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()) {
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->isThumb1Only())) {
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::BIT_CONVERT, 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())
677 setTargetDAGCombine(ISD::OR);
679 setStackPointerRegisterToSaveRestore(ARM::SP);
681 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
682 setSchedulingPreference(Sched::RegPressure);
684 setSchedulingPreference(Sched::Hybrid);
686 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
688 // On ARM arguments smaller than 4 bytes are extended, so all arguments
689 // are at least 4 bytes aligned.
690 setMinStackArgumentAlignment(4);
692 benefitFromCodePlacementOpt = true;
695 std::pair<const TargetRegisterClass*, uint8_t>
696 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
697 const TargetRegisterClass *RRC = 0;
699 switch (VT.getSimpleVT().SimpleTy) {
701 return TargetLowering::findRepresentativeClass(VT);
702 // Use DPR as representative register class for all floating point
703 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
704 // the cost is 1 for both f32 and f64.
705 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
706 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
707 RRC = ARM::DPRRegisterClass;
709 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
710 case MVT::v4f32: case MVT::v2f64:
711 RRC = ARM::DPRRegisterClass;
715 RRC = ARM::DPRRegisterClass;
719 RRC = ARM::DPRRegisterClass;
723 return std::make_pair(RRC, Cost);
726 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
729 case ARMISD::Wrapper: return "ARMISD::Wrapper";
730 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
731 case ARMISD::CALL: return "ARMISD::CALL";
732 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
733 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
734 case ARMISD::tCALL: return "ARMISD::tCALL";
735 case ARMISD::BRCOND: return "ARMISD::BRCOND";
736 case ARMISD::BR_JT: return "ARMISD::BR_JT";
737 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
738 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
739 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
740 case ARMISD::CMP: return "ARMISD::CMP";
741 case ARMISD::CMPZ: return "ARMISD::CMPZ";
742 case ARMISD::CMPFP: return "ARMISD::CMPFP";
743 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
744 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
745 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
746 case ARMISD::CMOV: return "ARMISD::CMOV";
747 case ARMISD::CNEG: return "ARMISD::CNEG";
749 case ARMISD::RBIT: return "ARMISD::RBIT";
751 case ARMISD::FTOSI: return "ARMISD::FTOSI";
752 case ARMISD::FTOUI: return "ARMISD::FTOUI";
753 case ARMISD::SITOF: return "ARMISD::SITOF";
754 case ARMISD::UITOF: return "ARMISD::UITOF";
756 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
757 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
758 case ARMISD::RRX: return "ARMISD::RRX";
760 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
761 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
763 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
764 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
765 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
767 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
769 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
771 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
773 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
774 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
776 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
778 case ARMISD::VCEQ: return "ARMISD::VCEQ";
779 case ARMISD::VCGE: return "ARMISD::VCGE";
780 case ARMISD::VCGEU: return "ARMISD::VCGEU";
781 case ARMISD::VCGT: return "ARMISD::VCGT";
782 case ARMISD::VCGTU: return "ARMISD::VCGTU";
783 case ARMISD::VTST: return "ARMISD::VTST";
785 case ARMISD::VSHL: return "ARMISD::VSHL";
786 case ARMISD::VSHRs: return "ARMISD::VSHRs";
787 case ARMISD::VSHRu: return "ARMISD::VSHRu";
788 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
789 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
790 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
791 case ARMISD::VSHRN: return "ARMISD::VSHRN";
792 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
793 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
794 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
795 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
796 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
797 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
798 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
799 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
800 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
801 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
802 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
803 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
804 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
805 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
806 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
807 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
808 case ARMISD::VDUP: return "ARMISD::VDUP";
809 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
810 case ARMISD::VEXT: return "ARMISD::VEXT";
811 case ARMISD::VREV64: return "ARMISD::VREV64";
812 case ARMISD::VREV32: return "ARMISD::VREV32";
813 case ARMISD::VREV16: return "ARMISD::VREV16";
814 case ARMISD::VZIP: return "ARMISD::VZIP";
815 case ARMISD::VUZP: return "ARMISD::VUZP";
816 case ARMISD::VTRN: return "ARMISD::VTRN";
817 case ARMISD::VMULLs: return "ARMISD::VMULLs";
818 case ARMISD::VMULLu: return "ARMISD::VMULLu";
819 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
820 case ARMISD::FMAX: return "ARMISD::FMAX";
821 case ARMISD::FMIN: return "ARMISD::FMIN";
822 case ARMISD::BFI: return "ARMISD::BFI";
823 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
827 /// getRegClassFor - Return the register class that should be used for the
828 /// specified value type.
829 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
830 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
831 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
832 // load / store 4 to 8 consecutive D registers.
833 if (Subtarget->hasNEON()) {
834 if (VT == MVT::v4i64)
835 return ARM::QQPRRegisterClass;
836 else if (VT == MVT::v8i64)
837 return ARM::QQQQPRRegisterClass;
839 return TargetLowering::getRegClassFor(VT);
842 // Create a fast isel object.
844 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
845 return ARM::createFastISel(funcInfo);
848 /// getFunctionAlignment - Return the Log2 alignment of this function.
849 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
850 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
853 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
854 /// be used for loads / stores from the global.
855 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
856 return (Subtarget->isThumb1Only() ? 127 : 4095);
859 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
860 unsigned NumVals = N->getNumValues();
862 return Sched::RegPressure;
864 for (unsigned i = 0; i != NumVals; ++i) {
865 EVT VT = N->getValueType(i);
866 if (VT == MVT::Flag || VT == MVT::Other)
868 if (VT.isFloatingPoint() || VT.isVector())
869 return Sched::Latency;
872 if (!N->isMachineOpcode())
873 return Sched::RegPressure;
875 // Load are scheduled for latency even if there instruction itinerary
877 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
878 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
880 if (TID.getNumDefs() == 0)
881 return Sched::RegPressure;
882 if (!Itins->isEmpty() &&
883 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
884 return Sched::Latency;
886 return Sched::RegPressure;
890 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
891 MachineFunction &MF) const {
892 switch (RC->getID()) {
895 case ARM::tGPRRegClassID:
896 return RegInfo->hasFP(MF) ? 4 : 5;
897 case ARM::GPRRegClassID: {
898 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
899 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
901 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
902 case ARM::DPRRegClassID:
907 //===----------------------------------------------------------------------===//
909 //===----------------------------------------------------------------------===//
911 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
912 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
914 default: llvm_unreachable("Unknown condition code!");
915 case ISD::SETNE: return ARMCC::NE;
916 case ISD::SETEQ: return ARMCC::EQ;
917 case ISD::SETGT: return ARMCC::GT;
918 case ISD::SETGE: return ARMCC::GE;
919 case ISD::SETLT: return ARMCC::LT;
920 case ISD::SETLE: return ARMCC::LE;
921 case ISD::SETUGT: return ARMCC::HI;
922 case ISD::SETUGE: return ARMCC::HS;
923 case ISD::SETULT: return ARMCC::LO;
924 case ISD::SETULE: return ARMCC::LS;
928 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
929 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
930 ARMCC::CondCodes &CondCode2) {
931 CondCode2 = ARMCC::AL;
933 default: llvm_unreachable("Unknown FP condition!");
935 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
937 case ISD::SETOGT: CondCode = ARMCC::GT; break;
939 case ISD::SETOGE: CondCode = ARMCC::GE; break;
940 case ISD::SETOLT: CondCode = ARMCC::MI; break;
941 case ISD::SETOLE: CondCode = ARMCC::LS; break;
942 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
943 case ISD::SETO: CondCode = ARMCC::VC; break;
944 case ISD::SETUO: CondCode = ARMCC::VS; break;
945 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
946 case ISD::SETUGT: CondCode = ARMCC::HI; break;
947 case ISD::SETUGE: CondCode = ARMCC::PL; break;
949 case ISD::SETULT: CondCode = ARMCC::LT; break;
951 case ISD::SETULE: CondCode = ARMCC::LE; break;
953 case ISD::SETUNE: CondCode = ARMCC::NE; break;
957 //===----------------------------------------------------------------------===//
958 // Calling Convention Implementation
959 //===----------------------------------------------------------------------===//
961 #include "ARMGenCallingConv.inc"
963 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
964 /// given CallingConvention value.
965 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
967 bool isVarArg) const {
970 llvm_unreachable("Unsupported calling convention");
971 case CallingConv::Fast:
972 if (Subtarget->hasVFP2() && !isVarArg) {
973 if (!Subtarget->isAAPCS_ABI())
974 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
975 // For AAPCS ABI targets, just use VFP variant of the calling convention.
976 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
979 case CallingConv::C: {
980 // Use target triple & subtarget features to do actual dispatch.
981 if (!Subtarget->isAAPCS_ABI())
982 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
983 else if (Subtarget->hasVFP2() &&
984 FloatABIType == FloatABI::Hard && !isVarArg)
985 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
986 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
988 case CallingConv::ARM_AAPCS_VFP:
989 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
990 case CallingConv::ARM_AAPCS:
991 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
992 case CallingConv::ARM_APCS:
993 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
997 /// LowerCallResult - Lower the result values of a call into the
998 /// appropriate copies out of appropriate physical registers.
1000 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1001 CallingConv::ID CallConv, bool isVarArg,
1002 const SmallVectorImpl<ISD::InputArg> &Ins,
1003 DebugLoc dl, SelectionDAG &DAG,
1004 SmallVectorImpl<SDValue> &InVals) const {
1006 // Assign locations to each value returned by this call.
1007 SmallVector<CCValAssign, 16> RVLocs;
1008 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1009 RVLocs, *DAG.getContext());
1010 CCInfo.AnalyzeCallResult(Ins,
1011 CCAssignFnForNode(CallConv, /* Return*/ true,
1014 // Copy all of the result registers out of their specified physreg.
1015 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1016 CCValAssign VA = RVLocs[i];
1019 if (VA.needsCustom()) {
1020 // Handle f64 or half of a v2f64.
1021 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1023 Chain = Lo.getValue(1);
1024 InFlag = Lo.getValue(2);
1025 VA = RVLocs[++i]; // skip ahead to next loc
1026 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1028 Chain = Hi.getValue(1);
1029 InFlag = Hi.getValue(2);
1030 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1032 if (VA.getLocVT() == MVT::v2f64) {
1033 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1034 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1035 DAG.getConstant(0, MVT::i32));
1037 VA = RVLocs[++i]; // skip ahead to next loc
1038 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1039 Chain = Lo.getValue(1);
1040 InFlag = Lo.getValue(2);
1041 VA = RVLocs[++i]; // skip ahead to next loc
1042 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1043 Chain = Hi.getValue(1);
1044 InFlag = Hi.getValue(2);
1045 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1046 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1047 DAG.getConstant(1, MVT::i32));
1050 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1052 Chain = Val.getValue(1);
1053 InFlag = Val.getValue(2);
1056 switch (VA.getLocInfo()) {
1057 default: llvm_unreachable("Unknown loc info!");
1058 case CCValAssign::Full: break;
1059 case CCValAssign::BCvt:
1060 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1064 InVals.push_back(Val);
1070 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1071 /// by "Src" to address "Dst" of size "Size". Alignment information is
1072 /// specified by the specific parameter attribute. The copy will be passed as
1073 /// a byval function parameter.
1074 /// Sometimes what we are copying is the end of a larger object, the part that
1075 /// does not fit in registers.
1077 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1078 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1080 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1081 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1082 /*isVolatile=*/false, /*AlwaysInline=*/false,
1083 MachinePointerInfo(0), MachinePointerInfo(0));
1086 /// LowerMemOpCallTo - Store the argument to the stack.
1088 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1089 SDValue StackPtr, SDValue Arg,
1090 DebugLoc dl, SelectionDAG &DAG,
1091 const CCValAssign &VA,
1092 ISD::ArgFlagsTy Flags) const {
1093 unsigned LocMemOffset = VA.getLocMemOffset();
1094 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1095 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1096 if (Flags.isByVal())
1097 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1099 return DAG.getStore(Chain, dl, Arg, PtrOff,
1100 MachinePointerInfo::getStack(LocMemOffset),
1104 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1105 SDValue Chain, SDValue &Arg,
1106 RegsToPassVector &RegsToPass,
1107 CCValAssign &VA, CCValAssign &NextVA,
1109 SmallVector<SDValue, 8> &MemOpChains,
1110 ISD::ArgFlagsTy Flags) const {
1112 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1113 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1114 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1116 if (NextVA.isRegLoc())
1117 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1119 assert(NextVA.isMemLoc());
1120 if (StackPtr.getNode() == 0)
1121 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1123 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1129 /// LowerCall - Lowering a call into a callseq_start <-
1130 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1133 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1134 CallingConv::ID CallConv, bool isVarArg,
1136 const SmallVectorImpl<ISD::OutputArg> &Outs,
1137 const SmallVectorImpl<SDValue> &OutVals,
1138 const SmallVectorImpl<ISD::InputArg> &Ins,
1139 DebugLoc dl, SelectionDAG &DAG,
1140 SmallVectorImpl<SDValue> &InVals) const {
1141 MachineFunction &MF = DAG.getMachineFunction();
1142 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1143 bool IsSibCall = false;
1144 // Temporarily disable tail calls so things don't break.
1145 if (!EnableARMTailCalls)
1148 // Check if it's really possible to do a tail call.
1149 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1150 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1151 Outs, OutVals, Ins, DAG);
1152 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1153 // detected sibcalls.
1160 // Analyze operands of the call, assigning locations to each operand.
1161 SmallVector<CCValAssign, 16> ArgLocs;
1162 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1164 CCInfo.AnalyzeCallOperands(Outs,
1165 CCAssignFnForNode(CallConv, /* Return*/ false,
1168 // Get a count of how many bytes are to be pushed on the stack.
1169 unsigned NumBytes = CCInfo.getNextStackOffset();
1171 // For tail calls, memory operands are available in our caller's stack.
1175 // Adjust the stack pointer for the new arguments...
1176 // These operations are automatically eliminated by the prolog/epilog pass
1178 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1180 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1182 RegsToPassVector RegsToPass;
1183 SmallVector<SDValue, 8> MemOpChains;
1185 // Walk the register/memloc assignments, inserting copies/loads. In the case
1186 // of tail call optimization, arguments are handled later.
1187 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1189 ++i, ++realArgIdx) {
1190 CCValAssign &VA = ArgLocs[i];
1191 SDValue Arg = OutVals[realArgIdx];
1192 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1194 // Promote the value if needed.
1195 switch (VA.getLocInfo()) {
1196 default: llvm_unreachable("Unknown loc info!");
1197 case CCValAssign::Full: break;
1198 case CCValAssign::SExt:
1199 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1201 case CCValAssign::ZExt:
1202 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1204 case CCValAssign::AExt:
1205 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1207 case CCValAssign::BCvt:
1208 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1212 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1213 if (VA.needsCustom()) {
1214 if (VA.getLocVT() == MVT::v2f64) {
1215 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1216 DAG.getConstant(0, MVT::i32));
1217 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1218 DAG.getConstant(1, MVT::i32));
1220 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1221 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1223 VA = ArgLocs[++i]; // skip ahead to next loc
1224 if (VA.isRegLoc()) {
1225 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1226 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1228 assert(VA.isMemLoc());
1230 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1231 dl, DAG, VA, Flags));
1234 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1235 StackPtr, MemOpChains, Flags);
1237 } else if (VA.isRegLoc()) {
1238 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1239 } else if (!IsSibCall) {
1240 assert(VA.isMemLoc());
1242 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1243 dl, DAG, VA, Flags));
1247 if (!MemOpChains.empty())
1248 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1249 &MemOpChains[0], MemOpChains.size());
1251 // Build a sequence of copy-to-reg nodes chained together with token chain
1252 // and flag operands which copy the outgoing args into the appropriate regs.
1254 // Tail call byval lowering might overwrite argument registers so in case of
1255 // tail call optimization the copies to registers are lowered later.
1257 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1258 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1259 RegsToPass[i].second, InFlag);
1260 InFlag = Chain.getValue(1);
1263 // For tail calls lower the arguments to the 'real' stack slot.
1265 // Force all the incoming stack arguments to be loaded from the stack
1266 // before any new outgoing arguments are stored to the stack, because the
1267 // outgoing stack slots may alias the incoming argument stack slots, and
1268 // the alias isn't otherwise explicit. This is slightly more conservative
1269 // than necessary, because it means that each store effectively depends
1270 // on every argument instead of just those arguments it would clobber.
1272 // Do not flag preceeding copytoreg stuff together with the following stuff.
1274 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1275 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1276 RegsToPass[i].second, InFlag);
1277 InFlag = Chain.getValue(1);
1282 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1283 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1284 // node so that legalize doesn't hack it.
1285 bool isDirect = false;
1286 bool isARMFunc = false;
1287 bool isLocalARMFunc = false;
1288 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1290 if (EnableARMLongCalls) {
1291 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1292 && "long-calls with non-static relocation model!");
1293 // Handle a global address or an external symbol. If it's not one of
1294 // those, the target's already in a register, so we don't need to do
1296 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1297 const GlobalValue *GV = G->getGlobal();
1298 // Create a constant pool entry for the callee address
1299 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1300 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1303 // Get the address of the callee into a register
1304 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1305 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1306 Callee = DAG.getLoad(getPointerTy(), dl,
1307 DAG.getEntryNode(), CPAddr,
1308 MachinePointerInfo::getConstantPool(),
1310 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1311 const char *Sym = S->getSymbol();
1313 // Create a constant pool entry for the callee address
1314 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1315 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1316 Sym, ARMPCLabelIndex, 0);
1317 // Get the address of the callee into a register
1318 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1319 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1320 Callee = DAG.getLoad(getPointerTy(), dl,
1321 DAG.getEntryNode(), CPAddr,
1322 MachinePointerInfo::getConstantPool(),
1325 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1326 const GlobalValue *GV = G->getGlobal();
1328 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1329 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1330 getTargetMachine().getRelocationModel() != Reloc::Static;
1331 isARMFunc = !Subtarget->isThumb() || isStub;
1332 // ARM call to a local ARM function is predicable.
1333 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1334 // tBX takes a register source operand.
1335 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1336 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1337 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1340 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1341 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1342 Callee = DAG.getLoad(getPointerTy(), dl,
1343 DAG.getEntryNode(), CPAddr,
1344 MachinePointerInfo::getConstantPool(),
1346 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1347 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1348 getPointerTy(), Callee, PICLabel);
1350 // On ELF targets for PIC code, direct calls should go through the PLT
1351 unsigned OpFlags = 0;
1352 if (Subtarget->isTargetELF() &&
1353 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1354 OpFlags = ARMII::MO_PLT;
1355 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1357 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1359 bool isStub = Subtarget->isTargetDarwin() &&
1360 getTargetMachine().getRelocationModel() != Reloc::Static;
1361 isARMFunc = !Subtarget->isThumb() || isStub;
1362 // tBX takes a register source operand.
1363 const char *Sym = S->getSymbol();
1364 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1365 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1366 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1367 Sym, ARMPCLabelIndex, 4);
1368 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1369 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1370 Callee = DAG.getLoad(getPointerTy(), dl,
1371 DAG.getEntryNode(), CPAddr,
1372 MachinePointerInfo::getConstantPool(),
1374 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1375 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1376 getPointerTy(), Callee, PICLabel);
1378 unsigned OpFlags = 0;
1379 // On ELF targets for PIC code, direct calls should go through the PLT
1380 if (Subtarget->isTargetELF() &&
1381 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1382 OpFlags = ARMII::MO_PLT;
1383 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1387 // FIXME: handle tail calls differently.
1389 if (Subtarget->isThumb()) {
1390 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1391 CallOpc = ARMISD::CALL_NOLINK;
1393 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1395 CallOpc = (isDirect || Subtarget->hasV5TOps())
1396 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1397 : ARMISD::CALL_NOLINK;
1400 std::vector<SDValue> Ops;
1401 Ops.push_back(Chain);
1402 Ops.push_back(Callee);
1404 // Add argument registers to the end of the list so that they are known live
1406 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1407 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1408 RegsToPass[i].second.getValueType()));
1410 if (InFlag.getNode())
1411 Ops.push_back(InFlag);
1413 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1415 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1417 // Returns a chain and a flag for retval copy to use.
1418 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1419 InFlag = Chain.getValue(1);
1421 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1422 DAG.getIntPtrConstant(0, true), InFlag);
1424 InFlag = Chain.getValue(1);
1426 // Handle result values, copying them out of physregs into vregs that we
1428 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1432 /// MatchingStackOffset - Return true if the given stack call argument is
1433 /// already available in the same position (relatively) of the caller's
1434 /// incoming argument stack.
1436 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1437 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1438 const ARMInstrInfo *TII) {
1439 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1441 if (Arg.getOpcode() == ISD::CopyFromReg) {
1442 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1443 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1445 MachineInstr *Def = MRI->getVRegDef(VR);
1448 if (!Flags.isByVal()) {
1449 if (!TII->isLoadFromStackSlot(Def, FI))
1454 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1455 if (Flags.isByVal())
1456 // ByVal argument is passed in as a pointer but it's now being
1457 // dereferenced. e.g.
1458 // define @foo(%struct.X* %A) {
1459 // tail call @bar(%struct.X* byval %A)
1462 SDValue Ptr = Ld->getBasePtr();
1463 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1466 FI = FINode->getIndex();
1470 assert(FI != INT_MAX);
1471 if (!MFI->isFixedObjectIndex(FI))
1473 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1476 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1477 /// for tail call optimization. Targets which want to do tail call
1478 /// optimization should implement this function.
1480 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1481 CallingConv::ID CalleeCC,
1483 bool isCalleeStructRet,
1484 bool isCallerStructRet,
1485 const SmallVectorImpl<ISD::OutputArg> &Outs,
1486 const SmallVectorImpl<SDValue> &OutVals,
1487 const SmallVectorImpl<ISD::InputArg> &Ins,
1488 SelectionDAG& DAG) const {
1489 const Function *CallerF = DAG.getMachineFunction().getFunction();
1490 CallingConv::ID CallerCC = CallerF->getCallingConv();
1491 bool CCMatch = CallerCC == CalleeCC;
1493 // Look for obvious safe cases to perform tail call optimization that do not
1494 // require ABI changes. This is what gcc calls sibcall.
1496 // Do not sibcall optimize vararg calls unless the call site is not passing
1498 if (isVarArg && !Outs.empty())
1501 // Also avoid sibcall optimization if either caller or callee uses struct
1502 // return semantics.
1503 if (isCalleeStructRet || isCallerStructRet)
1506 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1507 // emitEpilogue is not ready for them.
1508 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1509 // LR. This means if we need to reload LR, it takes an extra instructions,
1510 // which outweighs the value of the tail call; but here we don't know yet
1511 // whether LR is going to be used. Probably the right approach is to
1512 // generate the tail call here and turn it back into CALL/RET in
1513 // emitEpilogue if LR is used.
1514 if (Subtarget->isThumb1Only())
1517 // For the moment, we can only do this to functions defined in this
1518 // compilation, or to indirect calls. A Thumb B to an ARM function,
1519 // or vice versa, is not easily fixed up in the linker unlike BL.
1520 // (We could do this by loading the address of the callee into a register;
1521 // that is an extra instruction over the direct call and burns a register
1522 // as well, so is not likely to be a win.)
1524 // It might be safe to remove this restriction on non-Darwin.
1526 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1527 // but we need to make sure there are enough registers; the only valid
1528 // registers are the 4 used for parameters. We don't currently do this
1530 if (isa<ExternalSymbolSDNode>(Callee))
1533 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1534 const GlobalValue *GV = G->getGlobal();
1535 if (GV->isDeclaration() || GV->isWeakForLinker())
1539 // If the calling conventions do not match, then we'd better make sure the
1540 // results are returned in the same way as what the caller expects.
1542 SmallVector<CCValAssign, 16> RVLocs1;
1543 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1544 RVLocs1, *DAG.getContext());
1545 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1547 SmallVector<CCValAssign, 16> RVLocs2;
1548 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1549 RVLocs2, *DAG.getContext());
1550 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1552 if (RVLocs1.size() != RVLocs2.size())
1554 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1555 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1557 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1559 if (RVLocs1[i].isRegLoc()) {
1560 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1563 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1569 // If the callee takes no arguments then go on to check the results of the
1571 if (!Outs.empty()) {
1572 // Check if stack adjustment is needed. For now, do not do this if any
1573 // argument is passed on the stack.
1574 SmallVector<CCValAssign, 16> ArgLocs;
1575 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1576 ArgLocs, *DAG.getContext());
1577 CCInfo.AnalyzeCallOperands(Outs,
1578 CCAssignFnForNode(CalleeCC, false, isVarArg));
1579 if (CCInfo.getNextStackOffset()) {
1580 MachineFunction &MF = DAG.getMachineFunction();
1582 // Check if the arguments are already laid out in the right way as
1583 // the caller's fixed stack objects.
1584 MachineFrameInfo *MFI = MF.getFrameInfo();
1585 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1586 const ARMInstrInfo *TII =
1587 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1588 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1590 ++i, ++realArgIdx) {
1591 CCValAssign &VA = ArgLocs[i];
1592 EVT RegVT = VA.getLocVT();
1593 SDValue Arg = OutVals[realArgIdx];
1594 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1595 if (VA.getLocInfo() == CCValAssign::Indirect)
1597 if (VA.needsCustom()) {
1598 // f64 and vector types are split into multiple registers or
1599 // register/stack-slot combinations. The types will not match
1600 // the registers; give up on memory f64 refs until we figure
1601 // out what to do about this.
1604 if (!ArgLocs[++i].isRegLoc())
1606 if (RegVT == MVT::v2f64) {
1607 if (!ArgLocs[++i].isRegLoc())
1609 if (!ArgLocs[++i].isRegLoc())
1612 } else if (!VA.isRegLoc()) {
1613 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1625 ARMTargetLowering::LowerReturn(SDValue Chain,
1626 CallingConv::ID CallConv, bool isVarArg,
1627 const SmallVectorImpl<ISD::OutputArg> &Outs,
1628 const SmallVectorImpl<SDValue> &OutVals,
1629 DebugLoc dl, SelectionDAG &DAG) const {
1631 // CCValAssign - represent the assignment of the return value to a location.
1632 SmallVector<CCValAssign, 16> RVLocs;
1634 // CCState - Info about the registers and stack slots.
1635 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1638 // Analyze outgoing return values.
1639 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1642 // If this is the first return lowered for this function, add
1643 // the regs to the liveout set for the function.
1644 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1645 for (unsigned i = 0; i != RVLocs.size(); ++i)
1646 if (RVLocs[i].isRegLoc())
1647 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1652 // Copy the result values into the output registers.
1653 for (unsigned i = 0, realRVLocIdx = 0;
1655 ++i, ++realRVLocIdx) {
1656 CCValAssign &VA = RVLocs[i];
1657 assert(VA.isRegLoc() && "Can only return in registers!");
1659 SDValue Arg = OutVals[realRVLocIdx];
1661 switch (VA.getLocInfo()) {
1662 default: llvm_unreachable("Unknown loc info!");
1663 case CCValAssign::Full: break;
1664 case CCValAssign::BCvt:
1665 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1669 if (VA.needsCustom()) {
1670 if (VA.getLocVT() == MVT::v2f64) {
1671 // Extract the first half and return it in two registers.
1672 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1673 DAG.getConstant(0, MVT::i32));
1674 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1675 DAG.getVTList(MVT::i32, MVT::i32), Half);
1677 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1678 Flag = Chain.getValue(1);
1679 VA = RVLocs[++i]; // skip ahead to next loc
1680 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1681 HalfGPRs.getValue(1), Flag);
1682 Flag = Chain.getValue(1);
1683 VA = RVLocs[++i]; // skip ahead to next loc
1685 // Extract the 2nd half and fall through to handle it as an f64 value.
1686 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1687 DAG.getConstant(1, MVT::i32));
1689 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1691 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1692 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1693 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1694 Flag = Chain.getValue(1);
1695 VA = RVLocs[++i]; // skip ahead to next loc
1696 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1699 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1701 // Guarantee that all emitted copies are
1702 // stuck together, avoiding something bad.
1703 Flag = Chain.getValue(1);
1708 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1710 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1715 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1716 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1717 // one of the above mentioned nodes. It has to be wrapped because otherwise
1718 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1719 // be used to form addressing mode. These wrapped nodes will be selected
1721 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1722 EVT PtrVT = Op.getValueType();
1723 // FIXME there is no actual debug info here
1724 DebugLoc dl = Op.getDebugLoc();
1725 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1727 if (CP->isMachineConstantPoolEntry())
1728 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1729 CP->getAlignment());
1731 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1732 CP->getAlignment());
1733 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1736 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1737 return MachineJumpTableInfo::EK_Inline;
1740 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1741 SelectionDAG &DAG) const {
1742 MachineFunction &MF = DAG.getMachineFunction();
1743 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1744 unsigned ARMPCLabelIndex = 0;
1745 DebugLoc DL = Op.getDebugLoc();
1746 EVT PtrVT = getPointerTy();
1747 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1748 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1750 if (RelocM == Reloc::Static) {
1751 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1753 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1754 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1755 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1756 ARMCP::CPBlockAddress,
1758 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1760 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1761 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1762 MachinePointerInfo::getConstantPool(),
1764 if (RelocM == Reloc::Static)
1766 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1767 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1770 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1772 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1773 SelectionDAG &DAG) const {
1774 DebugLoc dl = GA->getDebugLoc();
1775 EVT PtrVT = getPointerTy();
1776 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1777 MachineFunction &MF = DAG.getMachineFunction();
1778 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1779 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1780 ARMConstantPoolValue *CPV =
1781 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1782 ARMCP::CPValue, PCAdj, "tlsgd", true);
1783 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1784 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1785 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1786 MachinePointerInfo::getConstantPool(),
1788 SDValue Chain = Argument.getValue(1);
1790 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1791 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1793 // call __tls_get_addr.
1796 Entry.Node = Argument;
1797 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1798 Args.push_back(Entry);
1799 // FIXME: is there useful debug info available here?
1800 std::pair<SDValue, SDValue> CallResult =
1801 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1802 false, false, false, false,
1803 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1804 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1805 return CallResult.first;
1808 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1809 // "local exec" model.
1811 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1812 SelectionDAG &DAG) const {
1813 const GlobalValue *GV = GA->getGlobal();
1814 DebugLoc dl = GA->getDebugLoc();
1816 SDValue Chain = DAG.getEntryNode();
1817 EVT PtrVT = getPointerTy();
1818 // Get the Thread Pointer
1819 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1821 if (GV->isDeclaration()) {
1822 MachineFunction &MF = DAG.getMachineFunction();
1823 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1824 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1825 // Initial exec model.
1826 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1827 ARMConstantPoolValue *CPV =
1828 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1829 ARMCP::CPValue, PCAdj, "gottpoff", true);
1830 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1831 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1832 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1833 MachinePointerInfo::getConstantPool(),
1835 Chain = Offset.getValue(1);
1837 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1838 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1840 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1841 MachinePointerInfo::getConstantPool(),
1845 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1846 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1847 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1848 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1849 MachinePointerInfo::getConstantPool(),
1853 // The address of the thread local variable is the add of the thread
1854 // pointer with the offset of the variable.
1855 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1859 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1860 // TODO: implement the "local dynamic" model
1861 assert(Subtarget->isTargetELF() &&
1862 "TLS not implemented for non-ELF targets");
1863 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1864 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1865 // otherwise use the "Local Exec" TLS Model
1866 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1867 return LowerToTLSGeneralDynamicModel(GA, DAG);
1869 return LowerToTLSExecModels(GA, DAG);
1872 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1873 SelectionDAG &DAG) const {
1874 EVT PtrVT = getPointerTy();
1875 DebugLoc dl = Op.getDebugLoc();
1876 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1877 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1878 if (RelocM == Reloc::PIC_) {
1879 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1880 ARMConstantPoolValue *CPV =
1881 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1882 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1883 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1884 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1886 MachinePointerInfo::getConstantPool(),
1888 SDValue Chain = Result.getValue(1);
1889 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1890 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1892 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1893 MachinePointerInfo::getGOT(), false, false, 0);
1896 // If we have T2 ops, we can materialize the address directly via movt/movw
1897 // pair. This is always cheaper.
1898 if (Subtarget->useMovt()) {
1899 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1900 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1902 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1903 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1904 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1905 MachinePointerInfo::getConstantPool(),
1911 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1912 SelectionDAG &DAG) const {
1913 MachineFunction &MF = DAG.getMachineFunction();
1914 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1915 unsigned ARMPCLabelIndex = 0;
1916 EVT PtrVT = getPointerTy();
1917 DebugLoc dl = Op.getDebugLoc();
1918 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1919 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1921 if (RelocM == Reloc::Static)
1922 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1924 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1925 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1926 ARMConstantPoolValue *CPV =
1927 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1928 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1930 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1932 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1933 MachinePointerInfo::getConstantPool(),
1935 SDValue Chain = Result.getValue(1);
1937 if (RelocM == Reloc::PIC_) {
1938 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1939 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1942 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1943 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1949 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1950 SelectionDAG &DAG) const {
1951 assert(Subtarget->isTargetELF() &&
1952 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1953 MachineFunction &MF = DAG.getMachineFunction();
1954 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1955 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1956 EVT PtrVT = getPointerTy();
1957 DebugLoc dl = Op.getDebugLoc();
1958 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1959 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1960 "_GLOBAL_OFFSET_TABLE_",
1961 ARMPCLabelIndex, PCAdj);
1962 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1963 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1964 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1965 MachinePointerInfo::getConstantPool(),
1967 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1968 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1972 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
1974 DebugLoc dl = Op.getDebugLoc();
1975 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
1976 Op.getOperand(0), Op.getOperand(1));
1980 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1981 DebugLoc dl = Op.getDebugLoc();
1982 SDValue Val = DAG.getConstant(0, MVT::i32);
1983 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1984 Op.getOperand(1), Val);
1988 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1989 DebugLoc dl = Op.getDebugLoc();
1990 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1991 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1995 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1996 const ARMSubtarget *Subtarget) const {
1997 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1998 DebugLoc dl = Op.getDebugLoc();
2000 default: return SDValue(); // Don't custom lower most intrinsics.
2001 case Intrinsic::arm_thread_pointer: {
2002 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2003 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2005 case Intrinsic::eh_sjlj_lsda: {
2006 MachineFunction &MF = DAG.getMachineFunction();
2007 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2008 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
2009 EVT PtrVT = getPointerTy();
2010 DebugLoc dl = Op.getDebugLoc();
2011 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2013 unsigned PCAdj = (RelocM != Reloc::PIC_)
2014 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2015 ARMConstantPoolValue *CPV =
2016 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2017 ARMCP::CPLSDA, PCAdj);
2018 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2019 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2021 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2022 MachinePointerInfo::getConstantPool(),
2025 if (RelocM == Reloc::PIC_) {
2026 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2027 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2034 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2035 const ARMSubtarget *Subtarget) {
2036 DebugLoc dl = Op.getDebugLoc();
2037 if (!Subtarget->hasDataBarrier()) {
2038 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2039 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2041 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2042 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2043 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2044 DAG.getConstant(0, MVT::i32));
2047 SDValue Op5 = Op.getOperand(5);
2048 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2049 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2050 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2051 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2053 ARM_MB::MemBOpt DMBOpt;
2054 if (isDeviceBarrier)
2055 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2057 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2058 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2059 DAG.getConstant(DMBOpt, MVT::i32));
2062 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2063 const ARMSubtarget *Subtarget) {
2064 // ARM pre v5TE and Thumb1 does not have preload instructions.
2065 if (!(Subtarget->isThumb2() ||
2066 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2067 // Just preserve the chain.
2068 return Op.getOperand(0);
2070 DebugLoc dl = Op.getDebugLoc();
2071 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2073 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2074 // ARMv7 with MP extension has PLDW.
2075 return Op.getOperand(0);
2077 if (Subtarget->isThumb())
2079 isRead = ~isRead & 1;
2080 unsigned isData = Subtarget->isThumb() ? 0 : 1;
2082 // Currently there is no intrinsic that matches pli.
2083 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2084 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2085 DAG.getConstant(isData, MVT::i32));
2088 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2089 MachineFunction &MF = DAG.getMachineFunction();
2090 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2092 // vastart just stores the address of the VarArgsFrameIndex slot into the
2093 // memory location argument.
2094 DebugLoc dl = Op.getDebugLoc();
2095 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2096 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2097 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2098 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2099 MachinePointerInfo(SV), false, false, 0);
2103 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2104 SDValue &Root, SelectionDAG &DAG,
2105 DebugLoc dl) const {
2106 MachineFunction &MF = DAG.getMachineFunction();
2107 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2109 TargetRegisterClass *RC;
2110 if (AFI->isThumb1OnlyFunction())
2111 RC = ARM::tGPRRegisterClass;
2113 RC = ARM::GPRRegisterClass;
2115 // Transform the arguments stored in physical registers into virtual ones.
2116 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2117 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2120 if (NextVA.isMemLoc()) {
2121 MachineFrameInfo *MFI = MF.getFrameInfo();
2122 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2124 // Create load node to retrieve arguments from the stack.
2125 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2126 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2127 MachinePointerInfo::getFixedStack(FI),
2130 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2131 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2134 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2138 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2139 CallingConv::ID CallConv, bool isVarArg,
2140 const SmallVectorImpl<ISD::InputArg>
2142 DebugLoc dl, SelectionDAG &DAG,
2143 SmallVectorImpl<SDValue> &InVals)
2146 MachineFunction &MF = DAG.getMachineFunction();
2147 MachineFrameInfo *MFI = MF.getFrameInfo();
2149 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2151 // Assign locations to all of the incoming arguments.
2152 SmallVector<CCValAssign, 16> ArgLocs;
2153 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2155 CCInfo.AnalyzeFormalArguments(Ins,
2156 CCAssignFnForNode(CallConv, /* Return*/ false,
2159 SmallVector<SDValue, 16> ArgValues;
2161 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2162 CCValAssign &VA = ArgLocs[i];
2164 // Arguments stored in registers.
2165 if (VA.isRegLoc()) {
2166 EVT RegVT = VA.getLocVT();
2169 if (VA.needsCustom()) {
2170 // f64 and vector types are split up into multiple registers or
2171 // combinations of registers and stack slots.
2172 if (VA.getLocVT() == MVT::v2f64) {
2173 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2175 VA = ArgLocs[++i]; // skip ahead to next loc
2177 if (VA.isMemLoc()) {
2178 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2179 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2180 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2181 MachinePointerInfo::getFixedStack(FI),
2184 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2187 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2188 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2189 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2190 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2191 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2193 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2196 TargetRegisterClass *RC;
2198 if (RegVT == MVT::f32)
2199 RC = ARM::SPRRegisterClass;
2200 else if (RegVT == MVT::f64)
2201 RC = ARM::DPRRegisterClass;
2202 else if (RegVT == MVT::v2f64)
2203 RC = ARM::QPRRegisterClass;
2204 else if (RegVT == MVT::i32)
2205 RC = (AFI->isThumb1OnlyFunction() ?
2206 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2208 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2210 // Transform the arguments in physical registers into virtual ones.
2211 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2212 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2215 // If this is an 8 or 16-bit value, it is really passed promoted
2216 // to 32 bits. Insert an assert[sz]ext to capture this, then
2217 // truncate to the right size.
2218 switch (VA.getLocInfo()) {
2219 default: llvm_unreachable("Unknown loc info!");
2220 case CCValAssign::Full: break;
2221 case CCValAssign::BCvt:
2222 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2224 case CCValAssign::SExt:
2225 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2226 DAG.getValueType(VA.getValVT()));
2227 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2229 case CCValAssign::ZExt:
2230 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2231 DAG.getValueType(VA.getValVT()));
2232 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2236 InVals.push_back(ArgValue);
2238 } else { // VA.isRegLoc()
2241 assert(VA.isMemLoc());
2242 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2244 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2245 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2247 // Create load nodes to retrieve arguments from the stack.
2248 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2249 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2250 MachinePointerInfo::getFixedStack(FI),
2257 static const unsigned GPRArgRegs[] = {
2258 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2261 unsigned NumGPRs = CCInfo.getFirstUnallocated
2262 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2264 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2265 unsigned VARegSize = (4 - NumGPRs) * 4;
2266 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2267 unsigned ArgOffset = CCInfo.getNextStackOffset();
2268 if (VARegSaveSize) {
2269 // If this function is vararg, store any remaining integer argument regs
2270 // to their spots on the stack so that they may be loaded by deferencing
2271 // the result of va_next.
2272 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2273 AFI->setVarArgsFrameIndex(
2274 MFI->CreateFixedObject(VARegSaveSize,
2275 ArgOffset + VARegSaveSize - VARegSize,
2277 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2280 SmallVector<SDValue, 4> MemOps;
2281 for (; NumGPRs < 4; ++NumGPRs) {
2282 TargetRegisterClass *RC;
2283 if (AFI->isThumb1OnlyFunction())
2284 RC = ARM::tGPRRegisterClass;
2286 RC = ARM::GPRRegisterClass;
2288 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2289 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2291 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2292 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2294 MemOps.push_back(Store);
2295 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2296 DAG.getConstant(4, getPointerTy()));
2298 if (!MemOps.empty())
2299 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2300 &MemOps[0], MemOps.size());
2302 // This will point to the next argument passed via stack.
2303 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2309 /// isFloatingPointZero - Return true if this is +0.0.
2310 static bool isFloatingPointZero(SDValue Op) {
2311 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2312 return CFP->getValueAPF().isPosZero();
2313 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2314 // Maybe this has already been legalized into the constant pool?
2315 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2316 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2317 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2318 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2319 return CFP->getValueAPF().isPosZero();
2325 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2326 /// the given operands.
2328 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2329 SDValue &ARMcc, SelectionDAG &DAG,
2330 DebugLoc dl) const {
2331 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2332 unsigned C = RHSC->getZExtValue();
2333 if (!isLegalICmpImmediate(C)) {
2334 // Constant does not fit, try adjusting it by one?
2339 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2340 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2341 RHS = DAG.getConstant(C-1, MVT::i32);
2346 if (C != 0 && isLegalICmpImmediate(C-1)) {
2347 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2348 RHS = DAG.getConstant(C-1, MVT::i32);
2353 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2354 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2355 RHS = DAG.getConstant(C+1, MVT::i32);
2360 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2361 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2362 RHS = DAG.getConstant(C+1, MVT::i32);
2369 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2370 ARMISD::NodeType CompareType;
2373 CompareType = ARMISD::CMP;
2378 CompareType = ARMISD::CMPZ;
2381 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2382 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2385 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2387 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2388 DebugLoc dl) const {
2390 if (!isFloatingPointZero(RHS))
2391 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2393 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2394 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2397 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2398 SDValue Cond = Op.getOperand(0);
2399 SDValue SelectTrue = Op.getOperand(1);
2400 SDValue SelectFalse = Op.getOperand(2);
2401 DebugLoc dl = Op.getDebugLoc();
2405 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2406 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2408 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2409 const ConstantSDNode *CMOVTrue =
2410 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2411 const ConstantSDNode *CMOVFalse =
2412 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2414 if (CMOVTrue && CMOVFalse) {
2415 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2416 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2420 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2422 False = SelectFalse;
2423 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2428 if (True.getNode() && False.getNode()) {
2429 EVT VT = Cond.getValueType();
2430 SDValue ARMcc = Cond.getOperand(2);
2431 SDValue CCR = Cond.getOperand(3);
2432 SDValue Cmp = Cond.getOperand(4);
2433 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2438 return DAG.getSelectCC(dl, Cond,
2439 DAG.getConstant(0, Cond.getValueType()),
2440 SelectTrue, SelectFalse, ISD::SETNE);
2443 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2444 EVT VT = Op.getValueType();
2445 SDValue LHS = Op.getOperand(0);
2446 SDValue RHS = Op.getOperand(1);
2447 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2448 SDValue TrueVal = Op.getOperand(2);
2449 SDValue FalseVal = Op.getOperand(3);
2450 DebugLoc dl = Op.getDebugLoc();
2452 if (LHS.getValueType() == MVT::i32) {
2454 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2455 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2456 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2459 ARMCC::CondCodes CondCode, CondCode2;
2460 FPCCToARMCC(CC, CondCode, CondCode2);
2462 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2463 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2464 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2465 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2467 if (CondCode2 != ARMCC::AL) {
2468 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2469 // FIXME: Needs another CMP because flag can have but one use.
2470 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2471 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2472 Result, TrueVal, ARMcc2, CCR, Cmp2);
2477 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2478 /// to morph to an integer compare sequence.
2479 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2480 const ARMSubtarget *Subtarget) {
2481 SDNode *N = Op.getNode();
2482 if (!N->hasOneUse())
2483 // Otherwise it requires moving the value from fp to integer registers.
2485 if (!N->getNumValues())
2487 EVT VT = Op.getValueType();
2488 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2489 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2490 // vmrs are very slow, e.g. cortex-a8.
2493 if (isFloatingPointZero(Op)) {
2497 return ISD::isNormalLoad(N);
2500 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2501 if (isFloatingPointZero(Op))
2502 return DAG.getConstant(0, MVT::i32);
2504 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2505 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2506 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2507 Ld->isVolatile(), Ld->isNonTemporal(),
2508 Ld->getAlignment());
2510 llvm_unreachable("Unknown VFP cmp argument!");
2513 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2514 SDValue &RetVal1, SDValue &RetVal2) {
2515 if (isFloatingPointZero(Op)) {
2516 RetVal1 = DAG.getConstant(0, MVT::i32);
2517 RetVal2 = DAG.getConstant(0, MVT::i32);
2521 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2522 SDValue Ptr = Ld->getBasePtr();
2523 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2524 Ld->getChain(), Ptr,
2525 Ld->getPointerInfo(),
2526 Ld->isVolatile(), Ld->isNonTemporal(),
2527 Ld->getAlignment());
2529 EVT PtrType = Ptr.getValueType();
2530 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2531 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2532 PtrType, Ptr, DAG.getConstant(4, PtrType));
2533 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2534 Ld->getChain(), NewPtr,
2535 Ld->getPointerInfo().getWithOffset(4),
2536 Ld->isVolatile(), Ld->isNonTemporal(),
2541 llvm_unreachable("Unknown VFP cmp argument!");
2544 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2545 /// f32 and even f64 comparisons to integer ones.
2547 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2548 SDValue Chain = Op.getOperand(0);
2549 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2550 SDValue LHS = Op.getOperand(2);
2551 SDValue RHS = Op.getOperand(3);
2552 SDValue Dest = Op.getOperand(4);
2553 DebugLoc dl = Op.getDebugLoc();
2555 bool SeenZero = false;
2556 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2557 canChangeToInt(RHS, SeenZero, Subtarget) &&
2558 // If one of the operand is zero, it's safe to ignore the NaN case since
2559 // we only care about equality comparisons.
2560 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2561 // If unsafe fp math optimization is enabled and there are no othter uses of
2562 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2563 // to an integer comparison.
2564 if (CC == ISD::SETOEQ)
2566 else if (CC == ISD::SETUNE)
2570 if (LHS.getValueType() == MVT::f32) {
2571 LHS = bitcastf32Toi32(LHS, DAG);
2572 RHS = bitcastf32Toi32(RHS, DAG);
2573 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2574 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2575 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2576 Chain, Dest, ARMcc, CCR, Cmp);
2581 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2582 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2583 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2584 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2585 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2586 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2587 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2593 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2594 SDValue Chain = Op.getOperand(0);
2595 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2596 SDValue LHS = Op.getOperand(2);
2597 SDValue RHS = Op.getOperand(3);
2598 SDValue Dest = Op.getOperand(4);
2599 DebugLoc dl = Op.getDebugLoc();
2601 if (LHS.getValueType() == MVT::i32) {
2603 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2604 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2605 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2606 Chain, Dest, ARMcc, CCR, Cmp);
2609 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2612 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2613 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2614 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2615 if (Result.getNode())
2619 ARMCC::CondCodes CondCode, CondCode2;
2620 FPCCToARMCC(CC, CondCode, CondCode2);
2622 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2623 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2624 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2625 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2626 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2627 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2628 if (CondCode2 != ARMCC::AL) {
2629 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2630 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2631 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2636 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2637 SDValue Chain = Op.getOperand(0);
2638 SDValue Table = Op.getOperand(1);
2639 SDValue Index = Op.getOperand(2);
2640 DebugLoc dl = Op.getDebugLoc();
2642 EVT PTy = getPointerTy();
2643 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2644 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2645 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2646 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2647 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2648 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2649 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2650 if (Subtarget->isThumb2()) {
2651 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2652 // which does another jump to the destination. This also makes it easier
2653 // to translate it to TBB / TBH later.
2654 // FIXME: This might not work if the function is extremely large.
2655 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2656 Addr, Op.getOperand(2), JTI, UId);
2658 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2659 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2660 MachinePointerInfo::getJumpTable(),
2662 Chain = Addr.getValue(1);
2663 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2664 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2666 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2667 MachinePointerInfo::getJumpTable(), false, false, 0);
2668 Chain = Addr.getValue(1);
2669 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2673 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2674 DebugLoc dl = Op.getDebugLoc();
2677 switch (Op.getOpcode()) {
2679 assert(0 && "Invalid opcode!");
2680 case ISD::FP_TO_SINT:
2681 Opc = ARMISD::FTOSI;
2683 case ISD::FP_TO_UINT:
2684 Opc = ARMISD::FTOUI;
2687 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2688 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2691 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2692 EVT VT = Op.getValueType();
2693 DebugLoc dl = Op.getDebugLoc();
2696 switch (Op.getOpcode()) {
2698 assert(0 && "Invalid opcode!");
2699 case ISD::SINT_TO_FP:
2700 Opc = ARMISD::SITOF;
2702 case ISD::UINT_TO_FP:
2703 Opc = ARMISD::UITOF;
2707 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2708 return DAG.getNode(Opc, dl, VT, Op);
2711 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2712 // Implement fcopysign with a fabs and a conditional fneg.
2713 SDValue Tmp0 = Op.getOperand(0);
2714 SDValue Tmp1 = Op.getOperand(1);
2715 DebugLoc dl = Op.getDebugLoc();
2716 EVT VT = Op.getValueType();
2717 EVT SrcVT = Tmp1.getValueType();
2718 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2719 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2720 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2721 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2722 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2723 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2726 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2727 MachineFunction &MF = DAG.getMachineFunction();
2728 MachineFrameInfo *MFI = MF.getFrameInfo();
2729 MFI->setReturnAddressIsTaken(true);
2731 EVT VT = Op.getValueType();
2732 DebugLoc dl = Op.getDebugLoc();
2733 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2735 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2736 SDValue Offset = DAG.getConstant(4, MVT::i32);
2737 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2738 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2739 MachinePointerInfo(), false, false, 0);
2742 // Return LR, which contains the return address. Mark it an implicit live-in.
2743 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2744 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2747 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2748 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2749 MFI->setFrameAddressIsTaken(true);
2751 EVT VT = Op.getValueType();
2752 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2753 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2754 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2755 ? ARM::R7 : ARM::R11;
2756 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2758 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2759 MachinePointerInfo(),
2764 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2765 /// expand a bit convert where either the source or destination type is i64 to
2766 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2767 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2768 /// vectors), since the legalizer won't know what to do with that.
2769 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2770 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2771 DebugLoc dl = N->getDebugLoc();
2772 SDValue Op = N->getOperand(0);
2774 // This function is only supposed to be called for i64 types, either as the
2775 // source or destination of the bit convert.
2776 EVT SrcVT = Op.getValueType();
2777 EVT DstVT = N->getValueType(0);
2778 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2779 "ExpandBIT_CONVERT called for non-i64 type");
2781 // Turn i64->f64 into VMOVDRR.
2782 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2783 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2784 DAG.getConstant(0, MVT::i32));
2785 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2786 DAG.getConstant(1, MVT::i32));
2787 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2788 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2791 // Turn f64->i64 into VMOVRRD.
2792 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2793 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2794 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2795 // Merge the pieces into a single i64 value.
2796 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2802 /// getZeroVector - Returns a vector of specified type with all zero elements.
2803 /// Zero vectors are used to represent vector negation and in those cases
2804 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2805 /// not support i64 elements, so sometimes the zero vectors will need to be
2806 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2808 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2809 assert(VT.isVector() && "Expected a vector type");
2810 // The canonical modified immediate encoding of a zero vector is....0!
2811 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2812 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2813 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2814 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2817 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2818 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2819 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2820 SelectionDAG &DAG) const {
2821 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2822 EVT VT = Op.getValueType();
2823 unsigned VTBits = VT.getSizeInBits();
2824 DebugLoc dl = Op.getDebugLoc();
2825 SDValue ShOpLo = Op.getOperand(0);
2826 SDValue ShOpHi = Op.getOperand(1);
2827 SDValue ShAmt = Op.getOperand(2);
2829 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2831 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2833 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2834 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2835 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2836 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2837 DAG.getConstant(VTBits, MVT::i32));
2838 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2839 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2840 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2842 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2843 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2845 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2846 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2849 SDValue Ops[2] = { Lo, Hi };
2850 return DAG.getMergeValues(Ops, 2, dl);
2853 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2854 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2855 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2856 SelectionDAG &DAG) const {
2857 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2858 EVT VT = Op.getValueType();
2859 unsigned VTBits = VT.getSizeInBits();
2860 DebugLoc dl = Op.getDebugLoc();
2861 SDValue ShOpLo = Op.getOperand(0);
2862 SDValue ShOpHi = Op.getOperand(1);
2863 SDValue ShAmt = Op.getOperand(2);
2866 assert(Op.getOpcode() == ISD::SHL_PARTS);
2867 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2868 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2869 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2870 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2871 DAG.getConstant(VTBits, MVT::i32));
2872 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2873 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2875 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2876 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2877 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2879 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2880 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2883 SDValue Ops[2] = { Lo, Hi };
2884 return DAG.getMergeValues(Ops, 2, dl);
2887 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2888 SelectionDAG &DAG) const {
2889 // The rounding mode is in bits 23:22 of the FPSCR.
2890 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2891 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2892 // so that the shift + and get folded into a bitfield extract.
2893 DebugLoc dl = Op.getDebugLoc();
2894 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2895 DAG.getConstant(Intrinsic::arm_get_fpscr,
2897 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2898 DAG.getConstant(1U << 22, MVT::i32));
2899 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2900 DAG.getConstant(22, MVT::i32));
2901 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2902 DAG.getConstant(3, MVT::i32));
2905 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2906 const ARMSubtarget *ST) {
2907 EVT VT = N->getValueType(0);
2908 DebugLoc dl = N->getDebugLoc();
2910 if (!ST->hasV6T2Ops())
2913 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2914 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2917 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2918 const ARMSubtarget *ST) {
2919 EVT VT = N->getValueType(0);
2920 DebugLoc dl = N->getDebugLoc();
2922 // Lower vector shifts on NEON to use VSHL.
2923 if (VT.isVector()) {
2924 assert(ST->hasNEON() && "unexpected vector shift");
2926 // Left shifts translate directly to the vshiftu intrinsic.
2927 if (N->getOpcode() == ISD::SHL)
2928 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2929 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2930 N->getOperand(0), N->getOperand(1));
2932 assert((N->getOpcode() == ISD::SRA ||
2933 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2935 // NEON uses the same intrinsics for both left and right shifts. For
2936 // right shifts, the shift amounts are negative, so negate the vector of
2938 EVT ShiftVT = N->getOperand(1).getValueType();
2939 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2940 getZeroVector(ShiftVT, DAG, dl),
2942 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2943 Intrinsic::arm_neon_vshifts :
2944 Intrinsic::arm_neon_vshiftu);
2945 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2946 DAG.getConstant(vshiftInt, MVT::i32),
2947 N->getOperand(0), NegatedCount);
2950 // We can get here for a node like i32 = ISD::SHL i32, i64
2954 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2955 "Unknown shift to lower!");
2957 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2958 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2959 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2962 // If we are in thumb mode, we don't have RRX.
2963 if (ST->isThumb1Only()) return SDValue();
2965 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2966 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2967 DAG.getConstant(0, MVT::i32));
2968 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2969 DAG.getConstant(1, MVT::i32));
2971 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2972 // captures the result into a carry flag.
2973 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2974 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2976 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2977 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2979 // Merge the pieces into a single i64 value.
2980 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2983 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2984 SDValue TmpOp0, TmpOp1;
2985 bool Invert = false;
2989 SDValue Op0 = Op.getOperand(0);
2990 SDValue Op1 = Op.getOperand(1);
2991 SDValue CC = Op.getOperand(2);
2992 EVT VT = Op.getValueType();
2993 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2994 DebugLoc dl = Op.getDebugLoc();
2996 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2997 switch (SetCCOpcode) {
2998 default: llvm_unreachable("Illegal FP comparison"); break;
3000 case ISD::SETNE: Invert = true; // Fallthrough
3002 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3004 case ISD::SETLT: Swap = true; // Fallthrough
3006 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3008 case ISD::SETLE: Swap = true; // Fallthrough
3010 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3011 case ISD::SETUGE: Swap = true; // Fallthrough
3012 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3013 case ISD::SETUGT: Swap = true; // Fallthrough
3014 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3015 case ISD::SETUEQ: Invert = true; // Fallthrough
3017 // Expand this to (OLT | OGT).
3021 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3022 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3024 case ISD::SETUO: Invert = true; // Fallthrough
3026 // Expand this to (OLT | OGE).
3030 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3031 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3035 // Integer comparisons.
3036 switch (SetCCOpcode) {
3037 default: llvm_unreachable("Illegal integer comparison"); break;
3038 case ISD::SETNE: Invert = true;
3039 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3040 case ISD::SETLT: Swap = true;
3041 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3042 case ISD::SETLE: Swap = true;
3043 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3044 case ISD::SETULT: Swap = true;
3045 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3046 case ISD::SETULE: Swap = true;
3047 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3050 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3051 if (Opc == ARMISD::VCEQ) {
3054 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3056 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3059 // Ignore bitconvert.
3060 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
3061 AndOp = AndOp.getOperand(0);
3063 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3065 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
3066 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
3073 std::swap(Op0, Op1);
3075 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3078 Result = DAG.getNOT(dl, Result, VT);
3083 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3084 /// valid vector constant for a NEON instruction with a "modified immediate"
3085 /// operand (e.g., VMOV). If so, return the encoded value.
3086 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3087 unsigned SplatBitSize, SelectionDAG &DAG,
3088 EVT &VT, bool is128Bits, bool isVMOV) {
3089 unsigned OpCmode, Imm;
3091 // SplatBitSize is set to the smallest size that splats the vector, so a
3092 // zero vector will always have SplatBitSize == 8. However, NEON modified
3093 // immediate instructions others than VMOV do not support the 8-bit encoding
3094 // of a zero vector, and the default encoding of zero is supposed to be the
3099 switch (SplatBitSize) {
3103 // Any 1-byte value is OK. Op=0, Cmode=1110.
3104 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3107 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3111 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3112 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3113 if ((SplatBits & ~0xff) == 0) {
3114 // Value = 0x00nn: Op=x, Cmode=100x.
3119 if ((SplatBits & ~0xff00) == 0) {
3120 // Value = 0xnn00: Op=x, Cmode=101x.
3122 Imm = SplatBits >> 8;
3128 // NEON's 32-bit VMOV supports splat values where:
3129 // * only one byte is nonzero, or
3130 // * the least significant byte is 0xff and the second byte is nonzero, or
3131 // * the least significant 2 bytes are 0xff and the third is nonzero.
3132 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3133 if ((SplatBits & ~0xff) == 0) {
3134 // Value = 0x000000nn: Op=x, Cmode=000x.
3139 if ((SplatBits & ~0xff00) == 0) {
3140 // Value = 0x0000nn00: Op=x, Cmode=001x.
3142 Imm = SplatBits >> 8;
3145 if ((SplatBits & ~0xff0000) == 0) {
3146 // Value = 0x00nn0000: Op=x, Cmode=010x.
3148 Imm = SplatBits >> 16;
3151 if ((SplatBits & ~0xff000000) == 0) {
3152 // Value = 0xnn000000: Op=x, Cmode=011x.
3154 Imm = SplatBits >> 24;
3158 if ((SplatBits & ~0xffff) == 0 &&
3159 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3160 // Value = 0x0000nnff: Op=x, Cmode=1100.
3162 Imm = SplatBits >> 8;
3167 if ((SplatBits & ~0xffffff) == 0 &&
3168 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3169 // Value = 0x00nnffff: Op=x, Cmode=1101.
3171 Imm = SplatBits >> 16;
3172 SplatBits |= 0xffff;
3176 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3177 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3178 // VMOV.I32. A (very) minor optimization would be to replicate the value
3179 // and fall through here to test for a valid 64-bit splat. But, then the
3180 // caller would also need to check and handle the change in size.
3186 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3187 uint64_t BitMask = 0xff;
3189 unsigned ImmMask = 1;
3191 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3192 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3195 } else if ((SplatBits & BitMask) != 0) {
3201 // Op=1, Cmode=1110.
3204 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3209 llvm_unreachable("unexpected size for isNEONModifiedImm");
3213 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3214 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3217 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3218 bool &ReverseVEXT, unsigned &Imm) {
3219 unsigned NumElts = VT.getVectorNumElements();
3220 ReverseVEXT = false;
3222 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3228 // If this is a VEXT shuffle, the immediate value is the index of the first
3229 // element. The other shuffle indices must be the successive elements after
3231 unsigned ExpectedElt = Imm;
3232 for (unsigned i = 1; i < NumElts; ++i) {
3233 // Increment the expected index. If it wraps around, it may still be
3234 // a VEXT but the source vectors must be swapped.
3236 if (ExpectedElt == NumElts * 2) {
3241 if (M[i] < 0) continue; // ignore UNDEF indices
3242 if (ExpectedElt != static_cast<unsigned>(M[i]))
3246 // Adjust the index value if the source operands will be swapped.
3253 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3254 /// instruction with the specified blocksize. (The order of the elements
3255 /// within each block of the vector is reversed.)
3256 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3257 unsigned BlockSize) {
3258 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3259 "Only possible block sizes for VREV are: 16, 32, 64");
3261 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3265 unsigned NumElts = VT.getVectorNumElements();
3266 unsigned BlockElts = M[0] + 1;
3267 // If the first shuffle index is UNDEF, be optimistic.
3269 BlockElts = BlockSize / EltSz;
3271 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3274 for (unsigned i = 0; i < NumElts; ++i) {
3275 if (M[i] < 0) continue; // ignore UNDEF indices
3276 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3283 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3284 unsigned &WhichResult) {
3285 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3289 unsigned NumElts = VT.getVectorNumElements();
3290 WhichResult = (M[0] == 0 ? 0 : 1);
3291 for (unsigned i = 0; i < NumElts; i += 2) {
3292 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3293 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3299 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3300 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3301 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3302 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3303 unsigned &WhichResult) {
3304 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3308 unsigned NumElts = VT.getVectorNumElements();
3309 WhichResult = (M[0] == 0 ? 0 : 1);
3310 for (unsigned i = 0; i < NumElts; i += 2) {
3311 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3312 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3318 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3319 unsigned &WhichResult) {
3320 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3324 unsigned NumElts = VT.getVectorNumElements();
3325 WhichResult = (M[0] == 0 ? 0 : 1);
3326 for (unsigned i = 0; i != NumElts; ++i) {
3327 if (M[i] < 0) continue; // ignore UNDEF indices
3328 if ((unsigned) M[i] != 2 * i + WhichResult)
3332 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3333 if (VT.is64BitVector() && EltSz == 32)
3339 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3340 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3341 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3342 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3343 unsigned &WhichResult) {
3344 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3348 unsigned Half = VT.getVectorNumElements() / 2;
3349 WhichResult = (M[0] == 0 ? 0 : 1);
3350 for (unsigned j = 0; j != 2; ++j) {
3351 unsigned Idx = WhichResult;
3352 for (unsigned i = 0; i != Half; ++i) {
3353 int MIdx = M[i + j * Half];
3354 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3360 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3361 if (VT.is64BitVector() && EltSz == 32)
3367 static bool isVZIPMask(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 unsigned Idx = WhichResult * NumElts / 2;
3376 for (unsigned i = 0; i != NumElts; i += 2) {
3377 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3378 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3383 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3384 if (VT.is64BitVector() && EltSz == 32)
3390 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3391 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3392 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3393 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3394 unsigned &WhichResult) {
3395 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3399 unsigned NumElts = VT.getVectorNumElements();
3400 WhichResult = (M[0] == 0 ? 0 : 1);
3401 unsigned Idx = WhichResult * NumElts / 2;
3402 for (unsigned i = 0; i != NumElts; i += 2) {
3403 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3404 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3409 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3410 if (VT.is64BitVector() && EltSz == 32)
3416 // If N is an integer constant that can be moved into a register in one
3417 // instruction, return an SDValue of such a constant (will become a MOV
3418 // instruction). Otherwise return null.
3419 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3420 const ARMSubtarget *ST, DebugLoc dl) {
3422 if (!isa<ConstantSDNode>(N))
3424 Val = cast<ConstantSDNode>(N)->getZExtValue();
3426 if (ST->isThumb1Only()) {
3427 if (Val <= 255 || ~Val <= 255)
3428 return DAG.getConstant(Val, MVT::i32);
3430 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3431 return DAG.getConstant(Val, MVT::i32);
3436 // If this is a case we can't handle, return null and let the default
3437 // expansion code take care of it.
3438 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3439 const ARMSubtarget *ST) {
3440 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3441 DebugLoc dl = Op.getDebugLoc();
3442 EVT VT = Op.getValueType();
3444 APInt SplatBits, SplatUndef;
3445 unsigned SplatBitSize;
3447 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3448 if (SplatBitSize <= 64) {
3449 // Check if an immediate VMOV works.
3451 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3452 SplatUndef.getZExtValue(), SplatBitSize,
3453 DAG, VmovVT, VT.is128BitVector(), true);
3454 if (Val.getNode()) {
3455 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3456 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3459 // Try an immediate VMVN.
3460 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3461 ((1LL << SplatBitSize) - 1));
3462 Val = isNEONModifiedImm(NegatedImm,
3463 SplatUndef.getZExtValue(), SplatBitSize,
3464 DAG, VmovVT, VT.is128BitVector(), false);
3465 if (Val.getNode()) {
3466 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3467 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3472 // Scan through the operands to see if only one value is used.
3473 unsigned NumElts = VT.getVectorNumElements();
3474 bool isOnlyLowElement = true;
3475 bool usesOnlyOneValue = true;
3476 bool isConstant = true;
3478 for (unsigned i = 0; i < NumElts; ++i) {
3479 SDValue V = Op.getOperand(i);
3480 if (V.getOpcode() == ISD::UNDEF)
3483 isOnlyLowElement = false;
3484 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3487 if (!Value.getNode())
3489 else if (V != Value)
3490 usesOnlyOneValue = false;
3493 if (!Value.getNode())
3494 return DAG.getUNDEF(VT);
3496 if (isOnlyLowElement)
3497 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3499 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3501 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3502 // i32 and try again.
3503 if (usesOnlyOneValue && EltSize <= 32) {
3505 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3506 if (VT.getVectorElementType().isFloatingPoint()) {
3507 SmallVector<SDValue, 8> Ops;
3508 for (unsigned i = 0; i < NumElts; ++i)
3509 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3511 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3513 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3515 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3517 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3519 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3522 // If all elements are constants and the case above didn't get hit, fall back
3523 // to the default expansion, which will generate a load from the constant
3528 // Vectors with 32- or 64-bit elements can be built by directly assigning
3529 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3530 // will be legalized.
3531 if (EltSize >= 32) {
3532 // Do the expansion with floating-point types, since that is what the VFP
3533 // registers are defined to use, and since i64 is not legal.
3534 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3535 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3536 SmallVector<SDValue, 8> Ops;
3537 for (unsigned i = 0; i < NumElts; ++i)
3538 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3539 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3540 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3546 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3547 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3548 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3549 /// are assumed to be legal.
3551 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3553 if (VT.getVectorNumElements() == 4 &&
3554 (VT.is128BitVector() || VT.is64BitVector())) {
3555 unsigned PFIndexes[4];
3556 for (unsigned i = 0; i != 4; ++i) {
3560 PFIndexes[i] = M[i];
3563 // Compute the index in the perfect shuffle table.
3564 unsigned PFTableIndex =
3565 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3566 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3567 unsigned Cost = (PFEntry >> 30);
3574 unsigned Imm, WhichResult;
3576 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3577 return (EltSize >= 32 ||
3578 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3579 isVREVMask(M, VT, 64) ||
3580 isVREVMask(M, VT, 32) ||
3581 isVREVMask(M, VT, 16) ||
3582 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3583 isVTRNMask(M, VT, WhichResult) ||
3584 isVUZPMask(M, VT, WhichResult) ||
3585 isVZIPMask(M, VT, WhichResult) ||
3586 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3587 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3588 isVZIP_v_undef_Mask(M, VT, WhichResult));
3591 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3592 /// the specified operations to build the shuffle.
3593 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3594 SDValue RHS, SelectionDAG &DAG,
3596 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3597 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3598 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3601 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3610 OP_VUZPL, // VUZP, left result
3611 OP_VUZPR, // VUZP, right result
3612 OP_VZIPL, // VZIP, left result
3613 OP_VZIPR, // VZIP, right result
3614 OP_VTRNL, // VTRN, left result
3615 OP_VTRNR // VTRN, right result
3618 if (OpNum == OP_COPY) {
3619 if (LHSID == (1*9+2)*9+3) return LHS;
3620 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3624 SDValue OpLHS, OpRHS;
3625 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3626 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3627 EVT VT = OpLHS.getValueType();
3630 default: llvm_unreachable("Unknown shuffle opcode!");
3632 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3637 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3638 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3642 return DAG.getNode(ARMISD::VEXT, dl, VT,
3644 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3647 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3648 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3651 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3652 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3655 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3656 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3660 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3661 SDValue V1 = Op.getOperand(0);
3662 SDValue V2 = Op.getOperand(1);
3663 DebugLoc dl = Op.getDebugLoc();
3664 EVT VT = Op.getValueType();
3665 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3666 SmallVector<int, 8> ShuffleMask;
3668 // Convert shuffles that are directly supported on NEON to target-specific
3669 // DAG nodes, instead of keeping them as shuffles and matching them again
3670 // during code selection. This is more efficient and avoids the possibility
3671 // of inconsistencies between legalization and selection.
3672 // FIXME: floating-point vectors should be canonicalized to integer vectors
3673 // of the same time so that they get CSEd properly.
3674 SVN->getMask(ShuffleMask);
3676 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3677 if (EltSize <= 32) {
3678 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3679 int Lane = SVN->getSplatIndex();
3680 // If this is undef splat, generate it via "just" vdup, if possible.
3681 if (Lane == -1) Lane = 0;
3683 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3684 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3686 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3687 DAG.getConstant(Lane, MVT::i32));
3692 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3695 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3696 DAG.getConstant(Imm, MVT::i32));
3699 if (isVREVMask(ShuffleMask, VT, 64))
3700 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3701 if (isVREVMask(ShuffleMask, VT, 32))
3702 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3703 if (isVREVMask(ShuffleMask, VT, 16))
3704 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3706 // Check for Neon shuffles that modify both input vectors in place.
3707 // If both results are used, i.e., if there are two shuffles with the same
3708 // source operands and with masks corresponding to both results of one of
3709 // these operations, DAG memoization will ensure that a single node is
3710 // used for both shuffles.
3711 unsigned WhichResult;
3712 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3713 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3714 V1, V2).getValue(WhichResult);
3715 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3716 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3717 V1, V2).getValue(WhichResult);
3718 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3719 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3720 V1, V2).getValue(WhichResult);
3722 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3723 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3724 V1, V1).getValue(WhichResult);
3725 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3726 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3727 V1, V1).getValue(WhichResult);
3728 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3729 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3730 V1, V1).getValue(WhichResult);
3733 // If the shuffle is not directly supported and it has 4 elements, use
3734 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3735 unsigned NumElts = VT.getVectorNumElements();
3737 unsigned PFIndexes[4];
3738 for (unsigned i = 0; i != 4; ++i) {
3739 if (ShuffleMask[i] < 0)
3742 PFIndexes[i] = ShuffleMask[i];
3745 // Compute the index in the perfect shuffle table.
3746 unsigned PFTableIndex =
3747 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3748 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3749 unsigned Cost = (PFEntry >> 30);
3752 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3755 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3756 if (EltSize >= 32) {
3757 // Do the expansion with floating-point types, since that is what the VFP
3758 // registers are defined to use, and since i64 is not legal.
3759 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3760 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3761 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3762 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3763 SmallVector<SDValue, 8> Ops;
3764 for (unsigned i = 0; i < NumElts; ++i) {
3765 if (ShuffleMask[i] < 0)
3766 Ops.push_back(DAG.getUNDEF(EltVT));
3768 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3769 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3770 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3773 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3774 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3780 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3781 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
3782 SDValue Lane = Op.getOperand(1);
3783 if (!isa<ConstantSDNode>(Lane))
3786 SDValue Vec = Op.getOperand(0);
3787 if (Op.getValueType() == MVT::i32 &&
3788 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
3789 DebugLoc dl = Op.getDebugLoc();
3790 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3796 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3797 // The only time a CONCAT_VECTORS operation can have legal types is when
3798 // two 64-bit vectors are concatenated to a 128-bit vector.
3799 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3800 "unexpected CONCAT_VECTORS");
3801 DebugLoc dl = Op.getDebugLoc();
3802 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3803 SDValue Op0 = Op.getOperand(0);
3804 SDValue Op1 = Op.getOperand(1);
3805 if (Op0.getOpcode() != ISD::UNDEF)
3806 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3807 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3808 DAG.getIntPtrConstant(0));
3809 if (Op1.getOpcode() != ISD::UNDEF)
3810 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3811 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3812 DAG.getIntPtrConstant(1));
3813 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3816 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3817 /// an extending load, return the unextended value.
3818 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3819 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3820 return N->getOperand(0);
3821 LoadSDNode *LD = cast<LoadSDNode>(N);
3822 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3823 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3824 LD->isNonTemporal(), LD->getAlignment());
3827 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3828 // Multiplications are only custom-lowered for 128-bit vectors so that
3829 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3830 EVT VT = Op.getValueType();
3831 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3832 SDNode *N0 = Op.getOperand(0).getNode();
3833 SDNode *N1 = Op.getOperand(1).getNode();
3834 unsigned NewOpc = 0;
3835 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3836 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3837 NewOpc = ARMISD::VMULLs;
3838 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3839 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3840 NewOpc = ARMISD::VMULLu;
3841 } else if (VT == MVT::v2i64) {
3842 // Fall through to expand this. It is not legal.
3845 // Other vector multiplications are legal.
3849 // Legalize to a VMULL instruction.
3850 DebugLoc DL = Op.getDebugLoc();
3851 SDValue Op0 = SkipExtension(N0, DAG);
3852 SDValue Op1 = SkipExtension(N1, DAG);
3854 assert(Op0.getValueType().is64BitVector() &&
3855 Op1.getValueType().is64BitVector() &&
3856 "unexpected types for extended operands to VMULL");
3857 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3860 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3861 switch (Op.getOpcode()) {
3862 default: llvm_unreachable("Don't know how to custom lower this!");
3863 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3864 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3865 case ISD::GlobalAddress:
3866 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3867 LowerGlobalAddressELF(Op, DAG);
3868 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3869 case ISD::SELECT: return LowerSELECT(Op, DAG);
3870 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3871 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3872 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3873 case ISD::VASTART: return LowerVASTART(Op, DAG);
3874 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3875 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
3876 case ISD::SINT_TO_FP:
3877 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3878 case ISD::FP_TO_SINT:
3879 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3880 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3881 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3882 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3883 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3884 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3885 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3886 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
3887 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3889 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3892 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3893 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3894 case ISD::SRL_PARTS:
3895 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3896 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3897 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3898 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3899 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3900 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3901 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3902 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3903 case ISD::MUL: return LowerMUL(Op, DAG);
3908 /// ReplaceNodeResults - Replace the results of node with an illegal result
3909 /// type with new values built out of custom code.
3910 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3911 SmallVectorImpl<SDValue>&Results,
3912 SelectionDAG &DAG) const {
3914 switch (N->getOpcode()) {
3916 llvm_unreachable("Don't know how to custom expand this!");
3918 case ISD::BIT_CONVERT:
3919 Res = ExpandBIT_CONVERT(N, DAG);
3923 Res = LowerShift(N, DAG, Subtarget);
3927 Results.push_back(Res);
3930 //===----------------------------------------------------------------------===//
3931 // ARM Scheduler Hooks
3932 //===----------------------------------------------------------------------===//
3935 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3936 MachineBasicBlock *BB,
3937 unsigned Size) const {
3938 unsigned dest = MI->getOperand(0).getReg();
3939 unsigned ptr = MI->getOperand(1).getReg();
3940 unsigned oldval = MI->getOperand(2).getReg();
3941 unsigned newval = MI->getOperand(3).getReg();
3942 unsigned scratch = BB->getParent()->getRegInfo()
3943 .createVirtualRegister(ARM::GPRRegisterClass);
3944 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3945 DebugLoc dl = MI->getDebugLoc();
3946 bool isThumb2 = Subtarget->isThumb2();
3948 unsigned ldrOpc, strOpc;
3950 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3952 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3953 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3956 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3957 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3960 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3961 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3965 MachineFunction *MF = BB->getParent();
3966 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3967 MachineFunction::iterator It = BB;
3968 ++It; // insert the new blocks after the current block
3970 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3971 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3972 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3973 MF->insert(It, loop1MBB);
3974 MF->insert(It, loop2MBB);
3975 MF->insert(It, exitMBB);
3977 // Transfer the remainder of BB and its successor edges to exitMBB.
3978 exitMBB->splice(exitMBB->begin(), BB,
3979 llvm::next(MachineBasicBlock::iterator(MI)),
3981 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3985 // fallthrough --> loop1MBB
3986 BB->addSuccessor(loop1MBB);
3989 // ldrex dest, [ptr]
3993 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3994 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3995 .addReg(dest).addReg(oldval));
3996 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3997 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3998 BB->addSuccessor(loop2MBB);
3999 BB->addSuccessor(exitMBB);
4002 // strex scratch, newval, [ptr]
4006 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4008 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4009 .addReg(scratch).addImm(0));
4010 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4011 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4012 BB->addSuccessor(loop1MBB);
4013 BB->addSuccessor(exitMBB);
4019 MI->eraseFromParent(); // The instruction is gone now.
4025 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4026 unsigned Size, unsigned BinOpcode) const {
4027 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4028 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4030 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4031 MachineFunction *MF = BB->getParent();
4032 MachineFunction::iterator It = BB;
4035 unsigned dest = MI->getOperand(0).getReg();
4036 unsigned ptr = MI->getOperand(1).getReg();
4037 unsigned incr = MI->getOperand(2).getReg();
4038 DebugLoc dl = MI->getDebugLoc();
4040 bool isThumb2 = Subtarget->isThumb2();
4041 unsigned ldrOpc, strOpc;
4043 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4045 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4046 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4049 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4050 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4053 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4054 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4058 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4059 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4060 MF->insert(It, loopMBB);
4061 MF->insert(It, exitMBB);
4063 // Transfer the remainder of BB and its successor edges to exitMBB.
4064 exitMBB->splice(exitMBB->begin(), BB,
4065 llvm::next(MachineBasicBlock::iterator(MI)),
4067 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4069 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4070 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4071 unsigned scratch2 = (!BinOpcode) ? incr :
4072 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4076 // fallthrough --> loopMBB
4077 BB->addSuccessor(loopMBB);
4081 // <binop> scratch2, dest, incr
4082 // strex scratch, scratch2, ptr
4085 // fallthrough --> exitMBB
4087 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4089 // operand order needs to go the other way for NAND
4090 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4091 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4092 addReg(incr).addReg(dest)).addReg(0);
4094 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4095 addReg(dest).addReg(incr)).addReg(0);
4098 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4100 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4101 .addReg(scratch).addImm(0));
4102 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4103 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4105 BB->addSuccessor(loopMBB);
4106 BB->addSuccessor(exitMBB);
4112 MI->eraseFromParent(); // The instruction is gone now.
4118 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4119 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4120 E = MBB->succ_end(); I != E; ++I)
4123 llvm_unreachable("Expecting a BB with two successors!");
4127 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4128 MachineBasicBlock *BB) const {
4129 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4130 DebugLoc dl = MI->getDebugLoc();
4131 bool isThumb2 = Subtarget->isThumb2();
4132 switch (MI->getOpcode()) {
4135 llvm_unreachable("Unexpected instr type to insert");
4137 case ARM::ATOMIC_LOAD_ADD_I8:
4138 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4139 case ARM::ATOMIC_LOAD_ADD_I16:
4140 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4141 case ARM::ATOMIC_LOAD_ADD_I32:
4142 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4144 case ARM::ATOMIC_LOAD_AND_I8:
4145 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4146 case ARM::ATOMIC_LOAD_AND_I16:
4147 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4148 case ARM::ATOMIC_LOAD_AND_I32:
4149 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4151 case ARM::ATOMIC_LOAD_OR_I8:
4152 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4153 case ARM::ATOMIC_LOAD_OR_I16:
4154 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4155 case ARM::ATOMIC_LOAD_OR_I32:
4156 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4158 case ARM::ATOMIC_LOAD_XOR_I8:
4159 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4160 case ARM::ATOMIC_LOAD_XOR_I16:
4161 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4162 case ARM::ATOMIC_LOAD_XOR_I32:
4163 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4165 case ARM::ATOMIC_LOAD_NAND_I8:
4166 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4167 case ARM::ATOMIC_LOAD_NAND_I16:
4168 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4169 case ARM::ATOMIC_LOAD_NAND_I32:
4170 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4172 case ARM::ATOMIC_LOAD_SUB_I8:
4173 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4174 case ARM::ATOMIC_LOAD_SUB_I16:
4175 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4176 case ARM::ATOMIC_LOAD_SUB_I32:
4177 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4179 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4180 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4181 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4183 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4184 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4185 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4187 case ARM::tMOVCCr_pseudo: {
4188 // To "insert" a SELECT_CC instruction, we actually have to insert the
4189 // diamond control-flow pattern. The incoming instruction knows the
4190 // destination vreg to set, the condition code register to branch on, the
4191 // true/false values to select between, and a branch opcode to use.
4192 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4193 MachineFunction::iterator It = BB;
4199 // cmpTY ccX, r1, r2
4201 // fallthrough --> copy0MBB
4202 MachineBasicBlock *thisMBB = BB;
4203 MachineFunction *F = BB->getParent();
4204 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4205 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4206 F->insert(It, copy0MBB);
4207 F->insert(It, sinkMBB);
4209 // Transfer the remainder of BB and its successor edges to sinkMBB.
4210 sinkMBB->splice(sinkMBB->begin(), BB,
4211 llvm::next(MachineBasicBlock::iterator(MI)),
4213 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4215 BB->addSuccessor(copy0MBB);
4216 BB->addSuccessor(sinkMBB);
4218 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4219 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4222 // %FalseValue = ...
4223 // # fallthrough to sinkMBB
4226 // Update machine-CFG edges
4227 BB->addSuccessor(sinkMBB);
4230 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4233 BuildMI(*BB, BB->begin(), dl,
4234 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4235 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4236 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4238 MI->eraseFromParent(); // The pseudo instruction is gone now.
4243 case ARM::BCCZi64: {
4244 // Compare both parts that make up the double comparison separately for
4246 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4248 unsigned LHS1 = MI->getOperand(1).getReg();
4249 unsigned LHS2 = MI->getOperand(2).getReg();
4251 AddDefaultPred(BuildMI(BB, dl,
4252 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4253 .addReg(LHS1).addImm(0));
4254 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4255 .addReg(LHS2).addImm(0)
4256 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4258 unsigned RHS1 = MI->getOperand(3).getReg();
4259 unsigned RHS2 = MI->getOperand(4).getReg();
4260 AddDefaultPred(BuildMI(BB, dl,
4261 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4262 .addReg(LHS1).addReg(RHS1));
4263 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4264 .addReg(LHS2).addReg(RHS2)
4265 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4268 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4269 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4270 if (MI->getOperand(0).getImm() == ARMCC::NE)
4271 std::swap(destMBB, exitMBB);
4273 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4274 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4275 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4278 MI->eraseFromParent(); // The pseudo instruction is gone now.
4284 //===----------------------------------------------------------------------===//
4285 // ARM Optimization Hooks
4286 //===----------------------------------------------------------------------===//
4289 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4290 TargetLowering::DAGCombinerInfo &DCI) {
4291 SelectionDAG &DAG = DCI.DAG;
4292 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4293 EVT VT = N->getValueType(0);
4294 unsigned Opc = N->getOpcode();
4295 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4296 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4297 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4298 ISD::CondCode CC = ISD::SETCC_INVALID;
4301 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4303 SDValue CCOp = Slct.getOperand(0);
4304 if (CCOp.getOpcode() == ISD::SETCC)
4305 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4308 bool DoXform = false;
4310 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4313 if (LHS.getOpcode() == ISD::Constant &&
4314 cast<ConstantSDNode>(LHS)->isNullValue()) {
4316 } else if (CC != ISD::SETCC_INVALID &&
4317 RHS.getOpcode() == ISD::Constant &&
4318 cast<ConstantSDNode>(RHS)->isNullValue()) {
4319 std::swap(LHS, RHS);
4320 SDValue Op0 = Slct.getOperand(0);
4321 EVT OpVT = isSlctCC ? Op0.getValueType() :
4322 Op0.getOperand(0).getValueType();
4323 bool isInt = OpVT.isInteger();
4324 CC = ISD::getSetCCInverse(CC, isInt);
4326 if (!TLI.isCondCodeLegal(CC, OpVT))
4327 return SDValue(); // Inverse operator isn't legal.
4334 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4336 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4337 Slct.getOperand(0), Slct.getOperand(1), CC);
4338 SDValue CCOp = Slct.getOperand(0);
4340 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4341 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4342 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4343 CCOp, OtherOp, Result);
4348 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4349 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4350 /// called with the default operands, and if that fails, with commuted
4352 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4353 TargetLowering::DAGCombinerInfo &DCI) {
4354 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4355 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4356 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4357 if (Result.getNode()) return Result;
4362 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4364 static SDValue PerformADDCombine(SDNode *N,
4365 TargetLowering::DAGCombinerInfo &DCI) {
4366 SDValue N0 = N->getOperand(0);
4367 SDValue N1 = N->getOperand(1);
4369 // First try with the default operand order.
4370 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4371 if (Result.getNode())
4374 // If that didn't work, try again with the operands commuted.
4375 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4378 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4380 static SDValue PerformSUBCombine(SDNode *N,
4381 TargetLowering::DAGCombinerInfo &DCI) {
4382 SDValue N0 = N->getOperand(0);
4383 SDValue N1 = N->getOperand(1);
4385 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4386 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4387 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4388 if (Result.getNode()) return Result;
4394 static SDValue PerformMULCombine(SDNode *N,
4395 TargetLowering::DAGCombinerInfo &DCI,
4396 const ARMSubtarget *Subtarget) {
4397 SelectionDAG &DAG = DCI.DAG;
4399 if (Subtarget->isThumb1Only())
4402 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4405 EVT VT = N->getValueType(0);
4409 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4413 uint64_t MulAmt = C->getZExtValue();
4414 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4415 ShiftAmt = ShiftAmt & (32 - 1);
4416 SDValue V = N->getOperand(0);
4417 DebugLoc DL = N->getDebugLoc();
4420 MulAmt >>= ShiftAmt;
4421 if (isPowerOf2_32(MulAmt - 1)) {
4422 // (mul x, 2^N + 1) => (add (shl x, N), x)
4423 Res = DAG.getNode(ISD::ADD, DL, VT,
4424 V, DAG.getNode(ISD::SHL, DL, VT,
4425 V, DAG.getConstant(Log2_32(MulAmt-1),
4427 } else if (isPowerOf2_32(MulAmt + 1)) {
4428 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4429 Res = DAG.getNode(ISD::SUB, DL, VT,
4430 DAG.getNode(ISD::SHL, DL, VT,
4431 V, DAG.getConstant(Log2_32(MulAmt+1),
4438 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4439 DAG.getConstant(ShiftAmt, MVT::i32));
4441 // Do not add new nodes to DAG combiner worklist.
4442 DCI.CombineTo(N, Res, false);
4446 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4447 static SDValue PerformORCombine(SDNode *N,
4448 TargetLowering::DAGCombinerInfo &DCI,
4449 const ARMSubtarget *Subtarget) {
4450 // Attempt to use immediate-form VORR
4451 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
4452 DebugLoc dl = N->getDebugLoc();
4453 EVT VT = N->getValueType(0);
4454 SelectionDAG &DAG = DCI.DAG;
4456 APInt SplatBits, SplatUndef;
4457 unsigned SplatBitSize;
4459 if (BVN && Subtarget->hasNEON() &&
4460 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4461 if (SplatBitSize <= 64) {
4463 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4464 SplatUndef.getZExtValue(), SplatBitSize,
4465 DAG, VorrVT, VT.is128BitVector(), false);
4466 if (Val.getNode()) {
4468 DAG.getNode(ISD::BIT_CONVERT, dl, VorrVT, N->getOperand(0));
4469 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
4470 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vorr);
4475 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4478 // BFI is only available on V6T2+
4479 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4482 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4483 DebugLoc DL = N->getDebugLoc();
4484 // 1) or (and A, mask), val => ARMbfi A, val, mask
4485 // iff (val & mask) == val
4487 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4488 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4489 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4490 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4491 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4492 // (i.e., copy a bitfield value into another bitfield of the same width)
4493 if (N0.getOpcode() != ISD::AND)
4500 // The value and the mask need to be constants so we can verify this is
4501 // actually a bitfield set. If the mask is 0xffff, we can do better
4502 // via a movt instruction, so don't use BFI in that case.
4503 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4506 unsigned Mask = C->getZExtValue();
4510 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4511 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4512 unsigned Val = C->getZExtValue();
4513 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4515 Val >>= CountTrailingZeros_32(~Mask);
4517 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4518 DAG.getConstant(Val, MVT::i32),
4519 DAG.getConstant(Mask, MVT::i32));
4521 // Do not add new nodes to DAG combiner worklist.
4522 DCI.CombineTo(N, Res, false);
4523 } else if (N1.getOpcode() == ISD::AND) {
4524 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4525 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4528 unsigned Mask2 = C->getZExtValue();
4530 if (ARM::isBitFieldInvertedMask(Mask) &&
4531 ARM::isBitFieldInvertedMask(~Mask2) &&
4532 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4533 // The pack halfword instruction works better for masks that fit it,
4534 // so use that when it's available.
4535 if (Subtarget->hasT2ExtractPack() &&
4536 (Mask == 0xffff || Mask == 0xffff0000))
4539 unsigned lsb = CountTrailingZeros_32(Mask2);
4540 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4541 DAG.getConstant(lsb, MVT::i32));
4542 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4543 DAG.getConstant(Mask, MVT::i32));
4544 // Do not add new nodes to DAG combiner worklist.
4545 DCI.CombineTo(N, Res, false);
4546 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4547 ARM::isBitFieldInvertedMask(Mask2) &&
4548 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4549 // The pack halfword instruction works better for masks that fit it,
4550 // so use that when it's available.
4551 if (Subtarget->hasT2ExtractPack() &&
4552 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4555 unsigned lsb = CountTrailingZeros_32(Mask);
4556 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4557 DAG.getConstant(lsb, MVT::i32));
4558 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4559 DAG.getConstant(Mask2, MVT::i32));
4560 // Do not add new nodes to DAG combiner worklist.
4561 DCI.CombineTo(N, Res, false);
4568 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4569 /// ARMISD::VMOVRRD.
4570 static SDValue PerformVMOVRRDCombine(SDNode *N,
4571 TargetLowering::DAGCombinerInfo &DCI) {
4572 // vmovrrd(vmovdrr x, y) -> x,y
4573 SDValue InDouble = N->getOperand(0);
4574 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4575 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4579 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4580 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4581 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4582 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4583 SDValue Op0 = N->getOperand(0);
4584 SDValue Op1 = N->getOperand(1);
4585 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4586 Op0 = Op0.getOperand(0);
4587 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4588 Op1 = Op1.getOperand(0);
4589 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4590 Op0.getNode() == Op1.getNode() &&
4591 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4592 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4593 N->getValueType(0), Op0.getOperand(0));
4597 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4598 /// ISD::BUILD_VECTOR.
4599 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4600 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4601 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4602 // into a pair of GPRs, which is fine when the value is used as a scalar,
4603 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4604 if (N->getNumOperands() == 2)
4605 return PerformVMOVDRRCombine(N, DAG);
4610 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
4611 /// ISD::VECTOR_SHUFFLE.
4612 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
4613 // The LLVM shufflevector instruction does not require the shuffle mask
4614 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
4615 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
4616 // operands do not match the mask length, they are extended by concatenating
4617 // them with undef vectors. That is probably the right thing for other
4618 // targets, but for NEON it is better to concatenate two double-register
4619 // size vector operands into a single quad-register size vector. Do that
4620 // transformation here:
4621 // shuffle(concat(v1, undef), concat(v2, undef)) ->
4622 // shuffle(concat(v1, v2), undef)
4623 SDValue Op0 = N->getOperand(0);
4624 SDValue Op1 = N->getOperand(1);
4625 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
4626 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
4627 Op0.getNumOperands() != 2 ||
4628 Op1.getNumOperands() != 2)
4630 SDValue Concat0Op1 = Op0.getOperand(1);
4631 SDValue Concat1Op1 = Op1.getOperand(1);
4632 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
4633 Concat1Op1.getOpcode() != ISD::UNDEF)
4635 // Skip the transformation if any of the types are illegal.
4636 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4637 EVT VT = N->getValueType(0);
4638 if (!TLI.isTypeLegal(VT) ||
4639 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
4640 !TLI.isTypeLegal(Concat1Op1.getValueType()))
4643 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
4644 Op0.getOperand(0), Op1.getOperand(0));
4645 // Translate the shuffle mask.
4646 SmallVector<int, 16> NewMask;
4647 unsigned NumElts = VT.getVectorNumElements();
4648 unsigned HalfElts = NumElts/2;
4649 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
4650 for (unsigned n = 0; n < NumElts; ++n) {
4651 int MaskElt = SVN->getMaskElt(n);
4653 if (MaskElt < (int)HalfElts)
4655 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
4656 NewElt = HalfElts + MaskElt - NumElts;
4657 NewMask.push_back(NewElt);
4659 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
4660 DAG.getUNDEF(VT), NewMask.data());
4663 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4664 /// ARMISD::VDUPLANE.
4665 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
4666 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4668 SDValue Op = N->getOperand(0);
4669 EVT VT = N->getValueType(0);
4671 // Ignore bit_converts.
4672 while (Op.getOpcode() == ISD::BIT_CONVERT)
4673 Op = Op.getOperand(0);
4674 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4677 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4678 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4679 // The canonical VMOV for a zero vector uses a 32-bit element size.
4680 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4682 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4684 if (EltSize > VT.getVectorElementType().getSizeInBits())
4687 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4690 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4691 /// operand of a vector shift operation, where all the elements of the
4692 /// build_vector must have the same constant integer value.
4693 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4694 // Ignore bit_converts.
4695 while (Op.getOpcode() == ISD::BIT_CONVERT)
4696 Op = Op.getOperand(0);
4697 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4698 APInt SplatBits, SplatUndef;
4699 unsigned SplatBitSize;
4701 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4702 HasAnyUndefs, ElementBits) ||
4703 SplatBitSize > ElementBits)
4705 Cnt = SplatBits.getSExtValue();
4709 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4710 /// operand of a vector shift left operation. That value must be in the range:
4711 /// 0 <= Value < ElementBits for a left shift; or
4712 /// 0 <= Value <= ElementBits for a long left shift.
4713 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4714 assert(VT.isVector() && "vector shift count is not a vector type");
4715 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4716 if (! getVShiftImm(Op, ElementBits, Cnt))
4718 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4721 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4722 /// operand of a vector shift right operation. For a shift opcode, the value
4723 /// is positive, but for an intrinsic the value count must be negative. The
4724 /// absolute value must be in the range:
4725 /// 1 <= |Value| <= ElementBits for a right shift; or
4726 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4727 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4729 assert(VT.isVector() && "vector shift count is not a vector type");
4730 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4731 if (! getVShiftImm(Op, ElementBits, Cnt))
4735 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4738 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4739 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4740 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4743 // Don't do anything for most intrinsics.
4746 // Vector shifts: check for immediate versions and lower them.
4747 // Note: This is done during DAG combining instead of DAG legalizing because
4748 // the build_vectors for 64-bit vector element shift counts are generally
4749 // not legal, and it is hard to see their values after they get legalized to
4750 // loads from a constant pool.
4751 case Intrinsic::arm_neon_vshifts:
4752 case Intrinsic::arm_neon_vshiftu:
4753 case Intrinsic::arm_neon_vshiftls:
4754 case Intrinsic::arm_neon_vshiftlu:
4755 case Intrinsic::arm_neon_vshiftn:
4756 case Intrinsic::arm_neon_vrshifts:
4757 case Intrinsic::arm_neon_vrshiftu:
4758 case Intrinsic::arm_neon_vrshiftn:
4759 case Intrinsic::arm_neon_vqshifts:
4760 case Intrinsic::arm_neon_vqshiftu:
4761 case Intrinsic::arm_neon_vqshiftsu:
4762 case Intrinsic::arm_neon_vqshiftns:
4763 case Intrinsic::arm_neon_vqshiftnu:
4764 case Intrinsic::arm_neon_vqshiftnsu:
4765 case Intrinsic::arm_neon_vqrshiftns:
4766 case Intrinsic::arm_neon_vqrshiftnu:
4767 case Intrinsic::arm_neon_vqrshiftnsu: {
4768 EVT VT = N->getOperand(1).getValueType();
4770 unsigned VShiftOpc = 0;
4773 case Intrinsic::arm_neon_vshifts:
4774 case Intrinsic::arm_neon_vshiftu:
4775 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4776 VShiftOpc = ARMISD::VSHL;
4779 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4780 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4781 ARMISD::VSHRs : ARMISD::VSHRu);
4786 case Intrinsic::arm_neon_vshiftls:
4787 case Intrinsic::arm_neon_vshiftlu:
4788 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4790 llvm_unreachable("invalid shift count for vshll intrinsic");
4792 case Intrinsic::arm_neon_vrshifts:
4793 case Intrinsic::arm_neon_vrshiftu:
4794 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4798 case Intrinsic::arm_neon_vqshifts:
4799 case Intrinsic::arm_neon_vqshiftu:
4800 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4804 case Intrinsic::arm_neon_vqshiftsu:
4805 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4807 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4809 case Intrinsic::arm_neon_vshiftn:
4810 case Intrinsic::arm_neon_vrshiftn:
4811 case Intrinsic::arm_neon_vqshiftns:
4812 case Intrinsic::arm_neon_vqshiftnu:
4813 case Intrinsic::arm_neon_vqshiftnsu:
4814 case Intrinsic::arm_neon_vqrshiftns:
4815 case Intrinsic::arm_neon_vqrshiftnu:
4816 case Intrinsic::arm_neon_vqrshiftnsu:
4817 // Narrowing shifts require an immediate right shift.
4818 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4820 llvm_unreachable("invalid shift count for narrowing vector shift "
4824 llvm_unreachable("unhandled vector shift");
4828 case Intrinsic::arm_neon_vshifts:
4829 case Intrinsic::arm_neon_vshiftu:
4830 // Opcode already set above.
4832 case Intrinsic::arm_neon_vshiftls:
4833 case Intrinsic::arm_neon_vshiftlu:
4834 if (Cnt == VT.getVectorElementType().getSizeInBits())
4835 VShiftOpc = ARMISD::VSHLLi;
4837 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4838 ARMISD::VSHLLs : ARMISD::VSHLLu);
4840 case Intrinsic::arm_neon_vshiftn:
4841 VShiftOpc = ARMISD::VSHRN; break;
4842 case Intrinsic::arm_neon_vrshifts:
4843 VShiftOpc = ARMISD::VRSHRs; break;
4844 case Intrinsic::arm_neon_vrshiftu:
4845 VShiftOpc = ARMISD::VRSHRu; break;
4846 case Intrinsic::arm_neon_vrshiftn:
4847 VShiftOpc = ARMISD::VRSHRN; break;
4848 case Intrinsic::arm_neon_vqshifts:
4849 VShiftOpc = ARMISD::VQSHLs; break;
4850 case Intrinsic::arm_neon_vqshiftu:
4851 VShiftOpc = ARMISD::VQSHLu; break;
4852 case Intrinsic::arm_neon_vqshiftsu:
4853 VShiftOpc = ARMISD::VQSHLsu; break;
4854 case Intrinsic::arm_neon_vqshiftns:
4855 VShiftOpc = ARMISD::VQSHRNs; break;
4856 case Intrinsic::arm_neon_vqshiftnu:
4857 VShiftOpc = ARMISD::VQSHRNu; break;
4858 case Intrinsic::arm_neon_vqshiftnsu:
4859 VShiftOpc = ARMISD::VQSHRNsu; break;
4860 case Intrinsic::arm_neon_vqrshiftns:
4861 VShiftOpc = ARMISD::VQRSHRNs; break;
4862 case Intrinsic::arm_neon_vqrshiftnu:
4863 VShiftOpc = ARMISD::VQRSHRNu; break;
4864 case Intrinsic::arm_neon_vqrshiftnsu:
4865 VShiftOpc = ARMISD::VQRSHRNsu; break;
4868 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4869 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4872 case Intrinsic::arm_neon_vshiftins: {
4873 EVT VT = N->getOperand(1).getValueType();
4875 unsigned VShiftOpc = 0;
4877 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4878 VShiftOpc = ARMISD::VSLI;
4879 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4880 VShiftOpc = ARMISD::VSRI;
4882 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4885 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4886 N->getOperand(1), N->getOperand(2),
4887 DAG.getConstant(Cnt, MVT::i32));
4890 case Intrinsic::arm_neon_vqrshifts:
4891 case Intrinsic::arm_neon_vqrshiftu:
4892 // No immediate versions of these to check for.
4899 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4900 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4901 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4902 /// vector element shift counts are generally not legal, and it is hard to see
4903 /// their values after they get legalized to loads from a constant pool.
4904 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4905 const ARMSubtarget *ST) {
4906 EVT VT = N->getValueType(0);
4908 // Nothing to be done for scalar shifts.
4909 if (! VT.isVector())
4912 assert(ST->hasNEON() && "unexpected vector shift");
4915 switch (N->getOpcode()) {
4916 default: llvm_unreachable("unexpected shift opcode");
4919 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4920 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4921 DAG.getConstant(Cnt, MVT::i32));
4926 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4927 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4928 ARMISD::VSHRs : ARMISD::VSHRu);
4929 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4930 DAG.getConstant(Cnt, MVT::i32));
4936 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4937 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4938 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4939 const ARMSubtarget *ST) {
4940 SDValue N0 = N->getOperand(0);
4942 // Check for sign- and zero-extensions of vector extract operations of 8-
4943 // and 16-bit vector elements. NEON supports these directly. They are
4944 // handled during DAG combining because type legalization will promote them
4945 // to 32-bit types and it is messy to recognize the operations after that.
4946 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4947 SDValue Vec = N0.getOperand(0);
4948 SDValue Lane = N0.getOperand(1);
4949 EVT VT = N->getValueType(0);
4950 EVT EltVT = N0.getValueType();
4951 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4953 if (VT == MVT::i32 &&
4954 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4955 TLI.isTypeLegal(Vec.getValueType()) &&
4956 isa<ConstantSDNode>(Lane)) {
4959 switch (N->getOpcode()) {
4960 default: llvm_unreachable("unexpected opcode");
4961 case ISD::SIGN_EXTEND:
4962 Opc = ARMISD::VGETLANEs;
4964 case ISD::ZERO_EXTEND:
4965 case ISD::ANY_EXTEND:
4966 Opc = ARMISD::VGETLANEu;
4969 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4976 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4977 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4978 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4979 const ARMSubtarget *ST) {
4980 // If the target supports NEON, try to use vmax/vmin instructions for f32
4981 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4982 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4983 // a NaN; only do the transformation when it matches that behavior.
4985 // For now only do this when using NEON for FP operations; if using VFP, it
4986 // is not obvious that the benefit outweighs the cost of switching to the
4988 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4989 N->getValueType(0) != MVT::f32)
4992 SDValue CondLHS = N->getOperand(0);
4993 SDValue CondRHS = N->getOperand(1);
4994 SDValue LHS = N->getOperand(2);
4995 SDValue RHS = N->getOperand(3);
4996 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4998 unsigned Opcode = 0;
5000 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
5001 IsReversed = false; // x CC y ? x : y
5002 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
5003 IsReversed = true ; // x CC y ? y : x
5017 // If LHS is NaN, an ordered comparison will be false and the result will
5018 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
5019 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5020 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
5021 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5023 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
5024 // will return -0, so vmin can only be used for unsafe math or if one of
5025 // the operands is known to be nonzero.
5026 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
5028 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5030 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
5039 // If LHS is NaN, an ordered comparison will be false and the result will
5040 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
5041 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5042 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
5043 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5045 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
5046 // will return +0, so vmax can only be used for unsafe math or if one of
5047 // the operands is known to be nonzero.
5048 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
5050 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5052 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
5058 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
5061 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
5062 DAGCombinerInfo &DCI) const {
5063 switch (N->getOpcode()) {
5065 case ISD::ADD: return PerformADDCombine(N, DCI);
5066 case ISD::SUB: return PerformSUBCombine(N, DCI);
5067 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
5068 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
5069 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
5070 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
5071 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
5072 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
5073 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
5074 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
5077 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
5078 case ISD::SIGN_EXTEND:
5079 case ISD::ZERO_EXTEND:
5080 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
5081 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
5086 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
5087 if (!Subtarget->allowsUnalignedMem())
5090 switch (VT.getSimpleVT().SimpleTy) {
5097 // FIXME: VLD1 etc with standard alignment is legal.
5101 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
5106 switch (VT.getSimpleVT().SimpleTy) {
5107 default: return false;
5122 if ((V & (Scale - 1)) != 0)
5125 return V == (V & ((1LL << 5) - 1));
5128 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
5129 const ARMSubtarget *Subtarget) {
5136 switch (VT.getSimpleVT().SimpleTy) {
5137 default: return false;
5142 // + imm12 or - imm8
5144 return V == (V & ((1LL << 8) - 1));
5145 return V == (V & ((1LL << 12) - 1));
5148 // Same as ARM mode. FIXME: NEON?
5149 if (!Subtarget->hasVFP2())
5154 return V == (V & ((1LL << 8) - 1));
5158 /// isLegalAddressImmediate - Return true if the integer value can be used
5159 /// as the offset of the target addressing mode for load / store of the
5161 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5162 const ARMSubtarget *Subtarget) {
5169 if (Subtarget->isThumb1Only())
5170 return isLegalT1AddressImmediate(V, VT);
5171 else if (Subtarget->isThumb2())
5172 return isLegalT2AddressImmediate(V, VT, Subtarget);
5177 switch (VT.getSimpleVT().SimpleTy) {
5178 default: return false;
5183 return V == (V & ((1LL << 12) - 1));
5186 return V == (V & ((1LL << 8) - 1));
5189 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5194 return V == (V & ((1LL << 8) - 1));
5198 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5200 int Scale = AM.Scale;
5204 switch (VT.getSimpleVT().SimpleTy) {
5205 default: return false;
5214 return Scale == 2 || Scale == 4 || Scale == 8;
5217 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5221 // Note, we allow "void" uses (basically, uses that aren't loads or
5222 // stores), because arm allows folding a scale into many arithmetic
5223 // operations. This should be made more precise and revisited later.
5225 // Allow r << imm, but the imm has to be a multiple of two.
5226 if (Scale & 1) return false;
5227 return isPowerOf2_32(Scale);
5231 /// isLegalAddressingMode - Return true if the addressing mode represented
5232 /// by AM is legal for this target, for a load/store of the specified type.
5233 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5234 const Type *Ty) const {
5235 EVT VT = getValueType(Ty, true);
5236 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5239 // Can never fold addr of global into load/store.
5244 case 0: // no scale reg, must be "r+i" or "r", or "i".
5247 if (Subtarget->isThumb1Only())
5251 // ARM doesn't support any R+R*scale+imm addr modes.
5258 if (Subtarget->isThumb2())
5259 return isLegalT2ScaledAddressingMode(AM, VT);
5261 int Scale = AM.Scale;
5262 switch (VT.getSimpleVT().SimpleTy) {
5263 default: return false;
5267 if (Scale < 0) Scale = -Scale;
5271 return isPowerOf2_32(Scale & ~1);
5275 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5280 // Note, we allow "void" uses (basically, uses that aren't loads or
5281 // stores), because arm allows folding a scale into many arithmetic
5282 // operations. This should be made more precise and revisited later.
5284 // Allow r << imm, but the imm has to be a multiple of two.
5285 if (Scale & 1) return false;
5286 return isPowerOf2_32(Scale);
5293 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5294 /// icmp immediate, that is the target has icmp instructions which can compare
5295 /// a register against the immediate without having to materialize the
5296 /// immediate into a register.
5297 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5298 if (!Subtarget->isThumb())
5299 return ARM_AM::getSOImmVal(Imm) != -1;
5300 if (Subtarget->isThumb2())
5301 return ARM_AM::getT2SOImmVal(Imm) != -1;
5302 return Imm >= 0 && Imm <= 255;
5305 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5306 bool isSEXTLoad, SDValue &Base,
5307 SDValue &Offset, bool &isInc,
5308 SelectionDAG &DAG) {
5309 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5312 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5314 Base = Ptr->getOperand(0);
5315 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5316 int RHSC = (int)RHS->getZExtValue();
5317 if (RHSC < 0 && RHSC > -256) {
5318 assert(Ptr->getOpcode() == ISD::ADD);
5320 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5324 isInc = (Ptr->getOpcode() == ISD::ADD);
5325 Offset = Ptr->getOperand(1);
5327 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5329 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5330 int RHSC = (int)RHS->getZExtValue();
5331 if (RHSC < 0 && RHSC > -0x1000) {
5332 assert(Ptr->getOpcode() == ISD::ADD);
5334 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5335 Base = Ptr->getOperand(0);
5340 if (Ptr->getOpcode() == ISD::ADD) {
5342 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5343 if (ShOpcVal != ARM_AM::no_shift) {
5344 Base = Ptr->getOperand(1);
5345 Offset = Ptr->getOperand(0);
5347 Base = Ptr->getOperand(0);
5348 Offset = Ptr->getOperand(1);
5353 isInc = (Ptr->getOpcode() == ISD::ADD);
5354 Base = Ptr->getOperand(0);
5355 Offset = Ptr->getOperand(1);
5359 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5363 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5364 bool isSEXTLoad, SDValue &Base,
5365 SDValue &Offset, bool &isInc,
5366 SelectionDAG &DAG) {
5367 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5370 Base = Ptr->getOperand(0);
5371 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5372 int RHSC = (int)RHS->getZExtValue();
5373 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5374 assert(Ptr->getOpcode() == ISD::ADD);
5376 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5378 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5379 isInc = Ptr->getOpcode() == ISD::ADD;
5380 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5388 /// getPreIndexedAddressParts - returns true by value, base pointer and
5389 /// offset pointer and addressing mode by reference if the node's address
5390 /// can be legally represented as pre-indexed load / store address.
5392 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5394 ISD::MemIndexedMode &AM,
5395 SelectionDAG &DAG) const {
5396 if (Subtarget->isThumb1Only())
5401 bool isSEXTLoad = false;
5402 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5403 Ptr = LD->getBasePtr();
5404 VT = LD->getMemoryVT();
5405 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5406 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5407 Ptr = ST->getBasePtr();
5408 VT = ST->getMemoryVT();
5413 bool isLegal = false;
5414 if (Subtarget->isThumb2())
5415 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5416 Offset, isInc, DAG);
5418 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5419 Offset, isInc, DAG);
5423 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5427 /// getPostIndexedAddressParts - returns true by value, base pointer and
5428 /// offset pointer and addressing mode by reference if this node can be
5429 /// combined with a load / store to form a post-indexed load / store.
5430 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5433 ISD::MemIndexedMode &AM,
5434 SelectionDAG &DAG) const {
5435 if (Subtarget->isThumb1Only())
5440 bool isSEXTLoad = false;
5441 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5442 VT = LD->getMemoryVT();
5443 Ptr = LD->getBasePtr();
5444 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5445 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5446 VT = ST->getMemoryVT();
5447 Ptr = ST->getBasePtr();
5452 bool isLegal = false;
5453 if (Subtarget->isThumb2())
5454 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5457 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5463 // Swap base ptr and offset to catch more post-index load / store when
5464 // it's legal. In Thumb2 mode, offset must be an immediate.
5465 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5466 !Subtarget->isThumb2())
5467 std::swap(Base, Offset);
5469 // Post-indexed load / store update the base pointer.
5474 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5478 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5482 const SelectionDAG &DAG,
5483 unsigned Depth) const {
5484 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5485 switch (Op.getOpcode()) {
5487 case ARMISD::CMOV: {
5488 // Bits are known zero/one if known on the LHS and RHS.
5489 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5490 if (KnownZero == 0 && KnownOne == 0) return;
5492 APInt KnownZeroRHS, KnownOneRHS;
5493 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5494 KnownZeroRHS, KnownOneRHS, Depth+1);
5495 KnownZero &= KnownZeroRHS;
5496 KnownOne &= KnownOneRHS;
5502 //===----------------------------------------------------------------------===//
5503 // ARM Inline Assembly Support
5504 //===----------------------------------------------------------------------===//
5506 /// getConstraintType - Given a constraint letter, return the type of
5507 /// constraint it is for this target.
5508 ARMTargetLowering::ConstraintType
5509 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5510 if (Constraint.size() == 1) {
5511 switch (Constraint[0]) {
5513 case 'l': return C_RegisterClass;
5514 case 'w': return C_RegisterClass;
5517 return TargetLowering::getConstraintType(Constraint);
5520 /// Examine constraint type and operand type and determine a weight value.
5521 /// This object must already have been set up with the operand type
5522 /// and the current alternative constraint selected.
5523 TargetLowering::ConstraintWeight
5524 ARMTargetLowering::getSingleConstraintMatchWeight(
5525 AsmOperandInfo &info, const char *constraint) const {
5526 ConstraintWeight weight = CW_Invalid;
5527 Value *CallOperandVal = info.CallOperandVal;
5528 // If we don't have a value, we can't do a match,
5529 // but allow it at the lowest weight.
5530 if (CallOperandVal == NULL)
5532 const Type *type = CallOperandVal->getType();
5533 // Look at the constraint type.
5534 switch (*constraint) {
5536 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5539 if (type->isIntegerTy()) {
5540 if (Subtarget->isThumb())
5541 weight = CW_SpecificReg;
5543 weight = CW_Register;
5547 if (type->isFloatingPointTy())
5548 weight = CW_Register;
5554 std::pair<unsigned, const TargetRegisterClass*>
5555 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5557 if (Constraint.size() == 1) {
5558 // GCC ARM Constraint Letters
5559 switch (Constraint[0]) {
5561 if (Subtarget->isThumb())
5562 return std::make_pair(0U, ARM::tGPRRegisterClass);
5564 return std::make_pair(0U, ARM::GPRRegisterClass);
5566 return std::make_pair(0U, ARM::GPRRegisterClass);
5569 return std::make_pair(0U, ARM::SPRRegisterClass);
5570 if (VT.getSizeInBits() == 64)
5571 return std::make_pair(0U, ARM::DPRRegisterClass);
5572 if (VT.getSizeInBits() == 128)
5573 return std::make_pair(0U, ARM::QPRRegisterClass);
5577 if (StringRef("{cc}").equals_lower(Constraint))
5578 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5580 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5583 std::vector<unsigned> ARMTargetLowering::
5584 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5586 if (Constraint.size() != 1)
5587 return std::vector<unsigned>();
5589 switch (Constraint[0]) { // GCC ARM Constraint Letters
5592 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5593 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5596 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5597 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5598 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5599 ARM::R12, ARM::LR, 0);
5602 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5603 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5604 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5605 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5606 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5607 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5608 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5609 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5610 if (VT.getSizeInBits() == 64)
5611 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5612 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5613 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5614 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5615 if (VT.getSizeInBits() == 128)
5616 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5617 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5621 return std::vector<unsigned>();
5624 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5625 /// vector. If it is invalid, don't add anything to Ops.
5626 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5628 std::vector<SDValue>&Ops,
5629 SelectionDAG &DAG) const {
5630 SDValue Result(0, 0);
5632 switch (Constraint) {
5634 case 'I': case 'J': case 'K': case 'L':
5635 case 'M': case 'N': case 'O':
5636 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5640 int64_t CVal64 = C->getSExtValue();
5641 int CVal = (int) CVal64;
5642 // None of these constraints allow values larger than 32 bits. Check
5643 // that the value fits in an int.
5647 switch (Constraint) {
5649 if (Subtarget->isThumb1Only()) {
5650 // This must be a constant between 0 and 255, for ADD
5652 if (CVal >= 0 && CVal <= 255)
5654 } else if (Subtarget->isThumb2()) {
5655 // A constant that can be used as an immediate value in a
5656 // data-processing instruction.
5657 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5660 // A constant that can be used as an immediate value in a
5661 // data-processing instruction.
5662 if (ARM_AM::getSOImmVal(CVal) != -1)
5668 if (Subtarget->isThumb()) { // FIXME thumb2
5669 // This must be a constant between -255 and -1, for negated ADD
5670 // immediates. This can be used in GCC with an "n" modifier that
5671 // prints the negated value, for use with SUB instructions. It is
5672 // not useful otherwise but is implemented for compatibility.
5673 if (CVal >= -255 && CVal <= -1)
5676 // This must be a constant between -4095 and 4095. It is not clear
5677 // what this constraint is intended for. Implemented for
5678 // compatibility with GCC.
5679 if (CVal >= -4095 && CVal <= 4095)
5685 if (Subtarget->isThumb1Only()) {
5686 // A 32-bit value where only one byte has a nonzero value. Exclude
5687 // zero to match GCC. This constraint is used by GCC internally for
5688 // constants that can be loaded with a move/shift combination.
5689 // It is not useful otherwise but is implemented for compatibility.
5690 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5692 } else if (Subtarget->isThumb2()) {
5693 // A constant whose bitwise inverse can be used as an immediate
5694 // value in a data-processing instruction. This can be used in GCC
5695 // with a "B" modifier that prints the inverted value, for use with
5696 // BIC and MVN instructions. It is not useful otherwise but is
5697 // implemented for compatibility.
5698 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5701 // A constant whose bitwise inverse can be used as an immediate
5702 // value in a data-processing instruction. This can be used in GCC
5703 // with a "B" modifier that prints the inverted value, for use with
5704 // BIC and MVN instructions. It is not useful otherwise but is
5705 // implemented for compatibility.
5706 if (ARM_AM::getSOImmVal(~CVal) != -1)
5712 if (Subtarget->isThumb1Only()) {
5713 // This must be a constant between -7 and 7,
5714 // for 3-operand ADD/SUB immediate instructions.
5715 if (CVal >= -7 && CVal < 7)
5717 } else if (Subtarget->isThumb2()) {
5718 // A constant whose negation can be used as an immediate value in a
5719 // data-processing instruction. This can be used in GCC with an "n"
5720 // modifier that prints the negated value, for use with SUB
5721 // instructions. It is not useful otherwise but is implemented for
5723 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5726 // A constant whose negation can be used as an immediate value in a
5727 // data-processing instruction. This can be used in GCC with an "n"
5728 // modifier that prints the negated value, for use with SUB
5729 // instructions. It is not useful otherwise but is implemented for
5731 if (ARM_AM::getSOImmVal(-CVal) != -1)
5737 if (Subtarget->isThumb()) { // FIXME thumb2
5738 // This must be a multiple of 4 between 0 and 1020, for
5739 // ADD sp + immediate.
5740 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5743 // A power of two or a constant between 0 and 32. This is used in
5744 // GCC for the shift amount on shifted register operands, but it is
5745 // useful in general for any shift amounts.
5746 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5752 if (Subtarget->isThumb()) { // FIXME thumb2
5753 // This must be a constant between 0 and 31, for shift amounts.
5754 if (CVal >= 0 && CVal <= 31)
5760 if (Subtarget->isThumb()) { // FIXME thumb2
5761 // This must be a multiple of 4 between -508 and 508, for
5762 // ADD/SUB sp = sp + immediate.
5763 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5768 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5772 if (Result.getNode()) {
5773 Ops.push_back(Result);
5776 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5780 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5781 // The ARM target isn't yet aware of offsets.
5785 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5786 APInt Imm = FPImm.bitcastToAPInt();
5787 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5788 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5789 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5791 // We can handle 4 bits of mantissa.
5792 // mantissa = (16+UInt(e:f:g:h))/16.
5793 if (Mantissa & 0x7ffff)
5796 if ((Mantissa & 0xf) != Mantissa)
5799 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5800 if (Exp < -3 || Exp > 4)
5802 Exp = ((Exp+3) & 0x7) ^ 4;
5804 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5807 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5808 APInt Imm = FPImm.bitcastToAPInt();
5809 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5810 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5811 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5813 // We can handle 4 bits of mantissa.
5814 // mantissa = (16+UInt(e:f:g:h))/16.
5815 if (Mantissa & 0xffffffffffffLL)
5818 if ((Mantissa & 0xf) != Mantissa)
5821 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5822 if (Exp < -3 || Exp > 4)
5824 Exp = ((Exp+3) & 0x7) ^ 4;
5826 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5829 bool ARM::isBitFieldInvertedMask(unsigned v) {
5830 if (v == 0xffffffff)
5832 // there can be 1's on either or both "outsides", all the "inside"
5834 unsigned int lsb = 0, msb = 31;
5835 while (v & (1 << msb)) --msb;
5836 while (v & (1 << lsb)) ++lsb;
5837 for (unsigned int i = lsb; i <= msb; ++i) {
5844 /// isFPImmLegal - Returns true if the target can instruction select the
5845 /// specified FP immediate natively. If false, the legalizer will
5846 /// materialize the FP immediate as a load from a constant pool.
5847 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5848 if (!Subtarget->hasVFP3())
5851 return ARM::getVFPf32Imm(Imm) != -1;
5853 return ARM::getVFPf64Imm(Imm) != -1;
5857 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5858 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5859 /// specified in the intrinsic calls.
5860 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5862 unsigned Intrinsic) const {
5863 switch (Intrinsic) {
5864 case Intrinsic::arm_neon_vld1:
5865 case Intrinsic::arm_neon_vld2:
5866 case Intrinsic::arm_neon_vld3:
5867 case Intrinsic::arm_neon_vld4:
5868 case Intrinsic::arm_neon_vld2lane:
5869 case Intrinsic::arm_neon_vld3lane:
5870 case Intrinsic::arm_neon_vld4lane: {
5871 Info.opc = ISD::INTRINSIC_W_CHAIN;
5872 // Conservatively set memVT to the entire set of vectors loaded.
5873 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
5874 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5875 Info.ptrVal = I.getArgOperand(0);
5877 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5878 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5879 Info.vol = false; // volatile loads with NEON intrinsics not supported
5880 Info.readMem = true;
5881 Info.writeMem = false;
5884 case Intrinsic::arm_neon_vst1:
5885 case Intrinsic::arm_neon_vst2:
5886 case Intrinsic::arm_neon_vst3:
5887 case Intrinsic::arm_neon_vst4:
5888 case Intrinsic::arm_neon_vst2lane:
5889 case Intrinsic::arm_neon_vst3lane:
5890 case Intrinsic::arm_neon_vst4lane: {
5891 Info.opc = ISD::INTRINSIC_VOID;
5892 // Conservatively set memVT to the entire set of vectors stored.
5893 unsigned NumElts = 0;
5894 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5895 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
5896 if (!ArgTy->isVectorTy())
5898 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
5900 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5901 Info.ptrVal = I.getArgOperand(0);
5903 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5904 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5905 Info.vol = false; // volatile stores with NEON intrinsics not supported
5906 Info.readMem = false;
5907 Info.writeMem = true;