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 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
105 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
106 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
107 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
108 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
109 setTruncStoreAction(VT.getSimpleVT(),
110 (MVT::SimpleValueType)InnerVT, Expand);
112 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
114 // Promote all bit-wise operations.
115 if (VT.isInteger() && VT != PromotedBitwiseVT) {
116 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
117 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
118 PromotedBitwiseVT.getSimpleVT());
119 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
120 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
121 PromotedBitwiseVT.getSimpleVT());
122 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
123 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
124 PromotedBitwiseVT.getSimpleVT());
127 // Neon does not support vector divide/remainder operations.
128 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
129 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
130 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
131 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
132 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
133 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
136 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
137 addRegisterClass(VT, ARM::DPRRegisterClass);
138 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
141 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
142 addRegisterClass(VT, ARM::QPRRegisterClass);
143 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
146 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
147 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
148 return new TargetLoweringObjectFileMachO();
150 return new ARMElfTargetObjectFile();
153 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
154 : TargetLowering(TM, createTLOF(TM)) {
155 Subtarget = &TM.getSubtarget<ARMSubtarget>();
156 RegInfo = TM.getRegisterInfo();
157 Itins = TM.getInstrItineraryData();
159 if (Subtarget->isTargetDarwin()) {
160 // Uses VFP for Thumb libfuncs if available.
161 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
162 // Single-precision floating-point arithmetic.
163 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
164 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
165 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
166 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
168 // Double-precision floating-point arithmetic.
169 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
170 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
171 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
172 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
174 // Single-precision comparisons.
175 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
176 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
177 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
178 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
179 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
180 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
181 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
182 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
184 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
193 // Double-precision comparisons.
194 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
195 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
196 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
197 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
198 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
199 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
200 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
201 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
203 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
212 // Floating-point to integer conversions.
213 // i64 conversions are done via library routines even when generating VFP
214 // instructions, so use the same ones.
215 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
216 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
217 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
218 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
220 // Conversions between floating types.
221 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
222 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
224 // Integer to floating-point conversions.
225 // i64 conversions are done via library routines even when generating VFP
226 // instructions, so use the same ones.
227 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
228 // e.g., __floatunsidf vs. __floatunssidfvfp.
229 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
230 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
231 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
232 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
236 // These libcalls are not available in 32-bit.
237 setLibcallName(RTLIB::SHL_I128, 0);
238 setLibcallName(RTLIB::SRL_I128, 0);
239 setLibcallName(RTLIB::SRA_I128, 0);
241 if (Subtarget->isAAPCS_ABI()) {
242 // Double-precision floating-point arithmetic helper functions
243 // RTABI chapter 4.1.2, Table 2
244 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
245 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
246 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
247 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
248 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
249 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
250 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
251 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
253 // Double-precision floating-point comparison helper functions
254 // RTABI chapter 4.1.2, Table 3
255 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
256 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
257 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
258 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
259 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
260 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
261 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
262 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
263 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
264 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
265 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
266 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
267 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
268 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
269 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
270 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
271 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
275 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
276 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
277 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
278 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
280 // Single-precision floating-point arithmetic helper functions
281 // RTABI chapter 4.1.2, Table 4
282 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
283 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
284 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
285 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
286 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
287 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
288 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
289 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
291 // Single-precision floating-point comparison helper functions
292 // RTABI chapter 4.1.2, Table 5
293 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
294 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
295 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
296 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
297 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
298 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
299 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
300 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
301 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
302 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
303 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
304 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
305 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
306 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
307 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
308 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
309 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
313 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
314 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
315 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
316 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
318 // Floating-point to integer conversions.
319 // RTABI chapter 4.1.2, Table 6
320 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
321 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
322 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
323 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
324 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
325 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
326 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
327 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
328 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
337 // Conversions between floating types.
338 // RTABI chapter 4.1.2, Table 7
339 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
340 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
341 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
342 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
344 // Integer to floating-point conversions.
345 // RTABI chapter 4.1.2, Table 8
346 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
347 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
348 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
349 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
350 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
351 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
352 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
353 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
354 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
361 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
363 // Long long helper functions
364 // RTABI chapter 4.2, Table 9
365 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
366 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
367 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
368 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
369 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
370 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
371 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
378 // Integer division functions
379 // RTABI chapter 4.3.1
380 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
381 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
382 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
383 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
384 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
385 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
386 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
394 if (Subtarget->isThumb1Only())
395 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
397 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
398 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
399 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
400 if (!Subtarget->isFPOnlySP())
401 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
403 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
406 if (Subtarget->hasNEON()) {
407 addDRTypeForNEON(MVT::v2f32);
408 addDRTypeForNEON(MVT::v8i8);
409 addDRTypeForNEON(MVT::v4i16);
410 addDRTypeForNEON(MVT::v2i32);
411 addDRTypeForNEON(MVT::v1i64);
413 addQRTypeForNEON(MVT::v4f32);
414 addQRTypeForNEON(MVT::v2f64);
415 addQRTypeForNEON(MVT::v16i8);
416 addQRTypeForNEON(MVT::v8i16);
417 addQRTypeForNEON(MVT::v4i32);
418 addQRTypeForNEON(MVT::v2i64);
420 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
421 // neither Neon nor VFP support any arithmetic operations on it.
422 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
423 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
424 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
425 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
426 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
427 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
428 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
429 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
430 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
431 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
432 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
433 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
434 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
435 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
436 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
437 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
438 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
439 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
440 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
441 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
442 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
443 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
444 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
445 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
447 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
449 // Neon does not support some operations on v1i64 and v2i64 types.
450 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
451 // Custom handling for some quad-vector types to detect VMULL.
452 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
453 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
454 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
455 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
456 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
458 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
459 setTargetDAGCombine(ISD::SHL);
460 setTargetDAGCombine(ISD::SRL);
461 setTargetDAGCombine(ISD::SRA);
462 setTargetDAGCombine(ISD::SIGN_EXTEND);
463 setTargetDAGCombine(ISD::ZERO_EXTEND);
464 setTargetDAGCombine(ISD::ANY_EXTEND);
465 setTargetDAGCombine(ISD::SELECT_CC);
466 setTargetDAGCombine(ISD::BUILD_VECTOR);
467 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
470 computeRegisterProperties();
472 // ARM does not have f32 extending load.
473 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
475 // ARM does not have i1 sign extending load.
476 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
478 // ARM supports all 4 flavors of integer indexed load / store.
479 if (!Subtarget->isThumb1Only()) {
480 for (unsigned im = (unsigned)ISD::PRE_INC;
481 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
482 setIndexedLoadAction(im, MVT::i1, Legal);
483 setIndexedLoadAction(im, MVT::i8, Legal);
484 setIndexedLoadAction(im, MVT::i16, Legal);
485 setIndexedLoadAction(im, MVT::i32, Legal);
486 setIndexedStoreAction(im, MVT::i1, Legal);
487 setIndexedStoreAction(im, MVT::i8, Legal);
488 setIndexedStoreAction(im, MVT::i16, Legal);
489 setIndexedStoreAction(im, MVT::i32, Legal);
493 // i64 operation support.
494 if (Subtarget->isThumb1Only()) {
495 setOperationAction(ISD::MUL, MVT::i64, Expand);
496 setOperationAction(ISD::MULHU, MVT::i32, Expand);
497 setOperationAction(ISD::MULHS, MVT::i32, Expand);
498 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
499 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
501 setOperationAction(ISD::MUL, MVT::i64, Expand);
502 setOperationAction(ISD::MULHU, MVT::i32, Expand);
503 if (!Subtarget->hasV6Ops())
504 setOperationAction(ISD::MULHS, MVT::i32, Expand);
506 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
507 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
508 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
509 setOperationAction(ISD::SRL, MVT::i64, Custom);
510 setOperationAction(ISD::SRA, MVT::i64, Custom);
512 // ARM does not have ROTL.
513 setOperationAction(ISD::ROTL, MVT::i32, Expand);
514 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
515 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
516 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
517 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
519 // Only ARMv6 has BSWAP.
520 if (!Subtarget->hasV6Ops())
521 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
523 // These are expanded into libcalls.
524 if (!Subtarget->hasDivide()) {
525 // v7M has a hardware divider
526 setOperationAction(ISD::SDIV, MVT::i32, Expand);
527 setOperationAction(ISD::UDIV, MVT::i32, Expand);
529 setOperationAction(ISD::SREM, MVT::i32, Expand);
530 setOperationAction(ISD::UREM, MVT::i32, Expand);
531 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
532 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
534 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
535 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
536 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
537 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
538 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
540 setOperationAction(ISD::TRAP, MVT::Other, Legal);
542 // Use the default implementation.
543 setOperationAction(ISD::VASTART, MVT::Other, Custom);
544 setOperationAction(ISD::VAARG, MVT::Other, Expand);
545 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
546 setOperationAction(ISD::VAEND, MVT::Other, Expand);
547 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
548 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
549 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
550 // FIXME: Shouldn't need this, since no register is used, but the legalizer
551 // doesn't yet know how to not do that for SjLj.
552 setExceptionSelectorRegister(ARM::R0);
553 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
554 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
555 // the default expansion.
556 if (Subtarget->hasDataBarrier() ||
557 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
558 // membarrier needs custom lowering; the rest are legal and handled
560 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
562 // Set them all for expansion, which will force libcalls.
563 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
564 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
565 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
566 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
567 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
568 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
569 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
570 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
571 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
572 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
573 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
584 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
585 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
586 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
587 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
588 // Since the libcalls include locking, fold in the fences
589 setShouldFoldAtomicFences(true);
591 // 64-bit versions are always libcalls (for now)
592 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
593 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
595 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
596 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
601 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
603 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
604 if (!Subtarget->hasV6Ops()) {
605 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
606 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
608 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
610 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
611 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
612 // iff target supports vfp2.
613 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
614 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
617 // We want to custom lower some of our intrinsics.
618 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
619 if (Subtarget->isTargetDarwin()) {
620 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
621 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
622 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
625 setOperationAction(ISD::SETCC, MVT::i32, Expand);
626 setOperationAction(ISD::SETCC, MVT::f32, Expand);
627 setOperationAction(ISD::SETCC, MVT::f64, Expand);
628 setOperationAction(ISD::SELECT, MVT::i32, Custom);
629 setOperationAction(ISD::SELECT, MVT::f32, Custom);
630 setOperationAction(ISD::SELECT, MVT::f64, Custom);
631 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
632 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
633 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
635 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
636 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
637 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
638 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
639 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
641 // We don't support sin/cos/fmod/copysign/pow
642 setOperationAction(ISD::FSIN, MVT::f64, Expand);
643 setOperationAction(ISD::FSIN, MVT::f32, Expand);
644 setOperationAction(ISD::FCOS, MVT::f32, Expand);
645 setOperationAction(ISD::FCOS, MVT::f64, Expand);
646 setOperationAction(ISD::FREM, MVT::f64, Expand);
647 setOperationAction(ISD::FREM, MVT::f32, Expand);
648 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
649 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
650 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
652 setOperationAction(ISD::FPOW, MVT::f64, Expand);
653 setOperationAction(ISD::FPOW, MVT::f32, Expand);
655 // Various VFP goodness
656 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
657 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
658 if (Subtarget->hasVFP2()) {
659 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
660 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
661 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
662 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
664 // Special handling for half-precision FP.
665 if (!Subtarget->hasFP16()) {
666 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
667 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
671 // We have target-specific dag combine patterns for the following nodes:
672 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
673 setTargetDAGCombine(ISD::ADD);
674 setTargetDAGCombine(ISD::SUB);
675 setTargetDAGCombine(ISD::MUL);
677 if (Subtarget->hasV6T2Ops())
678 setTargetDAGCombine(ISD::OR);
680 setStackPointerRegisterToSaveRestore(ARM::SP);
682 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
683 setSchedulingPreference(Sched::RegPressure);
685 setSchedulingPreference(Sched::Hybrid);
687 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
689 // On ARM arguments smaller than 4 bytes are extended, so all arguments
690 // are at least 4 bytes aligned.
691 setMinStackArgumentAlignment(4);
693 benefitFromCodePlacementOpt = true;
696 std::pair<const TargetRegisterClass*, uint8_t>
697 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
698 const TargetRegisterClass *RRC = 0;
700 switch (VT.getSimpleVT().SimpleTy) {
702 return TargetLowering::findRepresentativeClass(VT);
703 // Use DPR as representative register class for all floating point
704 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
705 // the cost is 1 for both f32 and f64.
706 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
707 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
708 RRC = ARM::DPRRegisterClass;
710 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
711 case MVT::v4f32: case MVT::v2f64:
712 RRC = ARM::DPRRegisterClass;
716 RRC = ARM::DPRRegisterClass;
720 RRC = ARM::DPRRegisterClass;
724 return std::make_pair(RRC, Cost);
727 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
730 case ARMISD::Wrapper: return "ARMISD::Wrapper";
731 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
732 case ARMISD::CALL: return "ARMISD::CALL";
733 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
734 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
735 case ARMISD::tCALL: return "ARMISD::tCALL";
736 case ARMISD::BRCOND: return "ARMISD::BRCOND";
737 case ARMISD::BR_JT: return "ARMISD::BR_JT";
738 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
739 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
740 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
741 case ARMISD::CMP: return "ARMISD::CMP";
742 case ARMISD::CMPZ: return "ARMISD::CMPZ";
743 case ARMISD::CMPFP: return "ARMISD::CMPFP";
744 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
745 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
746 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
747 case ARMISD::CMOV: return "ARMISD::CMOV";
748 case ARMISD::CNEG: return "ARMISD::CNEG";
750 case ARMISD::RBIT: return "ARMISD::RBIT";
752 case ARMISD::FTOSI: return "ARMISD::FTOSI";
753 case ARMISD::FTOUI: return "ARMISD::FTOUI";
754 case ARMISD::SITOF: return "ARMISD::SITOF";
755 case ARMISD::UITOF: return "ARMISD::UITOF";
757 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
758 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
759 case ARMISD::RRX: return "ARMISD::RRX";
761 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
762 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
764 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
765 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
766 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
768 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
770 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
772 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
774 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
775 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
777 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
779 case ARMISD::VCEQ: return "ARMISD::VCEQ";
780 case ARMISD::VCGE: return "ARMISD::VCGE";
781 case ARMISD::VCGEU: return "ARMISD::VCGEU";
782 case ARMISD::VCGT: return "ARMISD::VCGT";
783 case ARMISD::VCGTU: return "ARMISD::VCGTU";
784 case ARMISD::VTST: return "ARMISD::VTST";
786 case ARMISD::VSHL: return "ARMISD::VSHL";
787 case ARMISD::VSHRs: return "ARMISD::VSHRs";
788 case ARMISD::VSHRu: return "ARMISD::VSHRu";
789 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
790 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
791 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
792 case ARMISD::VSHRN: return "ARMISD::VSHRN";
793 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
794 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
795 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
796 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
797 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
798 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
799 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
800 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
801 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
802 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
803 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
804 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
805 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
806 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
807 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
808 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
809 case ARMISD::VDUP: return "ARMISD::VDUP";
810 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
811 case ARMISD::VEXT: return "ARMISD::VEXT";
812 case ARMISD::VREV64: return "ARMISD::VREV64";
813 case ARMISD::VREV32: return "ARMISD::VREV32";
814 case ARMISD::VREV16: return "ARMISD::VREV16";
815 case ARMISD::VZIP: return "ARMISD::VZIP";
816 case ARMISD::VUZP: return "ARMISD::VUZP";
817 case ARMISD::VTRN: return "ARMISD::VTRN";
818 case ARMISD::VMULLs: return "ARMISD::VMULLs";
819 case ARMISD::VMULLu: return "ARMISD::VMULLu";
820 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
821 case ARMISD::FMAX: return "ARMISD::FMAX";
822 case ARMISD::FMIN: return "ARMISD::FMIN";
823 case ARMISD::BFI: return "ARMISD::BFI";
824 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
828 /// getRegClassFor - Return the register class that should be used for the
829 /// specified value type.
830 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
831 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
832 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
833 // load / store 4 to 8 consecutive D registers.
834 if (Subtarget->hasNEON()) {
835 if (VT == MVT::v4i64)
836 return ARM::QQPRRegisterClass;
837 else if (VT == MVT::v8i64)
838 return ARM::QQQQPRRegisterClass;
840 return TargetLowering::getRegClassFor(VT);
843 // Create a fast isel object.
845 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
846 return ARM::createFastISel(funcInfo);
849 /// getFunctionAlignment - Return the Log2 alignment of this function.
850 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
851 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
854 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
855 /// be used for loads / stores from the global.
856 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
857 return (Subtarget->isThumb1Only() ? 127 : 4095);
860 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
861 unsigned NumVals = N->getNumValues();
863 return Sched::RegPressure;
865 for (unsigned i = 0; i != NumVals; ++i) {
866 EVT VT = N->getValueType(i);
867 if (VT == MVT::Flag || VT == MVT::Other)
869 if (VT.isFloatingPoint() || VT.isVector())
870 return Sched::Latency;
873 if (!N->isMachineOpcode())
874 return Sched::RegPressure;
876 // Load are scheduled for latency even if there instruction itinerary
878 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
879 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
881 if (TID.getNumDefs() == 0)
882 return Sched::RegPressure;
883 if (!Itins->isEmpty() &&
884 Itins->getOperandCycle(TID.getSchedClass(), 0) > 2)
885 return Sched::Latency;
887 return Sched::RegPressure;
891 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
892 MachineFunction &MF) const {
893 switch (RC->getID()) {
896 case ARM::tGPRRegClassID:
897 return RegInfo->hasFP(MF) ? 4 : 5;
898 case ARM::GPRRegClassID: {
899 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
900 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
902 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
903 case ARM::DPRRegClassID:
908 //===----------------------------------------------------------------------===//
910 //===----------------------------------------------------------------------===//
912 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
913 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
915 default: llvm_unreachable("Unknown condition code!");
916 case ISD::SETNE: return ARMCC::NE;
917 case ISD::SETEQ: return ARMCC::EQ;
918 case ISD::SETGT: return ARMCC::GT;
919 case ISD::SETGE: return ARMCC::GE;
920 case ISD::SETLT: return ARMCC::LT;
921 case ISD::SETLE: return ARMCC::LE;
922 case ISD::SETUGT: return ARMCC::HI;
923 case ISD::SETUGE: return ARMCC::HS;
924 case ISD::SETULT: return ARMCC::LO;
925 case ISD::SETULE: return ARMCC::LS;
929 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
930 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
931 ARMCC::CondCodes &CondCode2) {
932 CondCode2 = ARMCC::AL;
934 default: llvm_unreachable("Unknown FP condition!");
936 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
938 case ISD::SETOGT: CondCode = ARMCC::GT; break;
940 case ISD::SETOGE: CondCode = ARMCC::GE; break;
941 case ISD::SETOLT: CondCode = ARMCC::MI; break;
942 case ISD::SETOLE: CondCode = ARMCC::LS; break;
943 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
944 case ISD::SETO: CondCode = ARMCC::VC; break;
945 case ISD::SETUO: CondCode = ARMCC::VS; break;
946 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
947 case ISD::SETUGT: CondCode = ARMCC::HI; break;
948 case ISD::SETUGE: CondCode = ARMCC::PL; break;
950 case ISD::SETULT: CondCode = ARMCC::LT; break;
952 case ISD::SETULE: CondCode = ARMCC::LE; break;
954 case ISD::SETUNE: CondCode = ARMCC::NE; break;
958 //===----------------------------------------------------------------------===//
959 // Calling Convention Implementation
960 //===----------------------------------------------------------------------===//
962 #include "ARMGenCallingConv.inc"
964 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
965 /// given CallingConvention value.
966 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
968 bool isVarArg) const {
971 llvm_unreachable("Unsupported calling convention");
972 case CallingConv::Fast:
973 if (Subtarget->hasVFP2() && !isVarArg) {
974 if (!Subtarget->isAAPCS_ABI())
975 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
976 // For AAPCS ABI targets, just use VFP variant of the calling convention.
977 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
980 case CallingConv::C: {
981 // Use target triple & subtarget features to do actual dispatch.
982 if (!Subtarget->isAAPCS_ABI())
983 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
984 else if (Subtarget->hasVFP2() &&
985 FloatABIType == FloatABI::Hard && !isVarArg)
986 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
987 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
989 case CallingConv::ARM_AAPCS_VFP:
990 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
991 case CallingConv::ARM_AAPCS:
992 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
993 case CallingConv::ARM_APCS:
994 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
998 /// LowerCallResult - Lower the result values of a call into the
999 /// appropriate copies out of appropriate physical registers.
1001 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1002 CallingConv::ID CallConv, bool isVarArg,
1003 const SmallVectorImpl<ISD::InputArg> &Ins,
1004 DebugLoc dl, SelectionDAG &DAG,
1005 SmallVectorImpl<SDValue> &InVals) const {
1007 // Assign locations to each value returned by this call.
1008 SmallVector<CCValAssign, 16> RVLocs;
1009 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1010 RVLocs, *DAG.getContext());
1011 CCInfo.AnalyzeCallResult(Ins,
1012 CCAssignFnForNode(CallConv, /* Return*/ true,
1015 // Copy all of the result registers out of their specified physreg.
1016 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1017 CCValAssign VA = RVLocs[i];
1020 if (VA.needsCustom()) {
1021 // Handle f64 or half of a v2f64.
1022 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1024 Chain = Lo.getValue(1);
1025 InFlag = Lo.getValue(2);
1026 VA = RVLocs[++i]; // skip ahead to next loc
1027 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1029 Chain = Hi.getValue(1);
1030 InFlag = Hi.getValue(2);
1031 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1033 if (VA.getLocVT() == MVT::v2f64) {
1034 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1035 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1036 DAG.getConstant(0, MVT::i32));
1038 VA = RVLocs[++i]; // skip ahead to next loc
1039 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1040 Chain = Lo.getValue(1);
1041 InFlag = Lo.getValue(2);
1042 VA = RVLocs[++i]; // skip ahead to next loc
1043 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1044 Chain = Hi.getValue(1);
1045 InFlag = Hi.getValue(2);
1046 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1047 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1048 DAG.getConstant(1, MVT::i32));
1051 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1053 Chain = Val.getValue(1);
1054 InFlag = Val.getValue(2);
1057 switch (VA.getLocInfo()) {
1058 default: llvm_unreachable("Unknown loc info!");
1059 case CCValAssign::Full: break;
1060 case CCValAssign::BCvt:
1061 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1065 InVals.push_back(Val);
1071 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1072 /// by "Src" to address "Dst" of size "Size". Alignment information is
1073 /// specified by the specific parameter attribute. The copy will be passed as
1074 /// a byval function parameter.
1075 /// Sometimes what we are copying is the end of a larger object, the part that
1076 /// does not fit in registers.
1078 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1079 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1081 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1082 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1083 /*isVolatile=*/false, /*AlwaysInline=*/false,
1084 MachinePointerInfo(0), MachinePointerInfo(0));
1087 /// LowerMemOpCallTo - Store the argument to the stack.
1089 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1090 SDValue StackPtr, SDValue Arg,
1091 DebugLoc dl, SelectionDAG &DAG,
1092 const CCValAssign &VA,
1093 ISD::ArgFlagsTy Flags) const {
1094 unsigned LocMemOffset = VA.getLocMemOffset();
1095 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1096 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1097 if (Flags.isByVal())
1098 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1100 return DAG.getStore(Chain, dl, Arg, PtrOff,
1101 MachinePointerInfo::getStack(LocMemOffset),
1105 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1106 SDValue Chain, SDValue &Arg,
1107 RegsToPassVector &RegsToPass,
1108 CCValAssign &VA, CCValAssign &NextVA,
1110 SmallVector<SDValue, 8> &MemOpChains,
1111 ISD::ArgFlagsTy Flags) const {
1113 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1114 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1115 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1117 if (NextVA.isRegLoc())
1118 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1120 assert(NextVA.isMemLoc());
1121 if (StackPtr.getNode() == 0)
1122 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1124 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1130 /// LowerCall - Lowering a call into a callseq_start <-
1131 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1134 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1135 CallingConv::ID CallConv, bool isVarArg,
1137 const SmallVectorImpl<ISD::OutputArg> &Outs,
1138 const SmallVectorImpl<SDValue> &OutVals,
1139 const SmallVectorImpl<ISD::InputArg> &Ins,
1140 DebugLoc dl, SelectionDAG &DAG,
1141 SmallVectorImpl<SDValue> &InVals) const {
1142 MachineFunction &MF = DAG.getMachineFunction();
1143 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1144 bool IsSibCall = false;
1145 // Temporarily disable tail calls so things don't break.
1146 if (!EnableARMTailCalls)
1149 // Check if it's really possible to do a tail call.
1150 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1151 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1152 Outs, OutVals, Ins, DAG);
1153 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1154 // detected sibcalls.
1161 // Analyze operands of the call, assigning locations to each operand.
1162 SmallVector<CCValAssign, 16> ArgLocs;
1163 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1165 CCInfo.AnalyzeCallOperands(Outs,
1166 CCAssignFnForNode(CallConv, /* Return*/ false,
1169 // Get a count of how many bytes are to be pushed on the stack.
1170 unsigned NumBytes = CCInfo.getNextStackOffset();
1172 // For tail calls, memory operands are available in our caller's stack.
1176 // Adjust the stack pointer for the new arguments...
1177 // These operations are automatically eliminated by the prolog/epilog pass
1179 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1181 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1183 RegsToPassVector RegsToPass;
1184 SmallVector<SDValue, 8> MemOpChains;
1186 // Walk the register/memloc assignments, inserting copies/loads. In the case
1187 // of tail call optimization, arguments are handled later.
1188 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1190 ++i, ++realArgIdx) {
1191 CCValAssign &VA = ArgLocs[i];
1192 SDValue Arg = OutVals[realArgIdx];
1193 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1195 // Promote the value if needed.
1196 switch (VA.getLocInfo()) {
1197 default: llvm_unreachable("Unknown loc info!");
1198 case CCValAssign::Full: break;
1199 case CCValAssign::SExt:
1200 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1202 case CCValAssign::ZExt:
1203 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1205 case CCValAssign::AExt:
1206 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1208 case CCValAssign::BCvt:
1209 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1213 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1214 if (VA.needsCustom()) {
1215 if (VA.getLocVT() == MVT::v2f64) {
1216 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1217 DAG.getConstant(0, MVT::i32));
1218 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1219 DAG.getConstant(1, MVT::i32));
1221 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1222 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1224 VA = ArgLocs[++i]; // skip ahead to next loc
1225 if (VA.isRegLoc()) {
1226 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1227 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1229 assert(VA.isMemLoc());
1231 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1232 dl, DAG, VA, Flags));
1235 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1236 StackPtr, MemOpChains, Flags);
1238 } else if (VA.isRegLoc()) {
1239 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1240 } else if (!IsSibCall) {
1241 assert(VA.isMemLoc());
1243 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1244 dl, DAG, VA, Flags));
1248 if (!MemOpChains.empty())
1249 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1250 &MemOpChains[0], MemOpChains.size());
1252 // Build a sequence of copy-to-reg nodes chained together with token chain
1253 // and flag operands which copy the outgoing args into the appropriate regs.
1255 // Tail call byval lowering might overwrite argument registers so in case of
1256 // tail call optimization the copies to registers are lowered later.
1258 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1259 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1260 RegsToPass[i].second, InFlag);
1261 InFlag = Chain.getValue(1);
1264 // For tail calls lower the arguments to the 'real' stack slot.
1266 // Force all the incoming stack arguments to be loaded from the stack
1267 // before any new outgoing arguments are stored to the stack, because the
1268 // outgoing stack slots may alias the incoming argument stack slots, and
1269 // the alias isn't otherwise explicit. This is slightly more conservative
1270 // than necessary, because it means that each store effectively depends
1271 // on every argument instead of just those arguments it would clobber.
1273 // Do not flag preceeding copytoreg stuff together with the following stuff.
1275 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1276 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1277 RegsToPass[i].second, InFlag);
1278 InFlag = Chain.getValue(1);
1283 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1284 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1285 // node so that legalize doesn't hack it.
1286 bool isDirect = false;
1287 bool isARMFunc = false;
1288 bool isLocalARMFunc = false;
1289 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1291 if (EnableARMLongCalls) {
1292 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1293 && "long-calls with non-static relocation model!");
1294 // Handle a global address or an external symbol. If it's not one of
1295 // those, the target's already in a register, so we don't need to do
1297 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1298 const GlobalValue *GV = G->getGlobal();
1299 // Create a constant pool entry for the callee address
1300 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1301 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1304 // Get the address of the callee into a register
1305 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1306 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1307 Callee = DAG.getLoad(getPointerTy(), dl,
1308 DAG.getEntryNode(), CPAddr,
1309 MachinePointerInfo::getConstantPool(),
1311 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1312 const char *Sym = S->getSymbol();
1314 // Create a constant pool entry for the callee address
1315 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1316 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1317 Sym, ARMPCLabelIndex, 0);
1318 // Get the address of the callee into a register
1319 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1320 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1321 Callee = DAG.getLoad(getPointerTy(), dl,
1322 DAG.getEntryNode(), CPAddr,
1323 MachinePointerInfo::getConstantPool(),
1326 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1327 const GlobalValue *GV = G->getGlobal();
1329 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1330 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1331 getTargetMachine().getRelocationModel() != Reloc::Static;
1332 isARMFunc = !Subtarget->isThumb() || isStub;
1333 // ARM call to a local ARM function is predicable.
1334 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1335 // tBX takes a register source operand.
1336 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1337 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1338 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1341 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1342 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1343 Callee = DAG.getLoad(getPointerTy(), dl,
1344 DAG.getEntryNode(), CPAddr,
1345 MachinePointerInfo::getConstantPool(),
1347 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1348 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1349 getPointerTy(), Callee, PICLabel);
1351 // On ELF targets for PIC code, direct calls should go through the PLT
1352 unsigned OpFlags = 0;
1353 if (Subtarget->isTargetELF() &&
1354 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1355 OpFlags = ARMII::MO_PLT;
1356 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1358 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1360 bool isStub = Subtarget->isTargetDarwin() &&
1361 getTargetMachine().getRelocationModel() != Reloc::Static;
1362 isARMFunc = !Subtarget->isThumb() || isStub;
1363 // tBX takes a register source operand.
1364 const char *Sym = S->getSymbol();
1365 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1366 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1367 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1368 Sym, ARMPCLabelIndex, 4);
1369 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1370 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1371 Callee = DAG.getLoad(getPointerTy(), dl,
1372 DAG.getEntryNode(), CPAddr,
1373 MachinePointerInfo::getConstantPool(),
1375 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1376 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1377 getPointerTy(), Callee, PICLabel);
1379 unsigned OpFlags = 0;
1380 // On ELF targets for PIC code, direct calls should go through the PLT
1381 if (Subtarget->isTargetELF() &&
1382 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1383 OpFlags = ARMII::MO_PLT;
1384 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1388 // FIXME: handle tail calls differently.
1390 if (Subtarget->isThumb()) {
1391 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1392 CallOpc = ARMISD::CALL_NOLINK;
1394 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1396 CallOpc = (isDirect || Subtarget->hasV5TOps())
1397 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1398 : ARMISD::CALL_NOLINK;
1401 std::vector<SDValue> Ops;
1402 Ops.push_back(Chain);
1403 Ops.push_back(Callee);
1405 // Add argument registers to the end of the list so that they are known live
1407 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1408 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1409 RegsToPass[i].second.getValueType()));
1411 if (InFlag.getNode())
1412 Ops.push_back(InFlag);
1414 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1416 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1418 // Returns a chain and a flag for retval copy to use.
1419 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1420 InFlag = Chain.getValue(1);
1422 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1423 DAG.getIntPtrConstant(0, true), InFlag);
1425 InFlag = Chain.getValue(1);
1427 // Handle result values, copying them out of physregs into vregs that we
1429 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1433 /// MatchingStackOffset - Return true if the given stack call argument is
1434 /// already available in the same position (relatively) of the caller's
1435 /// incoming argument stack.
1437 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1438 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1439 const ARMInstrInfo *TII) {
1440 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1442 if (Arg.getOpcode() == ISD::CopyFromReg) {
1443 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1444 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1446 MachineInstr *Def = MRI->getVRegDef(VR);
1449 if (!Flags.isByVal()) {
1450 if (!TII->isLoadFromStackSlot(Def, FI))
1455 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1456 if (Flags.isByVal())
1457 // ByVal argument is passed in as a pointer but it's now being
1458 // dereferenced. e.g.
1459 // define @foo(%struct.X* %A) {
1460 // tail call @bar(%struct.X* byval %A)
1463 SDValue Ptr = Ld->getBasePtr();
1464 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1467 FI = FINode->getIndex();
1471 assert(FI != INT_MAX);
1472 if (!MFI->isFixedObjectIndex(FI))
1474 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1477 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1478 /// for tail call optimization. Targets which want to do tail call
1479 /// optimization should implement this function.
1481 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1482 CallingConv::ID CalleeCC,
1484 bool isCalleeStructRet,
1485 bool isCallerStructRet,
1486 const SmallVectorImpl<ISD::OutputArg> &Outs,
1487 const SmallVectorImpl<SDValue> &OutVals,
1488 const SmallVectorImpl<ISD::InputArg> &Ins,
1489 SelectionDAG& DAG) const {
1490 const Function *CallerF = DAG.getMachineFunction().getFunction();
1491 CallingConv::ID CallerCC = CallerF->getCallingConv();
1492 bool CCMatch = CallerCC == CalleeCC;
1494 // Look for obvious safe cases to perform tail call optimization that do not
1495 // require ABI changes. This is what gcc calls sibcall.
1497 // Do not sibcall optimize vararg calls unless the call site is not passing
1499 if (isVarArg && !Outs.empty())
1502 // Also avoid sibcall optimization if either caller or callee uses struct
1503 // return semantics.
1504 if (isCalleeStructRet || isCallerStructRet)
1507 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1508 // emitEpilogue is not ready for them.
1509 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1510 // LR. This means if we need to reload LR, it takes an extra instructions,
1511 // which outweighs the value of the tail call; but here we don't know yet
1512 // whether LR is going to be used. Probably the right approach is to
1513 // generate the tail call here and turn it back into CALL/RET in
1514 // emitEpilogue if LR is used.
1515 if (Subtarget->isThumb1Only())
1518 // For the moment, we can only do this to functions defined in this
1519 // compilation, or to indirect calls. A Thumb B to an ARM function,
1520 // or vice versa, is not easily fixed up in the linker unlike BL.
1521 // (We could do this by loading the address of the callee into a register;
1522 // that is an extra instruction over the direct call and burns a register
1523 // as well, so is not likely to be a win.)
1525 // It might be safe to remove this restriction on non-Darwin.
1527 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1528 // but we need to make sure there are enough registers; the only valid
1529 // registers are the 4 used for parameters. We don't currently do this
1531 if (isa<ExternalSymbolSDNode>(Callee))
1534 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1535 const GlobalValue *GV = G->getGlobal();
1536 if (GV->isDeclaration() || GV->isWeakForLinker())
1540 // If the calling conventions do not match, then we'd better make sure the
1541 // results are returned in the same way as what the caller expects.
1543 SmallVector<CCValAssign, 16> RVLocs1;
1544 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1545 RVLocs1, *DAG.getContext());
1546 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1548 SmallVector<CCValAssign, 16> RVLocs2;
1549 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1550 RVLocs2, *DAG.getContext());
1551 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1553 if (RVLocs1.size() != RVLocs2.size())
1555 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1556 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1558 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1560 if (RVLocs1[i].isRegLoc()) {
1561 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1564 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1570 // If the callee takes no arguments then go on to check the results of the
1572 if (!Outs.empty()) {
1573 // Check if stack adjustment is needed. For now, do not do this if any
1574 // argument is passed on the stack.
1575 SmallVector<CCValAssign, 16> ArgLocs;
1576 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1577 ArgLocs, *DAG.getContext());
1578 CCInfo.AnalyzeCallOperands(Outs,
1579 CCAssignFnForNode(CalleeCC, false, isVarArg));
1580 if (CCInfo.getNextStackOffset()) {
1581 MachineFunction &MF = DAG.getMachineFunction();
1583 // Check if the arguments are already laid out in the right way as
1584 // the caller's fixed stack objects.
1585 MachineFrameInfo *MFI = MF.getFrameInfo();
1586 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1587 const ARMInstrInfo *TII =
1588 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1589 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1591 ++i, ++realArgIdx) {
1592 CCValAssign &VA = ArgLocs[i];
1593 EVT RegVT = VA.getLocVT();
1594 SDValue Arg = OutVals[realArgIdx];
1595 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1596 if (VA.getLocInfo() == CCValAssign::Indirect)
1598 if (VA.needsCustom()) {
1599 // f64 and vector types are split into multiple registers or
1600 // register/stack-slot combinations. The types will not match
1601 // the registers; give up on memory f64 refs until we figure
1602 // out what to do about this.
1605 if (!ArgLocs[++i].isRegLoc())
1607 if (RegVT == MVT::v2f64) {
1608 if (!ArgLocs[++i].isRegLoc())
1610 if (!ArgLocs[++i].isRegLoc())
1613 } else if (!VA.isRegLoc()) {
1614 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1626 ARMTargetLowering::LowerReturn(SDValue Chain,
1627 CallingConv::ID CallConv, bool isVarArg,
1628 const SmallVectorImpl<ISD::OutputArg> &Outs,
1629 const SmallVectorImpl<SDValue> &OutVals,
1630 DebugLoc dl, SelectionDAG &DAG) const {
1632 // CCValAssign - represent the assignment of the return value to a location.
1633 SmallVector<CCValAssign, 16> RVLocs;
1635 // CCState - Info about the registers and stack slots.
1636 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1639 // Analyze outgoing return values.
1640 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1643 // If this is the first return lowered for this function, add
1644 // the regs to the liveout set for the function.
1645 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1646 for (unsigned i = 0; i != RVLocs.size(); ++i)
1647 if (RVLocs[i].isRegLoc())
1648 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1653 // Copy the result values into the output registers.
1654 for (unsigned i = 0, realRVLocIdx = 0;
1656 ++i, ++realRVLocIdx) {
1657 CCValAssign &VA = RVLocs[i];
1658 assert(VA.isRegLoc() && "Can only return in registers!");
1660 SDValue Arg = OutVals[realRVLocIdx];
1662 switch (VA.getLocInfo()) {
1663 default: llvm_unreachable("Unknown loc info!");
1664 case CCValAssign::Full: break;
1665 case CCValAssign::BCvt:
1666 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1670 if (VA.needsCustom()) {
1671 if (VA.getLocVT() == MVT::v2f64) {
1672 // Extract the first half and return it in two registers.
1673 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1674 DAG.getConstant(0, MVT::i32));
1675 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1676 DAG.getVTList(MVT::i32, MVT::i32), Half);
1678 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1679 Flag = Chain.getValue(1);
1680 VA = RVLocs[++i]; // skip ahead to next loc
1681 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1682 HalfGPRs.getValue(1), Flag);
1683 Flag = Chain.getValue(1);
1684 VA = RVLocs[++i]; // skip ahead to next loc
1686 // Extract the 2nd half and fall through to handle it as an f64 value.
1687 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1688 DAG.getConstant(1, MVT::i32));
1690 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1692 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1693 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1694 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1695 Flag = Chain.getValue(1);
1696 VA = RVLocs[++i]; // skip ahead to next loc
1697 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1700 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1702 // Guarantee that all emitted copies are
1703 // stuck together, avoiding something bad.
1704 Flag = Chain.getValue(1);
1709 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1711 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1716 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1717 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1718 // one of the above mentioned nodes. It has to be wrapped because otherwise
1719 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1720 // be used to form addressing mode. These wrapped nodes will be selected
1722 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1723 EVT PtrVT = Op.getValueType();
1724 // FIXME there is no actual debug info here
1725 DebugLoc dl = Op.getDebugLoc();
1726 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1728 if (CP->isMachineConstantPoolEntry())
1729 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1730 CP->getAlignment());
1732 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1733 CP->getAlignment());
1734 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1737 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1738 return MachineJumpTableInfo::EK_Inline;
1741 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1742 SelectionDAG &DAG) const {
1743 MachineFunction &MF = DAG.getMachineFunction();
1744 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1745 unsigned ARMPCLabelIndex = 0;
1746 DebugLoc DL = Op.getDebugLoc();
1747 EVT PtrVT = getPointerTy();
1748 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1749 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1751 if (RelocM == Reloc::Static) {
1752 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1754 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1755 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1756 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1757 ARMCP::CPBlockAddress,
1759 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1761 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1762 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1763 MachinePointerInfo::getConstantPool(),
1765 if (RelocM == Reloc::Static)
1767 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1768 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1771 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1773 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1774 SelectionDAG &DAG) const {
1775 DebugLoc dl = GA->getDebugLoc();
1776 EVT PtrVT = getPointerTy();
1777 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1778 MachineFunction &MF = DAG.getMachineFunction();
1779 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1780 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1781 ARMConstantPoolValue *CPV =
1782 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1783 ARMCP::CPValue, PCAdj, "tlsgd", true);
1784 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1785 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1786 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1787 MachinePointerInfo::getConstantPool(),
1789 SDValue Chain = Argument.getValue(1);
1791 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1792 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1794 // call __tls_get_addr.
1797 Entry.Node = Argument;
1798 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1799 Args.push_back(Entry);
1800 // FIXME: is there useful debug info available here?
1801 std::pair<SDValue, SDValue> CallResult =
1802 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1803 false, false, false, false,
1804 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1805 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1806 return CallResult.first;
1809 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1810 // "local exec" model.
1812 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1813 SelectionDAG &DAG) const {
1814 const GlobalValue *GV = GA->getGlobal();
1815 DebugLoc dl = GA->getDebugLoc();
1817 SDValue Chain = DAG.getEntryNode();
1818 EVT PtrVT = getPointerTy();
1819 // Get the Thread Pointer
1820 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1822 if (GV->isDeclaration()) {
1823 MachineFunction &MF = DAG.getMachineFunction();
1824 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1825 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1826 // Initial exec model.
1827 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1828 ARMConstantPoolValue *CPV =
1829 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1830 ARMCP::CPValue, PCAdj, "gottpoff", true);
1831 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1832 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1833 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1834 MachinePointerInfo::getConstantPool(),
1836 Chain = Offset.getValue(1);
1838 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1839 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1841 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1842 MachinePointerInfo::getConstantPool(),
1846 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1847 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1848 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1849 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1850 MachinePointerInfo::getConstantPool(),
1854 // The address of the thread local variable is the add of the thread
1855 // pointer with the offset of the variable.
1856 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1860 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1861 // TODO: implement the "local dynamic" model
1862 assert(Subtarget->isTargetELF() &&
1863 "TLS not implemented for non-ELF targets");
1864 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1865 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1866 // otherwise use the "Local Exec" TLS Model
1867 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1868 return LowerToTLSGeneralDynamicModel(GA, DAG);
1870 return LowerToTLSExecModels(GA, DAG);
1873 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1874 SelectionDAG &DAG) const {
1875 EVT PtrVT = getPointerTy();
1876 DebugLoc dl = Op.getDebugLoc();
1877 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1878 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1879 if (RelocM == Reloc::PIC_) {
1880 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1881 ARMConstantPoolValue *CPV =
1882 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1883 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1884 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1885 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1887 MachinePointerInfo::getConstantPool(),
1889 SDValue Chain = Result.getValue(1);
1890 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1891 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1893 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1894 MachinePointerInfo::getGOT(), false, false, 0);
1897 // If we have T2 ops, we can materialize the address directly via movt/movw
1898 // pair. This is always cheaper.
1899 if (Subtarget->useMovt()) {
1900 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1901 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1903 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1904 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1905 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1906 MachinePointerInfo::getConstantPool(),
1912 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1913 SelectionDAG &DAG) const {
1914 MachineFunction &MF = DAG.getMachineFunction();
1915 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1916 unsigned ARMPCLabelIndex = 0;
1917 EVT PtrVT = getPointerTy();
1918 DebugLoc dl = Op.getDebugLoc();
1919 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1920 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1922 if (RelocM == Reloc::Static)
1923 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1925 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1926 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1927 ARMConstantPoolValue *CPV =
1928 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1929 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1931 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1933 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1934 MachinePointerInfo::getConstantPool(),
1936 SDValue Chain = Result.getValue(1);
1938 if (RelocM == Reloc::PIC_) {
1939 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1940 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1943 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1944 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1950 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1951 SelectionDAG &DAG) const {
1952 assert(Subtarget->isTargetELF() &&
1953 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1954 MachineFunction &MF = DAG.getMachineFunction();
1955 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1956 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1957 EVT PtrVT = getPointerTy();
1958 DebugLoc dl = Op.getDebugLoc();
1959 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1960 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1961 "_GLOBAL_OFFSET_TABLE_",
1962 ARMPCLabelIndex, PCAdj);
1963 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1964 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1965 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1966 MachinePointerInfo::getConstantPool(),
1968 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1969 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1973 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
1975 DebugLoc dl = Op.getDebugLoc();
1976 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
1977 Op.getOperand(0), Op.getOperand(1));
1981 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1982 DebugLoc dl = Op.getDebugLoc();
1983 SDValue Val = DAG.getConstant(0, MVT::i32);
1984 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1985 Op.getOperand(1), Val);
1989 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1990 DebugLoc dl = Op.getDebugLoc();
1991 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1992 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1996 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1997 const ARMSubtarget *Subtarget) const {
1998 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1999 DebugLoc dl = Op.getDebugLoc();
2001 default: return SDValue(); // Don't custom lower most intrinsics.
2002 case Intrinsic::arm_thread_pointer: {
2003 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2004 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2006 case Intrinsic::eh_sjlj_lsda: {
2007 MachineFunction &MF = DAG.getMachineFunction();
2008 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2009 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
2010 EVT PtrVT = getPointerTy();
2011 DebugLoc dl = Op.getDebugLoc();
2012 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2014 unsigned PCAdj = (RelocM != Reloc::PIC_)
2015 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2016 ARMConstantPoolValue *CPV =
2017 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2018 ARMCP::CPLSDA, PCAdj);
2019 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2020 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2022 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2023 MachinePointerInfo::getConstantPool(),
2026 if (RelocM == Reloc::PIC_) {
2027 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2028 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2035 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2036 const ARMSubtarget *Subtarget) {
2037 DebugLoc dl = Op.getDebugLoc();
2038 if (!Subtarget->hasDataBarrier()) {
2039 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2040 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2042 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2043 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2044 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2045 DAG.getConstant(0, MVT::i32));
2048 SDValue Op5 = Op.getOperand(5);
2049 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2050 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2051 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2052 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2054 ARM_MB::MemBOpt DMBOpt;
2055 if (isDeviceBarrier)
2056 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2058 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2059 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2060 DAG.getConstant(DMBOpt, MVT::i32));
2063 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2064 const ARMSubtarget *Subtarget) {
2065 // ARM pre v5TE and Thumb1 does not have preload instructions.
2066 if (!(Subtarget->isThumb2() ||
2067 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2068 // Just preserve the chain.
2069 return Op.getOperand(0);
2071 DebugLoc dl = Op.getDebugLoc();
2072 unsigned Flavor = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
2074 if (!Subtarget->hasV7Ops())
2075 return Op.getOperand(0);
2076 else if (Flavor == 2 && !Subtarget->hasMPExtension())
2077 return Op.getOperand(0);
2080 if (Subtarget->isThumb())
2082 Flavor = ~Flavor & 0x3;
2084 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2085 Op.getOperand(1), DAG.getConstant(Flavor, 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 static SDValue LowerOR(SDValue Op, SelectionDAG &DAG) {
3437 SDValue Op1 = Op.getOperand(1);
3438 while (Op1.getOpcode() == ISD::BIT_CONVERT && Op1.getOperand(0) != Op1)
3439 Op1 = Op1.getOperand(0);
3440 if (Op1.getOpcode() != ARMISD::VMOVIMM) return Op;
3442 ConstantSDNode* TargetConstant = cast<ConstantSDNode>(Op1.getOperand(0));
3443 uint32_t ConstVal = TargetConstant->getZExtValue();
3445 // FIXME: VORRIMM only supports immediate encodings of 16 and 32 bit size.
3446 // In theory for VMOVIMMs whose value is already encoded as with an
3447 // 8 bit encoding, we could re-encode it as a 16 or 32 bit immediate.
3448 EVT VorrVT = Op1.getValueType();
3449 EVT EltVT = VorrVT.getVectorElementType();
3450 if (EltVT != MVT::i16 && EltVT != MVT::i32) return Op;
3453 SDValue OrConst = DAG.getTargetConstant(ConstVal, MVT::i32);
3455 DebugLoc dl = Op.getDebugLoc();
3456 EVT VT = Op.getValueType();
3457 SDValue toTy = DAG.getNode(ISD::BIT_CONVERT, dl, VorrVT, Op.getOperand(0));
3458 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, toTy, OrConst);
3459 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vorr);
3462 // If this is a case we can't handle, return null and let the default
3463 // expansion code take care of it.
3464 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3465 const ARMSubtarget *ST) {
3466 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3467 DebugLoc dl = Op.getDebugLoc();
3468 EVT VT = Op.getValueType();
3470 APInt SplatBits, SplatUndef;
3471 unsigned SplatBitSize;
3473 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3474 if (SplatBitSize <= 64) {
3475 // Check if an immediate VMOV works.
3477 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3478 SplatUndef.getZExtValue(), SplatBitSize,
3479 DAG, VmovVT, VT.is128BitVector(), true);
3480 if (Val.getNode()) {
3481 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3482 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3485 // Try an immediate VMVN.
3486 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3487 ((1LL << SplatBitSize) - 1));
3488 Val = isNEONModifiedImm(NegatedImm,
3489 SplatUndef.getZExtValue(), SplatBitSize,
3490 DAG, VmovVT, VT.is128BitVector(), false);
3491 if (Val.getNode()) {
3492 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3493 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3498 // Scan through the operands to see if only one value is used.
3499 unsigned NumElts = VT.getVectorNumElements();
3500 bool isOnlyLowElement = true;
3501 bool usesOnlyOneValue = true;
3502 bool isConstant = true;
3504 for (unsigned i = 0; i < NumElts; ++i) {
3505 SDValue V = Op.getOperand(i);
3506 if (V.getOpcode() == ISD::UNDEF)
3509 isOnlyLowElement = false;
3510 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3513 if (!Value.getNode())
3515 else if (V != Value)
3516 usesOnlyOneValue = false;
3519 if (!Value.getNode())
3520 return DAG.getUNDEF(VT);
3522 if (isOnlyLowElement)
3523 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3525 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3527 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3528 // i32 and try again.
3529 if (usesOnlyOneValue && EltSize <= 32) {
3531 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3532 if (VT.getVectorElementType().isFloatingPoint()) {
3533 SmallVector<SDValue, 8> Ops;
3534 for (unsigned i = 0; i < NumElts; ++i)
3535 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3537 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3539 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3541 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3543 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3545 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3548 // If all elements are constants and the case above didn't get hit, fall back
3549 // to the default expansion, which will generate a load from the constant
3554 // Vectors with 32- or 64-bit elements can be built by directly assigning
3555 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3556 // will be legalized.
3557 if (EltSize >= 32) {
3558 // Do the expansion with floating-point types, since that is what the VFP
3559 // registers are defined to use, and since i64 is not legal.
3560 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3561 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3562 SmallVector<SDValue, 8> Ops;
3563 for (unsigned i = 0; i < NumElts; ++i)
3564 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3565 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3566 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3572 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3573 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3574 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3575 /// are assumed to be legal.
3577 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3579 if (VT.getVectorNumElements() == 4 &&
3580 (VT.is128BitVector() || VT.is64BitVector())) {
3581 unsigned PFIndexes[4];
3582 for (unsigned i = 0; i != 4; ++i) {
3586 PFIndexes[i] = M[i];
3589 // Compute the index in the perfect shuffle table.
3590 unsigned PFTableIndex =
3591 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3592 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3593 unsigned Cost = (PFEntry >> 30);
3600 unsigned Imm, WhichResult;
3602 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3603 return (EltSize >= 32 ||
3604 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3605 isVREVMask(M, VT, 64) ||
3606 isVREVMask(M, VT, 32) ||
3607 isVREVMask(M, VT, 16) ||
3608 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3609 isVTRNMask(M, VT, WhichResult) ||
3610 isVUZPMask(M, VT, WhichResult) ||
3611 isVZIPMask(M, VT, WhichResult) ||
3612 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3613 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3614 isVZIP_v_undef_Mask(M, VT, WhichResult));
3617 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3618 /// the specified operations to build the shuffle.
3619 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3620 SDValue RHS, SelectionDAG &DAG,
3622 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3623 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3624 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3627 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3636 OP_VUZPL, // VUZP, left result
3637 OP_VUZPR, // VUZP, right result
3638 OP_VZIPL, // VZIP, left result
3639 OP_VZIPR, // VZIP, right result
3640 OP_VTRNL, // VTRN, left result
3641 OP_VTRNR // VTRN, right result
3644 if (OpNum == OP_COPY) {
3645 if (LHSID == (1*9+2)*9+3) return LHS;
3646 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3650 SDValue OpLHS, OpRHS;
3651 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3652 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3653 EVT VT = OpLHS.getValueType();
3656 default: llvm_unreachable("Unknown shuffle opcode!");
3658 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3663 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3664 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3668 return DAG.getNode(ARMISD::VEXT, dl, VT,
3670 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3673 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3674 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3677 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3678 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3681 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3682 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3686 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3687 SDValue V1 = Op.getOperand(0);
3688 SDValue V2 = Op.getOperand(1);
3689 DebugLoc dl = Op.getDebugLoc();
3690 EVT VT = Op.getValueType();
3691 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3692 SmallVector<int, 8> ShuffleMask;
3694 // Convert shuffles that are directly supported on NEON to target-specific
3695 // DAG nodes, instead of keeping them as shuffles and matching them again
3696 // during code selection. This is more efficient and avoids the possibility
3697 // of inconsistencies between legalization and selection.
3698 // FIXME: floating-point vectors should be canonicalized to integer vectors
3699 // of the same time so that they get CSEd properly.
3700 SVN->getMask(ShuffleMask);
3702 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3703 if (EltSize <= 32) {
3704 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3705 int Lane = SVN->getSplatIndex();
3706 // If this is undef splat, generate it via "just" vdup, if possible.
3707 if (Lane == -1) Lane = 0;
3709 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3710 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3712 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3713 DAG.getConstant(Lane, MVT::i32));
3718 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3721 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3722 DAG.getConstant(Imm, MVT::i32));
3725 if (isVREVMask(ShuffleMask, VT, 64))
3726 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3727 if (isVREVMask(ShuffleMask, VT, 32))
3728 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3729 if (isVREVMask(ShuffleMask, VT, 16))
3730 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3732 // Check for Neon shuffles that modify both input vectors in place.
3733 // If both results are used, i.e., if there are two shuffles with the same
3734 // source operands and with masks corresponding to both results of one of
3735 // these operations, DAG memoization will ensure that a single node is
3736 // used for both shuffles.
3737 unsigned WhichResult;
3738 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3739 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3740 V1, V2).getValue(WhichResult);
3741 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3742 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3743 V1, V2).getValue(WhichResult);
3744 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3745 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3746 V1, V2).getValue(WhichResult);
3748 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3749 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3750 V1, V1).getValue(WhichResult);
3751 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3752 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3753 V1, V1).getValue(WhichResult);
3754 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3755 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3756 V1, V1).getValue(WhichResult);
3759 // If the shuffle is not directly supported and it has 4 elements, use
3760 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3761 unsigned NumElts = VT.getVectorNumElements();
3763 unsigned PFIndexes[4];
3764 for (unsigned i = 0; i != 4; ++i) {
3765 if (ShuffleMask[i] < 0)
3768 PFIndexes[i] = ShuffleMask[i];
3771 // Compute the index in the perfect shuffle table.
3772 unsigned PFTableIndex =
3773 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3774 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3775 unsigned Cost = (PFEntry >> 30);
3778 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3781 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3782 if (EltSize >= 32) {
3783 // Do the expansion with floating-point types, since that is what the VFP
3784 // registers are defined to use, and since i64 is not legal.
3785 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3786 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3787 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3788 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3789 SmallVector<SDValue, 8> Ops;
3790 for (unsigned i = 0; i < NumElts; ++i) {
3791 if (ShuffleMask[i] < 0)
3792 Ops.push_back(DAG.getUNDEF(EltVT));
3794 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3795 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3796 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3799 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3800 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3806 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3807 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
3808 SDValue Lane = Op.getOperand(1);
3809 if (!isa<ConstantSDNode>(Lane))
3812 SDValue Vec = Op.getOperand(0);
3813 if (Op.getValueType() == MVT::i32 &&
3814 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
3815 DebugLoc dl = Op.getDebugLoc();
3816 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3822 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3823 // The only time a CONCAT_VECTORS operation can have legal types is when
3824 // two 64-bit vectors are concatenated to a 128-bit vector.
3825 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3826 "unexpected CONCAT_VECTORS");
3827 DebugLoc dl = Op.getDebugLoc();
3828 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3829 SDValue Op0 = Op.getOperand(0);
3830 SDValue Op1 = Op.getOperand(1);
3831 if (Op0.getOpcode() != ISD::UNDEF)
3832 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3833 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3834 DAG.getIntPtrConstant(0));
3835 if (Op1.getOpcode() != ISD::UNDEF)
3836 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3837 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3838 DAG.getIntPtrConstant(1));
3839 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3842 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3843 /// an extending load, return the unextended value.
3844 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3845 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3846 return N->getOperand(0);
3847 LoadSDNode *LD = cast<LoadSDNode>(N);
3848 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3849 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3850 LD->isNonTemporal(), LD->getAlignment());
3853 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3854 // Multiplications are only custom-lowered for 128-bit vectors so that
3855 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3856 EVT VT = Op.getValueType();
3857 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3858 SDNode *N0 = Op.getOperand(0).getNode();
3859 SDNode *N1 = Op.getOperand(1).getNode();
3860 unsigned NewOpc = 0;
3861 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3862 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3863 NewOpc = ARMISD::VMULLs;
3864 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3865 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3866 NewOpc = ARMISD::VMULLu;
3867 } else if (VT == MVT::v2i64) {
3868 // Fall through to expand this. It is not legal.
3871 // Other vector multiplications are legal.
3875 // Legalize to a VMULL instruction.
3876 DebugLoc DL = Op.getDebugLoc();
3877 SDValue Op0 = SkipExtension(N0, DAG);
3878 SDValue Op1 = SkipExtension(N1, DAG);
3880 assert(Op0.getValueType().is64BitVector() &&
3881 Op1.getValueType().is64BitVector() &&
3882 "unexpected types for extended operands to VMULL");
3883 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3886 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3887 switch (Op.getOpcode()) {
3888 default: llvm_unreachable("Don't know how to custom lower this!");
3889 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3890 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3891 case ISD::GlobalAddress:
3892 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3893 LowerGlobalAddressELF(Op, DAG);
3894 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3895 case ISD::SELECT: return LowerSELECT(Op, DAG);
3896 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3897 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3898 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3899 case ISD::VASTART: return LowerVASTART(Op, DAG);
3900 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3901 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
3902 case ISD::SINT_TO_FP:
3903 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3904 case ISD::FP_TO_SINT:
3905 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3906 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3907 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3908 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3909 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3910 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3911 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3912 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
3913 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3915 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3918 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3919 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3920 case ISD::SRL_PARTS:
3921 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3922 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3923 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3924 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3925 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3926 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3927 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3928 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3929 case ISD::MUL: return LowerMUL(Op, DAG);
3930 case ISD::OR: return LowerOR(Op, DAG);
3935 /// ReplaceNodeResults - Replace the results of node with an illegal result
3936 /// type with new values built out of custom code.
3937 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3938 SmallVectorImpl<SDValue>&Results,
3939 SelectionDAG &DAG) const {
3941 switch (N->getOpcode()) {
3943 llvm_unreachable("Don't know how to custom expand this!");
3945 case ISD::BIT_CONVERT:
3946 Res = ExpandBIT_CONVERT(N, DAG);
3950 Res = LowerShift(N, DAG, Subtarget);
3954 Results.push_back(Res);
3957 //===----------------------------------------------------------------------===//
3958 // ARM Scheduler Hooks
3959 //===----------------------------------------------------------------------===//
3962 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3963 MachineBasicBlock *BB,
3964 unsigned Size) const {
3965 unsigned dest = MI->getOperand(0).getReg();
3966 unsigned ptr = MI->getOperand(1).getReg();
3967 unsigned oldval = MI->getOperand(2).getReg();
3968 unsigned newval = MI->getOperand(3).getReg();
3969 unsigned scratch = BB->getParent()->getRegInfo()
3970 .createVirtualRegister(ARM::GPRRegisterClass);
3971 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3972 DebugLoc dl = MI->getDebugLoc();
3973 bool isThumb2 = Subtarget->isThumb2();
3975 unsigned ldrOpc, strOpc;
3977 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3979 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3980 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3983 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3984 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3987 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3988 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3992 MachineFunction *MF = BB->getParent();
3993 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3994 MachineFunction::iterator It = BB;
3995 ++It; // insert the new blocks after the current block
3997 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3998 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3999 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4000 MF->insert(It, loop1MBB);
4001 MF->insert(It, loop2MBB);
4002 MF->insert(It, exitMBB);
4004 // Transfer the remainder of BB and its successor edges to exitMBB.
4005 exitMBB->splice(exitMBB->begin(), BB,
4006 llvm::next(MachineBasicBlock::iterator(MI)),
4008 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4012 // fallthrough --> loop1MBB
4013 BB->addSuccessor(loop1MBB);
4016 // ldrex dest, [ptr]
4020 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4021 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4022 .addReg(dest).addReg(oldval));
4023 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4024 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4025 BB->addSuccessor(loop2MBB);
4026 BB->addSuccessor(exitMBB);
4029 // strex scratch, newval, [ptr]
4033 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
4035 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4036 .addReg(scratch).addImm(0));
4037 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4038 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4039 BB->addSuccessor(loop1MBB);
4040 BB->addSuccessor(exitMBB);
4046 MI->eraseFromParent(); // The instruction is gone now.
4052 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4053 unsigned Size, unsigned BinOpcode) const {
4054 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4055 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4057 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4058 MachineFunction *MF = BB->getParent();
4059 MachineFunction::iterator It = BB;
4062 unsigned dest = MI->getOperand(0).getReg();
4063 unsigned ptr = MI->getOperand(1).getReg();
4064 unsigned incr = MI->getOperand(2).getReg();
4065 DebugLoc dl = MI->getDebugLoc();
4067 bool isThumb2 = Subtarget->isThumb2();
4068 unsigned ldrOpc, strOpc;
4070 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4072 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4073 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4076 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4077 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4080 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4081 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4085 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4086 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4087 MF->insert(It, loopMBB);
4088 MF->insert(It, exitMBB);
4090 // Transfer the remainder of BB and its successor edges to exitMBB.
4091 exitMBB->splice(exitMBB->begin(), BB,
4092 llvm::next(MachineBasicBlock::iterator(MI)),
4094 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4096 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4097 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4098 unsigned scratch2 = (!BinOpcode) ? incr :
4099 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4103 // fallthrough --> loopMBB
4104 BB->addSuccessor(loopMBB);
4108 // <binop> scratch2, dest, incr
4109 // strex scratch, scratch2, ptr
4112 // fallthrough --> exitMBB
4114 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4116 // operand order needs to go the other way for NAND
4117 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4118 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4119 addReg(incr).addReg(dest)).addReg(0);
4121 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4122 addReg(dest).addReg(incr)).addReg(0);
4125 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4127 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4128 .addReg(scratch).addImm(0));
4129 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4130 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4132 BB->addSuccessor(loopMBB);
4133 BB->addSuccessor(exitMBB);
4139 MI->eraseFromParent(); // The instruction is gone now.
4145 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4146 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4147 E = MBB->succ_end(); I != E; ++I)
4150 llvm_unreachable("Expecting a BB with two successors!");
4154 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4155 MachineBasicBlock *BB) const {
4156 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4157 DebugLoc dl = MI->getDebugLoc();
4158 bool isThumb2 = Subtarget->isThumb2();
4159 switch (MI->getOpcode()) {
4162 llvm_unreachable("Unexpected instr type to insert");
4164 case ARM::ATOMIC_LOAD_ADD_I8:
4165 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4166 case ARM::ATOMIC_LOAD_ADD_I16:
4167 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4168 case ARM::ATOMIC_LOAD_ADD_I32:
4169 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4171 case ARM::ATOMIC_LOAD_AND_I8:
4172 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4173 case ARM::ATOMIC_LOAD_AND_I16:
4174 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4175 case ARM::ATOMIC_LOAD_AND_I32:
4176 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4178 case ARM::ATOMIC_LOAD_OR_I8:
4179 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4180 case ARM::ATOMIC_LOAD_OR_I16:
4181 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4182 case ARM::ATOMIC_LOAD_OR_I32:
4183 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4185 case ARM::ATOMIC_LOAD_XOR_I8:
4186 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4187 case ARM::ATOMIC_LOAD_XOR_I16:
4188 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4189 case ARM::ATOMIC_LOAD_XOR_I32:
4190 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4192 case ARM::ATOMIC_LOAD_NAND_I8:
4193 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4194 case ARM::ATOMIC_LOAD_NAND_I16:
4195 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4196 case ARM::ATOMIC_LOAD_NAND_I32:
4197 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4199 case ARM::ATOMIC_LOAD_SUB_I8:
4200 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4201 case ARM::ATOMIC_LOAD_SUB_I16:
4202 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4203 case ARM::ATOMIC_LOAD_SUB_I32:
4204 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4206 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4207 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4208 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4210 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4211 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4212 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4214 case ARM::tMOVCCr_pseudo: {
4215 // To "insert" a SELECT_CC instruction, we actually have to insert the
4216 // diamond control-flow pattern. The incoming instruction knows the
4217 // destination vreg to set, the condition code register to branch on, the
4218 // true/false values to select between, and a branch opcode to use.
4219 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4220 MachineFunction::iterator It = BB;
4226 // cmpTY ccX, r1, r2
4228 // fallthrough --> copy0MBB
4229 MachineBasicBlock *thisMBB = BB;
4230 MachineFunction *F = BB->getParent();
4231 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4232 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4233 F->insert(It, copy0MBB);
4234 F->insert(It, sinkMBB);
4236 // Transfer the remainder of BB and its successor edges to sinkMBB.
4237 sinkMBB->splice(sinkMBB->begin(), BB,
4238 llvm::next(MachineBasicBlock::iterator(MI)),
4240 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4242 BB->addSuccessor(copy0MBB);
4243 BB->addSuccessor(sinkMBB);
4245 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4246 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4249 // %FalseValue = ...
4250 // # fallthrough to sinkMBB
4253 // Update machine-CFG edges
4254 BB->addSuccessor(sinkMBB);
4257 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4260 BuildMI(*BB, BB->begin(), dl,
4261 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4262 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4263 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4265 MI->eraseFromParent(); // The pseudo instruction is gone now.
4270 case ARM::BCCZi64: {
4271 // Compare both parts that make up the double comparison separately for
4273 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4275 unsigned LHS1 = MI->getOperand(1).getReg();
4276 unsigned LHS2 = MI->getOperand(2).getReg();
4278 AddDefaultPred(BuildMI(BB, dl,
4279 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4280 .addReg(LHS1).addImm(0));
4281 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4282 .addReg(LHS2).addImm(0)
4283 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4285 unsigned RHS1 = MI->getOperand(3).getReg();
4286 unsigned RHS2 = MI->getOperand(4).getReg();
4287 AddDefaultPred(BuildMI(BB, dl,
4288 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4289 .addReg(LHS1).addReg(RHS1));
4290 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4291 .addReg(LHS2).addReg(RHS2)
4292 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4295 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4296 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4297 if (MI->getOperand(0).getImm() == ARMCC::NE)
4298 std::swap(destMBB, exitMBB);
4300 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4301 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4302 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4305 MI->eraseFromParent(); // The pseudo instruction is gone now.
4311 //===----------------------------------------------------------------------===//
4312 // ARM Optimization Hooks
4313 //===----------------------------------------------------------------------===//
4316 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4317 TargetLowering::DAGCombinerInfo &DCI) {
4318 SelectionDAG &DAG = DCI.DAG;
4319 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4320 EVT VT = N->getValueType(0);
4321 unsigned Opc = N->getOpcode();
4322 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4323 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4324 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4325 ISD::CondCode CC = ISD::SETCC_INVALID;
4328 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4330 SDValue CCOp = Slct.getOperand(0);
4331 if (CCOp.getOpcode() == ISD::SETCC)
4332 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4335 bool DoXform = false;
4337 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4340 if (LHS.getOpcode() == ISD::Constant &&
4341 cast<ConstantSDNode>(LHS)->isNullValue()) {
4343 } else if (CC != ISD::SETCC_INVALID &&
4344 RHS.getOpcode() == ISD::Constant &&
4345 cast<ConstantSDNode>(RHS)->isNullValue()) {
4346 std::swap(LHS, RHS);
4347 SDValue Op0 = Slct.getOperand(0);
4348 EVT OpVT = isSlctCC ? Op0.getValueType() :
4349 Op0.getOperand(0).getValueType();
4350 bool isInt = OpVT.isInteger();
4351 CC = ISD::getSetCCInverse(CC, isInt);
4353 if (!TLI.isCondCodeLegal(CC, OpVT))
4354 return SDValue(); // Inverse operator isn't legal.
4361 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4363 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4364 Slct.getOperand(0), Slct.getOperand(1), CC);
4365 SDValue CCOp = Slct.getOperand(0);
4367 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4368 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4369 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4370 CCOp, OtherOp, Result);
4375 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4376 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4377 /// called with the default operands, and if that fails, with commuted
4379 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4380 TargetLowering::DAGCombinerInfo &DCI) {
4381 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4382 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4383 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4384 if (Result.getNode()) return Result;
4389 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4391 static SDValue PerformADDCombine(SDNode *N,
4392 TargetLowering::DAGCombinerInfo &DCI) {
4393 SDValue N0 = N->getOperand(0);
4394 SDValue N1 = N->getOperand(1);
4396 // First try with the default operand order.
4397 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4398 if (Result.getNode())
4401 // If that didn't work, try again with the operands commuted.
4402 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4405 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4407 static SDValue PerformSUBCombine(SDNode *N,
4408 TargetLowering::DAGCombinerInfo &DCI) {
4409 SDValue N0 = N->getOperand(0);
4410 SDValue N1 = N->getOperand(1);
4412 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4413 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4414 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4415 if (Result.getNode()) return Result;
4421 static SDValue PerformMULCombine(SDNode *N,
4422 TargetLowering::DAGCombinerInfo &DCI,
4423 const ARMSubtarget *Subtarget) {
4424 SelectionDAG &DAG = DCI.DAG;
4426 if (Subtarget->isThumb1Only())
4429 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4432 EVT VT = N->getValueType(0);
4436 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4440 uint64_t MulAmt = C->getZExtValue();
4441 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4442 ShiftAmt = ShiftAmt & (32 - 1);
4443 SDValue V = N->getOperand(0);
4444 DebugLoc DL = N->getDebugLoc();
4447 MulAmt >>= ShiftAmt;
4448 if (isPowerOf2_32(MulAmt - 1)) {
4449 // (mul x, 2^N + 1) => (add (shl x, N), x)
4450 Res = DAG.getNode(ISD::ADD, DL, VT,
4451 V, DAG.getNode(ISD::SHL, DL, VT,
4452 V, DAG.getConstant(Log2_32(MulAmt-1),
4454 } else if (isPowerOf2_32(MulAmt + 1)) {
4455 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4456 Res = DAG.getNode(ISD::SUB, DL, VT,
4457 DAG.getNode(ISD::SHL, DL, VT,
4458 V, DAG.getConstant(Log2_32(MulAmt+1),
4465 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4466 DAG.getConstant(ShiftAmt, MVT::i32));
4468 // Do not add new nodes to DAG combiner worklist.
4469 DCI.CombineTo(N, Res, false);
4473 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4474 static SDValue PerformORCombine(SDNode *N,
4475 TargetLowering::DAGCombinerInfo &DCI,
4476 const ARMSubtarget *Subtarget) {
4477 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4480 // BFI is only available on V6T2+
4481 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4484 SelectionDAG &DAG = DCI.DAG;
4485 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4486 DebugLoc DL = N->getDebugLoc();
4487 // 1) or (and A, mask), val => ARMbfi A, val, mask
4488 // iff (val & mask) == val
4490 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4491 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4492 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4493 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4494 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4495 // (i.e., copy a bitfield value into another bitfield of the same width)
4496 if (N0.getOpcode() != ISD::AND)
4499 EVT VT = N->getValueType(0);
4504 // The value and the mask need to be constants so we can verify this is
4505 // actually a bitfield set. If the mask is 0xffff, we can do better
4506 // via a movt instruction, so don't use BFI in that case.
4507 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4510 unsigned Mask = C->getZExtValue();
4514 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4515 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4516 unsigned Val = C->getZExtValue();
4517 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4519 Val >>= CountTrailingZeros_32(~Mask);
4521 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4522 DAG.getConstant(Val, MVT::i32),
4523 DAG.getConstant(Mask, MVT::i32));
4525 // Do not add new nodes to DAG combiner worklist.
4526 DCI.CombineTo(N, Res, false);
4527 } else if (N1.getOpcode() == ISD::AND) {
4528 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4529 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4532 unsigned Mask2 = C->getZExtValue();
4534 if (ARM::isBitFieldInvertedMask(Mask) &&
4535 ARM::isBitFieldInvertedMask(~Mask2) &&
4536 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4537 // The pack halfword instruction works better for masks that fit it,
4538 // so use that when it's available.
4539 if (Subtarget->hasT2ExtractPack() &&
4540 (Mask == 0xffff || Mask == 0xffff0000))
4543 unsigned lsb = CountTrailingZeros_32(Mask2);
4544 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4545 DAG.getConstant(lsb, MVT::i32));
4546 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4547 DAG.getConstant(Mask, MVT::i32));
4548 // Do not add new nodes to DAG combiner worklist.
4549 DCI.CombineTo(N, Res, false);
4550 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4551 ARM::isBitFieldInvertedMask(Mask2) &&
4552 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4553 // The pack halfword instruction works better for masks that fit it,
4554 // so use that when it's available.
4555 if (Subtarget->hasT2ExtractPack() &&
4556 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4559 unsigned lsb = CountTrailingZeros_32(Mask);
4560 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4561 DAG.getConstant(lsb, MVT::i32));
4562 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4563 DAG.getConstant(Mask2, MVT::i32));
4564 // Do not add new nodes to DAG combiner worklist.
4565 DCI.CombineTo(N, Res, false);
4572 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4573 /// ARMISD::VMOVRRD.
4574 static SDValue PerformVMOVRRDCombine(SDNode *N,
4575 TargetLowering::DAGCombinerInfo &DCI) {
4576 // vmovrrd(vmovdrr x, y) -> x,y
4577 SDValue InDouble = N->getOperand(0);
4578 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4579 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4583 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4584 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4585 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4586 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4587 SDValue Op0 = N->getOperand(0);
4588 SDValue Op1 = N->getOperand(1);
4589 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4590 Op0 = Op0.getOperand(0);
4591 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4592 Op1 = Op1.getOperand(0);
4593 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4594 Op0.getNode() == Op1.getNode() &&
4595 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4596 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4597 N->getValueType(0), Op0.getOperand(0));
4601 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4602 /// ISD::BUILD_VECTOR.
4603 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4604 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4605 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4606 // into a pair of GPRs, which is fine when the value is used as a scalar,
4607 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4608 if (N->getNumOperands() == 2)
4609 return PerformVMOVDRRCombine(N, DAG);
4614 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
4615 /// ISD::VECTOR_SHUFFLE.
4616 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
4617 // The LLVM shufflevector instruction does not require the shuffle mask
4618 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
4619 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
4620 // operands do not match the mask length, they are extended by concatenating
4621 // them with undef vectors. That is probably the right thing for other
4622 // targets, but for NEON it is better to concatenate two double-register
4623 // size vector operands into a single quad-register size vector. Do that
4624 // transformation here:
4625 // shuffle(concat(v1, undef), concat(v2, undef)) ->
4626 // shuffle(concat(v1, v2), undef)
4627 SDValue Op0 = N->getOperand(0);
4628 SDValue Op1 = N->getOperand(1);
4629 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
4630 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
4631 Op0.getNumOperands() != 2 ||
4632 Op1.getNumOperands() != 2)
4634 SDValue Concat0Op1 = Op0.getOperand(1);
4635 SDValue Concat1Op1 = Op1.getOperand(1);
4636 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
4637 Concat1Op1.getOpcode() != ISD::UNDEF)
4639 // Skip the transformation if any of the types are illegal.
4640 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4641 EVT VT = N->getValueType(0);
4642 if (!TLI.isTypeLegal(VT) ||
4643 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
4644 !TLI.isTypeLegal(Concat1Op1.getValueType()))
4647 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
4648 Op0.getOperand(0), Op1.getOperand(0));
4649 // Translate the shuffle mask.
4650 SmallVector<int, 16> NewMask;
4651 unsigned NumElts = VT.getVectorNumElements();
4652 unsigned HalfElts = NumElts/2;
4653 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
4654 for (unsigned n = 0; n < NumElts; ++n) {
4655 int MaskElt = SVN->getMaskElt(n);
4657 if (MaskElt < (int)HalfElts)
4659 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
4660 NewElt = HalfElts + MaskElt - NumElts;
4661 NewMask.push_back(NewElt);
4663 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
4664 DAG.getUNDEF(VT), NewMask.data());
4667 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4668 /// ARMISD::VDUPLANE.
4669 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
4670 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4672 SDValue Op = N->getOperand(0);
4673 EVT VT = N->getValueType(0);
4675 // Ignore bit_converts.
4676 while (Op.getOpcode() == ISD::BIT_CONVERT)
4677 Op = Op.getOperand(0);
4678 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4681 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4682 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4683 // The canonical VMOV for a zero vector uses a 32-bit element size.
4684 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4686 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4688 if (EltSize > VT.getVectorElementType().getSizeInBits())
4691 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4694 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4695 /// operand of a vector shift operation, where all the elements of the
4696 /// build_vector must have the same constant integer value.
4697 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4698 // Ignore bit_converts.
4699 while (Op.getOpcode() == ISD::BIT_CONVERT)
4700 Op = Op.getOperand(0);
4701 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4702 APInt SplatBits, SplatUndef;
4703 unsigned SplatBitSize;
4705 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4706 HasAnyUndefs, ElementBits) ||
4707 SplatBitSize > ElementBits)
4709 Cnt = SplatBits.getSExtValue();
4713 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4714 /// operand of a vector shift left operation. That value must be in the range:
4715 /// 0 <= Value < ElementBits for a left shift; or
4716 /// 0 <= Value <= ElementBits for a long left shift.
4717 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4718 assert(VT.isVector() && "vector shift count is not a vector type");
4719 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4720 if (! getVShiftImm(Op, ElementBits, Cnt))
4722 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4725 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4726 /// operand of a vector shift right operation. For a shift opcode, the value
4727 /// is positive, but for an intrinsic the value count must be negative. The
4728 /// absolute value must be in the range:
4729 /// 1 <= |Value| <= ElementBits for a right shift; or
4730 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4731 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4733 assert(VT.isVector() && "vector shift count is not a vector type");
4734 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4735 if (! getVShiftImm(Op, ElementBits, Cnt))
4739 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4742 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4743 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4744 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4747 // Don't do anything for most intrinsics.
4750 // Vector shifts: check for immediate versions and lower them.
4751 // Note: This is done during DAG combining instead of DAG legalizing because
4752 // the build_vectors for 64-bit vector element shift counts are generally
4753 // not legal, and it is hard to see their values after they get legalized to
4754 // loads from a constant pool.
4755 case Intrinsic::arm_neon_vshifts:
4756 case Intrinsic::arm_neon_vshiftu:
4757 case Intrinsic::arm_neon_vshiftls:
4758 case Intrinsic::arm_neon_vshiftlu:
4759 case Intrinsic::arm_neon_vshiftn:
4760 case Intrinsic::arm_neon_vrshifts:
4761 case Intrinsic::arm_neon_vrshiftu:
4762 case Intrinsic::arm_neon_vrshiftn:
4763 case Intrinsic::arm_neon_vqshifts:
4764 case Intrinsic::arm_neon_vqshiftu:
4765 case Intrinsic::arm_neon_vqshiftsu:
4766 case Intrinsic::arm_neon_vqshiftns:
4767 case Intrinsic::arm_neon_vqshiftnu:
4768 case Intrinsic::arm_neon_vqshiftnsu:
4769 case Intrinsic::arm_neon_vqrshiftns:
4770 case Intrinsic::arm_neon_vqrshiftnu:
4771 case Intrinsic::arm_neon_vqrshiftnsu: {
4772 EVT VT = N->getOperand(1).getValueType();
4774 unsigned VShiftOpc = 0;
4777 case Intrinsic::arm_neon_vshifts:
4778 case Intrinsic::arm_neon_vshiftu:
4779 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4780 VShiftOpc = ARMISD::VSHL;
4783 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4784 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4785 ARMISD::VSHRs : ARMISD::VSHRu);
4790 case Intrinsic::arm_neon_vshiftls:
4791 case Intrinsic::arm_neon_vshiftlu:
4792 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4794 llvm_unreachable("invalid shift count for vshll intrinsic");
4796 case Intrinsic::arm_neon_vrshifts:
4797 case Intrinsic::arm_neon_vrshiftu:
4798 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4802 case Intrinsic::arm_neon_vqshifts:
4803 case Intrinsic::arm_neon_vqshiftu:
4804 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4808 case Intrinsic::arm_neon_vqshiftsu:
4809 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4811 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4813 case Intrinsic::arm_neon_vshiftn:
4814 case Intrinsic::arm_neon_vrshiftn:
4815 case Intrinsic::arm_neon_vqshiftns:
4816 case Intrinsic::arm_neon_vqshiftnu:
4817 case Intrinsic::arm_neon_vqshiftnsu:
4818 case Intrinsic::arm_neon_vqrshiftns:
4819 case Intrinsic::arm_neon_vqrshiftnu:
4820 case Intrinsic::arm_neon_vqrshiftnsu:
4821 // Narrowing shifts require an immediate right shift.
4822 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4824 llvm_unreachable("invalid shift count for narrowing vector shift "
4828 llvm_unreachable("unhandled vector shift");
4832 case Intrinsic::arm_neon_vshifts:
4833 case Intrinsic::arm_neon_vshiftu:
4834 // Opcode already set above.
4836 case Intrinsic::arm_neon_vshiftls:
4837 case Intrinsic::arm_neon_vshiftlu:
4838 if (Cnt == VT.getVectorElementType().getSizeInBits())
4839 VShiftOpc = ARMISD::VSHLLi;
4841 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4842 ARMISD::VSHLLs : ARMISD::VSHLLu);
4844 case Intrinsic::arm_neon_vshiftn:
4845 VShiftOpc = ARMISD::VSHRN; break;
4846 case Intrinsic::arm_neon_vrshifts:
4847 VShiftOpc = ARMISD::VRSHRs; break;
4848 case Intrinsic::arm_neon_vrshiftu:
4849 VShiftOpc = ARMISD::VRSHRu; break;
4850 case Intrinsic::arm_neon_vrshiftn:
4851 VShiftOpc = ARMISD::VRSHRN; break;
4852 case Intrinsic::arm_neon_vqshifts:
4853 VShiftOpc = ARMISD::VQSHLs; break;
4854 case Intrinsic::arm_neon_vqshiftu:
4855 VShiftOpc = ARMISD::VQSHLu; break;
4856 case Intrinsic::arm_neon_vqshiftsu:
4857 VShiftOpc = ARMISD::VQSHLsu; break;
4858 case Intrinsic::arm_neon_vqshiftns:
4859 VShiftOpc = ARMISD::VQSHRNs; break;
4860 case Intrinsic::arm_neon_vqshiftnu:
4861 VShiftOpc = ARMISD::VQSHRNu; break;
4862 case Intrinsic::arm_neon_vqshiftnsu:
4863 VShiftOpc = ARMISD::VQSHRNsu; break;
4864 case Intrinsic::arm_neon_vqrshiftns:
4865 VShiftOpc = ARMISD::VQRSHRNs; break;
4866 case Intrinsic::arm_neon_vqrshiftnu:
4867 VShiftOpc = ARMISD::VQRSHRNu; break;
4868 case Intrinsic::arm_neon_vqrshiftnsu:
4869 VShiftOpc = ARMISD::VQRSHRNsu; break;
4872 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4873 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4876 case Intrinsic::arm_neon_vshiftins: {
4877 EVT VT = N->getOperand(1).getValueType();
4879 unsigned VShiftOpc = 0;
4881 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4882 VShiftOpc = ARMISD::VSLI;
4883 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4884 VShiftOpc = ARMISD::VSRI;
4886 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4889 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4890 N->getOperand(1), N->getOperand(2),
4891 DAG.getConstant(Cnt, MVT::i32));
4894 case Intrinsic::arm_neon_vqrshifts:
4895 case Intrinsic::arm_neon_vqrshiftu:
4896 // No immediate versions of these to check for.
4903 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4904 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4905 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4906 /// vector element shift counts are generally not legal, and it is hard to see
4907 /// their values after they get legalized to loads from a constant pool.
4908 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4909 const ARMSubtarget *ST) {
4910 EVT VT = N->getValueType(0);
4912 // Nothing to be done for scalar shifts.
4913 if (! VT.isVector())
4916 assert(ST->hasNEON() && "unexpected vector shift");
4919 switch (N->getOpcode()) {
4920 default: llvm_unreachable("unexpected shift opcode");
4923 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4924 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4925 DAG.getConstant(Cnt, MVT::i32));
4930 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4931 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4932 ARMISD::VSHRs : ARMISD::VSHRu);
4933 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4934 DAG.getConstant(Cnt, MVT::i32));
4940 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4941 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4942 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4943 const ARMSubtarget *ST) {
4944 SDValue N0 = N->getOperand(0);
4946 // Check for sign- and zero-extensions of vector extract operations of 8-
4947 // and 16-bit vector elements. NEON supports these directly. They are
4948 // handled during DAG combining because type legalization will promote them
4949 // to 32-bit types and it is messy to recognize the operations after that.
4950 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4951 SDValue Vec = N0.getOperand(0);
4952 SDValue Lane = N0.getOperand(1);
4953 EVT VT = N->getValueType(0);
4954 EVT EltVT = N0.getValueType();
4955 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4957 if (VT == MVT::i32 &&
4958 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4959 TLI.isTypeLegal(Vec.getValueType()) &&
4960 isa<ConstantSDNode>(Lane)) {
4963 switch (N->getOpcode()) {
4964 default: llvm_unreachable("unexpected opcode");
4965 case ISD::SIGN_EXTEND:
4966 Opc = ARMISD::VGETLANEs;
4968 case ISD::ZERO_EXTEND:
4969 case ISD::ANY_EXTEND:
4970 Opc = ARMISD::VGETLANEu;
4973 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4980 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4981 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4982 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4983 const ARMSubtarget *ST) {
4984 // If the target supports NEON, try to use vmax/vmin instructions for f32
4985 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4986 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4987 // a NaN; only do the transformation when it matches that behavior.
4989 // For now only do this when using NEON for FP operations; if using VFP, it
4990 // is not obvious that the benefit outweighs the cost of switching to the
4992 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4993 N->getValueType(0) != MVT::f32)
4996 SDValue CondLHS = N->getOperand(0);
4997 SDValue CondRHS = N->getOperand(1);
4998 SDValue LHS = N->getOperand(2);
4999 SDValue RHS = N->getOperand(3);
5000 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
5002 unsigned Opcode = 0;
5004 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
5005 IsReversed = false; // x CC y ? x : y
5006 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
5007 IsReversed = true ; // x CC y ? y : x
5021 // If LHS is NaN, an ordered comparison will be false and the result will
5022 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
5023 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5024 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
5025 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5027 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
5028 // will return -0, so vmin can only be used for unsafe math or if one of
5029 // the operands is known to be nonzero.
5030 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
5032 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5034 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
5043 // If LHS is NaN, an ordered comparison will be false and the result will
5044 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
5045 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
5046 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
5047 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
5049 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
5050 // will return +0, so vmax can only be used for unsafe math or if one of
5051 // the operands is known to be nonzero.
5052 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
5054 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
5056 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
5062 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
5065 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
5066 DAGCombinerInfo &DCI) const {
5067 switch (N->getOpcode()) {
5069 case ISD::ADD: return PerformADDCombine(N, DCI);
5070 case ISD::SUB: return PerformSUBCombine(N, DCI);
5071 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
5072 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
5073 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
5074 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
5075 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
5076 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
5077 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
5078 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
5081 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
5082 case ISD::SIGN_EXTEND:
5083 case ISD::ZERO_EXTEND:
5084 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
5085 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
5090 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
5091 if (!Subtarget->allowsUnalignedMem())
5094 switch (VT.getSimpleVT().SimpleTy) {
5101 // FIXME: VLD1 etc with standard alignment is legal.
5105 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
5110 switch (VT.getSimpleVT().SimpleTy) {
5111 default: return false;
5126 if ((V & (Scale - 1)) != 0)
5129 return V == (V & ((1LL << 5) - 1));
5132 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
5133 const ARMSubtarget *Subtarget) {
5140 switch (VT.getSimpleVT().SimpleTy) {
5141 default: return false;
5146 // + imm12 or - imm8
5148 return V == (V & ((1LL << 8) - 1));
5149 return V == (V & ((1LL << 12) - 1));
5152 // Same as ARM mode. FIXME: NEON?
5153 if (!Subtarget->hasVFP2())
5158 return V == (V & ((1LL << 8) - 1));
5162 /// isLegalAddressImmediate - Return true if the integer value can be used
5163 /// as the offset of the target addressing mode for load / store of the
5165 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5166 const ARMSubtarget *Subtarget) {
5173 if (Subtarget->isThumb1Only())
5174 return isLegalT1AddressImmediate(V, VT);
5175 else if (Subtarget->isThumb2())
5176 return isLegalT2AddressImmediate(V, VT, Subtarget);
5181 switch (VT.getSimpleVT().SimpleTy) {
5182 default: return false;
5187 return V == (V & ((1LL << 12) - 1));
5190 return V == (V & ((1LL << 8) - 1));
5193 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5198 return V == (V & ((1LL << 8) - 1));
5202 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5204 int Scale = AM.Scale;
5208 switch (VT.getSimpleVT().SimpleTy) {
5209 default: return false;
5218 return Scale == 2 || Scale == 4 || Scale == 8;
5221 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5225 // Note, we allow "void" uses (basically, uses that aren't loads or
5226 // stores), because arm allows folding a scale into many arithmetic
5227 // operations. This should be made more precise and revisited later.
5229 // Allow r << imm, but the imm has to be a multiple of two.
5230 if (Scale & 1) return false;
5231 return isPowerOf2_32(Scale);
5235 /// isLegalAddressingMode - Return true if the addressing mode represented
5236 /// by AM is legal for this target, for a load/store of the specified type.
5237 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5238 const Type *Ty) const {
5239 EVT VT = getValueType(Ty, true);
5240 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5243 // Can never fold addr of global into load/store.
5248 case 0: // no scale reg, must be "r+i" or "r", or "i".
5251 if (Subtarget->isThumb1Only())
5255 // ARM doesn't support any R+R*scale+imm addr modes.
5262 if (Subtarget->isThumb2())
5263 return isLegalT2ScaledAddressingMode(AM, VT);
5265 int Scale = AM.Scale;
5266 switch (VT.getSimpleVT().SimpleTy) {
5267 default: return false;
5271 if (Scale < 0) Scale = -Scale;
5275 return isPowerOf2_32(Scale & ~1);
5279 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5284 // Note, we allow "void" uses (basically, uses that aren't loads or
5285 // stores), because arm allows folding a scale into many arithmetic
5286 // operations. This should be made more precise and revisited later.
5288 // Allow r << imm, but the imm has to be a multiple of two.
5289 if (Scale & 1) return false;
5290 return isPowerOf2_32(Scale);
5297 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5298 /// icmp immediate, that is the target has icmp instructions which can compare
5299 /// a register against the immediate without having to materialize the
5300 /// immediate into a register.
5301 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5302 if (!Subtarget->isThumb())
5303 return ARM_AM::getSOImmVal(Imm) != -1;
5304 if (Subtarget->isThumb2())
5305 return ARM_AM::getT2SOImmVal(Imm) != -1;
5306 return Imm >= 0 && Imm <= 255;
5309 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5310 bool isSEXTLoad, SDValue &Base,
5311 SDValue &Offset, bool &isInc,
5312 SelectionDAG &DAG) {
5313 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5316 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5318 Base = Ptr->getOperand(0);
5319 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5320 int RHSC = (int)RHS->getZExtValue();
5321 if (RHSC < 0 && RHSC > -256) {
5322 assert(Ptr->getOpcode() == ISD::ADD);
5324 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5328 isInc = (Ptr->getOpcode() == ISD::ADD);
5329 Offset = Ptr->getOperand(1);
5331 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5333 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5334 int RHSC = (int)RHS->getZExtValue();
5335 if (RHSC < 0 && RHSC > -0x1000) {
5336 assert(Ptr->getOpcode() == ISD::ADD);
5338 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5339 Base = Ptr->getOperand(0);
5344 if (Ptr->getOpcode() == ISD::ADD) {
5346 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5347 if (ShOpcVal != ARM_AM::no_shift) {
5348 Base = Ptr->getOperand(1);
5349 Offset = Ptr->getOperand(0);
5351 Base = Ptr->getOperand(0);
5352 Offset = Ptr->getOperand(1);
5357 isInc = (Ptr->getOpcode() == ISD::ADD);
5358 Base = Ptr->getOperand(0);
5359 Offset = Ptr->getOperand(1);
5363 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5367 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5368 bool isSEXTLoad, SDValue &Base,
5369 SDValue &Offset, bool &isInc,
5370 SelectionDAG &DAG) {
5371 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5374 Base = Ptr->getOperand(0);
5375 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5376 int RHSC = (int)RHS->getZExtValue();
5377 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5378 assert(Ptr->getOpcode() == ISD::ADD);
5380 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5382 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5383 isInc = Ptr->getOpcode() == ISD::ADD;
5384 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5392 /// getPreIndexedAddressParts - returns true by value, base pointer and
5393 /// offset pointer and addressing mode by reference if the node's address
5394 /// can be legally represented as pre-indexed load / store address.
5396 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5398 ISD::MemIndexedMode &AM,
5399 SelectionDAG &DAG) const {
5400 if (Subtarget->isThumb1Only())
5405 bool isSEXTLoad = false;
5406 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5407 Ptr = LD->getBasePtr();
5408 VT = LD->getMemoryVT();
5409 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5410 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5411 Ptr = ST->getBasePtr();
5412 VT = ST->getMemoryVT();
5417 bool isLegal = false;
5418 if (Subtarget->isThumb2())
5419 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5420 Offset, isInc, DAG);
5422 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5423 Offset, isInc, DAG);
5427 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5431 /// getPostIndexedAddressParts - returns true by value, base pointer and
5432 /// offset pointer and addressing mode by reference if this node can be
5433 /// combined with a load / store to form a post-indexed load / store.
5434 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5437 ISD::MemIndexedMode &AM,
5438 SelectionDAG &DAG) const {
5439 if (Subtarget->isThumb1Only())
5444 bool isSEXTLoad = false;
5445 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5446 VT = LD->getMemoryVT();
5447 Ptr = LD->getBasePtr();
5448 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5449 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5450 VT = ST->getMemoryVT();
5451 Ptr = ST->getBasePtr();
5456 bool isLegal = false;
5457 if (Subtarget->isThumb2())
5458 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5461 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5467 // Swap base ptr and offset to catch more post-index load / store when
5468 // it's legal. In Thumb2 mode, offset must be an immediate.
5469 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5470 !Subtarget->isThumb2())
5471 std::swap(Base, Offset);
5473 // Post-indexed load / store update the base pointer.
5478 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5482 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5486 const SelectionDAG &DAG,
5487 unsigned Depth) const {
5488 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5489 switch (Op.getOpcode()) {
5491 case ARMISD::CMOV: {
5492 // Bits are known zero/one if known on the LHS and RHS.
5493 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5494 if (KnownZero == 0 && KnownOne == 0) return;
5496 APInt KnownZeroRHS, KnownOneRHS;
5497 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5498 KnownZeroRHS, KnownOneRHS, Depth+1);
5499 KnownZero &= KnownZeroRHS;
5500 KnownOne &= KnownOneRHS;
5506 //===----------------------------------------------------------------------===//
5507 // ARM Inline Assembly Support
5508 //===----------------------------------------------------------------------===//
5510 /// getConstraintType - Given a constraint letter, return the type of
5511 /// constraint it is for this target.
5512 ARMTargetLowering::ConstraintType
5513 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5514 if (Constraint.size() == 1) {
5515 switch (Constraint[0]) {
5517 case 'l': return C_RegisterClass;
5518 case 'w': return C_RegisterClass;
5521 return TargetLowering::getConstraintType(Constraint);
5524 /// Examine constraint type and operand type and determine a weight value.
5525 /// This object must already have been set up with the operand type
5526 /// and the current alternative constraint selected.
5527 TargetLowering::ConstraintWeight
5528 ARMTargetLowering::getSingleConstraintMatchWeight(
5529 AsmOperandInfo &info, const char *constraint) const {
5530 ConstraintWeight weight = CW_Invalid;
5531 Value *CallOperandVal = info.CallOperandVal;
5532 // If we don't have a value, we can't do a match,
5533 // but allow it at the lowest weight.
5534 if (CallOperandVal == NULL)
5536 const Type *type = CallOperandVal->getType();
5537 // Look at the constraint type.
5538 switch (*constraint) {
5540 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5543 if (type->isIntegerTy()) {
5544 if (Subtarget->isThumb())
5545 weight = CW_SpecificReg;
5547 weight = CW_Register;
5551 if (type->isFloatingPointTy())
5552 weight = CW_Register;
5558 std::pair<unsigned, const TargetRegisterClass*>
5559 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5561 if (Constraint.size() == 1) {
5562 // GCC ARM Constraint Letters
5563 switch (Constraint[0]) {
5565 if (Subtarget->isThumb())
5566 return std::make_pair(0U, ARM::tGPRRegisterClass);
5568 return std::make_pair(0U, ARM::GPRRegisterClass);
5570 return std::make_pair(0U, ARM::GPRRegisterClass);
5573 return std::make_pair(0U, ARM::SPRRegisterClass);
5574 if (VT.getSizeInBits() == 64)
5575 return std::make_pair(0U, ARM::DPRRegisterClass);
5576 if (VT.getSizeInBits() == 128)
5577 return std::make_pair(0U, ARM::QPRRegisterClass);
5581 if (StringRef("{cc}").equals_lower(Constraint))
5582 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5584 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5587 std::vector<unsigned> ARMTargetLowering::
5588 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5590 if (Constraint.size() != 1)
5591 return std::vector<unsigned>();
5593 switch (Constraint[0]) { // GCC ARM Constraint Letters
5596 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5597 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5600 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5601 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5602 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5603 ARM::R12, ARM::LR, 0);
5606 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5607 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5608 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5609 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5610 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5611 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5612 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5613 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5614 if (VT.getSizeInBits() == 64)
5615 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5616 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5617 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5618 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5619 if (VT.getSizeInBits() == 128)
5620 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5621 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5625 return std::vector<unsigned>();
5628 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5629 /// vector. If it is invalid, don't add anything to Ops.
5630 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5632 std::vector<SDValue>&Ops,
5633 SelectionDAG &DAG) const {
5634 SDValue Result(0, 0);
5636 switch (Constraint) {
5638 case 'I': case 'J': case 'K': case 'L':
5639 case 'M': case 'N': case 'O':
5640 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5644 int64_t CVal64 = C->getSExtValue();
5645 int CVal = (int) CVal64;
5646 // None of these constraints allow values larger than 32 bits. Check
5647 // that the value fits in an int.
5651 switch (Constraint) {
5653 if (Subtarget->isThumb1Only()) {
5654 // This must be a constant between 0 and 255, for ADD
5656 if (CVal >= 0 && CVal <= 255)
5658 } else if (Subtarget->isThumb2()) {
5659 // A constant that can be used as an immediate value in a
5660 // data-processing instruction.
5661 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5664 // A constant that can be used as an immediate value in a
5665 // data-processing instruction.
5666 if (ARM_AM::getSOImmVal(CVal) != -1)
5672 if (Subtarget->isThumb()) { // FIXME thumb2
5673 // This must be a constant between -255 and -1, for negated ADD
5674 // immediates. This can be used in GCC with an "n" modifier that
5675 // prints the negated value, for use with SUB instructions. It is
5676 // not useful otherwise but is implemented for compatibility.
5677 if (CVal >= -255 && CVal <= -1)
5680 // This must be a constant between -4095 and 4095. It is not clear
5681 // what this constraint is intended for. Implemented for
5682 // compatibility with GCC.
5683 if (CVal >= -4095 && CVal <= 4095)
5689 if (Subtarget->isThumb1Only()) {
5690 // A 32-bit value where only one byte has a nonzero value. Exclude
5691 // zero to match GCC. This constraint is used by GCC internally for
5692 // constants that can be loaded with a move/shift combination.
5693 // It is not useful otherwise but is implemented for compatibility.
5694 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5696 } else if (Subtarget->isThumb2()) {
5697 // A constant whose bitwise inverse can be used as an immediate
5698 // value in a data-processing instruction. This can be used in GCC
5699 // with a "B" modifier that prints the inverted value, for use with
5700 // BIC and MVN instructions. It is not useful otherwise but is
5701 // implemented for compatibility.
5702 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5705 // A constant whose bitwise inverse can be used as an immediate
5706 // value in a data-processing instruction. This can be used in GCC
5707 // with a "B" modifier that prints the inverted value, for use with
5708 // BIC and MVN instructions. It is not useful otherwise but is
5709 // implemented for compatibility.
5710 if (ARM_AM::getSOImmVal(~CVal) != -1)
5716 if (Subtarget->isThumb1Only()) {
5717 // This must be a constant between -7 and 7,
5718 // for 3-operand ADD/SUB immediate instructions.
5719 if (CVal >= -7 && CVal < 7)
5721 } else if (Subtarget->isThumb2()) {
5722 // A constant whose negation can be used as an immediate value in a
5723 // data-processing instruction. This can be used in GCC with an "n"
5724 // modifier that prints the negated value, for use with SUB
5725 // instructions. It is not useful otherwise but is implemented for
5727 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5730 // A constant whose negation can be used as an immediate value in a
5731 // data-processing instruction. This can be used in GCC with an "n"
5732 // modifier that prints the negated value, for use with SUB
5733 // instructions. It is not useful otherwise but is implemented for
5735 if (ARM_AM::getSOImmVal(-CVal) != -1)
5741 if (Subtarget->isThumb()) { // FIXME thumb2
5742 // This must be a multiple of 4 between 0 and 1020, for
5743 // ADD sp + immediate.
5744 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5747 // A power of two or a constant between 0 and 32. This is used in
5748 // GCC for the shift amount on shifted register operands, but it is
5749 // useful in general for any shift amounts.
5750 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5756 if (Subtarget->isThumb()) { // FIXME thumb2
5757 // This must be a constant between 0 and 31, for shift amounts.
5758 if (CVal >= 0 && CVal <= 31)
5764 if (Subtarget->isThumb()) { // FIXME thumb2
5765 // This must be a multiple of 4 between -508 and 508, for
5766 // ADD/SUB sp = sp + immediate.
5767 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5772 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5776 if (Result.getNode()) {
5777 Ops.push_back(Result);
5780 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5784 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5785 // The ARM target isn't yet aware of offsets.
5789 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5790 APInt Imm = FPImm.bitcastToAPInt();
5791 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5792 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5793 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5795 // We can handle 4 bits of mantissa.
5796 // mantissa = (16+UInt(e:f:g:h))/16.
5797 if (Mantissa & 0x7ffff)
5800 if ((Mantissa & 0xf) != Mantissa)
5803 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5804 if (Exp < -3 || Exp > 4)
5806 Exp = ((Exp+3) & 0x7) ^ 4;
5808 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5811 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5812 APInt Imm = FPImm.bitcastToAPInt();
5813 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5814 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5815 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5817 // We can handle 4 bits of mantissa.
5818 // mantissa = (16+UInt(e:f:g:h))/16.
5819 if (Mantissa & 0xffffffffffffLL)
5822 if ((Mantissa & 0xf) != Mantissa)
5825 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5826 if (Exp < -3 || Exp > 4)
5828 Exp = ((Exp+3) & 0x7) ^ 4;
5830 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5833 bool ARM::isBitFieldInvertedMask(unsigned v) {
5834 if (v == 0xffffffff)
5836 // there can be 1's on either or both "outsides", all the "inside"
5838 unsigned int lsb = 0, msb = 31;
5839 while (v & (1 << msb)) --msb;
5840 while (v & (1 << lsb)) ++lsb;
5841 for (unsigned int i = lsb; i <= msb; ++i) {
5848 /// isFPImmLegal - Returns true if the target can instruction select the
5849 /// specified FP immediate natively. If false, the legalizer will
5850 /// materialize the FP immediate as a load from a constant pool.
5851 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5852 if (!Subtarget->hasVFP3())
5855 return ARM::getVFPf32Imm(Imm) != -1;
5857 return ARM::getVFPf64Imm(Imm) != -1;
5861 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5862 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5863 /// specified in the intrinsic calls.
5864 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5866 unsigned Intrinsic) const {
5867 switch (Intrinsic) {
5868 case Intrinsic::arm_neon_vld1:
5869 case Intrinsic::arm_neon_vld2:
5870 case Intrinsic::arm_neon_vld3:
5871 case Intrinsic::arm_neon_vld4:
5872 case Intrinsic::arm_neon_vld2lane:
5873 case Intrinsic::arm_neon_vld3lane:
5874 case Intrinsic::arm_neon_vld4lane: {
5875 Info.opc = ISD::INTRINSIC_W_CHAIN;
5876 // Conservatively set memVT to the entire set of vectors loaded.
5877 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
5878 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5879 Info.ptrVal = I.getArgOperand(0);
5881 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5882 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5883 Info.vol = false; // volatile loads with NEON intrinsics not supported
5884 Info.readMem = true;
5885 Info.writeMem = false;
5888 case Intrinsic::arm_neon_vst1:
5889 case Intrinsic::arm_neon_vst2:
5890 case Intrinsic::arm_neon_vst3:
5891 case Intrinsic::arm_neon_vst4:
5892 case Intrinsic::arm_neon_vst2lane:
5893 case Intrinsic::arm_neon_vst3lane:
5894 case Intrinsic::arm_neon_vst4lane: {
5895 Info.opc = ISD::INTRINSIC_VOID;
5896 // Conservatively set memVT to the entire set of vectors stored.
5897 unsigned NumElts = 0;
5898 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5899 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
5900 if (!ArgTy->isVectorTy())
5902 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
5904 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5905 Info.ptrVal = I.getArgOperand(0);
5907 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5908 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5909 Info.vol = false; // volatile stores with NEON intrinsics not supported
5910 Info.readMem = false;
5911 Info.writeMem = true;