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");
58 // This option should go away when tail calls fully work.
60 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
61 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
65 EnableARMLongCalls("arm-long-calls", cl::Hidden,
66 cl::desc("Generate calls via indirect call instructions"),
70 ARMInterworking("arm-interworking", cl::Hidden,
71 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
75 EVT PromotedBitwiseVT) {
76 if (VT != PromotedLdStVT) {
77 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
78 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
79 PromotedLdStVT.getSimpleVT());
81 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
82 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
83 PromotedLdStVT.getSimpleVT());
86 EVT ElemTy = VT.getVectorElementType();
87 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
88 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
89 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
90 if (ElemTy != MVT::i32) {
91 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
94 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
96 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
97 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
98 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
99 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
100 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
101 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
102 if (VT.isInteger()) {
103 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
104 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
105 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
106 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
107 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
108 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
109 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
110 setTruncStoreAction(VT.getSimpleVT(),
111 (MVT::SimpleValueType)InnerVT, Expand);
113 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
115 // Promote all bit-wise operations.
116 if (VT.isInteger() && VT != PromotedBitwiseVT) {
117 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
118 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
119 PromotedBitwiseVT.getSimpleVT());
120 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
121 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
122 PromotedBitwiseVT.getSimpleVT());
123 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
124 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
125 PromotedBitwiseVT.getSimpleVT());
128 // Neon does not support vector divide/remainder operations.
129 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
130 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
131 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
132 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
133 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
134 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
137 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
138 addRegisterClass(VT, ARM::DPRRegisterClass);
139 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
142 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
143 addRegisterClass(VT, ARM::QPRRegisterClass);
144 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
147 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
148 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
149 return new TargetLoweringObjectFileMachO();
151 return new ARMElfTargetObjectFile();
154 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
155 : TargetLowering(TM, createTLOF(TM)) {
156 Subtarget = &TM.getSubtarget<ARMSubtarget>();
157 RegInfo = TM.getRegisterInfo();
158 Itins = TM.getInstrItineraryData();
160 if (Subtarget->isTargetDarwin()) {
161 // Uses VFP for Thumb libfuncs if available.
162 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
163 // Single-precision floating-point arithmetic.
164 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
165 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
166 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
167 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
169 // Double-precision floating-point arithmetic.
170 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
171 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
172 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
173 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
175 // Single-precision comparisons.
176 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
177 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
178 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
179 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
180 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
181 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
182 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
183 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
185 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
194 // Double-precision comparisons.
195 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
196 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
197 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
198 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
199 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
200 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
201 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
202 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
204 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
213 // Floating-point to integer conversions.
214 // i64 conversions are done via library routines even when generating VFP
215 // instructions, so use the same ones.
216 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
217 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
218 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
219 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
221 // Conversions between floating types.
222 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
223 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
225 // Integer to floating-point conversions.
226 // i64 conversions are done via library routines even when generating VFP
227 // instructions, so use the same ones.
228 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
229 // e.g., __floatunsidf vs. __floatunssidfvfp.
230 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
231 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
232 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
233 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
237 // These libcalls are not available in 32-bit.
238 setLibcallName(RTLIB::SHL_I128, 0);
239 setLibcallName(RTLIB::SRL_I128, 0);
240 setLibcallName(RTLIB::SRA_I128, 0);
242 if (Subtarget->isAAPCS_ABI()) {
243 // Double-precision floating-point arithmetic helper functions
244 // RTABI chapter 4.1.2, Table 2
245 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
246 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
247 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
248 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
249 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
250 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
251 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
252 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
254 // Double-precision floating-point comparison helper functions
255 // RTABI chapter 4.1.2, Table 3
256 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
257 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
258 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
259 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
260 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
261 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
262 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
263 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
264 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
265 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
266 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
267 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
268 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
269 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
270 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
271 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
272 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
275 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
276 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
277 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
278 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
279 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
281 // Single-precision floating-point arithmetic helper functions
282 // RTABI chapter 4.1.2, Table 4
283 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
284 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
285 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
286 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
287 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
288 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
289 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
292 // Single-precision floating-point comparison helper functions
293 // RTABI chapter 4.1.2, Table 5
294 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
295 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
296 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
297 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
298 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
299 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
300 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
301 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
302 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
303 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
304 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
305 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
306 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
307 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
308 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
309 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
310 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
313 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
314 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
315 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
316 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
317 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
319 // Floating-point to integer conversions.
320 // RTABI chapter 4.1.2, Table 6
321 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
322 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
323 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
324 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
325 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
326 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
327 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
328 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
329 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
338 // Conversions between floating types.
339 // RTABI chapter 4.1.2, Table 7
340 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
341 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
342 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
343 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
345 // Integer to floating-point conversions.
346 // RTABI chapter 4.1.2, Table 8
347 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
348 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
349 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
350 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
351 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
352 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
353 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
354 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
355 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
361 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
362 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
364 // Long long helper functions
365 // RTABI chapter 4.2, Table 9
366 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
367 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
368 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
369 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
370 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
371 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
372 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
379 // Integer division functions
380 // RTABI chapter 4.3.1
381 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
382 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
383 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
384 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
385 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
386 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
387 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
395 if (Subtarget->isThumb1Only())
396 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
398 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
399 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
400 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
401 if (!Subtarget->isFPOnlySP())
402 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
404 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
407 if (Subtarget->hasNEON()) {
408 addDRTypeForNEON(MVT::v2f32);
409 addDRTypeForNEON(MVT::v8i8);
410 addDRTypeForNEON(MVT::v4i16);
411 addDRTypeForNEON(MVT::v2i32);
412 addDRTypeForNEON(MVT::v1i64);
414 addQRTypeForNEON(MVT::v4f32);
415 addQRTypeForNEON(MVT::v2f64);
416 addQRTypeForNEON(MVT::v16i8);
417 addQRTypeForNEON(MVT::v8i16);
418 addQRTypeForNEON(MVT::v4i32);
419 addQRTypeForNEON(MVT::v2i64);
421 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
422 // neither Neon nor VFP support any arithmetic operations on it.
423 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
424 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
425 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
426 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
427 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
428 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
429 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
430 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
431 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
432 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
433 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
434 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
435 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
436 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
437 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
438 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
439 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
440 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
441 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
442 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
443 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
444 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
445 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
446 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
448 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
450 // Neon does not support some operations on v1i64 and v2i64 types.
451 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
452 // Custom handling for some quad-vector types to detect VMULL.
453 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
454 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
455 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
456 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
457 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
459 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
460 setTargetDAGCombine(ISD::SHL);
461 setTargetDAGCombine(ISD::SRL);
462 setTargetDAGCombine(ISD::SRA);
463 setTargetDAGCombine(ISD::SIGN_EXTEND);
464 setTargetDAGCombine(ISD::ZERO_EXTEND);
465 setTargetDAGCombine(ISD::ANY_EXTEND);
466 setTargetDAGCombine(ISD::SELECT_CC);
467 setTargetDAGCombine(ISD::BUILD_VECTOR);
468 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
469 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
470 setTargetDAGCombine(ISD::STORE);
473 computeRegisterProperties();
475 // ARM does not have f32 extending load.
476 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
478 // ARM does not have i1 sign extending load.
479 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
481 // ARM supports all 4 flavors of integer indexed load / store.
482 if (!Subtarget->isThumb1Only()) {
483 for (unsigned im = (unsigned)ISD::PRE_INC;
484 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
485 setIndexedLoadAction(im, MVT::i1, Legal);
486 setIndexedLoadAction(im, MVT::i8, Legal);
487 setIndexedLoadAction(im, MVT::i16, Legal);
488 setIndexedLoadAction(im, MVT::i32, Legal);
489 setIndexedStoreAction(im, MVT::i1, Legal);
490 setIndexedStoreAction(im, MVT::i8, Legal);
491 setIndexedStoreAction(im, MVT::i16, Legal);
492 setIndexedStoreAction(im, MVT::i32, Legal);
496 // i64 operation support.
497 if (Subtarget->isThumb1Only()) {
498 setOperationAction(ISD::MUL, MVT::i64, Expand);
499 setOperationAction(ISD::MULHU, MVT::i32, Expand);
500 setOperationAction(ISD::MULHS, MVT::i32, Expand);
501 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
502 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
504 setOperationAction(ISD::MUL, MVT::i64, Expand);
505 setOperationAction(ISD::MULHU, MVT::i32, Expand);
506 if (!Subtarget->hasV6Ops())
507 setOperationAction(ISD::MULHS, MVT::i32, Expand);
509 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
510 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
511 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
512 setOperationAction(ISD::SRL, MVT::i64, Custom);
513 setOperationAction(ISD::SRA, MVT::i64, Custom);
515 // ARM does not have ROTL.
516 setOperationAction(ISD::ROTL, MVT::i32, Expand);
517 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
518 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
519 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
520 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
522 // Only ARMv6 has BSWAP.
523 if (!Subtarget->hasV6Ops())
524 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
526 // These are expanded into libcalls.
527 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
528 // v7M has a hardware divider
529 setOperationAction(ISD::SDIV, MVT::i32, Expand);
530 setOperationAction(ISD::UDIV, MVT::i32, Expand);
532 setOperationAction(ISD::SREM, MVT::i32, Expand);
533 setOperationAction(ISD::UREM, MVT::i32, Expand);
534 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
535 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
537 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
538 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
539 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
540 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
541 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
543 setOperationAction(ISD::TRAP, MVT::Other, Legal);
545 // Use the default implementation.
546 setOperationAction(ISD::VASTART, MVT::Other, Custom);
547 setOperationAction(ISD::VAARG, MVT::Other, Expand);
548 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
549 setOperationAction(ISD::VAEND, MVT::Other, Expand);
550 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
551 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
552 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
553 // FIXME: Shouldn't need this, since no register is used, but the legalizer
554 // doesn't yet know how to not do that for SjLj.
555 setExceptionSelectorRegister(ARM::R0);
556 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
557 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
558 // the default expansion.
559 if (Subtarget->hasDataBarrier() ||
560 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
561 // membarrier needs custom lowering; the rest are legal and handled
563 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
565 // Set them all for expansion, which will force libcalls.
566 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
567 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
568 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
569 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
570 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
571 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
572 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
573 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
584 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
585 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
586 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
587 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
588 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
589 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
590 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
591 // Since the libcalls include locking, fold in the fences
592 setShouldFoldAtomicFences(true);
594 // 64-bit versions are always libcalls (for now)
595 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
596 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
600 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
601 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
602 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
604 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
606 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
607 if (!Subtarget->hasV6Ops()) {
608 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
609 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
611 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
613 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
614 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
615 // iff target supports vfp2.
616 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
617 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
620 // We want to custom lower some of our intrinsics.
621 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
622 if (Subtarget->isTargetDarwin()) {
623 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
624 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
625 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
628 setOperationAction(ISD::SETCC, MVT::i32, Expand);
629 setOperationAction(ISD::SETCC, MVT::f32, Expand);
630 setOperationAction(ISD::SETCC, MVT::f64, Expand);
631 setOperationAction(ISD::SELECT, MVT::i32, Custom);
632 setOperationAction(ISD::SELECT, MVT::f32, Custom);
633 setOperationAction(ISD::SELECT, MVT::f64, Custom);
634 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
635 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
636 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
638 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
639 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
640 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
641 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
642 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
644 // We don't support sin/cos/fmod/copysign/pow
645 setOperationAction(ISD::FSIN, MVT::f64, Expand);
646 setOperationAction(ISD::FSIN, MVT::f32, Expand);
647 setOperationAction(ISD::FCOS, MVT::f32, Expand);
648 setOperationAction(ISD::FCOS, MVT::f64, Expand);
649 setOperationAction(ISD::FREM, MVT::f64, Expand);
650 setOperationAction(ISD::FREM, MVT::f32, Expand);
651 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
652 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
653 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
655 setOperationAction(ISD::FPOW, MVT::f64, Expand);
656 setOperationAction(ISD::FPOW, MVT::f32, Expand);
658 // Various VFP goodness
659 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
660 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
661 if (Subtarget->hasVFP2()) {
662 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
663 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
664 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
665 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
667 // Special handling for half-precision FP.
668 if (!Subtarget->hasFP16()) {
669 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
670 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
674 // We have target-specific dag combine patterns for the following nodes:
675 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
676 setTargetDAGCombine(ISD::ADD);
677 setTargetDAGCombine(ISD::SUB);
678 setTargetDAGCombine(ISD::MUL);
680 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
681 setTargetDAGCombine(ISD::OR);
682 if (Subtarget->hasNEON())
683 setTargetDAGCombine(ISD::AND);
685 setStackPointerRegisterToSaveRestore(ARM::SP);
687 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
688 setSchedulingPreference(Sched::RegPressure);
690 setSchedulingPreference(Sched::Hybrid);
692 //// temporary - rewrite interface to use type
693 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
695 // On ARM arguments smaller than 4 bytes are extended, so all arguments
696 // are at least 4 bytes aligned.
697 setMinStackArgumentAlignment(4);
699 benefitFromCodePlacementOpt = true;
702 std::pair<const TargetRegisterClass*, uint8_t>
703 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
704 const TargetRegisterClass *RRC = 0;
706 switch (VT.getSimpleVT().SimpleTy) {
708 return TargetLowering::findRepresentativeClass(VT);
709 // Use DPR as representative register class for all floating point
710 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
711 // the cost is 1 for both f32 and f64.
712 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
713 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
714 RRC = ARM::DPRRegisterClass;
716 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
717 case MVT::v4f32: case MVT::v2f64:
718 RRC = ARM::DPRRegisterClass;
722 RRC = ARM::DPRRegisterClass;
726 RRC = ARM::DPRRegisterClass;
730 return std::make_pair(RRC, Cost);
733 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
736 case ARMISD::Wrapper: return "ARMISD::Wrapper";
737 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
738 case ARMISD::CALL: return "ARMISD::CALL";
739 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
740 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
741 case ARMISD::tCALL: return "ARMISD::tCALL";
742 case ARMISD::BRCOND: return "ARMISD::BRCOND";
743 case ARMISD::BR_JT: return "ARMISD::BR_JT";
744 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
745 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
746 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
747 case ARMISD::CMP: return "ARMISD::CMP";
748 case ARMISD::CMPZ: return "ARMISD::CMPZ";
749 case ARMISD::CMPFP: return "ARMISD::CMPFP";
750 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
751 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
752 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
753 case ARMISD::CMOV: return "ARMISD::CMOV";
754 case ARMISD::CNEG: return "ARMISD::CNEG";
756 case ARMISD::RBIT: return "ARMISD::RBIT";
758 case ARMISD::FTOSI: return "ARMISD::FTOSI";
759 case ARMISD::FTOUI: return "ARMISD::FTOUI";
760 case ARMISD::SITOF: return "ARMISD::SITOF";
761 case ARMISD::UITOF: return "ARMISD::UITOF";
763 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
764 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
765 case ARMISD::RRX: return "ARMISD::RRX";
767 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
768 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
770 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
771 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
772 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
774 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
776 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
778 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
780 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
781 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
783 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
785 case ARMISD::VCEQ: return "ARMISD::VCEQ";
786 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
787 case ARMISD::VCGE: return "ARMISD::VCGE";
788 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
789 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
790 case ARMISD::VCGEU: return "ARMISD::VCGEU";
791 case ARMISD::VCGT: return "ARMISD::VCGT";
792 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
793 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
794 case ARMISD::VCGTU: return "ARMISD::VCGTU";
795 case ARMISD::VTST: return "ARMISD::VTST";
797 case ARMISD::VSHL: return "ARMISD::VSHL";
798 case ARMISD::VSHRs: return "ARMISD::VSHRs";
799 case ARMISD::VSHRu: return "ARMISD::VSHRu";
800 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
801 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
802 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
803 case ARMISD::VSHRN: return "ARMISD::VSHRN";
804 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
805 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
806 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
807 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
808 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
809 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
810 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
811 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
812 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
813 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
814 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
815 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
816 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
817 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
818 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
819 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
820 case ARMISD::VDUP: return "ARMISD::VDUP";
821 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
822 case ARMISD::VEXT: return "ARMISD::VEXT";
823 case ARMISD::VREV64: return "ARMISD::VREV64";
824 case ARMISD::VREV32: return "ARMISD::VREV32";
825 case ARMISD::VREV16: return "ARMISD::VREV16";
826 case ARMISD::VZIP: return "ARMISD::VZIP";
827 case ARMISD::VUZP: return "ARMISD::VUZP";
828 case ARMISD::VTRN: return "ARMISD::VTRN";
829 case ARMISD::VMULLs: return "ARMISD::VMULLs";
830 case ARMISD::VMULLu: return "ARMISD::VMULLu";
831 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
832 case ARMISD::FMAX: return "ARMISD::FMAX";
833 case ARMISD::FMIN: return "ARMISD::FMIN";
834 case ARMISD::BFI: return "ARMISD::BFI";
835 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
836 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
837 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
838 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
839 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
843 /// getRegClassFor - Return the register class that should be used for the
844 /// specified value type.
845 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
846 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
847 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
848 // load / store 4 to 8 consecutive D registers.
849 if (Subtarget->hasNEON()) {
850 if (VT == MVT::v4i64)
851 return ARM::QQPRRegisterClass;
852 else if (VT == MVT::v8i64)
853 return ARM::QQQQPRRegisterClass;
855 return TargetLowering::getRegClassFor(VT);
858 // Create a fast isel object.
860 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
861 return ARM::createFastISel(funcInfo);
864 /// getFunctionAlignment - Return the Log2 alignment of this function.
865 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
866 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
869 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
870 /// be used for loads / stores from the global.
871 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
872 return (Subtarget->isThumb1Only() ? 127 : 4095);
875 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
876 unsigned NumVals = N->getNumValues();
878 return Sched::RegPressure;
880 for (unsigned i = 0; i != NumVals; ++i) {
881 EVT VT = N->getValueType(i);
882 if (VT == MVT::Glue || VT == MVT::Other)
884 if (VT.isFloatingPoint() || VT.isVector())
885 return Sched::Latency;
888 if (!N->isMachineOpcode())
889 return Sched::RegPressure;
891 // Load are scheduled for latency even if there instruction itinerary
893 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
894 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
896 if (TID.getNumDefs() == 0)
897 return Sched::RegPressure;
898 if (!Itins->isEmpty() &&
899 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
900 return Sched::Latency;
902 return Sched::RegPressure;
906 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
907 MachineFunction &MF) const {
908 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo();
910 switch (RC->getID()) {
913 case ARM::tGPRRegClassID:
914 return TFI->hasFP(MF) ? 4 : 5;
915 case ARM::GPRRegClassID: {
916 unsigned FP = TFI->hasFP(MF) ? 1 : 0;
917 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
919 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
920 case ARM::DPRRegClassID:
925 //===----------------------------------------------------------------------===//
927 //===----------------------------------------------------------------------===//
929 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
930 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
932 default: llvm_unreachable("Unknown condition code!");
933 case ISD::SETNE: return ARMCC::NE;
934 case ISD::SETEQ: return ARMCC::EQ;
935 case ISD::SETGT: return ARMCC::GT;
936 case ISD::SETGE: return ARMCC::GE;
937 case ISD::SETLT: return ARMCC::LT;
938 case ISD::SETLE: return ARMCC::LE;
939 case ISD::SETUGT: return ARMCC::HI;
940 case ISD::SETUGE: return ARMCC::HS;
941 case ISD::SETULT: return ARMCC::LO;
942 case ISD::SETULE: return ARMCC::LS;
946 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
947 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
948 ARMCC::CondCodes &CondCode2) {
949 CondCode2 = ARMCC::AL;
951 default: llvm_unreachable("Unknown FP condition!");
953 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
955 case ISD::SETOGT: CondCode = ARMCC::GT; break;
957 case ISD::SETOGE: CondCode = ARMCC::GE; break;
958 case ISD::SETOLT: CondCode = ARMCC::MI; break;
959 case ISD::SETOLE: CondCode = ARMCC::LS; break;
960 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
961 case ISD::SETO: CondCode = ARMCC::VC; break;
962 case ISD::SETUO: CondCode = ARMCC::VS; break;
963 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
964 case ISD::SETUGT: CondCode = ARMCC::HI; break;
965 case ISD::SETUGE: CondCode = ARMCC::PL; break;
967 case ISD::SETULT: CondCode = ARMCC::LT; break;
969 case ISD::SETULE: CondCode = ARMCC::LE; break;
971 case ISD::SETUNE: CondCode = ARMCC::NE; break;
975 //===----------------------------------------------------------------------===//
976 // Calling Convention Implementation
977 //===----------------------------------------------------------------------===//
979 #include "ARMGenCallingConv.inc"
981 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
982 /// given CallingConvention value.
983 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
985 bool isVarArg) const {
988 llvm_unreachable("Unsupported calling convention");
989 case CallingConv::Fast:
990 if (Subtarget->hasVFP2() && !isVarArg) {
991 if (!Subtarget->isAAPCS_ABI())
992 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
993 // For AAPCS ABI targets, just use VFP variant of the calling convention.
994 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
997 case CallingConv::C: {
998 // Use target triple & subtarget features to do actual dispatch.
999 if (!Subtarget->isAAPCS_ABI())
1000 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1001 else if (Subtarget->hasVFP2() &&
1002 FloatABIType == FloatABI::Hard && !isVarArg)
1003 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1004 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1006 case CallingConv::ARM_AAPCS_VFP:
1007 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1008 case CallingConv::ARM_AAPCS:
1009 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1010 case CallingConv::ARM_APCS:
1011 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1015 /// LowerCallResult - Lower the result values of a call into the
1016 /// appropriate copies out of appropriate physical registers.
1018 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1019 CallingConv::ID CallConv, bool isVarArg,
1020 const SmallVectorImpl<ISD::InputArg> &Ins,
1021 DebugLoc dl, SelectionDAG &DAG,
1022 SmallVectorImpl<SDValue> &InVals) const {
1024 // Assign locations to each value returned by this call.
1025 SmallVector<CCValAssign, 16> RVLocs;
1026 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1027 RVLocs, *DAG.getContext());
1028 CCInfo.AnalyzeCallResult(Ins,
1029 CCAssignFnForNode(CallConv, /* Return*/ true,
1032 // Copy all of the result registers out of their specified physreg.
1033 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1034 CCValAssign VA = RVLocs[i];
1037 if (VA.needsCustom()) {
1038 // Handle f64 or half of a v2f64.
1039 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1041 Chain = Lo.getValue(1);
1042 InFlag = Lo.getValue(2);
1043 VA = RVLocs[++i]; // skip ahead to next loc
1044 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1046 Chain = Hi.getValue(1);
1047 InFlag = Hi.getValue(2);
1048 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1050 if (VA.getLocVT() == MVT::v2f64) {
1051 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1052 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1053 DAG.getConstant(0, MVT::i32));
1055 VA = RVLocs[++i]; // skip ahead to next loc
1056 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1057 Chain = Lo.getValue(1);
1058 InFlag = Lo.getValue(2);
1059 VA = RVLocs[++i]; // skip ahead to next loc
1060 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1061 Chain = Hi.getValue(1);
1062 InFlag = Hi.getValue(2);
1063 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1064 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1065 DAG.getConstant(1, MVT::i32));
1068 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1070 Chain = Val.getValue(1);
1071 InFlag = Val.getValue(2);
1074 switch (VA.getLocInfo()) {
1075 default: llvm_unreachable("Unknown loc info!");
1076 case CCValAssign::Full: break;
1077 case CCValAssign::BCvt:
1078 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1082 InVals.push_back(Val);
1088 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1089 /// by "Src" to address "Dst" of size "Size". Alignment information is
1090 /// specified by the specific parameter attribute. The copy will be passed as
1091 /// a byval function parameter.
1092 /// Sometimes what we are copying is the end of a larger object, the part that
1093 /// does not fit in registers.
1095 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1096 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1098 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1099 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1100 /*isVolatile=*/false, /*AlwaysInline=*/false,
1101 MachinePointerInfo(0), MachinePointerInfo(0));
1104 /// LowerMemOpCallTo - Store the argument to the stack.
1106 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1107 SDValue StackPtr, SDValue Arg,
1108 DebugLoc dl, SelectionDAG &DAG,
1109 const CCValAssign &VA,
1110 ISD::ArgFlagsTy Flags) const {
1111 unsigned LocMemOffset = VA.getLocMemOffset();
1112 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1113 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1114 if (Flags.isByVal())
1115 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1117 return DAG.getStore(Chain, dl, Arg, PtrOff,
1118 MachinePointerInfo::getStack(LocMemOffset),
1122 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1123 SDValue Chain, SDValue &Arg,
1124 RegsToPassVector &RegsToPass,
1125 CCValAssign &VA, CCValAssign &NextVA,
1127 SmallVector<SDValue, 8> &MemOpChains,
1128 ISD::ArgFlagsTy Flags) const {
1130 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1131 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1132 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1134 if (NextVA.isRegLoc())
1135 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1137 assert(NextVA.isMemLoc());
1138 if (StackPtr.getNode() == 0)
1139 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1141 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1147 /// LowerCall - Lowering a call into a callseq_start <-
1148 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1151 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1152 CallingConv::ID CallConv, bool isVarArg,
1154 const SmallVectorImpl<ISD::OutputArg> &Outs,
1155 const SmallVectorImpl<SDValue> &OutVals,
1156 const SmallVectorImpl<ISD::InputArg> &Ins,
1157 DebugLoc dl, SelectionDAG &DAG,
1158 SmallVectorImpl<SDValue> &InVals) const {
1159 MachineFunction &MF = DAG.getMachineFunction();
1160 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1161 bool IsSibCall = false;
1162 // Temporarily disable tail calls so things don't break.
1163 if (!EnableARMTailCalls)
1166 // Check if it's really possible to do a tail call.
1167 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1168 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1169 Outs, OutVals, Ins, DAG);
1170 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1171 // detected sibcalls.
1178 // Analyze operands of the call, assigning locations to each operand.
1179 SmallVector<CCValAssign, 16> ArgLocs;
1180 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1182 CCInfo.AnalyzeCallOperands(Outs,
1183 CCAssignFnForNode(CallConv, /* Return*/ false,
1186 // Get a count of how many bytes are to be pushed on the stack.
1187 unsigned NumBytes = CCInfo.getNextStackOffset();
1189 // For tail calls, memory operands are available in our caller's stack.
1193 // Adjust the stack pointer for the new arguments...
1194 // These operations are automatically eliminated by the prolog/epilog pass
1196 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1198 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1200 RegsToPassVector RegsToPass;
1201 SmallVector<SDValue, 8> MemOpChains;
1203 // Walk the register/memloc assignments, inserting copies/loads. In the case
1204 // of tail call optimization, arguments are handled later.
1205 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1207 ++i, ++realArgIdx) {
1208 CCValAssign &VA = ArgLocs[i];
1209 SDValue Arg = OutVals[realArgIdx];
1210 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1212 // Promote the value if needed.
1213 switch (VA.getLocInfo()) {
1214 default: llvm_unreachable("Unknown loc info!");
1215 case CCValAssign::Full: break;
1216 case CCValAssign::SExt:
1217 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1219 case CCValAssign::ZExt:
1220 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1222 case CCValAssign::AExt:
1223 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1225 case CCValAssign::BCvt:
1226 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1230 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1231 if (VA.needsCustom()) {
1232 if (VA.getLocVT() == MVT::v2f64) {
1233 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1234 DAG.getConstant(0, MVT::i32));
1235 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1236 DAG.getConstant(1, MVT::i32));
1238 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1239 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1241 VA = ArgLocs[++i]; // skip ahead to next loc
1242 if (VA.isRegLoc()) {
1243 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1244 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1246 assert(VA.isMemLoc());
1248 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1249 dl, DAG, VA, Flags));
1252 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1253 StackPtr, MemOpChains, Flags);
1255 } else if (VA.isRegLoc()) {
1256 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1257 } else if (!IsSibCall) {
1258 assert(VA.isMemLoc());
1260 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1261 dl, DAG, VA, Flags));
1265 if (!MemOpChains.empty())
1266 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1267 &MemOpChains[0], MemOpChains.size());
1269 // Build a sequence of copy-to-reg nodes chained together with token chain
1270 // and flag operands which copy the outgoing args into the appropriate regs.
1272 // Tail call byval lowering might overwrite argument registers so in case of
1273 // tail call optimization the copies to registers are lowered later.
1275 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1276 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1277 RegsToPass[i].second, InFlag);
1278 InFlag = Chain.getValue(1);
1281 // For tail calls lower the arguments to the 'real' stack slot.
1283 // Force all the incoming stack arguments to be loaded from the stack
1284 // before any new outgoing arguments are stored to the stack, because the
1285 // outgoing stack slots may alias the incoming argument stack slots, and
1286 // the alias isn't otherwise explicit. This is slightly more conservative
1287 // than necessary, because it means that each store effectively depends
1288 // on every argument instead of just those arguments it would clobber.
1290 // Do not flag preceeding copytoreg stuff together with the following stuff.
1292 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1293 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1294 RegsToPass[i].second, InFlag);
1295 InFlag = Chain.getValue(1);
1300 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1301 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1302 // node so that legalize doesn't hack it.
1303 bool isDirect = false;
1304 bool isARMFunc = false;
1305 bool isLocalARMFunc = false;
1306 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1308 if (EnableARMLongCalls) {
1309 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1310 && "long-calls with non-static relocation model!");
1311 // Handle a global address or an external symbol. If it's not one of
1312 // those, the target's already in a register, so we don't need to do
1314 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1315 const GlobalValue *GV = G->getGlobal();
1316 // Create a constant pool entry for the callee address
1317 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1318 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1321 // Get the address of the callee into a register
1322 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1323 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1324 Callee = DAG.getLoad(getPointerTy(), dl,
1325 DAG.getEntryNode(), CPAddr,
1326 MachinePointerInfo::getConstantPool(),
1328 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1329 const char *Sym = S->getSymbol();
1331 // Create a constant pool entry for the callee address
1332 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1333 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1334 Sym, ARMPCLabelIndex, 0);
1335 // Get the address of the callee into a register
1336 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1337 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1338 Callee = DAG.getLoad(getPointerTy(), dl,
1339 DAG.getEntryNode(), CPAddr,
1340 MachinePointerInfo::getConstantPool(),
1343 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1344 const GlobalValue *GV = G->getGlobal();
1346 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1347 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1348 getTargetMachine().getRelocationModel() != Reloc::Static;
1349 isARMFunc = !Subtarget->isThumb() || isStub;
1350 // ARM call to a local ARM function is predicable.
1351 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1352 // tBX takes a register source operand.
1353 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1354 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1355 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1358 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1359 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1360 Callee = DAG.getLoad(getPointerTy(), dl,
1361 DAG.getEntryNode(), CPAddr,
1362 MachinePointerInfo::getConstantPool(),
1364 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1365 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1366 getPointerTy(), Callee, PICLabel);
1368 // On ELF targets for PIC code, direct calls should go through the PLT
1369 unsigned OpFlags = 0;
1370 if (Subtarget->isTargetELF() &&
1371 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1372 OpFlags = ARMII::MO_PLT;
1373 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1375 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1377 bool isStub = Subtarget->isTargetDarwin() &&
1378 getTargetMachine().getRelocationModel() != Reloc::Static;
1379 isARMFunc = !Subtarget->isThumb() || isStub;
1380 // tBX takes a register source operand.
1381 const char *Sym = S->getSymbol();
1382 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1383 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1384 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1385 Sym, ARMPCLabelIndex, 4);
1386 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1387 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1388 Callee = DAG.getLoad(getPointerTy(), dl,
1389 DAG.getEntryNode(), CPAddr,
1390 MachinePointerInfo::getConstantPool(),
1392 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1393 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1394 getPointerTy(), Callee, PICLabel);
1396 unsigned OpFlags = 0;
1397 // On ELF targets for PIC code, direct calls should go through the PLT
1398 if (Subtarget->isTargetELF() &&
1399 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1400 OpFlags = ARMII::MO_PLT;
1401 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1405 // FIXME: handle tail calls differently.
1407 if (Subtarget->isThumb()) {
1408 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1409 CallOpc = ARMISD::CALL_NOLINK;
1411 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1413 CallOpc = (isDirect || Subtarget->hasV5TOps())
1414 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1415 : ARMISD::CALL_NOLINK;
1418 std::vector<SDValue> Ops;
1419 Ops.push_back(Chain);
1420 Ops.push_back(Callee);
1422 // Add argument registers to the end of the list so that they are known live
1424 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1425 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1426 RegsToPass[i].second.getValueType()));
1428 if (InFlag.getNode())
1429 Ops.push_back(InFlag);
1431 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1433 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1435 // Returns a chain and a flag for retval copy to use.
1436 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1437 InFlag = Chain.getValue(1);
1439 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1440 DAG.getIntPtrConstant(0, true), InFlag);
1442 InFlag = Chain.getValue(1);
1444 // Handle result values, copying them out of physregs into vregs that we
1446 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1450 /// MatchingStackOffset - Return true if the given stack call argument is
1451 /// already available in the same position (relatively) of the caller's
1452 /// incoming argument stack.
1454 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1455 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1456 const ARMInstrInfo *TII) {
1457 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1459 if (Arg.getOpcode() == ISD::CopyFromReg) {
1460 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1461 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1463 MachineInstr *Def = MRI->getVRegDef(VR);
1466 if (!Flags.isByVal()) {
1467 if (!TII->isLoadFromStackSlot(Def, FI))
1472 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1473 if (Flags.isByVal())
1474 // ByVal argument is passed in as a pointer but it's now being
1475 // dereferenced. e.g.
1476 // define @foo(%struct.X* %A) {
1477 // tail call @bar(%struct.X* byval %A)
1480 SDValue Ptr = Ld->getBasePtr();
1481 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1484 FI = FINode->getIndex();
1488 assert(FI != INT_MAX);
1489 if (!MFI->isFixedObjectIndex(FI))
1491 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1494 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1495 /// for tail call optimization. Targets which want to do tail call
1496 /// optimization should implement this function.
1498 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1499 CallingConv::ID CalleeCC,
1501 bool isCalleeStructRet,
1502 bool isCallerStructRet,
1503 const SmallVectorImpl<ISD::OutputArg> &Outs,
1504 const SmallVectorImpl<SDValue> &OutVals,
1505 const SmallVectorImpl<ISD::InputArg> &Ins,
1506 SelectionDAG& DAG) const {
1507 const Function *CallerF = DAG.getMachineFunction().getFunction();
1508 CallingConv::ID CallerCC = CallerF->getCallingConv();
1509 bool CCMatch = CallerCC == CalleeCC;
1511 // Look for obvious safe cases to perform tail call optimization that do not
1512 // require ABI changes. This is what gcc calls sibcall.
1514 // Do not sibcall optimize vararg calls unless the call site is not passing
1516 if (isVarArg && !Outs.empty())
1519 // Also avoid sibcall optimization if either caller or callee uses struct
1520 // return semantics.
1521 if (isCalleeStructRet || isCallerStructRet)
1524 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1525 // emitEpilogue is not ready for them.
1526 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1527 // LR. This means if we need to reload LR, it takes an extra instructions,
1528 // which outweighs the value of the tail call; but here we don't know yet
1529 // whether LR is going to be used. Probably the right approach is to
1530 // generate the tail call here and turn it back into CALL/RET in
1531 // emitEpilogue if LR is used.
1533 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1534 // but we need to make sure there are enough registers; the only valid
1535 // registers are the 4 used for parameters. We don't currently do this
1537 if (Subtarget->isThumb1Only())
1540 // If the calling conventions do not match, then we'd better make sure the
1541 // results are returned in the same way as what the caller expects.
1543 SmallVector<CCValAssign, 16> RVLocs1;
1544 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1545 RVLocs1, *DAG.getContext());
1546 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1548 SmallVector<CCValAssign, 16> RVLocs2;
1549 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1550 RVLocs2, *DAG.getContext());
1551 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1553 if (RVLocs1.size() != RVLocs2.size())
1555 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1556 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1558 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1560 if (RVLocs1[i].isRegLoc()) {
1561 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1564 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1570 // If the callee takes no arguments then go on to check the results of the
1572 if (!Outs.empty()) {
1573 // Check if stack adjustment is needed. For now, do not do this if any
1574 // argument is passed on the stack.
1575 SmallVector<CCValAssign, 16> ArgLocs;
1576 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1577 ArgLocs, *DAG.getContext());
1578 CCInfo.AnalyzeCallOperands(Outs,
1579 CCAssignFnForNode(CalleeCC, false, isVarArg));
1580 if (CCInfo.getNextStackOffset()) {
1581 MachineFunction &MF = DAG.getMachineFunction();
1583 // Check if the arguments are already laid out in the right way as
1584 // the caller's fixed stack objects.
1585 MachineFrameInfo *MFI = MF.getFrameInfo();
1586 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1587 const ARMInstrInfo *TII =
1588 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1589 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1591 ++i, ++realArgIdx) {
1592 CCValAssign &VA = ArgLocs[i];
1593 EVT RegVT = VA.getLocVT();
1594 SDValue Arg = OutVals[realArgIdx];
1595 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1596 if (VA.getLocInfo() == CCValAssign::Indirect)
1598 if (VA.needsCustom()) {
1599 // f64 and vector types are split into multiple registers or
1600 // register/stack-slot combinations. The types will not match
1601 // the registers; give up on memory f64 refs until we figure
1602 // out what to do about this.
1605 if (!ArgLocs[++i].isRegLoc())
1607 if (RegVT == MVT::v2f64) {
1608 if (!ArgLocs[++i].isRegLoc())
1610 if (!ArgLocs[++i].isRegLoc())
1613 } else if (!VA.isRegLoc()) {
1614 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1626 ARMTargetLowering::LowerReturn(SDValue Chain,
1627 CallingConv::ID CallConv, bool isVarArg,
1628 const SmallVectorImpl<ISD::OutputArg> &Outs,
1629 const SmallVectorImpl<SDValue> &OutVals,
1630 DebugLoc dl, SelectionDAG &DAG) const {
1632 // CCValAssign - represent the assignment of the return value to a location.
1633 SmallVector<CCValAssign, 16> RVLocs;
1635 // CCState - Info about the registers and stack slots.
1636 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1639 // Analyze outgoing return values.
1640 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1643 // If this is the first return lowered for this function, add
1644 // the regs to the liveout set for the function.
1645 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1646 for (unsigned i = 0; i != RVLocs.size(); ++i)
1647 if (RVLocs[i].isRegLoc())
1648 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1653 // Copy the result values into the output registers.
1654 for (unsigned i = 0, realRVLocIdx = 0;
1656 ++i, ++realRVLocIdx) {
1657 CCValAssign &VA = RVLocs[i];
1658 assert(VA.isRegLoc() && "Can only return in registers!");
1660 SDValue Arg = OutVals[realRVLocIdx];
1662 switch (VA.getLocInfo()) {
1663 default: llvm_unreachable("Unknown loc info!");
1664 case CCValAssign::Full: break;
1665 case CCValAssign::BCvt:
1666 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1670 if (VA.needsCustom()) {
1671 if (VA.getLocVT() == MVT::v2f64) {
1672 // Extract the first half and return it in two registers.
1673 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1674 DAG.getConstant(0, MVT::i32));
1675 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1676 DAG.getVTList(MVT::i32, MVT::i32), Half);
1678 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1679 Flag = Chain.getValue(1);
1680 VA = RVLocs[++i]; // skip ahead to next loc
1681 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1682 HalfGPRs.getValue(1), Flag);
1683 Flag = Chain.getValue(1);
1684 VA = RVLocs[++i]; // skip ahead to next loc
1686 // Extract the 2nd half and fall through to handle it as an f64 value.
1687 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1688 DAG.getConstant(1, MVT::i32));
1690 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1692 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1693 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1694 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1695 Flag = Chain.getValue(1);
1696 VA = RVLocs[++i]; // skip ahead to next loc
1697 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1700 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1702 // Guarantee that all emitted copies are
1703 // stuck together, avoiding something bad.
1704 Flag = Chain.getValue(1);
1709 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1711 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1716 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1717 if (N->getNumValues() != 1)
1719 if (!N->hasNUsesOfValue(1, 0))
1722 unsigned NumCopies = 0;
1724 SDNode *Use = *N->use_begin();
1725 if (Use->getOpcode() == ISD::CopyToReg) {
1726 Copies[NumCopies++] = Use;
1727 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1728 // f64 returned in a pair of GPRs.
1729 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1731 if (UI->getOpcode() != ISD::CopyToReg)
1733 Copies[UI.getUse().getResNo()] = *UI;
1736 } else if (Use->getOpcode() == ISD::BITCAST) {
1737 // f32 returned in a single GPR.
1738 if (!Use->hasNUsesOfValue(1, 0))
1740 Use = *Use->use_begin();
1741 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1743 Copies[NumCopies++] = Use;
1748 if (NumCopies != 1 && NumCopies != 2)
1751 bool HasRet = false;
1752 for (unsigned i = 0; i < NumCopies; ++i) {
1753 SDNode *Copy = Copies[i];
1754 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1756 if (UI->getOpcode() == ISD::CopyToReg) {
1758 if (Use == Copies[0] || Use == Copies[1])
1762 if (UI->getOpcode() != ARMISD::RET_FLAG)
1771 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1772 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1773 // one of the above mentioned nodes. It has to be wrapped because otherwise
1774 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1775 // be used to form addressing mode. These wrapped nodes will be selected
1777 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1778 EVT PtrVT = Op.getValueType();
1779 // FIXME there is no actual debug info here
1780 DebugLoc dl = Op.getDebugLoc();
1781 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1783 if (CP->isMachineConstantPoolEntry())
1784 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1785 CP->getAlignment());
1787 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1788 CP->getAlignment());
1789 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1792 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1793 return MachineJumpTableInfo::EK_Inline;
1796 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1797 SelectionDAG &DAG) const {
1798 MachineFunction &MF = DAG.getMachineFunction();
1799 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1800 unsigned ARMPCLabelIndex = 0;
1801 DebugLoc DL = Op.getDebugLoc();
1802 EVT PtrVT = getPointerTy();
1803 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1804 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1806 if (RelocM == Reloc::Static) {
1807 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1809 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1810 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1811 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1812 ARMCP::CPBlockAddress,
1814 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1816 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1817 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1818 MachinePointerInfo::getConstantPool(),
1820 if (RelocM == Reloc::Static)
1822 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1823 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1826 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1828 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1829 SelectionDAG &DAG) const {
1830 DebugLoc dl = GA->getDebugLoc();
1831 EVT PtrVT = getPointerTy();
1832 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1833 MachineFunction &MF = DAG.getMachineFunction();
1834 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1835 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1836 ARMConstantPoolValue *CPV =
1837 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1838 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1839 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1840 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1841 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1842 MachinePointerInfo::getConstantPool(),
1844 SDValue Chain = Argument.getValue(1);
1846 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1847 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1849 // call __tls_get_addr.
1852 Entry.Node = Argument;
1853 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1854 Args.push_back(Entry);
1855 // FIXME: is there useful debug info available here?
1856 std::pair<SDValue, SDValue> CallResult =
1857 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1858 false, false, false, false,
1859 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1860 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1861 return CallResult.first;
1864 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1865 // "local exec" model.
1867 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1868 SelectionDAG &DAG) const {
1869 const GlobalValue *GV = GA->getGlobal();
1870 DebugLoc dl = GA->getDebugLoc();
1872 SDValue Chain = DAG.getEntryNode();
1873 EVT PtrVT = getPointerTy();
1874 // Get the Thread Pointer
1875 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1877 if (GV->isDeclaration()) {
1878 MachineFunction &MF = DAG.getMachineFunction();
1879 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1880 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1881 // Initial exec model.
1882 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1883 ARMConstantPoolValue *CPV =
1884 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1885 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1886 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1887 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1888 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1889 MachinePointerInfo::getConstantPool(),
1891 Chain = Offset.getValue(1);
1893 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1894 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1896 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1897 MachinePointerInfo::getConstantPool(),
1901 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
1902 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1903 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1904 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1905 MachinePointerInfo::getConstantPool(),
1909 // The address of the thread local variable is the add of the thread
1910 // pointer with the offset of the variable.
1911 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1915 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1916 // TODO: implement the "local dynamic" model
1917 assert(Subtarget->isTargetELF() &&
1918 "TLS not implemented for non-ELF targets");
1919 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1920 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1921 // otherwise use the "Local Exec" TLS Model
1922 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1923 return LowerToTLSGeneralDynamicModel(GA, DAG);
1925 return LowerToTLSExecModels(GA, DAG);
1928 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1929 SelectionDAG &DAG) const {
1930 EVT PtrVT = getPointerTy();
1931 DebugLoc dl = Op.getDebugLoc();
1932 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1933 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1934 if (RelocM == Reloc::PIC_) {
1935 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1936 ARMConstantPoolValue *CPV =
1937 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
1938 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1939 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1940 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1942 MachinePointerInfo::getConstantPool(),
1944 SDValue Chain = Result.getValue(1);
1945 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1946 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1948 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1949 MachinePointerInfo::getGOT(), false, false, 0);
1952 // If we have T2 ops, we can materialize the address directly via movt/movw
1953 // pair. This is always cheaper.
1954 if (Subtarget->useMovt()) {
1955 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1956 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1958 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1959 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1960 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1961 MachinePointerInfo::getConstantPool(),
1967 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1968 SelectionDAG &DAG) const {
1969 MachineFunction &MF = DAG.getMachineFunction();
1970 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1971 unsigned ARMPCLabelIndex = 0;
1972 EVT PtrVT = getPointerTy();
1973 DebugLoc dl = Op.getDebugLoc();
1974 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1975 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1977 if (RelocM == Reloc::Static)
1978 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1980 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1981 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1982 ARMConstantPoolValue *CPV =
1983 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1984 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1986 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1988 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1989 MachinePointerInfo::getConstantPool(),
1991 SDValue Chain = Result.getValue(1);
1993 if (RelocM == Reloc::PIC_) {
1994 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1995 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1998 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1999 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2005 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2006 SelectionDAG &DAG) const {
2007 assert(Subtarget->isTargetELF() &&
2008 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2009 MachineFunction &MF = DAG.getMachineFunction();
2010 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2011 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
2012 EVT PtrVT = getPointerTy();
2013 DebugLoc dl = Op.getDebugLoc();
2014 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2015 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2016 "_GLOBAL_OFFSET_TABLE_",
2017 ARMPCLabelIndex, PCAdj);
2018 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2019 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2020 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2021 MachinePointerInfo::getConstantPool(),
2023 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2024 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2028 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2030 DebugLoc dl = Op.getDebugLoc();
2031 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2032 Op.getOperand(0), Op.getOperand(1));
2036 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2037 DebugLoc dl = Op.getDebugLoc();
2038 SDValue Val = DAG.getConstant(0, MVT::i32);
2039 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2040 Op.getOperand(1), Val);
2044 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2045 DebugLoc dl = Op.getDebugLoc();
2046 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2047 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2051 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2052 const ARMSubtarget *Subtarget) const {
2053 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2054 DebugLoc dl = Op.getDebugLoc();
2056 default: return SDValue(); // Don't custom lower most intrinsics.
2057 case Intrinsic::arm_thread_pointer: {
2058 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2059 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2061 case Intrinsic::eh_sjlj_lsda: {
2062 MachineFunction &MF = DAG.getMachineFunction();
2063 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2064 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
2065 EVT PtrVT = getPointerTy();
2066 DebugLoc dl = Op.getDebugLoc();
2067 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2069 unsigned PCAdj = (RelocM != Reloc::PIC_)
2070 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2071 ARMConstantPoolValue *CPV =
2072 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2073 ARMCP::CPLSDA, PCAdj);
2074 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2075 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2077 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2078 MachinePointerInfo::getConstantPool(),
2081 if (RelocM == Reloc::PIC_) {
2082 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2083 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2090 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2091 const ARMSubtarget *Subtarget) {
2092 DebugLoc dl = Op.getDebugLoc();
2093 if (!Subtarget->hasDataBarrier()) {
2094 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2095 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2097 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2098 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2099 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2100 DAG.getConstant(0, MVT::i32));
2103 SDValue Op5 = Op.getOperand(5);
2104 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2105 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2106 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2107 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2109 ARM_MB::MemBOpt DMBOpt;
2110 if (isDeviceBarrier)
2111 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2113 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2114 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2115 DAG.getConstant(DMBOpt, MVT::i32));
2118 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2119 const ARMSubtarget *Subtarget) {
2120 // ARM pre v5TE and Thumb1 does not have preload instructions.
2121 if (!(Subtarget->isThumb2() ||
2122 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2123 // Just preserve the chain.
2124 return Op.getOperand(0);
2126 DebugLoc dl = Op.getDebugLoc();
2127 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2129 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2130 // ARMv7 with MP extension has PLDW.
2131 return Op.getOperand(0);
2133 if (Subtarget->isThumb())
2135 isRead = ~isRead & 1;
2136 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2138 // Currently there is no intrinsic that matches pli.
2139 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2140 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2141 DAG.getConstant(isData, MVT::i32));
2144 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2145 MachineFunction &MF = DAG.getMachineFunction();
2146 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2148 // vastart just stores the address of the VarArgsFrameIndex slot into the
2149 // memory location argument.
2150 DebugLoc dl = Op.getDebugLoc();
2151 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2152 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2153 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2154 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2155 MachinePointerInfo(SV), false, false, 0);
2159 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2160 SDValue &Root, SelectionDAG &DAG,
2161 DebugLoc dl) const {
2162 MachineFunction &MF = DAG.getMachineFunction();
2163 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2165 TargetRegisterClass *RC;
2166 if (AFI->isThumb1OnlyFunction())
2167 RC = ARM::tGPRRegisterClass;
2169 RC = ARM::GPRRegisterClass;
2171 // Transform the arguments stored in physical registers into virtual ones.
2172 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2173 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2176 if (NextVA.isMemLoc()) {
2177 MachineFrameInfo *MFI = MF.getFrameInfo();
2178 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2180 // Create load node to retrieve arguments from the stack.
2181 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2182 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2183 MachinePointerInfo::getFixedStack(FI),
2186 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2187 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2190 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2194 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2195 CallingConv::ID CallConv, bool isVarArg,
2196 const SmallVectorImpl<ISD::InputArg>
2198 DebugLoc dl, SelectionDAG &DAG,
2199 SmallVectorImpl<SDValue> &InVals)
2202 MachineFunction &MF = DAG.getMachineFunction();
2203 MachineFrameInfo *MFI = MF.getFrameInfo();
2205 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2207 // Assign locations to all of the incoming arguments.
2208 SmallVector<CCValAssign, 16> ArgLocs;
2209 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2211 CCInfo.AnalyzeFormalArguments(Ins,
2212 CCAssignFnForNode(CallConv, /* Return*/ false,
2215 SmallVector<SDValue, 16> ArgValues;
2217 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2218 CCValAssign &VA = ArgLocs[i];
2220 // Arguments stored in registers.
2221 if (VA.isRegLoc()) {
2222 EVT RegVT = VA.getLocVT();
2225 if (VA.needsCustom()) {
2226 // f64 and vector types are split up into multiple registers or
2227 // combinations of registers and stack slots.
2228 if (VA.getLocVT() == MVT::v2f64) {
2229 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2231 VA = ArgLocs[++i]; // skip ahead to next loc
2233 if (VA.isMemLoc()) {
2234 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2235 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2236 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2237 MachinePointerInfo::getFixedStack(FI),
2240 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2243 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2244 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2245 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2246 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2247 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2249 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2252 TargetRegisterClass *RC;
2254 if (RegVT == MVT::f32)
2255 RC = ARM::SPRRegisterClass;
2256 else if (RegVT == MVT::f64)
2257 RC = ARM::DPRRegisterClass;
2258 else if (RegVT == MVT::v2f64)
2259 RC = ARM::QPRRegisterClass;
2260 else if (RegVT == MVT::i32)
2261 RC = (AFI->isThumb1OnlyFunction() ?
2262 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2264 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2266 // Transform the arguments in physical registers into virtual ones.
2267 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2268 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2271 // If this is an 8 or 16-bit value, it is really passed promoted
2272 // to 32 bits. Insert an assert[sz]ext to capture this, then
2273 // truncate to the right size.
2274 switch (VA.getLocInfo()) {
2275 default: llvm_unreachable("Unknown loc info!");
2276 case CCValAssign::Full: break;
2277 case CCValAssign::BCvt:
2278 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2280 case CCValAssign::SExt:
2281 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2282 DAG.getValueType(VA.getValVT()));
2283 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2285 case CCValAssign::ZExt:
2286 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2287 DAG.getValueType(VA.getValVT()));
2288 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2292 InVals.push_back(ArgValue);
2294 } else { // VA.isRegLoc()
2297 assert(VA.isMemLoc());
2298 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2300 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2301 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2303 // Create load nodes to retrieve arguments from the stack.
2304 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2305 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2306 MachinePointerInfo::getFixedStack(FI),
2313 static const unsigned GPRArgRegs[] = {
2314 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2317 unsigned NumGPRs = CCInfo.getFirstUnallocated
2318 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2320 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2321 unsigned VARegSize = (4 - NumGPRs) * 4;
2322 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2323 unsigned ArgOffset = CCInfo.getNextStackOffset();
2324 if (VARegSaveSize) {
2325 // If this function is vararg, store any remaining integer argument regs
2326 // to their spots on the stack so that they may be loaded by deferencing
2327 // the result of va_next.
2328 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2329 AFI->setVarArgsFrameIndex(
2330 MFI->CreateFixedObject(VARegSaveSize,
2331 ArgOffset + VARegSaveSize - VARegSize,
2333 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2336 SmallVector<SDValue, 4> MemOps;
2337 for (; NumGPRs < 4; ++NumGPRs) {
2338 TargetRegisterClass *RC;
2339 if (AFI->isThumb1OnlyFunction())
2340 RC = ARM::tGPRRegisterClass;
2342 RC = ARM::GPRRegisterClass;
2344 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2345 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2347 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2348 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2350 MemOps.push_back(Store);
2351 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2352 DAG.getConstant(4, getPointerTy()));
2354 if (!MemOps.empty())
2355 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2356 &MemOps[0], MemOps.size());
2358 // This will point to the next argument passed via stack.
2359 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2365 /// isFloatingPointZero - Return true if this is +0.0.
2366 static bool isFloatingPointZero(SDValue Op) {
2367 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2368 return CFP->getValueAPF().isPosZero();
2369 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2370 // Maybe this has already been legalized into the constant pool?
2371 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2372 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2373 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2374 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2375 return CFP->getValueAPF().isPosZero();
2381 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2382 /// the given operands.
2384 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2385 SDValue &ARMcc, SelectionDAG &DAG,
2386 DebugLoc dl) const {
2387 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2388 unsigned C = RHSC->getZExtValue();
2389 if (!isLegalICmpImmediate(C)) {
2390 // Constant does not fit, try adjusting it by one?
2395 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2396 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2397 RHS = DAG.getConstant(C-1, MVT::i32);
2402 if (C != 0 && isLegalICmpImmediate(C-1)) {
2403 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2404 RHS = DAG.getConstant(C-1, MVT::i32);
2409 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2410 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2411 RHS = DAG.getConstant(C+1, MVT::i32);
2416 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2417 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2418 RHS = DAG.getConstant(C+1, MVT::i32);
2425 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2426 ARMISD::NodeType CompareType;
2429 CompareType = ARMISD::CMP;
2434 CompareType = ARMISD::CMPZ;
2437 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2438 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2441 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2443 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2444 DebugLoc dl) const {
2446 if (!isFloatingPointZero(RHS))
2447 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2449 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2450 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2453 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2454 SDValue Cond = Op.getOperand(0);
2455 SDValue SelectTrue = Op.getOperand(1);
2456 SDValue SelectFalse = Op.getOperand(2);
2457 DebugLoc dl = Op.getDebugLoc();
2461 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2462 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2464 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2465 const ConstantSDNode *CMOVTrue =
2466 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2467 const ConstantSDNode *CMOVFalse =
2468 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2470 if (CMOVTrue && CMOVFalse) {
2471 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2472 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2476 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2478 False = SelectFalse;
2479 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2484 if (True.getNode() && False.getNode()) {
2485 EVT VT = Cond.getValueType();
2486 SDValue ARMcc = Cond.getOperand(2);
2487 SDValue CCR = Cond.getOperand(3);
2488 SDValue Cmp = Cond.getOperand(4);
2489 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2494 return DAG.getSelectCC(dl, Cond,
2495 DAG.getConstant(0, Cond.getValueType()),
2496 SelectTrue, SelectFalse, ISD::SETNE);
2499 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2500 EVT VT = Op.getValueType();
2501 SDValue LHS = Op.getOperand(0);
2502 SDValue RHS = Op.getOperand(1);
2503 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2504 SDValue TrueVal = Op.getOperand(2);
2505 SDValue FalseVal = Op.getOperand(3);
2506 DebugLoc dl = Op.getDebugLoc();
2508 if (LHS.getValueType() == MVT::i32) {
2510 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2511 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2512 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2515 ARMCC::CondCodes CondCode, CondCode2;
2516 FPCCToARMCC(CC, CondCode, CondCode2);
2518 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2519 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2520 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2521 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2523 if (CondCode2 != ARMCC::AL) {
2524 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2525 // FIXME: Needs another CMP because flag can have but one use.
2526 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2527 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2528 Result, TrueVal, ARMcc2, CCR, Cmp2);
2533 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2534 /// to morph to an integer compare sequence.
2535 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2536 const ARMSubtarget *Subtarget) {
2537 SDNode *N = Op.getNode();
2538 if (!N->hasOneUse())
2539 // Otherwise it requires moving the value from fp to integer registers.
2541 if (!N->getNumValues())
2543 EVT VT = Op.getValueType();
2544 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2545 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2546 // vmrs are very slow, e.g. cortex-a8.
2549 if (isFloatingPointZero(Op)) {
2553 return ISD::isNormalLoad(N);
2556 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2557 if (isFloatingPointZero(Op))
2558 return DAG.getConstant(0, MVT::i32);
2560 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2561 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2562 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2563 Ld->isVolatile(), Ld->isNonTemporal(),
2564 Ld->getAlignment());
2566 llvm_unreachable("Unknown VFP cmp argument!");
2569 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2570 SDValue &RetVal1, SDValue &RetVal2) {
2571 if (isFloatingPointZero(Op)) {
2572 RetVal1 = DAG.getConstant(0, MVT::i32);
2573 RetVal2 = DAG.getConstant(0, MVT::i32);
2577 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2578 SDValue Ptr = Ld->getBasePtr();
2579 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2580 Ld->getChain(), Ptr,
2581 Ld->getPointerInfo(),
2582 Ld->isVolatile(), Ld->isNonTemporal(),
2583 Ld->getAlignment());
2585 EVT PtrType = Ptr.getValueType();
2586 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2587 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2588 PtrType, Ptr, DAG.getConstant(4, PtrType));
2589 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2590 Ld->getChain(), NewPtr,
2591 Ld->getPointerInfo().getWithOffset(4),
2592 Ld->isVolatile(), Ld->isNonTemporal(),
2597 llvm_unreachable("Unknown VFP cmp argument!");
2600 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2601 /// f32 and even f64 comparisons to integer ones.
2603 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2604 SDValue Chain = Op.getOperand(0);
2605 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2606 SDValue LHS = Op.getOperand(2);
2607 SDValue RHS = Op.getOperand(3);
2608 SDValue Dest = Op.getOperand(4);
2609 DebugLoc dl = Op.getDebugLoc();
2611 bool SeenZero = false;
2612 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2613 canChangeToInt(RHS, SeenZero, Subtarget) &&
2614 // If one of the operand is zero, it's safe to ignore the NaN case since
2615 // we only care about equality comparisons.
2616 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2617 // If unsafe fp math optimization is enabled and there are no othter uses of
2618 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2619 // to an integer comparison.
2620 if (CC == ISD::SETOEQ)
2622 else if (CC == ISD::SETUNE)
2626 if (LHS.getValueType() == MVT::f32) {
2627 LHS = bitcastf32Toi32(LHS, DAG);
2628 RHS = bitcastf32Toi32(RHS, DAG);
2629 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2630 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2631 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2632 Chain, Dest, ARMcc, CCR, Cmp);
2637 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2638 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2639 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2640 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2641 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2642 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2643 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2649 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2650 SDValue Chain = Op.getOperand(0);
2651 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2652 SDValue LHS = Op.getOperand(2);
2653 SDValue RHS = Op.getOperand(3);
2654 SDValue Dest = Op.getOperand(4);
2655 DebugLoc dl = Op.getDebugLoc();
2657 if (LHS.getValueType() == MVT::i32) {
2659 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2660 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2661 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2662 Chain, Dest, ARMcc, CCR, Cmp);
2665 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2668 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2669 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2670 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2671 if (Result.getNode())
2675 ARMCC::CondCodes CondCode, CondCode2;
2676 FPCCToARMCC(CC, CondCode, CondCode2);
2678 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2679 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2680 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2681 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2682 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2683 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2684 if (CondCode2 != ARMCC::AL) {
2685 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2686 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2687 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2692 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2693 SDValue Chain = Op.getOperand(0);
2694 SDValue Table = Op.getOperand(1);
2695 SDValue Index = Op.getOperand(2);
2696 DebugLoc dl = Op.getDebugLoc();
2698 EVT PTy = getPointerTy();
2699 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2700 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2701 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2702 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2703 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2704 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2705 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2706 if (Subtarget->isThumb2()) {
2707 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2708 // which does another jump to the destination. This also makes it easier
2709 // to translate it to TBB / TBH later.
2710 // FIXME: This might not work if the function is extremely large.
2711 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2712 Addr, Op.getOperand(2), JTI, UId);
2714 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2715 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2716 MachinePointerInfo::getJumpTable(),
2718 Chain = Addr.getValue(1);
2719 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2720 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2722 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2723 MachinePointerInfo::getJumpTable(), false, false, 0);
2724 Chain = Addr.getValue(1);
2725 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2729 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2730 DebugLoc dl = Op.getDebugLoc();
2733 switch (Op.getOpcode()) {
2735 assert(0 && "Invalid opcode!");
2736 case ISD::FP_TO_SINT:
2737 Opc = ARMISD::FTOSI;
2739 case ISD::FP_TO_UINT:
2740 Opc = ARMISD::FTOUI;
2743 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2744 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2747 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2748 EVT VT = Op.getValueType();
2749 DebugLoc dl = Op.getDebugLoc();
2752 switch (Op.getOpcode()) {
2754 assert(0 && "Invalid opcode!");
2755 case ISD::SINT_TO_FP:
2756 Opc = ARMISD::SITOF;
2758 case ISD::UINT_TO_FP:
2759 Opc = ARMISD::UITOF;
2763 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
2764 return DAG.getNode(Opc, dl, VT, Op);
2767 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2768 // Implement fcopysign with a fabs and a conditional fneg.
2769 SDValue Tmp0 = Op.getOperand(0);
2770 SDValue Tmp1 = Op.getOperand(1);
2771 DebugLoc dl = Op.getDebugLoc();
2772 EVT VT = Op.getValueType();
2773 EVT SrcVT = Tmp1.getValueType();
2774 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2775 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2776 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2777 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2778 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2779 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2782 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2783 MachineFunction &MF = DAG.getMachineFunction();
2784 MachineFrameInfo *MFI = MF.getFrameInfo();
2785 MFI->setReturnAddressIsTaken(true);
2787 EVT VT = Op.getValueType();
2788 DebugLoc dl = Op.getDebugLoc();
2789 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2791 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2792 SDValue Offset = DAG.getConstant(4, MVT::i32);
2793 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2794 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2795 MachinePointerInfo(), false, false, 0);
2798 // Return LR, which contains the return address. Mark it an implicit live-in.
2799 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2800 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2803 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2804 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2805 MFI->setFrameAddressIsTaken(true);
2807 EVT VT = Op.getValueType();
2808 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2809 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2810 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2811 ? ARM::R7 : ARM::R11;
2812 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2814 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2815 MachinePointerInfo(),
2820 /// ExpandBITCAST - If the target supports VFP, this function is called to
2821 /// expand a bit convert where either the source or destination type is i64 to
2822 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2823 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2824 /// vectors), since the legalizer won't know what to do with that.
2825 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
2826 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2827 DebugLoc dl = N->getDebugLoc();
2828 SDValue Op = N->getOperand(0);
2830 // This function is only supposed to be called for i64 types, either as the
2831 // source or destination of the bit convert.
2832 EVT SrcVT = Op.getValueType();
2833 EVT DstVT = N->getValueType(0);
2834 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2835 "ExpandBITCAST called for non-i64 type");
2837 // Turn i64->f64 into VMOVDRR.
2838 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2839 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2840 DAG.getConstant(0, MVT::i32));
2841 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2842 DAG.getConstant(1, MVT::i32));
2843 return DAG.getNode(ISD::BITCAST, dl, DstVT,
2844 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2847 // Turn f64->i64 into VMOVRRD.
2848 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2849 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2850 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2851 // Merge the pieces into a single i64 value.
2852 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2858 /// getZeroVector - Returns a vector of specified type with all zero elements.
2859 /// Zero vectors are used to represent vector negation and in those cases
2860 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2861 /// not support i64 elements, so sometimes the zero vectors will need to be
2862 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2864 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2865 assert(VT.isVector() && "Expected a vector type");
2866 // The canonical modified immediate encoding of a zero vector is....0!
2867 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2868 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2869 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2870 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
2873 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2874 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2875 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2876 SelectionDAG &DAG) const {
2877 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2878 EVT VT = Op.getValueType();
2879 unsigned VTBits = VT.getSizeInBits();
2880 DebugLoc dl = Op.getDebugLoc();
2881 SDValue ShOpLo = Op.getOperand(0);
2882 SDValue ShOpHi = Op.getOperand(1);
2883 SDValue ShAmt = Op.getOperand(2);
2885 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2887 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2889 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2890 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2891 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2892 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2893 DAG.getConstant(VTBits, MVT::i32));
2894 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2895 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2896 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2898 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2899 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2901 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2902 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2905 SDValue Ops[2] = { Lo, Hi };
2906 return DAG.getMergeValues(Ops, 2, dl);
2909 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2910 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2911 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2912 SelectionDAG &DAG) const {
2913 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2914 EVT VT = Op.getValueType();
2915 unsigned VTBits = VT.getSizeInBits();
2916 DebugLoc dl = Op.getDebugLoc();
2917 SDValue ShOpLo = Op.getOperand(0);
2918 SDValue ShOpHi = Op.getOperand(1);
2919 SDValue ShAmt = Op.getOperand(2);
2922 assert(Op.getOpcode() == ISD::SHL_PARTS);
2923 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2924 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2925 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2926 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2927 DAG.getConstant(VTBits, MVT::i32));
2928 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2929 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2931 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2932 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2933 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2935 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2936 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2939 SDValue Ops[2] = { Lo, Hi };
2940 return DAG.getMergeValues(Ops, 2, dl);
2943 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2944 SelectionDAG &DAG) const {
2945 // The rounding mode is in bits 23:22 of the FPSCR.
2946 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2947 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2948 // so that the shift + and get folded into a bitfield extract.
2949 DebugLoc dl = Op.getDebugLoc();
2950 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2951 DAG.getConstant(Intrinsic::arm_get_fpscr,
2953 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2954 DAG.getConstant(1U << 22, MVT::i32));
2955 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2956 DAG.getConstant(22, MVT::i32));
2957 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2958 DAG.getConstant(3, MVT::i32));
2961 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2962 const ARMSubtarget *ST) {
2963 EVT VT = N->getValueType(0);
2964 DebugLoc dl = N->getDebugLoc();
2966 if (!ST->hasV6T2Ops())
2969 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2970 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2973 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2974 const ARMSubtarget *ST) {
2975 EVT VT = N->getValueType(0);
2976 DebugLoc dl = N->getDebugLoc();
2981 // Lower vector shifts on NEON to use VSHL.
2982 assert(ST->hasNEON() && "unexpected vector shift");
2984 // Left shifts translate directly to the vshiftu intrinsic.
2985 if (N->getOpcode() == ISD::SHL)
2986 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2987 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2988 N->getOperand(0), N->getOperand(1));
2990 assert((N->getOpcode() == ISD::SRA ||
2991 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2993 // NEON uses the same intrinsics for both left and right shifts. For
2994 // right shifts, the shift amounts are negative, so negate the vector of
2996 EVT ShiftVT = N->getOperand(1).getValueType();
2997 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2998 getZeroVector(ShiftVT, DAG, dl),
3000 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3001 Intrinsic::arm_neon_vshifts :
3002 Intrinsic::arm_neon_vshiftu);
3003 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3004 DAG.getConstant(vshiftInt, MVT::i32),
3005 N->getOperand(0), NegatedCount);
3008 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3009 const ARMSubtarget *ST) {
3010 EVT VT = N->getValueType(0);
3011 DebugLoc dl = N->getDebugLoc();
3013 // We can get here for a node like i32 = ISD::SHL i32, i64
3017 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3018 "Unknown shift to lower!");
3020 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3021 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3022 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3025 // If we are in thumb mode, we don't have RRX.
3026 if (ST->isThumb1Only()) return SDValue();
3028 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3029 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3030 DAG.getConstant(0, MVT::i32));
3031 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3032 DAG.getConstant(1, MVT::i32));
3034 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3035 // captures the result into a carry flag.
3036 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3037 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3039 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3040 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3042 // Merge the pieces into a single i64 value.
3043 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3046 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3047 SDValue TmpOp0, TmpOp1;
3048 bool Invert = false;
3052 SDValue Op0 = Op.getOperand(0);
3053 SDValue Op1 = Op.getOperand(1);
3054 SDValue CC = Op.getOperand(2);
3055 EVT VT = Op.getValueType();
3056 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3057 DebugLoc dl = Op.getDebugLoc();
3059 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3060 switch (SetCCOpcode) {
3061 default: llvm_unreachable("Illegal FP comparison"); break;
3063 case ISD::SETNE: Invert = true; // Fallthrough
3065 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3067 case ISD::SETLT: Swap = true; // Fallthrough
3069 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3071 case ISD::SETLE: Swap = true; // Fallthrough
3073 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3074 case ISD::SETUGE: Swap = true; // Fallthrough
3075 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3076 case ISD::SETUGT: Swap = true; // Fallthrough
3077 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3078 case ISD::SETUEQ: Invert = true; // Fallthrough
3080 // Expand this to (OLT | OGT).
3084 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3085 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3087 case ISD::SETUO: Invert = true; // Fallthrough
3089 // Expand this to (OLT | OGE).
3093 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3094 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3098 // Integer comparisons.
3099 switch (SetCCOpcode) {
3100 default: llvm_unreachable("Illegal integer comparison"); break;
3101 case ISD::SETNE: Invert = true;
3102 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3103 case ISD::SETLT: Swap = true;
3104 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3105 case ISD::SETLE: Swap = true;
3106 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3107 case ISD::SETULT: Swap = true;
3108 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3109 case ISD::SETULE: Swap = true;
3110 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3113 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3114 if (Opc == ARMISD::VCEQ) {
3117 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3119 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3122 // Ignore bitconvert.
3123 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3124 AndOp = AndOp.getOperand(0);
3126 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3128 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3129 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3136 std::swap(Op0, Op1);
3138 // If one of the operands is a constant vector zero, attempt to fold the
3139 // comparison to a specialized compare-against-zero form.
3141 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3143 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3144 if (Opc == ARMISD::VCGE)
3145 Opc = ARMISD::VCLEZ;
3146 else if (Opc == ARMISD::VCGT)
3147 Opc = ARMISD::VCLTZ;
3152 if (SingleOp.getNode()) {
3155 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3157 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3159 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3161 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3163 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3165 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3168 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3172 Result = DAG.getNOT(dl, Result, VT);
3177 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3178 /// valid vector constant for a NEON instruction with a "modified immediate"
3179 /// operand (e.g., VMOV). If so, return the encoded value.
3180 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3181 unsigned SplatBitSize, SelectionDAG &DAG,
3182 EVT &VT, bool is128Bits, NEONModImmType type) {
3183 unsigned OpCmode, Imm;
3185 // SplatBitSize is set to the smallest size that splats the vector, so a
3186 // zero vector will always have SplatBitSize == 8. However, NEON modified
3187 // immediate instructions others than VMOV do not support the 8-bit encoding
3188 // of a zero vector, and the default encoding of zero is supposed to be the
3193 switch (SplatBitSize) {
3195 if (type != VMOVModImm)
3197 // Any 1-byte value is OK. Op=0, Cmode=1110.
3198 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3201 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3205 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3206 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3207 if ((SplatBits & ~0xff) == 0) {
3208 // Value = 0x00nn: Op=x, Cmode=100x.
3213 if ((SplatBits & ~0xff00) == 0) {
3214 // Value = 0xnn00: Op=x, Cmode=101x.
3216 Imm = SplatBits >> 8;
3222 // NEON's 32-bit VMOV supports splat values where:
3223 // * only one byte is nonzero, or
3224 // * the least significant byte is 0xff and the second byte is nonzero, or
3225 // * the least significant 2 bytes are 0xff and the third is nonzero.
3226 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3227 if ((SplatBits & ~0xff) == 0) {
3228 // Value = 0x000000nn: Op=x, Cmode=000x.
3233 if ((SplatBits & ~0xff00) == 0) {
3234 // Value = 0x0000nn00: Op=x, Cmode=001x.
3236 Imm = SplatBits >> 8;
3239 if ((SplatBits & ~0xff0000) == 0) {
3240 // Value = 0x00nn0000: Op=x, Cmode=010x.
3242 Imm = SplatBits >> 16;
3245 if ((SplatBits & ~0xff000000) == 0) {
3246 // Value = 0xnn000000: Op=x, Cmode=011x.
3248 Imm = SplatBits >> 24;
3252 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3253 if (type == OtherModImm) return SDValue();
3255 if ((SplatBits & ~0xffff) == 0 &&
3256 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3257 // Value = 0x0000nnff: Op=x, Cmode=1100.
3259 Imm = SplatBits >> 8;
3264 if ((SplatBits & ~0xffffff) == 0 &&
3265 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3266 // Value = 0x00nnffff: Op=x, Cmode=1101.
3268 Imm = SplatBits >> 16;
3269 SplatBits |= 0xffff;
3273 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3274 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3275 // VMOV.I32. A (very) minor optimization would be to replicate the value
3276 // and fall through here to test for a valid 64-bit splat. But, then the
3277 // caller would also need to check and handle the change in size.
3281 if (type != VMOVModImm)
3283 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3284 uint64_t BitMask = 0xff;
3286 unsigned ImmMask = 1;
3288 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3289 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3292 } else if ((SplatBits & BitMask) != 0) {
3298 // Op=1, Cmode=1110.
3301 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3306 llvm_unreachable("unexpected size for isNEONModifiedImm");
3310 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3311 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3314 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3315 bool &ReverseVEXT, unsigned &Imm) {
3316 unsigned NumElts = VT.getVectorNumElements();
3317 ReverseVEXT = false;
3319 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3325 // If this is a VEXT shuffle, the immediate value is the index of the first
3326 // element. The other shuffle indices must be the successive elements after
3328 unsigned ExpectedElt = Imm;
3329 for (unsigned i = 1; i < NumElts; ++i) {
3330 // Increment the expected index. If it wraps around, it may still be
3331 // a VEXT but the source vectors must be swapped.
3333 if (ExpectedElt == NumElts * 2) {
3338 if (M[i] < 0) continue; // ignore UNDEF indices
3339 if (ExpectedElt != static_cast<unsigned>(M[i]))
3343 // Adjust the index value if the source operands will be swapped.
3350 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3351 /// instruction with the specified blocksize. (The order of the elements
3352 /// within each block of the vector is reversed.)
3353 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3354 unsigned BlockSize) {
3355 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3356 "Only possible block sizes for VREV are: 16, 32, 64");
3358 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3362 unsigned NumElts = VT.getVectorNumElements();
3363 unsigned BlockElts = M[0] + 1;
3364 // If the first shuffle index is UNDEF, be optimistic.
3366 BlockElts = BlockSize / EltSz;
3368 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3371 for (unsigned i = 0; i < NumElts; ++i) {
3372 if (M[i] < 0) continue; // ignore UNDEF indices
3373 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3380 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3381 unsigned &WhichResult) {
3382 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3386 unsigned NumElts = VT.getVectorNumElements();
3387 WhichResult = (M[0] == 0 ? 0 : 1);
3388 for (unsigned i = 0; i < NumElts; i += 2) {
3389 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3390 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3396 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3397 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3398 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3399 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3400 unsigned &WhichResult) {
3401 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3405 unsigned NumElts = VT.getVectorNumElements();
3406 WhichResult = (M[0] == 0 ? 0 : 1);
3407 for (unsigned i = 0; i < NumElts; i += 2) {
3408 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3409 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3415 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3416 unsigned &WhichResult) {
3417 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3421 unsigned NumElts = VT.getVectorNumElements();
3422 WhichResult = (M[0] == 0 ? 0 : 1);
3423 for (unsigned i = 0; i != NumElts; ++i) {
3424 if (M[i] < 0) continue; // ignore UNDEF indices
3425 if ((unsigned) M[i] != 2 * i + WhichResult)
3429 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3430 if (VT.is64BitVector() && EltSz == 32)
3436 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3437 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3438 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3439 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3440 unsigned &WhichResult) {
3441 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3445 unsigned Half = VT.getVectorNumElements() / 2;
3446 WhichResult = (M[0] == 0 ? 0 : 1);
3447 for (unsigned j = 0; j != 2; ++j) {
3448 unsigned Idx = WhichResult;
3449 for (unsigned i = 0; i != Half; ++i) {
3450 int MIdx = M[i + j * Half];
3451 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3457 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3458 if (VT.is64BitVector() && EltSz == 32)
3464 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3465 unsigned &WhichResult) {
3466 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3470 unsigned NumElts = VT.getVectorNumElements();
3471 WhichResult = (M[0] == 0 ? 0 : 1);
3472 unsigned Idx = WhichResult * NumElts / 2;
3473 for (unsigned i = 0; i != NumElts; i += 2) {
3474 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3475 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3480 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3481 if (VT.is64BitVector() && EltSz == 32)
3487 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3488 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3489 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3490 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3491 unsigned &WhichResult) {
3492 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3496 unsigned NumElts = VT.getVectorNumElements();
3497 WhichResult = (M[0] == 0 ? 0 : 1);
3498 unsigned Idx = WhichResult * NumElts / 2;
3499 for (unsigned i = 0; i != NumElts; i += 2) {
3500 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3501 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3506 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3507 if (VT.is64BitVector() && EltSz == 32)
3513 // If N is an integer constant that can be moved into a register in one
3514 // instruction, return an SDValue of such a constant (will become a MOV
3515 // instruction). Otherwise return null.
3516 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3517 const ARMSubtarget *ST, DebugLoc dl) {
3519 if (!isa<ConstantSDNode>(N))
3521 Val = cast<ConstantSDNode>(N)->getZExtValue();
3523 if (ST->isThumb1Only()) {
3524 if (Val <= 255 || ~Val <= 255)
3525 return DAG.getConstant(Val, MVT::i32);
3527 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3528 return DAG.getConstant(Val, MVT::i32);
3533 // If this is a case we can't handle, return null and let the default
3534 // expansion code take care of it.
3535 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3536 const ARMSubtarget *ST) const {
3537 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3538 DebugLoc dl = Op.getDebugLoc();
3539 EVT VT = Op.getValueType();
3541 APInt SplatBits, SplatUndef;
3542 unsigned SplatBitSize;
3544 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3545 if (SplatBitSize <= 64) {
3546 // Check if an immediate VMOV works.
3548 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3549 SplatUndef.getZExtValue(), SplatBitSize,
3550 DAG, VmovVT, VT.is128BitVector(),
3552 if (Val.getNode()) {
3553 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3554 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3557 // Try an immediate VMVN.
3558 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3559 ((1LL << SplatBitSize) - 1));
3560 Val = isNEONModifiedImm(NegatedImm,
3561 SplatUndef.getZExtValue(), SplatBitSize,
3562 DAG, VmovVT, VT.is128BitVector(),
3564 if (Val.getNode()) {
3565 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3566 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3571 // Scan through the operands to see if only one value is used.
3572 unsigned NumElts = VT.getVectorNumElements();
3573 bool isOnlyLowElement = true;
3574 bool usesOnlyOneValue = true;
3575 bool isConstant = true;
3577 for (unsigned i = 0; i < NumElts; ++i) {
3578 SDValue V = Op.getOperand(i);
3579 if (V.getOpcode() == ISD::UNDEF)
3582 isOnlyLowElement = false;
3583 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3586 if (!Value.getNode())
3588 else if (V != Value)
3589 usesOnlyOneValue = false;
3592 if (!Value.getNode())
3593 return DAG.getUNDEF(VT);
3595 if (isOnlyLowElement)
3596 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3598 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3600 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3601 // i32 and try again.
3602 if (usesOnlyOneValue && EltSize <= 32) {
3604 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3605 if (VT.getVectorElementType().isFloatingPoint()) {
3606 SmallVector<SDValue, 8> Ops;
3607 for (unsigned i = 0; i < NumElts; ++i)
3608 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3610 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3611 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3612 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3614 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3616 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3618 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3621 // If all elements are constants and the case above didn't get hit, fall back
3622 // to the default expansion, which will generate a load from the constant
3627 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3629 SDValue shuffle = ReconstructShuffle(Op, DAG);
3630 if (shuffle != SDValue())
3634 // Vectors with 32- or 64-bit elements can be built by directly assigning
3635 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3636 // will be legalized.
3637 if (EltSize >= 32) {
3638 // Do the expansion with floating-point types, since that is what the VFP
3639 // registers are defined to use, and since i64 is not legal.
3640 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3641 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3642 SmallVector<SDValue, 8> Ops;
3643 for (unsigned i = 0; i < NumElts; ++i)
3644 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3645 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3646 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3652 // Gather data to see if the operation can be modelled as a
3653 // shuffle in combination with VEXTs.
3654 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const {
3655 DebugLoc dl = Op.getDebugLoc();
3656 EVT VT = Op.getValueType();
3657 unsigned NumElts = VT.getVectorNumElements();
3659 SmallVector<SDValue, 2> SourceVecs;
3660 SmallVector<unsigned, 2> MinElts;
3661 SmallVector<unsigned, 2> MaxElts;
3663 for (unsigned i = 0; i < NumElts; ++i) {
3664 SDValue V = Op.getOperand(i);
3665 if (V.getOpcode() == ISD::UNDEF)
3667 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
3668 // A shuffle can only come from building a vector from various
3669 // elements of other vectors.
3673 // Record this extraction against the appropriate vector if possible...
3674 SDValue SourceVec = V.getOperand(0);
3675 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
3676 bool FoundSource = false;
3677 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
3678 if (SourceVecs[j] == SourceVec) {
3679 if (MinElts[j] > EltNo)
3681 if (MaxElts[j] < EltNo)
3688 // Or record a new source if not...
3690 SourceVecs.push_back(SourceVec);
3691 MinElts.push_back(EltNo);
3692 MaxElts.push_back(EltNo);
3696 // Currently only do something sane when at most two source vectors
3698 if (SourceVecs.size() > 2)
3701 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
3702 int VEXTOffsets[2] = {0, 0};
3704 // This loop extracts the usage patterns of the source vectors
3705 // and prepares appropriate SDValues for a shuffle if possible.
3706 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
3707 if (SourceVecs[i].getValueType() == VT) {
3708 // No VEXT necessary
3709 ShuffleSrcs[i] = SourceVecs[i];
3712 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
3713 // It probably isn't worth padding out a smaller vector just to
3714 // break it down again in a shuffle.
3718 // Since only 64-bit and 128-bit vectors are legal on ARM and
3719 // we've eliminated the other cases...
3720 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
3721 "unexpected vector sizes in ReconstructShuffle");
3723 if (MaxElts[i] - MinElts[i] >= NumElts) {
3724 // Span too large for a VEXT to cope
3728 if (MinElts[i] >= NumElts) {
3729 // The extraction can just take the second half
3730 VEXTOffsets[i] = NumElts;
3731 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3732 DAG.getIntPtrConstant(NumElts));
3733 } else if (MaxElts[i] < NumElts) {
3734 // The extraction can just take the first half
3736 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3737 DAG.getIntPtrConstant(0));
3739 // An actual VEXT is needed
3740 VEXTOffsets[i] = MinElts[i];
3741 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3742 DAG.getIntPtrConstant(0));
3743 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
3744 DAG.getIntPtrConstant(NumElts));
3745 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
3746 DAG.getConstant(VEXTOffsets[i], MVT::i32));
3750 SmallVector<int, 8> Mask;
3752 for (unsigned i = 0; i < NumElts; ++i) {
3753 SDValue Entry = Op.getOperand(i);
3754 if (Entry.getOpcode() == ISD::UNDEF) {
3759 SDValue ExtractVec = Entry.getOperand(0);
3760 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
3761 if (ExtractVec == SourceVecs[0]) {
3762 Mask.push_back(ExtractElt - VEXTOffsets[0]);
3764 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
3768 // Final check before we try to produce nonsense...
3769 if (isShuffleMaskLegal(Mask, VT))
3770 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1], &Mask[0]);
3775 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3776 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3777 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3778 /// are assumed to be legal.
3780 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3782 if (VT.getVectorNumElements() == 4 &&
3783 (VT.is128BitVector() || VT.is64BitVector())) {
3784 unsigned PFIndexes[4];
3785 for (unsigned i = 0; i != 4; ++i) {
3789 PFIndexes[i] = M[i];
3792 // Compute the index in the perfect shuffle table.
3793 unsigned PFTableIndex =
3794 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3795 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3796 unsigned Cost = (PFEntry >> 30);
3803 unsigned Imm, WhichResult;
3805 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3806 return (EltSize >= 32 ||
3807 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3808 isVREVMask(M, VT, 64) ||
3809 isVREVMask(M, VT, 32) ||
3810 isVREVMask(M, VT, 16) ||
3811 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3812 isVTRNMask(M, VT, WhichResult) ||
3813 isVUZPMask(M, VT, WhichResult) ||
3814 isVZIPMask(M, VT, WhichResult) ||
3815 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3816 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3817 isVZIP_v_undef_Mask(M, VT, WhichResult));
3820 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3821 /// the specified operations to build the shuffle.
3822 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3823 SDValue RHS, SelectionDAG &DAG,
3825 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3826 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3827 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3830 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3839 OP_VUZPL, // VUZP, left result
3840 OP_VUZPR, // VUZP, right result
3841 OP_VZIPL, // VZIP, left result
3842 OP_VZIPR, // VZIP, right result
3843 OP_VTRNL, // VTRN, left result
3844 OP_VTRNR // VTRN, right result
3847 if (OpNum == OP_COPY) {
3848 if (LHSID == (1*9+2)*9+3) return LHS;
3849 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3853 SDValue OpLHS, OpRHS;
3854 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3855 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3856 EVT VT = OpLHS.getValueType();
3859 default: llvm_unreachable("Unknown shuffle opcode!");
3861 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3866 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3867 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3871 return DAG.getNode(ARMISD::VEXT, dl, VT,
3873 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3876 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3877 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3880 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3881 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3884 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3885 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3889 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3890 SDValue V1 = Op.getOperand(0);
3891 SDValue V2 = Op.getOperand(1);
3892 DebugLoc dl = Op.getDebugLoc();
3893 EVT VT = Op.getValueType();
3894 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3895 SmallVector<int, 8> ShuffleMask;
3897 // Convert shuffles that are directly supported on NEON to target-specific
3898 // DAG nodes, instead of keeping them as shuffles and matching them again
3899 // during code selection. This is more efficient and avoids the possibility
3900 // of inconsistencies between legalization and selection.
3901 // FIXME: floating-point vectors should be canonicalized to integer vectors
3902 // of the same time so that they get CSEd properly.
3903 SVN->getMask(ShuffleMask);
3905 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3906 if (EltSize <= 32) {
3907 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3908 int Lane = SVN->getSplatIndex();
3909 // If this is undef splat, generate it via "just" vdup, if possible.
3910 if (Lane == -1) Lane = 0;
3912 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3913 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3915 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3916 DAG.getConstant(Lane, MVT::i32));
3921 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3924 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3925 DAG.getConstant(Imm, MVT::i32));
3928 if (isVREVMask(ShuffleMask, VT, 64))
3929 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3930 if (isVREVMask(ShuffleMask, VT, 32))
3931 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3932 if (isVREVMask(ShuffleMask, VT, 16))
3933 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3935 // Check for Neon shuffles that modify both input vectors in place.
3936 // If both results are used, i.e., if there are two shuffles with the same
3937 // source operands and with masks corresponding to both results of one of
3938 // these operations, DAG memoization will ensure that a single node is
3939 // used for both shuffles.
3940 unsigned WhichResult;
3941 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3942 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3943 V1, V2).getValue(WhichResult);
3944 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3945 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3946 V1, V2).getValue(WhichResult);
3947 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3948 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3949 V1, V2).getValue(WhichResult);
3951 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3952 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3953 V1, V1).getValue(WhichResult);
3954 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3955 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3956 V1, V1).getValue(WhichResult);
3957 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3958 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3959 V1, V1).getValue(WhichResult);
3962 // If the shuffle is not directly supported and it has 4 elements, use
3963 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3964 unsigned NumElts = VT.getVectorNumElements();
3966 unsigned PFIndexes[4];
3967 for (unsigned i = 0; i != 4; ++i) {
3968 if (ShuffleMask[i] < 0)
3971 PFIndexes[i] = ShuffleMask[i];
3974 // Compute the index in the perfect shuffle table.
3975 unsigned PFTableIndex =
3976 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3977 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3978 unsigned Cost = (PFEntry >> 30);
3981 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3984 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3985 if (EltSize >= 32) {
3986 // Do the expansion with floating-point types, since that is what the VFP
3987 // registers are defined to use, and since i64 is not legal.
3988 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3989 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3990 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
3991 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
3992 SmallVector<SDValue, 8> Ops;
3993 for (unsigned i = 0; i < NumElts; ++i) {
3994 if (ShuffleMask[i] < 0)
3995 Ops.push_back(DAG.getUNDEF(EltVT));
3997 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3998 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3999 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4002 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4003 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4009 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4010 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4011 SDValue Lane = Op.getOperand(1);
4012 if (!isa<ConstantSDNode>(Lane))
4015 SDValue Vec = Op.getOperand(0);
4016 if (Op.getValueType() == MVT::i32 &&
4017 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4018 DebugLoc dl = Op.getDebugLoc();
4019 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4025 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4026 // The only time a CONCAT_VECTORS operation can have legal types is when
4027 // two 64-bit vectors are concatenated to a 128-bit vector.
4028 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4029 "unexpected CONCAT_VECTORS");
4030 DebugLoc dl = Op.getDebugLoc();
4031 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4032 SDValue Op0 = Op.getOperand(0);
4033 SDValue Op1 = Op.getOperand(1);
4034 if (Op0.getOpcode() != ISD::UNDEF)
4035 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4036 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4037 DAG.getIntPtrConstant(0));
4038 if (Op1.getOpcode() != ISD::UNDEF)
4039 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4040 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4041 DAG.getIntPtrConstant(1));
4042 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4045 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4046 /// element has been zero/sign-extended, depending on the isSigned parameter,
4047 /// from an integer type half its size.
4048 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4050 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4051 EVT VT = N->getValueType(0);
4052 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4053 SDNode *BVN = N->getOperand(0).getNode();
4054 if (BVN->getValueType(0) != MVT::v4i32 ||
4055 BVN->getOpcode() != ISD::BUILD_VECTOR)
4057 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4058 unsigned HiElt = 1 - LoElt;
4059 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4060 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4061 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4062 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4063 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4066 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4067 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4070 if (Hi0->isNullValue() && Hi1->isNullValue())
4076 if (N->getOpcode() != ISD::BUILD_VECTOR)
4079 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4080 SDNode *Elt = N->getOperand(i).getNode();
4081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4082 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4083 unsigned HalfSize = EltSize / 2;
4085 int64_t SExtVal = C->getSExtValue();
4086 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4089 if ((C->getZExtValue() >> HalfSize) != 0)
4100 /// isSignExtended - Check if a node is a vector value that is sign-extended
4101 /// or a constant BUILD_VECTOR with sign-extended elements.
4102 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4103 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4105 if (isExtendedBUILD_VECTOR(N, DAG, true))
4110 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4111 /// or a constant BUILD_VECTOR with zero-extended elements.
4112 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4113 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4115 if (isExtendedBUILD_VECTOR(N, DAG, false))
4120 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4121 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4122 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4123 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4124 return N->getOperand(0);
4125 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4126 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4127 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4128 LD->isNonTemporal(), LD->getAlignment());
4129 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4130 // have been legalized as a BITCAST from v4i32.
4131 if (N->getOpcode() == ISD::BITCAST) {
4132 SDNode *BVN = N->getOperand(0).getNode();
4133 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4134 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4135 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4136 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4137 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4139 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4140 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4141 EVT VT = N->getValueType(0);
4142 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4143 unsigned NumElts = VT.getVectorNumElements();
4144 MVT TruncVT = MVT::getIntegerVT(EltSize);
4145 SmallVector<SDValue, 8> Ops;
4146 for (unsigned i = 0; i != NumElts; ++i) {
4147 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4148 const APInt &CInt = C->getAPIntValue();
4149 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4151 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4152 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4155 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4156 // Multiplications are only custom-lowered for 128-bit vectors so that
4157 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4158 EVT VT = Op.getValueType();
4159 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4160 SDNode *N0 = Op.getOperand(0).getNode();
4161 SDNode *N1 = Op.getOperand(1).getNode();
4162 unsigned NewOpc = 0;
4163 if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
4164 NewOpc = ARMISD::VMULLs;
4165 else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
4166 NewOpc = ARMISD::VMULLu;
4167 else if (VT == MVT::v2i64)
4168 // Fall through to expand this. It is not legal.
4171 // Other vector multiplications are legal.
4174 // Legalize to a VMULL instruction.
4175 DebugLoc DL = Op.getDebugLoc();
4176 SDValue Op0 = SkipExtension(N0, DAG);
4177 SDValue Op1 = SkipExtension(N1, DAG);
4179 assert(Op0.getValueType().is64BitVector() &&
4180 Op1.getValueType().is64BitVector() &&
4181 "unexpected types for extended operands to VMULL");
4182 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4185 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4186 switch (Op.getOpcode()) {
4187 default: llvm_unreachable("Don't know how to custom lower this!");
4188 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4189 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4190 case ISD::GlobalAddress:
4191 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4192 LowerGlobalAddressELF(Op, DAG);
4193 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4194 case ISD::SELECT: return LowerSELECT(Op, DAG);
4195 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4196 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4197 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4198 case ISD::VASTART: return LowerVASTART(Op, DAG);
4199 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4200 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4201 case ISD::SINT_TO_FP:
4202 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4203 case ISD::FP_TO_SINT:
4204 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4205 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4206 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4207 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4208 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4209 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4210 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4211 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4212 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4214 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4217 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4218 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4219 case ISD::SRL_PARTS:
4220 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4221 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4222 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4223 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4224 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4225 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4226 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4227 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4228 case ISD::MUL: return LowerMUL(Op, DAG);
4233 /// ReplaceNodeResults - Replace the results of node with an illegal result
4234 /// type with new values built out of custom code.
4235 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4236 SmallVectorImpl<SDValue>&Results,
4237 SelectionDAG &DAG) const {
4239 switch (N->getOpcode()) {
4241 llvm_unreachable("Don't know how to custom expand this!");
4244 Res = ExpandBITCAST(N, DAG);
4248 Res = Expand64BitShift(N, DAG, Subtarget);
4252 Results.push_back(Res);
4255 //===----------------------------------------------------------------------===//
4256 // ARM Scheduler Hooks
4257 //===----------------------------------------------------------------------===//
4260 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4261 MachineBasicBlock *BB,
4262 unsigned Size) const {
4263 unsigned dest = MI->getOperand(0).getReg();
4264 unsigned ptr = MI->getOperand(1).getReg();
4265 unsigned oldval = MI->getOperand(2).getReg();
4266 unsigned newval = MI->getOperand(3).getReg();
4267 unsigned scratch = BB->getParent()->getRegInfo()
4268 .createVirtualRegister(ARM::GPRRegisterClass);
4269 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4270 DebugLoc dl = MI->getDebugLoc();
4271 bool isThumb2 = Subtarget->isThumb2();
4273 unsigned ldrOpc, strOpc;
4275 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4277 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4278 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
4281 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4282 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4285 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4286 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4290 MachineFunction *MF = BB->getParent();
4291 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4292 MachineFunction::iterator It = BB;
4293 ++It; // insert the new blocks after the current block
4295 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4296 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4297 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4298 MF->insert(It, loop1MBB);
4299 MF->insert(It, loop2MBB);
4300 MF->insert(It, exitMBB);
4302 // Transfer the remainder of BB and its successor edges to exitMBB.
4303 exitMBB->splice(exitMBB->begin(), BB,
4304 llvm::next(MachineBasicBlock::iterator(MI)),
4306 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4310 // fallthrough --> loop1MBB
4311 BB->addSuccessor(loop1MBB);
4314 // ldrex dest, [ptr]
4318 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4319 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4320 .addReg(dest).addReg(oldval));
4321 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4322 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4323 BB->addSuccessor(loop2MBB);
4324 BB->addSuccessor(exitMBB);
4327 // strex scratch, newval, [ptr]
4331 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4333 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4334 .addReg(scratch).addImm(0));
4335 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4336 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4337 BB->addSuccessor(loop1MBB);
4338 BB->addSuccessor(exitMBB);
4344 MI->eraseFromParent(); // The instruction is gone now.
4350 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4351 unsigned Size, unsigned BinOpcode) const {
4352 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4353 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4355 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4356 MachineFunction *MF = BB->getParent();
4357 MachineFunction::iterator It = BB;
4360 unsigned dest = MI->getOperand(0).getReg();
4361 unsigned ptr = MI->getOperand(1).getReg();
4362 unsigned incr = MI->getOperand(2).getReg();
4363 DebugLoc dl = MI->getDebugLoc();
4365 bool isThumb2 = Subtarget->isThumb2();
4366 unsigned ldrOpc, strOpc;
4368 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4370 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4371 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4374 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4375 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4378 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4379 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4383 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4384 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4385 MF->insert(It, loopMBB);
4386 MF->insert(It, exitMBB);
4388 // Transfer the remainder of BB and its successor edges to exitMBB.
4389 exitMBB->splice(exitMBB->begin(), BB,
4390 llvm::next(MachineBasicBlock::iterator(MI)),
4392 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4394 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4395 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4396 unsigned scratch2 = (!BinOpcode) ? incr :
4397 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4401 // fallthrough --> loopMBB
4402 BB->addSuccessor(loopMBB);
4406 // <binop> scratch2, dest, incr
4407 // strex scratch, scratch2, ptr
4410 // fallthrough --> exitMBB
4412 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4414 // operand order needs to go the other way for NAND
4415 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4416 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4417 addReg(incr).addReg(dest)).addReg(0);
4419 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4420 addReg(dest).addReg(incr)).addReg(0);
4423 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4425 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4426 .addReg(scratch).addImm(0));
4427 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4428 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4430 BB->addSuccessor(loopMBB);
4431 BB->addSuccessor(exitMBB);
4437 MI->eraseFromParent(); // The instruction is gone now.
4443 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4444 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4445 E = MBB->succ_end(); I != E; ++I)
4448 llvm_unreachable("Expecting a BB with two successors!");
4452 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4453 MachineBasicBlock *BB) const {
4454 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4455 DebugLoc dl = MI->getDebugLoc();
4456 bool isThumb2 = Subtarget->isThumb2();
4457 switch (MI->getOpcode()) {
4460 llvm_unreachable("Unexpected instr type to insert");
4462 case ARM::ATOMIC_LOAD_ADD_I8:
4463 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4464 case ARM::ATOMIC_LOAD_ADD_I16:
4465 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4466 case ARM::ATOMIC_LOAD_ADD_I32:
4467 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4469 case ARM::ATOMIC_LOAD_AND_I8:
4470 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4471 case ARM::ATOMIC_LOAD_AND_I16:
4472 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4473 case ARM::ATOMIC_LOAD_AND_I32:
4474 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4476 case ARM::ATOMIC_LOAD_OR_I8:
4477 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4478 case ARM::ATOMIC_LOAD_OR_I16:
4479 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4480 case ARM::ATOMIC_LOAD_OR_I32:
4481 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4483 case ARM::ATOMIC_LOAD_XOR_I8:
4484 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4485 case ARM::ATOMIC_LOAD_XOR_I16:
4486 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4487 case ARM::ATOMIC_LOAD_XOR_I32:
4488 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4490 case ARM::ATOMIC_LOAD_NAND_I8:
4491 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4492 case ARM::ATOMIC_LOAD_NAND_I16:
4493 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4494 case ARM::ATOMIC_LOAD_NAND_I32:
4495 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4497 case ARM::ATOMIC_LOAD_SUB_I8:
4498 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4499 case ARM::ATOMIC_LOAD_SUB_I16:
4500 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4501 case ARM::ATOMIC_LOAD_SUB_I32:
4502 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4504 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4505 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4506 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4508 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4509 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4510 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4512 case ARM::tMOVCCr_pseudo: {
4513 // To "insert" a SELECT_CC instruction, we actually have to insert the
4514 // diamond control-flow pattern. The incoming instruction knows the
4515 // destination vreg to set, the condition code register to branch on, the
4516 // true/false values to select between, and a branch opcode to use.
4517 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4518 MachineFunction::iterator It = BB;
4524 // cmpTY ccX, r1, r2
4526 // fallthrough --> copy0MBB
4527 MachineBasicBlock *thisMBB = BB;
4528 MachineFunction *F = BB->getParent();
4529 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4530 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4531 F->insert(It, copy0MBB);
4532 F->insert(It, sinkMBB);
4534 // Transfer the remainder of BB and its successor edges to sinkMBB.
4535 sinkMBB->splice(sinkMBB->begin(), BB,
4536 llvm::next(MachineBasicBlock::iterator(MI)),
4538 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4540 BB->addSuccessor(copy0MBB);
4541 BB->addSuccessor(sinkMBB);
4543 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4544 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4547 // %FalseValue = ...
4548 // # fallthrough to sinkMBB
4551 // Update machine-CFG edges
4552 BB->addSuccessor(sinkMBB);
4555 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4558 BuildMI(*BB, BB->begin(), dl,
4559 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4560 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4561 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4563 MI->eraseFromParent(); // The pseudo instruction is gone now.
4568 case ARM::BCCZi64: {
4569 // If there is an unconditional branch to the other successor, remove it.
4570 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
4572 // Compare both parts that make up the double comparison separately for
4574 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4576 unsigned LHS1 = MI->getOperand(1).getReg();
4577 unsigned LHS2 = MI->getOperand(2).getReg();
4579 AddDefaultPred(BuildMI(BB, dl,
4580 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4581 .addReg(LHS1).addImm(0));
4582 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4583 .addReg(LHS2).addImm(0)
4584 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4586 unsigned RHS1 = MI->getOperand(3).getReg();
4587 unsigned RHS2 = MI->getOperand(4).getReg();
4588 AddDefaultPred(BuildMI(BB, dl,
4589 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4590 .addReg(LHS1).addReg(RHS1));
4591 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4592 .addReg(LHS2).addReg(RHS2)
4593 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4596 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4597 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4598 if (MI->getOperand(0).getImm() == ARMCC::NE)
4599 std::swap(destMBB, exitMBB);
4601 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4602 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4603 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4606 MI->eraseFromParent(); // The pseudo instruction is gone now.
4612 //===----------------------------------------------------------------------===//
4613 // ARM Optimization Hooks
4614 //===----------------------------------------------------------------------===//
4617 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4618 TargetLowering::DAGCombinerInfo &DCI) {
4619 SelectionDAG &DAG = DCI.DAG;
4620 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4621 EVT VT = N->getValueType(0);
4622 unsigned Opc = N->getOpcode();
4623 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4624 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4625 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4626 ISD::CondCode CC = ISD::SETCC_INVALID;
4629 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4631 SDValue CCOp = Slct.getOperand(0);
4632 if (CCOp.getOpcode() == ISD::SETCC)
4633 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4636 bool DoXform = false;
4638 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4641 if (LHS.getOpcode() == ISD::Constant &&
4642 cast<ConstantSDNode>(LHS)->isNullValue()) {
4644 } else if (CC != ISD::SETCC_INVALID &&
4645 RHS.getOpcode() == ISD::Constant &&
4646 cast<ConstantSDNode>(RHS)->isNullValue()) {
4647 std::swap(LHS, RHS);
4648 SDValue Op0 = Slct.getOperand(0);
4649 EVT OpVT = isSlctCC ? Op0.getValueType() :
4650 Op0.getOperand(0).getValueType();
4651 bool isInt = OpVT.isInteger();
4652 CC = ISD::getSetCCInverse(CC, isInt);
4654 if (!TLI.isCondCodeLegal(CC, OpVT))
4655 return SDValue(); // Inverse operator isn't legal.
4662 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4664 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4665 Slct.getOperand(0), Slct.getOperand(1), CC);
4666 SDValue CCOp = Slct.getOperand(0);
4668 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4669 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4670 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4671 CCOp, OtherOp, Result);
4676 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4677 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4678 /// called with the default operands, and if that fails, with commuted
4680 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4681 TargetLowering::DAGCombinerInfo &DCI) {
4682 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4683 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4684 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4685 if (Result.getNode()) return Result;
4690 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4692 static SDValue PerformADDCombine(SDNode *N,
4693 TargetLowering::DAGCombinerInfo &DCI) {
4694 SDValue N0 = N->getOperand(0);
4695 SDValue N1 = N->getOperand(1);
4697 // First try with the default operand order.
4698 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4699 if (Result.getNode())
4702 // If that didn't work, try again with the operands commuted.
4703 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4706 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4708 static SDValue PerformSUBCombine(SDNode *N,
4709 TargetLowering::DAGCombinerInfo &DCI) {
4710 SDValue N0 = N->getOperand(0);
4711 SDValue N1 = N->getOperand(1);
4713 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4714 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4715 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4716 if (Result.getNode()) return Result;
4722 static SDValue PerformMULCombine(SDNode *N,
4723 TargetLowering::DAGCombinerInfo &DCI,
4724 const ARMSubtarget *Subtarget) {
4725 SelectionDAG &DAG = DCI.DAG;
4727 if (Subtarget->isThumb1Only())
4730 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4733 EVT VT = N->getValueType(0);
4737 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4741 uint64_t MulAmt = C->getZExtValue();
4742 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4743 ShiftAmt = ShiftAmt & (32 - 1);
4744 SDValue V = N->getOperand(0);
4745 DebugLoc DL = N->getDebugLoc();
4748 MulAmt >>= ShiftAmt;
4749 if (isPowerOf2_32(MulAmt - 1)) {
4750 // (mul x, 2^N + 1) => (add (shl x, N), x)
4751 Res = DAG.getNode(ISD::ADD, DL, VT,
4752 V, DAG.getNode(ISD::SHL, DL, VT,
4753 V, DAG.getConstant(Log2_32(MulAmt-1),
4755 } else if (isPowerOf2_32(MulAmt + 1)) {
4756 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4757 Res = DAG.getNode(ISD::SUB, DL, VT,
4758 DAG.getNode(ISD::SHL, DL, VT,
4759 V, DAG.getConstant(Log2_32(MulAmt+1),
4766 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4767 DAG.getConstant(ShiftAmt, MVT::i32));
4769 // Do not add new nodes to DAG combiner worklist.
4770 DCI.CombineTo(N, Res, false);
4774 static SDValue PerformANDCombine(SDNode *N,
4775 TargetLowering::DAGCombinerInfo &DCI) {
4776 // Attempt to use immediate-form VBIC
4777 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
4778 DebugLoc dl = N->getDebugLoc();
4779 EVT VT = N->getValueType(0);
4780 SelectionDAG &DAG = DCI.DAG;
4782 APInt SplatBits, SplatUndef;
4783 unsigned SplatBitSize;
4786 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4787 if (SplatBitSize <= 64) {
4789 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
4790 SplatUndef.getZExtValue(), SplatBitSize,
4791 DAG, VbicVT, VT.is128BitVector(),
4793 if (Val.getNode()) {
4795 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
4796 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
4797 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
4805 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4806 static SDValue PerformORCombine(SDNode *N,
4807 TargetLowering::DAGCombinerInfo &DCI,
4808 const ARMSubtarget *Subtarget) {
4809 // Attempt to use immediate-form VORR
4810 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
4811 DebugLoc dl = N->getDebugLoc();
4812 EVT VT = N->getValueType(0);
4813 SelectionDAG &DAG = DCI.DAG;
4815 APInt SplatBits, SplatUndef;
4816 unsigned SplatBitSize;
4818 if (BVN && Subtarget->hasNEON() &&
4819 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4820 if (SplatBitSize <= 64) {
4822 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4823 SplatUndef.getZExtValue(), SplatBitSize,
4824 DAG, VorrVT, VT.is128BitVector(),
4826 if (Val.getNode()) {
4828 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
4829 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
4830 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
4835 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4838 // BFI is only available on V6T2+
4839 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4842 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4843 DebugLoc DL = N->getDebugLoc();
4844 // 1) or (and A, mask), val => ARMbfi A, val, mask
4845 // iff (val & mask) == val
4847 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4848 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4849 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4850 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4851 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4852 // (i.e., copy a bitfield value into another bitfield of the same width)
4853 if (N0.getOpcode() != ISD::AND)
4859 SDValue N00 = N0.getOperand(0);
4861 // The value and the mask need to be constants so we can verify this is
4862 // actually a bitfield set. If the mask is 0xffff, we can do better
4863 // via a movt instruction, so don't use BFI in that case.
4864 SDValue MaskOp = N0.getOperand(1);
4865 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
4868 unsigned Mask = MaskC->getZExtValue();
4872 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4873 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4875 unsigned Val = N1C->getZExtValue();
4876 if ((Val & ~Mask) != Val)
4879 if (ARM::isBitFieldInvertedMask(Mask)) {
4880 Val >>= CountTrailingZeros_32(~Mask);
4882 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
4883 DAG.getConstant(Val, MVT::i32),
4884 DAG.getConstant(Mask, MVT::i32));
4886 // Do not add new nodes to DAG combiner worklist.
4887 DCI.CombineTo(N, Res, false);
4890 } else if (N1.getOpcode() == ISD::AND) {
4891 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4892 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4895 unsigned Mask2 = N11C->getZExtValue();
4897 if (ARM::isBitFieldInvertedMask(Mask) &&
4898 ARM::isBitFieldInvertedMask(~Mask2) &&
4899 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4900 // The pack halfword instruction works better for masks that fit it,
4901 // so use that when it's available.
4902 if (Subtarget->hasT2ExtractPack() &&
4903 (Mask == 0xffff || Mask == 0xffff0000))
4906 unsigned lsb = CountTrailingZeros_32(Mask2);
4907 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4908 DAG.getConstant(lsb, MVT::i32));
4909 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
4910 DAG.getConstant(Mask, MVT::i32));
4911 // Do not add new nodes to DAG combiner worklist.
4912 DCI.CombineTo(N, Res, false);
4914 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4915 ARM::isBitFieldInvertedMask(Mask2) &&
4916 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4917 // The pack halfword instruction works better for masks that fit it,
4918 // so use that when it's available.
4919 if (Subtarget->hasT2ExtractPack() &&
4920 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4923 unsigned lsb = CountTrailingZeros_32(Mask);
4924 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
4925 DAG.getConstant(lsb, MVT::i32));
4926 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4927 DAG.getConstant(Mask2, MVT::i32));
4928 // Do not add new nodes to DAG combiner worklist.
4929 DCI.CombineTo(N, Res, false);
4934 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
4935 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
4936 ARM::isBitFieldInvertedMask(~Mask)) {
4937 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
4938 // where lsb(mask) == #shamt and masked bits of B are known zero.
4939 SDValue ShAmt = N00.getOperand(1);
4940 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4941 unsigned LSB = CountTrailingZeros_32(Mask);
4945 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
4946 DAG.getConstant(~Mask, MVT::i32));
4948 // Do not add new nodes to DAG combiner worklist.
4949 DCI.CombineTo(N, Res, false);
4955 /// PerformBFICombine - (bfi A, (and B, C1), C2) -> (bfi A, B, C2) iff
4957 static SDValue PerformBFICombine(SDNode *N,
4958 TargetLowering::DAGCombinerInfo &DCI) {
4959 SDValue N1 = N->getOperand(1);
4960 if (N1.getOpcode() == ISD::AND) {
4961 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4964 unsigned Mask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
4965 unsigned Mask2 = N11C->getZExtValue();
4966 if ((Mask & Mask2) == Mask2)
4967 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
4968 N->getOperand(0), N1.getOperand(0),
4974 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4975 /// ARMISD::VMOVRRD.
4976 static SDValue PerformVMOVRRDCombine(SDNode *N,
4977 TargetLowering::DAGCombinerInfo &DCI) {
4978 // vmovrrd(vmovdrr x, y) -> x,y
4979 SDValue InDouble = N->getOperand(0);
4980 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4981 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4985 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4986 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4987 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4988 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4989 SDValue Op0 = N->getOperand(0);
4990 SDValue Op1 = N->getOperand(1);
4991 if (Op0.getOpcode() == ISD::BITCAST)
4992 Op0 = Op0.getOperand(0);
4993 if (Op1.getOpcode() == ISD::BITCAST)
4994 Op1 = Op1.getOperand(0);
4995 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4996 Op0.getNode() == Op1.getNode() &&
4997 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4998 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
4999 N->getValueType(0), Op0.getOperand(0));
5003 /// PerformSTORECombine - Target-specific dag combine xforms for
5005 static SDValue PerformSTORECombine(SDNode *N,
5006 TargetLowering::DAGCombinerInfo &DCI) {
5007 // Bitcast an i64 store extracted from a vector to f64.
5008 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5009 StoreSDNode *St = cast<StoreSDNode>(N);
5010 SDValue StVal = St->getValue();
5011 if (!ISD::isNormalStore(St) || St->isVolatile() ||
5012 StVal.getValueType() != MVT::i64 ||
5013 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5016 SelectionDAG &DAG = DCI.DAG;
5017 DebugLoc dl = StVal.getDebugLoc();
5018 SDValue IntVec = StVal.getOperand(0);
5019 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5020 IntVec.getValueType().getVectorNumElements());
5021 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
5022 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5023 Vec, StVal.getOperand(1));
5024 dl = N->getDebugLoc();
5025 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
5026 // Make the DAGCombiner fold the bitcasts.
5027 DCI.AddToWorklist(Vec.getNode());
5028 DCI.AddToWorklist(ExtElt.getNode());
5029 DCI.AddToWorklist(V.getNode());
5030 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
5031 St->getPointerInfo(), St->isVolatile(),
5032 St->isNonTemporal(), St->getAlignment(),
5036 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
5037 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
5038 /// i64 vector to have f64 elements, since the value can then be loaded
5039 /// directly into a VFP register.
5040 static bool hasNormalLoadOperand(SDNode *N) {
5041 unsigned NumElts = N->getValueType(0).getVectorNumElements();
5042 for (unsigned i = 0; i < NumElts; ++i) {
5043 SDNode *Elt = N->getOperand(i).getNode();
5044 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
5050 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
5051 /// ISD::BUILD_VECTOR.
5052 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
5053 TargetLowering::DAGCombinerInfo &DCI){
5054 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
5055 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
5056 // into a pair of GPRs, which is fine when the value is used as a scalar,
5057 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
5058 SelectionDAG &DAG = DCI.DAG;
5059 if (N->getNumOperands() == 2) {
5060 SDValue RV = PerformVMOVDRRCombine(N, DAG);
5065 // Load i64 elements as f64 values so that type legalization does not split
5066 // them up into i32 values.
5067 EVT VT = N->getValueType(0);
5068 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
5070 DebugLoc dl = N->getDebugLoc();
5071 SmallVector<SDValue, 8> Ops;
5072 unsigned NumElts = VT.getVectorNumElements();
5073 for (unsigned i = 0; i < NumElts; ++i) {
5074 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
5076 // Make the DAGCombiner fold the bitcast.
5077 DCI.AddToWorklist(V.getNode());
5079 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
5080 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
5081 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
5084 /// PerformInsertEltCombine - Target-specific dag combine xforms for
5085 /// ISD::INSERT_VECTOR_ELT.
5086 static SDValue PerformInsertEltCombine(SDNode *N,
5087 TargetLowering::DAGCombinerInfo &DCI) {
5088 // Bitcast an i64 load inserted into a vector to f64.
5089 // Otherwise, the i64 value will be legalized to a pair of i32 values.
5090 EVT VT = N->getValueType(0);
5091 SDNode *Elt = N->getOperand(1).getNode();
5092 if (VT.getVectorElementType() != MVT::i64 ||
5093 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
5096 SelectionDAG &DAG = DCI.DAG;
5097 DebugLoc dl = N->getDebugLoc();
5098 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
5099 VT.getVectorNumElements());
5100 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
5101 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
5102 // Make the DAGCombiner fold the bitcasts.
5103 DCI.AddToWorklist(Vec.getNode());
5104 DCI.AddToWorklist(V.getNode());
5105 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
5106 Vec, V, N->getOperand(2));
5107 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
5110 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
5111 /// ISD::VECTOR_SHUFFLE.
5112 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
5113 // The LLVM shufflevector instruction does not require the shuffle mask
5114 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
5115 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
5116 // operands do not match the mask length, they are extended by concatenating
5117 // them with undef vectors. That is probably the right thing for other
5118 // targets, but for NEON it is better to concatenate two double-register
5119 // size vector operands into a single quad-register size vector. Do that
5120 // transformation here:
5121 // shuffle(concat(v1, undef), concat(v2, undef)) ->
5122 // shuffle(concat(v1, v2), undef)
5123 SDValue Op0 = N->getOperand(0);
5124 SDValue Op1 = N->getOperand(1);
5125 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
5126 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
5127 Op0.getNumOperands() != 2 ||
5128 Op1.getNumOperands() != 2)
5130 SDValue Concat0Op1 = Op0.getOperand(1);
5131 SDValue Concat1Op1 = Op1.getOperand(1);
5132 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
5133 Concat1Op1.getOpcode() != ISD::UNDEF)
5135 // Skip the transformation if any of the types are illegal.
5136 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5137 EVT VT = N->getValueType(0);
5138 if (!TLI.isTypeLegal(VT) ||
5139 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
5140 !TLI.isTypeLegal(Concat1Op1.getValueType()))
5143 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5144 Op0.getOperand(0), Op1.getOperand(0));
5145 // Translate the shuffle mask.
5146 SmallVector<int, 16> NewMask;
5147 unsigned NumElts = VT.getVectorNumElements();
5148 unsigned HalfElts = NumElts/2;
5149 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
5150 for (unsigned n = 0; n < NumElts; ++n) {
5151 int MaskElt = SVN->getMaskElt(n);
5153 if (MaskElt < (int)HalfElts)
5155 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
5156 NewElt = HalfElts + MaskElt - NumElts;
5157 NewMask.push_back(NewElt);
5159 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
5160 DAG.getUNDEF(VT), NewMask.data());
5163 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
5164 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
5165 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
5167 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
5168 SelectionDAG &DAG = DCI.DAG;
5169 EVT VT = N->getValueType(0);
5170 // vldN-dup instructions only support 64-bit vectors for N > 1.
5171 if (!VT.is64BitVector())
5174 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
5175 SDNode *VLD = N->getOperand(0).getNode();
5176 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
5178 unsigned NumVecs = 0;
5179 unsigned NewOpc = 0;
5180 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
5181 if (IntNo == Intrinsic::arm_neon_vld2lane) {
5183 NewOpc = ARMISD::VLD2DUP;
5184 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
5186 NewOpc = ARMISD::VLD3DUP;
5187 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
5189 NewOpc = ARMISD::VLD4DUP;
5194 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
5195 // numbers match the load.
5196 unsigned VLDLaneNo =
5197 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
5198 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5200 // Ignore uses of the chain result.
5201 if (UI.getUse().getResNo() == NumVecs)
5204 if (User->getOpcode() != ARMISD::VDUPLANE ||
5205 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
5209 // Create the vldN-dup node.
5212 for (n = 0; n < NumVecs; ++n)
5214 Tys[n] = MVT::Other;
5215 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
5216 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
5217 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
5218 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
5219 Ops, 2, VLDMemInt->getMemoryVT(),
5220 VLDMemInt->getMemOperand());
5223 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
5225 unsigned ResNo = UI.getUse().getResNo();
5226 // Ignore uses of the chain result.
5227 if (ResNo == NumVecs)
5230 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
5233 // Now the vldN-lane intrinsic is dead except for its chain result.
5234 // Update uses of the chain.
5235 std::vector<SDValue> VLDDupResults;
5236 for (unsigned n = 0; n < NumVecs; ++n)
5237 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
5238 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
5239 DCI.CombineTo(VLD, VLDDupResults);
5244 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
5245 /// ARMISD::VDUPLANE.
5246 static SDValue PerformVDUPLANECombine(SDNode *N,
5247 TargetLowering::DAGCombinerInfo &DCI) {
5248 SDValue Op = N->getOperand(0);
5250 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
5251 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
5252 if (CombineVLDDUP(N, DCI))
5253 return SDValue(N, 0);
5255 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
5256 // redundant. Ignore bit_converts for now; element sizes are checked below.
5257 while (Op.getOpcode() == ISD::BITCAST)
5258 Op = Op.getOperand(0);
5259 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
5262 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
5263 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
5264 // The canonical VMOV for a zero vector uses a 32-bit element size.
5265 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5267 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
5269 EVT VT = N->getValueType(0);
5270 if (EltSize > VT.getVectorElementType().getSizeInBits())
5273 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
5276 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5277 /// operand of a vector shift operation, where all the elements of the
5278 /// build_vector must have the same constant integer value.
5279 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5280 // Ignore bit_converts.
5281 while (Op.getOpcode() == ISD::BITCAST)
5282 Op = Op.getOperand(0);
5283 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5284 APInt SplatBits, SplatUndef;
5285 unsigned SplatBitSize;
5287 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5288 HasAnyUndefs, ElementBits) ||
5289 SplatBitSize > ElementBits)
5291 Cnt = SplatBits.getSExtValue();
5295 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5296 /// operand of a vector shift left operation. That value must be in the range:
5297 /// 0 <= Value < ElementBits for a left shift; or
5298 /// 0 <= Value <= ElementBits for a long left shift.
5299 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5300 assert(VT.isVector() && "vector shift count is not a vector type");
5301 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5302 if (! getVShiftImm(Op, ElementBits, Cnt))
5304 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
5307 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5308 /// operand of a vector shift right operation. For a shift opcode, the value
5309 /// is positive, but for an intrinsic the value count must be negative. The
5310 /// absolute value must be in the range:
5311 /// 1 <= |Value| <= ElementBits for a right shift; or
5312 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5313 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5315 assert(VT.isVector() && "vector shift count is not a vector type");
5316 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5317 if (! getVShiftImm(Op, ElementBits, Cnt))
5321 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
5324 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
5325 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
5326 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5329 // Don't do anything for most intrinsics.
5332 // Vector shifts: check for immediate versions and lower them.
5333 // Note: This is done during DAG combining instead of DAG legalizing because
5334 // the build_vectors for 64-bit vector element shift counts are generally
5335 // not legal, and it is hard to see their values after they get legalized to
5336 // loads from a constant pool.
5337 case Intrinsic::arm_neon_vshifts:
5338 case Intrinsic::arm_neon_vshiftu:
5339 case Intrinsic::arm_neon_vshiftls:
5340 case Intrinsic::arm_neon_vshiftlu:
5341 case Intrinsic::arm_neon_vshiftn:
5342 case Intrinsic::arm_neon_vrshifts:
5343 case Intrinsic::arm_neon_vrshiftu:
5344 case Intrinsic::arm_neon_vrshiftn:
5345 case Intrinsic::arm_neon_vqshifts:
5346 case Intrinsic::arm_neon_vqshiftu:
5347 case Intrinsic::arm_neon_vqshiftsu:
5348 case Intrinsic::arm_neon_vqshiftns:
5349 case Intrinsic::arm_neon_vqshiftnu:
5350 case Intrinsic::arm_neon_vqshiftnsu:
5351 case Intrinsic::arm_neon_vqrshiftns:
5352 case Intrinsic::arm_neon_vqrshiftnu:
5353 case Intrinsic::arm_neon_vqrshiftnsu: {
5354 EVT VT = N->getOperand(1).getValueType();
5356 unsigned VShiftOpc = 0;
5359 case Intrinsic::arm_neon_vshifts:
5360 case Intrinsic::arm_neon_vshiftu:
5361 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
5362 VShiftOpc = ARMISD::VSHL;
5365 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
5366 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
5367 ARMISD::VSHRs : ARMISD::VSHRu);
5372 case Intrinsic::arm_neon_vshiftls:
5373 case Intrinsic::arm_neon_vshiftlu:
5374 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
5376 llvm_unreachable("invalid shift count for vshll intrinsic");
5378 case Intrinsic::arm_neon_vrshifts:
5379 case Intrinsic::arm_neon_vrshiftu:
5380 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
5384 case Intrinsic::arm_neon_vqshifts:
5385 case Intrinsic::arm_neon_vqshiftu:
5386 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5390 case Intrinsic::arm_neon_vqshiftsu:
5391 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5393 llvm_unreachable("invalid shift count for vqshlu intrinsic");
5395 case Intrinsic::arm_neon_vshiftn:
5396 case Intrinsic::arm_neon_vrshiftn:
5397 case Intrinsic::arm_neon_vqshiftns:
5398 case Intrinsic::arm_neon_vqshiftnu:
5399 case Intrinsic::arm_neon_vqshiftnsu:
5400 case Intrinsic::arm_neon_vqrshiftns:
5401 case Intrinsic::arm_neon_vqrshiftnu:
5402 case Intrinsic::arm_neon_vqrshiftnsu:
5403 // Narrowing shifts require an immediate right shift.
5404 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
5406 llvm_unreachable("invalid shift count for narrowing vector shift "
5410 llvm_unreachable("unhandled vector shift");
5414 case Intrinsic::arm_neon_vshifts:
5415 case Intrinsic::arm_neon_vshiftu:
5416 // Opcode already set above.
5418 case Intrinsic::arm_neon_vshiftls:
5419 case Intrinsic::arm_neon_vshiftlu:
5420 if (Cnt == VT.getVectorElementType().getSizeInBits())
5421 VShiftOpc = ARMISD::VSHLLi;
5423 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
5424 ARMISD::VSHLLs : ARMISD::VSHLLu);
5426 case Intrinsic::arm_neon_vshiftn:
5427 VShiftOpc = ARMISD::VSHRN; break;
5428 case Intrinsic::arm_neon_vrshifts:
5429 VShiftOpc = ARMISD::VRSHRs; break;
5430 case Intrinsic::arm_neon_vrshiftu:
5431 VShiftOpc = ARMISD::VRSHRu; break;
5432 case Intrinsic::arm_neon_vrshiftn:
5433 VShiftOpc = ARMISD::VRSHRN; break;
5434 case Intrinsic::arm_neon_vqshifts:
5435 VShiftOpc = ARMISD::VQSHLs; break;
5436 case Intrinsic::arm_neon_vqshiftu:
5437 VShiftOpc = ARMISD::VQSHLu; break;
5438 case Intrinsic::arm_neon_vqshiftsu:
5439 VShiftOpc = ARMISD::VQSHLsu; break;
5440 case Intrinsic::arm_neon_vqshiftns:
5441 VShiftOpc = ARMISD::VQSHRNs; break;
5442 case Intrinsic::arm_neon_vqshiftnu:
5443 VShiftOpc = ARMISD::VQSHRNu; break;
5444 case Intrinsic::arm_neon_vqshiftnsu:
5445 VShiftOpc = ARMISD::VQSHRNsu; break;
5446 case Intrinsic::arm_neon_vqrshiftns:
5447 VShiftOpc = ARMISD::VQRSHRNs; break;
5448 case Intrinsic::arm_neon_vqrshiftnu:
5449 VShiftOpc = ARMISD::VQRSHRNu; break;
5450 case Intrinsic::arm_neon_vqrshiftnsu:
5451 VShiftOpc = ARMISD::VQRSHRNsu; break;
5454 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5455 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
5458 case Intrinsic::arm_neon_vshiftins: {
5459 EVT VT = N->getOperand(1).getValueType();
5461 unsigned VShiftOpc = 0;
5463 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
5464 VShiftOpc = ARMISD::VSLI;
5465 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
5466 VShiftOpc = ARMISD::VSRI;
5468 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
5471 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5472 N->getOperand(1), N->getOperand(2),
5473 DAG.getConstant(Cnt, MVT::i32));
5476 case Intrinsic::arm_neon_vqrshifts:
5477 case Intrinsic::arm_neon_vqrshiftu:
5478 // No immediate versions of these to check for.
5485 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
5486 /// lowers them. As with the vector shift intrinsics, this is done during DAG
5487 /// combining instead of DAG legalizing because the build_vectors for 64-bit
5488 /// vector element shift counts are generally not legal, and it is hard to see
5489 /// their values after they get legalized to loads from a constant pool.
5490 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
5491 const ARMSubtarget *ST) {
5492 EVT VT = N->getValueType(0);
5494 // Nothing to be done for scalar shifts.
5495 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5496 if (!VT.isVector() || !TLI.isTypeLegal(VT))
5499 assert(ST->hasNEON() && "unexpected vector shift");
5502 switch (N->getOpcode()) {
5503 default: llvm_unreachable("unexpected shift opcode");
5506 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
5507 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
5508 DAG.getConstant(Cnt, MVT::i32));
5513 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
5514 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
5515 ARMISD::VSHRs : ARMISD::VSHRu);
5516 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
5517 DAG.getConstant(Cnt, MVT::i32));
5523 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
5524 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
5525 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
5526 const ARMSubtarget *ST) {
5527 SDValue N0 = N->getOperand(0);
5529 // Check for sign- and zero-extensions of vector extract operations of 8-
5530 // and 16-bit vector elements. NEON supports these directly. They are
5531 // handled during DAG combining because type legalization will promote them
5532 // to 32-bit types and it is messy to recognize the operations after that.
5533 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
5534 SDValue Vec = N0.getOperand(0);
5535 SDValue Lane = N0.getOperand(1);
5536 EVT VT = N->getValueType(0);
5537 EVT EltVT = N0.getValueType();
5538 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5540 if (VT == MVT::i32 &&
5541 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
5542 TLI.isTypeLegal(Vec.getValueType()) &&
5543 isa<ConstantSDNode>(Lane)) {
5546 switch (N->getOpcode()) {
5547 default: llvm_unreachable("unexpected opcode");
5548 case ISD::SIGN_EXTEND:
5549 Opc = ARMISD::VGETLANEs;
5551 case ISD::ZERO_EXTEND:
5552 case ISD::ANY_EXTEND:
5553 Opc = ARMISD::VGETLANEu;
5556 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
5563 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
5564 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
5565 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
5566 const ARMSubtarget *ST) {
5567 // If the target supports NEON, try to use vmax/vmin instructions for f32
5568 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
5569 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
5570 // a NaN; only do the transformation when it matches that behavior.
5572 // For now only do this when using NEON for FP operations; if using VFP, it
5573 // is not obvious that the benefit outweighs the cost of switching to the
5575 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
5576 N->getValueType(0) != MVT::f32)
5579 SDValue CondLHS = N->getOperand(0);
5580 SDValue CondRHS = N->getOperand(1);
5581 SDValue LHS = N->getOperand(2);
5582 SDValue RHS = N->getOperand(3);
5583 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
5585 unsigned Opcode = 0;
5587 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
5588 IsReversed = false; // x CC y ? x : y
5589 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
5590 IsReversed = true ; // x CC y ? y : x
5604 // If LHS is NaN, an ordered comparison will be false and the result will
5605 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
5606 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5607 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
5608 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5610 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
5611 // will return -0, so vmin can only be used for unsafe math or if one of
5612 // the operands is known to be nonzero.
5613 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
5615 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5617 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
5626 // If LHS is NaN, an ordered comparison will be false and the result will
5627 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
5628 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5629 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
5630 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5632 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
5633 // will return +0, so vmax can only be used for unsafe math or if one of
5634 // the operands is known to be nonzero.
5635 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
5637 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5639 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
5645 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
5648 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
5649 DAGCombinerInfo &DCI) const {
5650 switch (N->getOpcode()) {
5652 case ISD::ADD: return PerformADDCombine(N, DCI);
5653 case ISD::SUB: return PerformSUBCombine(N, DCI);
5654 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
5655 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
5656 case ISD::AND: return PerformANDCombine(N, DCI);
5657 case ARMISD::BFI: return PerformBFICombine(N, DCI);
5658 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
5659 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
5660 case ISD::STORE: return PerformSTORECombine(N, DCI);
5661 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
5662 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
5663 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
5664 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
5665 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
5668 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
5669 case ISD::SIGN_EXTEND:
5670 case ISD::ZERO_EXTEND:
5671 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
5672 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
5677 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
5678 if (!Subtarget->allowsUnalignedMem())
5681 switch (VT.getSimpleVT().SimpleTy) {
5688 // FIXME: VLD1 etc with standard alignment is legal.
5692 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
5697 switch (VT.getSimpleVT().SimpleTy) {
5698 default: return false;
5713 if ((V & (Scale - 1)) != 0)
5716 return V == (V & ((1LL << 5) - 1));
5719 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
5720 const ARMSubtarget *Subtarget) {
5727 switch (VT.getSimpleVT().SimpleTy) {
5728 default: return false;
5733 // + imm12 or - imm8
5735 return V == (V & ((1LL << 8) - 1));
5736 return V == (V & ((1LL << 12) - 1));
5739 // Same as ARM mode. FIXME: NEON?
5740 if (!Subtarget->hasVFP2())
5745 return V == (V & ((1LL << 8) - 1));
5749 /// isLegalAddressImmediate - Return true if the integer value can be used
5750 /// as the offset of the target addressing mode for load / store of the
5752 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5753 const ARMSubtarget *Subtarget) {
5760 if (Subtarget->isThumb1Only())
5761 return isLegalT1AddressImmediate(V, VT);
5762 else if (Subtarget->isThumb2())
5763 return isLegalT2AddressImmediate(V, VT, Subtarget);
5768 switch (VT.getSimpleVT().SimpleTy) {
5769 default: return false;
5774 return V == (V & ((1LL << 12) - 1));
5777 return V == (V & ((1LL << 8) - 1));
5780 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5785 return V == (V & ((1LL << 8) - 1));
5789 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5791 int Scale = AM.Scale;
5795 switch (VT.getSimpleVT().SimpleTy) {
5796 default: return false;
5805 return Scale == 2 || Scale == 4 || Scale == 8;
5808 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5812 // Note, we allow "void" uses (basically, uses that aren't loads or
5813 // stores), because arm allows folding a scale into many arithmetic
5814 // operations. This should be made more precise and revisited later.
5816 // Allow r << imm, but the imm has to be a multiple of two.
5817 if (Scale & 1) return false;
5818 return isPowerOf2_32(Scale);
5822 /// isLegalAddressingMode - Return true if the addressing mode represented
5823 /// by AM is legal for this target, for a load/store of the specified type.
5824 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5825 const Type *Ty) const {
5826 EVT VT = getValueType(Ty, true);
5827 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5830 // Can never fold addr of global into load/store.
5835 case 0: // no scale reg, must be "r+i" or "r", or "i".
5838 if (Subtarget->isThumb1Only())
5842 // ARM doesn't support any R+R*scale+imm addr modes.
5849 if (Subtarget->isThumb2())
5850 return isLegalT2ScaledAddressingMode(AM, VT);
5852 int Scale = AM.Scale;
5853 switch (VT.getSimpleVT().SimpleTy) {
5854 default: return false;
5858 if (Scale < 0) Scale = -Scale;
5862 return isPowerOf2_32(Scale & ~1);
5866 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5871 // Note, we allow "void" uses (basically, uses that aren't loads or
5872 // stores), because arm allows folding a scale into many arithmetic
5873 // operations. This should be made more precise and revisited later.
5875 // Allow r << imm, but the imm has to be a multiple of two.
5876 if (Scale & 1) return false;
5877 return isPowerOf2_32(Scale);
5884 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5885 /// icmp immediate, that is the target has icmp instructions which can compare
5886 /// a register against the immediate without having to materialize the
5887 /// immediate into a register.
5888 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5889 if (!Subtarget->isThumb())
5890 return ARM_AM::getSOImmVal(Imm) != -1;
5891 if (Subtarget->isThumb2())
5892 return ARM_AM::getT2SOImmVal(Imm) != -1;
5893 return Imm >= 0 && Imm <= 255;
5896 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5897 bool isSEXTLoad, SDValue &Base,
5898 SDValue &Offset, bool &isInc,
5899 SelectionDAG &DAG) {
5900 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5903 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5905 Base = Ptr->getOperand(0);
5906 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5907 int RHSC = (int)RHS->getZExtValue();
5908 if (RHSC < 0 && RHSC > -256) {
5909 assert(Ptr->getOpcode() == ISD::ADD);
5911 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5915 isInc = (Ptr->getOpcode() == ISD::ADD);
5916 Offset = Ptr->getOperand(1);
5918 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5920 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5921 int RHSC = (int)RHS->getZExtValue();
5922 if (RHSC < 0 && RHSC > -0x1000) {
5923 assert(Ptr->getOpcode() == ISD::ADD);
5925 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5926 Base = Ptr->getOperand(0);
5931 if (Ptr->getOpcode() == ISD::ADD) {
5933 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5934 if (ShOpcVal != ARM_AM::no_shift) {
5935 Base = Ptr->getOperand(1);
5936 Offset = Ptr->getOperand(0);
5938 Base = Ptr->getOperand(0);
5939 Offset = Ptr->getOperand(1);
5944 isInc = (Ptr->getOpcode() == ISD::ADD);
5945 Base = Ptr->getOperand(0);
5946 Offset = Ptr->getOperand(1);
5950 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5954 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5955 bool isSEXTLoad, SDValue &Base,
5956 SDValue &Offset, bool &isInc,
5957 SelectionDAG &DAG) {
5958 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5961 Base = Ptr->getOperand(0);
5962 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5963 int RHSC = (int)RHS->getZExtValue();
5964 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5965 assert(Ptr->getOpcode() == ISD::ADD);
5967 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5969 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5970 isInc = Ptr->getOpcode() == ISD::ADD;
5971 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5979 /// getPreIndexedAddressParts - returns true by value, base pointer and
5980 /// offset pointer and addressing mode by reference if the node's address
5981 /// can be legally represented as pre-indexed load / store address.
5983 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5985 ISD::MemIndexedMode &AM,
5986 SelectionDAG &DAG) const {
5987 if (Subtarget->isThumb1Only())
5992 bool isSEXTLoad = false;
5993 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5994 Ptr = LD->getBasePtr();
5995 VT = LD->getMemoryVT();
5996 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5997 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5998 Ptr = ST->getBasePtr();
5999 VT = ST->getMemoryVT();
6004 bool isLegal = false;
6005 if (Subtarget->isThumb2())
6006 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6007 Offset, isInc, DAG);
6009 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
6010 Offset, isInc, DAG);
6014 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
6018 /// getPostIndexedAddressParts - returns true by value, base pointer and
6019 /// offset pointer and addressing mode by reference if this node can be
6020 /// combined with a load / store to form a post-indexed load / store.
6021 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
6024 ISD::MemIndexedMode &AM,
6025 SelectionDAG &DAG) const {
6026 if (Subtarget->isThumb1Only())
6031 bool isSEXTLoad = false;
6032 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6033 VT = LD->getMemoryVT();
6034 Ptr = LD->getBasePtr();
6035 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
6036 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6037 VT = ST->getMemoryVT();
6038 Ptr = ST->getBasePtr();
6043 bool isLegal = false;
6044 if (Subtarget->isThumb2())
6045 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6048 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
6054 // Swap base ptr and offset to catch more post-index load / store when
6055 // it's legal. In Thumb2 mode, offset must be an immediate.
6056 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
6057 !Subtarget->isThumb2())
6058 std::swap(Base, Offset);
6060 // Post-indexed load / store update the base pointer.
6065 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
6069 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6073 const SelectionDAG &DAG,
6074 unsigned Depth) const {
6075 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
6076 switch (Op.getOpcode()) {
6078 case ARMISD::CMOV: {
6079 // Bits are known zero/one if known on the LHS and RHS.
6080 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
6081 if (KnownZero == 0 && KnownOne == 0) return;
6083 APInt KnownZeroRHS, KnownOneRHS;
6084 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
6085 KnownZeroRHS, KnownOneRHS, Depth+1);
6086 KnownZero &= KnownZeroRHS;
6087 KnownOne &= KnownOneRHS;
6093 //===----------------------------------------------------------------------===//
6094 // ARM Inline Assembly Support
6095 //===----------------------------------------------------------------------===//
6097 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
6098 // Looking for "rev" which is V6+.
6099 if (!Subtarget->hasV6Ops())
6102 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
6103 std::string AsmStr = IA->getAsmString();
6104 SmallVector<StringRef, 4> AsmPieces;
6105 SplitString(AsmStr, AsmPieces, ";\n");
6107 switch (AsmPieces.size()) {
6108 default: return false;
6110 AsmStr = AsmPieces[0];
6112 SplitString(AsmStr, AsmPieces, " \t,");
6115 if (AsmPieces.size() == 3 &&
6116 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
6117 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
6118 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
6119 if (Ty && Ty->getBitWidth() == 32)
6120 return IntrinsicLowering::LowerToByteSwap(CI);
6128 /// getConstraintType - Given a constraint letter, return the type of
6129 /// constraint it is for this target.
6130 ARMTargetLowering::ConstraintType
6131 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
6132 if (Constraint.size() == 1) {
6133 switch (Constraint[0]) {
6135 case 'l': return C_RegisterClass;
6136 case 'w': return C_RegisterClass;
6139 return TargetLowering::getConstraintType(Constraint);
6142 /// Examine constraint type and operand type and determine a weight value.
6143 /// This object must already have been set up with the operand type
6144 /// and the current alternative constraint selected.
6145 TargetLowering::ConstraintWeight
6146 ARMTargetLowering::getSingleConstraintMatchWeight(
6147 AsmOperandInfo &info, const char *constraint) const {
6148 ConstraintWeight weight = CW_Invalid;
6149 Value *CallOperandVal = info.CallOperandVal;
6150 // If we don't have a value, we can't do a match,
6151 // but allow it at the lowest weight.
6152 if (CallOperandVal == NULL)
6154 const Type *type = CallOperandVal->getType();
6155 // Look at the constraint type.
6156 switch (*constraint) {
6158 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6161 if (type->isIntegerTy()) {
6162 if (Subtarget->isThumb())
6163 weight = CW_SpecificReg;
6165 weight = CW_Register;
6169 if (type->isFloatingPointTy())
6170 weight = CW_Register;
6176 std::pair<unsigned, const TargetRegisterClass*>
6177 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6179 if (Constraint.size() == 1) {
6180 // GCC ARM Constraint Letters
6181 switch (Constraint[0]) {
6183 if (Subtarget->isThumb())
6184 return std::make_pair(0U, ARM::tGPRRegisterClass);
6186 return std::make_pair(0U, ARM::GPRRegisterClass);
6188 return std::make_pair(0U, ARM::GPRRegisterClass);
6191 return std::make_pair(0U, ARM::SPRRegisterClass);
6192 if (VT.getSizeInBits() == 64)
6193 return std::make_pair(0U, ARM::DPRRegisterClass);
6194 if (VT.getSizeInBits() == 128)
6195 return std::make_pair(0U, ARM::QPRRegisterClass);
6199 if (StringRef("{cc}").equals_lower(Constraint))
6200 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
6202 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6205 std::vector<unsigned> ARMTargetLowering::
6206 getRegClassForInlineAsmConstraint(const std::string &Constraint,
6208 if (Constraint.size() != 1)
6209 return std::vector<unsigned>();
6211 switch (Constraint[0]) { // GCC ARM Constraint Letters
6214 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6215 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6218 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
6219 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
6220 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
6221 ARM::R12, ARM::LR, 0);
6224 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
6225 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
6226 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
6227 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
6228 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
6229 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
6230 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
6231 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
6232 if (VT.getSizeInBits() == 64)
6233 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
6234 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
6235 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
6236 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
6237 if (VT.getSizeInBits() == 128)
6238 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
6239 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
6243 return std::vector<unsigned>();
6246 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6247 /// vector. If it is invalid, don't add anything to Ops.
6248 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6250 std::vector<SDValue>&Ops,
6251 SelectionDAG &DAG) const {
6252 SDValue Result(0, 0);
6254 switch (Constraint) {
6256 case 'I': case 'J': case 'K': case 'L':
6257 case 'M': case 'N': case 'O':
6258 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
6262 int64_t CVal64 = C->getSExtValue();
6263 int CVal = (int) CVal64;
6264 // None of these constraints allow values larger than 32 bits. Check
6265 // that the value fits in an int.
6269 switch (Constraint) {
6271 if (Subtarget->isThumb1Only()) {
6272 // This must be a constant between 0 and 255, for ADD
6274 if (CVal >= 0 && CVal <= 255)
6276 } else if (Subtarget->isThumb2()) {
6277 // A constant that can be used as an immediate value in a
6278 // data-processing instruction.
6279 if (ARM_AM::getT2SOImmVal(CVal) != -1)
6282 // A constant that can be used as an immediate value in a
6283 // data-processing instruction.
6284 if (ARM_AM::getSOImmVal(CVal) != -1)
6290 if (Subtarget->isThumb()) { // FIXME thumb2
6291 // This must be a constant between -255 and -1, for negated ADD
6292 // immediates. This can be used in GCC with an "n" modifier that
6293 // prints the negated value, for use with SUB instructions. It is
6294 // not useful otherwise but is implemented for compatibility.
6295 if (CVal >= -255 && CVal <= -1)
6298 // This must be a constant between -4095 and 4095. It is not clear
6299 // what this constraint is intended for. Implemented for
6300 // compatibility with GCC.
6301 if (CVal >= -4095 && CVal <= 4095)
6307 if (Subtarget->isThumb1Only()) {
6308 // A 32-bit value where only one byte has a nonzero value. Exclude
6309 // zero to match GCC. This constraint is used by GCC internally for
6310 // constants that can be loaded with a move/shift combination.
6311 // It is not useful otherwise but is implemented for compatibility.
6312 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
6314 } else if (Subtarget->isThumb2()) {
6315 // A constant whose bitwise inverse can be used as an immediate
6316 // value in a data-processing instruction. This can be used in GCC
6317 // with a "B" modifier that prints the inverted value, for use with
6318 // BIC and MVN instructions. It is not useful otherwise but is
6319 // implemented for compatibility.
6320 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
6323 // A constant whose bitwise inverse can be used as an immediate
6324 // value in a data-processing instruction. This can be used in GCC
6325 // with a "B" modifier that prints the inverted value, for use with
6326 // BIC and MVN instructions. It is not useful otherwise but is
6327 // implemented for compatibility.
6328 if (ARM_AM::getSOImmVal(~CVal) != -1)
6334 if (Subtarget->isThumb1Only()) {
6335 // This must be a constant between -7 and 7,
6336 // for 3-operand ADD/SUB immediate instructions.
6337 if (CVal >= -7 && CVal < 7)
6339 } else if (Subtarget->isThumb2()) {
6340 // A constant whose negation can be used as an immediate value in a
6341 // data-processing instruction. This can be used in GCC with an "n"
6342 // modifier that prints the negated value, for use with SUB
6343 // instructions. It is not useful otherwise but is implemented for
6345 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
6348 // A constant whose negation can be used as an immediate value in a
6349 // data-processing instruction. This can be used in GCC with an "n"
6350 // modifier that prints the negated value, for use with SUB
6351 // instructions. It is not useful otherwise but is implemented for
6353 if (ARM_AM::getSOImmVal(-CVal) != -1)
6359 if (Subtarget->isThumb()) { // FIXME thumb2
6360 // This must be a multiple of 4 between 0 and 1020, for
6361 // ADD sp + immediate.
6362 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
6365 // A power of two or a constant between 0 and 32. This is used in
6366 // GCC for the shift amount on shifted register operands, but it is
6367 // useful in general for any shift amounts.
6368 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
6374 if (Subtarget->isThumb()) { // FIXME thumb2
6375 // This must be a constant between 0 and 31, for shift amounts.
6376 if (CVal >= 0 && CVal <= 31)
6382 if (Subtarget->isThumb()) { // FIXME thumb2
6383 // This must be a multiple of 4 between -508 and 508, for
6384 // ADD/SUB sp = sp + immediate.
6385 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
6390 Result = DAG.getTargetConstant(CVal, Op.getValueType());
6394 if (Result.getNode()) {
6395 Ops.push_back(Result);
6398 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6402 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6403 // The ARM target isn't yet aware of offsets.
6407 int ARM::getVFPf32Imm(const APFloat &FPImm) {
6408 APInt Imm = FPImm.bitcastToAPInt();
6409 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
6410 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
6411 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
6413 // We can handle 4 bits of mantissa.
6414 // mantissa = (16+UInt(e:f:g:h))/16.
6415 if (Mantissa & 0x7ffff)
6418 if ((Mantissa & 0xf) != Mantissa)
6421 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6422 if (Exp < -3 || Exp > 4)
6424 Exp = ((Exp+3) & 0x7) ^ 4;
6426 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6429 int ARM::getVFPf64Imm(const APFloat &FPImm) {
6430 APInt Imm = FPImm.bitcastToAPInt();
6431 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
6432 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
6433 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
6435 // We can handle 4 bits of mantissa.
6436 // mantissa = (16+UInt(e:f:g:h))/16.
6437 if (Mantissa & 0xffffffffffffLL)
6440 if ((Mantissa & 0xf) != Mantissa)
6443 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6444 if (Exp < -3 || Exp > 4)
6446 Exp = ((Exp+3) & 0x7) ^ 4;
6448 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6451 bool ARM::isBitFieldInvertedMask(unsigned v) {
6452 if (v == 0xffffffff)
6454 // there can be 1's on either or both "outsides", all the "inside"
6456 unsigned int lsb = 0, msb = 31;
6457 while (v & (1 << msb)) --msb;
6458 while (v & (1 << lsb)) ++lsb;
6459 for (unsigned int i = lsb; i <= msb; ++i) {
6466 /// isFPImmLegal - Returns true if the target can instruction select the
6467 /// specified FP immediate natively. If false, the legalizer will
6468 /// materialize the FP immediate as a load from a constant pool.
6469 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
6470 if (!Subtarget->hasVFP3())
6473 return ARM::getVFPf32Imm(Imm) != -1;
6475 return ARM::getVFPf64Imm(Imm) != -1;
6479 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6480 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6481 /// specified in the intrinsic calls.
6482 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6484 unsigned Intrinsic) const {
6485 switch (Intrinsic) {
6486 case Intrinsic::arm_neon_vld1:
6487 case Intrinsic::arm_neon_vld2:
6488 case Intrinsic::arm_neon_vld3:
6489 case Intrinsic::arm_neon_vld4:
6490 case Intrinsic::arm_neon_vld2lane:
6491 case Intrinsic::arm_neon_vld3lane:
6492 case Intrinsic::arm_neon_vld4lane: {
6493 Info.opc = ISD::INTRINSIC_W_CHAIN;
6494 // Conservatively set memVT to the entire set of vectors loaded.
6495 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
6496 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6497 Info.ptrVal = I.getArgOperand(0);
6499 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6500 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6501 Info.vol = false; // volatile loads with NEON intrinsics not supported
6502 Info.readMem = true;
6503 Info.writeMem = false;
6506 case Intrinsic::arm_neon_vst1:
6507 case Intrinsic::arm_neon_vst2:
6508 case Intrinsic::arm_neon_vst3:
6509 case Intrinsic::arm_neon_vst4:
6510 case Intrinsic::arm_neon_vst2lane:
6511 case Intrinsic::arm_neon_vst3lane:
6512 case Intrinsic::arm_neon_vst4lane: {
6513 Info.opc = ISD::INTRINSIC_VOID;
6514 // Conservatively set memVT to the entire set of vectors stored.
6515 unsigned NumElts = 0;
6516 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6517 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
6518 if (!ArgTy->isVectorTy())
6520 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
6522 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6523 Info.ptrVal = I.getArgOperand(0);
6525 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6526 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6527 Info.vol = false; // volatile stores with NEON intrinsics not supported
6528 Info.readMem = false;
6529 Info.writeMem = true;