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/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/PseudoSourceValue.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
54 STATISTIC(NumTailCalls, "Number of tail calls");
56 // This option should go away when tail calls fully work.
58 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
59 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
63 EnableARMLongCalls("arm-long-calls", cl::Hidden,
64 cl::desc("Generate calls via indirect call instructions"),
68 ARMInterworking("arm-interworking", cl::Hidden,
69 cl::desc("Enable / disable ARM interworking (for debugging only)"),
72 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
73 EVT PromotedBitwiseVT) {
74 if (VT != PromotedLdStVT) {
75 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
76 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
77 PromotedLdStVT.getSimpleVT());
79 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
80 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
81 PromotedLdStVT.getSimpleVT());
84 EVT ElemTy = VT.getVectorElementType();
85 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
86 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
87 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
88 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
89 if (ElemTy != MVT::i32) {
90 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
91 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
95 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
97 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
98 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
99 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
100 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
101 if (VT.isInteger()) {
102 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
103 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
104 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
105 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
106 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
108 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
110 // Promote all bit-wise operations.
111 if (VT.isInteger() && VT != PromotedBitwiseVT) {
112 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
113 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
114 PromotedBitwiseVT.getSimpleVT());
115 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
116 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
117 PromotedBitwiseVT.getSimpleVT());
118 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
119 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
120 PromotedBitwiseVT.getSimpleVT());
123 // Neon does not support vector divide/remainder operations.
124 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
125 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
128 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
129 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
132 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
133 addRegisterClass(VT, ARM::DPRRegisterClass);
134 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
137 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
138 addRegisterClass(VT, ARM::QPRRegisterClass);
139 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
142 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
143 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
144 return new TargetLoweringObjectFileMachO();
146 return new ARMElfTargetObjectFile();
149 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
150 : TargetLowering(TM, createTLOF(TM)) {
151 Subtarget = &TM.getSubtarget<ARMSubtarget>();
152 RegInfo = TM.getRegisterInfo();
153 Itins = TM.getInstrItineraryData();
155 if (Subtarget->isTargetDarwin()) {
156 // Uses VFP for Thumb libfuncs if available.
157 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
158 // Single-precision floating-point arithmetic.
159 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
160 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
161 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
162 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
164 // Double-precision floating-point arithmetic.
165 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
166 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
167 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
168 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
170 // Single-precision comparisons.
171 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
172 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
173 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
174 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
175 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
176 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
177 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
178 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
180 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
181 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
182 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
183 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
189 // Double-precision comparisons.
190 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
191 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
192 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
193 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
194 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
195 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
196 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
197 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
199 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
208 // Floating-point to integer conversions.
209 // i64 conversions are done via library routines even when generating VFP
210 // instructions, so use the same ones.
211 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
212 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
213 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
214 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
216 // Conversions between floating types.
217 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
218 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
220 // Integer to floating-point conversions.
221 // i64 conversions are done via library routines even when generating VFP
222 // instructions, so use the same ones.
223 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
224 // e.g., __floatunsidf vs. __floatunssidfvfp.
225 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
226 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
227 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
228 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
232 // These libcalls are not available in 32-bit.
233 setLibcallName(RTLIB::SHL_I128, 0);
234 setLibcallName(RTLIB::SRL_I128, 0);
235 setLibcallName(RTLIB::SRA_I128, 0);
237 if (Subtarget->isAAPCS_ABI()) {
238 // Double-precision floating-point arithmetic helper functions
239 // RTABI chapter 4.1.2, Table 2
240 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
241 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
242 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
243 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
244 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
245 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
246 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
247 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
249 // Double-precision floating-point comparison helper functions
250 // RTABI chapter 4.1.2, Table 3
251 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
252 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
253 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
254 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
255 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
256 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
257 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
258 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
259 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
260 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
261 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
262 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
263 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
264 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
265 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
266 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
267 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
270 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
276 // Single-precision floating-point arithmetic helper functions
277 // RTABI chapter 4.1.2, Table 4
278 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
279 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
280 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
281 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
282 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
283 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
284 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
285 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
287 // Single-precision floating-point comparison helper functions
288 // RTABI chapter 4.1.2, Table 5
289 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
290 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
291 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
292 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
293 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
294 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
295 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
296 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
297 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
298 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
299 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
300 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
301 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
302 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
303 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
304 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
305 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
308 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
314 // Floating-point to integer conversions.
315 // RTABI chapter 4.1.2, Table 6
316 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
317 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
318 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
319 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
320 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
321 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
322 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
323 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
324 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
325 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
326 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
327 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
333 // Conversions between floating types.
334 // RTABI chapter 4.1.2, Table 7
335 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
336 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
337 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
340 // Integer to floating-point conversions.
341 // RTABI chapter 4.1.2, Table 8
342 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
343 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
344 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
345 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
346 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
347 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
348 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
349 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
350 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
359 // Long long helper functions
360 // RTABI chapter 4.2, Table 9
361 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
362 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
363 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
364 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
365 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
366 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
367 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
368 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
369 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
370 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
371 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
374 // Integer division functions
375 // RTABI chapter 4.3.1
376 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
377 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
378 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
379 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
380 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
381 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
382 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
384 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
385 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
386 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
390 if (Subtarget->isThumb1Only())
391 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
393 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
394 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
395 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
396 if (!Subtarget->isFPOnlySP())
397 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
399 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
402 if (Subtarget->hasNEON()) {
403 addDRTypeForNEON(MVT::v2f32);
404 addDRTypeForNEON(MVT::v8i8);
405 addDRTypeForNEON(MVT::v4i16);
406 addDRTypeForNEON(MVT::v2i32);
407 addDRTypeForNEON(MVT::v1i64);
409 addQRTypeForNEON(MVT::v4f32);
410 addQRTypeForNEON(MVT::v2f64);
411 addQRTypeForNEON(MVT::v16i8);
412 addQRTypeForNEON(MVT::v8i16);
413 addQRTypeForNEON(MVT::v4i32);
414 addQRTypeForNEON(MVT::v2i64);
416 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
417 // neither Neon nor VFP support any arithmetic operations on it.
418 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
419 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
420 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
421 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
422 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
423 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
424 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
425 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
426 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
427 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
428 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
429 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
430 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
431 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
432 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
433 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
434 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
435 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
436 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
437 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
438 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
439 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
440 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
441 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
443 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
445 // Neon does not support some operations on v1i64 and v2i64 types.
446 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
447 // Custom handling for some quad-vector types to detect VMULL.
448 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
449 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
450 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
451 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
452 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
454 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
455 setTargetDAGCombine(ISD::SHL);
456 setTargetDAGCombine(ISD::SRL);
457 setTargetDAGCombine(ISD::SRA);
458 setTargetDAGCombine(ISD::SIGN_EXTEND);
459 setTargetDAGCombine(ISD::ZERO_EXTEND);
460 setTargetDAGCombine(ISD::ANY_EXTEND);
461 setTargetDAGCombine(ISD::SELECT_CC);
462 setTargetDAGCombine(ISD::BUILD_VECTOR);
463 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
466 computeRegisterProperties();
468 // ARM does not have f32 extending load.
469 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
471 // ARM does not have i1 sign extending load.
472 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
474 // ARM supports all 4 flavors of integer indexed load / store.
475 if (!Subtarget->isThumb1Only()) {
476 for (unsigned im = (unsigned)ISD::PRE_INC;
477 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
478 setIndexedLoadAction(im, MVT::i1, Legal);
479 setIndexedLoadAction(im, MVT::i8, Legal);
480 setIndexedLoadAction(im, MVT::i16, Legal);
481 setIndexedLoadAction(im, MVT::i32, Legal);
482 setIndexedStoreAction(im, MVT::i1, Legal);
483 setIndexedStoreAction(im, MVT::i8, Legal);
484 setIndexedStoreAction(im, MVT::i16, Legal);
485 setIndexedStoreAction(im, MVT::i32, Legal);
489 // i64 operation support.
490 if (Subtarget->isThumb1Only()) {
491 setOperationAction(ISD::MUL, MVT::i64, Expand);
492 setOperationAction(ISD::MULHU, MVT::i32, Expand);
493 setOperationAction(ISD::MULHS, MVT::i32, Expand);
494 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
495 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
497 setOperationAction(ISD::MUL, MVT::i64, Expand);
498 setOperationAction(ISD::MULHU, MVT::i32, Expand);
499 if (!Subtarget->hasV6Ops())
500 setOperationAction(ISD::MULHS, MVT::i32, Expand);
502 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
503 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
504 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
505 setOperationAction(ISD::SRL, MVT::i64, Custom);
506 setOperationAction(ISD::SRA, MVT::i64, Custom);
508 // ARM does not have ROTL.
509 setOperationAction(ISD::ROTL, MVT::i32, Expand);
510 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
511 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
512 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
513 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
515 // Only ARMv6 has BSWAP.
516 if (!Subtarget->hasV6Ops())
517 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
519 // These are expanded into libcalls.
520 if (!Subtarget->hasDivide()) {
521 // v7M has a hardware divider
522 setOperationAction(ISD::SDIV, MVT::i32, Expand);
523 setOperationAction(ISD::UDIV, MVT::i32, Expand);
525 setOperationAction(ISD::SREM, MVT::i32, Expand);
526 setOperationAction(ISD::UREM, MVT::i32, Expand);
527 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
528 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
530 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
531 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
532 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
533 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
534 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
536 setOperationAction(ISD::TRAP, MVT::Other, Legal);
538 // Use the default implementation.
539 setOperationAction(ISD::VASTART, MVT::Other, Custom);
540 setOperationAction(ISD::VAARG, MVT::Other, Expand);
541 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
542 setOperationAction(ISD::VAEND, MVT::Other, Expand);
543 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
544 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
545 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
546 // FIXME: Shouldn't need this, since no register is used, but the legalizer
547 // doesn't yet know how to not do that for SjLj.
548 setExceptionSelectorRegister(ARM::R0);
549 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
550 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
551 // the default expansion.
552 if (Subtarget->hasDataBarrier() ||
553 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
554 // membarrier needs custom lowering; the rest are legal and handled
556 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
558 // Set them all for expansion, which will force libcalls.
559 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
560 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
561 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
562 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
563 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
564 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
565 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
566 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
567 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
568 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
569 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
570 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
571 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
572 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
573 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
584 // Since the libcalls include locking, fold in the fences
585 setShouldFoldAtomicFences(true);
587 // 64-bit versions are always libcalls (for now)
588 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
589 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
590 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
591 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
592 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
593 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
595 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
597 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
598 if (!Subtarget->hasV6Ops()) {
599 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
600 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
602 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
604 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
605 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
606 // iff target supports vfp2.
607 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
608 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
611 // We want to custom lower some of our intrinsics.
612 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
613 if (Subtarget->isTargetDarwin()) {
614 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
615 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
616 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
619 setOperationAction(ISD::SETCC, MVT::i32, Expand);
620 setOperationAction(ISD::SETCC, MVT::f32, Expand);
621 setOperationAction(ISD::SETCC, MVT::f64, Expand);
622 setOperationAction(ISD::SELECT, MVT::i32, Custom);
623 setOperationAction(ISD::SELECT, MVT::f32, Custom);
624 setOperationAction(ISD::SELECT, MVT::f64, Custom);
625 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
626 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
627 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
629 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
630 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
631 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
632 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
633 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
635 // We don't support sin/cos/fmod/copysign/pow
636 setOperationAction(ISD::FSIN, MVT::f64, Expand);
637 setOperationAction(ISD::FSIN, MVT::f32, Expand);
638 setOperationAction(ISD::FCOS, MVT::f32, Expand);
639 setOperationAction(ISD::FCOS, MVT::f64, Expand);
640 setOperationAction(ISD::FREM, MVT::f64, Expand);
641 setOperationAction(ISD::FREM, MVT::f32, Expand);
642 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
643 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
644 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
646 setOperationAction(ISD::FPOW, MVT::f64, Expand);
647 setOperationAction(ISD::FPOW, MVT::f32, Expand);
649 // Various VFP goodness
650 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
651 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
652 if (Subtarget->hasVFP2()) {
653 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
654 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
655 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
656 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
658 // Special handling for half-precision FP.
659 if (!Subtarget->hasFP16()) {
660 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
661 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
665 // We have target-specific dag combine patterns for the following nodes:
666 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
667 setTargetDAGCombine(ISD::ADD);
668 setTargetDAGCombine(ISD::SUB);
669 setTargetDAGCombine(ISD::MUL);
671 if (Subtarget->hasV6T2Ops())
672 setTargetDAGCombine(ISD::OR);
674 setStackPointerRegisterToSaveRestore(ARM::SP);
676 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
677 setSchedulingPreference(Sched::RegPressure);
679 setSchedulingPreference(Sched::Hybrid);
681 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
683 // On ARM arguments smaller than 4 bytes are extended, so all arguments
684 // are at least 4 bytes aligned.
685 setMinStackArgumentAlignment(4);
687 benefitFromCodePlacementOpt = true;
690 std::pair<const TargetRegisterClass*, uint8_t>
691 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
692 const TargetRegisterClass *RRC = 0;
694 switch (VT.getSimpleVT().SimpleTy) {
696 return TargetLowering::findRepresentativeClass(VT);
697 // Use DPR as representative register class for all floating point
698 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
699 // the cost is 1 for both f32 and f64.
700 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
701 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
702 RRC = ARM::DPRRegisterClass;
704 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
705 case MVT::v4f32: case MVT::v2f64:
706 RRC = ARM::DPRRegisterClass;
710 RRC = ARM::DPRRegisterClass;
714 RRC = ARM::DPRRegisterClass;
718 return std::make_pair(RRC, Cost);
721 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
724 case ARMISD::Wrapper: return "ARMISD::Wrapper";
725 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
726 case ARMISD::CALL: return "ARMISD::CALL";
727 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
728 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
729 case ARMISD::tCALL: return "ARMISD::tCALL";
730 case ARMISD::BRCOND: return "ARMISD::BRCOND";
731 case ARMISD::BR_JT: return "ARMISD::BR_JT";
732 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
733 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
734 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
735 case ARMISD::CMP: return "ARMISD::CMP";
736 case ARMISD::CMPZ: return "ARMISD::CMPZ";
737 case ARMISD::CMPFP: return "ARMISD::CMPFP";
738 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
739 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
740 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
741 case ARMISD::CMOV: return "ARMISD::CMOV";
742 case ARMISD::CNEG: return "ARMISD::CNEG";
744 case ARMISD::RBIT: return "ARMISD::RBIT";
746 case ARMISD::FTOSI: return "ARMISD::FTOSI";
747 case ARMISD::FTOUI: return "ARMISD::FTOUI";
748 case ARMISD::SITOF: return "ARMISD::SITOF";
749 case ARMISD::UITOF: return "ARMISD::UITOF";
751 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
752 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
753 case ARMISD::RRX: return "ARMISD::RRX";
755 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
756 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
758 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
759 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
760 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
762 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
764 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
766 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
768 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
769 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
771 case ARMISD::VCEQ: return "ARMISD::VCEQ";
772 case ARMISD::VCGE: return "ARMISD::VCGE";
773 case ARMISD::VCGEU: return "ARMISD::VCGEU";
774 case ARMISD::VCGT: return "ARMISD::VCGT";
775 case ARMISD::VCGTU: return "ARMISD::VCGTU";
776 case ARMISD::VTST: return "ARMISD::VTST";
778 case ARMISD::VSHL: return "ARMISD::VSHL";
779 case ARMISD::VSHRs: return "ARMISD::VSHRs";
780 case ARMISD::VSHRu: return "ARMISD::VSHRu";
781 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
782 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
783 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
784 case ARMISD::VSHRN: return "ARMISD::VSHRN";
785 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
786 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
787 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
788 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
789 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
790 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
791 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
792 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
793 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
794 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
795 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
796 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
797 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
798 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
799 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
800 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
801 case ARMISD::VDUP: return "ARMISD::VDUP";
802 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
803 case ARMISD::VEXT: return "ARMISD::VEXT";
804 case ARMISD::VREV64: return "ARMISD::VREV64";
805 case ARMISD::VREV32: return "ARMISD::VREV32";
806 case ARMISD::VREV16: return "ARMISD::VREV16";
807 case ARMISD::VZIP: return "ARMISD::VZIP";
808 case ARMISD::VUZP: return "ARMISD::VUZP";
809 case ARMISD::VTRN: return "ARMISD::VTRN";
810 case ARMISD::VMULLs: return "ARMISD::VMULLs";
811 case ARMISD::VMULLu: return "ARMISD::VMULLu";
812 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
813 case ARMISD::FMAX: return "ARMISD::FMAX";
814 case ARMISD::FMIN: return "ARMISD::FMIN";
815 case ARMISD::BFI: return "ARMISD::BFI";
819 /// getRegClassFor - Return the register class that should be used for the
820 /// specified value type.
821 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
822 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
823 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
824 // load / store 4 to 8 consecutive D registers.
825 if (Subtarget->hasNEON()) {
826 if (VT == MVT::v4i64)
827 return ARM::QQPRRegisterClass;
828 else if (VT == MVT::v8i64)
829 return ARM::QQQQPRRegisterClass;
831 return TargetLowering::getRegClassFor(VT);
834 // Create a fast isel object.
836 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
837 return ARM::createFastISel(funcInfo);
840 /// getFunctionAlignment - Return the Log2 alignment of this function.
841 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
842 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
845 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
846 /// be used for loads / stores from the global.
847 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
848 return (Subtarget->isThumb1Only() ? 127 : 4095);
851 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
852 unsigned NumVals = N->getNumValues();
854 return Sched::RegPressure;
856 for (unsigned i = 0; i != NumVals; ++i) {
857 EVT VT = N->getValueType(i);
858 if (VT.isFloatingPoint() || VT.isVector())
859 return Sched::Latency;
862 if (!N->isMachineOpcode())
863 return Sched::RegPressure;
865 // Load are scheduled for latency even if there instruction itinerary
867 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
868 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
870 return Sched::Latency;
872 if (!Itins->isEmpty() && Itins->getStageLatency(TID.getSchedClass()) > 2)
873 return Sched::Latency;
874 return Sched::RegPressure;
878 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
879 MachineFunction &MF) const {
880 switch (RC->getID()) {
883 case ARM::tGPRRegClassID:
884 return RegInfo->hasFP(MF) ? 4 : 5;
885 case ARM::GPRRegClassID: {
886 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
887 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
889 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
890 case ARM::DPRRegClassID:
895 //===----------------------------------------------------------------------===//
897 //===----------------------------------------------------------------------===//
899 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
900 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
902 default: llvm_unreachable("Unknown condition code!");
903 case ISD::SETNE: return ARMCC::NE;
904 case ISD::SETEQ: return ARMCC::EQ;
905 case ISD::SETGT: return ARMCC::GT;
906 case ISD::SETGE: return ARMCC::GE;
907 case ISD::SETLT: return ARMCC::LT;
908 case ISD::SETLE: return ARMCC::LE;
909 case ISD::SETUGT: return ARMCC::HI;
910 case ISD::SETUGE: return ARMCC::HS;
911 case ISD::SETULT: return ARMCC::LO;
912 case ISD::SETULE: return ARMCC::LS;
916 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
917 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
918 ARMCC::CondCodes &CondCode2) {
919 CondCode2 = ARMCC::AL;
921 default: llvm_unreachable("Unknown FP condition!");
923 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
925 case ISD::SETOGT: CondCode = ARMCC::GT; break;
927 case ISD::SETOGE: CondCode = ARMCC::GE; break;
928 case ISD::SETOLT: CondCode = ARMCC::MI; break;
929 case ISD::SETOLE: CondCode = ARMCC::LS; break;
930 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
931 case ISD::SETO: CondCode = ARMCC::VC; break;
932 case ISD::SETUO: CondCode = ARMCC::VS; break;
933 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
934 case ISD::SETUGT: CondCode = ARMCC::HI; break;
935 case ISD::SETUGE: CondCode = ARMCC::PL; break;
937 case ISD::SETULT: CondCode = ARMCC::LT; break;
939 case ISD::SETULE: CondCode = ARMCC::LE; break;
941 case ISD::SETUNE: CondCode = ARMCC::NE; break;
945 //===----------------------------------------------------------------------===//
946 // Calling Convention Implementation
947 //===----------------------------------------------------------------------===//
949 #include "ARMGenCallingConv.inc"
951 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
952 /// given CallingConvention value.
953 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
955 bool isVarArg) const {
958 llvm_unreachable("Unsupported calling convention");
959 case CallingConv::Fast:
960 if (Subtarget->hasVFP2() && !isVarArg) {
961 if (!Subtarget->isAAPCS_ABI())
962 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
963 // For AAPCS ABI targets, just use VFP variant of the calling convention.
964 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
967 case CallingConv::C: {
968 // Use target triple & subtarget features to do actual dispatch.
969 if (!Subtarget->isAAPCS_ABI())
970 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
971 else if (Subtarget->hasVFP2() &&
972 FloatABIType == FloatABI::Hard && !isVarArg)
973 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
974 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
976 case CallingConv::ARM_AAPCS_VFP:
977 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
978 case CallingConv::ARM_AAPCS:
979 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
980 case CallingConv::ARM_APCS:
981 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
985 /// LowerCallResult - Lower the result values of a call into the
986 /// appropriate copies out of appropriate physical registers.
988 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
989 CallingConv::ID CallConv, bool isVarArg,
990 const SmallVectorImpl<ISD::InputArg> &Ins,
991 DebugLoc dl, SelectionDAG &DAG,
992 SmallVectorImpl<SDValue> &InVals) const {
994 // Assign locations to each value returned by this call.
995 SmallVector<CCValAssign, 16> RVLocs;
996 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
997 RVLocs, *DAG.getContext());
998 CCInfo.AnalyzeCallResult(Ins,
999 CCAssignFnForNode(CallConv, /* Return*/ true,
1002 // Copy all of the result registers out of their specified physreg.
1003 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1004 CCValAssign VA = RVLocs[i];
1007 if (VA.needsCustom()) {
1008 // Handle f64 or half of a v2f64.
1009 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1011 Chain = Lo.getValue(1);
1012 InFlag = Lo.getValue(2);
1013 VA = RVLocs[++i]; // skip ahead to next loc
1014 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1016 Chain = Hi.getValue(1);
1017 InFlag = Hi.getValue(2);
1018 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1020 if (VA.getLocVT() == MVT::v2f64) {
1021 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1022 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1023 DAG.getConstant(0, MVT::i32));
1025 VA = RVLocs[++i]; // skip ahead to next loc
1026 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1027 Chain = Lo.getValue(1);
1028 InFlag = Lo.getValue(2);
1029 VA = RVLocs[++i]; // skip ahead to next loc
1030 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1031 Chain = Hi.getValue(1);
1032 InFlag = Hi.getValue(2);
1033 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1034 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1035 DAG.getConstant(1, MVT::i32));
1038 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1040 Chain = Val.getValue(1);
1041 InFlag = Val.getValue(2);
1044 switch (VA.getLocInfo()) {
1045 default: llvm_unreachable("Unknown loc info!");
1046 case CCValAssign::Full: break;
1047 case CCValAssign::BCvt:
1048 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1052 InVals.push_back(Val);
1058 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1059 /// by "Src" to address "Dst" of size "Size". Alignment information is
1060 /// specified by the specific parameter attribute. The copy will be passed as
1061 /// a byval function parameter.
1062 /// Sometimes what we are copying is the end of a larger object, the part that
1063 /// does not fit in registers.
1065 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1066 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1068 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1069 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1070 /*isVolatile=*/false, /*AlwaysInline=*/false,
1071 MachinePointerInfo(0), MachinePointerInfo(0));
1074 /// LowerMemOpCallTo - Store the argument to the stack.
1076 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1077 SDValue StackPtr, SDValue Arg,
1078 DebugLoc dl, SelectionDAG &DAG,
1079 const CCValAssign &VA,
1080 ISD::ArgFlagsTy Flags) const {
1081 unsigned LocMemOffset = VA.getLocMemOffset();
1082 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1083 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1084 if (Flags.isByVal())
1085 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1087 return DAG.getStore(Chain, dl, Arg, PtrOff,
1088 MachinePointerInfo::getStack(LocMemOffset),
1092 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1093 SDValue Chain, SDValue &Arg,
1094 RegsToPassVector &RegsToPass,
1095 CCValAssign &VA, CCValAssign &NextVA,
1097 SmallVector<SDValue, 8> &MemOpChains,
1098 ISD::ArgFlagsTy Flags) const {
1100 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1101 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1102 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1104 if (NextVA.isRegLoc())
1105 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1107 assert(NextVA.isMemLoc());
1108 if (StackPtr.getNode() == 0)
1109 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1111 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1117 /// LowerCall - Lowering a call into a callseq_start <-
1118 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1121 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1122 CallingConv::ID CallConv, bool isVarArg,
1124 const SmallVectorImpl<ISD::OutputArg> &Outs,
1125 const SmallVectorImpl<SDValue> &OutVals,
1126 const SmallVectorImpl<ISD::InputArg> &Ins,
1127 DebugLoc dl, SelectionDAG &DAG,
1128 SmallVectorImpl<SDValue> &InVals) const {
1129 MachineFunction &MF = DAG.getMachineFunction();
1130 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1131 bool IsSibCall = false;
1132 // Temporarily disable tail calls so things don't break.
1133 if (!EnableARMTailCalls)
1136 // Check if it's really possible to do a tail call.
1137 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1138 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1139 Outs, OutVals, Ins, DAG);
1140 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1141 // detected sibcalls.
1148 // Analyze operands of the call, assigning locations to each operand.
1149 SmallVector<CCValAssign, 16> ArgLocs;
1150 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1152 CCInfo.AnalyzeCallOperands(Outs,
1153 CCAssignFnForNode(CallConv, /* Return*/ false,
1156 // Get a count of how many bytes are to be pushed on the stack.
1157 unsigned NumBytes = CCInfo.getNextStackOffset();
1159 // For tail calls, memory operands are available in our caller's stack.
1163 // Adjust the stack pointer for the new arguments...
1164 // These operations are automatically eliminated by the prolog/epilog pass
1166 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1168 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1170 RegsToPassVector RegsToPass;
1171 SmallVector<SDValue, 8> MemOpChains;
1173 // Walk the register/memloc assignments, inserting copies/loads. In the case
1174 // of tail call optimization, arguments are handled later.
1175 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1177 ++i, ++realArgIdx) {
1178 CCValAssign &VA = ArgLocs[i];
1179 SDValue Arg = OutVals[realArgIdx];
1180 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1182 // Promote the value if needed.
1183 switch (VA.getLocInfo()) {
1184 default: llvm_unreachable("Unknown loc info!");
1185 case CCValAssign::Full: break;
1186 case CCValAssign::SExt:
1187 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1189 case CCValAssign::ZExt:
1190 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1192 case CCValAssign::AExt:
1193 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1195 case CCValAssign::BCvt:
1196 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1200 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1201 if (VA.needsCustom()) {
1202 if (VA.getLocVT() == MVT::v2f64) {
1203 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1204 DAG.getConstant(0, MVT::i32));
1205 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1206 DAG.getConstant(1, MVT::i32));
1208 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1209 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1211 VA = ArgLocs[++i]; // skip ahead to next loc
1212 if (VA.isRegLoc()) {
1213 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1214 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1216 assert(VA.isMemLoc());
1218 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1219 dl, DAG, VA, Flags));
1222 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1223 StackPtr, MemOpChains, Flags);
1225 } else if (VA.isRegLoc()) {
1226 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1227 } else if (!IsSibCall) {
1228 assert(VA.isMemLoc());
1230 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1231 dl, DAG, VA, Flags));
1235 if (!MemOpChains.empty())
1236 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1237 &MemOpChains[0], MemOpChains.size());
1239 // Build a sequence of copy-to-reg nodes chained together with token chain
1240 // and flag operands which copy the outgoing args into the appropriate regs.
1242 // Tail call byval lowering might overwrite argument registers so in case of
1243 // tail call optimization the copies to registers are lowered later.
1245 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1246 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1247 RegsToPass[i].second, InFlag);
1248 InFlag = Chain.getValue(1);
1251 // For tail calls lower the arguments to the 'real' stack slot.
1253 // Force all the incoming stack arguments to be loaded from the stack
1254 // before any new outgoing arguments are stored to the stack, because the
1255 // outgoing stack slots may alias the incoming argument stack slots, and
1256 // the alias isn't otherwise explicit. This is slightly more conservative
1257 // than necessary, because it means that each store effectively depends
1258 // on every argument instead of just those arguments it would clobber.
1260 // Do not flag preceeding copytoreg stuff together with the following stuff.
1262 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1263 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1264 RegsToPass[i].second, InFlag);
1265 InFlag = Chain.getValue(1);
1270 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1271 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1272 // node so that legalize doesn't hack it.
1273 bool isDirect = false;
1274 bool isARMFunc = false;
1275 bool isLocalARMFunc = false;
1276 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1278 if (EnableARMLongCalls) {
1279 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1280 && "long-calls with non-static relocation model!");
1281 // Handle a global address or an external symbol. If it's not one of
1282 // those, the target's already in a register, so we don't need to do
1284 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1285 const GlobalValue *GV = G->getGlobal();
1286 // Create a constant pool entry for the callee address
1287 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1288 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1291 // Get the address of the callee into a register
1292 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1293 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1294 Callee = DAG.getLoad(getPointerTy(), dl,
1295 DAG.getEntryNode(), CPAddr,
1296 MachinePointerInfo::getConstantPool(),
1298 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1299 const char *Sym = S->getSymbol();
1301 // Create a constant pool entry for the callee address
1302 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1303 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1304 Sym, ARMPCLabelIndex, 0);
1305 // Get the address of the callee into a register
1306 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1307 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1308 Callee = DAG.getLoad(getPointerTy(), dl,
1309 DAG.getEntryNode(), CPAddr,
1310 MachinePointerInfo::getConstantPool(),
1313 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1314 const GlobalValue *GV = G->getGlobal();
1316 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1317 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1318 getTargetMachine().getRelocationModel() != Reloc::Static;
1319 isARMFunc = !Subtarget->isThumb() || isStub;
1320 // ARM call to a local ARM function is predicable.
1321 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1322 // tBX takes a register source operand.
1323 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1324 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1325 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1328 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1329 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1330 Callee = DAG.getLoad(getPointerTy(), dl,
1331 DAG.getEntryNode(), CPAddr,
1332 MachinePointerInfo::getConstantPool(),
1334 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1335 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1336 getPointerTy(), Callee, PICLabel);
1338 // On ELF targets for PIC code, direct calls should go through the PLT
1339 unsigned OpFlags = 0;
1340 if (Subtarget->isTargetELF() &&
1341 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1342 OpFlags = ARMII::MO_PLT;
1343 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1345 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1347 bool isStub = Subtarget->isTargetDarwin() &&
1348 getTargetMachine().getRelocationModel() != Reloc::Static;
1349 isARMFunc = !Subtarget->isThumb() || isStub;
1350 // tBX takes a register source operand.
1351 const char *Sym = S->getSymbol();
1352 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1353 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1354 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1355 Sym, ARMPCLabelIndex, 4);
1356 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1357 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1358 Callee = DAG.getLoad(getPointerTy(), dl,
1359 DAG.getEntryNode(), CPAddr,
1360 MachinePointerInfo::getConstantPool(),
1362 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1363 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1364 getPointerTy(), Callee, PICLabel);
1366 unsigned OpFlags = 0;
1367 // On ELF targets for PIC code, direct calls should go through the PLT
1368 if (Subtarget->isTargetELF() &&
1369 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1370 OpFlags = ARMII::MO_PLT;
1371 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1375 // FIXME: handle tail calls differently.
1377 if (Subtarget->isThumb()) {
1378 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1379 CallOpc = ARMISD::CALL_NOLINK;
1381 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1383 CallOpc = (isDirect || Subtarget->hasV5TOps())
1384 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1385 : ARMISD::CALL_NOLINK;
1388 std::vector<SDValue> Ops;
1389 Ops.push_back(Chain);
1390 Ops.push_back(Callee);
1392 // Add argument registers to the end of the list so that they are known live
1394 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1395 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1396 RegsToPass[i].second.getValueType()));
1398 if (InFlag.getNode())
1399 Ops.push_back(InFlag);
1401 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1403 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1405 // Returns a chain and a flag for retval copy to use.
1406 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1407 InFlag = Chain.getValue(1);
1409 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1410 DAG.getIntPtrConstant(0, true), InFlag);
1412 InFlag = Chain.getValue(1);
1414 // Handle result values, copying them out of physregs into vregs that we
1416 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1420 /// MatchingStackOffset - Return true if the given stack call argument is
1421 /// already available in the same position (relatively) of the caller's
1422 /// incoming argument stack.
1424 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1425 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1426 const ARMInstrInfo *TII) {
1427 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1429 if (Arg.getOpcode() == ISD::CopyFromReg) {
1430 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1431 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1433 MachineInstr *Def = MRI->getVRegDef(VR);
1436 if (!Flags.isByVal()) {
1437 if (!TII->isLoadFromStackSlot(Def, FI))
1442 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1443 if (Flags.isByVal())
1444 // ByVal argument is passed in as a pointer but it's now being
1445 // dereferenced. e.g.
1446 // define @foo(%struct.X* %A) {
1447 // tail call @bar(%struct.X* byval %A)
1450 SDValue Ptr = Ld->getBasePtr();
1451 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1454 FI = FINode->getIndex();
1458 assert(FI != INT_MAX);
1459 if (!MFI->isFixedObjectIndex(FI))
1461 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1464 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1465 /// for tail call optimization. Targets which want to do tail call
1466 /// optimization should implement this function.
1468 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1469 CallingConv::ID CalleeCC,
1471 bool isCalleeStructRet,
1472 bool isCallerStructRet,
1473 const SmallVectorImpl<ISD::OutputArg> &Outs,
1474 const SmallVectorImpl<SDValue> &OutVals,
1475 const SmallVectorImpl<ISD::InputArg> &Ins,
1476 SelectionDAG& DAG) const {
1477 const Function *CallerF = DAG.getMachineFunction().getFunction();
1478 CallingConv::ID CallerCC = CallerF->getCallingConv();
1479 bool CCMatch = CallerCC == CalleeCC;
1481 // Look for obvious safe cases to perform tail call optimization that do not
1482 // require ABI changes. This is what gcc calls sibcall.
1484 // Do not sibcall optimize vararg calls unless the call site is not passing
1486 if (isVarArg && !Outs.empty())
1489 // Also avoid sibcall optimization if either caller or callee uses struct
1490 // return semantics.
1491 if (isCalleeStructRet || isCallerStructRet)
1494 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1495 // emitEpilogue is not ready for them.
1496 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1497 // LR. This means if we need to reload LR, it takes an extra instructions,
1498 // which outweighs the value of the tail call; but here we don't know yet
1499 // whether LR is going to be used. Probably the right approach is to
1500 // generate the tail call here and turn it back into CALL/RET in
1501 // emitEpilogue if LR is used.
1502 if (Subtarget->isThumb1Only())
1505 // For the moment, we can only do this to functions defined in this
1506 // compilation, or to indirect calls. A Thumb B to an ARM function,
1507 // or vice versa, is not easily fixed up in the linker unlike BL.
1508 // (We could do this by loading the address of the callee into a register;
1509 // that is an extra instruction over the direct call and burns a register
1510 // as well, so is not likely to be a win.)
1512 // It might be safe to remove this restriction on non-Darwin.
1514 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1515 // but we need to make sure there are enough registers; the only valid
1516 // registers are the 4 used for parameters. We don't currently do this
1518 if (isa<ExternalSymbolSDNode>(Callee))
1521 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1522 const GlobalValue *GV = G->getGlobal();
1523 if (GV->isDeclaration() || GV->isWeakForLinker())
1527 // If the calling conventions do not match, then we'd better make sure the
1528 // results are returned in the same way as what the caller expects.
1530 SmallVector<CCValAssign, 16> RVLocs1;
1531 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1532 RVLocs1, *DAG.getContext());
1533 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1535 SmallVector<CCValAssign, 16> RVLocs2;
1536 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1537 RVLocs2, *DAG.getContext());
1538 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1540 if (RVLocs1.size() != RVLocs2.size())
1542 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1543 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1545 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1547 if (RVLocs1[i].isRegLoc()) {
1548 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1551 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1557 // If the callee takes no arguments then go on to check the results of the
1559 if (!Outs.empty()) {
1560 // Check if stack adjustment is needed. For now, do not do this if any
1561 // argument is passed on the stack.
1562 SmallVector<CCValAssign, 16> ArgLocs;
1563 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1564 ArgLocs, *DAG.getContext());
1565 CCInfo.AnalyzeCallOperands(Outs,
1566 CCAssignFnForNode(CalleeCC, false, isVarArg));
1567 if (CCInfo.getNextStackOffset()) {
1568 MachineFunction &MF = DAG.getMachineFunction();
1570 // Check if the arguments are already laid out in the right way as
1571 // the caller's fixed stack objects.
1572 MachineFrameInfo *MFI = MF.getFrameInfo();
1573 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1574 const ARMInstrInfo *TII =
1575 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1576 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1578 ++i, ++realArgIdx) {
1579 CCValAssign &VA = ArgLocs[i];
1580 EVT RegVT = VA.getLocVT();
1581 SDValue Arg = OutVals[realArgIdx];
1582 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1583 if (VA.getLocInfo() == CCValAssign::Indirect)
1585 if (VA.needsCustom()) {
1586 // f64 and vector types are split into multiple registers or
1587 // register/stack-slot combinations. The types will not match
1588 // the registers; give up on memory f64 refs until we figure
1589 // out what to do about this.
1592 if (!ArgLocs[++i].isRegLoc())
1594 if (RegVT == MVT::v2f64) {
1595 if (!ArgLocs[++i].isRegLoc())
1597 if (!ArgLocs[++i].isRegLoc())
1600 } else if (!VA.isRegLoc()) {
1601 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1613 ARMTargetLowering::LowerReturn(SDValue Chain,
1614 CallingConv::ID CallConv, bool isVarArg,
1615 const SmallVectorImpl<ISD::OutputArg> &Outs,
1616 const SmallVectorImpl<SDValue> &OutVals,
1617 DebugLoc dl, SelectionDAG &DAG) const {
1619 // CCValAssign - represent the assignment of the return value to a location.
1620 SmallVector<CCValAssign, 16> RVLocs;
1622 // CCState - Info about the registers and stack slots.
1623 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1626 // Analyze outgoing return values.
1627 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1630 // If this is the first return lowered for this function, add
1631 // the regs to the liveout set for the function.
1632 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1633 for (unsigned i = 0; i != RVLocs.size(); ++i)
1634 if (RVLocs[i].isRegLoc())
1635 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1640 // Copy the result values into the output registers.
1641 for (unsigned i = 0, realRVLocIdx = 0;
1643 ++i, ++realRVLocIdx) {
1644 CCValAssign &VA = RVLocs[i];
1645 assert(VA.isRegLoc() && "Can only return in registers!");
1647 SDValue Arg = OutVals[realRVLocIdx];
1649 switch (VA.getLocInfo()) {
1650 default: llvm_unreachable("Unknown loc info!");
1651 case CCValAssign::Full: break;
1652 case CCValAssign::BCvt:
1653 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1657 if (VA.needsCustom()) {
1658 if (VA.getLocVT() == MVT::v2f64) {
1659 // Extract the first half and return it in two registers.
1660 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1661 DAG.getConstant(0, MVT::i32));
1662 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1663 DAG.getVTList(MVT::i32, MVT::i32), Half);
1665 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1666 Flag = Chain.getValue(1);
1667 VA = RVLocs[++i]; // skip ahead to next loc
1668 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1669 HalfGPRs.getValue(1), Flag);
1670 Flag = Chain.getValue(1);
1671 VA = RVLocs[++i]; // skip ahead to next loc
1673 // Extract the 2nd half and fall through to handle it as an f64 value.
1674 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1675 DAG.getConstant(1, MVT::i32));
1677 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1679 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1680 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1681 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1682 Flag = Chain.getValue(1);
1683 VA = RVLocs[++i]; // skip ahead to next loc
1684 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1687 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1689 // Guarantee that all emitted copies are
1690 // stuck together, avoiding something bad.
1691 Flag = Chain.getValue(1);
1696 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1698 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1703 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1704 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1705 // one of the above mentioned nodes. It has to be wrapped because otherwise
1706 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1707 // be used to form addressing mode. These wrapped nodes will be selected
1709 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1710 EVT PtrVT = Op.getValueType();
1711 // FIXME there is no actual debug info here
1712 DebugLoc dl = Op.getDebugLoc();
1713 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1715 if (CP->isMachineConstantPoolEntry())
1716 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1717 CP->getAlignment());
1719 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1720 CP->getAlignment());
1721 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1724 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1725 return MachineJumpTableInfo::EK_Inline;
1728 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1729 SelectionDAG &DAG) const {
1730 MachineFunction &MF = DAG.getMachineFunction();
1731 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1732 unsigned ARMPCLabelIndex = 0;
1733 DebugLoc DL = Op.getDebugLoc();
1734 EVT PtrVT = getPointerTy();
1735 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1736 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1738 if (RelocM == Reloc::Static) {
1739 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1741 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1742 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1743 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1744 ARMCP::CPBlockAddress,
1746 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1748 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1749 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1750 MachinePointerInfo::getConstantPool(),
1752 if (RelocM == Reloc::Static)
1754 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1755 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1758 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1760 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1761 SelectionDAG &DAG) const {
1762 DebugLoc dl = GA->getDebugLoc();
1763 EVT PtrVT = getPointerTy();
1764 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1765 MachineFunction &MF = DAG.getMachineFunction();
1766 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1767 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1768 ARMConstantPoolValue *CPV =
1769 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1770 ARMCP::CPValue, PCAdj, "tlsgd", true);
1771 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1772 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1773 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1774 MachinePointerInfo::getConstantPool(),
1776 SDValue Chain = Argument.getValue(1);
1778 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1779 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1781 // call __tls_get_addr.
1784 Entry.Node = Argument;
1785 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1786 Args.push_back(Entry);
1787 // FIXME: is there useful debug info available here?
1788 std::pair<SDValue, SDValue> CallResult =
1789 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1790 false, false, false, false,
1791 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1792 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1793 return CallResult.first;
1796 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1797 // "local exec" model.
1799 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1800 SelectionDAG &DAG) const {
1801 const GlobalValue *GV = GA->getGlobal();
1802 DebugLoc dl = GA->getDebugLoc();
1804 SDValue Chain = DAG.getEntryNode();
1805 EVT PtrVT = getPointerTy();
1806 // Get the Thread Pointer
1807 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1809 if (GV->isDeclaration()) {
1810 MachineFunction &MF = DAG.getMachineFunction();
1811 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1812 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1813 // Initial exec model.
1814 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1815 ARMConstantPoolValue *CPV =
1816 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1817 ARMCP::CPValue, PCAdj, "gottpoff", true);
1818 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1819 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1820 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1821 MachinePointerInfo::getConstantPool(),
1823 Chain = Offset.getValue(1);
1825 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1826 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1828 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1829 MachinePointerInfo::getConstantPool(),
1833 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1834 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1835 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1836 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1837 MachinePointerInfo::getConstantPool(),
1841 // The address of the thread local variable is the add of the thread
1842 // pointer with the offset of the variable.
1843 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1847 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1848 // TODO: implement the "local dynamic" model
1849 assert(Subtarget->isTargetELF() &&
1850 "TLS not implemented for non-ELF targets");
1851 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1852 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1853 // otherwise use the "Local Exec" TLS Model
1854 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1855 return LowerToTLSGeneralDynamicModel(GA, DAG);
1857 return LowerToTLSExecModels(GA, DAG);
1860 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1861 SelectionDAG &DAG) const {
1862 EVT PtrVT = getPointerTy();
1863 DebugLoc dl = Op.getDebugLoc();
1864 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1865 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1866 if (RelocM == Reloc::PIC_) {
1867 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1868 ARMConstantPoolValue *CPV =
1869 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1870 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1871 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1872 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1874 MachinePointerInfo::getConstantPool(),
1876 SDValue Chain = Result.getValue(1);
1877 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1878 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1880 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1881 MachinePointerInfo::getGOT(), false, false, 0);
1884 // If we have T2 ops, we can materialize the address directly via movt/movw
1885 // pair. This is always cheaper.
1886 if (Subtarget->useMovt()) {
1887 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1888 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1890 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1891 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1892 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1893 MachinePointerInfo::getConstantPool(),
1899 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1900 SelectionDAG &DAG) const {
1901 MachineFunction &MF = DAG.getMachineFunction();
1902 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1903 unsigned ARMPCLabelIndex = 0;
1904 EVT PtrVT = getPointerTy();
1905 DebugLoc dl = Op.getDebugLoc();
1906 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1907 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1909 if (RelocM == Reloc::Static)
1910 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1912 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1913 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1914 ARMConstantPoolValue *CPV =
1915 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1916 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1918 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1920 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1921 MachinePointerInfo::getConstantPool(),
1923 SDValue Chain = Result.getValue(1);
1925 if (RelocM == Reloc::PIC_) {
1926 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1927 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1930 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1931 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1937 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1938 SelectionDAG &DAG) const {
1939 assert(Subtarget->isTargetELF() &&
1940 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1941 MachineFunction &MF = DAG.getMachineFunction();
1942 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1943 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1944 EVT PtrVT = getPointerTy();
1945 DebugLoc dl = Op.getDebugLoc();
1946 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1947 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1948 "_GLOBAL_OFFSET_TABLE_",
1949 ARMPCLabelIndex, PCAdj);
1950 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1951 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1952 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1953 MachinePointerInfo::getConstantPool(),
1955 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1956 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1960 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
1962 DebugLoc dl = Op.getDebugLoc();
1963 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
1964 Op.getOperand(0), Op.getOperand(1));
1968 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1969 DebugLoc dl = Op.getDebugLoc();
1970 SDValue Val = DAG.getConstant(0, MVT::i32);
1971 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1972 Op.getOperand(1), Val);
1976 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1977 DebugLoc dl = Op.getDebugLoc();
1978 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1979 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1983 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1984 const ARMSubtarget *Subtarget) const {
1985 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1986 DebugLoc dl = Op.getDebugLoc();
1988 default: return SDValue(); // Don't custom lower most intrinsics.
1989 case Intrinsic::arm_thread_pointer: {
1990 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1991 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1993 case Intrinsic::eh_sjlj_lsda: {
1994 MachineFunction &MF = DAG.getMachineFunction();
1995 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1996 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1997 EVT PtrVT = getPointerTy();
1998 DebugLoc dl = Op.getDebugLoc();
1999 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2001 unsigned PCAdj = (RelocM != Reloc::PIC_)
2002 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2003 ARMConstantPoolValue *CPV =
2004 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2005 ARMCP::CPLSDA, PCAdj);
2006 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2007 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2009 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2010 MachinePointerInfo::getConstantPool(),
2013 if (RelocM == Reloc::PIC_) {
2014 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2015 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2022 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2023 const ARMSubtarget *Subtarget) {
2024 DebugLoc dl = Op.getDebugLoc();
2025 SDValue Op5 = Op.getOperand(5);
2026 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
2027 // Some subtargets which have dmb and dsb instructions can handle barriers
2028 // directly. Some ARMv6 cpus can support them with the help of mcr
2029 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
2031 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
2032 if (Subtarget->hasDataBarrier())
2033 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2035 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2036 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2037 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2038 DAG.getConstant(0, MVT::i32));
2042 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2043 MachineFunction &MF = DAG.getMachineFunction();
2044 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2046 // vastart just stores the address of the VarArgsFrameIndex slot into the
2047 // memory location argument.
2048 DebugLoc dl = Op.getDebugLoc();
2049 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2050 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2051 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2052 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2053 MachinePointerInfo(SV), false, false, 0);
2057 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2058 SDValue &Root, SelectionDAG &DAG,
2059 DebugLoc dl) const {
2060 MachineFunction &MF = DAG.getMachineFunction();
2061 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2063 TargetRegisterClass *RC;
2064 if (AFI->isThumb1OnlyFunction())
2065 RC = ARM::tGPRRegisterClass;
2067 RC = ARM::GPRRegisterClass;
2069 // Transform the arguments stored in physical registers into virtual ones.
2070 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2071 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2074 if (NextVA.isMemLoc()) {
2075 MachineFrameInfo *MFI = MF.getFrameInfo();
2076 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2078 // Create load node to retrieve arguments from the stack.
2079 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2080 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2081 MachinePointerInfo::getFixedStack(FI),
2084 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2085 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2088 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2092 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2093 CallingConv::ID CallConv, bool isVarArg,
2094 const SmallVectorImpl<ISD::InputArg>
2096 DebugLoc dl, SelectionDAG &DAG,
2097 SmallVectorImpl<SDValue> &InVals)
2100 MachineFunction &MF = DAG.getMachineFunction();
2101 MachineFrameInfo *MFI = MF.getFrameInfo();
2103 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2105 // Assign locations to all of the incoming arguments.
2106 SmallVector<CCValAssign, 16> ArgLocs;
2107 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2109 CCInfo.AnalyzeFormalArguments(Ins,
2110 CCAssignFnForNode(CallConv, /* Return*/ false,
2113 SmallVector<SDValue, 16> ArgValues;
2115 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2116 CCValAssign &VA = ArgLocs[i];
2118 // Arguments stored in registers.
2119 if (VA.isRegLoc()) {
2120 EVT RegVT = VA.getLocVT();
2123 if (VA.needsCustom()) {
2124 // f64 and vector types are split up into multiple registers or
2125 // combinations of registers and stack slots.
2126 if (VA.getLocVT() == MVT::v2f64) {
2127 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2129 VA = ArgLocs[++i]; // skip ahead to next loc
2131 if (VA.isMemLoc()) {
2132 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2133 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2134 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2135 MachinePointerInfo::getFixedStack(FI),
2138 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2141 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2142 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2143 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2144 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2145 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2147 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2150 TargetRegisterClass *RC;
2152 if (RegVT == MVT::f32)
2153 RC = ARM::SPRRegisterClass;
2154 else if (RegVT == MVT::f64)
2155 RC = ARM::DPRRegisterClass;
2156 else if (RegVT == MVT::v2f64)
2157 RC = ARM::QPRRegisterClass;
2158 else if (RegVT == MVT::i32)
2159 RC = (AFI->isThumb1OnlyFunction() ?
2160 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2162 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2164 // Transform the arguments in physical registers into virtual ones.
2165 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2166 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2169 // If this is an 8 or 16-bit value, it is really passed promoted
2170 // to 32 bits. Insert an assert[sz]ext to capture this, then
2171 // truncate to the right size.
2172 switch (VA.getLocInfo()) {
2173 default: llvm_unreachable("Unknown loc info!");
2174 case CCValAssign::Full: break;
2175 case CCValAssign::BCvt:
2176 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2178 case CCValAssign::SExt:
2179 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2180 DAG.getValueType(VA.getValVT()));
2181 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2183 case CCValAssign::ZExt:
2184 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2185 DAG.getValueType(VA.getValVT()));
2186 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2190 InVals.push_back(ArgValue);
2192 } else { // VA.isRegLoc()
2195 assert(VA.isMemLoc());
2196 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2198 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2199 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2201 // Create load nodes to retrieve arguments from the stack.
2202 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2203 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2204 MachinePointerInfo::getFixedStack(FI),
2211 static const unsigned GPRArgRegs[] = {
2212 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2215 unsigned NumGPRs = CCInfo.getFirstUnallocated
2216 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2218 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2219 unsigned VARegSize = (4 - NumGPRs) * 4;
2220 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2221 unsigned ArgOffset = CCInfo.getNextStackOffset();
2222 if (VARegSaveSize) {
2223 // If this function is vararg, store any remaining integer argument regs
2224 // to their spots on the stack so that they may be loaded by deferencing
2225 // the result of va_next.
2226 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2227 AFI->setVarArgsFrameIndex(
2228 MFI->CreateFixedObject(VARegSaveSize,
2229 ArgOffset + VARegSaveSize - VARegSize,
2231 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2234 SmallVector<SDValue, 4> MemOps;
2235 for (; NumGPRs < 4; ++NumGPRs) {
2236 TargetRegisterClass *RC;
2237 if (AFI->isThumb1OnlyFunction())
2238 RC = ARM::tGPRRegisterClass;
2240 RC = ARM::GPRRegisterClass;
2242 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2243 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2245 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2246 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2248 MemOps.push_back(Store);
2249 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2250 DAG.getConstant(4, getPointerTy()));
2252 if (!MemOps.empty())
2253 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2254 &MemOps[0], MemOps.size());
2256 // This will point to the next argument passed via stack.
2257 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2263 /// isFloatingPointZero - Return true if this is +0.0.
2264 static bool isFloatingPointZero(SDValue Op) {
2265 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2266 return CFP->getValueAPF().isPosZero();
2267 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2268 // Maybe this has already been legalized into the constant pool?
2269 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2270 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2271 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2272 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2273 return CFP->getValueAPF().isPosZero();
2279 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2280 /// the given operands.
2282 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2283 SDValue &ARMcc, SelectionDAG &DAG,
2284 DebugLoc dl) const {
2285 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2286 unsigned C = RHSC->getZExtValue();
2287 if (!isLegalICmpImmediate(C)) {
2288 // Constant does not fit, try adjusting it by one?
2293 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2294 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2295 RHS = DAG.getConstant(C-1, MVT::i32);
2300 if (C != 0 && isLegalICmpImmediate(C-1)) {
2301 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2302 RHS = DAG.getConstant(C-1, MVT::i32);
2307 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2308 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2309 RHS = DAG.getConstant(C+1, MVT::i32);
2314 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2315 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2316 RHS = DAG.getConstant(C+1, MVT::i32);
2323 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2324 ARMISD::NodeType CompareType;
2327 CompareType = ARMISD::CMP;
2332 CompareType = ARMISD::CMPZ;
2335 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2336 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2339 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2341 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2342 DebugLoc dl) const {
2344 if (!isFloatingPointZero(RHS))
2345 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2347 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2348 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2351 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2352 SDValue Cond = Op.getOperand(0);
2353 SDValue SelectTrue = Op.getOperand(1);
2354 SDValue SelectFalse = Op.getOperand(2);
2355 DebugLoc dl = Op.getDebugLoc();
2359 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2360 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2362 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2363 const ConstantSDNode *CMOVTrue =
2364 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2365 const ConstantSDNode *CMOVFalse =
2366 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2368 if (CMOVTrue && CMOVFalse) {
2369 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2370 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2374 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2376 False = SelectFalse;
2377 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2382 if (True.getNode() && False.getNode()) {
2383 EVT VT = Cond.getValueType();
2384 SDValue ARMcc = Cond.getOperand(2);
2385 SDValue CCR = Cond.getOperand(3);
2386 SDValue Cmp = Cond.getOperand(4);
2387 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2392 return DAG.getSelectCC(dl, Cond,
2393 DAG.getConstant(0, Cond.getValueType()),
2394 SelectTrue, SelectFalse, ISD::SETNE);
2397 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2398 EVT VT = Op.getValueType();
2399 SDValue LHS = Op.getOperand(0);
2400 SDValue RHS = Op.getOperand(1);
2401 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2402 SDValue TrueVal = Op.getOperand(2);
2403 SDValue FalseVal = Op.getOperand(3);
2404 DebugLoc dl = Op.getDebugLoc();
2406 if (LHS.getValueType() == MVT::i32) {
2408 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2409 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2410 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2413 ARMCC::CondCodes CondCode, CondCode2;
2414 FPCCToARMCC(CC, CondCode, CondCode2);
2416 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2417 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2418 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2419 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2421 if (CondCode2 != ARMCC::AL) {
2422 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2423 // FIXME: Needs another CMP because flag can have but one use.
2424 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2425 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2426 Result, TrueVal, ARMcc2, CCR, Cmp2);
2431 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2432 /// to morph to an integer compare sequence.
2433 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2434 const ARMSubtarget *Subtarget) {
2435 SDNode *N = Op.getNode();
2436 if (!N->hasOneUse())
2437 // Otherwise it requires moving the value from fp to integer registers.
2439 if (!N->getNumValues())
2441 EVT VT = Op.getValueType();
2442 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2443 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2444 // vmrs are very slow, e.g. cortex-a8.
2447 if (isFloatingPointZero(Op)) {
2451 return ISD::isNormalLoad(N);
2454 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2455 if (isFloatingPointZero(Op))
2456 return DAG.getConstant(0, MVT::i32);
2458 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2459 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2460 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2461 Ld->isVolatile(), Ld->isNonTemporal(),
2462 Ld->getAlignment());
2464 llvm_unreachable("Unknown VFP cmp argument!");
2467 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2468 SDValue &RetVal1, SDValue &RetVal2) {
2469 if (isFloatingPointZero(Op)) {
2470 RetVal1 = DAG.getConstant(0, MVT::i32);
2471 RetVal2 = DAG.getConstant(0, MVT::i32);
2475 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2476 SDValue Ptr = Ld->getBasePtr();
2477 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2478 Ld->getChain(), Ptr,
2479 Ld->getPointerInfo(),
2480 Ld->isVolatile(), Ld->isNonTemporal(),
2481 Ld->getAlignment());
2483 EVT PtrType = Ptr.getValueType();
2484 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2485 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2486 PtrType, Ptr, DAG.getConstant(4, PtrType));
2487 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2488 Ld->getChain(), NewPtr,
2489 Ld->getPointerInfo().getWithOffset(4),
2490 Ld->isVolatile(), Ld->isNonTemporal(),
2495 llvm_unreachable("Unknown VFP cmp argument!");
2498 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2499 /// f32 and even f64 comparisons to integer ones.
2501 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2502 SDValue Chain = Op.getOperand(0);
2503 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2504 SDValue LHS = Op.getOperand(2);
2505 SDValue RHS = Op.getOperand(3);
2506 SDValue Dest = Op.getOperand(4);
2507 DebugLoc dl = Op.getDebugLoc();
2509 bool SeenZero = false;
2510 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2511 canChangeToInt(RHS, SeenZero, Subtarget) &&
2512 // If one of the operand is zero, it's safe to ignore the NaN case since
2513 // we only care about equality comparisons.
2514 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2515 // If unsafe fp math optimization is enabled and there are no othter uses of
2516 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2517 // to an integer comparison.
2518 if (CC == ISD::SETOEQ)
2520 else if (CC == ISD::SETUNE)
2524 if (LHS.getValueType() == MVT::f32) {
2525 LHS = bitcastf32Toi32(LHS, DAG);
2526 RHS = bitcastf32Toi32(RHS, DAG);
2527 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2528 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2529 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2530 Chain, Dest, ARMcc, CCR, Cmp);
2535 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2536 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2537 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2538 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2539 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2540 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2541 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2547 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2548 SDValue Chain = Op.getOperand(0);
2549 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2550 SDValue LHS = Op.getOperand(2);
2551 SDValue RHS = Op.getOperand(3);
2552 SDValue Dest = Op.getOperand(4);
2553 DebugLoc dl = Op.getDebugLoc();
2555 if (LHS.getValueType() == MVT::i32) {
2557 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2558 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2559 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2560 Chain, Dest, ARMcc, CCR, Cmp);
2563 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2566 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2567 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2568 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2569 if (Result.getNode())
2573 ARMCC::CondCodes CondCode, CondCode2;
2574 FPCCToARMCC(CC, CondCode, CondCode2);
2576 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2577 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2578 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2579 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2580 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2581 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2582 if (CondCode2 != ARMCC::AL) {
2583 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2584 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2585 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2590 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2591 SDValue Chain = Op.getOperand(0);
2592 SDValue Table = Op.getOperand(1);
2593 SDValue Index = Op.getOperand(2);
2594 DebugLoc dl = Op.getDebugLoc();
2596 EVT PTy = getPointerTy();
2597 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2598 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2599 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2600 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2601 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2602 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2603 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2604 if (Subtarget->isThumb2()) {
2605 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2606 // which does another jump to the destination. This also makes it easier
2607 // to translate it to TBB / TBH later.
2608 // FIXME: This might not work if the function is extremely large.
2609 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2610 Addr, Op.getOperand(2), JTI, UId);
2612 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2613 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2614 MachinePointerInfo::getJumpTable(),
2616 Chain = Addr.getValue(1);
2617 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2618 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2620 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2621 MachinePointerInfo::getJumpTable(), false, false, 0);
2622 Chain = Addr.getValue(1);
2623 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2627 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2628 DebugLoc dl = Op.getDebugLoc();
2631 switch (Op.getOpcode()) {
2633 assert(0 && "Invalid opcode!");
2634 case ISD::FP_TO_SINT:
2635 Opc = ARMISD::FTOSI;
2637 case ISD::FP_TO_UINT:
2638 Opc = ARMISD::FTOUI;
2641 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2642 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2645 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2646 EVT VT = Op.getValueType();
2647 DebugLoc dl = Op.getDebugLoc();
2650 switch (Op.getOpcode()) {
2652 assert(0 && "Invalid opcode!");
2653 case ISD::SINT_TO_FP:
2654 Opc = ARMISD::SITOF;
2656 case ISD::UINT_TO_FP:
2657 Opc = ARMISD::UITOF;
2661 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2662 return DAG.getNode(Opc, dl, VT, Op);
2665 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2666 // Implement fcopysign with a fabs and a conditional fneg.
2667 SDValue Tmp0 = Op.getOperand(0);
2668 SDValue Tmp1 = Op.getOperand(1);
2669 DebugLoc dl = Op.getDebugLoc();
2670 EVT VT = Op.getValueType();
2671 EVT SrcVT = Tmp1.getValueType();
2672 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2673 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2674 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2675 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2676 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2677 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2680 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2681 MachineFunction &MF = DAG.getMachineFunction();
2682 MachineFrameInfo *MFI = MF.getFrameInfo();
2683 MFI->setReturnAddressIsTaken(true);
2685 EVT VT = Op.getValueType();
2686 DebugLoc dl = Op.getDebugLoc();
2687 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2689 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2690 SDValue Offset = DAG.getConstant(4, MVT::i32);
2691 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2692 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2693 MachinePointerInfo(), false, false, 0);
2696 // Return LR, which contains the return address. Mark it an implicit live-in.
2697 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2698 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2701 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2702 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2703 MFI->setFrameAddressIsTaken(true);
2705 EVT VT = Op.getValueType();
2706 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2707 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2708 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2709 ? ARM::R7 : ARM::R11;
2710 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2712 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2713 MachinePointerInfo(),
2718 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2719 /// expand a bit convert where either the source or destination type is i64 to
2720 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2721 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2722 /// vectors), since the legalizer won't know what to do with that.
2723 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2724 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2725 DebugLoc dl = N->getDebugLoc();
2726 SDValue Op = N->getOperand(0);
2728 // This function is only supposed to be called for i64 types, either as the
2729 // source or destination of the bit convert.
2730 EVT SrcVT = Op.getValueType();
2731 EVT DstVT = N->getValueType(0);
2732 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2733 "ExpandBIT_CONVERT called for non-i64 type");
2735 // Turn i64->f64 into VMOVDRR.
2736 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2737 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2738 DAG.getConstant(0, MVT::i32));
2739 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2740 DAG.getConstant(1, MVT::i32));
2741 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2742 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2745 // Turn f64->i64 into VMOVRRD.
2746 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2747 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2748 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2749 // Merge the pieces into a single i64 value.
2750 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2756 /// getZeroVector - Returns a vector of specified type with all zero elements.
2757 /// Zero vectors are used to represent vector negation and in those cases
2758 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2759 /// not support i64 elements, so sometimes the zero vectors will need to be
2760 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2762 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2763 assert(VT.isVector() && "Expected a vector type");
2764 // The canonical modified immediate encoding of a zero vector is....0!
2765 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2766 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2767 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2768 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2771 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2772 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2773 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2774 SelectionDAG &DAG) const {
2775 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2776 EVT VT = Op.getValueType();
2777 unsigned VTBits = VT.getSizeInBits();
2778 DebugLoc dl = Op.getDebugLoc();
2779 SDValue ShOpLo = Op.getOperand(0);
2780 SDValue ShOpHi = Op.getOperand(1);
2781 SDValue ShAmt = Op.getOperand(2);
2783 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2785 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2787 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2788 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2789 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2790 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2791 DAG.getConstant(VTBits, MVT::i32));
2792 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2793 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2794 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2796 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2797 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2799 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2800 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2803 SDValue Ops[2] = { Lo, Hi };
2804 return DAG.getMergeValues(Ops, 2, dl);
2807 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2808 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2809 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2810 SelectionDAG &DAG) const {
2811 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2812 EVT VT = Op.getValueType();
2813 unsigned VTBits = VT.getSizeInBits();
2814 DebugLoc dl = Op.getDebugLoc();
2815 SDValue ShOpLo = Op.getOperand(0);
2816 SDValue ShOpHi = Op.getOperand(1);
2817 SDValue ShAmt = Op.getOperand(2);
2820 assert(Op.getOpcode() == ISD::SHL_PARTS);
2821 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2822 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2823 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2824 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2825 DAG.getConstant(VTBits, MVT::i32));
2826 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2827 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2829 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2830 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2831 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2833 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2834 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2837 SDValue Ops[2] = { Lo, Hi };
2838 return DAG.getMergeValues(Ops, 2, dl);
2841 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2842 SelectionDAG &DAG) const {
2843 // The rounding mode is in bits 23:22 of the FPSCR.
2844 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2845 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2846 // so that the shift + and get folded into a bitfield extract.
2847 DebugLoc dl = Op.getDebugLoc();
2848 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2849 DAG.getConstant(Intrinsic::arm_get_fpscr,
2851 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2852 DAG.getConstant(1U << 22, MVT::i32));
2853 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2854 DAG.getConstant(22, MVT::i32));
2855 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2856 DAG.getConstant(3, MVT::i32));
2859 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2860 const ARMSubtarget *ST) {
2861 EVT VT = N->getValueType(0);
2862 DebugLoc dl = N->getDebugLoc();
2864 if (!ST->hasV6T2Ops())
2867 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2868 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2871 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2872 const ARMSubtarget *ST) {
2873 EVT VT = N->getValueType(0);
2874 DebugLoc dl = N->getDebugLoc();
2876 // Lower vector shifts on NEON to use VSHL.
2877 if (VT.isVector()) {
2878 assert(ST->hasNEON() && "unexpected vector shift");
2880 // Left shifts translate directly to the vshiftu intrinsic.
2881 if (N->getOpcode() == ISD::SHL)
2882 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2883 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2884 N->getOperand(0), N->getOperand(1));
2886 assert((N->getOpcode() == ISD::SRA ||
2887 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2889 // NEON uses the same intrinsics for both left and right shifts. For
2890 // right shifts, the shift amounts are negative, so negate the vector of
2892 EVT ShiftVT = N->getOperand(1).getValueType();
2893 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2894 getZeroVector(ShiftVT, DAG, dl),
2896 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2897 Intrinsic::arm_neon_vshifts :
2898 Intrinsic::arm_neon_vshiftu);
2899 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2900 DAG.getConstant(vshiftInt, MVT::i32),
2901 N->getOperand(0), NegatedCount);
2904 // We can get here for a node like i32 = ISD::SHL i32, i64
2908 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2909 "Unknown shift to lower!");
2911 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2912 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2913 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2916 // If we are in thumb mode, we don't have RRX.
2917 if (ST->isThumb1Only()) return SDValue();
2919 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2920 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2921 DAG.getConstant(0, MVT::i32));
2922 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2923 DAG.getConstant(1, MVT::i32));
2925 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2926 // captures the result into a carry flag.
2927 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2928 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2930 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2931 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2933 // Merge the pieces into a single i64 value.
2934 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2937 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2938 SDValue TmpOp0, TmpOp1;
2939 bool Invert = false;
2943 SDValue Op0 = Op.getOperand(0);
2944 SDValue Op1 = Op.getOperand(1);
2945 SDValue CC = Op.getOperand(2);
2946 EVT VT = Op.getValueType();
2947 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2948 DebugLoc dl = Op.getDebugLoc();
2950 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2951 switch (SetCCOpcode) {
2952 default: llvm_unreachable("Illegal FP comparison"); break;
2954 case ISD::SETNE: Invert = true; // Fallthrough
2956 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2958 case ISD::SETLT: Swap = true; // Fallthrough
2960 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2962 case ISD::SETLE: Swap = true; // Fallthrough
2964 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2965 case ISD::SETUGE: Swap = true; // Fallthrough
2966 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2967 case ISD::SETUGT: Swap = true; // Fallthrough
2968 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2969 case ISD::SETUEQ: Invert = true; // Fallthrough
2971 // Expand this to (OLT | OGT).
2975 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2976 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2978 case ISD::SETUO: Invert = true; // Fallthrough
2980 // Expand this to (OLT | OGE).
2984 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2985 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2989 // Integer comparisons.
2990 switch (SetCCOpcode) {
2991 default: llvm_unreachable("Illegal integer comparison"); break;
2992 case ISD::SETNE: Invert = true;
2993 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2994 case ISD::SETLT: Swap = true;
2995 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2996 case ISD::SETLE: Swap = true;
2997 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2998 case ISD::SETULT: Swap = true;
2999 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3000 case ISD::SETULE: Swap = true;
3001 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3004 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3005 if (Opc == ARMISD::VCEQ) {
3008 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3010 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3013 // Ignore bitconvert.
3014 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
3015 AndOp = AndOp.getOperand(0);
3017 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3019 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
3020 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
3027 std::swap(Op0, Op1);
3029 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3032 Result = DAG.getNOT(dl, Result, VT);
3037 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3038 /// valid vector constant for a NEON instruction with a "modified immediate"
3039 /// operand (e.g., VMOV). If so, return the encoded value.
3040 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3041 unsigned SplatBitSize, SelectionDAG &DAG,
3042 EVT &VT, bool is128Bits, bool isVMOV) {
3043 unsigned OpCmode, Imm;
3045 // SplatBitSize is set to the smallest size that splats the vector, so a
3046 // zero vector will always have SplatBitSize == 8. However, NEON modified
3047 // immediate instructions others than VMOV do not support the 8-bit encoding
3048 // of a zero vector, and the default encoding of zero is supposed to be the
3053 switch (SplatBitSize) {
3057 // Any 1-byte value is OK. Op=0, Cmode=1110.
3058 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3061 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3065 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3066 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3067 if ((SplatBits & ~0xff) == 0) {
3068 // Value = 0x00nn: Op=x, Cmode=100x.
3073 if ((SplatBits & ~0xff00) == 0) {
3074 // Value = 0xnn00: Op=x, Cmode=101x.
3076 Imm = SplatBits >> 8;
3082 // NEON's 32-bit VMOV supports splat values where:
3083 // * only one byte is nonzero, or
3084 // * the least significant byte is 0xff and the second byte is nonzero, or
3085 // * the least significant 2 bytes are 0xff and the third is nonzero.
3086 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3087 if ((SplatBits & ~0xff) == 0) {
3088 // Value = 0x000000nn: Op=x, Cmode=000x.
3093 if ((SplatBits & ~0xff00) == 0) {
3094 // Value = 0x0000nn00: Op=x, Cmode=001x.
3096 Imm = SplatBits >> 8;
3099 if ((SplatBits & ~0xff0000) == 0) {
3100 // Value = 0x00nn0000: Op=x, Cmode=010x.
3102 Imm = SplatBits >> 16;
3105 if ((SplatBits & ~0xff000000) == 0) {
3106 // Value = 0xnn000000: Op=x, Cmode=011x.
3108 Imm = SplatBits >> 24;
3112 if ((SplatBits & ~0xffff) == 0 &&
3113 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3114 // Value = 0x0000nnff: Op=x, Cmode=1100.
3116 Imm = SplatBits >> 8;
3121 if ((SplatBits & ~0xffffff) == 0 &&
3122 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3123 // Value = 0x00nnffff: Op=x, Cmode=1101.
3125 Imm = SplatBits >> 16;
3126 SplatBits |= 0xffff;
3130 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3131 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3132 // VMOV.I32. A (very) minor optimization would be to replicate the value
3133 // and fall through here to test for a valid 64-bit splat. But, then the
3134 // caller would also need to check and handle the change in size.
3140 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3141 uint64_t BitMask = 0xff;
3143 unsigned ImmMask = 1;
3145 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3146 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3149 } else if ((SplatBits & BitMask) != 0) {
3155 // Op=1, Cmode=1110.
3158 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3163 llvm_unreachable("unexpected size for isNEONModifiedImm");
3167 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3168 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3171 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3172 bool &ReverseVEXT, unsigned &Imm) {
3173 unsigned NumElts = VT.getVectorNumElements();
3174 ReverseVEXT = false;
3176 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3182 // If this is a VEXT shuffle, the immediate value is the index of the first
3183 // element. The other shuffle indices must be the successive elements after
3185 unsigned ExpectedElt = Imm;
3186 for (unsigned i = 1; i < NumElts; ++i) {
3187 // Increment the expected index. If it wraps around, it may still be
3188 // a VEXT but the source vectors must be swapped.
3190 if (ExpectedElt == NumElts * 2) {
3195 if (M[i] < 0) continue; // ignore UNDEF indices
3196 if (ExpectedElt != static_cast<unsigned>(M[i]))
3200 // Adjust the index value if the source operands will be swapped.
3207 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3208 /// instruction with the specified blocksize. (The order of the elements
3209 /// within each block of the vector is reversed.)
3210 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3211 unsigned BlockSize) {
3212 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3213 "Only possible block sizes for VREV are: 16, 32, 64");
3215 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3219 unsigned NumElts = VT.getVectorNumElements();
3220 unsigned BlockElts = M[0] + 1;
3221 // If the first shuffle index is UNDEF, be optimistic.
3223 BlockElts = BlockSize / EltSz;
3225 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3228 for (unsigned i = 0; i < NumElts; ++i) {
3229 if (M[i] < 0) continue; // ignore UNDEF indices
3230 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3237 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3238 unsigned &WhichResult) {
3239 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3243 unsigned NumElts = VT.getVectorNumElements();
3244 WhichResult = (M[0] == 0 ? 0 : 1);
3245 for (unsigned i = 0; i < NumElts; i += 2) {
3246 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3247 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3253 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3254 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3255 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3256 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3257 unsigned &WhichResult) {
3258 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3262 unsigned NumElts = VT.getVectorNumElements();
3263 WhichResult = (M[0] == 0 ? 0 : 1);
3264 for (unsigned i = 0; i < NumElts; i += 2) {
3265 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3266 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3272 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3273 unsigned &WhichResult) {
3274 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3278 unsigned NumElts = VT.getVectorNumElements();
3279 WhichResult = (M[0] == 0 ? 0 : 1);
3280 for (unsigned i = 0; i != NumElts; ++i) {
3281 if (M[i] < 0) continue; // ignore UNDEF indices
3282 if ((unsigned) M[i] != 2 * i + WhichResult)
3286 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3287 if (VT.is64BitVector() && EltSz == 32)
3293 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3294 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3295 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3296 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3297 unsigned &WhichResult) {
3298 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3302 unsigned Half = VT.getVectorNumElements() / 2;
3303 WhichResult = (M[0] == 0 ? 0 : 1);
3304 for (unsigned j = 0; j != 2; ++j) {
3305 unsigned Idx = WhichResult;
3306 for (unsigned i = 0; i != Half; ++i) {
3307 int MIdx = M[i + j * Half];
3308 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3314 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3315 if (VT.is64BitVector() && EltSz == 32)
3321 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3322 unsigned &WhichResult) {
3323 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3327 unsigned NumElts = VT.getVectorNumElements();
3328 WhichResult = (M[0] == 0 ? 0 : 1);
3329 unsigned Idx = WhichResult * NumElts / 2;
3330 for (unsigned i = 0; i != NumElts; i += 2) {
3331 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3332 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3337 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3338 if (VT.is64BitVector() && EltSz == 32)
3344 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3345 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3346 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3347 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3348 unsigned &WhichResult) {
3349 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3353 unsigned NumElts = VT.getVectorNumElements();
3354 WhichResult = (M[0] == 0 ? 0 : 1);
3355 unsigned Idx = WhichResult * NumElts / 2;
3356 for (unsigned i = 0; i != NumElts; i += 2) {
3357 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3358 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3363 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3364 if (VT.is64BitVector() && EltSz == 32)
3370 // If N is an integer constant that can be moved into a register in one
3371 // instruction, return an SDValue of such a constant (will become a MOV
3372 // instruction). Otherwise return null.
3373 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3374 const ARMSubtarget *ST, DebugLoc dl) {
3376 if (!isa<ConstantSDNode>(N))
3378 Val = cast<ConstantSDNode>(N)->getZExtValue();
3380 if (ST->isThumb1Only()) {
3381 if (Val <= 255 || ~Val <= 255)
3382 return DAG.getConstant(Val, MVT::i32);
3384 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3385 return DAG.getConstant(Val, MVT::i32);
3390 // If this is a case we can't handle, return null and let the default
3391 // expansion code take care of it.
3392 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3393 const ARMSubtarget *ST) {
3394 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3395 DebugLoc dl = Op.getDebugLoc();
3396 EVT VT = Op.getValueType();
3398 APInt SplatBits, SplatUndef;
3399 unsigned SplatBitSize;
3401 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3402 if (SplatBitSize <= 64) {
3403 // Check if an immediate VMOV works.
3405 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3406 SplatUndef.getZExtValue(), SplatBitSize,
3407 DAG, VmovVT, VT.is128BitVector(), true);
3408 if (Val.getNode()) {
3409 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3410 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3413 // Try an immediate VMVN.
3414 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3415 ((1LL << SplatBitSize) - 1));
3416 Val = isNEONModifiedImm(NegatedImm,
3417 SplatUndef.getZExtValue(), SplatBitSize,
3418 DAG, VmovVT, VT.is128BitVector(), false);
3419 if (Val.getNode()) {
3420 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3421 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3426 // Scan through the operands to see if only one value is used.
3427 unsigned NumElts = VT.getVectorNumElements();
3428 bool isOnlyLowElement = true;
3429 bool usesOnlyOneValue = true;
3430 bool isConstant = true;
3432 for (unsigned i = 0; i < NumElts; ++i) {
3433 SDValue V = Op.getOperand(i);
3434 if (V.getOpcode() == ISD::UNDEF)
3437 isOnlyLowElement = false;
3438 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3441 if (!Value.getNode())
3443 else if (V != Value)
3444 usesOnlyOneValue = false;
3447 if (!Value.getNode())
3448 return DAG.getUNDEF(VT);
3450 if (isOnlyLowElement)
3451 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3453 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3455 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3456 // i32 and try again.
3457 if (usesOnlyOneValue && EltSize <= 32) {
3459 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3460 if (VT.getVectorElementType().isFloatingPoint()) {
3461 SmallVector<SDValue, 8> Ops;
3462 for (unsigned i = 0; i < NumElts; ++i)
3463 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3465 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3467 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3469 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3471 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3473 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3476 // If all elements are constants and the case above didn't get hit, fall back
3477 // to the default expansion, which will generate a load from the constant
3482 // Vectors with 32- or 64-bit elements can be built by directly assigning
3483 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3484 // will be legalized.
3485 if (EltSize >= 32) {
3486 // Do the expansion with floating-point types, since that is what the VFP
3487 // registers are defined to use, and since i64 is not legal.
3488 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3489 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3490 SmallVector<SDValue, 8> Ops;
3491 for (unsigned i = 0; i < NumElts; ++i)
3492 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3493 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3494 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3500 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3501 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3502 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3503 /// are assumed to be legal.
3505 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3507 if (VT.getVectorNumElements() == 4 &&
3508 (VT.is128BitVector() || VT.is64BitVector())) {
3509 unsigned PFIndexes[4];
3510 for (unsigned i = 0; i != 4; ++i) {
3514 PFIndexes[i] = M[i];
3517 // Compute the index in the perfect shuffle table.
3518 unsigned PFTableIndex =
3519 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3520 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3521 unsigned Cost = (PFEntry >> 30);
3528 unsigned Imm, WhichResult;
3530 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3531 return (EltSize >= 32 ||
3532 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3533 isVREVMask(M, VT, 64) ||
3534 isVREVMask(M, VT, 32) ||
3535 isVREVMask(M, VT, 16) ||
3536 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3537 isVTRNMask(M, VT, WhichResult) ||
3538 isVUZPMask(M, VT, WhichResult) ||
3539 isVZIPMask(M, VT, WhichResult) ||
3540 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3541 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3542 isVZIP_v_undef_Mask(M, VT, WhichResult));
3545 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3546 /// the specified operations to build the shuffle.
3547 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3548 SDValue RHS, SelectionDAG &DAG,
3550 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3551 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3552 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3555 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3564 OP_VUZPL, // VUZP, left result
3565 OP_VUZPR, // VUZP, right result
3566 OP_VZIPL, // VZIP, left result
3567 OP_VZIPR, // VZIP, right result
3568 OP_VTRNL, // VTRN, left result
3569 OP_VTRNR // VTRN, right result
3572 if (OpNum == OP_COPY) {
3573 if (LHSID == (1*9+2)*9+3) return LHS;
3574 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3578 SDValue OpLHS, OpRHS;
3579 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3580 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3581 EVT VT = OpLHS.getValueType();
3584 default: llvm_unreachable("Unknown shuffle opcode!");
3586 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3591 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3592 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3596 return DAG.getNode(ARMISD::VEXT, dl, VT,
3598 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3601 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3602 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3605 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3606 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3609 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3610 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3614 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3615 SDValue V1 = Op.getOperand(0);
3616 SDValue V2 = Op.getOperand(1);
3617 DebugLoc dl = Op.getDebugLoc();
3618 EVT VT = Op.getValueType();
3619 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3620 SmallVector<int, 8> ShuffleMask;
3622 // Convert shuffles that are directly supported on NEON to target-specific
3623 // DAG nodes, instead of keeping them as shuffles and matching them again
3624 // during code selection. This is more efficient and avoids the possibility
3625 // of inconsistencies between legalization and selection.
3626 // FIXME: floating-point vectors should be canonicalized to integer vectors
3627 // of the same time so that they get CSEd properly.
3628 SVN->getMask(ShuffleMask);
3630 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3631 if (EltSize <= 32) {
3632 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3633 int Lane = SVN->getSplatIndex();
3634 // If this is undef splat, generate it via "just" vdup, if possible.
3635 if (Lane == -1) Lane = 0;
3637 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3638 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3640 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3641 DAG.getConstant(Lane, MVT::i32));
3646 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3649 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3650 DAG.getConstant(Imm, MVT::i32));
3653 if (isVREVMask(ShuffleMask, VT, 64))
3654 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3655 if (isVREVMask(ShuffleMask, VT, 32))
3656 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3657 if (isVREVMask(ShuffleMask, VT, 16))
3658 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3660 // Check for Neon shuffles that modify both input vectors in place.
3661 // If both results are used, i.e., if there are two shuffles with the same
3662 // source operands and with masks corresponding to both results of one of
3663 // these operations, DAG memoization will ensure that a single node is
3664 // used for both shuffles.
3665 unsigned WhichResult;
3666 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3667 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3668 V1, V2).getValue(WhichResult);
3669 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3670 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3671 V1, V2).getValue(WhichResult);
3672 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3673 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3674 V1, V2).getValue(WhichResult);
3676 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3677 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3678 V1, V1).getValue(WhichResult);
3679 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3680 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3681 V1, V1).getValue(WhichResult);
3682 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3683 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3684 V1, V1).getValue(WhichResult);
3687 // If the shuffle is not directly supported and it has 4 elements, use
3688 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3689 unsigned NumElts = VT.getVectorNumElements();
3691 unsigned PFIndexes[4];
3692 for (unsigned i = 0; i != 4; ++i) {
3693 if (ShuffleMask[i] < 0)
3696 PFIndexes[i] = ShuffleMask[i];
3699 // Compute the index in the perfect shuffle table.
3700 unsigned PFTableIndex =
3701 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3702 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3703 unsigned Cost = (PFEntry >> 30);
3706 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3709 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3710 if (EltSize >= 32) {
3711 // Do the expansion with floating-point types, since that is what the VFP
3712 // registers are defined to use, and since i64 is not legal.
3713 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3714 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3715 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3716 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3717 SmallVector<SDValue, 8> Ops;
3718 for (unsigned i = 0; i < NumElts; ++i) {
3719 if (ShuffleMask[i] < 0)
3720 Ops.push_back(DAG.getUNDEF(EltVT));
3722 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3723 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3724 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3727 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3728 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3734 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3735 EVT VT = Op.getValueType();
3736 DebugLoc dl = Op.getDebugLoc();
3737 SDValue Vec = Op.getOperand(0);
3738 SDValue Lane = Op.getOperand(1);
3739 assert(VT == MVT::i32 &&
3740 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3741 "unexpected type for custom-lowering vector extract");
3742 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3745 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3746 // The only time a CONCAT_VECTORS operation can have legal types is when
3747 // two 64-bit vectors are concatenated to a 128-bit vector.
3748 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3749 "unexpected CONCAT_VECTORS");
3750 DebugLoc dl = Op.getDebugLoc();
3751 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3752 SDValue Op0 = Op.getOperand(0);
3753 SDValue Op1 = Op.getOperand(1);
3754 if (Op0.getOpcode() != ISD::UNDEF)
3755 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3756 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3757 DAG.getIntPtrConstant(0));
3758 if (Op1.getOpcode() != ISD::UNDEF)
3759 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3760 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3761 DAG.getIntPtrConstant(1));
3762 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3765 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3766 /// an extending load, return the unextended value.
3767 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3768 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3769 return N->getOperand(0);
3770 LoadSDNode *LD = cast<LoadSDNode>(N);
3771 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3772 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3773 LD->isNonTemporal(), LD->getAlignment());
3776 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3777 // Multiplications are only custom-lowered for 128-bit vectors so that
3778 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3779 EVT VT = Op.getValueType();
3780 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3781 SDNode *N0 = Op.getOperand(0).getNode();
3782 SDNode *N1 = Op.getOperand(1).getNode();
3783 unsigned NewOpc = 0;
3784 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3785 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3786 NewOpc = ARMISD::VMULLs;
3787 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3788 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3789 NewOpc = ARMISD::VMULLu;
3790 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3791 // Fall through to expand this. It is not legal.
3794 // Other vector multiplications are legal.
3798 // Legalize to a VMULL instruction.
3799 DebugLoc DL = Op.getDebugLoc();
3800 SDValue Op0 = SkipExtension(N0, DAG);
3801 SDValue Op1 = SkipExtension(N1, DAG);
3803 assert(Op0.getValueType().is64BitVector() &&
3804 Op1.getValueType().is64BitVector() &&
3805 "unexpected types for extended operands to VMULL");
3806 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3809 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3810 switch (Op.getOpcode()) {
3811 default: llvm_unreachable("Don't know how to custom lower this!");
3812 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3813 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3814 case ISD::GlobalAddress:
3815 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3816 LowerGlobalAddressELF(Op, DAG);
3817 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3818 case ISD::SELECT: return LowerSELECT(Op, DAG);
3819 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3820 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3821 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3822 case ISD::VASTART: return LowerVASTART(Op, DAG);
3823 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3824 case ISD::SINT_TO_FP:
3825 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3826 case ISD::FP_TO_SINT:
3827 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3828 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3829 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3830 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3831 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3832 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3833 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3834 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
3835 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3837 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3840 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3841 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3842 case ISD::SRL_PARTS:
3843 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3844 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3845 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3846 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3847 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3848 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3849 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3850 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3851 case ISD::MUL: return LowerMUL(Op, DAG);
3856 /// ReplaceNodeResults - Replace the results of node with an illegal result
3857 /// type with new values built out of custom code.
3858 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3859 SmallVectorImpl<SDValue>&Results,
3860 SelectionDAG &DAG) const {
3862 switch (N->getOpcode()) {
3864 llvm_unreachable("Don't know how to custom expand this!");
3866 case ISD::BIT_CONVERT:
3867 Res = ExpandBIT_CONVERT(N, DAG);
3871 Res = LowerShift(N, DAG, Subtarget);
3875 Results.push_back(Res);
3878 //===----------------------------------------------------------------------===//
3879 // ARM Scheduler Hooks
3880 //===----------------------------------------------------------------------===//
3883 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3884 MachineBasicBlock *BB,
3885 unsigned Size) const {
3886 unsigned dest = MI->getOperand(0).getReg();
3887 unsigned ptr = MI->getOperand(1).getReg();
3888 unsigned oldval = MI->getOperand(2).getReg();
3889 unsigned newval = MI->getOperand(3).getReg();
3890 unsigned scratch = BB->getParent()->getRegInfo()
3891 .createVirtualRegister(ARM::GPRRegisterClass);
3892 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3893 DebugLoc dl = MI->getDebugLoc();
3894 bool isThumb2 = Subtarget->isThumb2();
3896 unsigned ldrOpc, strOpc;
3898 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3900 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3901 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3904 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3905 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3908 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3909 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3913 MachineFunction *MF = BB->getParent();
3914 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3915 MachineFunction::iterator It = BB;
3916 ++It; // insert the new blocks after the current block
3918 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3919 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3920 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3921 MF->insert(It, loop1MBB);
3922 MF->insert(It, loop2MBB);
3923 MF->insert(It, exitMBB);
3925 // Transfer the remainder of BB and its successor edges to exitMBB.
3926 exitMBB->splice(exitMBB->begin(), BB,
3927 llvm::next(MachineBasicBlock::iterator(MI)),
3929 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3933 // fallthrough --> loop1MBB
3934 BB->addSuccessor(loop1MBB);
3937 // ldrex dest, [ptr]
3941 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3942 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3943 .addReg(dest).addReg(oldval));
3944 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3945 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3946 BB->addSuccessor(loop2MBB);
3947 BB->addSuccessor(exitMBB);
3950 // strex scratch, newval, [ptr]
3954 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3956 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3957 .addReg(scratch).addImm(0));
3958 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3959 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3960 BB->addSuccessor(loop1MBB);
3961 BB->addSuccessor(exitMBB);
3967 MI->eraseFromParent(); // The instruction is gone now.
3973 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3974 unsigned Size, unsigned BinOpcode) const {
3975 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3976 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3978 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3979 MachineFunction *MF = BB->getParent();
3980 MachineFunction::iterator It = BB;
3983 unsigned dest = MI->getOperand(0).getReg();
3984 unsigned ptr = MI->getOperand(1).getReg();
3985 unsigned incr = MI->getOperand(2).getReg();
3986 DebugLoc dl = MI->getDebugLoc();
3988 bool isThumb2 = Subtarget->isThumb2();
3989 unsigned ldrOpc, strOpc;
3991 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3993 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3994 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3997 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3998 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4001 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4002 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4006 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4007 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4008 MF->insert(It, loopMBB);
4009 MF->insert(It, exitMBB);
4011 // Transfer the remainder of BB and its successor edges to exitMBB.
4012 exitMBB->splice(exitMBB->begin(), BB,
4013 llvm::next(MachineBasicBlock::iterator(MI)),
4015 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4017 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4018 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4019 unsigned scratch2 = (!BinOpcode) ? incr :
4020 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4024 // fallthrough --> loopMBB
4025 BB->addSuccessor(loopMBB);
4029 // <binop> scratch2, dest, incr
4030 // strex scratch, scratch2, ptr
4033 // fallthrough --> exitMBB
4035 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4037 // operand order needs to go the other way for NAND
4038 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4039 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4040 addReg(incr).addReg(dest)).addReg(0);
4042 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4043 addReg(dest).addReg(incr)).addReg(0);
4046 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4048 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4049 .addReg(scratch).addImm(0));
4050 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4051 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4053 BB->addSuccessor(loopMBB);
4054 BB->addSuccessor(exitMBB);
4060 MI->eraseFromParent(); // The instruction is gone now.
4066 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4067 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4068 E = MBB->succ_end(); I != E; ++I)
4071 llvm_unreachable("Expecting a BB with two successors!");
4075 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4076 MachineBasicBlock *BB) const {
4077 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4078 DebugLoc dl = MI->getDebugLoc();
4079 bool isThumb2 = Subtarget->isThumb2();
4080 switch (MI->getOpcode()) {
4083 llvm_unreachable("Unexpected instr type to insert");
4085 case ARM::ATOMIC_LOAD_ADD_I8:
4086 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4087 case ARM::ATOMIC_LOAD_ADD_I16:
4088 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4089 case ARM::ATOMIC_LOAD_ADD_I32:
4090 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4092 case ARM::ATOMIC_LOAD_AND_I8:
4093 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4094 case ARM::ATOMIC_LOAD_AND_I16:
4095 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4096 case ARM::ATOMIC_LOAD_AND_I32:
4097 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4099 case ARM::ATOMIC_LOAD_OR_I8:
4100 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4101 case ARM::ATOMIC_LOAD_OR_I16:
4102 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4103 case ARM::ATOMIC_LOAD_OR_I32:
4104 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4106 case ARM::ATOMIC_LOAD_XOR_I8:
4107 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4108 case ARM::ATOMIC_LOAD_XOR_I16:
4109 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4110 case ARM::ATOMIC_LOAD_XOR_I32:
4111 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4113 case ARM::ATOMIC_LOAD_NAND_I8:
4114 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4115 case ARM::ATOMIC_LOAD_NAND_I16:
4116 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4117 case ARM::ATOMIC_LOAD_NAND_I32:
4118 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4120 case ARM::ATOMIC_LOAD_SUB_I8:
4121 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4122 case ARM::ATOMIC_LOAD_SUB_I16:
4123 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4124 case ARM::ATOMIC_LOAD_SUB_I32:
4125 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4127 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4128 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4129 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4131 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4132 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4133 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4135 case ARM::tMOVCCr_pseudo: {
4136 // To "insert" a SELECT_CC instruction, we actually have to insert the
4137 // diamond control-flow pattern. The incoming instruction knows the
4138 // destination vreg to set, the condition code register to branch on, the
4139 // true/false values to select between, and a branch opcode to use.
4140 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4141 MachineFunction::iterator It = BB;
4147 // cmpTY ccX, r1, r2
4149 // fallthrough --> copy0MBB
4150 MachineBasicBlock *thisMBB = BB;
4151 MachineFunction *F = BB->getParent();
4152 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4153 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4154 F->insert(It, copy0MBB);
4155 F->insert(It, sinkMBB);
4157 // Transfer the remainder of BB and its successor edges to sinkMBB.
4158 sinkMBB->splice(sinkMBB->begin(), BB,
4159 llvm::next(MachineBasicBlock::iterator(MI)),
4161 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4163 BB->addSuccessor(copy0MBB);
4164 BB->addSuccessor(sinkMBB);
4166 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4167 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4170 // %FalseValue = ...
4171 // # fallthrough to sinkMBB
4174 // Update machine-CFG edges
4175 BB->addSuccessor(sinkMBB);
4178 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4181 BuildMI(*BB, BB->begin(), dl,
4182 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4183 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4184 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4186 MI->eraseFromParent(); // The pseudo instruction is gone now.
4191 case ARM::BCCZi64: {
4192 // Compare both parts that make up the double comparison separately for
4194 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4196 unsigned LHS1 = MI->getOperand(1).getReg();
4197 unsigned LHS2 = MI->getOperand(2).getReg();
4199 AddDefaultPred(BuildMI(BB, dl,
4200 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4201 .addReg(LHS1).addImm(0));
4202 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4203 .addReg(LHS2).addImm(0)
4204 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4206 unsigned RHS1 = MI->getOperand(3).getReg();
4207 unsigned RHS2 = MI->getOperand(4).getReg();
4208 AddDefaultPred(BuildMI(BB, dl,
4209 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4210 .addReg(LHS1).addReg(RHS1));
4211 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4212 .addReg(LHS2).addReg(RHS2)
4213 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4216 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4217 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4218 if (MI->getOperand(0).getImm() == ARMCC::NE)
4219 std::swap(destMBB, exitMBB);
4221 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4222 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4223 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4226 MI->eraseFromParent(); // The pseudo instruction is gone now.
4232 //===----------------------------------------------------------------------===//
4233 // ARM Optimization Hooks
4234 //===----------------------------------------------------------------------===//
4237 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4238 TargetLowering::DAGCombinerInfo &DCI) {
4239 SelectionDAG &DAG = DCI.DAG;
4240 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4241 EVT VT = N->getValueType(0);
4242 unsigned Opc = N->getOpcode();
4243 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4244 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4245 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4246 ISD::CondCode CC = ISD::SETCC_INVALID;
4249 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4251 SDValue CCOp = Slct.getOperand(0);
4252 if (CCOp.getOpcode() == ISD::SETCC)
4253 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4256 bool DoXform = false;
4258 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4261 if (LHS.getOpcode() == ISD::Constant &&
4262 cast<ConstantSDNode>(LHS)->isNullValue()) {
4264 } else if (CC != ISD::SETCC_INVALID &&
4265 RHS.getOpcode() == ISD::Constant &&
4266 cast<ConstantSDNode>(RHS)->isNullValue()) {
4267 std::swap(LHS, RHS);
4268 SDValue Op0 = Slct.getOperand(0);
4269 EVT OpVT = isSlctCC ? Op0.getValueType() :
4270 Op0.getOperand(0).getValueType();
4271 bool isInt = OpVT.isInteger();
4272 CC = ISD::getSetCCInverse(CC, isInt);
4274 if (!TLI.isCondCodeLegal(CC, OpVT))
4275 return SDValue(); // Inverse operator isn't legal.
4282 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4284 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4285 Slct.getOperand(0), Slct.getOperand(1), CC);
4286 SDValue CCOp = Slct.getOperand(0);
4288 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4289 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4290 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4291 CCOp, OtherOp, Result);
4296 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4297 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4298 /// called with the default operands, and if that fails, with commuted
4300 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4301 TargetLowering::DAGCombinerInfo &DCI) {
4302 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4303 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4304 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4305 if (Result.getNode()) return Result;
4310 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4312 static SDValue PerformADDCombine(SDNode *N,
4313 TargetLowering::DAGCombinerInfo &DCI) {
4314 SDValue N0 = N->getOperand(0);
4315 SDValue N1 = N->getOperand(1);
4317 // First try with the default operand order.
4318 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4319 if (Result.getNode())
4322 // If that didn't work, try again with the operands commuted.
4323 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4326 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4328 static SDValue PerformSUBCombine(SDNode *N,
4329 TargetLowering::DAGCombinerInfo &DCI) {
4330 SDValue N0 = N->getOperand(0);
4331 SDValue N1 = N->getOperand(1);
4333 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4334 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4335 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4336 if (Result.getNode()) return Result;
4342 static SDValue PerformMULCombine(SDNode *N,
4343 TargetLowering::DAGCombinerInfo &DCI,
4344 const ARMSubtarget *Subtarget) {
4345 SelectionDAG &DAG = DCI.DAG;
4347 if (Subtarget->isThumb1Only())
4350 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4353 EVT VT = N->getValueType(0);
4357 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4361 uint64_t MulAmt = C->getZExtValue();
4362 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4363 ShiftAmt = ShiftAmt & (32 - 1);
4364 SDValue V = N->getOperand(0);
4365 DebugLoc DL = N->getDebugLoc();
4368 MulAmt >>= ShiftAmt;
4369 if (isPowerOf2_32(MulAmt - 1)) {
4370 // (mul x, 2^N + 1) => (add (shl x, N), x)
4371 Res = DAG.getNode(ISD::ADD, DL, VT,
4372 V, DAG.getNode(ISD::SHL, DL, VT,
4373 V, DAG.getConstant(Log2_32(MulAmt-1),
4375 } else if (isPowerOf2_32(MulAmt + 1)) {
4376 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4377 Res = DAG.getNode(ISD::SUB, DL, VT,
4378 DAG.getNode(ISD::SHL, DL, VT,
4379 V, DAG.getConstant(Log2_32(MulAmt+1),
4386 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4387 DAG.getConstant(ShiftAmt, MVT::i32));
4389 // Do not add new nodes to DAG combiner worklist.
4390 DCI.CombineTo(N, Res, false);
4394 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4395 static SDValue PerformORCombine(SDNode *N,
4396 TargetLowering::DAGCombinerInfo &DCI,
4397 const ARMSubtarget *Subtarget) {
4398 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4401 // BFI is only available on V6T2+
4402 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4405 SelectionDAG &DAG = DCI.DAG;
4406 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4407 DebugLoc DL = N->getDebugLoc();
4408 // 1) or (and A, mask), val => ARMbfi A, val, mask
4409 // iff (val & mask) == val
4411 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4412 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4413 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4414 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4415 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4416 // (i.e., copy a bitfield value into another bitfield of the same width)
4417 if (N0.getOpcode() != ISD::AND)
4420 EVT VT = N->getValueType(0);
4425 // The value and the mask need to be constants so we can verify this is
4426 // actually a bitfield set. If the mask is 0xffff, we can do better
4427 // via a movt instruction, so don't use BFI in that case.
4428 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4431 unsigned Mask = C->getZExtValue();
4435 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4436 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4437 unsigned Val = C->getZExtValue();
4438 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4440 Val >>= CountTrailingZeros_32(~Mask);
4442 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4443 DAG.getConstant(Val, MVT::i32),
4444 DAG.getConstant(Mask, MVT::i32));
4446 // Do not add new nodes to DAG combiner worklist.
4447 DCI.CombineTo(N, Res, false);
4448 } else if (N1.getOpcode() == ISD::AND) {
4449 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4450 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4453 unsigned Mask2 = C->getZExtValue();
4455 if (ARM::isBitFieldInvertedMask(Mask) &&
4456 ARM::isBitFieldInvertedMask(~Mask2) &&
4457 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4458 // The pack halfword instruction works better for masks that fit it,
4459 // so use that when it's available.
4460 if (Subtarget->hasT2ExtractPack() &&
4461 (Mask == 0xffff || Mask == 0xffff0000))
4464 unsigned lsb = CountTrailingZeros_32(Mask2);
4465 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4466 DAG.getConstant(lsb, MVT::i32));
4467 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4468 DAG.getConstant(Mask, MVT::i32));
4469 // Do not add new nodes to DAG combiner worklist.
4470 DCI.CombineTo(N, Res, false);
4471 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4472 ARM::isBitFieldInvertedMask(Mask2) &&
4473 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4474 // The pack halfword instruction works better for masks that fit it,
4475 // so use that when it's available.
4476 if (Subtarget->hasT2ExtractPack() &&
4477 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4480 unsigned lsb = CountTrailingZeros_32(Mask);
4481 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4482 DAG.getConstant(lsb, MVT::i32));
4483 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4484 DAG.getConstant(Mask2, MVT::i32));
4485 // Do not add new nodes to DAG combiner worklist.
4486 DCI.CombineTo(N, Res, false);
4493 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4494 /// ARMISD::VMOVRRD.
4495 static SDValue PerformVMOVRRDCombine(SDNode *N,
4496 TargetLowering::DAGCombinerInfo &DCI) {
4497 // vmovrrd(vmovdrr x, y) -> x,y
4498 SDValue InDouble = N->getOperand(0);
4499 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4500 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4504 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4505 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4506 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4507 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4508 SDValue Op0 = N->getOperand(0);
4509 SDValue Op1 = N->getOperand(1);
4510 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4511 Op0 = Op0.getOperand(0);
4512 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4513 Op1 = Op1.getOperand(0);
4514 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4515 Op0.getNode() == Op1.getNode() &&
4516 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4517 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4518 N->getValueType(0), Op0.getOperand(0));
4522 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4523 /// ISD::BUILD_VECTOR.
4524 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4525 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4526 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4527 // into a pair of GPRs, which is fine when the value is used as a scalar,
4528 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4529 if (N->getNumOperands() == 2)
4530 return PerformVMOVDRRCombine(N, DAG);
4535 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
4536 /// ISD::VECTOR_SHUFFLE.
4537 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
4538 // The LLVM shufflevector instruction does not require the shuffle mask
4539 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
4540 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
4541 // operands do not match the mask length, they are extended by concatenating
4542 // them with undef vectors. That is probably the right thing for other
4543 // targets, but for NEON it is better to concatenate two double-register
4544 // size vector operands into a single quad-register size vector. Do that
4545 // transformation here:
4546 // shuffle(concat(v1, undef), concat(v2, undef)) ->
4547 // shuffle(concat(v1, v2), undef)
4548 SDValue Op0 = N->getOperand(0);
4549 SDValue Op1 = N->getOperand(1);
4550 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
4551 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
4552 Op0.getNumOperands() != 2 ||
4553 Op1.getNumOperands() != 2)
4555 SDValue Concat0Op1 = Op0.getOperand(1);
4556 SDValue Concat1Op1 = Op1.getOperand(1);
4557 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
4558 Concat1Op1.getOpcode() != ISD::UNDEF)
4560 // Skip the transformation if any of the types are illegal.
4561 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4562 EVT VT = N->getValueType(0);
4563 if (!TLI.isTypeLegal(VT) ||
4564 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
4565 !TLI.isTypeLegal(Concat1Op1.getValueType()))
4568 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
4569 Op0.getOperand(0), Op1.getOperand(0));
4570 // Translate the shuffle mask.
4571 SmallVector<int, 16> NewMask;
4572 unsigned NumElts = VT.getVectorNumElements();
4573 unsigned HalfElts = NumElts/2;
4574 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
4575 for (unsigned n = 0; n < NumElts; ++n) {
4576 int MaskElt = SVN->getMaskElt(n);
4578 if (MaskElt < (int)HalfElts)
4580 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
4581 NewElt = HalfElts + MaskElt - NumElts;
4582 NewMask.push_back(NewElt);
4584 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
4585 DAG.getUNDEF(VT), NewMask.data());
4588 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4589 /// ARMISD::VDUPLANE.
4590 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
4591 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4593 SDValue Op = N->getOperand(0);
4594 EVT VT = N->getValueType(0);
4596 // Ignore bit_converts.
4597 while (Op.getOpcode() == ISD::BIT_CONVERT)
4598 Op = Op.getOperand(0);
4599 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4602 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4603 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4604 // The canonical VMOV for a zero vector uses a 32-bit element size.
4605 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4607 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4609 if (EltSize > VT.getVectorElementType().getSizeInBits())
4612 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4615 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4616 /// operand of a vector shift operation, where all the elements of the
4617 /// build_vector must have the same constant integer value.
4618 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4619 // Ignore bit_converts.
4620 while (Op.getOpcode() == ISD::BIT_CONVERT)
4621 Op = Op.getOperand(0);
4622 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4623 APInt SplatBits, SplatUndef;
4624 unsigned SplatBitSize;
4626 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4627 HasAnyUndefs, ElementBits) ||
4628 SplatBitSize > ElementBits)
4630 Cnt = SplatBits.getSExtValue();
4634 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4635 /// operand of a vector shift left operation. That value must be in the range:
4636 /// 0 <= Value < ElementBits for a left shift; or
4637 /// 0 <= Value <= ElementBits for a long left shift.
4638 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4639 assert(VT.isVector() && "vector shift count is not a vector type");
4640 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4641 if (! getVShiftImm(Op, ElementBits, Cnt))
4643 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4646 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4647 /// operand of a vector shift right operation. For a shift opcode, the value
4648 /// is positive, but for an intrinsic the value count must be negative. The
4649 /// absolute value must be in the range:
4650 /// 1 <= |Value| <= ElementBits for a right shift; or
4651 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4652 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4654 assert(VT.isVector() && "vector shift count is not a vector type");
4655 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4656 if (! getVShiftImm(Op, ElementBits, Cnt))
4660 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4663 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4664 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4665 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4668 // Don't do anything for most intrinsics.
4671 // Vector shifts: check for immediate versions and lower them.
4672 // Note: This is done during DAG combining instead of DAG legalizing because
4673 // the build_vectors for 64-bit vector element shift counts are generally
4674 // not legal, and it is hard to see their values after they get legalized to
4675 // loads from a constant pool.
4676 case Intrinsic::arm_neon_vshifts:
4677 case Intrinsic::arm_neon_vshiftu:
4678 case Intrinsic::arm_neon_vshiftls:
4679 case Intrinsic::arm_neon_vshiftlu:
4680 case Intrinsic::arm_neon_vshiftn:
4681 case Intrinsic::arm_neon_vrshifts:
4682 case Intrinsic::arm_neon_vrshiftu:
4683 case Intrinsic::arm_neon_vrshiftn:
4684 case Intrinsic::arm_neon_vqshifts:
4685 case Intrinsic::arm_neon_vqshiftu:
4686 case Intrinsic::arm_neon_vqshiftsu:
4687 case Intrinsic::arm_neon_vqshiftns:
4688 case Intrinsic::arm_neon_vqshiftnu:
4689 case Intrinsic::arm_neon_vqshiftnsu:
4690 case Intrinsic::arm_neon_vqrshiftns:
4691 case Intrinsic::arm_neon_vqrshiftnu:
4692 case Intrinsic::arm_neon_vqrshiftnsu: {
4693 EVT VT = N->getOperand(1).getValueType();
4695 unsigned VShiftOpc = 0;
4698 case Intrinsic::arm_neon_vshifts:
4699 case Intrinsic::arm_neon_vshiftu:
4700 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4701 VShiftOpc = ARMISD::VSHL;
4704 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4705 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4706 ARMISD::VSHRs : ARMISD::VSHRu);
4711 case Intrinsic::arm_neon_vshiftls:
4712 case Intrinsic::arm_neon_vshiftlu:
4713 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4715 llvm_unreachable("invalid shift count for vshll intrinsic");
4717 case Intrinsic::arm_neon_vrshifts:
4718 case Intrinsic::arm_neon_vrshiftu:
4719 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4723 case Intrinsic::arm_neon_vqshifts:
4724 case Intrinsic::arm_neon_vqshiftu:
4725 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4729 case Intrinsic::arm_neon_vqshiftsu:
4730 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4732 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4734 case Intrinsic::arm_neon_vshiftn:
4735 case Intrinsic::arm_neon_vrshiftn:
4736 case Intrinsic::arm_neon_vqshiftns:
4737 case Intrinsic::arm_neon_vqshiftnu:
4738 case Intrinsic::arm_neon_vqshiftnsu:
4739 case Intrinsic::arm_neon_vqrshiftns:
4740 case Intrinsic::arm_neon_vqrshiftnu:
4741 case Intrinsic::arm_neon_vqrshiftnsu:
4742 // Narrowing shifts require an immediate right shift.
4743 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4745 llvm_unreachable("invalid shift count for narrowing vector shift "
4749 llvm_unreachable("unhandled vector shift");
4753 case Intrinsic::arm_neon_vshifts:
4754 case Intrinsic::arm_neon_vshiftu:
4755 // Opcode already set above.
4757 case Intrinsic::arm_neon_vshiftls:
4758 case Intrinsic::arm_neon_vshiftlu:
4759 if (Cnt == VT.getVectorElementType().getSizeInBits())
4760 VShiftOpc = ARMISD::VSHLLi;
4762 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4763 ARMISD::VSHLLs : ARMISD::VSHLLu);
4765 case Intrinsic::arm_neon_vshiftn:
4766 VShiftOpc = ARMISD::VSHRN; break;
4767 case Intrinsic::arm_neon_vrshifts:
4768 VShiftOpc = ARMISD::VRSHRs; break;
4769 case Intrinsic::arm_neon_vrshiftu:
4770 VShiftOpc = ARMISD::VRSHRu; break;
4771 case Intrinsic::arm_neon_vrshiftn:
4772 VShiftOpc = ARMISD::VRSHRN; break;
4773 case Intrinsic::arm_neon_vqshifts:
4774 VShiftOpc = ARMISD::VQSHLs; break;
4775 case Intrinsic::arm_neon_vqshiftu:
4776 VShiftOpc = ARMISD::VQSHLu; break;
4777 case Intrinsic::arm_neon_vqshiftsu:
4778 VShiftOpc = ARMISD::VQSHLsu; break;
4779 case Intrinsic::arm_neon_vqshiftns:
4780 VShiftOpc = ARMISD::VQSHRNs; break;
4781 case Intrinsic::arm_neon_vqshiftnu:
4782 VShiftOpc = ARMISD::VQSHRNu; break;
4783 case Intrinsic::arm_neon_vqshiftnsu:
4784 VShiftOpc = ARMISD::VQSHRNsu; break;
4785 case Intrinsic::arm_neon_vqrshiftns:
4786 VShiftOpc = ARMISD::VQRSHRNs; break;
4787 case Intrinsic::arm_neon_vqrshiftnu:
4788 VShiftOpc = ARMISD::VQRSHRNu; break;
4789 case Intrinsic::arm_neon_vqrshiftnsu:
4790 VShiftOpc = ARMISD::VQRSHRNsu; break;
4793 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4794 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4797 case Intrinsic::arm_neon_vshiftins: {
4798 EVT VT = N->getOperand(1).getValueType();
4800 unsigned VShiftOpc = 0;
4802 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4803 VShiftOpc = ARMISD::VSLI;
4804 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4805 VShiftOpc = ARMISD::VSRI;
4807 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4810 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4811 N->getOperand(1), N->getOperand(2),
4812 DAG.getConstant(Cnt, MVT::i32));
4815 case Intrinsic::arm_neon_vqrshifts:
4816 case Intrinsic::arm_neon_vqrshiftu:
4817 // No immediate versions of these to check for.
4824 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4825 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4826 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4827 /// vector element shift counts are generally not legal, and it is hard to see
4828 /// their values after they get legalized to loads from a constant pool.
4829 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4830 const ARMSubtarget *ST) {
4831 EVT VT = N->getValueType(0);
4833 // Nothing to be done for scalar shifts.
4834 if (! VT.isVector())
4837 assert(ST->hasNEON() && "unexpected vector shift");
4840 switch (N->getOpcode()) {
4841 default: llvm_unreachable("unexpected shift opcode");
4844 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4845 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4846 DAG.getConstant(Cnt, MVT::i32));
4851 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4852 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4853 ARMISD::VSHRs : ARMISD::VSHRu);
4854 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4855 DAG.getConstant(Cnt, MVT::i32));
4861 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4862 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4863 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4864 const ARMSubtarget *ST) {
4865 SDValue N0 = N->getOperand(0);
4867 // Check for sign- and zero-extensions of vector extract operations of 8-
4868 // and 16-bit vector elements. NEON supports these directly. They are
4869 // handled during DAG combining because type legalization will promote them
4870 // to 32-bit types and it is messy to recognize the operations after that.
4871 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4872 SDValue Vec = N0.getOperand(0);
4873 SDValue Lane = N0.getOperand(1);
4874 EVT VT = N->getValueType(0);
4875 EVT EltVT = N0.getValueType();
4876 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4878 if (VT == MVT::i32 &&
4879 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4880 TLI.isTypeLegal(Vec.getValueType())) {
4883 switch (N->getOpcode()) {
4884 default: llvm_unreachable("unexpected opcode");
4885 case ISD::SIGN_EXTEND:
4886 Opc = ARMISD::VGETLANEs;
4888 case ISD::ZERO_EXTEND:
4889 case ISD::ANY_EXTEND:
4890 Opc = ARMISD::VGETLANEu;
4893 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4900 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4901 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4902 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4903 const ARMSubtarget *ST) {
4904 // If the target supports NEON, try to use vmax/vmin instructions for f32
4905 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4906 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4907 // a NaN; only do the transformation when it matches that behavior.
4909 // For now only do this when using NEON for FP operations; if using VFP, it
4910 // is not obvious that the benefit outweighs the cost of switching to the
4912 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4913 N->getValueType(0) != MVT::f32)
4916 SDValue CondLHS = N->getOperand(0);
4917 SDValue CondRHS = N->getOperand(1);
4918 SDValue LHS = N->getOperand(2);
4919 SDValue RHS = N->getOperand(3);
4920 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4922 unsigned Opcode = 0;
4924 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4925 IsReversed = false; // x CC y ? x : y
4926 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4927 IsReversed = true ; // x CC y ? y : x
4941 // If LHS is NaN, an ordered comparison will be false and the result will
4942 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4943 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4944 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4945 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4947 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4948 // will return -0, so vmin can only be used for unsafe math or if one of
4949 // the operands is known to be nonzero.
4950 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4952 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4954 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4963 // If LHS is NaN, an ordered comparison will be false and the result will
4964 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4965 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4966 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4967 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4969 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4970 // will return +0, so vmax can only be used for unsafe math or if one of
4971 // the operands is known to be nonzero.
4972 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4974 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4976 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4982 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4985 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4986 DAGCombinerInfo &DCI) const {
4987 switch (N->getOpcode()) {
4989 case ISD::ADD: return PerformADDCombine(N, DCI);
4990 case ISD::SUB: return PerformSUBCombine(N, DCI);
4991 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4992 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4993 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4994 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
4995 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
4996 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
4997 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
4998 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
5001 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
5002 case ISD::SIGN_EXTEND:
5003 case ISD::ZERO_EXTEND:
5004 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
5005 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
5010 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
5011 if (!Subtarget->allowsUnalignedMem())
5014 switch (VT.getSimpleVT().SimpleTy) {
5021 // FIXME: VLD1 etc with standard alignment is legal.
5025 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
5030 switch (VT.getSimpleVT().SimpleTy) {
5031 default: return false;
5046 if ((V & (Scale - 1)) != 0)
5049 return V == (V & ((1LL << 5) - 1));
5052 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
5053 const ARMSubtarget *Subtarget) {
5060 switch (VT.getSimpleVT().SimpleTy) {
5061 default: return false;
5066 // + imm12 or - imm8
5068 return V == (V & ((1LL << 8) - 1));
5069 return V == (V & ((1LL << 12) - 1));
5072 // Same as ARM mode. FIXME: NEON?
5073 if (!Subtarget->hasVFP2())
5078 return V == (V & ((1LL << 8) - 1));
5082 /// isLegalAddressImmediate - Return true if the integer value can be used
5083 /// as the offset of the target addressing mode for load / store of the
5085 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5086 const ARMSubtarget *Subtarget) {
5093 if (Subtarget->isThumb1Only())
5094 return isLegalT1AddressImmediate(V, VT);
5095 else if (Subtarget->isThumb2())
5096 return isLegalT2AddressImmediate(V, VT, Subtarget);
5101 switch (VT.getSimpleVT().SimpleTy) {
5102 default: return false;
5107 return V == (V & ((1LL << 12) - 1));
5110 return V == (V & ((1LL << 8) - 1));
5113 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5118 return V == (V & ((1LL << 8) - 1));
5122 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5124 int Scale = AM.Scale;
5128 switch (VT.getSimpleVT().SimpleTy) {
5129 default: return false;
5138 return Scale == 2 || Scale == 4 || Scale == 8;
5141 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5145 // Note, we allow "void" uses (basically, uses that aren't loads or
5146 // stores), because arm allows folding a scale into many arithmetic
5147 // operations. This should be made more precise and revisited later.
5149 // Allow r << imm, but the imm has to be a multiple of two.
5150 if (Scale & 1) return false;
5151 return isPowerOf2_32(Scale);
5155 /// isLegalAddressingMode - Return true if the addressing mode represented
5156 /// by AM is legal for this target, for a load/store of the specified type.
5157 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5158 const Type *Ty) const {
5159 EVT VT = getValueType(Ty, true);
5160 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5163 // Can never fold addr of global into load/store.
5168 case 0: // no scale reg, must be "r+i" or "r", or "i".
5171 if (Subtarget->isThumb1Only())
5175 // ARM doesn't support any R+R*scale+imm addr modes.
5182 if (Subtarget->isThumb2())
5183 return isLegalT2ScaledAddressingMode(AM, VT);
5185 int Scale = AM.Scale;
5186 switch (VT.getSimpleVT().SimpleTy) {
5187 default: return false;
5191 if (Scale < 0) Scale = -Scale;
5195 return isPowerOf2_32(Scale & ~1);
5199 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5204 // Note, we allow "void" uses (basically, uses that aren't loads or
5205 // stores), because arm allows folding a scale into many arithmetic
5206 // operations. This should be made more precise and revisited later.
5208 // Allow r << imm, but the imm has to be a multiple of two.
5209 if (Scale & 1) return false;
5210 return isPowerOf2_32(Scale);
5217 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5218 /// icmp immediate, that is the target has icmp instructions which can compare
5219 /// a register against the immediate without having to materialize the
5220 /// immediate into a register.
5221 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5222 if (!Subtarget->isThumb())
5223 return ARM_AM::getSOImmVal(Imm) != -1;
5224 if (Subtarget->isThumb2())
5225 return ARM_AM::getT2SOImmVal(Imm) != -1;
5226 return Imm >= 0 && Imm <= 255;
5229 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5230 bool isSEXTLoad, SDValue &Base,
5231 SDValue &Offset, bool &isInc,
5232 SelectionDAG &DAG) {
5233 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5236 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5238 Base = Ptr->getOperand(0);
5239 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5240 int RHSC = (int)RHS->getZExtValue();
5241 if (RHSC < 0 && RHSC > -256) {
5242 assert(Ptr->getOpcode() == ISD::ADD);
5244 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5248 isInc = (Ptr->getOpcode() == ISD::ADD);
5249 Offset = Ptr->getOperand(1);
5251 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5253 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5254 int RHSC = (int)RHS->getZExtValue();
5255 if (RHSC < 0 && RHSC > -0x1000) {
5256 assert(Ptr->getOpcode() == ISD::ADD);
5258 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5259 Base = Ptr->getOperand(0);
5264 if (Ptr->getOpcode() == ISD::ADD) {
5266 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5267 if (ShOpcVal != ARM_AM::no_shift) {
5268 Base = Ptr->getOperand(1);
5269 Offset = Ptr->getOperand(0);
5271 Base = Ptr->getOperand(0);
5272 Offset = Ptr->getOperand(1);
5277 isInc = (Ptr->getOpcode() == ISD::ADD);
5278 Base = Ptr->getOperand(0);
5279 Offset = Ptr->getOperand(1);
5283 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5287 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5288 bool isSEXTLoad, SDValue &Base,
5289 SDValue &Offset, bool &isInc,
5290 SelectionDAG &DAG) {
5291 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5294 Base = Ptr->getOperand(0);
5295 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5296 int RHSC = (int)RHS->getZExtValue();
5297 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5298 assert(Ptr->getOpcode() == ISD::ADD);
5300 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5302 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5303 isInc = Ptr->getOpcode() == ISD::ADD;
5304 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5312 /// getPreIndexedAddressParts - returns true by value, base pointer and
5313 /// offset pointer and addressing mode by reference if the node's address
5314 /// can be legally represented as pre-indexed load / store address.
5316 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5318 ISD::MemIndexedMode &AM,
5319 SelectionDAG &DAG) const {
5320 if (Subtarget->isThumb1Only())
5325 bool isSEXTLoad = false;
5326 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5327 Ptr = LD->getBasePtr();
5328 VT = LD->getMemoryVT();
5329 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5330 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5331 Ptr = ST->getBasePtr();
5332 VT = ST->getMemoryVT();
5337 bool isLegal = false;
5338 if (Subtarget->isThumb2())
5339 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5340 Offset, isInc, DAG);
5342 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5343 Offset, isInc, DAG);
5347 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5351 /// getPostIndexedAddressParts - returns true by value, base pointer and
5352 /// offset pointer and addressing mode by reference if this node can be
5353 /// combined with a load / store to form a post-indexed load / store.
5354 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5357 ISD::MemIndexedMode &AM,
5358 SelectionDAG &DAG) const {
5359 if (Subtarget->isThumb1Only())
5364 bool isSEXTLoad = false;
5365 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5366 VT = LD->getMemoryVT();
5367 Ptr = LD->getBasePtr();
5368 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5369 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5370 VT = ST->getMemoryVT();
5371 Ptr = ST->getBasePtr();
5376 bool isLegal = false;
5377 if (Subtarget->isThumb2())
5378 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5381 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5387 // Swap base ptr and offset to catch more post-index load / store when
5388 // it's legal. In Thumb2 mode, offset must be an immediate.
5389 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5390 !Subtarget->isThumb2())
5391 std::swap(Base, Offset);
5393 // Post-indexed load / store update the base pointer.
5398 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5402 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5406 const SelectionDAG &DAG,
5407 unsigned Depth) const {
5408 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5409 switch (Op.getOpcode()) {
5411 case ARMISD::CMOV: {
5412 // Bits are known zero/one if known on the LHS and RHS.
5413 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5414 if (KnownZero == 0 && KnownOne == 0) return;
5416 APInt KnownZeroRHS, KnownOneRHS;
5417 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5418 KnownZeroRHS, KnownOneRHS, Depth+1);
5419 KnownZero &= KnownZeroRHS;
5420 KnownOne &= KnownOneRHS;
5426 //===----------------------------------------------------------------------===//
5427 // ARM Inline Assembly Support
5428 //===----------------------------------------------------------------------===//
5430 /// getConstraintType - Given a constraint letter, return the type of
5431 /// constraint it is for this target.
5432 ARMTargetLowering::ConstraintType
5433 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5434 if (Constraint.size() == 1) {
5435 switch (Constraint[0]) {
5437 case 'l': return C_RegisterClass;
5438 case 'w': return C_RegisterClass;
5441 return TargetLowering::getConstraintType(Constraint);
5444 std::pair<unsigned, const TargetRegisterClass*>
5445 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5447 if (Constraint.size() == 1) {
5448 // GCC ARM Constraint Letters
5449 switch (Constraint[0]) {
5451 if (Subtarget->isThumb())
5452 return std::make_pair(0U, ARM::tGPRRegisterClass);
5454 return std::make_pair(0U, ARM::GPRRegisterClass);
5456 return std::make_pair(0U, ARM::GPRRegisterClass);
5459 return std::make_pair(0U, ARM::SPRRegisterClass);
5460 if (VT.getSizeInBits() == 64)
5461 return std::make_pair(0U, ARM::DPRRegisterClass);
5462 if (VT.getSizeInBits() == 128)
5463 return std::make_pair(0U, ARM::QPRRegisterClass);
5467 if (StringRef("{cc}").equals_lower(Constraint))
5468 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5470 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5473 std::vector<unsigned> ARMTargetLowering::
5474 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5476 if (Constraint.size() != 1)
5477 return std::vector<unsigned>();
5479 switch (Constraint[0]) { // GCC ARM Constraint Letters
5482 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5483 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5486 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5487 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5488 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5489 ARM::R12, ARM::LR, 0);
5492 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5493 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5494 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5495 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5496 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5497 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5498 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5499 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5500 if (VT.getSizeInBits() == 64)
5501 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5502 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5503 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5504 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5505 if (VT.getSizeInBits() == 128)
5506 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5507 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5511 return std::vector<unsigned>();
5514 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5515 /// vector. If it is invalid, don't add anything to Ops.
5516 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5518 std::vector<SDValue>&Ops,
5519 SelectionDAG &DAG) const {
5520 SDValue Result(0, 0);
5522 switch (Constraint) {
5524 case 'I': case 'J': case 'K': case 'L':
5525 case 'M': case 'N': case 'O':
5526 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5530 int64_t CVal64 = C->getSExtValue();
5531 int CVal = (int) CVal64;
5532 // None of these constraints allow values larger than 32 bits. Check
5533 // that the value fits in an int.
5537 switch (Constraint) {
5539 if (Subtarget->isThumb1Only()) {
5540 // This must be a constant between 0 and 255, for ADD
5542 if (CVal >= 0 && CVal <= 255)
5544 } else if (Subtarget->isThumb2()) {
5545 // A constant that can be used as an immediate value in a
5546 // data-processing instruction.
5547 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5550 // A constant that can be used as an immediate value in a
5551 // data-processing instruction.
5552 if (ARM_AM::getSOImmVal(CVal) != -1)
5558 if (Subtarget->isThumb()) { // FIXME thumb2
5559 // This must be a constant between -255 and -1, for negated ADD
5560 // immediates. This can be used in GCC with an "n" modifier that
5561 // prints the negated value, for use with SUB instructions. It is
5562 // not useful otherwise but is implemented for compatibility.
5563 if (CVal >= -255 && CVal <= -1)
5566 // This must be a constant between -4095 and 4095. It is not clear
5567 // what this constraint is intended for. Implemented for
5568 // compatibility with GCC.
5569 if (CVal >= -4095 && CVal <= 4095)
5575 if (Subtarget->isThumb1Only()) {
5576 // A 32-bit value where only one byte has a nonzero value. Exclude
5577 // zero to match GCC. This constraint is used by GCC internally for
5578 // constants that can be loaded with a move/shift combination.
5579 // It is not useful otherwise but is implemented for compatibility.
5580 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5582 } else if (Subtarget->isThumb2()) {
5583 // A constant whose bitwise inverse can be used as an immediate
5584 // value in a data-processing instruction. This can be used in GCC
5585 // with a "B" modifier that prints the inverted value, for use with
5586 // BIC and MVN instructions. It is not useful otherwise but is
5587 // implemented for compatibility.
5588 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5591 // A constant whose bitwise inverse can be used as an immediate
5592 // value in a data-processing instruction. This can be used in GCC
5593 // with a "B" modifier that prints the inverted value, for use with
5594 // BIC and MVN instructions. It is not useful otherwise but is
5595 // implemented for compatibility.
5596 if (ARM_AM::getSOImmVal(~CVal) != -1)
5602 if (Subtarget->isThumb1Only()) {
5603 // This must be a constant between -7 and 7,
5604 // for 3-operand ADD/SUB immediate instructions.
5605 if (CVal >= -7 && CVal < 7)
5607 } else if (Subtarget->isThumb2()) {
5608 // A constant whose negation can be used as an immediate value in a
5609 // data-processing instruction. This can be used in GCC with an "n"
5610 // modifier that prints the negated value, for use with SUB
5611 // instructions. It is not useful otherwise but is implemented for
5613 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5616 // A constant whose negation can be used as an immediate value in a
5617 // data-processing instruction. This can be used in GCC with an "n"
5618 // modifier that prints the negated value, for use with SUB
5619 // instructions. It is not useful otherwise but is implemented for
5621 if (ARM_AM::getSOImmVal(-CVal) != -1)
5627 if (Subtarget->isThumb()) { // FIXME thumb2
5628 // This must be a multiple of 4 between 0 and 1020, for
5629 // ADD sp + immediate.
5630 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5633 // A power of two or a constant between 0 and 32. This is used in
5634 // GCC for the shift amount on shifted register operands, but it is
5635 // useful in general for any shift amounts.
5636 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5642 if (Subtarget->isThumb()) { // FIXME thumb2
5643 // This must be a constant between 0 and 31, for shift amounts.
5644 if (CVal >= 0 && CVal <= 31)
5650 if (Subtarget->isThumb()) { // FIXME thumb2
5651 // This must be a multiple of 4 between -508 and 508, for
5652 // ADD/SUB sp = sp + immediate.
5653 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5658 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5662 if (Result.getNode()) {
5663 Ops.push_back(Result);
5666 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5670 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5671 // The ARM target isn't yet aware of offsets.
5675 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5676 APInt Imm = FPImm.bitcastToAPInt();
5677 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5678 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5679 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5681 // We can handle 4 bits of mantissa.
5682 // mantissa = (16+UInt(e:f:g:h))/16.
5683 if (Mantissa & 0x7ffff)
5686 if ((Mantissa & 0xf) != Mantissa)
5689 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5690 if (Exp < -3 || Exp > 4)
5692 Exp = ((Exp+3) & 0x7) ^ 4;
5694 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5697 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5698 APInt Imm = FPImm.bitcastToAPInt();
5699 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5700 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5701 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5703 // We can handle 4 bits of mantissa.
5704 // mantissa = (16+UInt(e:f:g:h))/16.
5705 if (Mantissa & 0xffffffffffffLL)
5708 if ((Mantissa & 0xf) != Mantissa)
5711 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5712 if (Exp < -3 || Exp > 4)
5714 Exp = ((Exp+3) & 0x7) ^ 4;
5716 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5719 bool ARM::isBitFieldInvertedMask(unsigned v) {
5720 if (v == 0xffffffff)
5722 // there can be 1's on either or both "outsides", all the "inside"
5724 unsigned int lsb = 0, msb = 31;
5725 while (v & (1 << msb)) --msb;
5726 while (v & (1 << lsb)) ++lsb;
5727 for (unsigned int i = lsb; i <= msb; ++i) {
5734 /// isFPImmLegal - Returns true if the target can instruction select the
5735 /// specified FP immediate natively. If false, the legalizer will
5736 /// materialize the FP immediate as a load from a constant pool.
5737 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5738 if (!Subtarget->hasVFP3())
5741 return ARM::getVFPf32Imm(Imm) != -1;
5743 return ARM::getVFPf64Imm(Imm) != -1;
5747 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5748 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5749 /// specified in the intrinsic calls.
5750 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5752 unsigned Intrinsic) const {
5753 switch (Intrinsic) {
5754 case Intrinsic::arm_neon_vld1:
5755 case Intrinsic::arm_neon_vld2:
5756 case Intrinsic::arm_neon_vld3:
5757 case Intrinsic::arm_neon_vld4:
5758 case Intrinsic::arm_neon_vld2lane:
5759 case Intrinsic::arm_neon_vld3lane:
5760 case Intrinsic::arm_neon_vld4lane: {
5761 Info.opc = ISD::INTRINSIC_W_CHAIN;
5762 // Conservatively set memVT to the entire set of vectors loaded.
5763 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
5764 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5765 Info.ptrVal = I.getArgOperand(0);
5767 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5768 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5769 Info.vol = false; // volatile loads with NEON intrinsics not supported
5770 Info.readMem = true;
5771 Info.writeMem = false;
5774 case Intrinsic::arm_neon_vst1:
5775 case Intrinsic::arm_neon_vst2:
5776 case Intrinsic::arm_neon_vst3:
5777 case Intrinsic::arm_neon_vst4:
5778 case Intrinsic::arm_neon_vst2lane:
5779 case Intrinsic::arm_neon_vst3lane:
5780 case Intrinsic::arm_neon_vst4lane: {
5781 Info.opc = ISD::INTRINSIC_VOID;
5782 // Conservatively set memVT to the entire set of vectors stored.
5783 unsigned NumElts = 0;
5784 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5785 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
5786 if (!ArgTy->isVectorTy())
5788 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
5790 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5791 Info.ptrVal = I.getArgOperand(0);
5793 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5794 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5795 Info.vol = false; // volatile stores with NEON intrinsics not supported
5796 Info.readMem = false;
5797 Info.writeMem = true;