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 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
88 if (ElemTy != MVT::i32) {
89 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
90 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
91 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
94 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
95 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
97 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
98 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
99 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
100 if (VT.isInteger()) {
101 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
102 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
103 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
104 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
105 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
106 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
107 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
108 setTruncStoreAction(VT.getSimpleVT(),
109 (MVT::SimpleValueType)InnerVT, Expand);
111 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
113 // Promote all bit-wise operations.
114 if (VT.isInteger() && VT != PromotedBitwiseVT) {
115 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
116 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
117 PromotedBitwiseVT.getSimpleVT());
118 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
119 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
120 PromotedBitwiseVT.getSimpleVT());
121 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
122 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
123 PromotedBitwiseVT.getSimpleVT());
126 // Neon does not support vector divide/remainder operations.
127 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
128 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
129 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
130 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
131 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
132 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
135 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
136 addRegisterClass(VT, ARM::DPRRegisterClass);
137 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
140 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
141 addRegisterClass(VT, ARM::QPRRegisterClass);
142 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
145 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
146 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
147 return new TargetLoweringObjectFileMachO();
149 return new ARMElfTargetObjectFile();
152 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
153 : TargetLowering(TM, createTLOF(TM)) {
154 Subtarget = &TM.getSubtarget<ARMSubtarget>();
155 RegInfo = TM.getRegisterInfo();
156 Itins = TM.getInstrItineraryData();
158 if (Subtarget->isTargetDarwin()) {
159 // Uses VFP for Thumb libfuncs if available.
160 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
161 // Single-precision floating-point arithmetic.
162 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
163 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
164 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
165 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
167 // Double-precision floating-point arithmetic.
168 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
169 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
170 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
171 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
173 // Single-precision comparisons.
174 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
175 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
176 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
177 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
178 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
179 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
180 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
181 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
183 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
192 // Double-precision comparisons.
193 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
194 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
195 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
196 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
197 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
198 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
199 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
200 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
202 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
211 // Floating-point to integer conversions.
212 // i64 conversions are done via library routines even when generating VFP
213 // instructions, so use the same ones.
214 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
215 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
216 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
217 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
219 // Conversions between floating types.
220 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
221 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
223 // Integer to floating-point conversions.
224 // i64 conversions are done via library routines even when generating VFP
225 // instructions, so use the same ones.
226 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
227 // e.g., __floatunsidf vs. __floatunssidfvfp.
228 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
229 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
230 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
231 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
235 // These libcalls are not available in 32-bit.
236 setLibcallName(RTLIB::SHL_I128, 0);
237 setLibcallName(RTLIB::SRL_I128, 0);
238 setLibcallName(RTLIB::SRA_I128, 0);
240 if (Subtarget->isAAPCS_ABI()) {
241 // Double-precision floating-point arithmetic helper functions
242 // RTABI chapter 4.1.2, Table 2
243 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
244 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
245 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
246 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
247 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
248 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
249 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
250 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
252 // Double-precision floating-point comparison helper functions
253 // RTABI chapter 4.1.2, Table 3
254 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
255 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
256 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
257 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
258 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
259 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
260 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
261 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
262 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
263 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
264 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
265 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
266 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
267 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
268 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
269 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
270 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
275 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
276 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
277 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
279 // Single-precision floating-point arithmetic helper functions
280 // RTABI chapter 4.1.2, Table 4
281 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
282 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
283 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
284 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
285 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
286 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
287 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
288 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
290 // Single-precision floating-point comparison helper functions
291 // RTABI chapter 4.1.2, Table 5
292 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
293 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
294 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
295 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
296 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
297 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
298 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
299 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
300 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
301 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
302 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
303 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
304 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
305 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
306 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
307 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
308 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
313 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
314 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
315 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
317 // Floating-point to integer conversions.
318 // RTABI chapter 4.1.2, Table 6
319 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
320 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
321 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
322 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
323 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
324 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
325 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
326 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
327 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
336 // Conversions between floating types.
337 // RTABI chapter 4.1.2, Table 7
338 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
339 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
340 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
341 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
343 // Integer to floating-point conversions.
344 // RTABI chapter 4.1.2, Table 8
345 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
346 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
347 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
348 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
349 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
350 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
351 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
352 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
353 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
362 // Long long helper functions
363 // RTABI chapter 4.2, Table 9
364 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
365 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
366 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
367 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
368 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
369 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
370 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
371 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
377 // Integer division functions
378 // RTABI chapter 4.3.1
379 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
380 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
381 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
382 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
383 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
384 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
385 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
386 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
393 if (Subtarget->isThumb1Only())
394 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
396 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
397 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
398 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
399 if (!Subtarget->isFPOnlySP())
400 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
402 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
405 if (Subtarget->hasNEON()) {
406 addDRTypeForNEON(MVT::v2f32);
407 addDRTypeForNEON(MVT::v8i8);
408 addDRTypeForNEON(MVT::v4i16);
409 addDRTypeForNEON(MVT::v2i32);
410 addDRTypeForNEON(MVT::v1i64);
412 addQRTypeForNEON(MVT::v4f32);
413 addQRTypeForNEON(MVT::v2f64);
414 addQRTypeForNEON(MVT::v16i8);
415 addQRTypeForNEON(MVT::v8i16);
416 addQRTypeForNEON(MVT::v4i32);
417 addQRTypeForNEON(MVT::v2i64);
419 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
420 // neither Neon nor VFP support any arithmetic operations on it.
421 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
422 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
423 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
424 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
425 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
426 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
427 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
428 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
429 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
430 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
431 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
432 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
433 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
434 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
435 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
436 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
437 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
438 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
439 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
440 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
441 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
442 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
443 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
444 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
446 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
448 // Neon does not support some operations on v1i64 and v2i64 types.
449 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
450 // Custom handling for some quad-vector types to detect VMULL.
451 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
452 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
453 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
454 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
455 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
457 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
458 setTargetDAGCombine(ISD::SHL);
459 setTargetDAGCombine(ISD::SRL);
460 setTargetDAGCombine(ISD::SRA);
461 setTargetDAGCombine(ISD::SIGN_EXTEND);
462 setTargetDAGCombine(ISD::ZERO_EXTEND);
463 setTargetDAGCombine(ISD::ANY_EXTEND);
464 setTargetDAGCombine(ISD::SELECT_CC);
465 setTargetDAGCombine(ISD::BUILD_VECTOR);
466 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
469 computeRegisterProperties();
471 // ARM does not have f32 extending load.
472 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
474 // ARM does not have i1 sign extending load.
475 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
477 // ARM supports all 4 flavors of integer indexed load / store.
478 if (!Subtarget->isThumb1Only()) {
479 for (unsigned im = (unsigned)ISD::PRE_INC;
480 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
481 setIndexedLoadAction(im, MVT::i1, Legal);
482 setIndexedLoadAction(im, MVT::i8, Legal);
483 setIndexedLoadAction(im, MVT::i16, Legal);
484 setIndexedLoadAction(im, MVT::i32, Legal);
485 setIndexedStoreAction(im, MVT::i1, Legal);
486 setIndexedStoreAction(im, MVT::i8, Legal);
487 setIndexedStoreAction(im, MVT::i16, Legal);
488 setIndexedStoreAction(im, MVT::i32, Legal);
492 // i64 operation support.
493 if (Subtarget->isThumb1Only()) {
494 setOperationAction(ISD::MUL, MVT::i64, Expand);
495 setOperationAction(ISD::MULHU, MVT::i32, Expand);
496 setOperationAction(ISD::MULHS, MVT::i32, Expand);
497 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
498 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
500 setOperationAction(ISD::MUL, MVT::i64, Expand);
501 setOperationAction(ISD::MULHU, MVT::i32, Expand);
502 if (!Subtarget->hasV6Ops())
503 setOperationAction(ISD::MULHS, MVT::i32, Expand);
505 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
506 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
507 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
508 setOperationAction(ISD::SRL, MVT::i64, Custom);
509 setOperationAction(ISD::SRA, MVT::i64, Custom);
511 // ARM does not have ROTL.
512 setOperationAction(ISD::ROTL, MVT::i32, Expand);
513 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
514 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
515 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
516 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
518 // Only ARMv6 has BSWAP.
519 if (!Subtarget->hasV6Ops())
520 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
522 // These are expanded into libcalls.
523 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
524 // v7M has a hardware divider
525 setOperationAction(ISD::SDIV, MVT::i32, Expand);
526 setOperationAction(ISD::UDIV, MVT::i32, Expand);
528 setOperationAction(ISD::SREM, MVT::i32, Expand);
529 setOperationAction(ISD::UREM, MVT::i32, Expand);
530 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
531 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
533 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
534 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
535 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
536 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
537 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
539 setOperationAction(ISD::TRAP, MVT::Other, Legal);
541 // Use the default implementation.
542 setOperationAction(ISD::VASTART, MVT::Other, Custom);
543 setOperationAction(ISD::VAARG, MVT::Other, Expand);
544 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
545 setOperationAction(ISD::VAEND, MVT::Other, Expand);
546 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
547 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
548 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
549 // FIXME: Shouldn't need this, since no register is used, but the legalizer
550 // doesn't yet know how to not do that for SjLj.
551 setExceptionSelectorRegister(ARM::R0);
552 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
553 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
554 // the default expansion.
555 if (Subtarget->hasDataBarrier() ||
556 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
557 // membarrier needs custom lowering; the rest are legal and handled
559 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
561 // Set them all for expansion, which will force libcalls.
562 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
563 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
564 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
565 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
566 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
567 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
568 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
569 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
570 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
571 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
572 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
573 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
584 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
585 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
586 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
587 // Since the libcalls include locking, fold in the fences
588 setShouldFoldAtomicFences(true);
590 // 64-bit versions are always libcalls (for now)
591 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
592 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
593 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
595 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
596 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
600 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
602 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
603 if (!Subtarget->hasV6Ops()) {
604 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
605 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
607 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
609 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
610 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
611 // iff target supports vfp2.
612 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
613 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
616 // We want to custom lower some of our intrinsics.
617 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
618 if (Subtarget->isTargetDarwin()) {
619 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
620 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
621 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
624 setOperationAction(ISD::SETCC, MVT::i32, Expand);
625 setOperationAction(ISD::SETCC, MVT::f32, Expand);
626 setOperationAction(ISD::SETCC, MVT::f64, Expand);
627 setOperationAction(ISD::SELECT, MVT::i32, Custom);
628 setOperationAction(ISD::SELECT, MVT::f32, Custom);
629 setOperationAction(ISD::SELECT, MVT::f64, Custom);
630 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
631 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
632 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
634 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
635 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
636 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
637 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
638 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
640 // We don't support sin/cos/fmod/copysign/pow
641 setOperationAction(ISD::FSIN, MVT::f64, Expand);
642 setOperationAction(ISD::FSIN, MVT::f32, Expand);
643 setOperationAction(ISD::FCOS, MVT::f32, Expand);
644 setOperationAction(ISD::FCOS, MVT::f64, Expand);
645 setOperationAction(ISD::FREM, MVT::f64, Expand);
646 setOperationAction(ISD::FREM, MVT::f32, Expand);
647 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
648 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
649 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
651 setOperationAction(ISD::FPOW, MVT::f64, Expand);
652 setOperationAction(ISD::FPOW, MVT::f32, Expand);
654 // Various VFP goodness
655 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
656 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
657 if (Subtarget->hasVFP2()) {
658 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
659 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
660 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
661 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
663 // Special handling for half-precision FP.
664 if (!Subtarget->hasFP16()) {
665 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
666 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
670 // We have target-specific dag combine patterns for the following nodes:
671 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
672 setTargetDAGCombine(ISD::ADD);
673 setTargetDAGCombine(ISD::SUB);
674 setTargetDAGCombine(ISD::MUL);
676 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
677 setTargetDAGCombine(ISD::OR);
678 if (Subtarget->hasNEON())
679 setTargetDAGCombine(ISD::AND);
681 setStackPointerRegisterToSaveRestore(ARM::SP);
683 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
684 setSchedulingPreference(Sched::RegPressure);
686 setSchedulingPreference(Sched::Hybrid);
688 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
690 // On ARM arguments smaller than 4 bytes are extended, so all arguments
691 // are at least 4 bytes aligned.
692 setMinStackArgumentAlignment(4);
694 benefitFromCodePlacementOpt = true;
697 std::pair<const TargetRegisterClass*, uint8_t>
698 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
699 const TargetRegisterClass *RRC = 0;
701 switch (VT.getSimpleVT().SimpleTy) {
703 return TargetLowering::findRepresentativeClass(VT);
704 // Use DPR as representative register class for all floating point
705 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
706 // the cost is 1 for both f32 and f64.
707 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
708 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
709 RRC = ARM::DPRRegisterClass;
711 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
712 case MVT::v4f32: case MVT::v2f64:
713 RRC = ARM::DPRRegisterClass;
717 RRC = ARM::DPRRegisterClass;
721 RRC = ARM::DPRRegisterClass;
725 return std::make_pair(RRC, Cost);
728 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
731 case ARMISD::Wrapper: return "ARMISD::Wrapper";
732 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
733 case ARMISD::CALL: return "ARMISD::CALL";
734 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
735 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
736 case ARMISD::tCALL: return "ARMISD::tCALL";
737 case ARMISD::BRCOND: return "ARMISD::BRCOND";
738 case ARMISD::BR_JT: return "ARMISD::BR_JT";
739 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
740 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
741 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
742 case ARMISD::CMP: return "ARMISD::CMP";
743 case ARMISD::CMPZ: return "ARMISD::CMPZ";
744 case ARMISD::CMPFP: return "ARMISD::CMPFP";
745 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
746 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
747 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
748 case ARMISD::CMOV: return "ARMISD::CMOV";
749 case ARMISD::CNEG: return "ARMISD::CNEG";
751 case ARMISD::RBIT: return "ARMISD::RBIT";
753 case ARMISD::FTOSI: return "ARMISD::FTOSI";
754 case ARMISD::FTOUI: return "ARMISD::FTOUI";
755 case ARMISD::SITOF: return "ARMISD::SITOF";
756 case ARMISD::UITOF: return "ARMISD::UITOF";
758 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
759 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
760 case ARMISD::RRX: return "ARMISD::RRX";
762 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
763 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
765 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
766 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
767 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
769 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
771 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
773 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
775 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
776 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
778 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
780 case ARMISD::VCEQ: return "ARMISD::VCEQ";
781 case ARMISD::VCGE: return "ARMISD::VCGE";
782 case ARMISD::VCGEU: return "ARMISD::VCGEU";
783 case ARMISD::VCGT: return "ARMISD::VCGT";
784 case ARMISD::VCGTU: return "ARMISD::VCGTU";
785 case ARMISD::VTST: return "ARMISD::VTST";
787 case ARMISD::VSHL: return "ARMISD::VSHL";
788 case ARMISD::VSHRs: return "ARMISD::VSHRs";
789 case ARMISD::VSHRu: return "ARMISD::VSHRu";
790 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
791 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
792 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
793 case ARMISD::VSHRN: return "ARMISD::VSHRN";
794 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
795 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
796 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
797 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
798 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
799 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
800 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
801 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
802 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
803 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
804 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
805 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
806 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
807 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
808 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
809 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
810 case ARMISD::VDUP: return "ARMISD::VDUP";
811 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
812 case ARMISD::VEXT: return "ARMISD::VEXT";
813 case ARMISD::VREV64: return "ARMISD::VREV64";
814 case ARMISD::VREV32: return "ARMISD::VREV32";
815 case ARMISD::VREV16: return "ARMISD::VREV16";
816 case ARMISD::VZIP: return "ARMISD::VZIP";
817 case ARMISD::VUZP: return "ARMISD::VUZP";
818 case ARMISD::VTRN: return "ARMISD::VTRN";
819 case ARMISD::VMULLs: return "ARMISD::VMULLs";
820 case ARMISD::VMULLu: return "ARMISD::VMULLu";
821 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
822 case ARMISD::FMAX: return "ARMISD::FMAX";
823 case ARMISD::FMIN: return "ARMISD::FMIN";
824 case ARMISD::BFI: return "ARMISD::BFI";
825 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
826 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
827 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
828 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
829 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
833 /// getRegClassFor - Return the register class that should be used for the
834 /// specified value type.
835 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
836 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
837 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
838 // load / store 4 to 8 consecutive D registers.
839 if (Subtarget->hasNEON()) {
840 if (VT == MVT::v4i64)
841 return ARM::QQPRRegisterClass;
842 else if (VT == MVT::v8i64)
843 return ARM::QQQQPRRegisterClass;
845 return TargetLowering::getRegClassFor(VT);
848 // Create a fast isel object.
850 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
851 return ARM::createFastISel(funcInfo);
854 /// getFunctionAlignment - Return the Log2 alignment of this function.
855 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
856 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
859 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
860 /// be used for loads / stores from the global.
861 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
862 return (Subtarget->isThumb1Only() ? 127 : 4095);
865 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
866 unsigned NumVals = N->getNumValues();
868 return Sched::RegPressure;
870 for (unsigned i = 0; i != NumVals; ++i) {
871 EVT VT = N->getValueType(i);
872 if (VT == MVT::Flag || VT == MVT::Other)
874 if (VT.isFloatingPoint() || VT.isVector())
875 return Sched::Latency;
878 if (!N->isMachineOpcode())
879 return Sched::RegPressure;
881 // Load are scheduled for latency even if there instruction itinerary
883 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
884 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
886 if (TID.getNumDefs() == 0)
887 return Sched::RegPressure;
888 if (!Itins->isEmpty() &&
889 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
890 return Sched::Latency;
892 return Sched::RegPressure;
896 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
897 MachineFunction &MF) const {
898 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo();
900 switch (RC->getID()) {
903 case ARM::tGPRRegClassID:
904 return TFI->hasFP(MF) ? 4 : 5;
905 case ARM::GPRRegClassID: {
906 unsigned FP = TFI->hasFP(MF) ? 1 : 0;
907 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
909 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
910 case ARM::DPRRegClassID:
915 //===----------------------------------------------------------------------===//
917 //===----------------------------------------------------------------------===//
919 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
920 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
922 default: llvm_unreachable("Unknown condition code!");
923 case ISD::SETNE: return ARMCC::NE;
924 case ISD::SETEQ: return ARMCC::EQ;
925 case ISD::SETGT: return ARMCC::GT;
926 case ISD::SETGE: return ARMCC::GE;
927 case ISD::SETLT: return ARMCC::LT;
928 case ISD::SETLE: return ARMCC::LE;
929 case ISD::SETUGT: return ARMCC::HI;
930 case ISD::SETUGE: return ARMCC::HS;
931 case ISD::SETULT: return ARMCC::LO;
932 case ISD::SETULE: return ARMCC::LS;
936 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
937 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
938 ARMCC::CondCodes &CondCode2) {
939 CondCode2 = ARMCC::AL;
941 default: llvm_unreachable("Unknown FP condition!");
943 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
945 case ISD::SETOGT: CondCode = ARMCC::GT; break;
947 case ISD::SETOGE: CondCode = ARMCC::GE; break;
948 case ISD::SETOLT: CondCode = ARMCC::MI; break;
949 case ISD::SETOLE: CondCode = ARMCC::LS; break;
950 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
951 case ISD::SETO: CondCode = ARMCC::VC; break;
952 case ISD::SETUO: CondCode = ARMCC::VS; break;
953 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
954 case ISD::SETUGT: CondCode = ARMCC::HI; break;
955 case ISD::SETUGE: CondCode = ARMCC::PL; break;
957 case ISD::SETULT: CondCode = ARMCC::LT; break;
959 case ISD::SETULE: CondCode = ARMCC::LE; break;
961 case ISD::SETUNE: CondCode = ARMCC::NE; break;
965 //===----------------------------------------------------------------------===//
966 // Calling Convention Implementation
967 //===----------------------------------------------------------------------===//
969 #include "ARMGenCallingConv.inc"
971 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
972 /// given CallingConvention value.
973 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
975 bool isVarArg) const {
978 llvm_unreachable("Unsupported calling convention");
979 case CallingConv::Fast:
980 if (Subtarget->hasVFP2() && !isVarArg) {
981 if (!Subtarget->isAAPCS_ABI())
982 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
983 // For AAPCS ABI targets, just use VFP variant of the calling convention.
984 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
987 case CallingConv::C: {
988 // Use target triple & subtarget features to do actual dispatch.
989 if (!Subtarget->isAAPCS_ABI())
990 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
991 else if (Subtarget->hasVFP2() &&
992 FloatABIType == FloatABI::Hard && !isVarArg)
993 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
994 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
996 case CallingConv::ARM_AAPCS_VFP:
997 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
998 case CallingConv::ARM_AAPCS:
999 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1000 case CallingConv::ARM_APCS:
1001 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1005 /// LowerCallResult - Lower the result values of a call into the
1006 /// appropriate copies out of appropriate physical registers.
1008 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1009 CallingConv::ID CallConv, bool isVarArg,
1010 const SmallVectorImpl<ISD::InputArg> &Ins,
1011 DebugLoc dl, SelectionDAG &DAG,
1012 SmallVectorImpl<SDValue> &InVals) const {
1014 // Assign locations to each value returned by this call.
1015 SmallVector<CCValAssign, 16> RVLocs;
1016 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1017 RVLocs, *DAG.getContext());
1018 CCInfo.AnalyzeCallResult(Ins,
1019 CCAssignFnForNode(CallConv, /* Return*/ true,
1022 // Copy all of the result registers out of their specified physreg.
1023 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1024 CCValAssign VA = RVLocs[i];
1027 if (VA.needsCustom()) {
1028 // Handle f64 or half of a v2f64.
1029 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1031 Chain = Lo.getValue(1);
1032 InFlag = Lo.getValue(2);
1033 VA = RVLocs[++i]; // skip ahead to next loc
1034 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1036 Chain = Hi.getValue(1);
1037 InFlag = Hi.getValue(2);
1038 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1040 if (VA.getLocVT() == MVT::v2f64) {
1041 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1042 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1043 DAG.getConstant(0, MVT::i32));
1045 VA = RVLocs[++i]; // skip ahead to next loc
1046 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1047 Chain = Lo.getValue(1);
1048 InFlag = Lo.getValue(2);
1049 VA = RVLocs[++i]; // skip ahead to next loc
1050 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1051 Chain = Hi.getValue(1);
1052 InFlag = Hi.getValue(2);
1053 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1054 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1055 DAG.getConstant(1, MVT::i32));
1058 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1060 Chain = Val.getValue(1);
1061 InFlag = Val.getValue(2);
1064 switch (VA.getLocInfo()) {
1065 default: llvm_unreachable("Unknown loc info!");
1066 case CCValAssign::Full: break;
1067 case CCValAssign::BCvt:
1068 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1072 InVals.push_back(Val);
1078 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1079 /// by "Src" to address "Dst" of size "Size". Alignment information is
1080 /// specified by the specific parameter attribute. The copy will be passed as
1081 /// a byval function parameter.
1082 /// Sometimes what we are copying is the end of a larger object, the part that
1083 /// does not fit in registers.
1085 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1086 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1088 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1089 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1090 /*isVolatile=*/false, /*AlwaysInline=*/false,
1091 MachinePointerInfo(0), MachinePointerInfo(0));
1094 /// LowerMemOpCallTo - Store the argument to the stack.
1096 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1097 SDValue StackPtr, SDValue Arg,
1098 DebugLoc dl, SelectionDAG &DAG,
1099 const CCValAssign &VA,
1100 ISD::ArgFlagsTy Flags) const {
1101 unsigned LocMemOffset = VA.getLocMemOffset();
1102 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1103 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1104 if (Flags.isByVal())
1105 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1107 return DAG.getStore(Chain, dl, Arg, PtrOff,
1108 MachinePointerInfo::getStack(LocMemOffset),
1112 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1113 SDValue Chain, SDValue &Arg,
1114 RegsToPassVector &RegsToPass,
1115 CCValAssign &VA, CCValAssign &NextVA,
1117 SmallVector<SDValue, 8> &MemOpChains,
1118 ISD::ArgFlagsTy Flags) const {
1120 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1121 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1122 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1124 if (NextVA.isRegLoc())
1125 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1127 assert(NextVA.isMemLoc());
1128 if (StackPtr.getNode() == 0)
1129 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1131 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1137 /// LowerCall - Lowering a call into a callseq_start <-
1138 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1141 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1142 CallingConv::ID CallConv, bool isVarArg,
1144 const SmallVectorImpl<ISD::OutputArg> &Outs,
1145 const SmallVectorImpl<SDValue> &OutVals,
1146 const SmallVectorImpl<ISD::InputArg> &Ins,
1147 DebugLoc dl, SelectionDAG &DAG,
1148 SmallVectorImpl<SDValue> &InVals) const {
1149 MachineFunction &MF = DAG.getMachineFunction();
1150 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1151 bool IsSibCall = false;
1152 // Temporarily disable tail calls so things don't break.
1153 if (!EnableARMTailCalls)
1156 // Check if it's really possible to do a tail call.
1157 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1158 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1159 Outs, OutVals, Ins, DAG);
1160 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1161 // detected sibcalls.
1168 // Analyze operands of the call, assigning locations to each operand.
1169 SmallVector<CCValAssign, 16> ArgLocs;
1170 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1172 CCInfo.AnalyzeCallOperands(Outs,
1173 CCAssignFnForNode(CallConv, /* Return*/ false,
1176 // Get a count of how many bytes are to be pushed on the stack.
1177 unsigned NumBytes = CCInfo.getNextStackOffset();
1179 // For tail calls, memory operands are available in our caller's stack.
1183 // Adjust the stack pointer for the new arguments...
1184 // These operations are automatically eliminated by the prolog/epilog pass
1186 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1188 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1190 RegsToPassVector RegsToPass;
1191 SmallVector<SDValue, 8> MemOpChains;
1193 // Walk the register/memloc assignments, inserting copies/loads. In the case
1194 // of tail call optimization, arguments are handled later.
1195 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1197 ++i, ++realArgIdx) {
1198 CCValAssign &VA = ArgLocs[i];
1199 SDValue Arg = OutVals[realArgIdx];
1200 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1202 // Promote the value if needed.
1203 switch (VA.getLocInfo()) {
1204 default: llvm_unreachable("Unknown loc info!");
1205 case CCValAssign::Full: break;
1206 case CCValAssign::SExt:
1207 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1209 case CCValAssign::ZExt:
1210 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1212 case CCValAssign::AExt:
1213 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1215 case CCValAssign::BCvt:
1216 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1220 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1221 if (VA.needsCustom()) {
1222 if (VA.getLocVT() == MVT::v2f64) {
1223 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1224 DAG.getConstant(0, MVT::i32));
1225 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1226 DAG.getConstant(1, MVT::i32));
1228 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1229 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1231 VA = ArgLocs[++i]; // skip ahead to next loc
1232 if (VA.isRegLoc()) {
1233 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1234 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1236 assert(VA.isMemLoc());
1238 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1239 dl, DAG, VA, Flags));
1242 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1243 StackPtr, MemOpChains, Flags);
1245 } else if (VA.isRegLoc()) {
1246 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1247 } else if (!IsSibCall) {
1248 assert(VA.isMemLoc());
1250 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1251 dl, DAG, VA, Flags));
1255 if (!MemOpChains.empty())
1256 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1257 &MemOpChains[0], MemOpChains.size());
1259 // Build a sequence of copy-to-reg nodes chained together with token chain
1260 // and flag operands which copy the outgoing args into the appropriate regs.
1262 // Tail call byval lowering might overwrite argument registers so in case of
1263 // tail call optimization the copies to registers are lowered later.
1265 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1266 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1267 RegsToPass[i].second, InFlag);
1268 InFlag = Chain.getValue(1);
1271 // For tail calls lower the arguments to the 'real' stack slot.
1273 // Force all the incoming stack arguments to be loaded from the stack
1274 // before any new outgoing arguments are stored to the stack, because the
1275 // outgoing stack slots may alias the incoming argument stack slots, and
1276 // the alias isn't otherwise explicit. This is slightly more conservative
1277 // than necessary, because it means that each store effectively depends
1278 // on every argument instead of just those arguments it would clobber.
1280 // Do not flag preceeding copytoreg stuff together with the following stuff.
1282 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1283 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1284 RegsToPass[i].second, InFlag);
1285 InFlag = Chain.getValue(1);
1290 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1291 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1292 // node so that legalize doesn't hack it.
1293 bool isDirect = false;
1294 bool isARMFunc = false;
1295 bool isLocalARMFunc = false;
1296 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1298 if (EnableARMLongCalls) {
1299 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1300 && "long-calls with non-static relocation model!");
1301 // Handle a global address or an external symbol. If it's not one of
1302 // those, the target's already in a register, so we don't need to do
1304 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1305 const GlobalValue *GV = G->getGlobal();
1306 // Create a constant pool entry for the callee address
1307 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1308 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1311 // Get the address of the callee into a register
1312 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1313 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1314 Callee = DAG.getLoad(getPointerTy(), dl,
1315 DAG.getEntryNode(), CPAddr,
1316 MachinePointerInfo::getConstantPool(),
1318 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1319 const char *Sym = S->getSymbol();
1321 // Create a constant pool entry for the callee address
1322 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1323 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1324 Sym, ARMPCLabelIndex, 0);
1325 // Get the address of the callee into a register
1326 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1327 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1328 Callee = DAG.getLoad(getPointerTy(), dl,
1329 DAG.getEntryNode(), CPAddr,
1330 MachinePointerInfo::getConstantPool(),
1333 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1334 const GlobalValue *GV = G->getGlobal();
1336 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1337 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1338 getTargetMachine().getRelocationModel() != Reloc::Static;
1339 isARMFunc = !Subtarget->isThumb() || isStub;
1340 // ARM call to a local ARM function is predicable.
1341 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1342 // tBX takes a register source operand.
1343 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1344 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1345 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1348 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1349 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1350 Callee = DAG.getLoad(getPointerTy(), dl,
1351 DAG.getEntryNode(), CPAddr,
1352 MachinePointerInfo::getConstantPool(),
1354 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1355 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1356 getPointerTy(), Callee, PICLabel);
1358 // On ELF targets for PIC code, direct calls should go through the PLT
1359 unsigned OpFlags = 0;
1360 if (Subtarget->isTargetELF() &&
1361 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1362 OpFlags = ARMII::MO_PLT;
1363 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1365 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1367 bool isStub = Subtarget->isTargetDarwin() &&
1368 getTargetMachine().getRelocationModel() != Reloc::Static;
1369 isARMFunc = !Subtarget->isThumb() || isStub;
1370 // tBX takes a register source operand.
1371 const char *Sym = S->getSymbol();
1372 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1373 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1374 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1375 Sym, ARMPCLabelIndex, 4);
1376 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1377 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1378 Callee = DAG.getLoad(getPointerTy(), dl,
1379 DAG.getEntryNode(), CPAddr,
1380 MachinePointerInfo::getConstantPool(),
1382 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1383 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1384 getPointerTy(), Callee, PICLabel);
1386 unsigned OpFlags = 0;
1387 // On ELF targets for PIC code, direct calls should go through the PLT
1388 if (Subtarget->isTargetELF() &&
1389 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1390 OpFlags = ARMII::MO_PLT;
1391 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1395 // FIXME: handle tail calls differently.
1397 if (Subtarget->isThumb()) {
1398 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1399 CallOpc = ARMISD::CALL_NOLINK;
1401 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1403 CallOpc = (isDirect || Subtarget->hasV5TOps())
1404 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1405 : ARMISD::CALL_NOLINK;
1408 std::vector<SDValue> Ops;
1409 Ops.push_back(Chain);
1410 Ops.push_back(Callee);
1412 // Add argument registers to the end of the list so that they are known live
1414 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1415 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1416 RegsToPass[i].second.getValueType()));
1418 if (InFlag.getNode())
1419 Ops.push_back(InFlag);
1421 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1423 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1425 // Returns a chain and a flag for retval copy to use.
1426 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1427 InFlag = Chain.getValue(1);
1429 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1430 DAG.getIntPtrConstant(0, true), InFlag);
1432 InFlag = Chain.getValue(1);
1434 // Handle result values, copying them out of physregs into vregs that we
1436 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1440 /// MatchingStackOffset - Return true if the given stack call argument is
1441 /// already available in the same position (relatively) of the caller's
1442 /// incoming argument stack.
1444 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1445 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1446 const ARMInstrInfo *TII) {
1447 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1449 if (Arg.getOpcode() == ISD::CopyFromReg) {
1450 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1451 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1453 MachineInstr *Def = MRI->getVRegDef(VR);
1456 if (!Flags.isByVal()) {
1457 if (!TII->isLoadFromStackSlot(Def, FI))
1462 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1463 if (Flags.isByVal())
1464 // ByVal argument is passed in as a pointer but it's now being
1465 // dereferenced. e.g.
1466 // define @foo(%struct.X* %A) {
1467 // tail call @bar(%struct.X* byval %A)
1470 SDValue Ptr = Ld->getBasePtr();
1471 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1474 FI = FINode->getIndex();
1478 assert(FI != INT_MAX);
1479 if (!MFI->isFixedObjectIndex(FI))
1481 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1484 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1485 /// for tail call optimization. Targets which want to do tail call
1486 /// optimization should implement this function.
1488 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1489 CallingConv::ID CalleeCC,
1491 bool isCalleeStructRet,
1492 bool isCallerStructRet,
1493 const SmallVectorImpl<ISD::OutputArg> &Outs,
1494 const SmallVectorImpl<SDValue> &OutVals,
1495 const SmallVectorImpl<ISD::InputArg> &Ins,
1496 SelectionDAG& DAG) const {
1497 const Function *CallerF = DAG.getMachineFunction().getFunction();
1498 CallingConv::ID CallerCC = CallerF->getCallingConv();
1499 bool CCMatch = CallerCC == CalleeCC;
1501 // Look for obvious safe cases to perform tail call optimization that do not
1502 // require ABI changes. This is what gcc calls sibcall.
1504 // Do not sibcall optimize vararg calls unless the call site is not passing
1506 if (isVarArg && !Outs.empty())
1509 // Also avoid sibcall optimization if either caller or callee uses struct
1510 // return semantics.
1511 if (isCalleeStructRet || isCallerStructRet)
1514 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1515 // emitEpilogue is not ready for them.
1516 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1517 // LR. This means if we need to reload LR, it takes an extra instructions,
1518 // which outweighs the value of the tail call; but here we don't know yet
1519 // whether LR is going to be used. Probably the right approach is to
1520 // generate the tail call here and turn it back into CALL/RET in
1521 // emitEpilogue if LR is used.
1522 if (Subtarget->isThumb1Only())
1525 // For the moment, we can only do this to functions defined in this
1526 // compilation, or to indirect calls. A Thumb B to an ARM function,
1527 // or vice versa, is not easily fixed up in the linker unlike BL.
1528 // (We could do this by loading the address of the callee into a register;
1529 // that is an extra instruction over the direct call and burns a register
1530 // as well, so is not likely to be a win.)
1532 // It might be safe to remove this restriction on non-Darwin.
1534 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1535 // but we need to make sure there are enough registers; the only valid
1536 // registers are the 4 used for parameters. We don't currently do this
1538 if (isa<ExternalSymbolSDNode>(Callee))
1541 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1542 const GlobalValue *GV = G->getGlobal();
1543 if (GV->isDeclaration() || GV->isWeakForLinker())
1547 // If the calling conventions do not match, then we'd better make sure the
1548 // results are returned in the same way as what the caller expects.
1550 SmallVector<CCValAssign, 16> RVLocs1;
1551 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1552 RVLocs1, *DAG.getContext());
1553 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1555 SmallVector<CCValAssign, 16> RVLocs2;
1556 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1557 RVLocs2, *DAG.getContext());
1558 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1560 if (RVLocs1.size() != RVLocs2.size())
1562 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1563 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1565 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1567 if (RVLocs1[i].isRegLoc()) {
1568 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1571 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1577 // If the callee takes no arguments then go on to check the results of the
1579 if (!Outs.empty()) {
1580 // Check if stack adjustment is needed. For now, do not do this if any
1581 // argument is passed on the stack.
1582 SmallVector<CCValAssign, 16> ArgLocs;
1583 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1584 ArgLocs, *DAG.getContext());
1585 CCInfo.AnalyzeCallOperands(Outs,
1586 CCAssignFnForNode(CalleeCC, false, isVarArg));
1587 if (CCInfo.getNextStackOffset()) {
1588 MachineFunction &MF = DAG.getMachineFunction();
1590 // Check if the arguments are already laid out in the right way as
1591 // the caller's fixed stack objects.
1592 MachineFrameInfo *MFI = MF.getFrameInfo();
1593 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1594 const ARMInstrInfo *TII =
1595 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1596 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1598 ++i, ++realArgIdx) {
1599 CCValAssign &VA = ArgLocs[i];
1600 EVT RegVT = VA.getLocVT();
1601 SDValue Arg = OutVals[realArgIdx];
1602 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1603 if (VA.getLocInfo() == CCValAssign::Indirect)
1605 if (VA.needsCustom()) {
1606 // f64 and vector types are split into multiple registers or
1607 // register/stack-slot combinations. The types will not match
1608 // the registers; give up on memory f64 refs until we figure
1609 // out what to do about this.
1612 if (!ArgLocs[++i].isRegLoc())
1614 if (RegVT == MVT::v2f64) {
1615 if (!ArgLocs[++i].isRegLoc())
1617 if (!ArgLocs[++i].isRegLoc())
1620 } else if (!VA.isRegLoc()) {
1621 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1633 ARMTargetLowering::LowerReturn(SDValue Chain,
1634 CallingConv::ID CallConv, bool isVarArg,
1635 const SmallVectorImpl<ISD::OutputArg> &Outs,
1636 const SmallVectorImpl<SDValue> &OutVals,
1637 DebugLoc dl, SelectionDAG &DAG) const {
1639 // CCValAssign - represent the assignment of the return value to a location.
1640 SmallVector<CCValAssign, 16> RVLocs;
1642 // CCState - Info about the registers and stack slots.
1643 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1646 // Analyze outgoing return values.
1647 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1650 // If this is the first return lowered for this function, add
1651 // the regs to the liveout set for the function.
1652 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1653 for (unsigned i = 0; i != RVLocs.size(); ++i)
1654 if (RVLocs[i].isRegLoc())
1655 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1660 // Copy the result values into the output registers.
1661 for (unsigned i = 0, realRVLocIdx = 0;
1663 ++i, ++realRVLocIdx) {
1664 CCValAssign &VA = RVLocs[i];
1665 assert(VA.isRegLoc() && "Can only return in registers!");
1667 SDValue Arg = OutVals[realRVLocIdx];
1669 switch (VA.getLocInfo()) {
1670 default: llvm_unreachable("Unknown loc info!");
1671 case CCValAssign::Full: break;
1672 case CCValAssign::BCvt:
1673 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1677 if (VA.needsCustom()) {
1678 if (VA.getLocVT() == MVT::v2f64) {
1679 // Extract the first half and return it in two registers.
1680 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1681 DAG.getConstant(0, MVT::i32));
1682 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1683 DAG.getVTList(MVT::i32, MVT::i32), Half);
1685 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1686 Flag = Chain.getValue(1);
1687 VA = RVLocs[++i]; // skip ahead to next loc
1688 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1689 HalfGPRs.getValue(1), Flag);
1690 Flag = Chain.getValue(1);
1691 VA = RVLocs[++i]; // skip ahead to next loc
1693 // Extract the 2nd half and fall through to handle it as an f64 value.
1694 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1695 DAG.getConstant(1, MVT::i32));
1697 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1699 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1700 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1701 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1702 Flag = Chain.getValue(1);
1703 VA = RVLocs[++i]; // skip ahead to next loc
1704 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1707 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1709 // Guarantee that all emitted copies are
1710 // stuck together, avoiding something bad.
1711 Flag = Chain.getValue(1);
1716 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1718 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1723 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1724 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1725 // one of the above mentioned nodes. It has to be wrapped because otherwise
1726 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1727 // be used to form addressing mode. These wrapped nodes will be selected
1729 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1730 EVT PtrVT = Op.getValueType();
1731 // FIXME there is no actual debug info here
1732 DebugLoc dl = Op.getDebugLoc();
1733 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1735 if (CP->isMachineConstantPoolEntry())
1736 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1737 CP->getAlignment());
1739 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1740 CP->getAlignment());
1741 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1744 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1745 return MachineJumpTableInfo::EK_Inline;
1748 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1749 SelectionDAG &DAG) const {
1750 MachineFunction &MF = DAG.getMachineFunction();
1751 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1752 unsigned ARMPCLabelIndex = 0;
1753 DebugLoc DL = Op.getDebugLoc();
1754 EVT PtrVT = getPointerTy();
1755 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1756 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1758 if (RelocM == Reloc::Static) {
1759 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1761 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1762 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1763 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1764 ARMCP::CPBlockAddress,
1766 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1768 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1769 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1770 MachinePointerInfo::getConstantPool(),
1772 if (RelocM == Reloc::Static)
1774 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1775 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1778 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1780 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1781 SelectionDAG &DAG) const {
1782 DebugLoc dl = GA->getDebugLoc();
1783 EVT PtrVT = getPointerTy();
1784 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1785 MachineFunction &MF = DAG.getMachineFunction();
1786 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1787 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1788 ARMConstantPoolValue *CPV =
1789 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1790 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1791 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1792 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1793 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1794 MachinePointerInfo::getConstantPool(),
1796 SDValue Chain = Argument.getValue(1);
1798 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1799 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1801 // call __tls_get_addr.
1804 Entry.Node = Argument;
1805 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1806 Args.push_back(Entry);
1807 // FIXME: is there useful debug info available here?
1808 std::pair<SDValue, SDValue> CallResult =
1809 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1810 false, false, false, false,
1811 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1812 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1813 return CallResult.first;
1816 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1817 // "local exec" model.
1819 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1820 SelectionDAG &DAG) const {
1821 const GlobalValue *GV = GA->getGlobal();
1822 DebugLoc dl = GA->getDebugLoc();
1824 SDValue Chain = DAG.getEntryNode();
1825 EVT PtrVT = getPointerTy();
1826 // Get the Thread Pointer
1827 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1829 if (GV->isDeclaration()) {
1830 MachineFunction &MF = DAG.getMachineFunction();
1831 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1832 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1833 // Initial exec model.
1834 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1835 ARMConstantPoolValue *CPV =
1836 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1837 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1838 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1839 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1840 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1841 MachinePointerInfo::getConstantPool(),
1843 Chain = Offset.getValue(1);
1845 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1846 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1848 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1849 MachinePointerInfo::getConstantPool(),
1853 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
1854 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1855 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1856 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1857 MachinePointerInfo::getConstantPool(),
1861 // The address of the thread local variable is the add of the thread
1862 // pointer with the offset of the variable.
1863 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1867 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1868 // TODO: implement the "local dynamic" model
1869 assert(Subtarget->isTargetELF() &&
1870 "TLS not implemented for non-ELF targets");
1871 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1872 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1873 // otherwise use the "Local Exec" TLS Model
1874 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1875 return LowerToTLSGeneralDynamicModel(GA, DAG);
1877 return LowerToTLSExecModels(GA, DAG);
1880 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1881 SelectionDAG &DAG) const {
1882 EVT PtrVT = getPointerTy();
1883 DebugLoc dl = Op.getDebugLoc();
1884 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1885 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1886 if (RelocM == Reloc::PIC_) {
1887 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1888 ARMConstantPoolValue *CPV =
1889 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
1890 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1891 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1892 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1894 MachinePointerInfo::getConstantPool(),
1896 SDValue Chain = Result.getValue(1);
1897 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1898 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1900 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1901 MachinePointerInfo::getGOT(), false, false, 0);
1904 // If we have T2 ops, we can materialize the address directly via movt/movw
1905 // pair. This is always cheaper.
1906 if (Subtarget->useMovt()) {
1907 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1908 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1910 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1911 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1912 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1913 MachinePointerInfo::getConstantPool(),
1919 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1920 SelectionDAG &DAG) const {
1921 MachineFunction &MF = DAG.getMachineFunction();
1922 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1923 unsigned ARMPCLabelIndex = 0;
1924 EVT PtrVT = getPointerTy();
1925 DebugLoc dl = Op.getDebugLoc();
1926 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1927 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1929 if (RelocM == Reloc::Static)
1930 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1932 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1933 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1934 ARMConstantPoolValue *CPV =
1935 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1936 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1938 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1940 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1941 MachinePointerInfo::getConstantPool(),
1943 SDValue Chain = Result.getValue(1);
1945 if (RelocM == Reloc::PIC_) {
1946 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1947 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1950 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1951 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1957 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1958 SelectionDAG &DAG) const {
1959 assert(Subtarget->isTargetELF() &&
1960 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1961 MachineFunction &MF = DAG.getMachineFunction();
1962 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1963 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1964 EVT PtrVT = getPointerTy();
1965 DebugLoc dl = Op.getDebugLoc();
1966 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1967 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1968 "_GLOBAL_OFFSET_TABLE_",
1969 ARMPCLabelIndex, PCAdj);
1970 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1971 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1972 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1973 MachinePointerInfo::getConstantPool(),
1975 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1976 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1980 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
1982 DebugLoc dl = Op.getDebugLoc();
1983 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
1984 Op.getOperand(0), Op.getOperand(1));
1988 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1989 DebugLoc dl = Op.getDebugLoc();
1990 SDValue Val = DAG.getConstant(0, MVT::i32);
1991 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1992 Op.getOperand(1), Val);
1996 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1997 DebugLoc dl = Op.getDebugLoc();
1998 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1999 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2003 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2004 const ARMSubtarget *Subtarget) const {
2005 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2006 DebugLoc dl = Op.getDebugLoc();
2008 default: return SDValue(); // Don't custom lower most intrinsics.
2009 case Intrinsic::arm_thread_pointer: {
2010 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2011 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2013 case Intrinsic::eh_sjlj_lsda: {
2014 MachineFunction &MF = DAG.getMachineFunction();
2015 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2016 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
2017 EVT PtrVT = getPointerTy();
2018 DebugLoc dl = Op.getDebugLoc();
2019 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2021 unsigned PCAdj = (RelocM != Reloc::PIC_)
2022 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2023 ARMConstantPoolValue *CPV =
2024 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2025 ARMCP::CPLSDA, PCAdj);
2026 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2027 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2029 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2030 MachinePointerInfo::getConstantPool(),
2033 if (RelocM == Reloc::PIC_) {
2034 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2035 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2042 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2043 const ARMSubtarget *Subtarget) {
2044 DebugLoc dl = Op.getDebugLoc();
2045 if (!Subtarget->hasDataBarrier()) {
2046 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2047 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2049 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2050 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2051 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2052 DAG.getConstant(0, MVT::i32));
2055 SDValue Op5 = Op.getOperand(5);
2056 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2057 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2058 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2059 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2061 ARM_MB::MemBOpt DMBOpt;
2062 if (isDeviceBarrier)
2063 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2065 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2066 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2067 DAG.getConstant(DMBOpt, MVT::i32));
2070 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2071 const ARMSubtarget *Subtarget) {
2072 // ARM pre v5TE and Thumb1 does not have preload instructions.
2073 if (!(Subtarget->isThumb2() ||
2074 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2075 // Just preserve the chain.
2076 return Op.getOperand(0);
2078 DebugLoc dl = Op.getDebugLoc();
2079 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2081 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2082 // ARMv7 with MP extension has PLDW.
2083 return Op.getOperand(0);
2085 if (Subtarget->isThumb())
2087 isRead = ~isRead & 1;
2088 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2090 // Currently there is no intrinsic that matches pli.
2091 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2092 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2093 DAG.getConstant(isData, MVT::i32));
2096 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2097 MachineFunction &MF = DAG.getMachineFunction();
2098 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2100 // vastart just stores the address of the VarArgsFrameIndex slot into the
2101 // memory location argument.
2102 DebugLoc dl = Op.getDebugLoc();
2103 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2104 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2105 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2106 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2107 MachinePointerInfo(SV), false, false, 0);
2111 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2112 SDValue &Root, SelectionDAG &DAG,
2113 DebugLoc dl) const {
2114 MachineFunction &MF = DAG.getMachineFunction();
2115 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2117 TargetRegisterClass *RC;
2118 if (AFI->isThumb1OnlyFunction())
2119 RC = ARM::tGPRRegisterClass;
2121 RC = ARM::GPRRegisterClass;
2123 // Transform the arguments stored in physical registers into virtual ones.
2124 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2125 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2128 if (NextVA.isMemLoc()) {
2129 MachineFrameInfo *MFI = MF.getFrameInfo();
2130 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2132 // Create load node to retrieve arguments from the stack.
2133 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2134 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2135 MachinePointerInfo::getFixedStack(FI),
2138 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2139 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2142 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2146 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2147 CallingConv::ID CallConv, bool isVarArg,
2148 const SmallVectorImpl<ISD::InputArg>
2150 DebugLoc dl, SelectionDAG &DAG,
2151 SmallVectorImpl<SDValue> &InVals)
2154 MachineFunction &MF = DAG.getMachineFunction();
2155 MachineFrameInfo *MFI = MF.getFrameInfo();
2157 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2159 // Assign locations to all of the incoming arguments.
2160 SmallVector<CCValAssign, 16> ArgLocs;
2161 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2163 CCInfo.AnalyzeFormalArguments(Ins,
2164 CCAssignFnForNode(CallConv, /* Return*/ false,
2167 SmallVector<SDValue, 16> ArgValues;
2169 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2170 CCValAssign &VA = ArgLocs[i];
2172 // Arguments stored in registers.
2173 if (VA.isRegLoc()) {
2174 EVT RegVT = VA.getLocVT();
2177 if (VA.needsCustom()) {
2178 // f64 and vector types are split up into multiple registers or
2179 // combinations of registers and stack slots.
2180 if (VA.getLocVT() == MVT::v2f64) {
2181 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2183 VA = ArgLocs[++i]; // skip ahead to next loc
2185 if (VA.isMemLoc()) {
2186 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2187 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2188 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2189 MachinePointerInfo::getFixedStack(FI),
2192 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2195 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2196 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2197 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2198 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2199 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2201 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2204 TargetRegisterClass *RC;
2206 if (RegVT == MVT::f32)
2207 RC = ARM::SPRRegisterClass;
2208 else if (RegVT == MVT::f64)
2209 RC = ARM::DPRRegisterClass;
2210 else if (RegVT == MVT::v2f64)
2211 RC = ARM::QPRRegisterClass;
2212 else if (RegVT == MVT::i32)
2213 RC = (AFI->isThumb1OnlyFunction() ?
2214 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2216 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2218 // Transform the arguments in physical registers into virtual ones.
2219 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2220 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2223 // If this is an 8 or 16-bit value, it is really passed promoted
2224 // to 32 bits. Insert an assert[sz]ext to capture this, then
2225 // truncate to the right size.
2226 switch (VA.getLocInfo()) {
2227 default: llvm_unreachable("Unknown loc info!");
2228 case CCValAssign::Full: break;
2229 case CCValAssign::BCvt:
2230 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2232 case CCValAssign::SExt:
2233 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2234 DAG.getValueType(VA.getValVT()));
2235 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2237 case CCValAssign::ZExt:
2238 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2239 DAG.getValueType(VA.getValVT()));
2240 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2244 InVals.push_back(ArgValue);
2246 } else { // VA.isRegLoc()
2249 assert(VA.isMemLoc());
2250 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2252 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2253 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2255 // Create load nodes to retrieve arguments from the stack.
2256 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2257 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2258 MachinePointerInfo::getFixedStack(FI),
2265 static const unsigned GPRArgRegs[] = {
2266 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2269 unsigned NumGPRs = CCInfo.getFirstUnallocated
2270 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2272 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2273 unsigned VARegSize = (4 - NumGPRs) * 4;
2274 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2275 unsigned ArgOffset = CCInfo.getNextStackOffset();
2276 if (VARegSaveSize) {
2277 // If this function is vararg, store any remaining integer argument regs
2278 // to their spots on the stack so that they may be loaded by deferencing
2279 // the result of va_next.
2280 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2281 AFI->setVarArgsFrameIndex(
2282 MFI->CreateFixedObject(VARegSaveSize,
2283 ArgOffset + VARegSaveSize - VARegSize,
2285 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2288 SmallVector<SDValue, 4> MemOps;
2289 for (; NumGPRs < 4; ++NumGPRs) {
2290 TargetRegisterClass *RC;
2291 if (AFI->isThumb1OnlyFunction())
2292 RC = ARM::tGPRRegisterClass;
2294 RC = ARM::GPRRegisterClass;
2296 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2297 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2299 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2300 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2302 MemOps.push_back(Store);
2303 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2304 DAG.getConstant(4, getPointerTy()));
2306 if (!MemOps.empty())
2307 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2308 &MemOps[0], MemOps.size());
2310 // This will point to the next argument passed via stack.
2311 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2317 /// isFloatingPointZero - Return true if this is +0.0.
2318 static bool isFloatingPointZero(SDValue Op) {
2319 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2320 return CFP->getValueAPF().isPosZero();
2321 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2322 // Maybe this has already been legalized into the constant pool?
2323 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2324 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2325 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2326 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2327 return CFP->getValueAPF().isPosZero();
2333 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2334 /// the given operands.
2336 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2337 SDValue &ARMcc, SelectionDAG &DAG,
2338 DebugLoc dl) const {
2339 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2340 unsigned C = RHSC->getZExtValue();
2341 if (!isLegalICmpImmediate(C)) {
2342 // Constant does not fit, try adjusting it by one?
2347 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2348 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2349 RHS = DAG.getConstant(C-1, MVT::i32);
2354 if (C != 0 && isLegalICmpImmediate(C-1)) {
2355 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2356 RHS = DAG.getConstant(C-1, MVT::i32);
2361 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2362 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2363 RHS = DAG.getConstant(C+1, MVT::i32);
2368 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2369 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2370 RHS = DAG.getConstant(C+1, MVT::i32);
2377 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2378 ARMISD::NodeType CompareType;
2381 CompareType = ARMISD::CMP;
2386 CompareType = ARMISD::CMPZ;
2389 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2390 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2393 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2395 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2396 DebugLoc dl) const {
2398 if (!isFloatingPointZero(RHS))
2399 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2401 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2402 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2405 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2406 SDValue Cond = Op.getOperand(0);
2407 SDValue SelectTrue = Op.getOperand(1);
2408 SDValue SelectFalse = Op.getOperand(2);
2409 DebugLoc dl = Op.getDebugLoc();
2413 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2414 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2416 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2417 const ConstantSDNode *CMOVTrue =
2418 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2419 const ConstantSDNode *CMOVFalse =
2420 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2422 if (CMOVTrue && CMOVFalse) {
2423 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2424 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2428 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2430 False = SelectFalse;
2431 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2436 if (True.getNode() && False.getNode()) {
2437 EVT VT = Cond.getValueType();
2438 SDValue ARMcc = Cond.getOperand(2);
2439 SDValue CCR = Cond.getOperand(3);
2440 SDValue Cmp = Cond.getOperand(4);
2441 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2446 return DAG.getSelectCC(dl, Cond,
2447 DAG.getConstant(0, Cond.getValueType()),
2448 SelectTrue, SelectFalse, ISD::SETNE);
2451 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2452 EVT VT = Op.getValueType();
2453 SDValue LHS = Op.getOperand(0);
2454 SDValue RHS = Op.getOperand(1);
2455 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2456 SDValue TrueVal = Op.getOperand(2);
2457 SDValue FalseVal = Op.getOperand(3);
2458 DebugLoc dl = Op.getDebugLoc();
2460 if (LHS.getValueType() == MVT::i32) {
2462 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2463 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2464 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2467 ARMCC::CondCodes CondCode, CondCode2;
2468 FPCCToARMCC(CC, CondCode, CondCode2);
2470 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2471 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2472 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2473 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2475 if (CondCode2 != ARMCC::AL) {
2476 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2477 // FIXME: Needs another CMP because flag can have but one use.
2478 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2479 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2480 Result, TrueVal, ARMcc2, CCR, Cmp2);
2485 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2486 /// to morph to an integer compare sequence.
2487 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2488 const ARMSubtarget *Subtarget) {
2489 SDNode *N = Op.getNode();
2490 if (!N->hasOneUse())
2491 // Otherwise it requires moving the value from fp to integer registers.
2493 if (!N->getNumValues())
2495 EVT VT = Op.getValueType();
2496 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2497 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2498 // vmrs are very slow, e.g. cortex-a8.
2501 if (isFloatingPointZero(Op)) {
2505 return ISD::isNormalLoad(N);
2508 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2509 if (isFloatingPointZero(Op))
2510 return DAG.getConstant(0, MVT::i32);
2512 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2513 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2514 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2515 Ld->isVolatile(), Ld->isNonTemporal(),
2516 Ld->getAlignment());
2518 llvm_unreachable("Unknown VFP cmp argument!");
2521 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2522 SDValue &RetVal1, SDValue &RetVal2) {
2523 if (isFloatingPointZero(Op)) {
2524 RetVal1 = DAG.getConstant(0, MVT::i32);
2525 RetVal2 = DAG.getConstant(0, MVT::i32);
2529 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2530 SDValue Ptr = Ld->getBasePtr();
2531 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2532 Ld->getChain(), Ptr,
2533 Ld->getPointerInfo(),
2534 Ld->isVolatile(), Ld->isNonTemporal(),
2535 Ld->getAlignment());
2537 EVT PtrType = Ptr.getValueType();
2538 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2539 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2540 PtrType, Ptr, DAG.getConstant(4, PtrType));
2541 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2542 Ld->getChain(), NewPtr,
2543 Ld->getPointerInfo().getWithOffset(4),
2544 Ld->isVolatile(), Ld->isNonTemporal(),
2549 llvm_unreachable("Unknown VFP cmp argument!");
2552 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2553 /// f32 and even f64 comparisons to integer ones.
2555 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2556 SDValue Chain = Op.getOperand(0);
2557 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2558 SDValue LHS = Op.getOperand(2);
2559 SDValue RHS = Op.getOperand(3);
2560 SDValue Dest = Op.getOperand(4);
2561 DebugLoc dl = Op.getDebugLoc();
2563 bool SeenZero = false;
2564 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2565 canChangeToInt(RHS, SeenZero, Subtarget) &&
2566 // If one of the operand is zero, it's safe to ignore the NaN case since
2567 // we only care about equality comparisons.
2568 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2569 // If unsafe fp math optimization is enabled and there are no othter uses of
2570 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2571 // to an integer comparison.
2572 if (CC == ISD::SETOEQ)
2574 else if (CC == ISD::SETUNE)
2578 if (LHS.getValueType() == MVT::f32) {
2579 LHS = bitcastf32Toi32(LHS, DAG);
2580 RHS = bitcastf32Toi32(RHS, DAG);
2581 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2582 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2583 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2584 Chain, Dest, ARMcc, CCR, Cmp);
2589 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2590 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2591 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2592 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2593 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2594 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2595 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2601 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2602 SDValue Chain = Op.getOperand(0);
2603 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2604 SDValue LHS = Op.getOperand(2);
2605 SDValue RHS = Op.getOperand(3);
2606 SDValue Dest = Op.getOperand(4);
2607 DebugLoc dl = Op.getDebugLoc();
2609 if (LHS.getValueType() == MVT::i32) {
2611 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2612 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2613 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2614 Chain, Dest, ARMcc, CCR, Cmp);
2617 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2620 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2621 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2622 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2623 if (Result.getNode())
2627 ARMCC::CondCodes CondCode, CondCode2;
2628 FPCCToARMCC(CC, CondCode, CondCode2);
2630 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2631 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2632 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2633 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2634 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2635 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2636 if (CondCode2 != ARMCC::AL) {
2637 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2638 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2639 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2644 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2645 SDValue Chain = Op.getOperand(0);
2646 SDValue Table = Op.getOperand(1);
2647 SDValue Index = Op.getOperand(2);
2648 DebugLoc dl = Op.getDebugLoc();
2650 EVT PTy = getPointerTy();
2651 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2652 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2653 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2654 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2655 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2656 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2657 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2658 if (Subtarget->isThumb2()) {
2659 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2660 // which does another jump to the destination. This also makes it easier
2661 // to translate it to TBB / TBH later.
2662 // FIXME: This might not work if the function is extremely large.
2663 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2664 Addr, Op.getOperand(2), JTI, UId);
2666 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2667 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2668 MachinePointerInfo::getJumpTable(),
2670 Chain = Addr.getValue(1);
2671 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2672 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2674 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2675 MachinePointerInfo::getJumpTable(), false, false, 0);
2676 Chain = Addr.getValue(1);
2677 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2681 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2682 DebugLoc dl = Op.getDebugLoc();
2685 switch (Op.getOpcode()) {
2687 assert(0 && "Invalid opcode!");
2688 case ISD::FP_TO_SINT:
2689 Opc = ARMISD::FTOSI;
2691 case ISD::FP_TO_UINT:
2692 Opc = ARMISD::FTOUI;
2695 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2696 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
2699 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2700 EVT VT = Op.getValueType();
2701 DebugLoc dl = Op.getDebugLoc();
2704 switch (Op.getOpcode()) {
2706 assert(0 && "Invalid opcode!");
2707 case ISD::SINT_TO_FP:
2708 Opc = ARMISD::SITOF;
2710 case ISD::UINT_TO_FP:
2711 Opc = ARMISD::UITOF;
2715 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
2716 return DAG.getNode(Opc, dl, VT, Op);
2719 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2720 // Implement fcopysign with a fabs and a conditional fneg.
2721 SDValue Tmp0 = Op.getOperand(0);
2722 SDValue Tmp1 = Op.getOperand(1);
2723 DebugLoc dl = Op.getDebugLoc();
2724 EVT VT = Op.getValueType();
2725 EVT SrcVT = Tmp1.getValueType();
2726 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2727 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2728 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2729 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2730 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2731 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2734 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2735 MachineFunction &MF = DAG.getMachineFunction();
2736 MachineFrameInfo *MFI = MF.getFrameInfo();
2737 MFI->setReturnAddressIsTaken(true);
2739 EVT VT = Op.getValueType();
2740 DebugLoc dl = Op.getDebugLoc();
2741 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2743 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2744 SDValue Offset = DAG.getConstant(4, MVT::i32);
2745 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2746 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2747 MachinePointerInfo(), false, false, 0);
2750 // Return LR, which contains the return address. Mark it an implicit live-in.
2751 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2752 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2755 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2756 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2757 MFI->setFrameAddressIsTaken(true);
2759 EVT VT = Op.getValueType();
2760 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2761 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2762 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2763 ? ARM::R7 : ARM::R11;
2764 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2766 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2767 MachinePointerInfo(),
2772 /// ExpandBITCAST - If the target supports VFP, this function is called to
2773 /// expand a bit convert where either the source or destination type is i64 to
2774 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2775 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2776 /// vectors), since the legalizer won't know what to do with that.
2777 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
2778 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2779 DebugLoc dl = N->getDebugLoc();
2780 SDValue Op = N->getOperand(0);
2782 // This function is only supposed to be called for i64 types, either as the
2783 // source or destination of the bit convert.
2784 EVT SrcVT = Op.getValueType();
2785 EVT DstVT = N->getValueType(0);
2786 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2787 "ExpandBITCAST called for non-i64 type");
2789 // Turn i64->f64 into VMOVDRR.
2790 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2791 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2792 DAG.getConstant(0, MVT::i32));
2793 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2794 DAG.getConstant(1, MVT::i32));
2795 return DAG.getNode(ISD::BITCAST, dl, DstVT,
2796 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2799 // Turn f64->i64 into VMOVRRD.
2800 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2801 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2802 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2803 // Merge the pieces into a single i64 value.
2804 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2810 /// getZeroVector - Returns a vector of specified type with all zero elements.
2811 /// Zero vectors are used to represent vector negation and in those cases
2812 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2813 /// not support i64 elements, so sometimes the zero vectors will need to be
2814 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2816 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2817 assert(VT.isVector() && "Expected a vector type");
2818 // The canonical modified immediate encoding of a zero vector is....0!
2819 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2820 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2821 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2822 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
2825 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2826 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2827 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2828 SelectionDAG &DAG) const {
2829 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2830 EVT VT = Op.getValueType();
2831 unsigned VTBits = VT.getSizeInBits();
2832 DebugLoc dl = Op.getDebugLoc();
2833 SDValue ShOpLo = Op.getOperand(0);
2834 SDValue ShOpHi = Op.getOperand(1);
2835 SDValue ShAmt = Op.getOperand(2);
2837 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2839 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2841 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2842 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2843 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2844 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2845 DAG.getConstant(VTBits, MVT::i32));
2846 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2847 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2848 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2850 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2851 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2853 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2854 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2857 SDValue Ops[2] = { Lo, Hi };
2858 return DAG.getMergeValues(Ops, 2, dl);
2861 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2862 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2863 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2864 SelectionDAG &DAG) const {
2865 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2866 EVT VT = Op.getValueType();
2867 unsigned VTBits = VT.getSizeInBits();
2868 DebugLoc dl = Op.getDebugLoc();
2869 SDValue ShOpLo = Op.getOperand(0);
2870 SDValue ShOpHi = Op.getOperand(1);
2871 SDValue ShAmt = Op.getOperand(2);
2874 assert(Op.getOpcode() == ISD::SHL_PARTS);
2875 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2876 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2877 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2878 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2879 DAG.getConstant(VTBits, MVT::i32));
2880 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2881 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2883 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2884 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2885 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2887 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2888 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2891 SDValue Ops[2] = { Lo, Hi };
2892 return DAG.getMergeValues(Ops, 2, dl);
2895 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2896 SelectionDAG &DAG) const {
2897 // The rounding mode is in bits 23:22 of the FPSCR.
2898 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2899 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2900 // so that the shift + and get folded into a bitfield extract.
2901 DebugLoc dl = Op.getDebugLoc();
2902 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2903 DAG.getConstant(Intrinsic::arm_get_fpscr,
2905 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2906 DAG.getConstant(1U << 22, MVT::i32));
2907 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2908 DAG.getConstant(22, MVT::i32));
2909 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2910 DAG.getConstant(3, MVT::i32));
2913 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2914 const ARMSubtarget *ST) {
2915 EVT VT = N->getValueType(0);
2916 DebugLoc dl = N->getDebugLoc();
2918 if (!ST->hasV6T2Ops())
2921 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2922 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2925 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2926 const ARMSubtarget *ST) {
2927 EVT VT = N->getValueType(0);
2928 DebugLoc dl = N->getDebugLoc();
2933 // Lower vector shifts on NEON to use VSHL.
2934 assert(ST->hasNEON() && "unexpected vector shift");
2936 // Left shifts translate directly to the vshiftu intrinsic.
2937 if (N->getOpcode() == ISD::SHL)
2938 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2939 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2940 N->getOperand(0), N->getOperand(1));
2942 assert((N->getOpcode() == ISD::SRA ||
2943 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2945 // NEON uses the same intrinsics for both left and right shifts. For
2946 // right shifts, the shift amounts are negative, so negate the vector of
2948 EVT ShiftVT = N->getOperand(1).getValueType();
2949 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2950 getZeroVector(ShiftVT, DAG, dl),
2952 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2953 Intrinsic::arm_neon_vshifts :
2954 Intrinsic::arm_neon_vshiftu);
2955 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2956 DAG.getConstant(vshiftInt, MVT::i32),
2957 N->getOperand(0), NegatedCount);
2960 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
2961 const ARMSubtarget *ST) {
2962 EVT VT = N->getValueType(0);
2963 DebugLoc dl = N->getDebugLoc();
2965 // We can get here for a node like i32 = ISD::SHL i32, i64
2969 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2970 "Unknown shift to lower!");
2972 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2973 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2974 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2977 // If we are in thumb mode, we don't have RRX.
2978 if (ST->isThumb1Only()) return SDValue();
2980 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2981 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2982 DAG.getConstant(0, MVT::i32));
2983 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2984 DAG.getConstant(1, MVT::i32));
2986 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2987 // captures the result into a carry flag.
2988 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2989 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2991 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2992 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2994 // Merge the pieces into a single i64 value.
2995 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2998 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2999 SDValue TmpOp0, TmpOp1;
3000 bool Invert = false;
3004 SDValue Op0 = Op.getOperand(0);
3005 SDValue Op1 = Op.getOperand(1);
3006 SDValue CC = Op.getOperand(2);
3007 EVT VT = Op.getValueType();
3008 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3009 DebugLoc dl = Op.getDebugLoc();
3011 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3012 switch (SetCCOpcode) {
3013 default: llvm_unreachable("Illegal FP comparison"); break;
3015 case ISD::SETNE: Invert = true; // Fallthrough
3017 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3019 case ISD::SETLT: Swap = true; // Fallthrough
3021 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3023 case ISD::SETLE: Swap = true; // Fallthrough
3025 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3026 case ISD::SETUGE: Swap = true; // Fallthrough
3027 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3028 case ISD::SETUGT: Swap = true; // Fallthrough
3029 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3030 case ISD::SETUEQ: Invert = true; // Fallthrough
3032 // Expand this to (OLT | OGT).
3036 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3037 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3039 case ISD::SETUO: Invert = true; // Fallthrough
3041 // Expand this to (OLT | OGE).
3045 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3046 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3050 // Integer comparisons.
3051 switch (SetCCOpcode) {
3052 default: llvm_unreachable("Illegal integer comparison"); break;
3053 case ISD::SETNE: Invert = true;
3054 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3055 case ISD::SETLT: Swap = true;
3056 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3057 case ISD::SETLE: Swap = true;
3058 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3059 case ISD::SETULT: Swap = true;
3060 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3061 case ISD::SETULE: Swap = true;
3062 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3065 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3066 if (Opc == ARMISD::VCEQ) {
3069 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3071 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3074 // Ignore bitconvert.
3075 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3076 AndOp = AndOp.getOperand(0);
3078 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3080 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3081 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3088 std::swap(Op0, Op1);
3090 // If one of the operands is a constant vector zero, attempt to fold the
3091 // comparison to a specialized compare-against-zero form.
3093 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3095 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3096 if (Opc == ARMISD::VCGE)
3097 Opc = ARMISD::VCLEZ;
3098 else if (Opc == ARMISD::VCGT)
3099 Opc = ARMISD::VCLTZ;
3104 if (SingleOp.getNode()) {
3107 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3109 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3111 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3113 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3115 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3117 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3120 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3124 Result = DAG.getNOT(dl, Result, VT);
3129 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3130 /// valid vector constant for a NEON instruction with a "modified immediate"
3131 /// operand (e.g., VMOV). If so, return the encoded value.
3132 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3133 unsigned SplatBitSize, SelectionDAG &DAG,
3134 EVT &VT, bool is128Bits, NEONModImmType type) {
3135 unsigned OpCmode, Imm;
3137 // SplatBitSize is set to the smallest size that splats the vector, so a
3138 // zero vector will always have SplatBitSize == 8. However, NEON modified
3139 // immediate instructions others than VMOV do not support the 8-bit encoding
3140 // of a zero vector, and the default encoding of zero is supposed to be the
3145 switch (SplatBitSize) {
3147 if (type != VMOVModImm)
3149 // Any 1-byte value is OK. Op=0, Cmode=1110.
3150 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3153 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3157 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3158 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3159 if ((SplatBits & ~0xff) == 0) {
3160 // Value = 0x00nn: Op=x, Cmode=100x.
3165 if ((SplatBits & ~0xff00) == 0) {
3166 // Value = 0xnn00: Op=x, Cmode=101x.
3168 Imm = SplatBits >> 8;
3174 // NEON's 32-bit VMOV supports splat values where:
3175 // * only one byte is nonzero, or
3176 // * the least significant byte is 0xff and the second byte is nonzero, or
3177 // * the least significant 2 bytes are 0xff and the third is nonzero.
3178 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3179 if ((SplatBits & ~0xff) == 0) {
3180 // Value = 0x000000nn: Op=x, Cmode=000x.
3185 if ((SplatBits & ~0xff00) == 0) {
3186 // Value = 0x0000nn00: Op=x, Cmode=001x.
3188 Imm = SplatBits >> 8;
3191 if ((SplatBits & ~0xff0000) == 0) {
3192 // Value = 0x00nn0000: Op=x, Cmode=010x.
3194 Imm = SplatBits >> 16;
3197 if ((SplatBits & ~0xff000000) == 0) {
3198 // Value = 0xnn000000: Op=x, Cmode=011x.
3200 Imm = SplatBits >> 24;
3204 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3205 if (type == OtherModImm) return SDValue();
3207 if ((SplatBits & ~0xffff) == 0 &&
3208 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3209 // Value = 0x0000nnff: Op=x, Cmode=1100.
3211 Imm = SplatBits >> 8;
3216 if ((SplatBits & ~0xffffff) == 0 &&
3217 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3218 // Value = 0x00nnffff: Op=x, Cmode=1101.
3220 Imm = SplatBits >> 16;
3221 SplatBits |= 0xffff;
3225 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3226 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3227 // VMOV.I32. A (very) minor optimization would be to replicate the value
3228 // and fall through here to test for a valid 64-bit splat. But, then the
3229 // caller would also need to check and handle the change in size.
3233 if (type != VMOVModImm)
3235 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3236 uint64_t BitMask = 0xff;
3238 unsigned ImmMask = 1;
3240 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3241 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3244 } else if ((SplatBits & BitMask) != 0) {
3250 // Op=1, Cmode=1110.
3253 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3258 llvm_unreachable("unexpected size for isNEONModifiedImm");
3262 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3263 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3266 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3267 bool &ReverseVEXT, unsigned &Imm) {
3268 unsigned NumElts = VT.getVectorNumElements();
3269 ReverseVEXT = false;
3271 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3277 // If this is a VEXT shuffle, the immediate value is the index of the first
3278 // element. The other shuffle indices must be the successive elements after
3280 unsigned ExpectedElt = Imm;
3281 for (unsigned i = 1; i < NumElts; ++i) {
3282 // Increment the expected index. If it wraps around, it may still be
3283 // a VEXT but the source vectors must be swapped.
3285 if (ExpectedElt == NumElts * 2) {
3290 if (M[i] < 0) continue; // ignore UNDEF indices
3291 if (ExpectedElt != static_cast<unsigned>(M[i]))
3295 // Adjust the index value if the source operands will be swapped.
3302 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3303 /// instruction with the specified blocksize. (The order of the elements
3304 /// within each block of the vector is reversed.)
3305 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3306 unsigned BlockSize) {
3307 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3308 "Only possible block sizes for VREV are: 16, 32, 64");
3310 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3314 unsigned NumElts = VT.getVectorNumElements();
3315 unsigned BlockElts = M[0] + 1;
3316 // If the first shuffle index is UNDEF, be optimistic.
3318 BlockElts = BlockSize / EltSz;
3320 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3323 for (unsigned i = 0; i < NumElts; ++i) {
3324 if (M[i] < 0) continue; // ignore UNDEF indices
3325 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3332 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3333 unsigned &WhichResult) {
3334 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3338 unsigned NumElts = VT.getVectorNumElements();
3339 WhichResult = (M[0] == 0 ? 0 : 1);
3340 for (unsigned i = 0; i < NumElts; i += 2) {
3341 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3342 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3348 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3349 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3350 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3351 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3352 unsigned &WhichResult) {
3353 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3357 unsigned NumElts = VT.getVectorNumElements();
3358 WhichResult = (M[0] == 0 ? 0 : 1);
3359 for (unsigned i = 0; i < NumElts; i += 2) {
3360 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3361 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3367 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3368 unsigned &WhichResult) {
3369 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3373 unsigned NumElts = VT.getVectorNumElements();
3374 WhichResult = (M[0] == 0 ? 0 : 1);
3375 for (unsigned i = 0; i != NumElts; ++i) {
3376 if (M[i] < 0) continue; // ignore UNDEF indices
3377 if ((unsigned) M[i] != 2 * i + WhichResult)
3381 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3382 if (VT.is64BitVector() && EltSz == 32)
3388 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3389 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3390 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3391 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3392 unsigned &WhichResult) {
3393 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3397 unsigned Half = VT.getVectorNumElements() / 2;
3398 WhichResult = (M[0] == 0 ? 0 : 1);
3399 for (unsigned j = 0; j != 2; ++j) {
3400 unsigned Idx = WhichResult;
3401 for (unsigned i = 0; i != Half; ++i) {
3402 int MIdx = M[i + j * Half];
3403 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3409 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3410 if (VT.is64BitVector() && EltSz == 32)
3416 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3417 unsigned &WhichResult) {
3418 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3422 unsigned NumElts = VT.getVectorNumElements();
3423 WhichResult = (M[0] == 0 ? 0 : 1);
3424 unsigned Idx = WhichResult * NumElts / 2;
3425 for (unsigned i = 0; i != NumElts; i += 2) {
3426 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3427 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3432 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3433 if (VT.is64BitVector() && EltSz == 32)
3439 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3440 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3441 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3442 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3443 unsigned &WhichResult) {
3444 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3448 unsigned NumElts = VT.getVectorNumElements();
3449 WhichResult = (M[0] == 0 ? 0 : 1);
3450 unsigned Idx = WhichResult * NumElts / 2;
3451 for (unsigned i = 0; i != NumElts; i += 2) {
3452 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3453 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3458 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3459 if (VT.is64BitVector() && EltSz == 32)
3465 // If N is an integer constant that can be moved into a register in one
3466 // instruction, return an SDValue of such a constant (will become a MOV
3467 // instruction). Otherwise return null.
3468 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3469 const ARMSubtarget *ST, DebugLoc dl) {
3471 if (!isa<ConstantSDNode>(N))
3473 Val = cast<ConstantSDNode>(N)->getZExtValue();
3475 if (ST->isThumb1Only()) {
3476 if (Val <= 255 || ~Val <= 255)
3477 return DAG.getConstant(Val, MVT::i32);
3479 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3480 return DAG.getConstant(Val, MVT::i32);
3485 // If this is a case we can't handle, return null and let the default
3486 // expansion code take care of it.
3487 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3488 const ARMSubtarget *ST) {
3489 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3490 DebugLoc dl = Op.getDebugLoc();
3491 EVT VT = Op.getValueType();
3493 APInt SplatBits, SplatUndef;
3494 unsigned SplatBitSize;
3496 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3497 if (SplatBitSize <= 64) {
3498 // Check if an immediate VMOV works.
3500 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3501 SplatUndef.getZExtValue(), SplatBitSize,
3502 DAG, VmovVT, VT.is128BitVector(),
3504 if (Val.getNode()) {
3505 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3506 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3509 // Try an immediate VMVN.
3510 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3511 ((1LL << SplatBitSize) - 1));
3512 Val = isNEONModifiedImm(NegatedImm,
3513 SplatUndef.getZExtValue(), SplatBitSize,
3514 DAG, VmovVT, VT.is128BitVector(),
3516 if (Val.getNode()) {
3517 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3518 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3523 // Scan through the operands to see if only one value is used.
3524 unsigned NumElts = VT.getVectorNumElements();
3525 bool isOnlyLowElement = true;
3526 bool usesOnlyOneValue = true;
3527 bool isConstant = true;
3529 for (unsigned i = 0; i < NumElts; ++i) {
3530 SDValue V = Op.getOperand(i);
3531 if (V.getOpcode() == ISD::UNDEF)
3534 isOnlyLowElement = false;
3535 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3538 if (!Value.getNode())
3540 else if (V != Value)
3541 usesOnlyOneValue = false;
3544 if (!Value.getNode())
3545 return DAG.getUNDEF(VT);
3547 if (isOnlyLowElement)
3548 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3550 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3552 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3553 // i32 and try again.
3554 if (usesOnlyOneValue && EltSize <= 32) {
3556 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3557 if (VT.getVectorElementType().isFloatingPoint()) {
3558 SmallVector<SDValue, 8> Ops;
3559 for (unsigned i = 0; i < NumElts; ++i)
3560 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3562 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3563 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3564 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3566 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3568 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3570 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3573 // If all elements are constants and the case above didn't get hit, fall back
3574 // to the default expansion, which will generate a load from the constant
3579 // Vectors with 32- or 64-bit elements can be built by directly assigning
3580 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3581 // will be legalized.
3582 if (EltSize >= 32) {
3583 // Do the expansion with floating-point types, since that is what the VFP
3584 // registers are defined to use, and since i64 is not legal.
3585 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3586 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3587 SmallVector<SDValue, 8> Ops;
3588 for (unsigned i = 0; i < NumElts; ++i)
3589 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
3590 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3591 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3597 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3598 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3599 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3600 /// are assumed to be legal.
3602 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3604 if (VT.getVectorNumElements() == 4 &&
3605 (VT.is128BitVector() || VT.is64BitVector())) {
3606 unsigned PFIndexes[4];
3607 for (unsigned i = 0; i != 4; ++i) {
3611 PFIndexes[i] = M[i];
3614 // Compute the index in the perfect shuffle table.
3615 unsigned PFTableIndex =
3616 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3617 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3618 unsigned Cost = (PFEntry >> 30);
3625 unsigned Imm, WhichResult;
3627 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3628 return (EltSize >= 32 ||
3629 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3630 isVREVMask(M, VT, 64) ||
3631 isVREVMask(M, VT, 32) ||
3632 isVREVMask(M, VT, 16) ||
3633 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3634 isVTRNMask(M, VT, WhichResult) ||
3635 isVUZPMask(M, VT, WhichResult) ||
3636 isVZIPMask(M, VT, WhichResult) ||
3637 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3638 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3639 isVZIP_v_undef_Mask(M, VT, WhichResult));
3642 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3643 /// the specified operations to build the shuffle.
3644 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3645 SDValue RHS, SelectionDAG &DAG,
3647 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3648 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3649 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3652 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3661 OP_VUZPL, // VUZP, left result
3662 OP_VUZPR, // VUZP, right result
3663 OP_VZIPL, // VZIP, left result
3664 OP_VZIPR, // VZIP, right result
3665 OP_VTRNL, // VTRN, left result
3666 OP_VTRNR // VTRN, right result
3669 if (OpNum == OP_COPY) {
3670 if (LHSID == (1*9+2)*9+3) return LHS;
3671 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3675 SDValue OpLHS, OpRHS;
3676 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3677 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3678 EVT VT = OpLHS.getValueType();
3681 default: llvm_unreachable("Unknown shuffle opcode!");
3683 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3688 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3689 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3693 return DAG.getNode(ARMISD::VEXT, dl, VT,
3695 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3698 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3699 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3702 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3703 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3706 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3707 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3711 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3712 SDValue V1 = Op.getOperand(0);
3713 SDValue V2 = Op.getOperand(1);
3714 DebugLoc dl = Op.getDebugLoc();
3715 EVT VT = Op.getValueType();
3716 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3717 SmallVector<int, 8> ShuffleMask;
3719 // Convert shuffles that are directly supported on NEON to target-specific
3720 // DAG nodes, instead of keeping them as shuffles and matching them again
3721 // during code selection. This is more efficient and avoids the possibility
3722 // of inconsistencies between legalization and selection.
3723 // FIXME: floating-point vectors should be canonicalized to integer vectors
3724 // of the same time so that they get CSEd properly.
3725 SVN->getMask(ShuffleMask);
3727 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3728 if (EltSize <= 32) {
3729 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3730 int Lane = SVN->getSplatIndex();
3731 // If this is undef splat, generate it via "just" vdup, if possible.
3732 if (Lane == -1) Lane = 0;
3734 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3735 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3737 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3738 DAG.getConstant(Lane, MVT::i32));
3743 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3746 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3747 DAG.getConstant(Imm, MVT::i32));
3750 if (isVREVMask(ShuffleMask, VT, 64))
3751 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3752 if (isVREVMask(ShuffleMask, VT, 32))
3753 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3754 if (isVREVMask(ShuffleMask, VT, 16))
3755 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3757 // Check for Neon shuffles that modify both input vectors in place.
3758 // If both results are used, i.e., if there are two shuffles with the same
3759 // source operands and with masks corresponding to both results of one of
3760 // these operations, DAG memoization will ensure that a single node is
3761 // used for both shuffles.
3762 unsigned WhichResult;
3763 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3764 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3765 V1, V2).getValue(WhichResult);
3766 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3767 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3768 V1, V2).getValue(WhichResult);
3769 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3770 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3771 V1, V2).getValue(WhichResult);
3773 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3774 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3775 V1, V1).getValue(WhichResult);
3776 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3777 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3778 V1, V1).getValue(WhichResult);
3779 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3780 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3781 V1, V1).getValue(WhichResult);
3784 // If the shuffle is not directly supported and it has 4 elements, use
3785 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3786 unsigned NumElts = VT.getVectorNumElements();
3788 unsigned PFIndexes[4];
3789 for (unsigned i = 0; i != 4; ++i) {
3790 if (ShuffleMask[i] < 0)
3793 PFIndexes[i] = ShuffleMask[i];
3796 // Compute the index in the perfect shuffle table.
3797 unsigned PFTableIndex =
3798 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3799 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3800 unsigned Cost = (PFEntry >> 30);
3803 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3806 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3807 if (EltSize >= 32) {
3808 // Do the expansion with floating-point types, since that is what the VFP
3809 // registers are defined to use, and since i64 is not legal.
3810 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3811 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3812 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
3813 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
3814 SmallVector<SDValue, 8> Ops;
3815 for (unsigned i = 0; i < NumElts; ++i) {
3816 if (ShuffleMask[i] < 0)
3817 Ops.push_back(DAG.getUNDEF(EltVT));
3819 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3820 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3821 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3824 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3825 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3831 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3832 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
3833 SDValue Lane = Op.getOperand(1);
3834 if (!isa<ConstantSDNode>(Lane))
3837 SDValue Vec = Op.getOperand(0);
3838 if (Op.getValueType() == MVT::i32 &&
3839 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
3840 DebugLoc dl = Op.getDebugLoc();
3841 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3847 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3848 // The only time a CONCAT_VECTORS operation can have legal types is when
3849 // two 64-bit vectors are concatenated to a 128-bit vector.
3850 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3851 "unexpected CONCAT_VECTORS");
3852 DebugLoc dl = Op.getDebugLoc();
3853 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3854 SDValue Op0 = Op.getOperand(0);
3855 SDValue Op1 = Op.getOperand(1);
3856 if (Op0.getOpcode() != ISD::UNDEF)
3857 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3858 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
3859 DAG.getIntPtrConstant(0));
3860 if (Op1.getOpcode() != ISD::UNDEF)
3861 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3862 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
3863 DAG.getIntPtrConstant(1));
3864 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
3867 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
3868 /// element has been zero/sign-extended, depending on the isSigned parameter,
3869 /// from an integer type half its size.
3870 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
3872 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
3873 EVT VT = N->getValueType(0);
3874 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
3875 SDNode *BVN = N->getOperand(0).getNode();
3876 if (BVN->getValueType(0) != MVT::v4i32 ||
3877 BVN->getOpcode() != ISD::BUILD_VECTOR)
3879 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
3880 unsigned HiElt = 1 - LoElt;
3881 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
3882 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
3883 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
3884 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
3885 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
3888 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
3889 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
3892 if (Hi0->isNullValue() && Hi1->isNullValue())
3898 if (N->getOpcode() != ISD::BUILD_VECTOR)
3901 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
3902 SDNode *Elt = N->getOperand(i).getNode();
3903 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
3904 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3905 unsigned HalfSize = EltSize / 2;
3907 int64_t SExtVal = C->getSExtValue();
3908 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
3911 if ((C->getZExtValue() >> HalfSize) != 0)
3922 /// isSignExtended - Check if a node is a vector value that is sign-extended
3923 /// or a constant BUILD_VECTOR with sign-extended elements.
3924 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
3925 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
3927 if (isExtendedBUILD_VECTOR(N, DAG, true))
3932 /// isZeroExtended - Check if a node is a vector value that is zero-extended
3933 /// or a constant BUILD_VECTOR with zero-extended elements.
3934 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
3935 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
3937 if (isExtendedBUILD_VECTOR(N, DAG, false))
3942 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
3943 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
3944 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3945 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3946 return N->getOperand(0);
3947 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
3948 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3949 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3950 LD->isNonTemporal(), LD->getAlignment());
3951 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
3952 // have been legalized as a BITCAST from v4i32.
3953 if (N->getOpcode() == ISD::BITCAST) {
3954 SDNode *BVN = N->getOperand(0).getNode();
3955 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
3956 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
3957 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
3958 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
3959 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
3961 // Construct a new BUILD_VECTOR with elements truncated to half the size.
3962 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
3963 EVT VT = N->getValueType(0);
3964 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
3965 unsigned NumElts = VT.getVectorNumElements();
3966 MVT TruncVT = MVT::getIntegerVT(EltSize);
3967 SmallVector<SDValue, 8> Ops;
3968 for (unsigned i = 0; i != NumElts; ++i) {
3969 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
3970 const APInt &CInt = C->getAPIntValue();
3971 Ops.push_back(DAG.getConstant(APInt(CInt).trunc(EltSize), TruncVT));
3973 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
3974 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
3977 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3978 // Multiplications are only custom-lowered for 128-bit vectors so that
3979 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3980 EVT VT = Op.getValueType();
3981 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3982 SDNode *N0 = Op.getOperand(0).getNode();
3983 SDNode *N1 = Op.getOperand(1).getNode();
3984 unsigned NewOpc = 0;
3985 if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
3986 NewOpc = ARMISD::VMULLs;
3987 else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
3988 NewOpc = ARMISD::VMULLu;
3989 else if (VT == MVT::v2i64)
3990 // Fall through to expand this. It is not legal.
3993 // Other vector multiplications are legal.
3996 // Legalize to a VMULL instruction.
3997 DebugLoc DL = Op.getDebugLoc();
3998 SDValue Op0 = SkipExtension(N0, DAG);
3999 SDValue Op1 = SkipExtension(N1, DAG);
4001 assert(Op0.getValueType().is64BitVector() &&
4002 Op1.getValueType().is64BitVector() &&
4003 "unexpected types for extended operands to VMULL");
4004 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4007 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4008 switch (Op.getOpcode()) {
4009 default: llvm_unreachable("Don't know how to custom lower this!");
4010 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4011 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4012 case ISD::GlobalAddress:
4013 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4014 LowerGlobalAddressELF(Op, DAG);
4015 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4016 case ISD::SELECT: return LowerSELECT(Op, DAG);
4017 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4018 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4019 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4020 case ISD::VASTART: return LowerVASTART(Op, DAG);
4021 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4022 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4023 case ISD::SINT_TO_FP:
4024 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4025 case ISD::FP_TO_SINT:
4026 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4027 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4028 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4029 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4030 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4031 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4032 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4033 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4034 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4036 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4039 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4040 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4041 case ISD::SRL_PARTS:
4042 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4043 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4044 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4045 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4046 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4047 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4048 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4049 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4050 case ISD::MUL: return LowerMUL(Op, DAG);
4055 /// ReplaceNodeResults - Replace the results of node with an illegal result
4056 /// type with new values built out of custom code.
4057 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4058 SmallVectorImpl<SDValue>&Results,
4059 SelectionDAG &DAG) const {
4061 switch (N->getOpcode()) {
4063 llvm_unreachable("Don't know how to custom expand this!");
4066 Res = ExpandBITCAST(N, DAG);
4070 Res = Expand64BitShift(N, DAG, Subtarget);
4074 Results.push_back(Res);
4077 //===----------------------------------------------------------------------===//
4078 // ARM Scheduler Hooks
4079 //===----------------------------------------------------------------------===//
4082 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4083 MachineBasicBlock *BB,
4084 unsigned Size) const {
4085 unsigned dest = MI->getOperand(0).getReg();
4086 unsigned ptr = MI->getOperand(1).getReg();
4087 unsigned oldval = MI->getOperand(2).getReg();
4088 unsigned newval = MI->getOperand(3).getReg();
4089 unsigned scratch = BB->getParent()->getRegInfo()
4090 .createVirtualRegister(ARM::GPRRegisterClass);
4091 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4092 DebugLoc dl = MI->getDebugLoc();
4093 bool isThumb2 = Subtarget->isThumb2();
4095 unsigned ldrOpc, strOpc;
4097 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4099 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4100 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
4103 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4104 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4107 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4108 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4112 MachineFunction *MF = BB->getParent();
4113 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4114 MachineFunction::iterator It = BB;
4115 ++It; // insert the new blocks after the current block
4117 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4118 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
4119 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4120 MF->insert(It, loop1MBB);
4121 MF->insert(It, loop2MBB);
4122 MF->insert(It, exitMBB);
4124 // Transfer the remainder of BB and its successor edges to exitMBB.
4125 exitMBB->splice(exitMBB->begin(), BB,
4126 llvm::next(MachineBasicBlock::iterator(MI)),
4128 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4132 // fallthrough --> loop1MBB
4133 BB->addSuccessor(loop1MBB);
4136 // ldrex dest, [ptr]
4140 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4141 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4142 .addReg(dest).addReg(oldval));
4143 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4144 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4145 BB->addSuccessor(loop2MBB);
4146 BB->addSuccessor(exitMBB);
4149 // strex scratch, newval, [ptr]
4153 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4155 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4156 .addReg(scratch).addImm(0));
4157 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4158 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4159 BB->addSuccessor(loop1MBB);
4160 BB->addSuccessor(exitMBB);
4166 MI->eraseFromParent(); // The instruction is gone now.
4172 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4173 unsigned Size, unsigned BinOpcode) const {
4174 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4175 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4177 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4178 MachineFunction *MF = BB->getParent();
4179 MachineFunction::iterator It = BB;
4182 unsigned dest = MI->getOperand(0).getReg();
4183 unsigned ptr = MI->getOperand(1).getReg();
4184 unsigned incr = MI->getOperand(2).getReg();
4185 DebugLoc dl = MI->getDebugLoc();
4187 bool isThumb2 = Subtarget->isThumb2();
4188 unsigned ldrOpc, strOpc;
4190 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4192 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4193 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4196 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4197 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4200 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4201 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4205 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4206 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4207 MF->insert(It, loopMBB);
4208 MF->insert(It, exitMBB);
4210 // Transfer the remainder of BB and its successor edges to exitMBB.
4211 exitMBB->splice(exitMBB->begin(), BB,
4212 llvm::next(MachineBasicBlock::iterator(MI)),
4214 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4216 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4217 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4218 unsigned scratch2 = (!BinOpcode) ? incr :
4219 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4223 // fallthrough --> loopMBB
4224 BB->addSuccessor(loopMBB);
4228 // <binop> scratch2, dest, incr
4229 // strex scratch, scratch2, ptr
4232 // fallthrough --> exitMBB
4234 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4236 // operand order needs to go the other way for NAND
4237 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4238 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4239 addReg(incr).addReg(dest)).addReg(0);
4241 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4242 addReg(dest).addReg(incr)).addReg(0);
4245 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4247 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4248 .addReg(scratch).addImm(0));
4249 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4250 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4252 BB->addSuccessor(loopMBB);
4253 BB->addSuccessor(exitMBB);
4259 MI->eraseFromParent(); // The instruction is gone now.
4265 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4266 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4267 E = MBB->succ_end(); I != E; ++I)
4270 llvm_unreachable("Expecting a BB with two successors!");
4274 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4275 MachineBasicBlock *BB) const {
4276 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4277 DebugLoc dl = MI->getDebugLoc();
4278 bool isThumb2 = Subtarget->isThumb2();
4279 switch (MI->getOpcode()) {
4282 llvm_unreachable("Unexpected instr type to insert");
4284 case ARM::ATOMIC_LOAD_ADD_I8:
4285 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4286 case ARM::ATOMIC_LOAD_ADD_I16:
4287 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4288 case ARM::ATOMIC_LOAD_ADD_I32:
4289 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4291 case ARM::ATOMIC_LOAD_AND_I8:
4292 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4293 case ARM::ATOMIC_LOAD_AND_I16:
4294 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4295 case ARM::ATOMIC_LOAD_AND_I32:
4296 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4298 case ARM::ATOMIC_LOAD_OR_I8:
4299 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4300 case ARM::ATOMIC_LOAD_OR_I16:
4301 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4302 case ARM::ATOMIC_LOAD_OR_I32:
4303 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4305 case ARM::ATOMIC_LOAD_XOR_I8:
4306 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4307 case ARM::ATOMIC_LOAD_XOR_I16:
4308 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4309 case ARM::ATOMIC_LOAD_XOR_I32:
4310 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4312 case ARM::ATOMIC_LOAD_NAND_I8:
4313 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4314 case ARM::ATOMIC_LOAD_NAND_I16:
4315 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4316 case ARM::ATOMIC_LOAD_NAND_I32:
4317 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4319 case ARM::ATOMIC_LOAD_SUB_I8:
4320 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4321 case ARM::ATOMIC_LOAD_SUB_I16:
4322 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4323 case ARM::ATOMIC_LOAD_SUB_I32:
4324 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4326 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4327 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4328 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4330 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4331 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4332 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4334 case ARM::tMOVCCr_pseudo: {
4335 // To "insert" a SELECT_CC instruction, we actually have to insert the
4336 // diamond control-flow pattern. The incoming instruction knows the
4337 // destination vreg to set, the condition code register to branch on, the
4338 // true/false values to select between, and a branch opcode to use.
4339 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4340 MachineFunction::iterator It = BB;
4346 // cmpTY ccX, r1, r2
4348 // fallthrough --> copy0MBB
4349 MachineBasicBlock *thisMBB = BB;
4350 MachineFunction *F = BB->getParent();
4351 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4352 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4353 F->insert(It, copy0MBB);
4354 F->insert(It, sinkMBB);
4356 // Transfer the remainder of BB and its successor edges to sinkMBB.
4357 sinkMBB->splice(sinkMBB->begin(), BB,
4358 llvm::next(MachineBasicBlock::iterator(MI)),
4360 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4362 BB->addSuccessor(copy0MBB);
4363 BB->addSuccessor(sinkMBB);
4365 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4366 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4369 // %FalseValue = ...
4370 // # fallthrough to sinkMBB
4373 // Update machine-CFG edges
4374 BB->addSuccessor(sinkMBB);
4377 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4380 BuildMI(*BB, BB->begin(), dl,
4381 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4382 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4383 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4385 MI->eraseFromParent(); // The pseudo instruction is gone now.
4390 case ARM::BCCZi64: {
4391 // Compare both parts that make up the double comparison separately for
4393 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4395 unsigned LHS1 = MI->getOperand(1).getReg();
4396 unsigned LHS2 = MI->getOperand(2).getReg();
4398 AddDefaultPred(BuildMI(BB, dl,
4399 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4400 .addReg(LHS1).addImm(0));
4401 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4402 .addReg(LHS2).addImm(0)
4403 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4405 unsigned RHS1 = MI->getOperand(3).getReg();
4406 unsigned RHS2 = MI->getOperand(4).getReg();
4407 AddDefaultPred(BuildMI(BB, dl,
4408 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4409 .addReg(LHS1).addReg(RHS1));
4410 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4411 .addReg(LHS2).addReg(RHS2)
4412 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4415 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4416 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4417 if (MI->getOperand(0).getImm() == ARMCC::NE)
4418 std::swap(destMBB, exitMBB);
4420 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4421 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4422 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4425 MI->eraseFromParent(); // The pseudo instruction is gone now.
4431 //===----------------------------------------------------------------------===//
4432 // ARM Optimization Hooks
4433 //===----------------------------------------------------------------------===//
4436 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4437 TargetLowering::DAGCombinerInfo &DCI) {
4438 SelectionDAG &DAG = DCI.DAG;
4439 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4440 EVT VT = N->getValueType(0);
4441 unsigned Opc = N->getOpcode();
4442 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4443 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4444 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4445 ISD::CondCode CC = ISD::SETCC_INVALID;
4448 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4450 SDValue CCOp = Slct.getOperand(0);
4451 if (CCOp.getOpcode() == ISD::SETCC)
4452 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4455 bool DoXform = false;
4457 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4460 if (LHS.getOpcode() == ISD::Constant &&
4461 cast<ConstantSDNode>(LHS)->isNullValue()) {
4463 } else if (CC != ISD::SETCC_INVALID &&
4464 RHS.getOpcode() == ISD::Constant &&
4465 cast<ConstantSDNode>(RHS)->isNullValue()) {
4466 std::swap(LHS, RHS);
4467 SDValue Op0 = Slct.getOperand(0);
4468 EVT OpVT = isSlctCC ? Op0.getValueType() :
4469 Op0.getOperand(0).getValueType();
4470 bool isInt = OpVT.isInteger();
4471 CC = ISD::getSetCCInverse(CC, isInt);
4473 if (!TLI.isCondCodeLegal(CC, OpVT))
4474 return SDValue(); // Inverse operator isn't legal.
4481 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4483 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4484 Slct.getOperand(0), Slct.getOperand(1), CC);
4485 SDValue CCOp = Slct.getOperand(0);
4487 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4488 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4489 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4490 CCOp, OtherOp, Result);
4495 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4496 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4497 /// called with the default operands, and if that fails, with commuted
4499 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4500 TargetLowering::DAGCombinerInfo &DCI) {
4501 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4502 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4503 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4504 if (Result.getNode()) return Result;
4509 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4511 static SDValue PerformADDCombine(SDNode *N,
4512 TargetLowering::DAGCombinerInfo &DCI) {
4513 SDValue N0 = N->getOperand(0);
4514 SDValue N1 = N->getOperand(1);
4516 // First try with the default operand order.
4517 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4518 if (Result.getNode())
4521 // If that didn't work, try again with the operands commuted.
4522 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4525 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4527 static SDValue PerformSUBCombine(SDNode *N,
4528 TargetLowering::DAGCombinerInfo &DCI) {
4529 SDValue N0 = N->getOperand(0);
4530 SDValue N1 = N->getOperand(1);
4532 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4533 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4534 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4535 if (Result.getNode()) return Result;
4541 static SDValue PerformMULCombine(SDNode *N,
4542 TargetLowering::DAGCombinerInfo &DCI,
4543 const ARMSubtarget *Subtarget) {
4544 SelectionDAG &DAG = DCI.DAG;
4546 if (Subtarget->isThumb1Only())
4549 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4552 EVT VT = N->getValueType(0);
4556 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4560 uint64_t MulAmt = C->getZExtValue();
4561 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4562 ShiftAmt = ShiftAmt & (32 - 1);
4563 SDValue V = N->getOperand(0);
4564 DebugLoc DL = N->getDebugLoc();
4567 MulAmt >>= ShiftAmt;
4568 if (isPowerOf2_32(MulAmt - 1)) {
4569 // (mul x, 2^N + 1) => (add (shl x, N), x)
4570 Res = DAG.getNode(ISD::ADD, DL, VT,
4571 V, DAG.getNode(ISD::SHL, DL, VT,
4572 V, DAG.getConstant(Log2_32(MulAmt-1),
4574 } else if (isPowerOf2_32(MulAmt + 1)) {
4575 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4576 Res = DAG.getNode(ISD::SUB, DL, VT,
4577 DAG.getNode(ISD::SHL, DL, VT,
4578 V, DAG.getConstant(Log2_32(MulAmt+1),
4585 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4586 DAG.getConstant(ShiftAmt, MVT::i32));
4588 // Do not add new nodes to DAG combiner worklist.
4589 DCI.CombineTo(N, Res, false);
4593 static SDValue PerformANDCombine(SDNode *N,
4594 TargetLowering::DAGCombinerInfo &DCI) {
4595 // Attempt to use immediate-form VBIC
4596 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
4597 DebugLoc dl = N->getDebugLoc();
4598 EVT VT = N->getValueType(0);
4599 SelectionDAG &DAG = DCI.DAG;
4601 APInt SplatBits, SplatUndef;
4602 unsigned SplatBitSize;
4605 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4606 if (SplatBitSize <= 64) {
4608 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
4609 SplatUndef.getZExtValue(), SplatBitSize,
4610 DAG, VbicVT, VT.is128BitVector(),
4612 if (Val.getNode()) {
4614 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
4615 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
4616 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
4624 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4625 static SDValue PerformORCombine(SDNode *N,
4626 TargetLowering::DAGCombinerInfo &DCI,
4627 const ARMSubtarget *Subtarget) {
4628 // Attempt to use immediate-form VORR
4629 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
4630 DebugLoc dl = N->getDebugLoc();
4631 EVT VT = N->getValueType(0);
4632 SelectionDAG &DAG = DCI.DAG;
4634 APInt SplatBits, SplatUndef;
4635 unsigned SplatBitSize;
4637 if (BVN && Subtarget->hasNEON() &&
4638 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4639 if (SplatBitSize <= 64) {
4641 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4642 SplatUndef.getZExtValue(), SplatBitSize,
4643 DAG, VorrVT, VT.is128BitVector(),
4645 if (Val.getNode()) {
4647 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
4648 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
4649 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
4654 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4657 // BFI is only available on V6T2+
4658 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4661 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4662 DebugLoc DL = N->getDebugLoc();
4663 // 1) or (and A, mask), val => ARMbfi A, val, mask
4664 // iff (val & mask) == val
4666 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4667 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4668 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4669 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4670 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4671 // (i.e., copy a bitfield value into another bitfield of the same width)
4672 if (N0.getOpcode() != ISD::AND)
4679 // The value and the mask need to be constants so we can verify this is
4680 // actually a bitfield set. If the mask is 0xffff, we can do better
4681 // via a movt instruction, so don't use BFI in that case.
4682 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4685 unsigned Mask = C->getZExtValue();
4689 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4690 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4691 unsigned Val = C->getZExtValue();
4692 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4694 Val >>= CountTrailingZeros_32(~Mask);
4696 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4697 DAG.getConstant(Val, MVT::i32),
4698 DAG.getConstant(Mask, MVT::i32));
4700 // Do not add new nodes to DAG combiner worklist.
4701 DCI.CombineTo(N, Res, false);
4702 } else if (N1.getOpcode() == ISD::AND) {
4703 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4704 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4707 unsigned Mask2 = C->getZExtValue();
4709 if (ARM::isBitFieldInvertedMask(Mask) &&
4710 ARM::isBitFieldInvertedMask(~Mask2) &&
4711 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4712 // The pack halfword instruction works better for masks that fit it,
4713 // so use that when it's available.
4714 if (Subtarget->hasT2ExtractPack() &&
4715 (Mask == 0xffff || Mask == 0xffff0000))
4718 unsigned lsb = CountTrailingZeros_32(Mask2);
4719 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4720 DAG.getConstant(lsb, MVT::i32));
4721 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4722 DAG.getConstant(Mask, MVT::i32));
4723 // Do not add new nodes to DAG combiner worklist.
4724 DCI.CombineTo(N, Res, false);
4725 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4726 ARM::isBitFieldInvertedMask(Mask2) &&
4727 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4728 // The pack halfword instruction works better for masks that fit it,
4729 // so use that when it's available.
4730 if (Subtarget->hasT2ExtractPack() &&
4731 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4734 unsigned lsb = CountTrailingZeros_32(Mask);
4735 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4736 DAG.getConstant(lsb, MVT::i32));
4737 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4738 DAG.getConstant(Mask2, MVT::i32));
4739 // Do not add new nodes to DAG combiner worklist.
4740 DCI.CombineTo(N, Res, false);
4747 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4748 /// ARMISD::VMOVRRD.
4749 static SDValue PerformVMOVRRDCombine(SDNode *N,
4750 TargetLowering::DAGCombinerInfo &DCI) {
4751 // vmovrrd(vmovdrr x, y) -> x,y
4752 SDValue InDouble = N->getOperand(0);
4753 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4754 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4758 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4759 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4760 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4761 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4762 SDValue Op0 = N->getOperand(0);
4763 SDValue Op1 = N->getOperand(1);
4764 if (Op0.getOpcode() == ISD::BITCAST)
4765 Op0 = Op0.getOperand(0);
4766 if (Op1.getOpcode() == ISD::BITCAST)
4767 Op1 = Op1.getOperand(0);
4768 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4769 Op0.getNode() == Op1.getNode() &&
4770 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4771 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
4772 N->getValueType(0), Op0.getOperand(0));
4776 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4777 /// ISD::BUILD_VECTOR.
4778 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4779 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4780 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4781 // into a pair of GPRs, which is fine when the value is used as a scalar,
4782 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4783 if (N->getNumOperands() == 2)
4784 return PerformVMOVDRRCombine(N, DAG);
4789 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
4790 /// ISD::VECTOR_SHUFFLE.
4791 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
4792 // The LLVM shufflevector instruction does not require the shuffle mask
4793 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
4794 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
4795 // operands do not match the mask length, they are extended by concatenating
4796 // them with undef vectors. That is probably the right thing for other
4797 // targets, but for NEON it is better to concatenate two double-register
4798 // size vector operands into a single quad-register size vector. Do that
4799 // transformation here:
4800 // shuffle(concat(v1, undef), concat(v2, undef)) ->
4801 // shuffle(concat(v1, v2), undef)
4802 SDValue Op0 = N->getOperand(0);
4803 SDValue Op1 = N->getOperand(1);
4804 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
4805 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
4806 Op0.getNumOperands() != 2 ||
4807 Op1.getNumOperands() != 2)
4809 SDValue Concat0Op1 = Op0.getOperand(1);
4810 SDValue Concat1Op1 = Op1.getOperand(1);
4811 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
4812 Concat1Op1.getOpcode() != ISD::UNDEF)
4814 // Skip the transformation if any of the types are illegal.
4815 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4816 EVT VT = N->getValueType(0);
4817 if (!TLI.isTypeLegal(VT) ||
4818 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
4819 !TLI.isTypeLegal(Concat1Op1.getValueType()))
4822 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
4823 Op0.getOperand(0), Op1.getOperand(0));
4824 // Translate the shuffle mask.
4825 SmallVector<int, 16> NewMask;
4826 unsigned NumElts = VT.getVectorNumElements();
4827 unsigned HalfElts = NumElts/2;
4828 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
4829 for (unsigned n = 0; n < NumElts; ++n) {
4830 int MaskElt = SVN->getMaskElt(n);
4832 if (MaskElt < (int)HalfElts)
4834 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
4835 NewElt = HalfElts + MaskElt - NumElts;
4836 NewMask.push_back(NewElt);
4838 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
4839 DAG.getUNDEF(VT), NewMask.data());
4842 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
4843 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
4844 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
4846 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
4847 SelectionDAG &DAG = DCI.DAG;
4848 EVT VT = N->getValueType(0);
4849 // vldN-dup instructions only support 64-bit vectors for N > 1.
4850 if (!VT.is64BitVector())
4853 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
4854 SDNode *VLD = N->getOperand(0).getNode();
4855 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
4857 unsigned NumVecs = 0;
4858 unsigned NewOpc = 0;
4859 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
4860 if (IntNo == Intrinsic::arm_neon_vld2lane) {
4862 NewOpc = ARMISD::VLD2DUP;
4863 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
4865 NewOpc = ARMISD::VLD3DUP;
4866 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
4868 NewOpc = ARMISD::VLD4DUP;
4873 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
4874 // numbers match the load.
4875 unsigned VLDLaneNo =
4876 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
4877 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4879 // Ignore uses of the chain result.
4880 if (UI.getUse().getResNo() == NumVecs)
4883 if (User->getOpcode() != ARMISD::VDUPLANE ||
4884 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
4888 // Create the vldN-dup node.
4891 for (n = 0; n < NumVecs; ++n)
4893 Tys[n] = MVT::Other;
4894 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
4895 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
4896 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
4897 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
4898 Ops, 2, VLDMemInt->getMemoryVT(),
4899 VLDMemInt->getMemOperand());
4902 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4904 unsigned ResNo = UI.getUse().getResNo();
4905 // Ignore uses of the chain result.
4906 if (ResNo == NumVecs)
4909 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
4912 // Now the vldN-lane intrinsic is dead except for its chain result.
4913 // Update uses of the chain.
4914 std::vector<SDValue> VLDDupResults;
4915 for (unsigned n = 0; n < NumVecs; ++n)
4916 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
4917 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
4918 DCI.CombineTo(VLD, VLDDupResults);
4923 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4924 /// ARMISD::VDUPLANE.
4925 static SDValue PerformVDUPLANECombine(SDNode *N,
4926 TargetLowering::DAGCombinerInfo &DCI) {
4927 SDValue Op = N->getOperand(0);
4929 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
4930 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
4931 if (CombineVLDDUP(N, DCI))
4932 return SDValue(N, 0);
4934 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4935 // redundant. Ignore bit_converts for now; element sizes are checked below.
4936 while (Op.getOpcode() == ISD::BITCAST)
4937 Op = Op.getOperand(0);
4938 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4941 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4942 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4943 // The canonical VMOV for a zero vector uses a 32-bit element size.
4944 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4946 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4948 EVT VT = N->getValueType(0);
4949 if (EltSize > VT.getVectorElementType().getSizeInBits())
4952 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
4955 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4956 /// operand of a vector shift operation, where all the elements of the
4957 /// build_vector must have the same constant integer value.
4958 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4959 // Ignore bit_converts.
4960 while (Op.getOpcode() == ISD::BITCAST)
4961 Op = Op.getOperand(0);
4962 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4963 APInt SplatBits, SplatUndef;
4964 unsigned SplatBitSize;
4966 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4967 HasAnyUndefs, ElementBits) ||
4968 SplatBitSize > ElementBits)
4970 Cnt = SplatBits.getSExtValue();
4974 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4975 /// operand of a vector shift left operation. That value must be in the range:
4976 /// 0 <= Value < ElementBits for a left shift; or
4977 /// 0 <= Value <= ElementBits for a long left shift.
4978 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4979 assert(VT.isVector() && "vector shift count is not a vector type");
4980 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4981 if (! getVShiftImm(Op, ElementBits, Cnt))
4983 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4986 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4987 /// operand of a vector shift right operation. For a shift opcode, the value
4988 /// is positive, but for an intrinsic the value count must be negative. The
4989 /// absolute value must be in the range:
4990 /// 1 <= |Value| <= ElementBits for a right shift; or
4991 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4992 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4994 assert(VT.isVector() && "vector shift count is not a vector type");
4995 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4996 if (! getVShiftImm(Op, ElementBits, Cnt))
5000 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
5003 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
5004 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
5005 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5008 // Don't do anything for most intrinsics.
5011 // Vector shifts: check for immediate versions and lower them.
5012 // Note: This is done during DAG combining instead of DAG legalizing because
5013 // the build_vectors for 64-bit vector element shift counts are generally
5014 // not legal, and it is hard to see their values after they get legalized to
5015 // loads from a constant pool.
5016 case Intrinsic::arm_neon_vshifts:
5017 case Intrinsic::arm_neon_vshiftu:
5018 case Intrinsic::arm_neon_vshiftls:
5019 case Intrinsic::arm_neon_vshiftlu:
5020 case Intrinsic::arm_neon_vshiftn:
5021 case Intrinsic::arm_neon_vrshifts:
5022 case Intrinsic::arm_neon_vrshiftu:
5023 case Intrinsic::arm_neon_vrshiftn:
5024 case Intrinsic::arm_neon_vqshifts:
5025 case Intrinsic::arm_neon_vqshiftu:
5026 case Intrinsic::arm_neon_vqshiftsu:
5027 case Intrinsic::arm_neon_vqshiftns:
5028 case Intrinsic::arm_neon_vqshiftnu:
5029 case Intrinsic::arm_neon_vqshiftnsu:
5030 case Intrinsic::arm_neon_vqrshiftns:
5031 case Intrinsic::arm_neon_vqrshiftnu:
5032 case Intrinsic::arm_neon_vqrshiftnsu: {
5033 EVT VT = N->getOperand(1).getValueType();
5035 unsigned VShiftOpc = 0;
5038 case Intrinsic::arm_neon_vshifts:
5039 case Intrinsic::arm_neon_vshiftu:
5040 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
5041 VShiftOpc = ARMISD::VSHL;
5044 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
5045 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
5046 ARMISD::VSHRs : ARMISD::VSHRu);
5051 case Intrinsic::arm_neon_vshiftls:
5052 case Intrinsic::arm_neon_vshiftlu:
5053 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
5055 llvm_unreachable("invalid shift count for vshll intrinsic");
5057 case Intrinsic::arm_neon_vrshifts:
5058 case Intrinsic::arm_neon_vrshiftu:
5059 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
5063 case Intrinsic::arm_neon_vqshifts:
5064 case Intrinsic::arm_neon_vqshiftu:
5065 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5069 case Intrinsic::arm_neon_vqshiftsu:
5070 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
5072 llvm_unreachable("invalid shift count for vqshlu intrinsic");
5074 case Intrinsic::arm_neon_vshiftn:
5075 case Intrinsic::arm_neon_vrshiftn:
5076 case Intrinsic::arm_neon_vqshiftns:
5077 case Intrinsic::arm_neon_vqshiftnu:
5078 case Intrinsic::arm_neon_vqshiftnsu:
5079 case Intrinsic::arm_neon_vqrshiftns:
5080 case Intrinsic::arm_neon_vqrshiftnu:
5081 case Intrinsic::arm_neon_vqrshiftnsu:
5082 // Narrowing shifts require an immediate right shift.
5083 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
5085 llvm_unreachable("invalid shift count for narrowing vector shift "
5089 llvm_unreachable("unhandled vector shift");
5093 case Intrinsic::arm_neon_vshifts:
5094 case Intrinsic::arm_neon_vshiftu:
5095 // Opcode already set above.
5097 case Intrinsic::arm_neon_vshiftls:
5098 case Intrinsic::arm_neon_vshiftlu:
5099 if (Cnt == VT.getVectorElementType().getSizeInBits())
5100 VShiftOpc = ARMISD::VSHLLi;
5102 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
5103 ARMISD::VSHLLs : ARMISD::VSHLLu);
5105 case Intrinsic::arm_neon_vshiftn:
5106 VShiftOpc = ARMISD::VSHRN; break;
5107 case Intrinsic::arm_neon_vrshifts:
5108 VShiftOpc = ARMISD::VRSHRs; break;
5109 case Intrinsic::arm_neon_vrshiftu:
5110 VShiftOpc = ARMISD::VRSHRu; break;
5111 case Intrinsic::arm_neon_vrshiftn:
5112 VShiftOpc = ARMISD::VRSHRN; break;
5113 case Intrinsic::arm_neon_vqshifts:
5114 VShiftOpc = ARMISD::VQSHLs; break;
5115 case Intrinsic::arm_neon_vqshiftu:
5116 VShiftOpc = ARMISD::VQSHLu; break;
5117 case Intrinsic::arm_neon_vqshiftsu:
5118 VShiftOpc = ARMISD::VQSHLsu; break;
5119 case Intrinsic::arm_neon_vqshiftns:
5120 VShiftOpc = ARMISD::VQSHRNs; break;
5121 case Intrinsic::arm_neon_vqshiftnu:
5122 VShiftOpc = ARMISD::VQSHRNu; break;
5123 case Intrinsic::arm_neon_vqshiftnsu:
5124 VShiftOpc = ARMISD::VQSHRNsu; break;
5125 case Intrinsic::arm_neon_vqrshiftns:
5126 VShiftOpc = ARMISD::VQRSHRNs; break;
5127 case Intrinsic::arm_neon_vqrshiftnu:
5128 VShiftOpc = ARMISD::VQRSHRNu; break;
5129 case Intrinsic::arm_neon_vqrshiftnsu:
5130 VShiftOpc = ARMISD::VQRSHRNsu; break;
5133 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5134 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
5137 case Intrinsic::arm_neon_vshiftins: {
5138 EVT VT = N->getOperand(1).getValueType();
5140 unsigned VShiftOpc = 0;
5142 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
5143 VShiftOpc = ARMISD::VSLI;
5144 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
5145 VShiftOpc = ARMISD::VSRI;
5147 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
5150 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
5151 N->getOperand(1), N->getOperand(2),
5152 DAG.getConstant(Cnt, MVT::i32));
5155 case Intrinsic::arm_neon_vqrshifts:
5156 case Intrinsic::arm_neon_vqrshiftu:
5157 // No immediate versions of these to check for.
5164 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
5165 /// lowers them. As with the vector shift intrinsics, this is done during DAG
5166 /// combining instead of DAG legalizing because the build_vectors for 64-bit
5167 /// vector element shift counts are generally not legal, and it is hard to see
5168 /// their values after they get legalized to loads from a constant pool.
5169 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
5170 const ARMSubtarget *ST) {
5171 EVT VT = N->getValueType(0);
5173 // Nothing to be done for scalar shifts.
5174 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5175 if (!VT.isVector() || !TLI.isTypeLegal(VT))
5178 assert(ST->hasNEON() && "unexpected vector shift");
5181 switch (N->getOpcode()) {
5182 default: llvm_unreachable("unexpected shift opcode");
5185 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
5186 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
5187 DAG.getConstant(Cnt, MVT::i32));
5192 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
5193 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
5194 ARMISD::VSHRs : ARMISD::VSHRu);
5195 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
5196 DAG.getConstant(Cnt, MVT::i32));
5202 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
5203 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
5204 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
5205 const ARMSubtarget *ST) {
5206 SDValue N0 = N->getOperand(0);
5208 // Check for sign- and zero-extensions of vector extract operations of 8-
5209 // and 16-bit vector elements. NEON supports these directly. They are
5210 // handled during DAG combining because type legalization will promote them
5211 // to 32-bit types and it is messy to recognize the operations after that.
5212 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
5213 SDValue Vec = N0.getOperand(0);
5214 SDValue Lane = N0.getOperand(1);
5215 EVT VT = N->getValueType(0);
5216 EVT EltVT = N0.getValueType();
5217 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5219 if (VT == MVT::i32 &&
5220 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
5221 TLI.isTypeLegal(Vec.getValueType()) &&
5222 isa<ConstantSDNode>(Lane)) {
5225 switch (N->getOpcode()) {
5226 default: llvm_unreachable("unexpected opcode");
5227 case ISD::SIGN_EXTEND:
5228 Opc = ARMISD::VGETLANEs;
5230 case ISD::ZERO_EXTEND:
5231 case ISD::ANY_EXTEND:
5232 Opc = ARMISD::VGETLANEu;
5235 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
5242 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
5243 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
5244 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
5245 const ARMSubtarget *ST) {
5246 // If the target supports NEON, try to use vmax/vmin instructions for f32
5247 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
5248 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
5249 // a NaN; only do the transformation when it matches that behavior.
5251 // For now only do this when using NEON for FP operations; if using VFP, it
5252 // is not obvious that the benefit outweighs the cost of switching to the
5254 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
5255 N->getValueType(0) != MVT::f32)
5258 SDValue CondLHS = N->getOperand(0);
5259 SDValue CondRHS = N->getOperand(1);
5260 SDValue LHS = N->getOperand(2);
5261 SDValue RHS = N->getOperand(3);
5262 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
5264 unsigned Opcode = 0;
5266 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
5267 IsReversed = false; // x CC y ? x : y
5268 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
5269 IsReversed = true ; // x CC y ? y : x
5283 // If LHS is NaN, an ordered comparison will be false and the result will
5284 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
5285 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5286 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
5287 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5289 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
5290 // will return -0, so vmin can only be used for unsafe math or if one of
5291 // the operands is known to be nonzero.
5292 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
5294 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5296 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
5305 // If LHS is NaN, an ordered comparison will be false and the result will
5306 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
5307 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5308 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
5309 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5311 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
5312 // will return +0, so vmax can only be used for unsafe math or if one of
5313 // the operands is known to be nonzero.
5314 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
5316 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5318 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
5324 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
5327 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
5328 DAGCombinerInfo &DCI) const {
5329 switch (N->getOpcode()) {
5331 case ISD::ADD: return PerformADDCombine(N, DCI);
5332 case ISD::SUB: return PerformSUBCombine(N, DCI);
5333 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
5334 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
5335 case ISD::AND: return PerformANDCombine(N, DCI);
5336 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
5337 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
5338 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
5339 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
5340 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
5341 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
5344 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
5345 case ISD::SIGN_EXTEND:
5346 case ISD::ZERO_EXTEND:
5347 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
5348 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
5353 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
5354 if (!Subtarget->allowsUnalignedMem())
5357 switch (VT.getSimpleVT().SimpleTy) {
5364 // FIXME: VLD1 etc with standard alignment is legal.
5368 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
5373 switch (VT.getSimpleVT().SimpleTy) {
5374 default: return false;
5389 if ((V & (Scale - 1)) != 0)
5392 return V == (V & ((1LL << 5) - 1));
5395 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
5396 const ARMSubtarget *Subtarget) {
5403 switch (VT.getSimpleVT().SimpleTy) {
5404 default: return false;
5409 // + imm12 or - imm8
5411 return V == (V & ((1LL << 8) - 1));
5412 return V == (V & ((1LL << 12) - 1));
5415 // Same as ARM mode. FIXME: NEON?
5416 if (!Subtarget->hasVFP2())
5421 return V == (V & ((1LL << 8) - 1));
5425 /// isLegalAddressImmediate - Return true if the integer value can be used
5426 /// as the offset of the target addressing mode for load / store of the
5428 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5429 const ARMSubtarget *Subtarget) {
5436 if (Subtarget->isThumb1Only())
5437 return isLegalT1AddressImmediate(V, VT);
5438 else if (Subtarget->isThumb2())
5439 return isLegalT2AddressImmediate(V, VT, Subtarget);
5444 switch (VT.getSimpleVT().SimpleTy) {
5445 default: return false;
5450 return V == (V & ((1LL << 12) - 1));
5453 return V == (V & ((1LL << 8) - 1));
5456 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5461 return V == (V & ((1LL << 8) - 1));
5465 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5467 int Scale = AM.Scale;
5471 switch (VT.getSimpleVT().SimpleTy) {
5472 default: return false;
5481 return Scale == 2 || Scale == 4 || Scale == 8;
5484 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5488 // Note, we allow "void" uses (basically, uses that aren't loads or
5489 // stores), because arm allows folding a scale into many arithmetic
5490 // operations. This should be made more precise and revisited later.
5492 // Allow r << imm, but the imm has to be a multiple of two.
5493 if (Scale & 1) return false;
5494 return isPowerOf2_32(Scale);
5498 /// isLegalAddressingMode - Return true if the addressing mode represented
5499 /// by AM is legal for this target, for a load/store of the specified type.
5500 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5501 const Type *Ty) const {
5502 EVT VT = getValueType(Ty, true);
5503 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5506 // Can never fold addr of global into load/store.
5511 case 0: // no scale reg, must be "r+i" or "r", or "i".
5514 if (Subtarget->isThumb1Only())
5518 // ARM doesn't support any R+R*scale+imm addr modes.
5525 if (Subtarget->isThumb2())
5526 return isLegalT2ScaledAddressingMode(AM, VT);
5528 int Scale = AM.Scale;
5529 switch (VT.getSimpleVT().SimpleTy) {
5530 default: return false;
5534 if (Scale < 0) Scale = -Scale;
5538 return isPowerOf2_32(Scale & ~1);
5542 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5547 // Note, we allow "void" uses (basically, uses that aren't loads or
5548 // stores), because arm allows folding a scale into many arithmetic
5549 // operations. This should be made more precise and revisited later.
5551 // Allow r << imm, but the imm has to be a multiple of two.
5552 if (Scale & 1) return false;
5553 return isPowerOf2_32(Scale);
5560 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5561 /// icmp immediate, that is the target has icmp instructions which can compare
5562 /// a register against the immediate without having to materialize the
5563 /// immediate into a register.
5564 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5565 if (!Subtarget->isThumb())
5566 return ARM_AM::getSOImmVal(Imm) != -1;
5567 if (Subtarget->isThumb2())
5568 return ARM_AM::getT2SOImmVal(Imm) != -1;
5569 return Imm >= 0 && Imm <= 255;
5572 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5573 bool isSEXTLoad, SDValue &Base,
5574 SDValue &Offset, bool &isInc,
5575 SelectionDAG &DAG) {
5576 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5579 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5581 Base = Ptr->getOperand(0);
5582 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5583 int RHSC = (int)RHS->getZExtValue();
5584 if (RHSC < 0 && RHSC > -256) {
5585 assert(Ptr->getOpcode() == ISD::ADD);
5587 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5591 isInc = (Ptr->getOpcode() == ISD::ADD);
5592 Offset = Ptr->getOperand(1);
5594 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5596 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5597 int RHSC = (int)RHS->getZExtValue();
5598 if (RHSC < 0 && RHSC > -0x1000) {
5599 assert(Ptr->getOpcode() == ISD::ADD);
5601 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5602 Base = Ptr->getOperand(0);
5607 if (Ptr->getOpcode() == ISD::ADD) {
5609 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5610 if (ShOpcVal != ARM_AM::no_shift) {
5611 Base = Ptr->getOperand(1);
5612 Offset = Ptr->getOperand(0);
5614 Base = Ptr->getOperand(0);
5615 Offset = Ptr->getOperand(1);
5620 isInc = (Ptr->getOpcode() == ISD::ADD);
5621 Base = Ptr->getOperand(0);
5622 Offset = Ptr->getOperand(1);
5626 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5630 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5631 bool isSEXTLoad, SDValue &Base,
5632 SDValue &Offset, bool &isInc,
5633 SelectionDAG &DAG) {
5634 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5637 Base = Ptr->getOperand(0);
5638 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5639 int RHSC = (int)RHS->getZExtValue();
5640 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5641 assert(Ptr->getOpcode() == ISD::ADD);
5643 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5645 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5646 isInc = Ptr->getOpcode() == ISD::ADD;
5647 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5655 /// getPreIndexedAddressParts - returns true by value, base pointer and
5656 /// offset pointer and addressing mode by reference if the node's address
5657 /// can be legally represented as pre-indexed load / store address.
5659 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5661 ISD::MemIndexedMode &AM,
5662 SelectionDAG &DAG) const {
5663 if (Subtarget->isThumb1Only())
5668 bool isSEXTLoad = false;
5669 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5670 Ptr = LD->getBasePtr();
5671 VT = LD->getMemoryVT();
5672 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5673 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5674 Ptr = ST->getBasePtr();
5675 VT = ST->getMemoryVT();
5680 bool isLegal = false;
5681 if (Subtarget->isThumb2())
5682 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5683 Offset, isInc, DAG);
5685 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5686 Offset, isInc, DAG);
5690 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5694 /// getPostIndexedAddressParts - returns true by value, base pointer and
5695 /// offset pointer and addressing mode by reference if this node can be
5696 /// combined with a load / store to form a post-indexed load / store.
5697 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5700 ISD::MemIndexedMode &AM,
5701 SelectionDAG &DAG) const {
5702 if (Subtarget->isThumb1Only())
5707 bool isSEXTLoad = false;
5708 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5709 VT = LD->getMemoryVT();
5710 Ptr = LD->getBasePtr();
5711 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5712 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5713 VT = ST->getMemoryVT();
5714 Ptr = ST->getBasePtr();
5719 bool isLegal = false;
5720 if (Subtarget->isThumb2())
5721 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5724 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5730 // Swap base ptr and offset to catch more post-index load / store when
5731 // it's legal. In Thumb2 mode, offset must be an immediate.
5732 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5733 !Subtarget->isThumb2())
5734 std::swap(Base, Offset);
5736 // Post-indexed load / store update the base pointer.
5741 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5745 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5749 const SelectionDAG &DAG,
5750 unsigned Depth) const {
5751 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5752 switch (Op.getOpcode()) {
5754 case ARMISD::CMOV: {
5755 // Bits are known zero/one if known on the LHS and RHS.
5756 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5757 if (KnownZero == 0 && KnownOne == 0) return;
5759 APInt KnownZeroRHS, KnownOneRHS;
5760 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5761 KnownZeroRHS, KnownOneRHS, Depth+1);
5762 KnownZero &= KnownZeroRHS;
5763 KnownOne &= KnownOneRHS;
5769 //===----------------------------------------------------------------------===//
5770 // ARM Inline Assembly Support
5771 //===----------------------------------------------------------------------===//
5773 /// getConstraintType - Given a constraint letter, return the type of
5774 /// constraint it is for this target.
5775 ARMTargetLowering::ConstraintType
5776 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5777 if (Constraint.size() == 1) {
5778 switch (Constraint[0]) {
5780 case 'l': return C_RegisterClass;
5781 case 'w': return C_RegisterClass;
5784 return TargetLowering::getConstraintType(Constraint);
5787 /// Examine constraint type and operand type and determine a weight value.
5788 /// This object must already have been set up with the operand type
5789 /// and the current alternative constraint selected.
5790 TargetLowering::ConstraintWeight
5791 ARMTargetLowering::getSingleConstraintMatchWeight(
5792 AsmOperandInfo &info, const char *constraint) const {
5793 ConstraintWeight weight = CW_Invalid;
5794 Value *CallOperandVal = info.CallOperandVal;
5795 // If we don't have a value, we can't do a match,
5796 // but allow it at the lowest weight.
5797 if (CallOperandVal == NULL)
5799 const Type *type = CallOperandVal->getType();
5800 // Look at the constraint type.
5801 switch (*constraint) {
5803 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5806 if (type->isIntegerTy()) {
5807 if (Subtarget->isThumb())
5808 weight = CW_SpecificReg;
5810 weight = CW_Register;
5814 if (type->isFloatingPointTy())
5815 weight = CW_Register;
5821 std::pair<unsigned, const TargetRegisterClass*>
5822 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5824 if (Constraint.size() == 1) {
5825 // GCC ARM Constraint Letters
5826 switch (Constraint[0]) {
5828 if (Subtarget->isThumb())
5829 return std::make_pair(0U, ARM::tGPRRegisterClass);
5831 return std::make_pair(0U, ARM::GPRRegisterClass);
5833 return std::make_pair(0U, ARM::GPRRegisterClass);
5836 return std::make_pair(0U, ARM::SPRRegisterClass);
5837 if (VT.getSizeInBits() == 64)
5838 return std::make_pair(0U, ARM::DPRRegisterClass);
5839 if (VT.getSizeInBits() == 128)
5840 return std::make_pair(0U, ARM::QPRRegisterClass);
5844 if (StringRef("{cc}").equals_lower(Constraint))
5845 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5847 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5850 std::vector<unsigned> ARMTargetLowering::
5851 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5853 if (Constraint.size() != 1)
5854 return std::vector<unsigned>();
5856 switch (Constraint[0]) { // GCC ARM Constraint Letters
5859 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5860 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5863 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5864 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5865 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5866 ARM::R12, ARM::LR, 0);
5869 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5870 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5871 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5872 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5873 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5874 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5875 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5876 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5877 if (VT.getSizeInBits() == 64)
5878 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5879 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5880 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5881 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5882 if (VT.getSizeInBits() == 128)
5883 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5884 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5888 return std::vector<unsigned>();
5891 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5892 /// vector. If it is invalid, don't add anything to Ops.
5893 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5895 std::vector<SDValue>&Ops,
5896 SelectionDAG &DAG) const {
5897 SDValue Result(0, 0);
5899 switch (Constraint) {
5901 case 'I': case 'J': case 'K': case 'L':
5902 case 'M': case 'N': case 'O':
5903 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5907 int64_t CVal64 = C->getSExtValue();
5908 int CVal = (int) CVal64;
5909 // None of these constraints allow values larger than 32 bits. Check
5910 // that the value fits in an int.
5914 switch (Constraint) {
5916 if (Subtarget->isThumb1Only()) {
5917 // This must be a constant between 0 and 255, for ADD
5919 if (CVal >= 0 && CVal <= 255)
5921 } else if (Subtarget->isThumb2()) {
5922 // A constant that can be used as an immediate value in a
5923 // data-processing instruction.
5924 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5927 // A constant that can be used as an immediate value in a
5928 // data-processing instruction.
5929 if (ARM_AM::getSOImmVal(CVal) != -1)
5935 if (Subtarget->isThumb()) { // FIXME thumb2
5936 // This must be a constant between -255 and -1, for negated ADD
5937 // immediates. This can be used in GCC with an "n" modifier that
5938 // prints the negated value, for use with SUB instructions. It is
5939 // not useful otherwise but is implemented for compatibility.
5940 if (CVal >= -255 && CVal <= -1)
5943 // This must be a constant between -4095 and 4095. It is not clear
5944 // what this constraint is intended for. Implemented for
5945 // compatibility with GCC.
5946 if (CVal >= -4095 && CVal <= 4095)
5952 if (Subtarget->isThumb1Only()) {
5953 // A 32-bit value where only one byte has a nonzero value. Exclude
5954 // zero to match GCC. This constraint is used by GCC internally for
5955 // constants that can be loaded with a move/shift combination.
5956 // It is not useful otherwise but is implemented for compatibility.
5957 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5959 } else if (Subtarget->isThumb2()) {
5960 // A constant whose bitwise inverse can be used as an immediate
5961 // value in a data-processing instruction. This can be used in GCC
5962 // with a "B" modifier that prints the inverted value, for use with
5963 // BIC and MVN instructions. It is not useful otherwise but is
5964 // implemented for compatibility.
5965 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5968 // A constant whose bitwise inverse can be used as an immediate
5969 // value in a data-processing instruction. This can be used in GCC
5970 // with a "B" modifier that prints the inverted value, for use with
5971 // BIC and MVN instructions. It is not useful otherwise but is
5972 // implemented for compatibility.
5973 if (ARM_AM::getSOImmVal(~CVal) != -1)
5979 if (Subtarget->isThumb1Only()) {
5980 // This must be a constant between -7 and 7,
5981 // for 3-operand ADD/SUB immediate instructions.
5982 if (CVal >= -7 && CVal < 7)
5984 } else if (Subtarget->isThumb2()) {
5985 // A constant whose negation can be used as an immediate value in a
5986 // data-processing instruction. This can be used in GCC with an "n"
5987 // modifier that prints the negated value, for use with SUB
5988 // instructions. It is not useful otherwise but is implemented for
5990 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5993 // A constant whose negation can be used as an immediate value in a
5994 // data-processing instruction. This can be used in GCC with an "n"
5995 // modifier that prints the negated value, for use with SUB
5996 // instructions. It is not useful otherwise but is implemented for
5998 if (ARM_AM::getSOImmVal(-CVal) != -1)
6004 if (Subtarget->isThumb()) { // FIXME thumb2
6005 // This must be a multiple of 4 between 0 and 1020, for
6006 // ADD sp + immediate.
6007 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
6010 // A power of two or a constant between 0 and 32. This is used in
6011 // GCC for the shift amount on shifted register operands, but it is
6012 // useful in general for any shift amounts.
6013 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
6019 if (Subtarget->isThumb()) { // FIXME thumb2
6020 // This must be a constant between 0 and 31, for shift amounts.
6021 if (CVal >= 0 && CVal <= 31)
6027 if (Subtarget->isThumb()) { // FIXME thumb2
6028 // This must be a multiple of 4 between -508 and 508, for
6029 // ADD/SUB sp = sp + immediate.
6030 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
6035 Result = DAG.getTargetConstant(CVal, Op.getValueType());
6039 if (Result.getNode()) {
6040 Ops.push_back(Result);
6043 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6047 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6048 // The ARM target isn't yet aware of offsets.
6052 int ARM::getVFPf32Imm(const APFloat &FPImm) {
6053 APInt Imm = FPImm.bitcastToAPInt();
6054 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
6055 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
6056 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
6058 // We can handle 4 bits of mantissa.
6059 // mantissa = (16+UInt(e:f:g:h))/16.
6060 if (Mantissa & 0x7ffff)
6063 if ((Mantissa & 0xf) != Mantissa)
6066 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6067 if (Exp < -3 || Exp > 4)
6069 Exp = ((Exp+3) & 0x7) ^ 4;
6071 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6074 int ARM::getVFPf64Imm(const APFloat &FPImm) {
6075 APInt Imm = FPImm.bitcastToAPInt();
6076 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
6077 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
6078 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
6080 // We can handle 4 bits of mantissa.
6081 // mantissa = (16+UInt(e:f:g:h))/16.
6082 if (Mantissa & 0xffffffffffffLL)
6085 if ((Mantissa & 0xf) != Mantissa)
6088 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
6089 if (Exp < -3 || Exp > 4)
6091 Exp = ((Exp+3) & 0x7) ^ 4;
6093 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
6096 bool ARM::isBitFieldInvertedMask(unsigned v) {
6097 if (v == 0xffffffff)
6099 // there can be 1's on either or both "outsides", all the "inside"
6101 unsigned int lsb = 0, msb = 31;
6102 while (v & (1 << msb)) --msb;
6103 while (v & (1 << lsb)) ++lsb;
6104 for (unsigned int i = lsb; i <= msb; ++i) {
6111 /// isFPImmLegal - Returns true if the target can instruction select the
6112 /// specified FP immediate natively. If false, the legalizer will
6113 /// materialize the FP immediate as a load from a constant pool.
6114 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
6115 if (!Subtarget->hasVFP3())
6118 return ARM::getVFPf32Imm(Imm) != -1;
6120 return ARM::getVFPf64Imm(Imm) != -1;
6124 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
6125 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
6126 /// specified in the intrinsic calls.
6127 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
6129 unsigned Intrinsic) const {
6130 switch (Intrinsic) {
6131 case Intrinsic::arm_neon_vld1:
6132 case Intrinsic::arm_neon_vld2:
6133 case Intrinsic::arm_neon_vld3:
6134 case Intrinsic::arm_neon_vld4:
6135 case Intrinsic::arm_neon_vld2lane:
6136 case Intrinsic::arm_neon_vld3lane:
6137 case Intrinsic::arm_neon_vld4lane: {
6138 Info.opc = ISD::INTRINSIC_W_CHAIN;
6139 // Conservatively set memVT to the entire set of vectors loaded.
6140 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
6141 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6142 Info.ptrVal = I.getArgOperand(0);
6144 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6145 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6146 Info.vol = false; // volatile loads with NEON intrinsics not supported
6147 Info.readMem = true;
6148 Info.writeMem = false;
6151 case Intrinsic::arm_neon_vst1:
6152 case Intrinsic::arm_neon_vst2:
6153 case Intrinsic::arm_neon_vst3:
6154 case Intrinsic::arm_neon_vst4:
6155 case Intrinsic::arm_neon_vst2lane:
6156 case Intrinsic::arm_neon_vst3lane:
6157 case Intrinsic::arm_neon_vst4lane: {
6158 Info.opc = ISD::INTRINSIC_VOID;
6159 // Conservatively set memVT to the entire set of vectors stored.
6160 unsigned NumElts = 0;
6161 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
6162 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
6163 if (!ArgTy->isVectorTy())
6165 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
6167 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
6168 Info.ptrVal = I.getArgOperand(0);
6170 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
6171 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
6172 Info.vol = false; // volatile stores with NEON intrinsics not supported
6173 Info.readMem = false;
6174 Info.writeMem = true;