return false;
5851 return Scale == 2 || Scale == 4 || Scale == 8;
5854 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5858 // Note, we allow "void" uses (basically, uses that aren't loads or
5859 // stores), because arm allows folding a scale into many arithmetic
5860 // operations. This should be made more precise and revisited later.
5862 // Allow r << imm, but the imm has to be a multiple of two.
5863 if (Scale & 1) return false;
5864 return isPowerOf2_32(Scale);
5868 /// isLegalAddressingMode - Return true if the addressing mode represented
5869 /// by AM is legal for this target, for a load/store of the specified type.
5870 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5871 const Type *Ty) const {
5872 EVT VT = getValueType(Ty, true);
5873 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5876 // Can never fold addr of global into load/store.
5881 case 0: // no scale reg, must be "r+i" or "r", or "i".
5884 if (Subtarget->isThumb1Only())
5888 // ARM doesn't support any R+R*scale+imm addr modes.
5895 if (Subtarget->isThumb2())
5896 return isLegalT2ScaledAddressingMode(AM, VT);
5898 int Scale = AM.Scale;
5899 switch (VT.getSimpleVT().SimpleTy) {
5900 default: return false;
5904 if (Scale < 0) Scale = -Scale;
5908 return isPowerOf2_32(Scale & ~1);
5912 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5917 // Note, we allow "void" uses (basically, uses that aren't loads or
5918 // stores), because arm allows folding a scale into many arithmetic
5919 // operations. This should be made more precise and revisited later.
5921 // Allow r << imm, but the imm has to be a multiple of two.
5922 if (Scale & 1) return false;
5923 return isPowerOf2_32(Scale);
5930 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5931 /// icmp immediate, that is the target has icmp instructions which can compare
5932 /// a register against the immediate without having to materialize the
5933 /// immediate into a register.
5934 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5935 if (!Subtarget->isThumb())
5936 return ARM_AM::getSOImmVal(Imm) != -1;
5937 if (Subtarget->isThumb2())
5938 return ARM_AM::getT2SOImmVal(Imm) != -1;
5939 return Imm >= 0 && Imm <= 255;
5942 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5943 bool isSEXTLoad, SDValue &Base,
5944 SDValue &Offset, bool &isInc,
5945 SelectionDAG &DAG) {
5946 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5949 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5951 Base = Ptr->getOperand(0);
5952 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5953 int RHSC = (int)RHS->getZExtValue();
5954 if (RHSC < 0 && RHSC > -256) {
5955 assert(Ptr->getOpcode() == ISD::ADD);
5957 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5961 isInc = (Ptr->getOpcode() == ISD::ADD);
5962 Offset = Ptr->getOperand(1);
5964 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5966 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5967 int RHSC = (int)RHS->getZExtValue();
5968 if (RHSC < 0 && RHSC > -0x1000) {
5969 assert(Ptr->getOpcode() == ISD::ADD);
5971 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5972 Base = Ptr->getOperand(0);
5977 if (Ptr->getOpcode() == ISD::ADD) {
5979 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5980 if (ShOpcVal != ARM_AM::no_shift) {
5981 Base = Ptr->getOperand(1);
5982 Offset = Ptr->getOperand(0);
5984 Base = Ptr->getOperand(0);
5985 Offset = Ptr->getOperand(1);
5990 isInc = (Ptr->getOpcode() == ISD::ADD);
5991 Base = Ptr->getOperand(0);
5992 Offset = Ptr->getOperand(1);
5996 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
6000 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
6001 bool isSEXTLoad, SDValue &Base,
6002 SDValue &Offset, bool &isInc,
6003 SelectionDAG &DAG) {
6004 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
6007 Base = Ptr->getOperand(0);
6008 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
6009 int RHSC = (int)RHS->getZExtValue();
6010 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
6011 assert(Ptr->getOpcode() == ISD::ADD);
6013 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
6015 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
6016 isInc = Ptr->getOpcode() == ISD::ADD;
6017 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
6025 /// getPreIndexedAddressParts - returns true by value, base pointer and
6026 /// offset pointer and addressing mode by reference if the node's address
6027 /// can be legally represented as pre-indexed load / store address.
6029 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
6031 ISD::MemIndexedMode &AM,
6032 SelectionDAG &DAG) const {
6033 if (Subtarget->isThumb1Only())
6038 bool isSEXTLoad = false;
6039 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6040 Ptr = LD->getBasePtr();
6041 VT = LD->getMemoryVT();
6042 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6043 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6044 Ptr = ST->getBasePtr();
6045 VT = ST->getMemoryVT();
6050 bool isLegal = false;
6051 if (Subtarget->isThumb2())
6052 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6053 Offset, isInc, DAG);
6055 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6056 Offset, isInc, DAG);
6060 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
6064 /// getPostIndexedAddressParts - returns true by value, base pointer and
6065 /// offset pointer and addressing mode by reference if this node can be
6066 /// combined with a load / store to form a post-indexed load / store.
6067 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
6070 ISD::MemIndexedMode &AM,
6071 SelectionDAG &DAG) const {
6072 if (Subtarget->isThumb1Only())
6077 bool isSEXTLoad = false;
6078 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6079 VT = LD->getMemoryVT();
6080 Ptr = LD->getBasePtr();
6081 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6082 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6083 VT = ST->getMemoryVT();
6084 Ptr = ST->getBasePtr();
6089 bool isLegal = false;
6090 if (Subtarget->isThumb2())
6091 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6094 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6100 // Swap base ptr and offset to catch more post-index load / store when
6101 // it's legal. In Thumb2 mode, offset must be an immediate.
6102 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
6103 !Subtarget->isThumb2())
6104 std::swap(Base, Offset);
6106 // Post-indexed load / store update the base pointer.
6111 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
6115 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6119 const SelectionDAG &DAG,
6120 unsigned Depth) const {
6121 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
6122 switch (Op.getOpcode()) {
6124 case ARMISD::CMOV: {
6125 // Bits are known zero/one if known on the LHS and RHS.
6126 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
6127 if (KnownZero == 0 && KnownOne == 0) return;
6129 APInt KnownZeroRHS, KnownOneRHS;
6130 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
6131 KnownZeroRHS, KnownOneRHS, Depth+1);
6132 KnownZero &= KnownZeroRHS;
6133 KnownOne &= KnownOneRHS;
6139 //===----------------------------------------------------------------------===//
6140 // ARM Inline Assembly Support
6141 //===----------------------------------------------------------------------===//
6143 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
6144 // Looking for "rev" which is V6+.
6145 if (!Subtarget->hasV6Ops())
6148 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
6149 std::string AsmStr = IA->getAsmString();
6150 SmallVector<StringRef, 4> AsmPieces;
6151 SplitString(AsmStr, AsmPieces, ";\n");
6153 switch (AsmPieces.size()) {
6154 default: return false;
6156 AsmStr = AsmPieces[0];
6158 SplitString(AsmStr, AsmPieces, " \t,");
6161 if (AsmPieces.size() == 3 &&
6162 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
6163 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
6164 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
6165 if (Ty && Ty->getBitWidth() == 32)
6166 return IntrinsicLowering::LowerToByteSwap(CI);
6174 /// getConstraintType - Given a constraint letter, return the type of
6175 /// constraint it is for this target.
6176 ARMTargetLowering::ConstraintType
6177 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
6178 if (Constraint.size() == 1) {
6179 switch (Constraint[0]) {
6181 case 'l': return C_RegisterClass;
6182 case 'w': return C_RegisterClass;
6185 return TargetLowering::getConstraintType(Constraint);
6188 /// Examine constraint type and operand type and determine a weight value.
6189 /// This object must already have been set up with the operand type
6190 /// and the current alternative constraint selected.
6191 TargetLowering::ConstraintWeight
6192 ARMTargetLowering::getSingleConstraintMatchWeight(
6193 AsmOperandInfo &info, const char *constraint) const {
6194 ConstraintWeight weight = CW_Invalid;
6195 Value *CallOperandVal = info.CallOperandVal;
6196 // If we don't have a value, we can't do a match,
6197 // but allow it at the lowest weight.
6198 if (CallOperandVal == NULL)
6200 const Type *type = CallOperandVal->getType();
6201 // Look at the constraint type.
6202 switch (*constraint) {
6204 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6207 if (type->isIntegerTy()) {
6208 if (Subtarget->isThumb())
6209 weight = CW_SpecificReg;
6211 weight = CW_Register;
6215 if (type->isFloatingPointTy())
6216 weight = CW_Register;
6222 std::pair<unsigned, const TargetRegisterClass*>
6223 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6225 if (Constraint.size() == 1) {
6226 // GCC ARM Constraint Letters
6227 switch (Constraint[0]) {
6229 if (Subtarget->isThumb())
6230 return std::make_pair(0U, ARM::tGPRRegisterClass);
6232 return std::make_pair(0U, ARM::GPRRegisterClass);
6234 return std::make_pair(0U, ARM::GPRRegisterClass);
6237 return std::make_pair(0U, ARM::SPRRegisterClass);
6238 if (VT.getSizeInBits() == 64)
6239 return std::make_pair(0U, ARM::DPRRegisterClass);
6240 if (VT.getSizeInBits() == 128)
6241 return std::make_pair(0U, ARM::QPRRegisterClass);
6245 if (StringRef("{cc}").equals_lower(Constraint))
6246 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
6248 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6251 std::vector<unsigned> ARMTargetLowering::
6252 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6254 if (Constraint.size() != 1)
6255 return std::vector<unsigned>();
6257 switch (Constraint[0]) { // GCC ARM Constraint Letters
6260 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6261 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6264 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6265 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6266 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
6267 ARM::R12, ARM::LR, 0);
6270 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
6271 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
6272 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
6273 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
6274 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
6275 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
6276 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
6277 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
6278 if (VT.getSizeInBits() == 64)
6279 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
6280 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
6281 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
6282 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
6283 if (VT.getSizeInBits() == 128)
6284 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
6285 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
6289 return std::vector<unsigned>();
6292 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6293 /// vector. If it is invalid, don't add anything to Ops.
6294 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6296 std::vector<SDValue>&Ops,
6297 SelectionDAG &DAG) const {
6298 SDValue Result(0, 0);
6300 switch (Constraint) {
6302 case 'I': case 'J': case 'K': case 'L':
6303 case 'M': case 'N': case 'O':
6304 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
6308 int64_t CVal64 = C->getSExtValue();
6309 int CVal = (int) CVal64;
6310 // None of these constraints allow values larger than 32 bits. Check
6311 // that the value fits in an int.
6315 switch (Constraint) {
6317 if (Subtarget->isThumb1Only()) {
6318 // This must be a constant between 0 and 255, for ADD
6320 if (CVal >= 0 && CVal <= 255)
6322 } else if (Subtarget->isThumb2()) {
6323 // A constant that can be used as an immediate value in a
6324 // data-processing instruction.
6325 if (ARM_AM::getT2SOImmVal(CVal) != -1)
6328 // A constant that can be used as an immediate value in a
6329 // data-processing instruction.
6330 if (ARM_AM::getSOImmVal(CVal) != -1)
6336 if (Subtarget->isThumb()) { // FIXME thumb2
6337 // This must be a constant between -255 and -1, for negated ADD
6338 // immediates. This can be used in GCC with an "n" modifier that
6339 // prints the negated value, for use with SUB instructions. It is
6340 // not useful otherwise but is implemented for compatibility.
6341 if (CVal >= -255 && CVal <= -1)
6344 // This must be a constant between -4095 and 4095. It is not clear
6345 // what this constraint is intended for. Implemented for
6346 // compatibility with GCC.
6347 if (CVal >= -4095 && CVal <= 4095)
6353 if (Subtarget->isThumb1Only()) {
6354 // A 32-bit value where only one byte has a nonzero value. Exclude
6355 // zero to match GCC. This constraint is used by GCC internally for
6356 // constants that can be loaded with a move/shift combination.
6357 // It is not useful otherwise but is implemented for compatibility.
6358 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
6360 } else if (Subtarget->isThumb2()) {
6361 // A constant whose bitwise inverse can be used as an immediate
6362 // value in a data-processing instruction. This can be used in GCC
6363 // with a "B" modifier that prints the inverted value, for use with
6364 // BIC and MVN instructions. It is not useful otherwise but is
6365 // implemented for compatibility.
6366 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
6369 // A constant whose bitwise inverse can be used as an immediate
6370 // value in a data-processing instruction. This can be used in GCC
6371 // with a "B" modifier that prints the inverted value, for use with
6372 // BIC and MVN instructions. It is not useful otherwise but is
6373 // implemented for compatibility.
6374 if (ARM_AM::getSOImmVal(~CVal) != -1)
6380 if (Subtarget->isThumb1Only()) {
6381 // This must be a constant between -7 and 7,
6382 // for 3-operand ADD/SUB immediate instructions.
6383 if (CVal >= -7 && CVal < 7)
6385 } else if (Subtarget->isThumb2()) {
6386 // A constant whose negation can be used as an immediate value in a
6387 // data-processing instruction. This can be used in GCC with an "n"
6388 // modifier that prints the negated value, for use with SUB
6389 // instructions. It is not useful otherwise but is implemented for
6391 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
6394 // A constant whose negation can be used as an immediate value in a
6395 // data-processing instruction. This can be used in GCC with an "n"
6396 // modifier that prints the negated value, for use with SUB
6397 // instructions. It is not useful otherwise but is implemented for
6399 if (ARM_AM::getSOImmVal(-CVal) != -1)
6405 if (Subtarget->isThumb()) { // FIXME thumb2
6406 // This must be a multiple of 4 between 0 and 1020, for
6407 // ADD sp + immediate.
6408 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
6411 // A power of two or a constant between 0 and 32. This is used in
6412 // GCC for the shift amount on shifted register operands, but it is
6413 // useful in general for any shift amounts.
6414 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
6420 if (Subtarget->isThumb()) { // FIXME thumb2
6421 // This must be a constant between 0 and 31, for shift amounts.
6422 if (CVal >= 0 && CVal <= 31)
6428 if (Subtarget->isThumb()) { // FIXME thumb2
6429 // This must be a multiple of 4 between -508 and 508, for
6430 // ADD/SUB sp = sp + immediate.
6431 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
6436 Result = DAG.getTargetConstant(CVal, Op.getValueType());
6440 if (Result.getNode()) {
6441 Ops.push_back(Result);
6444 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6448 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6449 // The ARM target isn't yet aware of offsets.
6453 int ARM::getVFPf32Imm(const APFloat &FPImm) {
6454 APInt Imm = FPImm.bitcastToAPInt();
6455 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
6456 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
6457 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
6459 // We can handle 4 bits of mantissa.
6460 // mantissa = (16+UInt(e:f:g:h))/16.
6461 if (Mantissa & 0x7ffff)
6464 if ((Mantissa & 0xf) != Mantissa)
6467 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6468 if (Exp < -3 || Exp > 4)
6470 Exp = ((Exp+3) & 0x7) ^ 4;
6472 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6475 int ARM::getVFPf64Imm(const APFloat &FPImm) {
6476 APInt Imm = FPImm.bitcastToAPInt();
6477 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
6478 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
6479 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
6481 // We can handle 4 bits of mantissa.
6482 // mantissa = (16+UInt(e:f:g:h))/16.
6483 if (Mantissa & 0xffffffffffffLL)
6486 if ((Mantissa & 0xf) != Mantissa)
6489 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6490 if (Exp < -3 || Exp > 4)
6492 Exp = ((Exp+3) & 0x7) ^ 4;
6494 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6497 bool ARM::isBitFieldInvertedMask(unsigned v) {
6498 if (v == 0xffffffff)
6500 // there can be 1's on either or both "outsides", all the "inside"
6502 unsigned int lsb = 0, msb = 31;
6503 while (v & (1 << msb)) --msb;
6504 while (v & (1 << lsb)) ++lsb;
6505 for (unsigned int i = lsb; i <= msb; ++i) {
6512 /// isFPImmLegal - Returns true if the target can instruction select the
6513 /// specified FP immediate natively. If false, the legalizer will
6514 /// materialize the FP immediate as a load from a constant pool.
6515 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
6516 if (!Subtarget->hasVFP3())
6519 return ARM::getVFPf32Imm(Imm) != -1;
6521 return ARM::getVFPf64Imm(Imm) != -1;
6525 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6526 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6527 /// specified in the intrinsic calls.
6528 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6530 unsigned Intrinsic) const {
6531 switch (Intrinsic) {
6532 case Intrinsic::arm_neon_vld1:
6533 case Intrinsic::arm_neon_vld2:
6534 case Intrinsic::arm_neon_vld3:
6535 case Intrinsic::arm_neon_vld4:
6536 case Intrinsic::arm_neon_vld2lane:
6537 case Intrinsic::arm_neon_vld3lane:
6538 case Intrinsic::arm_neon_vld4lane: {
6539 Info.opc = ISD::INTRINSIC_W_CHAIN;
6540 // Conservatively set memVT to the entire set of vectors loaded.
6541 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
6542 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6543 Info.ptrVal = I.getArgOperand(0);
6545 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6546 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6547 Info.vol = false; // volatile loads with NEON intrinsics not supported
6548 Info.readMem = true;
6549 Info.writeMem = false;
6552 case Intrinsic::arm_neon_vst1:
6553 case Intrinsic::arm_neon_vst2:
6554 case Intrinsic::arm_neon_vst3:
6555 case Intrinsic::arm_neon_vst4:
6556 case Intrinsic::arm_neon_vst2lane:
6557 case Intrinsic::arm_neon_vst3lane:
6558 case Intrinsic::arm_neon_vst4lane: {
6559 Info.opc = ISD::INTRINSIC_VOID;
6560 // Conservatively set memVT to the entire set of vectors stored.
6561 unsigned NumElts = 0;
6562 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6563 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
6564 if (!ArgTy->isVectorTy())
6566 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
6568 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6569 Info.ptrVal = I.getArgOperand(0);
6571 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6572 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6573 Info.vol = false; // volatile stores with NEON intrinsics not supported
6574 Info.readMem = false;
6575 Info.writeMem = true;