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)"),
73 ARMFastCC("arm-fastcc", cl::Hidden,
74 cl::desc("Use AAPCS / AAPCS-VFP calling conventions for fastcc"));
76 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
77 EVT PromotedBitwiseVT) {
78 if (VT != PromotedLdStVT) {
79 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
80 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
81 PromotedLdStVT.getSimpleVT());
83 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
84 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
85 PromotedLdStVT.getSimpleVT());
88 EVT ElemTy = VT.getVectorElementType();
89 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
90 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
91 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
92 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
93 if (ElemTy != MVT::i32) {
94 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
95 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
96 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
97 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
99 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
100 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
101 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
102 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
103 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
104 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
105 if (VT.isInteger()) {
106 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
107 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
108 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
109 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
110 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), 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);
469 computeRegisterProperties();
471 // ARM does not have f32 extending load.
472 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
474 // ARM does not have i1 sign extending load.
475 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
477 // ARM supports all 4 flavors of integer indexed load / store.
478 if (!Subtarget->isThumb1Only()) {
479 for (unsigned im = (unsigned)ISD::PRE_INC;
480 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
481 setIndexedLoadAction(im, MVT::i1, Legal);
482 setIndexedLoadAction(im, MVT::i8, Legal);
483 setIndexedLoadAction(im, MVT::i16, Legal);
484 setIndexedLoadAction(im, MVT::i32, Legal);
485 setIndexedStoreAction(im, MVT::i1, Legal);
486 setIndexedStoreAction(im, MVT::i8, Legal);
487 setIndexedStoreAction(im, MVT::i16, Legal);
488 setIndexedStoreAction(im, MVT::i32, Legal);
492 // i64 operation support.
493 if (Subtarget->isThumb1Only()) {
494 setOperationAction(ISD::MUL, MVT::i64, Expand);
495 setOperationAction(ISD::MULHU, MVT::i32, Expand);
496 setOperationAction(ISD::MULHS, MVT::i32, Expand);
497 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
498 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
500 setOperationAction(ISD::MUL, MVT::i64, Expand);
501 setOperationAction(ISD::MULHU, MVT::i32, Expand);
502 if (!Subtarget->hasV6Ops())
503 setOperationAction(ISD::MULHS, MVT::i32, Expand);
505 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
506 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
507 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
508 setOperationAction(ISD::SRL, MVT::i64, Custom);
509 setOperationAction(ISD::SRA, MVT::i64, Custom);
511 // ARM does not have ROTL.
512 setOperationAction(ISD::ROTL, MVT::i32, Expand);
513 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
514 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
515 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
516 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
518 // Only ARMv6 has BSWAP.
519 if (!Subtarget->hasV6Ops())
520 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
522 // These are expanded into libcalls.
523 if (!Subtarget->hasDivide()) {
524 // v7M has a hardware divider
525 setOperationAction(ISD::SDIV, MVT::i32, Expand);
526 setOperationAction(ISD::UDIV, MVT::i32, Expand);
528 setOperationAction(ISD::SREM, MVT::i32, Expand);
529 setOperationAction(ISD::UREM, MVT::i32, Expand);
530 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
531 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
533 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
534 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
535 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
536 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
537 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
539 setOperationAction(ISD::TRAP, MVT::Other, Legal);
541 // Use the default implementation.
542 setOperationAction(ISD::VASTART, MVT::Other, Custom);
543 setOperationAction(ISD::VAARG, MVT::Other, Expand);
544 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
545 setOperationAction(ISD::VAEND, MVT::Other, Expand);
546 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
547 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
548 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
549 // FIXME: Shouldn't need this, since no register is used, but the legalizer
550 // doesn't yet know how to not do that for SjLj.
551 setExceptionSelectorRegister(ARM::R0);
552 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
553 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
554 // the default expansion.
555 if (Subtarget->hasDataBarrier() ||
556 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
557 // membarrier needs custom lowering; the rest are legal and handled
559 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
561 // Set them all for expansion, which will force libcalls.
562 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
563 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
564 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
565 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
566 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
567 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
568 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
569 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
570 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
571 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
572 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
573 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
584 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
585 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
586 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
587 // Since the libcalls include locking, fold in the fences
588 setShouldFoldAtomicFences(true);
590 // 64-bit versions are always libcalls (for now)
591 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
592 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
593 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
595 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
596 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
600 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
601 if (!Subtarget->hasV6Ops()) {
602 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
603 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
605 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
607 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
608 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
609 // iff target supports vfp2.
610 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
611 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
614 // We want to custom lower some of our intrinsics.
615 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
616 if (Subtarget->isTargetDarwin()) {
617 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
618 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
619 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
622 setOperationAction(ISD::SETCC, MVT::i32, Expand);
623 setOperationAction(ISD::SETCC, MVT::f32, Expand);
624 setOperationAction(ISD::SETCC, MVT::f64, Expand);
625 setOperationAction(ISD::SELECT, MVT::i32, Custom);
626 setOperationAction(ISD::SELECT, MVT::f32, Custom);
627 setOperationAction(ISD::SELECT, MVT::f64, Custom);
628 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
629 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
630 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
632 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
633 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
634 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
635 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
636 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
638 // We don't support sin/cos/fmod/copysign/pow
639 setOperationAction(ISD::FSIN, MVT::f64, Expand);
640 setOperationAction(ISD::FSIN, MVT::f32, Expand);
641 setOperationAction(ISD::FCOS, MVT::f32, Expand);
642 setOperationAction(ISD::FCOS, MVT::f64, Expand);
643 setOperationAction(ISD::FREM, MVT::f64, Expand);
644 setOperationAction(ISD::FREM, MVT::f32, Expand);
645 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
646 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
647 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
649 setOperationAction(ISD::FPOW, MVT::f64, Expand);
650 setOperationAction(ISD::FPOW, MVT::f32, Expand);
652 // Various VFP goodness
653 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
654 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
655 if (Subtarget->hasVFP2()) {
656 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
657 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
658 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
659 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
661 // Special handling for half-precision FP.
662 if (!Subtarget->hasFP16()) {
663 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
664 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
668 // We have target-specific dag combine patterns for the following nodes:
669 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
670 setTargetDAGCombine(ISD::ADD);
671 setTargetDAGCombine(ISD::SUB);
672 setTargetDAGCombine(ISD::MUL);
674 if (Subtarget->hasV6T2Ops())
675 setTargetDAGCombine(ISD::OR);
677 setStackPointerRegisterToSaveRestore(ARM::SP);
679 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
680 setSchedulingPreference(Sched::RegPressure);
682 setSchedulingPreference(Sched::Hybrid);
684 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
686 // On ARM arguments smaller than 4 bytes are extended, so all arguments
687 // are at least 4 bytes aligned.
688 setMinStackArgumentAlignment(4);
690 benefitFromCodePlacementOpt = true;
693 std::pair<const TargetRegisterClass*, uint8_t>
694 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
695 const TargetRegisterClass *RRC = 0;
697 switch (VT.getSimpleVT().SimpleTy) {
699 return TargetLowering::findRepresentativeClass(VT);
700 // Use DPR as representative register class for all floating point
701 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
702 // the cost is 1 for both f32 and f64.
703 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
704 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
705 RRC = ARM::DPRRegisterClass;
707 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
708 case MVT::v4f32: case MVT::v2f64:
709 RRC = ARM::DPRRegisterClass;
713 RRC = ARM::DPRRegisterClass;
717 RRC = ARM::DPRRegisterClass;
721 return std::make_pair(RRC, Cost);
724 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
727 case ARMISD::Wrapper: return "ARMISD::Wrapper";
728 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
729 case ARMISD::CALL: return "ARMISD::CALL";
730 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
731 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
732 case ARMISD::tCALL: return "ARMISD::tCALL";
733 case ARMISD::BRCOND: return "ARMISD::BRCOND";
734 case ARMISD::BR_JT: return "ARMISD::BR_JT";
735 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
736 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
737 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
738 case ARMISD::CMP: return "ARMISD::CMP";
739 case ARMISD::CMPZ: return "ARMISD::CMPZ";
740 case ARMISD::CMPFP: return "ARMISD::CMPFP";
741 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
742 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
743 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
744 case ARMISD::CMOV: return "ARMISD::CMOV";
745 case ARMISD::CNEG: return "ARMISD::CNEG";
747 case ARMISD::RBIT: return "ARMISD::RBIT";
749 case ARMISD::FTOSI: return "ARMISD::FTOSI";
750 case ARMISD::FTOUI: return "ARMISD::FTOUI";
751 case ARMISD::SITOF: return "ARMISD::SITOF";
752 case ARMISD::UITOF: return "ARMISD::UITOF";
754 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
755 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
756 case ARMISD::RRX: return "ARMISD::RRX";
758 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
759 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
761 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
762 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
763 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
765 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
767 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
769 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
771 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
772 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
774 case ARMISD::VCEQ: return "ARMISD::VCEQ";
775 case ARMISD::VCGE: return "ARMISD::VCGE";
776 case ARMISD::VCGEU: return "ARMISD::VCGEU";
777 case ARMISD::VCGT: return "ARMISD::VCGT";
778 case ARMISD::VCGTU: return "ARMISD::VCGTU";
779 case ARMISD::VTST: return "ARMISD::VTST";
781 case ARMISD::VSHL: return "ARMISD::VSHL";
782 case ARMISD::VSHRs: return "ARMISD::VSHRs";
783 case ARMISD::VSHRu: return "ARMISD::VSHRu";
784 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
785 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
786 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
787 case ARMISD::VSHRN: return "ARMISD::VSHRN";
788 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
789 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
790 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
791 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
792 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
793 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
794 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
795 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
796 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
797 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
798 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
799 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
800 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
801 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
802 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
803 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
804 case ARMISD::VDUP: return "ARMISD::VDUP";
805 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
806 case ARMISD::VEXT: return "ARMISD::VEXT";
807 case ARMISD::VREV64: return "ARMISD::VREV64";
808 case ARMISD::VREV32: return "ARMISD::VREV32";
809 case ARMISD::VREV16: return "ARMISD::VREV16";
810 case ARMISD::VZIP: return "ARMISD::VZIP";
811 case ARMISD::VUZP: return "ARMISD::VUZP";
812 case ARMISD::VTRN: return "ARMISD::VTRN";
813 case ARMISD::VMULLs: return "ARMISD::VMULLs";
814 case ARMISD::VMULLu: return "ARMISD::VMULLu";
815 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
816 case ARMISD::FMAX: return "ARMISD::FMAX";
817 case ARMISD::FMIN: return "ARMISD::FMIN";
818 case ARMISD::BFI: return "ARMISD::BFI";
822 /// getRegClassFor - Return the register class that should be used for the
823 /// specified value type.
824 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
825 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
826 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
827 // load / store 4 to 8 consecutive D registers.
828 if (Subtarget->hasNEON()) {
829 if (VT == MVT::v4i64)
830 return ARM::QQPRRegisterClass;
831 else if (VT == MVT::v8i64)
832 return ARM::QQQQPRRegisterClass;
834 return TargetLowering::getRegClassFor(VT);
837 // Create a fast isel object.
839 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
840 return ARM::createFastISel(funcInfo);
843 /// getFunctionAlignment - Return the Log2 alignment of this function.
844 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
845 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
848 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
849 /// be used for loads / stores from the global.
850 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
851 return (Subtarget->isThumb1Only() ? 127 : 4095);
854 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
855 unsigned NumVals = N->getNumValues();
857 return Sched::RegPressure;
859 for (unsigned i = 0; i != NumVals; ++i) {
860 EVT VT = N->getValueType(i);
861 if (VT.isFloatingPoint() || VT.isVector())
862 return Sched::Latency;
865 if (!N->isMachineOpcode())
866 return Sched::RegPressure;
868 // Load are scheduled for latency even if there instruction itinerary
870 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
871 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
873 return Sched::Latency;
875 if (!Itins->isEmpty() && Itins->getStageLatency(TID.getSchedClass()) > 2)
876 return Sched::Latency;
877 return Sched::RegPressure;
881 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
882 MachineFunction &MF) const {
883 switch (RC->getID()) {
886 case ARM::tGPRRegClassID:
887 return RegInfo->hasFP(MF) ? 4 : 5;
888 case ARM::GPRRegClassID: {
889 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
890 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
892 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
893 case ARM::DPRRegClassID:
898 //===----------------------------------------------------------------------===//
900 //===----------------------------------------------------------------------===//
902 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
903 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
905 default: llvm_unreachable("Unknown condition code!");
906 case ISD::SETNE: return ARMCC::NE;
907 case ISD::SETEQ: return ARMCC::EQ;
908 case ISD::SETGT: return ARMCC::GT;
909 case ISD::SETGE: return ARMCC::GE;
910 case ISD::SETLT: return ARMCC::LT;
911 case ISD::SETLE: return ARMCC::LE;
912 case ISD::SETUGT: return ARMCC::HI;
913 case ISD::SETUGE: return ARMCC::HS;
914 case ISD::SETULT: return ARMCC::LO;
915 case ISD::SETULE: return ARMCC::LS;
919 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
920 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
921 ARMCC::CondCodes &CondCode2) {
922 CondCode2 = ARMCC::AL;
924 default: llvm_unreachable("Unknown FP condition!");
926 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
928 case ISD::SETOGT: CondCode = ARMCC::GT; break;
930 case ISD::SETOGE: CondCode = ARMCC::GE; break;
931 case ISD::SETOLT: CondCode = ARMCC::MI; break;
932 case ISD::SETOLE: CondCode = ARMCC::LS; break;
933 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
934 case ISD::SETO: CondCode = ARMCC::VC; break;
935 case ISD::SETUO: CondCode = ARMCC::VS; break;
936 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
937 case ISD::SETUGT: CondCode = ARMCC::HI; break;
938 case ISD::SETUGE: CondCode = ARMCC::PL; break;
940 case ISD::SETULT: CondCode = ARMCC::LT; break;
942 case ISD::SETULE: CondCode = ARMCC::LE; break;
944 case ISD::SETUNE: CondCode = ARMCC::NE; break;
948 //===----------------------------------------------------------------------===//
949 // Calling Convention Implementation
950 //===----------------------------------------------------------------------===//
952 #include "ARMGenCallingConv.inc"
954 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
955 /// given CallingConvention value.
956 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
958 bool isVarArg) const {
961 llvm_unreachable("Unsupported calling convention");
962 case CallingConv::Fast:
963 if (ARMFastCC && Subtarget->hasVFP2() && !isVarArg) {
964 if (!Subtarget->isAAPCS_ABI())
965 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
966 // For AAPCS ABI targets, just use VFP variant of the calling convention.
967 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
970 case CallingConv::C: {
971 // Use target triple & subtarget features to do actual dispatch.
972 if (!Subtarget->isAAPCS_ABI())
973 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
974 else if (Subtarget->hasVFP2() &&
975 FloatABIType == FloatABI::Hard && !isVarArg)
976 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
977 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
979 case CallingConv::ARM_AAPCS_VFP:
980 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
981 case CallingConv::ARM_AAPCS:
982 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
983 case CallingConv::ARM_APCS:
984 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
988 /// LowerCallResult - Lower the result values of a call into the
989 /// appropriate copies out of appropriate physical registers.
991 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
992 CallingConv::ID CallConv, bool isVarArg,
993 const SmallVectorImpl<ISD::InputArg> &Ins,
994 DebugLoc dl, SelectionDAG &DAG,
995 SmallVectorImpl<SDValue> &InVals) const {
997 // Assign locations to each value returned by this call.
998 SmallVector<CCValAssign, 16> RVLocs;
999 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1000 RVLocs, *DAG.getContext());
1001 CCInfo.AnalyzeCallResult(Ins,
1002 CCAssignFnForNode(CallConv, /* Return*/ true,
1005 // Copy all of the result registers out of their specified physreg.
1006 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1007 CCValAssign VA = RVLocs[i];
1010 if (VA.needsCustom()) {
1011 // Handle f64 or half of a v2f64.
1012 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1014 Chain = Lo.getValue(1);
1015 InFlag = Lo.getValue(2);
1016 VA = RVLocs[++i]; // skip ahead to next loc
1017 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1019 Chain = Hi.getValue(1);
1020 InFlag = Hi.getValue(2);
1021 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1023 if (VA.getLocVT() == MVT::v2f64) {
1024 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1025 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1026 DAG.getConstant(0, MVT::i32));
1028 VA = RVLocs[++i]; // skip ahead to next loc
1029 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1030 Chain = Lo.getValue(1);
1031 InFlag = Lo.getValue(2);
1032 VA = RVLocs[++i]; // skip ahead to next loc
1033 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1034 Chain = Hi.getValue(1);
1035 InFlag = Hi.getValue(2);
1036 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1037 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1038 DAG.getConstant(1, MVT::i32));
1041 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1043 Chain = Val.getValue(1);
1044 InFlag = Val.getValue(2);
1047 switch (VA.getLocInfo()) {
1048 default: llvm_unreachable("Unknown loc info!");
1049 case CCValAssign::Full: break;
1050 case CCValAssign::BCvt:
1051 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1055 InVals.push_back(Val);
1061 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1062 /// by "Src" to address "Dst" of size "Size". Alignment information is
1063 /// specified by the specific parameter attribute. The copy will be passed as
1064 /// a byval function parameter.
1065 /// Sometimes what we are copying is the end of a larger object, the part that
1066 /// does not fit in registers.
1068 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1069 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1071 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1072 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1073 /*isVolatile=*/false, /*AlwaysInline=*/false,
1074 MachinePointerInfo(0), MachinePointerInfo(0));
1077 /// LowerMemOpCallTo - Store the argument to the stack.
1079 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1080 SDValue StackPtr, SDValue Arg,
1081 DebugLoc dl, SelectionDAG &DAG,
1082 const CCValAssign &VA,
1083 ISD::ArgFlagsTy Flags) const {
1084 unsigned LocMemOffset = VA.getLocMemOffset();
1085 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1086 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1087 if (Flags.isByVal())
1088 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1090 return DAG.getStore(Chain, dl, Arg, PtrOff,
1091 MachinePointerInfo::getStack(LocMemOffset),
1095 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1096 SDValue Chain, SDValue &Arg,
1097 RegsToPassVector &RegsToPass,
1098 CCValAssign &VA, CCValAssign &NextVA,
1100 SmallVector<SDValue, 8> &MemOpChains,
1101 ISD::ArgFlagsTy Flags) const {
1103 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1104 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1105 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1107 if (NextVA.isRegLoc())
1108 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1110 assert(NextVA.isMemLoc());
1111 if (StackPtr.getNode() == 0)
1112 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1114 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1120 /// LowerCall - Lowering a call into a callseq_start <-
1121 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1124 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1125 CallingConv::ID CallConv, bool isVarArg,
1127 const SmallVectorImpl<ISD::OutputArg> &Outs,
1128 const SmallVectorImpl<SDValue> &OutVals,
1129 const SmallVectorImpl<ISD::InputArg> &Ins,
1130 DebugLoc dl, SelectionDAG &DAG,
1131 SmallVectorImpl<SDValue> &InVals) const {
1132 MachineFunction &MF = DAG.getMachineFunction();
1133 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1134 bool IsSibCall = false;
1135 // Temporarily disable tail calls so things don't break.
1136 if (!EnableARMTailCalls)
1139 // Check if it's really possible to do a tail call.
1140 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1141 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1142 Outs, OutVals, Ins, DAG);
1143 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1144 // detected sibcalls.
1151 // Analyze operands of the call, assigning locations to each operand.
1152 SmallVector<CCValAssign, 16> ArgLocs;
1153 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1155 CCInfo.AnalyzeCallOperands(Outs,
1156 CCAssignFnForNode(CallConv, /* Return*/ false,
1159 // Get a count of how many bytes are to be pushed on the stack.
1160 unsigned NumBytes = CCInfo.getNextStackOffset();
1162 // For tail calls, memory operands are available in our caller's stack.
1166 // Adjust the stack pointer for the new arguments...
1167 // These operations are automatically eliminated by the prolog/epilog pass
1169 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1171 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1173 RegsToPassVector RegsToPass;
1174 SmallVector<SDValue, 8> MemOpChains;
1176 // Walk the register/memloc assignments, inserting copies/loads. In the case
1177 // of tail call optimization, arguments are handled later.
1178 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1180 ++i, ++realArgIdx) {
1181 CCValAssign &VA = ArgLocs[i];
1182 SDValue Arg = OutVals[realArgIdx];
1183 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1185 // Promote the value if needed.
1186 switch (VA.getLocInfo()) {
1187 default: llvm_unreachable("Unknown loc info!");
1188 case CCValAssign::Full: break;
1189 case CCValAssign::SExt:
1190 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1192 case CCValAssign::ZExt:
1193 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1195 case CCValAssign::AExt:
1196 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1198 case CCValAssign::BCvt:
1199 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1203 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1204 if (VA.needsCustom()) {
1205 if (VA.getLocVT() == MVT::v2f64) {
1206 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1207 DAG.getConstant(0, MVT::i32));
1208 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1209 DAG.getConstant(1, MVT::i32));
1211 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1212 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1214 VA = ArgLocs[++i]; // skip ahead to next loc
1215 if (VA.isRegLoc()) {
1216 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1217 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1219 assert(VA.isMemLoc());
1221 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1222 dl, DAG, VA, Flags));
1225 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1226 StackPtr, MemOpChains, Flags);
1228 } else if (VA.isRegLoc()) {
1229 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1230 } else if (!IsSibCall) {
1231 assert(VA.isMemLoc());
1233 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1234 dl, DAG, VA, Flags));
1238 if (!MemOpChains.empty())
1239 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1240 &MemOpChains[0], MemOpChains.size());
1242 // Build a sequence of copy-to-reg nodes chained together with token chain
1243 // and flag operands which copy the outgoing args into the appropriate regs.
1245 // Tail call byval lowering might overwrite argument registers so in case of
1246 // tail call optimization the copies to registers are lowered later.
1248 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1249 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1250 RegsToPass[i].second, InFlag);
1251 InFlag = Chain.getValue(1);
1254 // For tail calls lower the arguments to the 'real' stack slot.
1256 // Force all the incoming stack arguments to be loaded from the stack
1257 // before any new outgoing arguments are stored to the stack, because the
1258 // outgoing stack slots may alias the incoming argument stack slots, and
1259 // the alias isn't otherwise explicit. This is slightly more conservative
1260 // than necessary, because it means that each store effectively depends
1261 // on every argument instead of just those arguments it would clobber.
1263 // Do not flag preceeding copytoreg stuff together with the following stuff.
1265 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1266 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1267 RegsToPass[i].second, InFlag);
1268 InFlag = Chain.getValue(1);
1273 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1274 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1275 // node so that legalize doesn't hack it.
1276 bool isDirect = false;
1277 bool isARMFunc = false;
1278 bool isLocalARMFunc = false;
1279 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1281 if (EnableARMLongCalls) {
1282 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1283 && "long-calls with non-static relocation model!");
1284 // Handle a global address or an external symbol. If it's not one of
1285 // those, the target's already in a register, so we don't need to do
1287 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1288 const GlobalValue *GV = G->getGlobal();
1289 // Create a constant pool entry for the callee address
1290 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1291 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1294 // Get the address of the callee into a register
1295 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1296 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1297 Callee = DAG.getLoad(getPointerTy(), dl,
1298 DAG.getEntryNode(), CPAddr,
1299 MachinePointerInfo::getConstantPool(),
1301 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1302 const char *Sym = S->getSymbol();
1304 // Create a constant pool entry for the callee address
1305 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1306 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1307 Sym, ARMPCLabelIndex, 0);
1308 // Get the address of the callee into a register
1309 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1310 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1311 Callee = DAG.getLoad(getPointerTy(), dl,
1312 DAG.getEntryNode(), CPAddr,
1313 MachinePointerInfo::getConstantPool(),
1316 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1317 const GlobalValue *GV = G->getGlobal();
1319 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1320 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1321 getTargetMachine().getRelocationModel() != Reloc::Static;
1322 isARMFunc = !Subtarget->isThumb() || isStub;
1323 // ARM call to a local ARM function is predicable.
1324 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1325 // tBX takes a register source operand.
1326 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1327 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1328 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1331 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1332 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1333 Callee = DAG.getLoad(getPointerTy(), dl,
1334 DAG.getEntryNode(), CPAddr,
1335 MachinePointerInfo::getConstantPool(),
1337 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1338 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1339 getPointerTy(), Callee, PICLabel);
1341 // On ELF targets for PIC code, direct calls should go through the PLT
1342 unsigned OpFlags = 0;
1343 if (Subtarget->isTargetELF() &&
1344 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1345 OpFlags = ARMII::MO_PLT;
1346 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1348 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1350 bool isStub = Subtarget->isTargetDarwin() &&
1351 getTargetMachine().getRelocationModel() != Reloc::Static;
1352 isARMFunc = !Subtarget->isThumb() || isStub;
1353 // tBX takes a register source operand.
1354 const char *Sym = S->getSymbol();
1355 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1356 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1357 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1358 Sym, ARMPCLabelIndex, 4);
1359 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1360 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1361 Callee = DAG.getLoad(getPointerTy(), dl,
1362 DAG.getEntryNode(), CPAddr,
1363 MachinePointerInfo::getConstantPool(),
1365 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1366 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1367 getPointerTy(), Callee, PICLabel);
1369 unsigned OpFlags = 0;
1370 // On ELF targets for PIC code, direct calls should go through the PLT
1371 if (Subtarget->isTargetELF() &&
1372 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1373 OpFlags = ARMII::MO_PLT;
1374 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1378 // FIXME: handle tail calls differently.
1380 if (Subtarget->isThumb()) {
1381 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1382 CallOpc = ARMISD::CALL_NOLINK;
1384 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1386 CallOpc = (isDirect || Subtarget->hasV5TOps())
1387 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1388 : ARMISD::CALL_NOLINK;
1391 std::vector<SDValue> Ops;
1392 Ops.push_back(Chain);
1393 Ops.push_back(Callee);
1395 // Add argument registers to the end of the list so that they are known live
1397 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1398 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1399 RegsToPass[i].second.getValueType()));
1401 if (InFlag.getNode())
1402 Ops.push_back(InFlag);
1404 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1406 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1408 // Returns a chain and a flag for retval copy to use.
1409 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1410 InFlag = Chain.getValue(1);
1412 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1413 DAG.getIntPtrConstant(0, true), InFlag);
1415 InFlag = Chain.getValue(1);
1417 // Handle result values, copying them out of physregs into vregs that we
1419 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1423 /// MatchingStackOffset - Return true if the given stack call argument is
1424 /// already available in the same position (relatively) of the caller's
1425 /// incoming argument stack.
1427 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1428 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1429 const ARMInstrInfo *TII) {
1430 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1432 if (Arg.getOpcode() == ISD::CopyFromReg) {
1433 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1434 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1436 MachineInstr *Def = MRI->getVRegDef(VR);
1439 if (!Flags.isByVal()) {
1440 if (!TII->isLoadFromStackSlot(Def, FI))
1445 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1446 if (Flags.isByVal())
1447 // ByVal argument is passed in as a pointer but it's now being
1448 // dereferenced. e.g.
1449 // define @foo(%struct.X* %A) {
1450 // tail call @bar(%struct.X* byval %A)
1453 SDValue Ptr = Ld->getBasePtr();
1454 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1457 FI = FINode->getIndex();
1461 assert(FI != INT_MAX);
1462 if (!MFI->isFixedObjectIndex(FI))
1464 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1467 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1468 /// for tail call optimization. Targets which want to do tail call
1469 /// optimization should implement this function.
1471 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1472 CallingConv::ID CalleeCC,
1474 bool isCalleeStructRet,
1475 bool isCallerStructRet,
1476 const SmallVectorImpl<ISD::OutputArg> &Outs,
1477 const SmallVectorImpl<SDValue> &OutVals,
1478 const SmallVectorImpl<ISD::InputArg> &Ins,
1479 SelectionDAG& DAG) const {
1480 const Function *CallerF = DAG.getMachineFunction().getFunction();
1481 CallingConv::ID CallerCC = CallerF->getCallingConv();
1482 bool CCMatch = CallerCC == CalleeCC;
1484 // Look for obvious safe cases to perform tail call optimization that do not
1485 // require ABI changes. This is what gcc calls sibcall.
1487 // Do not sibcall optimize vararg calls unless the call site is not passing
1489 if (isVarArg && !Outs.empty())
1492 // Also avoid sibcall optimization if either caller or callee uses struct
1493 // return semantics.
1494 if (isCalleeStructRet || isCallerStructRet)
1497 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1498 // emitEpilogue is not ready for them.
1499 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1500 // LR. This means if we need to reload LR, it takes an extra instructions,
1501 // which outweighs the value of the tail call; but here we don't know yet
1502 // whether LR is going to be used. Probably the right approach is to
1503 // generate the tail call here and turn it back into CALL/RET in
1504 // emitEpilogue if LR is used.
1505 if (Subtarget->isThumb1Only())
1508 // For the moment, we can only do this to functions defined in this
1509 // compilation, or to indirect calls. A Thumb B to an ARM function,
1510 // or vice versa, is not easily fixed up in the linker unlike BL.
1511 // (We could do this by loading the address of the callee into a register;
1512 // that is an extra instruction over the direct call and burns a register
1513 // as well, so is not likely to be a win.)
1515 // It might be safe to remove this restriction on non-Darwin.
1517 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1518 // but we need to make sure there are enough registers; the only valid
1519 // registers are the 4 used for parameters. We don't currently do this
1521 if (isa<ExternalSymbolSDNode>(Callee))
1524 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1525 const GlobalValue *GV = G->getGlobal();
1526 if (GV->isDeclaration() || GV->isWeakForLinker())
1530 // If the calling conventions do not match, then we'd better make sure the
1531 // results are returned in the same way as what the caller expects.
1533 SmallVector<CCValAssign, 16> RVLocs1;
1534 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1535 RVLocs1, *DAG.getContext());
1536 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1538 SmallVector<CCValAssign, 16> RVLocs2;
1539 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1540 RVLocs2, *DAG.getContext());
1541 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1543 if (RVLocs1.size() != RVLocs2.size())
1545 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1546 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1548 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1550 if (RVLocs1[i].isRegLoc()) {
1551 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1554 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1560 // If the callee takes no arguments then go on to check the results of the
1562 if (!Outs.empty()) {
1563 // Check if stack adjustment is needed. For now, do not do this if any
1564 // argument is passed on the stack.
1565 SmallVector<CCValAssign, 16> ArgLocs;
1566 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1567 ArgLocs, *DAG.getContext());
1568 CCInfo.AnalyzeCallOperands(Outs,
1569 CCAssignFnForNode(CalleeCC, false, isVarArg));
1570 if (CCInfo.getNextStackOffset()) {
1571 MachineFunction &MF = DAG.getMachineFunction();
1573 // Check if the arguments are already laid out in the right way as
1574 // the caller's fixed stack objects.
1575 MachineFrameInfo *MFI = MF.getFrameInfo();
1576 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1577 const ARMInstrInfo *TII =
1578 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1579 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1581 ++i, ++realArgIdx) {
1582 CCValAssign &VA = ArgLocs[i];
1583 EVT RegVT = VA.getLocVT();
1584 SDValue Arg = OutVals[realArgIdx];
1585 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1586 if (VA.getLocInfo() == CCValAssign::Indirect)
1588 if (VA.needsCustom()) {
1589 // f64 and vector types are split into multiple registers or
1590 // register/stack-slot combinations. The types will not match
1591 // the registers; give up on memory f64 refs until we figure
1592 // out what to do about this.
1595 if (!ArgLocs[++i].isRegLoc())
1597 if (RegVT == MVT::v2f64) {
1598 if (!ArgLocs[++i].isRegLoc())
1600 if (!ArgLocs[++i].isRegLoc())
1603 } else if (!VA.isRegLoc()) {
1604 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1616 ARMTargetLowering::LowerReturn(SDValue Chain,
1617 CallingConv::ID CallConv, bool isVarArg,
1618 const SmallVectorImpl<ISD::OutputArg> &Outs,
1619 const SmallVectorImpl<SDValue> &OutVals,
1620 DebugLoc dl, SelectionDAG &DAG) const {
1622 // CCValAssign - represent the assignment of the return value to a location.
1623 SmallVector<CCValAssign, 16> RVLocs;
1625 // CCState - Info about the registers and stack slots.
1626 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1629 // Analyze outgoing return values.
1630 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1633 // If this is the first return lowered for this function, add
1634 // the regs to the liveout set for the function.
1635 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1636 for (unsigned i = 0; i != RVLocs.size(); ++i)
1637 if (RVLocs[i].isRegLoc())
1638 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1643 // Copy the result values into the output registers.
1644 for (unsigned i = 0, realRVLocIdx = 0;
1646 ++i, ++realRVLocIdx) {
1647 CCValAssign &VA = RVLocs[i];
1648 assert(VA.isRegLoc() && "Can only return in registers!");
1650 SDValue Arg = OutVals[realRVLocIdx];
1652 switch (VA.getLocInfo()) {
1653 default: llvm_unreachable("Unknown loc info!");
1654 case CCValAssign::Full: break;
1655 case CCValAssign::BCvt:
1656 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1660 if (VA.needsCustom()) {
1661 if (VA.getLocVT() == MVT::v2f64) {
1662 // Extract the first half and return it in two registers.
1663 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1664 DAG.getConstant(0, MVT::i32));
1665 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1666 DAG.getVTList(MVT::i32, MVT::i32), Half);
1668 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1669 Flag = Chain.getValue(1);
1670 VA = RVLocs[++i]; // skip ahead to next loc
1671 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1672 HalfGPRs.getValue(1), Flag);
1673 Flag = Chain.getValue(1);
1674 VA = RVLocs[++i]; // skip ahead to next loc
1676 // Extract the 2nd half and fall through to handle it as an f64 value.
1677 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1678 DAG.getConstant(1, MVT::i32));
1680 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1682 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1683 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1684 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1685 Flag = Chain.getValue(1);
1686 VA = RVLocs[++i]; // skip ahead to next loc
1687 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1690 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1692 // Guarantee that all emitted copies are
1693 // stuck together, avoiding something bad.
1694 Flag = Chain.getValue(1);
1699 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1701 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1706 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1707 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1708 // one of the above mentioned nodes. It has to be wrapped because otherwise
1709 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1710 // be used to form addressing mode. These wrapped nodes will be selected
1712 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1713 EVT PtrVT = Op.getValueType();
1714 // FIXME there is no actual debug info here
1715 DebugLoc dl = Op.getDebugLoc();
1716 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1718 if (CP->isMachineConstantPoolEntry())
1719 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1720 CP->getAlignment());
1722 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1723 CP->getAlignment());
1724 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1727 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1728 return MachineJumpTableInfo::EK_Inline;
1731 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1732 SelectionDAG &DAG) const {
1733 MachineFunction &MF = DAG.getMachineFunction();
1734 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1735 unsigned ARMPCLabelIndex = 0;
1736 DebugLoc DL = Op.getDebugLoc();
1737 EVT PtrVT = getPointerTy();
1738 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1739 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1741 if (RelocM == Reloc::Static) {
1742 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1744 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1745 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1746 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1747 ARMCP::CPBlockAddress,
1749 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1751 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1752 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1753 MachinePointerInfo::getConstantPool(),
1755 if (RelocM == Reloc::Static)
1757 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1758 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1761 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1763 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1764 SelectionDAG &DAG) const {
1765 DebugLoc dl = GA->getDebugLoc();
1766 EVT PtrVT = getPointerTy();
1767 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1768 MachineFunction &MF = DAG.getMachineFunction();
1769 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1770 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1771 ARMConstantPoolValue *CPV =
1772 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1773 ARMCP::CPValue, PCAdj, "tlsgd", true);
1774 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1775 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1776 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1777 MachinePointerInfo::getConstantPool(),
1779 SDValue Chain = Argument.getValue(1);
1781 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1782 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1784 // call __tls_get_addr.
1787 Entry.Node = Argument;
1788 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1789 Args.push_back(Entry);
1790 // FIXME: is there useful debug info available here?
1791 std::pair<SDValue, SDValue> CallResult =
1792 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1793 false, false, false, false,
1794 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1795 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1796 return CallResult.first;
1799 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1800 // "local exec" model.
1802 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1803 SelectionDAG &DAG) const {
1804 const GlobalValue *GV = GA->getGlobal();
1805 DebugLoc dl = GA->getDebugLoc();
1807 SDValue Chain = DAG.getEntryNode();
1808 EVT PtrVT = getPointerTy();
1809 // Get the Thread Pointer
1810 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1812 if (GV->isDeclaration()) {
1813 MachineFunction &MF = DAG.getMachineFunction();
1814 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1815 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1816 // Initial exec model.
1817 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1818 ARMConstantPoolValue *CPV =
1819 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1820 ARMCP::CPValue, PCAdj, "gottpoff", true);
1821 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1822 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1823 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1824 MachinePointerInfo::getConstantPool(),
1826 Chain = Offset.getValue(1);
1828 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1829 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1831 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1832 MachinePointerInfo::getConstantPool(),
1836 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1837 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1838 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1839 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1840 MachinePointerInfo::getConstantPool(),
1844 // The address of the thread local variable is the add of the thread
1845 // pointer with the offset of the variable.
1846 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1850 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1851 // TODO: implement the "local dynamic" model
1852 assert(Subtarget->isTargetELF() &&
1853 "TLS not implemented for non-ELF targets");
1854 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1855 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1856 // otherwise use the "Local Exec" TLS Model
1857 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1858 return LowerToTLSGeneralDynamicModel(GA, DAG);
1860 return LowerToTLSExecModels(GA, DAG);
1863 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1864 SelectionDAG &DAG) const {
1865 EVT PtrVT = getPointerTy();
1866 DebugLoc dl = Op.getDebugLoc();
1867 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1868 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1869 if (RelocM == Reloc::PIC_) {
1870 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1871 ARMConstantPoolValue *CPV =
1872 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1873 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1874 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1875 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1877 MachinePointerInfo::getConstantPool(),
1879 SDValue Chain = Result.getValue(1);
1880 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1881 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1883 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1884 MachinePointerInfo::getGOT(), false, false, 0);
1887 // If we have T2 ops, we can materialize the address directly via movt/movw
1888 // pair. This is always cheaper.
1889 if (Subtarget->useMovt()) {
1890 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1891 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1893 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1894 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1895 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1896 MachinePointerInfo::getConstantPool(),
1902 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1903 SelectionDAG &DAG) const {
1904 MachineFunction &MF = DAG.getMachineFunction();
1905 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1906 unsigned ARMPCLabelIndex = 0;
1907 EVT PtrVT = getPointerTy();
1908 DebugLoc dl = Op.getDebugLoc();
1909 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1910 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1912 if (RelocM == Reloc::Static)
1913 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1915 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1916 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1917 ARMConstantPoolValue *CPV =
1918 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1919 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1921 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1923 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1924 MachinePointerInfo::getConstantPool(),
1926 SDValue Chain = Result.getValue(1);
1928 if (RelocM == Reloc::PIC_) {
1929 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1930 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1933 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1934 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1940 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1941 SelectionDAG &DAG) const {
1942 assert(Subtarget->isTargetELF() &&
1943 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1944 MachineFunction &MF = DAG.getMachineFunction();
1945 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1946 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1947 EVT PtrVT = getPointerTy();
1948 DebugLoc dl = Op.getDebugLoc();
1949 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1950 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1951 "_GLOBAL_OFFSET_TABLE_",
1952 ARMPCLabelIndex, PCAdj);
1953 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1954 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1955 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1956 MachinePointerInfo::getConstantPool(),
1958 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1959 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1963 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
1965 DebugLoc dl = Op.getDebugLoc();
1966 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
1967 Op.getOperand(0), Op.getOperand(1));
1971 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1972 DebugLoc dl = Op.getDebugLoc();
1973 SDValue Val = DAG.getConstant(0, MVT::i32);
1974 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1975 Op.getOperand(1), Val);
1979 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1980 DebugLoc dl = Op.getDebugLoc();
1981 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1982 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1986 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1987 const ARMSubtarget *Subtarget) const {
1988 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1989 DebugLoc dl = Op.getDebugLoc();
1991 default: return SDValue(); // Don't custom lower most intrinsics.
1992 case Intrinsic::arm_thread_pointer: {
1993 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1994 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1996 case Intrinsic::eh_sjlj_lsda: {
1997 MachineFunction &MF = DAG.getMachineFunction();
1998 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1999 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
2000 EVT PtrVT = getPointerTy();
2001 DebugLoc dl = Op.getDebugLoc();
2002 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2004 unsigned PCAdj = (RelocM != Reloc::PIC_)
2005 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2006 ARMConstantPoolValue *CPV =
2007 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2008 ARMCP::CPLSDA, PCAdj);
2009 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2010 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2012 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2013 MachinePointerInfo::getConstantPool(),
2016 if (RelocM == Reloc::PIC_) {
2017 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2018 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2025 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2026 const ARMSubtarget *Subtarget) {
2027 DebugLoc dl = Op.getDebugLoc();
2028 SDValue Op5 = Op.getOperand(5);
2029 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
2030 // Some subtargets which have dmb and dsb instructions can handle barriers
2031 // directly. Some ARMv6 cpus can support them with the help of mcr
2032 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
2034 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
2035 if (Subtarget->hasDataBarrier())
2036 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2038 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2039 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2040 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2041 DAG.getConstant(0, MVT::i32));
2045 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2046 MachineFunction &MF = DAG.getMachineFunction();
2047 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2049 // vastart just stores the address of the VarArgsFrameIndex slot into the
2050 // memory location argument.
2051 DebugLoc dl = Op.getDebugLoc();
2052 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2053 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2054 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2055 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2056 MachinePointerInfo(SV), false, false, 0);
2060 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2061 SDValue &Root, SelectionDAG &DAG,
2062 DebugLoc dl) const {
2063 MachineFunction &MF = DAG.getMachineFunction();
2064 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2066 TargetRegisterClass *RC;
2067 if (AFI->isThumb1OnlyFunction())
2068 RC = ARM::tGPRRegisterClass;
2070 RC = ARM::GPRRegisterClass;
2072 // Transform the arguments stored in physical registers into virtual ones.
2073 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2074 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2077 if (NextVA.isMemLoc()) {
2078 MachineFrameInfo *MFI = MF.getFrameInfo();
2079 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2081 // Create load node to retrieve arguments from the stack.
2082 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2083 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2084 MachinePointerInfo::getFixedStack(FI),
2087 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2088 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2091 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2095 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2096 CallingConv::ID CallConv, bool isVarArg,
2097 const SmallVectorImpl<ISD::InputArg>
2099 DebugLoc dl, SelectionDAG &DAG,
2100 SmallVectorImpl<SDValue> &InVals)
2103 MachineFunction &MF = DAG.getMachineFunction();
2104 MachineFrameInfo *MFI = MF.getFrameInfo();
2106 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2108 // Assign locations to all of the incoming arguments.
2109 SmallVector<CCValAssign, 16> ArgLocs;
2110 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2112 CCInfo.AnalyzeFormalArguments(Ins,
2113 CCAssignFnForNode(CallConv, /* Return*/ false,
2116 SmallVector<SDValue, 16> ArgValues;
2118 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2119 CCValAssign &VA = ArgLocs[i];
2121 // Arguments stored in registers.
2122 if (VA.isRegLoc()) {
2123 EVT RegVT = VA.getLocVT();
2126 if (VA.needsCustom()) {
2127 // f64 and vector types are split up into multiple registers or
2128 // combinations of registers and stack slots.
2129 if (VA.getLocVT() == MVT::v2f64) {
2130 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2132 VA = ArgLocs[++i]; // skip ahead to next loc
2134 if (VA.isMemLoc()) {
2135 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2136 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2137 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2138 MachinePointerInfo::getFixedStack(FI),
2141 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2144 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2145 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2146 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2147 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2148 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2150 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2153 TargetRegisterClass *RC;
2155 if (RegVT == MVT::f32)
2156 RC = ARM::SPRRegisterClass;
2157 else if (RegVT == MVT::f64)
2158 RC = ARM::DPRRegisterClass;
2159 else if (RegVT == MVT::v2f64)
2160 RC = ARM::QPRRegisterClass;
2161 else if (RegVT == MVT::i32)
2162 RC = (AFI->isThumb1OnlyFunction() ?
2163 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2165 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2167 // Transform the arguments in physical registers into virtual ones.
2168 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2169 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2172 // If this is an 8 or 16-bit value, it is really passed promoted
2173 // to 32 bits. Insert an assert[sz]ext to capture this, then
2174 // truncate to the right size.
2175 switch (VA.getLocInfo()) {
2176 default: llvm_unreachable("Unknown loc info!");
2177 case CCValAssign::Full: break;
2178 case CCValAssign::BCvt:
2179 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2181 case CCValAssign::SExt:
2182 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2183 DAG.getValueType(VA.getValVT()));
2184 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2186 case CCValAssign::ZExt:
2187 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2188 DAG.getValueType(VA.getValVT()));
2189 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2193 InVals.push_back(ArgValue);
2195 } else { // VA.isRegLoc()
2198 assert(VA.isMemLoc());
2199 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2201 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2202 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2204 // Create load nodes to retrieve arguments from the stack.
2205 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2206 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2207 MachinePointerInfo::getFixedStack(FI),
2214 static const unsigned GPRArgRegs[] = {
2215 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2218 unsigned NumGPRs = CCInfo.getFirstUnallocated
2219 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2221 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2222 unsigned VARegSize = (4 - NumGPRs) * 4;
2223 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2224 unsigned ArgOffset = CCInfo.getNextStackOffset();
2225 if (VARegSaveSize) {
2226 // If this function is vararg, store any remaining integer argument regs
2227 // to their spots on the stack so that they may be loaded by deferencing
2228 // the result of va_next.
2229 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2230 AFI->setVarArgsFrameIndex(
2231 MFI->CreateFixedObject(VARegSaveSize,
2232 ArgOffset + VARegSaveSize - VARegSize,
2234 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2237 SmallVector<SDValue, 4> MemOps;
2238 for (; NumGPRs < 4; ++NumGPRs) {
2239 TargetRegisterClass *RC;
2240 if (AFI->isThumb1OnlyFunction())
2241 RC = ARM::tGPRRegisterClass;
2243 RC = ARM::GPRRegisterClass;
2245 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2246 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2248 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2249 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2251 MemOps.push_back(Store);
2252 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2253 DAG.getConstant(4, getPointerTy()));
2255 if (!MemOps.empty())
2256 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2257 &MemOps[0], MemOps.size());
2259 // This will point to the next argument passed via stack.
2260 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2266 /// isFloatingPointZero - Return true if this is +0.0.
2267 static bool isFloatingPointZero(SDValue Op) {
2268 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2269 return CFP->getValueAPF().isPosZero();
2270 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2271 // Maybe this has already been legalized into the constant pool?
2272 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2273 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2274 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2275 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2276 return CFP->getValueAPF().isPosZero();
2282 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2283 /// the given operands.
2285 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2286 SDValue &ARMcc, SelectionDAG &DAG,
2287 DebugLoc dl) const {
2288 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2289 unsigned C = RHSC->getZExtValue();
2290 if (!isLegalICmpImmediate(C)) {
2291 // Constant does not fit, try adjusting it by one?
2296 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2297 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2298 RHS = DAG.getConstant(C-1, MVT::i32);
2303 if (C != 0 && isLegalICmpImmediate(C-1)) {
2304 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2305 RHS = DAG.getConstant(C-1, MVT::i32);
2310 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2311 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2312 RHS = DAG.getConstant(C+1, MVT::i32);
2317 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2318 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2319 RHS = DAG.getConstant(C+1, MVT::i32);
2326 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2327 ARMISD::NodeType CompareType;
2330 CompareType = ARMISD::CMP;
2335 CompareType = ARMISD::CMPZ;
2338 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2339 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2342 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2344 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2345 DebugLoc dl) const {
2347 if (!isFloatingPointZero(RHS))
2348 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2350 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2351 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2354 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2355 SDValue Cond = Op.getOperand(0);
2356 SDValue SelectTrue = Op.getOperand(1);
2357 SDValue SelectFalse = Op.getOperand(2);
2358 DebugLoc dl = Op.getDebugLoc();
2362 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2363 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2365 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2366 const ConstantSDNode *CMOVTrue =
2367 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2368 const ConstantSDNode *CMOVFalse =
2369 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2371 if (CMOVTrue && CMOVFalse) {
2372 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2373 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2377 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2379 False = SelectFalse;
2380 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2385 if (True.getNode() && False.getNode()) {
2386 EVT VT = Cond.getValueType();
2387 SDValue ARMcc = Cond.getOperand(2);
2388 SDValue CCR = Cond.getOperand(3);
2389 SDValue Cmp = Cond.getOperand(4);
2390 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2395 return DAG.getSelectCC(dl, Cond,
2396 DAG.getConstant(0, Cond.getValueType()),
2397 SelectTrue, SelectFalse, ISD::SETNE);
2400 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2401 EVT VT = Op.getValueType();
2402 SDValue LHS = Op.getOperand(0);
2403 SDValue RHS = Op.getOperand(1);
2404 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2405 SDValue TrueVal = Op.getOperand(2);
2406 SDValue FalseVal = Op.getOperand(3);
2407 DebugLoc dl = Op.getDebugLoc();
2409 if (LHS.getValueType() == MVT::i32) {
2411 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2412 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2413 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2416 ARMCC::CondCodes CondCode, CondCode2;
2417 FPCCToARMCC(CC, CondCode, CondCode2);
2419 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2420 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2421 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2422 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2424 if (CondCode2 != ARMCC::AL) {
2425 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2426 // FIXME: Needs another CMP because flag can have but one use.
2427 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2428 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2429 Result, TrueVal, ARMcc2, CCR, Cmp2);
2434 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2435 /// to morph to an integer compare sequence.
2436 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2437 const ARMSubtarget *Subtarget) {
2438 SDNode *N = Op.getNode();
2439 if (!N->hasOneUse())
2440 // Otherwise it requires moving the value from fp to integer registers.
2442 if (!N->getNumValues())
2444 EVT VT = Op.getValueType();
2445 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2446 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2447 // vmrs are very slow, e.g. cortex-a8.
2450 if (isFloatingPointZero(Op)) {
2454 return ISD::isNormalLoad(N);
2457 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2458 if (isFloatingPointZero(Op))
2459 return DAG.getConstant(0, MVT::i32);
2461 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2462 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2463 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2464 Ld->isVolatile(), Ld->isNonTemporal(),
2465 Ld->getAlignment());
2467 llvm_unreachable("Unknown VFP cmp argument!");
2470 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2471 SDValue &RetVal1, SDValue &RetVal2) {
2472 if (isFloatingPointZero(Op)) {
2473 RetVal1 = DAG.getConstant(0, MVT::i32);
2474 RetVal2 = DAG.getConstant(0, MVT::i32);
2478 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2479 SDValue Ptr = Ld->getBasePtr();
2480 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2481 Ld->getChain(), Ptr,
2482 Ld->getPointerInfo(),
2483 Ld->isVolatile(), Ld->isNonTemporal(),
2484 Ld->getAlignment());
2486 EVT PtrType = Ptr.getValueType();
2487 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2488 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2489 PtrType, Ptr, DAG.getConstant(4, PtrType));
2490 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2491 Ld->getChain(), NewPtr,
2492 Ld->getPointerInfo().getWithOffset(4),
2493 Ld->isVolatile(), Ld->isNonTemporal(),
2498 llvm_unreachable("Unknown VFP cmp argument!");
2501 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2502 /// f32 and even f64 comparisons to integer ones.
2504 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2505 SDValue Chain = Op.getOperand(0);
2506 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2507 SDValue LHS = Op.getOperand(2);
2508 SDValue RHS = Op.getOperand(3);
2509 SDValue Dest = Op.getOperand(4);
2510 DebugLoc dl = Op.getDebugLoc();
2512 bool SeenZero = false;
2513 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2514 canChangeToInt(RHS, SeenZero, Subtarget) &&
2515 // If one of the operand is zero, it's safe to ignore the NaN case since
2516 // we only care about equality comparisons.
2517 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2518 // If unsafe fp math optimization is enabled and there are no othter uses of
2519 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2520 // to an integer comparison.
2521 if (CC == ISD::SETOEQ)
2523 else if (CC == ISD::SETUNE)
2527 if (LHS.getValueType() == MVT::f32) {
2528 LHS = bitcastf32Toi32(LHS, DAG);
2529 RHS = bitcastf32Toi32(RHS, DAG);
2530 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2531 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2532 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2533 Chain, Dest, ARMcc, CCR, Cmp);
2538 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2539 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2540 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2541 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2542 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2543 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2544 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2550 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2551 SDValue Chain = Op.getOperand(0);
2552 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2553 SDValue LHS = Op.getOperand(2);
2554 SDValue RHS = Op.getOperand(3);
2555 SDValue Dest = Op.getOperand(4);
2556 DebugLoc dl = Op.getDebugLoc();
2558 if (LHS.getValueType() == MVT::i32) {
2560 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2561 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2562 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2563 Chain, Dest, ARMcc, CCR, Cmp);
2566 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2569 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2570 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2571 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2572 if (Result.getNode())
2576 ARMCC::CondCodes CondCode, CondCode2;
2577 FPCCToARMCC(CC, CondCode, CondCode2);
2579 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2580 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2581 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2582 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2583 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2584 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2585 if (CondCode2 != ARMCC::AL) {
2586 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2587 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2588 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2593 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2594 SDValue Chain = Op.getOperand(0);
2595 SDValue Table = Op.getOperand(1);
2596 SDValue Index = Op.getOperand(2);
2597 DebugLoc dl = Op.getDebugLoc();
2599 EVT PTy = getPointerTy();
2600 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2601 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2602 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2603 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2604 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2605 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2606 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2607 if (Subtarget->isThumb2()) {
2608 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2609 // which does another jump to the destination. This also makes it easier
2610 // to translate it to TBB / TBH later.
2611 // FIXME: This might not work if the function is extremely large.
2612 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2613 Addr, Op.getOperand(2), JTI, UId);
2615 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2616 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2617 MachinePointerInfo::getJumpTable(),
2619 Chain = Addr.getValue(1);
2620 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2621 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2623 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2624 MachinePointerInfo::getJumpTable(), false, false, 0);
2625 Chain = Addr.getValue(1);
2626 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2630 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2631 DebugLoc dl = Op.getDebugLoc();
2634 switch (Op.getOpcode()) {
2636 assert(0 && "Invalid opcode!");
2637 case ISD::FP_TO_SINT:
2638 Opc = ARMISD::FTOSI;
2640 case ISD::FP_TO_UINT:
2641 Opc = ARMISD::FTOUI;
2644 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2645 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2648 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2649 EVT VT = Op.getValueType();
2650 DebugLoc dl = Op.getDebugLoc();
2653 switch (Op.getOpcode()) {
2655 assert(0 && "Invalid opcode!");
2656 case ISD::SINT_TO_FP:
2657 Opc = ARMISD::SITOF;
2659 case ISD::UINT_TO_FP:
2660 Opc = ARMISD::UITOF;
2664 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2665 return DAG.getNode(Opc, dl, VT, Op);
2668 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2669 // Implement fcopysign with a fabs and a conditional fneg.
2670 SDValue Tmp0 = Op.getOperand(0);
2671 SDValue Tmp1 = Op.getOperand(1);
2672 DebugLoc dl = Op.getDebugLoc();
2673 EVT VT = Op.getValueType();
2674 EVT SrcVT = Tmp1.getValueType();
2675 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2676 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2677 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2678 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2679 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2680 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2683 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2684 MachineFunction &MF = DAG.getMachineFunction();
2685 MachineFrameInfo *MFI = MF.getFrameInfo();
2686 MFI->setReturnAddressIsTaken(true);
2688 EVT VT = Op.getValueType();
2689 DebugLoc dl = Op.getDebugLoc();
2690 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2692 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2693 SDValue Offset = DAG.getConstant(4, MVT::i32);
2694 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2695 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2696 MachinePointerInfo(), false, false, 0);
2699 // Return LR, which contains the return address. Mark it an implicit live-in.
2700 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2701 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2704 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2705 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2706 MFI->setFrameAddressIsTaken(true);
2708 EVT VT = Op.getValueType();
2709 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2710 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2711 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2712 ? ARM::R7 : ARM::R11;
2713 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2715 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2716 MachinePointerInfo(),
2721 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2722 /// expand a bit convert where either the source or destination type is i64 to
2723 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2724 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2725 /// vectors), since the legalizer won't know what to do with that.
2726 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2727 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2728 DebugLoc dl = N->getDebugLoc();
2729 SDValue Op = N->getOperand(0);
2731 // This function is only supposed to be called for i64 types, either as the
2732 // source or destination of the bit convert.
2733 EVT SrcVT = Op.getValueType();
2734 EVT DstVT = N->getValueType(0);
2735 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2736 "ExpandBIT_CONVERT called for non-i64 type");
2738 // Turn i64->f64 into VMOVDRR.
2739 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2740 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2741 DAG.getConstant(0, MVT::i32));
2742 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2743 DAG.getConstant(1, MVT::i32));
2744 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2745 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2748 // Turn f64->i64 into VMOVRRD.
2749 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2750 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2751 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2752 // Merge the pieces into a single i64 value.
2753 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2759 /// getZeroVector - Returns a vector of specified type with all zero elements.
2760 /// Zero vectors are used to represent vector negation and in those cases
2761 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2762 /// not support i64 elements, so sometimes the zero vectors will need to be
2763 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2765 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2766 assert(VT.isVector() && "Expected a vector type");
2767 // The canonical modified immediate encoding of a zero vector is....0!
2768 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2769 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2770 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2771 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2774 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2775 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2776 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2777 SelectionDAG &DAG) const {
2778 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2779 EVT VT = Op.getValueType();
2780 unsigned VTBits = VT.getSizeInBits();
2781 DebugLoc dl = Op.getDebugLoc();
2782 SDValue ShOpLo = Op.getOperand(0);
2783 SDValue ShOpHi = Op.getOperand(1);
2784 SDValue ShAmt = Op.getOperand(2);
2786 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2788 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2790 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2791 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2792 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2793 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2794 DAG.getConstant(VTBits, MVT::i32));
2795 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2796 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2797 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2799 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2800 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2802 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2803 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2806 SDValue Ops[2] = { Lo, Hi };
2807 return DAG.getMergeValues(Ops, 2, dl);
2810 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2811 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2812 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2813 SelectionDAG &DAG) const {
2814 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2815 EVT VT = Op.getValueType();
2816 unsigned VTBits = VT.getSizeInBits();
2817 DebugLoc dl = Op.getDebugLoc();
2818 SDValue ShOpLo = Op.getOperand(0);
2819 SDValue ShOpHi = Op.getOperand(1);
2820 SDValue ShAmt = Op.getOperand(2);
2823 assert(Op.getOpcode() == ISD::SHL_PARTS);
2824 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2825 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2826 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2827 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2828 DAG.getConstant(VTBits, MVT::i32));
2829 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2830 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2832 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2833 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2834 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2836 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2837 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2840 SDValue Ops[2] = { Lo, Hi };
2841 return DAG.getMergeValues(Ops, 2, dl);
2844 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2845 SelectionDAG &DAG) const {
2846 // The rounding mode is in bits 23:22 of the FPSCR.
2847 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2848 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2849 // so that the shift + and get folded into a bitfield extract.
2850 DebugLoc dl = Op.getDebugLoc();
2851 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2852 DAG.getConstant(Intrinsic::arm_get_fpscr,
2854 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2855 DAG.getConstant(1U << 22, MVT::i32));
2856 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2857 DAG.getConstant(22, MVT::i32));
2858 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2859 DAG.getConstant(3, MVT::i32));
2862 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2863 const ARMSubtarget *ST) {
2864 EVT VT = N->getValueType(0);
2865 DebugLoc dl = N->getDebugLoc();
2867 if (!ST->hasV6T2Ops())
2870 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2871 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2874 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2875 const ARMSubtarget *ST) {
2876 EVT VT = N->getValueType(0);
2877 DebugLoc dl = N->getDebugLoc();
2879 // Lower vector shifts on NEON to use VSHL.
2880 if (VT.isVector()) {
2881 assert(ST->hasNEON() && "unexpected vector shift");
2883 // Left shifts translate directly to the vshiftu intrinsic.
2884 if (N->getOpcode() == ISD::SHL)
2885 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2886 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2887 N->getOperand(0), N->getOperand(1));
2889 assert((N->getOpcode() == ISD::SRA ||
2890 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2892 // NEON uses the same intrinsics for both left and right shifts. For
2893 // right shifts, the shift amounts are negative, so negate the vector of
2895 EVT ShiftVT = N->getOperand(1).getValueType();
2896 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2897 getZeroVector(ShiftVT, DAG, dl),
2899 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2900 Intrinsic::arm_neon_vshifts :
2901 Intrinsic::arm_neon_vshiftu);
2902 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2903 DAG.getConstant(vshiftInt, MVT::i32),
2904 N->getOperand(0), NegatedCount);
2907 // We can get here for a node like i32 = ISD::SHL i32, i64
2911 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2912 "Unknown shift to lower!");
2914 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2915 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2916 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2919 // If we are in thumb mode, we don't have RRX.
2920 if (ST->isThumb1Only()) return SDValue();
2922 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2923 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2924 DAG.getConstant(0, MVT::i32));
2925 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2926 DAG.getConstant(1, MVT::i32));
2928 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2929 // captures the result into a carry flag.
2930 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2931 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2933 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2934 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2936 // Merge the pieces into a single i64 value.
2937 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2940 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2941 SDValue TmpOp0, TmpOp1;
2942 bool Invert = false;
2946 SDValue Op0 = Op.getOperand(0);
2947 SDValue Op1 = Op.getOperand(1);
2948 SDValue CC = Op.getOperand(2);
2949 EVT VT = Op.getValueType();
2950 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2951 DebugLoc dl = Op.getDebugLoc();
2953 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2954 switch (SetCCOpcode) {
2955 default: llvm_unreachable("Illegal FP comparison"); break;
2957 case ISD::SETNE: Invert = true; // Fallthrough
2959 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2961 case ISD::SETLT: Swap = true; // Fallthrough
2963 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2965 case ISD::SETLE: Swap = true; // Fallthrough
2967 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2968 case ISD::SETUGE: Swap = true; // Fallthrough
2969 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2970 case ISD::SETUGT: Swap = true; // Fallthrough
2971 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2972 case ISD::SETUEQ: Invert = true; // Fallthrough
2974 // Expand this to (OLT | OGT).
2978 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2979 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2981 case ISD::SETUO: Invert = true; // Fallthrough
2983 // Expand this to (OLT | OGE).
2987 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2988 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2992 // Integer comparisons.
2993 switch (SetCCOpcode) {
2994 default: llvm_unreachable("Illegal integer comparison"); break;
2995 case ISD::SETNE: Invert = true;
2996 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2997 case ISD::SETLT: Swap = true;
2998 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2999 case ISD::SETLE: Swap = true;
3000 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3001 case ISD::SETULT: Swap = true;
3002 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3003 case ISD::SETULE: Swap = true;
3004 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3007 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3008 if (Opc == ARMISD::VCEQ) {
3011 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3013 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3016 // Ignore bitconvert.
3017 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
3018 AndOp = AndOp.getOperand(0);
3020 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3022 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
3023 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
3030 std::swap(Op0, Op1);
3032 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3035 Result = DAG.getNOT(dl, Result, VT);
3040 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3041 /// valid vector constant for a NEON instruction with a "modified immediate"
3042 /// operand (e.g., VMOV). If so, return the encoded value.
3043 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3044 unsigned SplatBitSize, SelectionDAG &DAG,
3045 EVT &VT, bool is128Bits, bool isVMOV) {
3046 unsigned OpCmode, Imm;
3048 // SplatBitSize is set to the smallest size that splats the vector, so a
3049 // zero vector will always have SplatBitSize == 8. However, NEON modified
3050 // immediate instructions others than VMOV do not support the 8-bit encoding
3051 // of a zero vector, and the default encoding of zero is supposed to be the
3056 switch (SplatBitSize) {
3060 // Any 1-byte value is OK. Op=0, Cmode=1110.
3061 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3064 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3068 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3069 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3070 if ((SplatBits & ~0xff) == 0) {
3071 // Value = 0x00nn: Op=x, Cmode=100x.
3076 if ((SplatBits & ~0xff00) == 0) {
3077 // Value = 0xnn00: Op=x, Cmode=101x.
3079 Imm = SplatBits >> 8;
3085 // NEON's 32-bit VMOV supports splat values where:
3086 // * only one byte is nonzero, or
3087 // * the least significant byte is 0xff and the second byte is nonzero, or
3088 // * the least significant 2 bytes are 0xff and the third is nonzero.
3089 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3090 if ((SplatBits & ~0xff) == 0) {
3091 // Value = 0x000000nn: Op=x, Cmode=000x.
3096 if ((SplatBits & ~0xff00) == 0) {
3097 // Value = 0x0000nn00: Op=x, Cmode=001x.
3099 Imm = SplatBits >> 8;
3102 if ((SplatBits & ~0xff0000) == 0) {
3103 // Value = 0x00nn0000: Op=x, Cmode=010x.
3105 Imm = SplatBits >> 16;
3108 if ((SplatBits & ~0xff000000) == 0) {
3109 // Value = 0xnn000000: Op=x, Cmode=011x.
3111 Imm = SplatBits >> 24;
3115 if ((SplatBits & ~0xffff) == 0 &&
3116 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3117 // Value = 0x0000nnff: Op=x, Cmode=1100.
3119 Imm = SplatBits >> 8;
3124 if ((SplatBits & ~0xffffff) == 0 &&
3125 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3126 // Value = 0x00nnffff: Op=x, Cmode=1101.
3128 Imm = SplatBits >> 16;
3129 SplatBits |= 0xffff;
3133 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3134 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3135 // VMOV.I32. A (very) minor optimization would be to replicate the value
3136 // and fall through here to test for a valid 64-bit splat. But, then the
3137 // caller would also need to check and handle the change in size.
3143 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3144 uint64_t BitMask = 0xff;
3146 unsigned ImmMask = 1;
3148 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3149 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3152 } else if ((SplatBits & BitMask) != 0) {
3158 // Op=1, Cmode=1110.
3161 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3166 llvm_unreachable("unexpected size for isNEONModifiedImm");
3170 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3171 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3174 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3175 bool &ReverseVEXT, unsigned &Imm) {
3176 unsigned NumElts = VT.getVectorNumElements();
3177 ReverseVEXT = false;
3179 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3185 // If this is a VEXT shuffle, the immediate value is the index of the first
3186 // element. The other shuffle indices must be the successive elements after
3188 unsigned ExpectedElt = Imm;
3189 for (unsigned i = 1; i < NumElts; ++i) {
3190 // Increment the expected index. If it wraps around, it may still be
3191 // a VEXT but the source vectors must be swapped.
3193 if (ExpectedElt == NumElts * 2) {
3198 if (M[i] < 0) continue; // ignore UNDEF indices
3199 if (ExpectedElt != static_cast<unsigned>(M[i]))
3203 // Adjust the index value if the source operands will be swapped.
3210 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3211 /// instruction with the specified blocksize. (The order of the elements
3212 /// within each block of the vector is reversed.)
3213 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3214 unsigned BlockSize) {
3215 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3216 "Only possible block sizes for VREV are: 16, 32, 64");
3218 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3222 unsigned NumElts = VT.getVectorNumElements();
3223 unsigned BlockElts = M[0] + 1;
3224 // If the first shuffle index is UNDEF, be optimistic.
3226 BlockElts = BlockSize / EltSz;
3228 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3231 for (unsigned i = 0; i < NumElts; ++i) {
3232 if (M[i] < 0) continue; // ignore UNDEF indices
3233 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3240 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3241 unsigned &WhichResult) {
3242 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3246 unsigned NumElts = VT.getVectorNumElements();
3247 WhichResult = (M[0] == 0 ? 0 : 1);
3248 for (unsigned i = 0; i < NumElts; i += 2) {
3249 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3250 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3256 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3257 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3258 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3259 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3260 unsigned &WhichResult) {
3261 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3265 unsigned NumElts = VT.getVectorNumElements();
3266 WhichResult = (M[0] == 0 ? 0 : 1);
3267 for (unsigned i = 0; i < NumElts; i += 2) {
3268 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3269 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3275 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3276 unsigned &WhichResult) {
3277 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3281 unsigned NumElts = VT.getVectorNumElements();
3282 WhichResult = (M[0] == 0 ? 0 : 1);
3283 for (unsigned i = 0; i != NumElts; ++i) {
3284 if (M[i] < 0) continue; // ignore UNDEF indices
3285 if ((unsigned) M[i] != 2 * i + WhichResult)
3289 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3290 if (VT.is64BitVector() && EltSz == 32)
3296 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3297 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3298 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3299 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3300 unsigned &WhichResult) {
3301 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3305 unsigned Half = VT.getVectorNumElements() / 2;
3306 WhichResult = (M[0] == 0 ? 0 : 1);
3307 for (unsigned j = 0; j != 2; ++j) {
3308 unsigned Idx = WhichResult;
3309 for (unsigned i = 0; i != Half; ++i) {
3310 int MIdx = M[i + j * Half];
3311 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3317 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3318 if (VT.is64BitVector() && EltSz == 32)
3324 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3325 unsigned &WhichResult) {
3326 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3330 unsigned NumElts = VT.getVectorNumElements();
3331 WhichResult = (M[0] == 0 ? 0 : 1);
3332 unsigned Idx = WhichResult * NumElts / 2;
3333 for (unsigned i = 0; i != NumElts; i += 2) {
3334 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3335 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3340 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3341 if (VT.is64BitVector() && EltSz == 32)
3347 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3348 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3349 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3350 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3351 unsigned &WhichResult) {
3352 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3356 unsigned NumElts = VT.getVectorNumElements();
3357 WhichResult = (M[0] == 0 ? 0 : 1);
3358 unsigned Idx = WhichResult * NumElts / 2;
3359 for (unsigned i = 0; i != NumElts; i += 2) {
3360 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3361 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3366 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3367 if (VT.is64BitVector() && EltSz == 32)
3373 // If N is an integer constant that can be moved into a register in one
3374 // instruction, return an SDValue of such a constant (will become a MOV
3375 // instruction). Otherwise return null.
3376 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3377 const ARMSubtarget *ST, DebugLoc dl) {
3379 if (!isa<ConstantSDNode>(N))
3381 Val = cast<ConstantSDNode>(N)->getZExtValue();
3383 if (ST->isThumb1Only()) {
3384 if (Val <= 255 || ~Val <= 255)
3385 return DAG.getConstant(Val, MVT::i32);
3387 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3388 return DAG.getConstant(Val, MVT::i32);
3393 // If this is a case we can't handle, return null and let the default
3394 // expansion code take care of it.
3395 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3396 const ARMSubtarget *ST) {
3397 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3398 DebugLoc dl = Op.getDebugLoc();
3399 EVT VT = Op.getValueType();
3401 APInt SplatBits, SplatUndef;
3402 unsigned SplatBitSize;
3404 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3405 if (SplatBitSize <= 64) {
3406 // Check if an immediate VMOV works.
3408 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3409 SplatUndef.getZExtValue(), SplatBitSize,
3410 DAG, VmovVT, VT.is128BitVector(), true);
3411 if (Val.getNode()) {
3412 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3413 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3416 // Try an immediate VMVN.
3417 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3418 ((1LL << SplatBitSize) - 1));
3419 Val = isNEONModifiedImm(NegatedImm,
3420 SplatUndef.getZExtValue(), SplatBitSize,
3421 DAG, VmovVT, VT.is128BitVector(), false);
3422 if (Val.getNode()) {
3423 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3424 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3429 // Scan through the operands to see if only one value is used.
3430 unsigned NumElts = VT.getVectorNumElements();
3431 bool isOnlyLowElement = true;
3432 bool usesOnlyOneValue = true;
3433 bool isConstant = true;
3435 for (unsigned i = 0; i < NumElts; ++i) {
3436 SDValue V = Op.getOperand(i);
3437 if (V.getOpcode() == ISD::UNDEF)
3440 isOnlyLowElement = false;
3441 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3444 if (!Value.getNode())
3446 else if (V != Value)
3447 usesOnlyOneValue = false;
3450 if (!Value.getNode())
3451 return DAG.getUNDEF(VT);
3453 if (isOnlyLowElement)
3454 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3456 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3458 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3459 // i32 and try again.
3460 if (usesOnlyOneValue && EltSize <= 32) {
3462 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3463 if (VT.getVectorElementType().isFloatingPoint()) {
3464 SmallVector<SDValue, 8> Ops;
3465 for (unsigned i = 0; i < NumElts; ++i)
3466 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3468 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3470 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3472 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3474 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3476 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3479 // If all elements are constants and the case above didn't get hit, fall back
3480 // to the default expansion, which will generate a load from the constant
3485 // Vectors with 32- or 64-bit elements can be built by directly assigning
3486 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3487 // will be legalized.
3488 if (EltSize >= 32) {
3489 // Do the expansion with floating-point types, since that is what the VFP
3490 // registers are defined to use, and since i64 is not legal.
3491 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3492 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3493 SmallVector<SDValue, 8> Ops;
3494 for (unsigned i = 0; i < NumElts; ++i)
3495 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3496 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3497 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3503 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3504 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3505 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3506 /// are assumed to be legal.
3508 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3510 if (VT.getVectorNumElements() == 4 &&
3511 (VT.is128BitVector() || VT.is64BitVector())) {
3512 unsigned PFIndexes[4];
3513 for (unsigned i = 0; i != 4; ++i) {
3517 PFIndexes[i] = M[i];
3520 // Compute the index in the perfect shuffle table.
3521 unsigned PFTableIndex =
3522 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3523 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3524 unsigned Cost = (PFEntry >> 30);
3531 unsigned Imm, WhichResult;
3533 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3534 return (EltSize >= 32 ||
3535 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3536 isVREVMask(M, VT, 64) ||
3537 isVREVMask(M, VT, 32) ||
3538 isVREVMask(M, VT, 16) ||
3539 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3540 isVTRNMask(M, VT, WhichResult) ||
3541 isVUZPMask(M, VT, WhichResult) ||
3542 isVZIPMask(M, VT, WhichResult) ||
3543 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3544 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3545 isVZIP_v_undef_Mask(M, VT, WhichResult));
3548 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3549 /// the specified operations to build the shuffle.
3550 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3551 SDValue RHS, SelectionDAG &DAG,
3553 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3554 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3555 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3558 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3567 OP_VUZPL, // VUZP, left result
3568 OP_VUZPR, // VUZP, right result
3569 OP_VZIPL, // VZIP, left result
3570 OP_VZIPR, // VZIP, right result
3571 OP_VTRNL, // VTRN, left result
3572 OP_VTRNR // VTRN, right result
3575 if (OpNum == OP_COPY) {
3576 if (LHSID == (1*9+2)*9+3) return LHS;
3577 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3581 SDValue OpLHS, OpRHS;
3582 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3583 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3584 EVT VT = OpLHS.getValueType();
3587 default: llvm_unreachable("Unknown shuffle opcode!");
3589 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3594 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3595 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3599 return DAG.getNode(ARMISD::VEXT, dl, VT,
3601 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3604 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3605 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3608 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3609 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3612 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3613 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3617 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3618 SDValue V1 = Op.getOperand(0);
3619 SDValue V2 = Op.getOperand(1);
3620 DebugLoc dl = Op.getDebugLoc();
3621 EVT VT = Op.getValueType();
3622 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3623 SmallVector<int, 8> ShuffleMask;
3625 // Convert shuffles that are directly supported on NEON to target-specific
3626 // DAG nodes, instead of keeping them as shuffles and matching them again
3627 // during code selection. This is more efficient and avoids the possibility
3628 // of inconsistencies between legalization and selection.
3629 // FIXME: floating-point vectors should be canonicalized to integer vectors
3630 // of the same time so that they get CSEd properly.
3631 SVN->getMask(ShuffleMask);
3633 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3634 if (EltSize <= 32) {
3635 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3636 int Lane = SVN->getSplatIndex();
3637 // If this is undef splat, generate it via "just" vdup, if possible.
3638 if (Lane == -1) Lane = 0;
3640 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3641 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3643 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3644 DAG.getConstant(Lane, MVT::i32));
3649 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3652 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3653 DAG.getConstant(Imm, MVT::i32));
3656 if (isVREVMask(ShuffleMask, VT, 64))
3657 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3658 if (isVREVMask(ShuffleMask, VT, 32))
3659 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3660 if (isVREVMask(ShuffleMask, VT, 16))
3661 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3663 // Check for Neon shuffles that modify both input vectors in place.
3664 // If both results are used, i.e., if there are two shuffles with the same
3665 // source operands and with masks corresponding to both results of one of
3666 // these operations, DAG memoization will ensure that a single node is
3667 // used for both shuffles.
3668 unsigned WhichResult;
3669 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3670 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3671 V1, V2).getValue(WhichResult);
3672 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3673 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3674 V1, V2).getValue(WhichResult);
3675 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3676 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3677 V1, V2).getValue(WhichResult);
3679 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3680 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3681 V1, V1).getValue(WhichResult);
3682 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3683 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3684 V1, V1).getValue(WhichResult);
3685 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3686 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3687 V1, V1).getValue(WhichResult);
3690 // If the shuffle is not directly supported and it has 4 elements, use
3691 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3692 unsigned NumElts = VT.getVectorNumElements();
3694 unsigned PFIndexes[4];
3695 for (unsigned i = 0; i != 4; ++i) {
3696 if (ShuffleMask[i] < 0)
3699 PFIndexes[i] = ShuffleMask[i];
3702 // Compute the index in the perfect shuffle table.
3703 unsigned PFTableIndex =
3704 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3705 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3706 unsigned Cost = (PFEntry >> 30);
3709 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3712 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3713 if (EltSize >= 32) {
3714 // Do the expansion with floating-point types, since that is what the VFP
3715 // registers are defined to use, and since i64 is not legal.
3716 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3717 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3718 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3719 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3720 SmallVector<SDValue, 8> Ops;
3721 for (unsigned i = 0; i < NumElts; ++i) {
3722 if (ShuffleMask[i] < 0)
3723 Ops.push_back(DAG.getUNDEF(EltVT));
3725 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3726 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3727 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3730 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3731 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3737 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3738 EVT VT = Op.getValueType();
3739 DebugLoc dl = Op.getDebugLoc();
3740 SDValue Vec = Op.getOperand(0);
3741 SDValue Lane = Op.getOperand(1);
3742 assert(VT == MVT::i32 &&
3743 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3744 "unexpected type for custom-lowering vector extract");
3745 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3748 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3749 // The only time a CONCAT_VECTORS operation can have legal types is when
3750 // two 64-bit vectors are concatenated to a 128-bit vector.
3751 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3752 "unexpected CONCAT_VECTORS");
3753 DebugLoc dl = Op.getDebugLoc();
3754 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3755 SDValue Op0 = Op.getOperand(0);
3756 SDValue Op1 = Op.getOperand(1);
3757 if (Op0.getOpcode() != ISD::UNDEF)
3758 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3759 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3760 DAG.getIntPtrConstant(0));
3761 if (Op1.getOpcode() != ISD::UNDEF)
3762 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3763 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3764 DAG.getIntPtrConstant(1));
3765 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3768 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3769 /// an extending load, return the unextended value.
3770 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3771 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3772 return N->getOperand(0);
3773 LoadSDNode *LD = cast<LoadSDNode>(N);
3774 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3775 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3776 LD->isNonTemporal(), LD->getAlignment());
3779 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3780 // Multiplications are only custom-lowered for 128-bit vectors so that
3781 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3782 EVT VT = Op.getValueType();
3783 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3784 SDNode *N0 = Op.getOperand(0).getNode();
3785 SDNode *N1 = Op.getOperand(1).getNode();
3786 unsigned NewOpc = 0;
3787 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3788 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3789 NewOpc = ARMISD::VMULLs;
3790 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3791 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3792 NewOpc = ARMISD::VMULLu;
3793 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3794 // Fall through to expand this. It is not legal.
3797 // Other vector multiplications are legal.
3801 // Legalize to a VMULL instruction.
3802 DebugLoc DL = Op.getDebugLoc();
3803 SDValue Op0 = SkipExtension(N0, DAG);
3804 SDValue Op1 = SkipExtension(N1, DAG);
3806 assert(Op0.getValueType().is64BitVector() &&
3807 Op1.getValueType().is64BitVector() &&
3808 "unexpected types for extended operands to VMULL");
3809 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3812 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3813 switch (Op.getOpcode()) {
3814 default: llvm_unreachable("Don't know how to custom lower this!");
3815 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3816 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3817 case ISD::GlobalAddress:
3818 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3819 LowerGlobalAddressELF(Op, DAG);
3820 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3821 case ISD::SELECT: return LowerSELECT(Op, DAG);
3822 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3823 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3824 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3825 case ISD::VASTART: return LowerVASTART(Op, DAG);
3826 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3827 case ISD::SINT_TO_FP:
3828 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3829 case ISD::FP_TO_SINT:
3830 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3831 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3832 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3833 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3834 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3835 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3836 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3837 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
3838 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3840 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3843 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3844 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3845 case ISD::SRL_PARTS:
3846 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3847 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3848 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3849 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3850 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3851 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3852 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3853 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3854 case ISD::MUL: return LowerMUL(Op, DAG);
3859 /// ReplaceNodeResults - Replace the results of node with an illegal result
3860 /// type with new values built out of custom code.
3861 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3862 SmallVectorImpl<SDValue>&Results,
3863 SelectionDAG &DAG) const {
3865 switch (N->getOpcode()) {
3867 llvm_unreachable("Don't know how to custom expand this!");
3869 case ISD::BIT_CONVERT:
3870 Res = ExpandBIT_CONVERT(N, DAG);
3874 Res = LowerShift(N, DAG, Subtarget);
3878 Results.push_back(Res);
3881 //===----------------------------------------------------------------------===//
3882 // ARM Scheduler Hooks
3883 //===----------------------------------------------------------------------===//
3886 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3887 MachineBasicBlock *BB,
3888 unsigned Size) const {
3889 unsigned dest = MI->getOperand(0).getReg();
3890 unsigned ptr = MI->getOperand(1).getReg();
3891 unsigned oldval = MI->getOperand(2).getReg();
3892 unsigned newval = MI->getOperand(3).getReg();
3893 unsigned scratch = BB->getParent()->getRegInfo()
3894 .createVirtualRegister(ARM::GPRRegisterClass);
3895 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3896 DebugLoc dl = MI->getDebugLoc();
3897 bool isThumb2 = Subtarget->isThumb2();
3899 unsigned ldrOpc, strOpc;
3901 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3903 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3904 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3907 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3908 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3911 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3912 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3916 MachineFunction *MF = BB->getParent();
3917 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3918 MachineFunction::iterator It = BB;
3919 ++It; // insert the new blocks after the current block
3921 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3922 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3923 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3924 MF->insert(It, loop1MBB);
3925 MF->insert(It, loop2MBB);
3926 MF->insert(It, exitMBB);
3928 // Transfer the remainder of BB and its successor edges to exitMBB.
3929 exitMBB->splice(exitMBB->begin(), BB,
3930 llvm::next(MachineBasicBlock::iterator(MI)),
3932 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3936 // fallthrough --> loop1MBB
3937 BB->addSuccessor(loop1MBB);
3940 // ldrex dest, [ptr]
3944 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3945 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3946 .addReg(dest).addReg(oldval));
3947 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3948 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3949 BB->addSuccessor(loop2MBB);
3950 BB->addSuccessor(exitMBB);
3953 // strex scratch, newval, [ptr]
3957 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3959 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3960 .addReg(scratch).addImm(0));
3961 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3962 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3963 BB->addSuccessor(loop1MBB);
3964 BB->addSuccessor(exitMBB);
3970 MI->eraseFromParent(); // The instruction is gone now.
3976 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3977 unsigned Size, unsigned BinOpcode) const {
3978 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3979 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3981 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3982 MachineFunction *MF = BB->getParent();
3983 MachineFunction::iterator It = BB;
3986 unsigned dest = MI->getOperand(0).getReg();
3987 unsigned ptr = MI->getOperand(1).getReg();
3988 unsigned incr = MI->getOperand(2).getReg();
3989 DebugLoc dl = MI->getDebugLoc();
3991 bool isThumb2 = Subtarget->isThumb2();
3992 unsigned ldrOpc, strOpc;
3994 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3996 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3997 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4000 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4001 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4004 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4005 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4009 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4010 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4011 MF->insert(It, loopMBB);
4012 MF->insert(It, exitMBB);
4014 // Transfer the remainder of BB and its successor edges to exitMBB.
4015 exitMBB->splice(exitMBB->begin(), BB,
4016 llvm::next(MachineBasicBlock::iterator(MI)),
4018 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4020 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4021 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4022 unsigned scratch2 = (!BinOpcode) ? incr :
4023 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4027 // fallthrough --> loopMBB
4028 BB->addSuccessor(loopMBB);
4032 // <binop> scratch2, dest, incr
4033 // strex scratch, scratch2, ptr
4036 // fallthrough --> exitMBB
4038 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4040 // operand order needs to go the other way for NAND
4041 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4042 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4043 addReg(incr).addReg(dest)).addReg(0);
4045 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4046 addReg(dest).addReg(incr)).addReg(0);
4049 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4051 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4052 .addReg(scratch).addImm(0));
4053 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4054 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4056 BB->addSuccessor(loopMBB);
4057 BB->addSuccessor(exitMBB);
4063 MI->eraseFromParent(); // The instruction is gone now.
4069 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4070 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4071 E = MBB->succ_end(); I != E; ++I)
4074 llvm_unreachable("Expecting a BB with two successors!");
4078 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4079 MachineBasicBlock *BB) const {
4080 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4081 DebugLoc dl = MI->getDebugLoc();
4082 bool isThumb2 = Subtarget->isThumb2();
4083 switch (MI->getOpcode()) {
4086 llvm_unreachable("Unexpected instr type to insert");
4088 case ARM::ATOMIC_LOAD_ADD_I8:
4089 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4090 case ARM::ATOMIC_LOAD_ADD_I16:
4091 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4092 case ARM::ATOMIC_LOAD_ADD_I32:
4093 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4095 case ARM::ATOMIC_LOAD_AND_I8:
4096 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4097 case ARM::ATOMIC_LOAD_AND_I16:
4098 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4099 case ARM::ATOMIC_LOAD_AND_I32:
4100 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4102 case ARM::ATOMIC_LOAD_OR_I8:
4103 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4104 case ARM::ATOMIC_LOAD_OR_I16:
4105 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4106 case ARM::ATOMIC_LOAD_OR_I32:
4107 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4109 case ARM::ATOMIC_LOAD_XOR_I8:
4110 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4111 case ARM::ATOMIC_LOAD_XOR_I16:
4112 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4113 case ARM::ATOMIC_LOAD_XOR_I32:
4114 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4116 case ARM::ATOMIC_LOAD_NAND_I8:
4117 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4118 case ARM::ATOMIC_LOAD_NAND_I16:
4119 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4120 case ARM::ATOMIC_LOAD_NAND_I32:
4121 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4123 case ARM::ATOMIC_LOAD_SUB_I8:
4124 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4125 case ARM::ATOMIC_LOAD_SUB_I16:
4126 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4127 case ARM::ATOMIC_LOAD_SUB_I32:
4128 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4130 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4131 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4132 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4134 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4135 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4136 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4138 case ARM::tMOVCCr_pseudo: {
4139 // To "insert" a SELECT_CC instruction, we actually have to insert the
4140 // diamond control-flow pattern. The incoming instruction knows the
4141 // destination vreg to set, the condition code register to branch on, the
4142 // true/false values to select between, and a branch opcode to use.
4143 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4144 MachineFunction::iterator It = BB;
4150 // cmpTY ccX, r1, r2
4152 // fallthrough --> copy0MBB
4153 MachineBasicBlock *thisMBB = BB;
4154 MachineFunction *F = BB->getParent();
4155 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4156 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4157 F->insert(It, copy0MBB);
4158 F->insert(It, sinkMBB);
4160 // Transfer the remainder of BB and its successor edges to sinkMBB.
4161 sinkMBB->splice(sinkMBB->begin(), BB,
4162 llvm::next(MachineBasicBlock::iterator(MI)),
4164 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4166 BB->addSuccessor(copy0MBB);
4167 BB->addSuccessor(sinkMBB);
4169 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4170 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4173 // %FalseValue = ...
4174 // # fallthrough to sinkMBB
4177 // Update machine-CFG edges
4178 BB->addSuccessor(sinkMBB);
4181 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4184 BuildMI(*BB, BB->begin(), dl,
4185 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4186 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4187 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4189 MI->eraseFromParent(); // The pseudo instruction is gone now.
4194 case ARM::BCCZi64: {
4195 // Compare both parts that make up the double comparison separately for
4197 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4199 unsigned LHS1 = MI->getOperand(1).getReg();
4200 unsigned LHS2 = MI->getOperand(2).getReg();
4202 AddDefaultPred(BuildMI(BB, dl,
4203 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4204 .addReg(LHS1).addImm(0));
4205 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4206 .addReg(LHS2).addImm(0)
4207 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4209 unsigned RHS1 = MI->getOperand(3).getReg();
4210 unsigned RHS2 = MI->getOperand(4).getReg();
4211 AddDefaultPred(BuildMI(BB, dl,
4212 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4213 .addReg(LHS1).addReg(RHS1));
4214 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4215 .addReg(LHS2).addReg(RHS2)
4216 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4219 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4220 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4221 if (MI->getOperand(0).getImm() == ARMCC::NE)
4222 std::swap(destMBB, exitMBB);
4224 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4225 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4226 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4229 MI->eraseFromParent(); // The pseudo instruction is gone now.
4235 //===----------------------------------------------------------------------===//
4236 // ARM Optimization Hooks
4237 //===----------------------------------------------------------------------===//
4240 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4241 TargetLowering::DAGCombinerInfo &DCI) {
4242 SelectionDAG &DAG = DCI.DAG;
4243 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4244 EVT VT = N->getValueType(0);
4245 unsigned Opc = N->getOpcode();
4246 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4247 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4248 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4249 ISD::CondCode CC = ISD::SETCC_INVALID;
4252 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4254 SDValue CCOp = Slct.getOperand(0);
4255 if (CCOp.getOpcode() == ISD::SETCC)
4256 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4259 bool DoXform = false;
4261 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4264 if (LHS.getOpcode() == ISD::Constant &&
4265 cast<ConstantSDNode>(LHS)->isNullValue()) {
4267 } else if (CC != ISD::SETCC_INVALID &&
4268 RHS.getOpcode() == ISD::Constant &&
4269 cast<ConstantSDNode>(RHS)->isNullValue()) {
4270 std::swap(LHS, RHS);
4271 SDValue Op0 = Slct.getOperand(0);
4272 EVT OpVT = isSlctCC ? Op0.getValueType() :
4273 Op0.getOperand(0).getValueType();
4274 bool isInt = OpVT.isInteger();
4275 CC = ISD::getSetCCInverse(CC, isInt);
4277 if (!TLI.isCondCodeLegal(CC, OpVT))
4278 return SDValue(); // Inverse operator isn't legal.
4285 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4287 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4288 Slct.getOperand(0), Slct.getOperand(1), CC);
4289 SDValue CCOp = Slct.getOperand(0);
4291 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4292 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4293 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4294 CCOp, OtherOp, Result);
4299 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4300 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4301 /// called with the default operands, and if that fails, with commuted
4303 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4304 TargetLowering::DAGCombinerInfo &DCI) {
4305 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4306 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4307 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4308 if (Result.getNode()) return Result;
4313 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4315 static SDValue PerformADDCombine(SDNode *N,
4316 TargetLowering::DAGCombinerInfo &DCI) {
4317 SDValue N0 = N->getOperand(0);
4318 SDValue N1 = N->getOperand(1);
4320 // First try with the default operand order.
4321 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4322 if (Result.getNode())
4325 // If that didn't work, try again with the operands commuted.
4326 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4329 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4331 static SDValue PerformSUBCombine(SDNode *N,
4332 TargetLowering::DAGCombinerInfo &DCI) {
4333 SDValue N0 = N->getOperand(0);
4334 SDValue N1 = N->getOperand(1);
4336 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4337 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4338 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4339 if (Result.getNode()) return Result;
4345 static SDValue PerformMULCombine(SDNode *N,
4346 TargetLowering::DAGCombinerInfo &DCI,
4347 const ARMSubtarget *Subtarget) {
4348 SelectionDAG &DAG = DCI.DAG;
4350 if (Subtarget->isThumb1Only())
4353 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4356 EVT VT = N->getValueType(0);
4360 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4364 uint64_t MulAmt = C->getZExtValue();
4365 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4366 ShiftAmt = ShiftAmt & (32 - 1);
4367 SDValue V = N->getOperand(0);
4368 DebugLoc DL = N->getDebugLoc();
4371 MulAmt >>= ShiftAmt;
4372 if (isPowerOf2_32(MulAmt - 1)) {
4373 // (mul x, 2^N + 1) => (add (shl x, N), x)
4374 Res = DAG.getNode(ISD::ADD, DL, VT,
4375 V, DAG.getNode(ISD::SHL, DL, VT,
4376 V, DAG.getConstant(Log2_32(MulAmt-1),
4378 } else if (isPowerOf2_32(MulAmt + 1)) {
4379 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4380 Res = DAG.getNode(ISD::SUB, DL, VT,
4381 DAG.getNode(ISD::SHL, DL, VT,
4382 V, DAG.getConstant(Log2_32(MulAmt+1),
4389 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4390 DAG.getConstant(ShiftAmt, MVT::i32));
4392 // Do not add new nodes to DAG combiner worklist.
4393 DCI.CombineTo(N, Res, false);
4397 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4398 static SDValue PerformORCombine(SDNode *N,
4399 TargetLowering::DAGCombinerInfo &DCI,
4400 const ARMSubtarget *Subtarget) {
4401 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4404 // BFI is only available on V6T2+
4405 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4408 SelectionDAG &DAG = DCI.DAG;
4409 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4410 DebugLoc DL = N->getDebugLoc();
4411 // 1) or (and A, mask), val => ARMbfi A, val, mask
4412 // iff (val & mask) == val
4414 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4415 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4416 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4417 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4418 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4419 // (i.e., copy a bitfield value into another bitfield of the same width)
4420 if (N0.getOpcode() != ISD::AND)
4423 EVT VT = N->getValueType(0);
4428 // The value and the mask need to be constants so we can verify this is
4429 // actually a bitfield set. If the mask is 0xffff, we can do better
4430 // via a movt instruction, so don't use BFI in that case.
4431 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4434 unsigned Mask = C->getZExtValue();
4438 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4439 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4440 unsigned Val = C->getZExtValue();
4441 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4443 Val >>= CountTrailingZeros_32(~Mask);
4445 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4446 DAG.getConstant(Val, MVT::i32),
4447 DAG.getConstant(Mask, MVT::i32));
4449 // Do not add new nodes to DAG combiner worklist.
4450 DCI.CombineTo(N, Res, false);
4451 } else if (N1.getOpcode() == ISD::AND) {
4452 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4453 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4456 unsigned Mask2 = C->getZExtValue();
4458 if (ARM::isBitFieldInvertedMask(Mask) &&
4459 ARM::isBitFieldInvertedMask(~Mask2) &&
4460 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4461 // The pack halfword instruction works better for masks that fit it,
4462 // so use that when it's available.
4463 if (Subtarget->hasT2ExtractPack() &&
4464 (Mask == 0xffff || Mask == 0xffff0000))
4467 unsigned lsb = CountTrailingZeros_32(Mask2);
4468 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4469 DAG.getConstant(lsb, MVT::i32));
4470 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4471 DAG.getConstant(Mask, MVT::i32));
4472 // Do not add new nodes to DAG combiner worklist.
4473 DCI.CombineTo(N, Res, false);
4474 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4475 ARM::isBitFieldInvertedMask(Mask2) &&
4476 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4477 // The pack halfword instruction works better for masks that fit it,
4478 // so use that when it's available.
4479 if (Subtarget->hasT2ExtractPack() &&
4480 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4483 unsigned lsb = CountTrailingZeros_32(Mask);
4484 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4485 DAG.getConstant(lsb, MVT::i32));
4486 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4487 DAG.getConstant(Mask2, MVT::i32));
4488 // Do not add new nodes to DAG combiner worklist.
4489 DCI.CombineTo(N, Res, false);
4496 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4497 /// ARMISD::VMOVRRD.
4498 static SDValue PerformVMOVRRDCombine(SDNode *N,
4499 TargetLowering::DAGCombinerInfo &DCI) {
4500 // vmovrrd(vmovdrr x, y) -> x,y
4501 SDValue InDouble = N->getOperand(0);
4502 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4503 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4507 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4508 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4509 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4510 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4511 SDValue Op0 = N->getOperand(0);
4512 SDValue Op1 = N->getOperand(1);
4513 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4514 Op0 = Op0.getOperand(0);
4515 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4516 Op1 = Op1.getOperand(0);
4517 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4518 Op0.getNode() == Op1.getNode() &&
4519 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4520 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4521 N->getValueType(0), Op0.getOperand(0));
4525 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4526 /// ISD::BUILD_VECTOR.
4527 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4528 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4529 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4530 // into a pair of GPRs, which is fine when the value is used as a scalar,
4531 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4532 if (N->getNumOperands() == 2)
4533 return PerformVMOVDRRCombine(N, DAG);
4538 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4539 /// ARMISD::VDUPLANE.
4540 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
4541 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4543 SDValue Op = N->getOperand(0);
4544 EVT VT = N->getValueType(0);
4546 // Ignore bit_converts.
4547 while (Op.getOpcode() == ISD::BIT_CONVERT)
4548 Op = Op.getOperand(0);
4549 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4552 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4553 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4554 // The canonical VMOV for a zero vector uses a 32-bit element size.
4555 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4557 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4559 if (EltSize > VT.getVectorElementType().getSizeInBits())
4562 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4565 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4566 /// operand of a vector shift operation, where all the elements of the
4567 /// build_vector must have the same constant integer value.
4568 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4569 // Ignore bit_converts.
4570 while (Op.getOpcode() == ISD::BIT_CONVERT)
4571 Op = Op.getOperand(0);
4572 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4573 APInt SplatBits, SplatUndef;
4574 unsigned SplatBitSize;
4576 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4577 HasAnyUndefs, ElementBits) ||
4578 SplatBitSize > ElementBits)
4580 Cnt = SplatBits.getSExtValue();
4584 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4585 /// operand of a vector shift left operation. That value must be in the range:
4586 /// 0 <= Value < ElementBits for a left shift; or
4587 /// 0 <= Value <= ElementBits for a long left shift.
4588 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4589 assert(VT.isVector() && "vector shift count is not a vector type");
4590 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4591 if (! getVShiftImm(Op, ElementBits, Cnt))
4593 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4596 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4597 /// operand of a vector shift right operation. For a shift opcode, the value
4598 /// is positive, but for an intrinsic the value count must be negative. The
4599 /// absolute value must be in the range:
4600 /// 1 <= |Value| <= ElementBits for a right shift; or
4601 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4602 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4604 assert(VT.isVector() && "vector shift count is not a vector type");
4605 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4606 if (! getVShiftImm(Op, ElementBits, Cnt))
4610 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4613 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4614 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4615 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4618 // Don't do anything for most intrinsics.
4621 // Vector shifts: check for immediate versions and lower them.
4622 // Note: This is done during DAG combining instead of DAG legalizing because
4623 // the build_vectors for 64-bit vector element shift counts are generally
4624 // not legal, and it is hard to see their values after they get legalized to
4625 // loads from a constant pool.
4626 case Intrinsic::arm_neon_vshifts:
4627 case Intrinsic::arm_neon_vshiftu:
4628 case Intrinsic::arm_neon_vshiftls:
4629 case Intrinsic::arm_neon_vshiftlu:
4630 case Intrinsic::arm_neon_vshiftn:
4631 case Intrinsic::arm_neon_vrshifts:
4632 case Intrinsic::arm_neon_vrshiftu:
4633 case Intrinsic::arm_neon_vrshiftn:
4634 case Intrinsic::arm_neon_vqshifts:
4635 case Intrinsic::arm_neon_vqshiftu:
4636 case Intrinsic::arm_neon_vqshiftsu:
4637 case Intrinsic::arm_neon_vqshiftns:
4638 case Intrinsic::arm_neon_vqshiftnu:
4639 case Intrinsic::arm_neon_vqshiftnsu:
4640 case Intrinsic::arm_neon_vqrshiftns:
4641 case Intrinsic::arm_neon_vqrshiftnu:
4642 case Intrinsic::arm_neon_vqrshiftnsu: {
4643 EVT VT = N->getOperand(1).getValueType();
4645 unsigned VShiftOpc = 0;
4648 case Intrinsic::arm_neon_vshifts:
4649 case Intrinsic::arm_neon_vshiftu:
4650 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4651 VShiftOpc = ARMISD::VSHL;
4654 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4655 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4656 ARMISD::VSHRs : ARMISD::VSHRu);
4661 case Intrinsic::arm_neon_vshiftls:
4662 case Intrinsic::arm_neon_vshiftlu:
4663 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4665 llvm_unreachable("invalid shift count for vshll intrinsic");
4667 case Intrinsic::arm_neon_vrshifts:
4668 case Intrinsic::arm_neon_vrshiftu:
4669 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4673 case Intrinsic::arm_neon_vqshifts:
4674 case Intrinsic::arm_neon_vqshiftu:
4675 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4679 case Intrinsic::arm_neon_vqshiftsu:
4680 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4682 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4684 case Intrinsic::arm_neon_vshiftn:
4685 case Intrinsic::arm_neon_vrshiftn:
4686 case Intrinsic::arm_neon_vqshiftns:
4687 case Intrinsic::arm_neon_vqshiftnu:
4688 case Intrinsic::arm_neon_vqshiftnsu:
4689 case Intrinsic::arm_neon_vqrshiftns:
4690 case Intrinsic::arm_neon_vqrshiftnu:
4691 case Intrinsic::arm_neon_vqrshiftnsu:
4692 // Narrowing shifts require an immediate right shift.
4693 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4695 llvm_unreachable("invalid shift count for narrowing vector shift "
4699 llvm_unreachable("unhandled vector shift");
4703 case Intrinsic::arm_neon_vshifts:
4704 case Intrinsic::arm_neon_vshiftu:
4705 // Opcode already set above.
4707 case Intrinsic::arm_neon_vshiftls:
4708 case Intrinsic::arm_neon_vshiftlu:
4709 if (Cnt == VT.getVectorElementType().getSizeInBits())
4710 VShiftOpc = ARMISD::VSHLLi;
4712 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4713 ARMISD::VSHLLs : ARMISD::VSHLLu);
4715 case Intrinsic::arm_neon_vshiftn:
4716 VShiftOpc = ARMISD::VSHRN; break;
4717 case Intrinsic::arm_neon_vrshifts:
4718 VShiftOpc = ARMISD::VRSHRs; break;
4719 case Intrinsic::arm_neon_vrshiftu:
4720 VShiftOpc = ARMISD::VRSHRu; break;
4721 case Intrinsic::arm_neon_vrshiftn:
4722 VShiftOpc = ARMISD::VRSHRN; break;
4723 case Intrinsic::arm_neon_vqshifts:
4724 VShiftOpc = ARMISD::VQSHLs; break;
4725 case Intrinsic::arm_neon_vqshiftu:
4726 VShiftOpc = ARMISD::VQSHLu; break;
4727 case Intrinsic::arm_neon_vqshiftsu:
4728 VShiftOpc = ARMISD::VQSHLsu; break;
4729 case Intrinsic::arm_neon_vqshiftns:
4730 VShiftOpc = ARMISD::VQSHRNs; break;
4731 case Intrinsic::arm_neon_vqshiftnu:
4732 VShiftOpc = ARMISD::VQSHRNu; break;
4733 case Intrinsic::arm_neon_vqshiftnsu:
4734 VShiftOpc = ARMISD::VQSHRNsu; break;
4735 case Intrinsic::arm_neon_vqrshiftns:
4736 VShiftOpc = ARMISD::VQRSHRNs; break;
4737 case Intrinsic::arm_neon_vqrshiftnu:
4738 VShiftOpc = ARMISD::VQRSHRNu; break;
4739 case Intrinsic::arm_neon_vqrshiftnsu:
4740 VShiftOpc = ARMISD::VQRSHRNsu; break;
4743 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4744 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4747 case Intrinsic::arm_neon_vshiftins: {
4748 EVT VT = N->getOperand(1).getValueType();
4750 unsigned VShiftOpc = 0;
4752 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4753 VShiftOpc = ARMISD::VSLI;
4754 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4755 VShiftOpc = ARMISD::VSRI;
4757 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4760 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4761 N->getOperand(1), N->getOperand(2),
4762 DAG.getConstant(Cnt, MVT::i32));
4765 case Intrinsic::arm_neon_vqrshifts:
4766 case Intrinsic::arm_neon_vqrshiftu:
4767 // No immediate versions of these to check for.
4774 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4775 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4776 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4777 /// vector element shift counts are generally not legal, and it is hard to see
4778 /// their values after they get legalized to loads from a constant pool.
4779 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4780 const ARMSubtarget *ST) {
4781 EVT VT = N->getValueType(0);
4783 // Nothing to be done for scalar shifts.
4784 if (! VT.isVector())
4787 assert(ST->hasNEON() && "unexpected vector shift");
4790 switch (N->getOpcode()) {
4791 default: llvm_unreachable("unexpected shift opcode");
4794 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4795 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4796 DAG.getConstant(Cnt, MVT::i32));
4801 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4802 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4803 ARMISD::VSHRs : ARMISD::VSHRu);
4804 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4805 DAG.getConstant(Cnt, MVT::i32));
4811 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4812 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4813 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4814 const ARMSubtarget *ST) {
4815 SDValue N0 = N->getOperand(0);
4817 // Check for sign- and zero-extensions of vector extract operations of 8-
4818 // and 16-bit vector elements. NEON supports these directly. They are
4819 // handled during DAG combining because type legalization will promote them
4820 // to 32-bit types and it is messy to recognize the operations after that.
4821 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4822 SDValue Vec = N0.getOperand(0);
4823 SDValue Lane = N0.getOperand(1);
4824 EVT VT = N->getValueType(0);
4825 EVT EltVT = N0.getValueType();
4826 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4828 if (VT == MVT::i32 &&
4829 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4830 TLI.isTypeLegal(Vec.getValueType())) {
4833 switch (N->getOpcode()) {
4834 default: llvm_unreachable("unexpected opcode");
4835 case ISD::SIGN_EXTEND:
4836 Opc = ARMISD::VGETLANEs;
4838 case ISD::ZERO_EXTEND:
4839 case ISD::ANY_EXTEND:
4840 Opc = ARMISD::VGETLANEu;
4843 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4850 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4851 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4852 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4853 const ARMSubtarget *ST) {
4854 // If the target supports NEON, try to use vmax/vmin instructions for f32
4855 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4856 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4857 // a NaN; only do the transformation when it matches that behavior.
4859 // For now only do this when using NEON for FP operations; if using VFP, it
4860 // is not obvious that the benefit outweighs the cost of switching to the
4862 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4863 N->getValueType(0) != MVT::f32)
4866 SDValue CondLHS = N->getOperand(0);
4867 SDValue CondRHS = N->getOperand(1);
4868 SDValue LHS = N->getOperand(2);
4869 SDValue RHS = N->getOperand(3);
4870 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4872 unsigned Opcode = 0;
4874 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4875 IsReversed = false; // x CC y ? x : y
4876 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4877 IsReversed = true ; // x CC y ? y : x
4891 // If LHS is NaN, an ordered comparison will be false and the result will
4892 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4893 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4894 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4895 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4897 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4898 // will return -0, so vmin can only be used for unsafe math or if one of
4899 // the operands is known to be nonzero.
4900 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4902 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4904 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4913 // If LHS is NaN, an ordered comparison will be false and the result will
4914 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4915 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4916 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4917 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4919 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4920 // will return +0, so vmax can only be used for unsafe math or if one of
4921 // the operands is known to be nonzero.
4922 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4924 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4926 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4932 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4935 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4936 DAGCombinerInfo &DCI) const {
4937 switch (N->getOpcode()) {
4939 case ISD::ADD: return PerformADDCombine(N, DCI);
4940 case ISD::SUB: return PerformSUBCombine(N, DCI);
4941 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4942 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4943 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4944 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
4945 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
4946 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
4947 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4950 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4951 case ISD::SIGN_EXTEND:
4952 case ISD::ZERO_EXTEND:
4953 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4954 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4959 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4960 if (!Subtarget->allowsUnalignedMem())
4963 switch (VT.getSimpleVT().SimpleTy) {
4970 // FIXME: VLD1 etc with standard alignment is legal.
4974 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4979 switch (VT.getSimpleVT().SimpleTy) {
4980 default: return false;
4995 if ((V & (Scale - 1)) != 0)
4998 return V == (V & ((1LL << 5) - 1));
5001 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
5002 const ARMSubtarget *Subtarget) {
5009 switch (VT.getSimpleVT().SimpleTy) {
5010 default: return false;
5015 // + imm12 or - imm8
5017 return V == (V & ((1LL << 8) - 1));
5018 return V == (V & ((1LL << 12) - 1));
5021 // Same as ARM mode. FIXME: NEON?
5022 if (!Subtarget->hasVFP2())
5027 return V == (V & ((1LL << 8) - 1));
5031 /// isLegalAddressImmediate - Return true if the integer value can be used
5032 /// as the offset of the target addressing mode for load / store of the
5034 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5035 const ARMSubtarget *Subtarget) {
5042 if (Subtarget->isThumb1Only())
5043 return isLegalT1AddressImmediate(V, VT);
5044 else if (Subtarget->isThumb2())
5045 return isLegalT2AddressImmediate(V, VT, Subtarget);
5050 switch (VT.getSimpleVT().SimpleTy) {
5051 default: return false;
5056 return V == (V & ((1LL << 12) - 1));
5059 return V == (V & ((1LL << 8) - 1));
5062 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5067 return V == (V & ((1LL << 8) - 1));
5071 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5073 int Scale = AM.Scale;
5077 switch (VT.getSimpleVT().SimpleTy) {
5078 default: return false;
5087 return Scale == 2 || Scale == 4 || Scale == 8;
5090 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5094 // Note, we allow "void" uses (basically, uses that aren't loads or
5095 // stores), because arm allows folding a scale into many arithmetic
5096 // operations. This should be made more precise and revisited later.
5098 // Allow r << imm, but the imm has to be a multiple of two.
5099 if (Scale & 1) return false;
5100 return isPowerOf2_32(Scale);
5104 /// isLegalAddressingMode - Return true if the addressing mode represented
5105 /// by AM is legal for this target, for a load/store of the specified type.
5106 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5107 const Type *Ty) const {
5108 EVT VT = getValueType(Ty, true);
5109 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5112 // Can never fold addr of global into load/store.
5117 case 0: // no scale reg, must be "r+i" or "r", or "i".
5120 if (Subtarget->isThumb1Only())
5124 // ARM doesn't support any R+R*scale+imm addr modes.
5131 if (Subtarget->isThumb2())
5132 return isLegalT2ScaledAddressingMode(AM, VT);
5134 int Scale = AM.Scale;
5135 switch (VT.getSimpleVT().SimpleTy) {
5136 default: return false;
5140 if (Scale < 0) Scale = -Scale;
5144 return isPowerOf2_32(Scale & ~1);
5148 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5153 // Note, we allow "void" uses (basically, uses that aren't loads or
5154 // stores), because arm allows folding a scale into many arithmetic
5155 // operations. This should be made more precise and revisited later.
5157 // Allow r << imm, but the imm has to be a multiple of two.
5158 if (Scale & 1) return false;
5159 return isPowerOf2_32(Scale);
5166 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5167 /// icmp immediate, that is the target has icmp instructions which can compare
5168 /// a register against the immediate without having to materialize the
5169 /// immediate into a register.
5170 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5171 if (!Subtarget->isThumb())
5172 return ARM_AM::getSOImmVal(Imm) != -1;
5173 if (Subtarget->isThumb2())
5174 return ARM_AM::getT2SOImmVal(Imm) != -1;
5175 return Imm >= 0 && Imm <= 255;
5178 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5179 bool isSEXTLoad, SDValue &Base,
5180 SDValue &Offset, bool &isInc,
5181 SelectionDAG &DAG) {
5182 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5185 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5187 Base = Ptr->getOperand(0);
5188 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5189 int RHSC = (int)RHS->getZExtValue();
5190 if (RHSC < 0 && RHSC > -256) {
5191 assert(Ptr->getOpcode() == ISD::ADD);
5193 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5197 isInc = (Ptr->getOpcode() == ISD::ADD);
5198 Offset = Ptr->getOperand(1);
5200 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5202 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5203 int RHSC = (int)RHS->getZExtValue();
5204 if (RHSC < 0 && RHSC > -0x1000) {
5205 assert(Ptr->getOpcode() == ISD::ADD);
5207 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5208 Base = Ptr->getOperand(0);
5213 if (Ptr->getOpcode() == ISD::ADD) {
5215 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5216 if (ShOpcVal != ARM_AM::no_shift) {
5217 Base = Ptr->getOperand(1);
5218 Offset = Ptr->getOperand(0);
5220 Base = Ptr->getOperand(0);
5221 Offset = Ptr->getOperand(1);
5226 isInc = (Ptr->getOpcode() == ISD::ADD);
5227 Base = Ptr->getOperand(0);
5228 Offset = Ptr->getOperand(1);
5232 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5236 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5237 bool isSEXTLoad, SDValue &Base,
5238 SDValue &Offset, bool &isInc,
5239 SelectionDAG &DAG) {
5240 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5243 Base = Ptr->getOperand(0);
5244 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5245 int RHSC = (int)RHS->getZExtValue();
5246 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5247 assert(Ptr->getOpcode() == ISD::ADD);
5249 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5251 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5252 isInc = Ptr->getOpcode() == ISD::ADD;
5253 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5261 /// getPreIndexedAddressParts - returns true by value, base pointer and
5262 /// offset pointer and addressing mode by reference if the node's address
5263 /// can be legally represented as pre-indexed load / store address.
5265 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5267 ISD::MemIndexedMode &AM,
5268 SelectionDAG &DAG) const {
5269 if (Subtarget->isThumb1Only())
5274 bool isSEXTLoad = false;
5275 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5276 Ptr = LD->getBasePtr();
5277 VT = LD->getMemoryVT();
5278 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5279 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5280 Ptr = ST->getBasePtr();
5281 VT = ST->getMemoryVT();
5286 bool isLegal = false;
5287 if (Subtarget->isThumb2())
5288 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5289 Offset, isInc, DAG);
5291 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5292 Offset, isInc, DAG);
5296 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5300 /// getPostIndexedAddressParts - returns true by value, base pointer and
5301 /// offset pointer and addressing mode by reference if this node can be
5302 /// combined with a load / store to form a post-indexed load / store.
5303 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5306 ISD::MemIndexedMode &AM,
5307 SelectionDAG &DAG) const {
5308 if (Subtarget->isThumb1Only())
5313 bool isSEXTLoad = false;
5314 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5315 VT = LD->getMemoryVT();
5316 Ptr = LD->getBasePtr();
5317 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5318 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5319 VT = ST->getMemoryVT();
5320 Ptr = ST->getBasePtr();
5325 bool isLegal = false;
5326 if (Subtarget->isThumb2())
5327 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5330 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5336 // Swap base ptr and offset to catch more post-index load / store when
5337 // it's legal. In Thumb2 mode, offset must be an immediate.
5338 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5339 !Subtarget->isThumb2())
5340 std::swap(Base, Offset);
5342 // Post-indexed load / store update the base pointer.
5347 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5351 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5355 const SelectionDAG &DAG,
5356 unsigned Depth) const {
5357 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5358 switch (Op.getOpcode()) {
5360 case ARMISD::CMOV: {
5361 // Bits are known zero/one if known on the LHS and RHS.
5362 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5363 if (KnownZero == 0 && KnownOne == 0) return;
5365 APInt KnownZeroRHS, KnownOneRHS;
5366 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5367 KnownZeroRHS, KnownOneRHS, Depth+1);
5368 KnownZero &= KnownZeroRHS;
5369 KnownOne &= KnownOneRHS;
5375 //===----------------------------------------------------------------------===//
5376 // ARM Inline Assembly Support
5377 //===----------------------------------------------------------------------===//
5379 /// getConstraintType - Given a constraint letter, return the type of
5380 /// constraint it is for this target.
5381 ARMTargetLowering::ConstraintType
5382 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5383 if (Constraint.size() == 1) {
5384 switch (Constraint[0]) {
5386 case 'l': return C_RegisterClass;
5387 case 'w': return C_RegisterClass;
5390 return TargetLowering::getConstraintType(Constraint);
5393 std::pair<unsigned, const TargetRegisterClass*>
5394 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5396 if (Constraint.size() == 1) {
5397 // GCC ARM Constraint Letters
5398 switch (Constraint[0]) {
5400 if (Subtarget->isThumb())
5401 return std::make_pair(0U, ARM::tGPRRegisterClass);
5403 return std::make_pair(0U, ARM::GPRRegisterClass);
5405 return std::make_pair(0U, ARM::GPRRegisterClass);
5408 return std::make_pair(0U, ARM::SPRRegisterClass);
5409 if (VT.getSizeInBits() == 64)
5410 return std::make_pair(0U, ARM::DPRRegisterClass);
5411 if (VT.getSizeInBits() == 128)
5412 return std::make_pair(0U, ARM::QPRRegisterClass);
5416 if (StringRef("{cc}").equals_lower(Constraint))
5417 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5419 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5422 std::vector<unsigned> ARMTargetLowering::
5423 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5425 if (Constraint.size() != 1)
5426 return std::vector<unsigned>();
5428 switch (Constraint[0]) { // GCC ARM Constraint Letters
5431 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5432 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5435 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5436 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5437 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5438 ARM::R12, ARM::LR, 0);
5441 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5442 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5443 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5444 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5445 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5446 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5447 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5448 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5449 if (VT.getSizeInBits() == 64)
5450 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5451 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5452 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5453 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5454 if (VT.getSizeInBits() == 128)
5455 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5456 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5460 return std::vector<unsigned>();
5463 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5464 /// vector. If it is invalid, don't add anything to Ops.
5465 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5467 std::vector<SDValue>&Ops,
5468 SelectionDAG &DAG) const {
5469 SDValue Result(0, 0);
5471 switch (Constraint) {
5473 case 'I': case 'J': case 'K': case 'L':
5474 case 'M': case 'N': case 'O':
5475 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5479 int64_t CVal64 = C->getSExtValue();
5480 int CVal = (int) CVal64;
5481 // None of these constraints allow values larger than 32 bits. Check
5482 // that the value fits in an int.
5486 switch (Constraint) {
5488 if (Subtarget->isThumb1Only()) {
5489 // This must be a constant between 0 and 255, for ADD
5491 if (CVal >= 0 && CVal <= 255)
5493 } else if (Subtarget->isThumb2()) {
5494 // A constant that can be used as an immediate value in a
5495 // data-processing instruction.
5496 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5499 // A constant that can be used as an immediate value in a
5500 // data-processing instruction.
5501 if (ARM_AM::getSOImmVal(CVal) != -1)
5507 if (Subtarget->isThumb()) { // FIXME thumb2
5508 // This must be a constant between -255 and -1, for negated ADD
5509 // immediates. This can be used in GCC with an "n" modifier that
5510 // prints the negated value, for use with SUB instructions. It is
5511 // not useful otherwise but is implemented for compatibility.
5512 if (CVal >= -255 && CVal <= -1)
5515 // This must be a constant between -4095 and 4095. It is not clear
5516 // what this constraint is intended for. Implemented for
5517 // compatibility with GCC.
5518 if (CVal >= -4095 && CVal <= 4095)
5524 if (Subtarget->isThumb1Only()) {
5525 // A 32-bit value where only one byte has a nonzero value. Exclude
5526 // zero to match GCC. This constraint is used by GCC internally for
5527 // constants that can be loaded with a move/shift combination.
5528 // It is not useful otherwise but is implemented for compatibility.
5529 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5531 } else if (Subtarget->isThumb2()) {
5532 // A constant whose bitwise inverse can be used as an immediate
5533 // value in a data-processing instruction. This can be used in GCC
5534 // with a "B" modifier that prints the inverted value, for use with
5535 // BIC and MVN instructions. It is not useful otherwise but is
5536 // implemented for compatibility.
5537 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5540 // A constant whose bitwise inverse can be used as an immediate
5541 // value in a data-processing instruction. This can be used in GCC
5542 // with a "B" modifier that prints the inverted value, for use with
5543 // BIC and MVN instructions. It is not useful otherwise but is
5544 // implemented for compatibility.
5545 if (ARM_AM::getSOImmVal(~CVal) != -1)
5551 if (Subtarget->isThumb1Only()) {
5552 // This must be a constant between -7 and 7,
5553 // for 3-operand ADD/SUB immediate instructions.
5554 if (CVal >= -7 && CVal < 7)
5556 } else if (Subtarget->isThumb2()) {
5557 // A constant whose negation can be used as an immediate value in a
5558 // data-processing instruction. This can be used in GCC with an "n"
5559 // modifier that prints the negated value, for use with SUB
5560 // instructions. It is not useful otherwise but is implemented for
5562 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5565 // A constant whose negation can be used as an immediate value in a
5566 // data-processing instruction. This can be used in GCC with an "n"
5567 // modifier that prints the negated value, for use with SUB
5568 // instructions. It is not useful otherwise but is implemented for
5570 if (ARM_AM::getSOImmVal(-CVal) != -1)
5576 if (Subtarget->isThumb()) { // FIXME thumb2
5577 // This must be a multiple of 4 between 0 and 1020, for
5578 // ADD sp + immediate.
5579 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5582 // A power of two or a constant between 0 and 32. This is used in
5583 // GCC for the shift amount on shifted register operands, but it is
5584 // useful in general for any shift amounts.
5585 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5591 if (Subtarget->isThumb()) { // FIXME thumb2
5592 // This must be a constant between 0 and 31, for shift amounts.
5593 if (CVal >= 0 && CVal <= 31)
5599 if (Subtarget->isThumb()) { // FIXME thumb2
5600 // This must be a multiple of 4 between -508 and 508, for
5601 // ADD/SUB sp = sp + immediate.
5602 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5607 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5611 if (Result.getNode()) {
5612 Ops.push_back(Result);
5615 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5619 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5620 // The ARM target isn't yet aware of offsets.
5624 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5625 APInt Imm = FPImm.bitcastToAPInt();
5626 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5627 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5628 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5630 // We can handle 4 bits of mantissa.
5631 // mantissa = (16+UInt(e:f:g:h))/16.
5632 if (Mantissa & 0x7ffff)
5635 if ((Mantissa & 0xf) != Mantissa)
5638 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5639 if (Exp < -3 || Exp > 4)
5641 Exp = ((Exp+3) & 0x7) ^ 4;
5643 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5646 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5647 APInt Imm = FPImm.bitcastToAPInt();
5648 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5649 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5650 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5652 // We can handle 4 bits of mantissa.
5653 // mantissa = (16+UInt(e:f:g:h))/16.
5654 if (Mantissa & 0xffffffffffffLL)
5657 if ((Mantissa & 0xf) != Mantissa)
5660 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5661 if (Exp < -3 || Exp > 4)
5663 Exp = ((Exp+3) & 0x7) ^ 4;
5665 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5668 bool ARM::isBitFieldInvertedMask(unsigned v) {
5669 if (v == 0xffffffff)
5671 // there can be 1's on either or both "outsides", all the "inside"
5673 unsigned int lsb = 0, msb = 31;
5674 while (v & (1 << msb)) --msb;
5675 while (v & (1 << lsb)) ++lsb;
5676 for (unsigned int i = lsb; i <= msb; ++i) {
5683 /// isFPImmLegal - Returns true if the target can instruction select the
5684 /// specified FP immediate natively. If false, the legalizer will
5685 /// materialize the FP immediate as a load from a constant pool.
5686 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5687 if (!Subtarget->hasVFP3())
5690 return ARM::getVFPf32Imm(Imm) != -1;
5692 return ARM::getVFPf64Imm(Imm) != -1;
5696 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5697 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5698 /// specified in the intrinsic calls.
5699 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5701 unsigned Intrinsic) const {
5702 switch (Intrinsic) {
5703 case Intrinsic::arm_neon_vld1:
5704 case Intrinsic::arm_neon_vld2:
5705 case Intrinsic::arm_neon_vld3:
5706 case Intrinsic::arm_neon_vld4:
5707 case Intrinsic::arm_neon_vld2lane:
5708 case Intrinsic::arm_neon_vld3lane:
5709 case Intrinsic::arm_neon_vld4lane: {
5710 Info.opc = ISD::INTRINSIC_W_CHAIN;
5711 // Conservatively set memVT to the entire set of vectors loaded.
5712 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
5713 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5714 Info.ptrVal = I.getArgOperand(0);
5716 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5717 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5718 Info.vol = false; // volatile loads with NEON intrinsics not supported
5719 Info.readMem = true;
5720 Info.writeMem = false;
5723 case Intrinsic::arm_neon_vst1:
5724 case Intrinsic::arm_neon_vst2:
5725 case Intrinsic::arm_neon_vst3:
5726 case Intrinsic::arm_neon_vst4:
5727 case Intrinsic::arm_neon_vst2lane:
5728 case Intrinsic::arm_neon_vst3lane:
5729 case Intrinsic::arm_neon_vst4lane: {
5730 Info.opc = ISD::INTRINSIC_VOID;
5731 // Conservatively set memVT to the entire set of vectors stored.
5732 unsigned NumElts = 0;
5733 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5734 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
5735 if (!ArgTy->isVectorTy())
5737 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
5739 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5740 Info.ptrVal = I.getArgOperand(0);
5742 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5743 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5744 Info.vol = false; // volatile stores with NEON intrinsics not supported
5745 Info.readMem = false;
5746 Info.writeMem = true;