1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMAddressingModes.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMISelLowering.h"
21 #include "ARMMachineFunctionInfo.h"
22 #include "ARMPerfectShuffle.h"
23 #include "ARMRegisterInfo.h"
24 #include "ARMSubtarget.h"
25 #include "ARMTargetMachine.h"
26 #include "ARMTargetObjectFile.h"
27 #include "llvm/CallingConv.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Instruction.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/Intrinsics.h"
34 #include "llvm/Type.h"
35 #include "llvm/CodeGen/CallingConvLower.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/PseudoSourceValue.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
54 STATISTIC(NumTailCalls, "Number of tail calls");
56 // This option should go away when tail calls fully work.
58 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
59 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
63 EnableARMLongCalls("arm-long-calls", cl::Hidden,
64 cl::desc("Generate calls via indirect call instructions"),
68 ARMInterworking("arm-interworking", cl::Hidden,
69 cl::desc("Enable / disable ARM interworking (for debugging only)"),
72 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
73 EVT PromotedBitwiseVT) {
74 if (VT != PromotedLdStVT) {
75 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
76 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
77 PromotedLdStVT.getSimpleVT());
79 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
80 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
81 PromotedLdStVT.getSimpleVT());
84 EVT ElemTy = VT.getVectorElementType();
85 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
86 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
87 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
88 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
89 if (ElemTy != MVT::i32) {
90 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
91 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
95 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
97 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
98 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
99 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
100 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
101 if (VT.isInteger()) {
102 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
103 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
104 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
105 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
106 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
108 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
110 // Promote all bit-wise operations.
111 if (VT.isInteger() && VT != PromotedBitwiseVT) {
112 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
113 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
114 PromotedBitwiseVT.getSimpleVT());
115 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
116 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
117 PromotedBitwiseVT.getSimpleVT());
118 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
119 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
120 PromotedBitwiseVT.getSimpleVT());
123 // Neon does not support vector divide/remainder operations.
124 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
125 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
128 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
129 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
132 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
133 addRegisterClass(VT, ARM::DPRRegisterClass);
134 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
137 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
138 addRegisterClass(VT, ARM::QPRRegisterClass);
139 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
142 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
143 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
144 return new TargetLoweringObjectFileMachO();
146 return new ARMElfTargetObjectFile();
149 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
150 : TargetLowering(TM, createTLOF(TM)) {
151 Subtarget = &TM.getSubtarget<ARMSubtarget>();
152 RegInfo = TM.getRegisterInfo();
153 Itins = TM.getInstrItineraryData();
155 if (Subtarget->isTargetDarwin()) {
156 // Uses VFP for Thumb libfuncs if available.
157 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
158 // Single-precision floating-point arithmetic.
159 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
160 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
161 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
162 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
164 // Double-precision floating-point arithmetic.
165 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
166 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
167 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
168 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
170 // Single-precision comparisons.
171 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
172 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
173 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
174 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
175 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
176 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
177 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
178 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
180 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
181 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
182 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
183 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
189 // Double-precision comparisons.
190 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
191 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
192 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
193 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
194 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
195 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
196 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
197 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
199 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
208 // Floating-point to integer conversions.
209 // i64 conversions are done via library routines even when generating VFP
210 // instructions, so use the same ones.
211 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
212 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
213 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
214 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
216 // Conversions between floating types.
217 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
218 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
220 // Integer to floating-point conversions.
221 // i64 conversions are done via library routines even when generating VFP
222 // instructions, so use the same ones.
223 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
224 // e.g., __floatunsidf vs. __floatunssidfvfp.
225 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
226 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
227 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
228 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
232 // These libcalls are not available in 32-bit.
233 setLibcallName(RTLIB::SHL_I128, 0);
234 setLibcallName(RTLIB::SRL_I128, 0);
235 setLibcallName(RTLIB::SRA_I128, 0);
237 if (Subtarget->isAAPCS_ABI()) {
238 // Double-precision floating-point arithmetic helper functions
239 // RTABI chapter 4.1.2, Table 2
240 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
241 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
242 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
243 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
244 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
245 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
246 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
247 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
249 // Double-precision floating-point comparison helper functions
250 // RTABI chapter 4.1.2, Table 3
251 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
252 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
253 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
254 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
255 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
256 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
257 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
258 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
259 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
260 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
261 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
262 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
263 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
264 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
265 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
266 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
267 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
270 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
276 // Single-precision floating-point arithmetic helper functions
277 // RTABI chapter 4.1.2, Table 4
278 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
279 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
280 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
281 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
282 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
283 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
284 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
285 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
287 // Single-precision floating-point comparison helper functions
288 // RTABI chapter 4.1.2, Table 5
289 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
290 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
291 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
292 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
293 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
294 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
295 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
296 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
297 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
298 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
299 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
300 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
301 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
302 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
303 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
304 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
305 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
308 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
314 // Floating-point to integer conversions.
315 // RTABI chapter 4.1.2, Table 6
316 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
317 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
318 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
319 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
320 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
321 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
322 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
323 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
324 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
325 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
326 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
327 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
333 // Conversions between floating types.
334 // RTABI chapter 4.1.2, Table 7
335 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
336 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
337 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
340 // Integer to floating-point conversions.
341 // RTABI chapter 4.1.2, Table 8
342 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
343 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
344 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
345 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
346 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
347 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
348 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
349 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
350 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
359 // Long long helper functions
360 // RTABI chapter 4.2, Table 9
361 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
362 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
363 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
364 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
365 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
366 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
367 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
368 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
369 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
370 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
371 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
374 // Integer division functions
375 // RTABI chapter 4.3.1
376 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
377 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
378 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
379 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
380 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
381 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
382 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
384 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
385 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
386 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
390 if (Subtarget->isThumb1Only())
391 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
393 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
394 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
395 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
396 if (!Subtarget->isFPOnlySP())
397 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
399 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
402 if (Subtarget->hasNEON()) {
403 addDRTypeForNEON(MVT::v2f32);
404 addDRTypeForNEON(MVT::v8i8);
405 addDRTypeForNEON(MVT::v4i16);
406 addDRTypeForNEON(MVT::v2i32);
407 addDRTypeForNEON(MVT::v1i64);
409 addQRTypeForNEON(MVT::v4f32);
410 addQRTypeForNEON(MVT::v2f64);
411 addQRTypeForNEON(MVT::v16i8);
412 addQRTypeForNEON(MVT::v8i16);
413 addQRTypeForNEON(MVT::v4i32);
414 addQRTypeForNEON(MVT::v2i64);
416 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
417 // neither Neon nor VFP support any arithmetic operations on it.
418 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
419 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
420 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
421 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
422 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
423 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
424 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
425 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
426 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
427 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
428 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
429 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
430 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
431 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
432 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
433 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
434 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
435 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
436 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
437 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
438 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
439 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
440 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
441 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
443 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
445 // Neon does not support some operations on v1i64 and v2i64 types.
446 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
447 // Custom handling for some quad-vector types to detect VMULL.
448 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
449 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
450 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
451 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
452 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
454 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
455 setTargetDAGCombine(ISD::SHL);
456 setTargetDAGCombine(ISD::SRL);
457 setTargetDAGCombine(ISD::SRA);
458 setTargetDAGCombine(ISD::SIGN_EXTEND);
459 setTargetDAGCombine(ISD::ZERO_EXTEND);
460 setTargetDAGCombine(ISD::ANY_EXTEND);
461 setTargetDAGCombine(ISD::SELECT_CC);
462 setTargetDAGCombine(ISD::BUILD_VECTOR);
465 computeRegisterProperties();
467 // ARM does not have f32 extending load.
468 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
470 // ARM does not have i1 sign extending load.
471 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
473 // ARM supports all 4 flavors of integer indexed load / store.
474 if (!Subtarget->isThumb1Only()) {
475 for (unsigned im = (unsigned)ISD::PRE_INC;
476 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
477 setIndexedLoadAction(im, MVT::i1, Legal);
478 setIndexedLoadAction(im, MVT::i8, Legal);
479 setIndexedLoadAction(im, MVT::i16, Legal);
480 setIndexedLoadAction(im, MVT::i32, Legal);
481 setIndexedStoreAction(im, MVT::i1, Legal);
482 setIndexedStoreAction(im, MVT::i8, Legal);
483 setIndexedStoreAction(im, MVT::i16, Legal);
484 setIndexedStoreAction(im, MVT::i32, Legal);
488 // i64 operation support.
489 if (Subtarget->isThumb1Only()) {
490 setOperationAction(ISD::MUL, MVT::i64, Expand);
491 setOperationAction(ISD::MULHU, MVT::i32, Expand);
492 setOperationAction(ISD::MULHS, MVT::i32, Expand);
493 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
494 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
496 setOperationAction(ISD::MUL, MVT::i64, Expand);
497 setOperationAction(ISD::MULHU, MVT::i32, Expand);
498 if (!Subtarget->hasV6Ops())
499 setOperationAction(ISD::MULHS, MVT::i32, Expand);
501 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
502 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
503 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
504 setOperationAction(ISD::SRL, MVT::i64, Custom);
505 setOperationAction(ISD::SRA, MVT::i64, Custom);
507 // ARM does not have ROTL.
508 setOperationAction(ISD::ROTL, MVT::i32, Expand);
509 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
510 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
511 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
512 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
514 // Only ARMv6 has BSWAP.
515 if (!Subtarget->hasV6Ops())
516 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
518 // These are expanded into libcalls.
519 if (!Subtarget->hasDivide()) {
520 // v7M has a hardware divider
521 setOperationAction(ISD::SDIV, MVT::i32, Expand);
522 setOperationAction(ISD::UDIV, MVT::i32, Expand);
524 setOperationAction(ISD::SREM, MVT::i32, Expand);
525 setOperationAction(ISD::UREM, MVT::i32, Expand);
526 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
527 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
529 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
530 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
531 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
532 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
533 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
535 setOperationAction(ISD::TRAP, MVT::Other, Legal);
537 // Use the default implementation.
538 setOperationAction(ISD::VASTART, MVT::Other, Custom);
539 setOperationAction(ISD::VAARG, MVT::Other, Expand);
540 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
541 setOperationAction(ISD::VAEND, MVT::Other, Expand);
542 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
543 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
544 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
545 // FIXME: Shouldn't need this, since no register is used, but the legalizer
546 // doesn't yet know how to not do that for SjLj.
547 setExceptionSelectorRegister(ARM::R0);
548 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
549 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
550 // the default expansion.
551 if (Subtarget->hasDataBarrier() ||
552 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
553 // membarrier needs custom lowering; the rest are legal and handled
555 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
557 // Set them all for expansion, which will force libcalls.
558 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
559 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
560 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
561 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
562 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
563 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
564 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
565 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
566 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
567 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
568 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
569 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
570 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
571 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
572 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
573 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
583 // Since the libcalls include locking, fold in the fences
584 setShouldFoldAtomicFences(true);
586 // 64-bit versions are always libcalls (for now)
587 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
588 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
589 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
590 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
591 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
592 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
593 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
594 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
596 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
597 if (!Subtarget->hasV6Ops()) {
598 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
599 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
601 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
603 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
604 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
605 // iff target supports vfp2.
606 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
607 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
610 // We want to custom lower some of our intrinsics.
611 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
612 if (Subtarget->isTargetDarwin()) {
613 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
614 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
615 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
618 setOperationAction(ISD::SETCC, MVT::i32, Expand);
619 setOperationAction(ISD::SETCC, MVT::f32, Expand);
620 setOperationAction(ISD::SETCC, MVT::f64, Expand);
621 setOperationAction(ISD::SELECT, MVT::i32, Custom);
622 setOperationAction(ISD::SELECT, MVT::f32, Custom);
623 setOperationAction(ISD::SELECT, MVT::f64, Custom);
624 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
625 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
626 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
628 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
629 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
630 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
631 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
632 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
634 // We don't support sin/cos/fmod/copysign/pow
635 setOperationAction(ISD::FSIN, MVT::f64, Expand);
636 setOperationAction(ISD::FSIN, MVT::f32, Expand);
637 setOperationAction(ISD::FCOS, MVT::f32, Expand);
638 setOperationAction(ISD::FCOS, MVT::f64, Expand);
639 setOperationAction(ISD::FREM, MVT::f64, Expand);
640 setOperationAction(ISD::FREM, MVT::f32, Expand);
641 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
642 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
643 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
645 setOperationAction(ISD::FPOW, MVT::f64, Expand);
646 setOperationAction(ISD::FPOW, MVT::f32, Expand);
648 // Various VFP goodness
649 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
650 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
651 if (Subtarget->hasVFP2()) {
652 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
653 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
654 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
655 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
657 // Special handling for half-precision FP.
658 if (!Subtarget->hasFP16()) {
659 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
660 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
664 // We have target-specific dag combine patterns for the following nodes:
665 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
666 setTargetDAGCombine(ISD::ADD);
667 setTargetDAGCombine(ISD::SUB);
668 setTargetDAGCombine(ISD::MUL);
670 if (Subtarget->hasV6T2Ops())
671 setTargetDAGCombine(ISD::OR);
673 setStackPointerRegisterToSaveRestore(ARM::SP);
675 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
676 setSchedulingPreference(Sched::RegPressure);
678 setSchedulingPreference(Sched::Hybrid);
680 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
682 // On ARM arguments smaller than 4 bytes are extended, so all arguments
683 // are at least 4 bytes aligned.
684 setMinStackArgumentAlignment(4);
686 benefitFromCodePlacementOpt = true;
689 std::pair<const TargetRegisterClass*, uint8_t>
690 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
691 const TargetRegisterClass *RRC = 0;
693 switch (VT.getSimpleVT().SimpleTy) {
695 return TargetLowering::findRepresentativeClass(VT);
696 // Use DPR as representative register class for all floating point
697 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
698 // the cost is 1 for both f32 and f64.
699 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
700 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
701 RRC = ARM::DPRRegisterClass;
703 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
704 case MVT::v4f32: case MVT::v2f64:
705 RRC = ARM::DPRRegisterClass;
709 RRC = ARM::DPRRegisterClass;
713 RRC = ARM::DPRRegisterClass;
717 return std::make_pair(RRC, Cost);
720 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
723 case ARMISD::Wrapper: return "ARMISD::Wrapper";
724 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
725 case ARMISD::CALL: return "ARMISD::CALL";
726 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
727 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
728 case ARMISD::tCALL: return "ARMISD::tCALL";
729 case ARMISD::BRCOND: return "ARMISD::BRCOND";
730 case ARMISD::BR_JT: return "ARMISD::BR_JT";
731 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
732 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
733 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
734 case ARMISD::CMP: return "ARMISD::CMP";
735 case ARMISD::CMPZ: return "ARMISD::CMPZ";
736 case ARMISD::CMPFP: return "ARMISD::CMPFP";
737 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
738 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
739 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
740 case ARMISD::CMOV: return "ARMISD::CMOV";
741 case ARMISD::CNEG: return "ARMISD::CNEG";
743 case ARMISD::RBIT: return "ARMISD::RBIT";
745 case ARMISD::FTOSI: return "ARMISD::FTOSI";
746 case ARMISD::FTOUI: return "ARMISD::FTOUI";
747 case ARMISD::SITOF: return "ARMISD::SITOF";
748 case ARMISD::UITOF: return "ARMISD::UITOF";
750 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
751 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
752 case ARMISD::RRX: return "ARMISD::RRX";
754 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
755 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
757 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
758 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
759 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
761 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
763 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
765 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
767 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
768 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
770 case ARMISD::VCEQ: return "ARMISD::VCEQ";
771 case ARMISD::VCGE: return "ARMISD::VCGE";
772 case ARMISD::VCGEU: return "ARMISD::VCGEU";
773 case ARMISD::VCGT: return "ARMISD::VCGT";
774 case ARMISD::VCGTU: return "ARMISD::VCGTU";
775 case ARMISD::VTST: return "ARMISD::VTST";
777 case ARMISD::VSHL: return "ARMISD::VSHL";
778 case ARMISD::VSHRs: return "ARMISD::VSHRs";
779 case ARMISD::VSHRu: return "ARMISD::VSHRu";
780 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
781 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
782 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
783 case ARMISD::VSHRN: return "ARMISD::VSHRN";
784 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
785 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
786 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
787 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
788 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
789 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
790 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
791 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
792 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
793 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
794 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
795 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
796 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
797 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
798 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
799 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
800 case ARMISD::VDUP: return "ARMISD::VDUP";
801 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
802 case ARMISD::VEXT: return "ARMISD::VEXT";
803 case ARMISD::VREV64: return "ARMISD::VREV64";
804 case ARMISD::VREV32: return "ARMISD::VREV32";
805 case ARMISD::VREV16: return "ARMISD::VREV16";
806 case ARMISD::VZIP: return "ARMISD::VZIP";
807 case ARMISD::VUZP: return "ARMISD::VUZP";
808 case ARMISD::VTRN: return "ARMISD::VTRN";
809 case ARMISD::VMULLs: return "ARMISD::VMULLs";
810 case ARMISD::VMULLu: return "ARMISD::VMULLu";
811 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
812 case ARMISD::FMAX: return "ARMISD::FMAX";
813 case ARMISD::FMIN: return "ARMISD::FMIN";
814 case ARMISD::BFI: return "ARMISD::BFI";
818 /// getRegClassFor - Return the register class that should be used for the
819 /// specified value type.
820 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
821 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
822 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
823 // load / store 4 to 8 consecutive D registers.
824 if (Subtarget->hasNEON()) {
825 if (VT == MVT::v4i64)
826 return ARM::QQPRRegisterClass;
827 else if (VT == MVT::v8i64)
828 return ARM::QQQQPRRegisterClass;
830 return TargetLowering::getRegClassFor(VT);
833 // Create a fast isel object.
835 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
836 return ARM::createFastISel(funcInfo);
839 /// getFunctionAlignment - Return the Log2 alignment of this function.
840 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
841 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
844 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
845 /// be used for loads / stores from the global.
846 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
847 return (Subtarget->isThumb1Only() ? 127 : 4095);
850 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
851 unsigned NumVals = N->getNumValues();
853 return Sched::RegPressure;
855 for (unsigned i = 0; i != NumVals; ++i) {
856 EVT VT = N->getValueType(i);
857 if (VT.isFloatingPoint() || VT.isVector())
858 return Sched::Latency;
861 if (!N->isMachineOpcode())
862 return Sched::RegPressure;
864 // Load are scheduled for latency even if there instruction itinerary
866 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
867 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
869 return Sched::Latency;
871 if (!Itins->isEmpty() && Itins->getStageLatency(TID.getSchedClass()) > 2)
872 return Sched::Latency;
873 return Sched::RegPressure;
877 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
878 MachineFunction &MF) const {
879 switch (RC->getID()) {
882 case ARM::tGPRRegClassID:
883 return RegInfo->hasFP(MF) ? 4 : 5;
884 case ARM::GPRRegClassID: {
885 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
886 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
888 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
889 case ARM::DPRRegClassID:
894 //===----------------------------------------------------------------------===//
896 //===----------------------------------------------------------------------===//
898 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
899 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
901 default: llvm_unreachable("Unknown condition code!");
902 case ISD::SETNE: return ARMCC::NE;
903 case ISD::SETEQ: return ARMCC::EQ;
904 case ISD::SETGT: return ARMCC::GT;
905 case ISD::SETGE: return ARMCC::GE;
906 case ISD::SETLT: return ARMCC::LT;
907 case ISD::SETLE: return ARMCC::LE;
908 case ISD::SETUGT: return ARMCC::HI;
909 case ISD::SETUGE: return ARMCC::HS;
910 case ISD::SETULT: return ARMCC::LO;
911 case ISD::SETULE: return ARMCC::LS;
915 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
916 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
917 ARMCC::CondCodes &CondCode2) {
918 CondCode2 = ARMCC::AL;
920 default: llvm_unreachable("Unknown FP condition!");
922 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
924 case ISD::SETOGT: CondCode = ARMCC::GT; break;
926 case ISD::SETOGE: CondCode = ARMCC::GE; break;
927 case ISD::SETOLT: CondCode = ARMCC::MI; break;
928 case ISD::SETOLE: CondCode = ARMCC::LS; break;
929 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
930 case ISD::SETO: CondCode = ARMCC::VC; break;
931 case ISD::SETUO: CondCode = ARMCC::VS; break;
932 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
933 case ISD::SETUGT: CondCode = ARMCC::HI; break;
934 case ISD::SETUGE: CondCode = ARMCC::PL; break;
936 case ISD::SETULT: CondCode = ARMCC::LT; break;
938 case ISD::SETULE: CondCode = ARMCC::LE; break;
940 case ISD::SETUNE: CondCode = ARMCC::NE; break;
944 //===----------------------------------------------------------------------===//
945 // Calling Convention Implementation
946 //===----------------------------------------------------------------------===//
948 #include "ARMGenCallingConv.inc"
950 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
951 /// given CallingConvention value.
952 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
954 bool isVarArg) const {
957 llvm_unreachable("Unsupported calling convention");
958 case CallingConv::Fast:
959 if (Subtarget->hasVFP2() && !isVarArg) {
960 if (!Subtarget->isAAPCS_ABI())
961 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
962 // For AAPCS ABI targets, just use VFP variant of the calling convention.
963 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
966 case CallingConv::C: {
967 // Use target triple & subtarget features to do actual dispatch.
968 if (!Subtarget->isAAPCS_ABI())
969 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
970 else if (Subtarget->hasVFP2() &&
971 FloatABIType == FloatABI::Hard && !isVarArg)
972 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
973 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
975 case CallingConv::ARM_AAPCS_VFP:
976 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
977 case CallingConv::ARM_AAPCS:
978 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
979 case CallingConv::ARM_APCS:
980 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
984 /// LowerCallResult - Lower the result values of a call into the
985 /// appropriate copies out of appropriate physical registers.
987 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
988 CallingConv::ID CallConv, bool isVarArg,
989 const SmallVectorImpl<ISD::InputArg> &Ins,
990 DebugLoc dl, SelectionDAG &DAG,
991 SmallVectorImpl<SDValue> &InVals) const {
993 // Assign locations to each value returned by this call.
994 SmallVector<CCValAssign, 16> RVLocs;
995 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
996 RVLocs, *DAG.getContext());
997 CCInfo.AnalyzeCallResult(Ins,
998 CCAssignFnForNode(CallConv, /* Return*/ true,
1001 // Copy all of the result registers out of their specified physreg.
1002 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1003 CCValAssign VA = RVLocs[i];
1006 if (VA.needsCustom()) {
1007 // Handle f64 or half of a v2f64.
1008 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1010 Chain = Lo.getValue(1);
1011 InFlag = Lo.getValue(2);
1012 VA = RVLocs[++i]; // skip ahead to next loc
1013 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1015 Chain = Hi.getValue(1);
1016 InFlag = Hi.getValue(2);
1017 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1019 if (VA.getLocVT() == MVT::v2f64) {
1020 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1021 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1022 DAG.getConstant(0, MVT::i32));
1024 VA = RVLocs[++i]; // skip ahead to next loc
1025 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1026 Chain = Lo.getValue(1);
1027 InFlag = Lo.getValue(2);
1028 VA = RVLocs[++i]; // skip ahead to next loc
1029 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1030 Chain = Hi.getValue(1);
1031 InFlag = Hi.getValue(2);
1032 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1033 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1034 DAG.getConstant(1, MVT::i32));
1037 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1039 Chain = Val.getValue(1);
1040 InFlag = Val.getValue(2);
1043 switch (VA.getLocInfo()) {
1044 default: llvm_unreachable("Unknown loc info!");
1045 case CCValAssign::Full: break;
1046 case CCValAssign::BCvt:
1047 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1051 InVals.push_back(Val);
1057 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1058 /// by "Src" to address "Dst" of size "Size". Alignment information is
1059 /// specified by the specific parameter attribute. The copy will be passed as
1060 /// a byval function parameter.
1061 /// Sometimes what we are copying is the end of a larger object, the part that
1062 /// does not fit in registers.
1064 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1065 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1067 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1068 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1069 /*isVolatile=*/false, /*AlwaysInline=*/false,
1070 MachinePointerInfo(0), MachinePointerInfo(0));
1073 /// LowerMemOpCallTo - Store the argument to the stack.
1075 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1076 SDValue StackPtr, SDValue Arg,
1077 DebugLoc dl, SelectionDAG &DAG,
1078 const CCValAssign &VA,
1079 ISD::ArgFlagsTy Flags) const {
1080 unsigned LocMemOffset = VA.getLocMemOffset();
1081 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1082 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1083 if (Flags.isByVal())
1084 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1086 return DAG.getStore(Chain, dl, Arg, PtrOff,
1087 MachinePointerInfo::getStack(LocMemOffset),
1091 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1092 SDValue Chain, SDValue &Arg,
1093 RegsToPassVector &RegsToPass,
1094 CCValAssign &VA, CCValAssign &NextVA,
1096 SmallVector<SDValue, 8> &MemOpChains,
1097 ISD::ArgFlagsTy Flags) const {
1099 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1100 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1101 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1103 if (NextVA.isRegLoc())
1104 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1106 assert(NextVA.isMemLoc());
1107 if (StackPtr.getNode() == 0)
1108 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1110 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1116 /// LowerCall - Lowering a call into a callseq_start <-
1117 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1120 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1121 CallingConv::ID CallConv, bool isVarArg,
1123 const SmallVectorImpl<ISD::OutputArg> &Outs,
1124 const SmallVectorImpl<SDValue> &OutVals,
1125 const SmallVectorImpl<ISD::InputArg> &Ins,
1126 DebugLoc dl, SelectionDAG &DAG,
1127 SmallVectorImpl<SDValue> &InVals) const {
1128 MachineFunction &MF = DAG.getMachineFunction();
1129 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1130 bool IsSibCall = false;
1131 // Temporarily disable tail calls so things don't break.
1132 if (!EnableARMTailCalls)
1135 // Check if it's really possible to do a tail call.
1136 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1137 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1138 Outs, OutVals, Ins, DAG);
1139 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1140 // detected sibcalls.
1147 // Analyze operands of the call, assigning locations to each operand.
1148 SmallVector<CCValAssign, 16> ArgLocs;
1149 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1151 CCInfo.AnalyzeCallOperands(Outs,
1152 CCAssignFnForNode(CallConv, /* Return*/ false,
1155 // Get a count of how many bytes are to be pushed on the stack.
1156 unsigned NumBytes = CCInfo.getNextStackOffset();
1158 // For tail calls, memory operands are available in our caller's stack.
1162 // Adjust the stack pointer for the new arguments...
1163 // These operations are automatically eliminated by the prolog/epilog pass
1165 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1167 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1169 RegsToPassVector RegsToPass;
1170 SmallVector<SDValue, 8> MemOpChains;
1172 // Walk the register/memloc assignments, inserting copies/loads. In the case
1173 // of tail call optimization, arguments are handled later.
1174 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1176 ++i, ++realArgIdx) {
1177 CCValAssign &VA = ArgLocs[i];
1178 SDValue Arg = OutVals[realArgIdx];
1179 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1181 // Promote the value if needed.
1182 switch (VA.getLocInfo()) {
1183 default: llvm_unreachable("Unknown loc info!");
1184 case CCValAssign::Full: break;
1185 case CCValAssign::SExt:
1186 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1188 case CCValAssign::ZExt:
1189 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1191 case CCValAssign::AExt:
1192 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1194 case CCValAssign::BCvt:
1195 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1199 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1200 if (VA.needsCustom()) {
1201 if (VA.getLocVT() == MVT::v2f64) {
1202 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1203 DAG.getConstant(0, MVT::i32));
1204 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1205 DAG.getConstant(1, MVT::i32));
1207 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1208 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1210 VA = ArgLocs[++i]; // skip ahead to next loc
1211 if (VA.isRegLoc()) {
1212 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1213 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1215 assert(VA.isMemLoc());
1217 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1218 dl, DAG, VA, Flags));
1221 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1222 StackPtr, MemOpChains, Flags);
1224 } else if (VA.isRegLoc()) {
1225 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1226 } else if (!IsSibCall) {
1227 assert(VA.isMemLoc());
1229 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1230 dl, DAG, VA, Flags));
1234 if (!MemOpChains.empty())
1235 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1236 &MemOpChains[0], MemOpChains.size());
1238 // Build a sequence of copy-to-reg nodes chained together with token chain
1239 // and flag operands which copy the outgoing args into the appropriate regs.
1241 // Tail call byval lowering might overwrite argument registers so in case of
1242 // tail call optimization the copies to registers are lowered later.
1244 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1245 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1246 RegsToPass[i].second, InFlag);
1247 InFlag = Chain.getValue(1);
1250 // For tail calls lower the arguments to the 'real' stack slot.
1252 // Force all the incoming stack arguments to be loaded from the stack
1253 // before any new outgoing arguments are stored to the stack, because the
1254 // outgoing stack slots may alias the incoming argument stack slots, and
1255 // the alias isn't otherwise explicit. This is slightly more conservative
1256 // than necessary, because it means that each store effectively depends
1257 // on every argument instead of just those arguments it would clobber.
1259 // Do not flag preceeding copytoreg stuff together with the following stuff.
1261 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1262 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1263 RegsToPass[i].second, InFlag);
1264 InFlag = Chain.getValue(1);
1269 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1270 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1271 // node so that legalize doesn't hack it.
1272 bool isDirect = false;
1273 bool isARMFunc = false;
1274 bool isLocalARMFunc = false;
1275 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1277 if (EnableARMLongCalls) {
1278 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1279 && "long-calls with non-static relocation model!");
1280 // Handle a global address or an external symbol. If it's not one of
1281 // those, the target's already in a register, so we don't need to do
1283 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1284 const GlobalValue *GV = G->getGlobal();
1285 // Create a constant pool entry for the callee address
1286 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1287 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1290 // Get the address of the callee into a register
1291 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1292 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1293 Callee = DAG.getLoad(getPointerTy(), dl,
1294 DAG.getEntryNode(), CPAddr,
1295 MachinePointerInfo::getConstantPool(),
1297 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1298 const char *Sym = S->getSymbol();
1300 // Create a constant pool entry for the callee address
1301 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1302 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1303 Sym, ARMPCLabelIndex, 0);
1304 // Get the address of the callee into a register
1305 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1306 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1307 Callee = DAG.getLoad(getPointerTy(), dl,
1308 DAG.getEntryNode(), CPAddr,
1309 MachinePointerInfo::getConstantPool(),
1312 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1313 const GlobalValue *GV = G->getGlobal();
1315 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1316 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1317 getTargetMachine().getRelocationModel() != Reloc::Static;
1318 isARMFunc = !Subtarget->isThumb() || isStub;
1319 // ARM call to a local ARM function is predicable.
1320 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1321 // tBX takes a register source operand.
1322 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1323 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1324 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1327 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1328 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1329 Callee = DAG.getLoad(getPointerTy(), dl,
1330 DAG.getEntryNode(), CPAddr,
1331 MachinePointerInfo::getConstantPool(),
1333 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1334 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1335 getPointerTy(), Callee, PICLabel);
1337 // On ELF targets for PIC code, direct calls should go through the PLT
1338 unsigned OpFlags = 0;
1339 if (Subtarget->isTargetELF() &&
1340 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1341 OpFlags = ARMII::MO_PLT;
1342 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1344 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1346 bool isStub = Subtarget->isTargetDarwin() &&
1347 getTargetMachine().getRelocationModel() != Reloc::Static;
1348 isARMFunc = !Subtarget->isThumb() || isStub;
1349 // tBX takes a register source operand.
1350 const char *Sym = S->getSymbol();
1351 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1352 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1353 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1354 Sym, ARMPCLabelIndex, 4);
1355 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1356 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1357 Callee = DAG.getLoad(getPointerTy(), dl,
1358 DAG.getEntryNode(), CPAddr,
1359 MachinePointerInfo::getConstantPool(),
1361 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1362 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1363 getPointerTy(), Callee, PICLabel);
1365 unsigned OpFlags = 0;
1366 // On ELF targets for PIC code, direct calls should go through the PLT
1367 if (Subtarget->isTargetELF() &&
1368 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1369 OpFlags = ARMII::MO_PLT;
1370 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1374 // FIXME: handle tail calls differently.
1376 if (Subtarget->isThumb()) {
1377 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1378 CallOpc = ARMISD::CALL_NOLINK;
1380 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1382 CallOpc = (isDirect || Subtarget->hasV5TOps())
1383 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1384 : ARMISD::CALL_NOLINK;
1387 std::vector<SDValue> Ops;
1388 Ops.push_back(Chain);
1389 Ops.push_back(Callee);
1391 // Add argument registers to the end of the list so that they are known live
1393 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1394 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1395 RegsToPass[i].second.getValueType()));
1397 if (InFlag.getNode())
1398 Ops.push_back(InFlag);
1400 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1402 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1404 // Returns a chain and a flag for retval copy to use.
1405 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1406 InFlag = Chain.getValue(1);
1408 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1409 DAG.getIntPtrConstant(0, true), InFlag);
1411 InFlag = Chain.getValue(1);
1413 // Handle result values, copying them out of physregs into vregs that we
1415 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1419 /// MatchingStackOffset - Return true if the given stack call argument is
1420 /// already available in the same position (relatively) of the caller's
1421 /// incoming argument stack.
1423 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1424 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1425 const ARMInstrInfo *TII) {
1426 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1428 if (Arg.getOpcode() == ISD::CopyFromReg) {
1429 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1430 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1432 MachineInstr *Def = MRI->getVRegDef(VR);
1435 if (!Flags.isByVal()) {
1436 if (!TII->isLoadFromStackSlot(Def, FI))
1441 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1442 if (Flags.isByVal())
1443 // ByVal argument is passed in as a pointer but it's now being
1444 // dereferenced. e.g.
1445 // define @foo(%struct.X* %A) {
1446 // tail call @bar(%struct.X* byval %A)
1449 SDValue Ptr = Ld->getBasePtr();
1450 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1453 FI = FINode->getIndex();
1457 assert(FI != INT_MAX);
1458 if (!MFI->isFixedObjectIndex(FI))
1460 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1463 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1464 /// for tail call optimization. Targets which want to do tail call
1465 /// optimization should implement this function.
1467 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1468 CallingConv::ID CalleeCC,
1470 bool isCalleeStructRet,
1471 bool isCallerStructRet,
1472 const SmallVectorImpl<ISD::OutputArg> &Outs,
1473 const SmallVectorImpl<SDValue> &OutVals,
1474 const SmallVectorImpl<ISD::InputArg> &Ins,
1475 SelectionDAG& DAG) const {
1476 const Function *CallerF = DAG.getMachineFunction().getFunction();
1477 CallingConv::ID CallerCC = CallerF->getCallingConv();
1478 bool CCMatch = CallerCC == CalleeCC;
1480 // Look for obvious safe cases to perform tail call optimization that do not
1481 // require ABI changes. This is what gcc calls sibcall.
1483 // Do not sibcall optimize vararg calls unless the call site is not passing
1485 if (isVarArg && !Outs.empty())
1488 // Also avoid sibcall optimization if either caller or callee uses struct
1489 // return semantics.
1490 if (isCalleeStructRet || isCallerStructRet)
1493 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1494 // emitEpilogue is not ready for them.
1495 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1496 // LR. This means if we need to reload LR, it takes an extra instructions,
1497 // which outweighs the value of the tail call; but here we don't know yet
1498 // whether LR is going to be used. Probably the right approach is to
1499 // generate the tail call here and turn it back into CALL/RET in
1500 // emitEpilogue if LR is used.
1501 if (Subtarget->isThumb1Only())
1504 // For the moment, we can only do this to functions defined in this
1505 // compilation, or to indirect calls. A Thumb B to an ARM function,
1506 // or vice versa, is not easily fixed up in the linker unlike BL.
1507 // (We could do this by loading the address of the callee into a register;
1508 // that is an extra instruction over the direct call and burns a register
1509 // as well, so is not likely to be a win.)
1511 // It might be safe to remove this restriction on non-Darwin.
1513 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1514 // but we need to make sure there are enough registers; the only valid
1515 // registers are the 4 used for parameters. We don't currently do this
1517 if (isa<ExternalSymbolSDNode>(Callee))
1520 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1521 const GlobalValue *GV = G->getGlobal();
1522 if (GV->isDeclaration() || GV->isWeakForLinker())
1526 // If the calling conventions do not match, then we'd better make sure the
1527 // results are returned in the same way as what the caller expects.
1529 SmallVector<CCValAssign, 16> RVLocs1;
1530 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1531 RVLocs1, *DAG.getContext());
1532 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1534 SmallVector<CCValAssign, 16> RVLocs2;
1535 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1536 RVLocs2, *DAG.getContext());
1537 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1539 if (RVLocs1.size() != RVLocs2.size())
1541 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1542 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1544 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1546 if (RVLocs1[i].isRegLoc()) {
1547 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1550 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1556 // If the callee takes no arguments then go on to check the results of the
1558 if (!Outs.empty()) {
1559 // Check if stack adjustment is needed. For now, do not do this if any
1560 // argument is passed on the stack.
1561 SmallVector<CCValAssign, 16> ArgLocs;
1562 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1563 ArgLocs, *DAG.getContext());
1564 CCInfo.AnalyzeCallOperands(Outs,
1565 CCAssignFnForNode(CalleeCC, false, isVarArg));
1566 if (CCInfo.getNextStackOffset()) {
1567 MachineFunction &MF = DAG.getMachineFunction();
1569 // Check if the arguments are already laid out in the right way as
1570 // the caller's fixed stack objects.
1571 MachineFrameInfo *MFI = MF.getFrameInfo();
1572 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1573 const ARMInstrInfo *TII =
1574 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1575 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1577 ++i, ++realArgIdx) {
1578 CCValAssign &VA = ArgLocs[i];
1579 EVT RegVT = VA.getLocVT();
1580 SDValue Arg = OutVals[realArgIdx];
1581 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1582 if (VA.getLocInfo() == CCValAssign::Indirect)
1584 if (VA.needsCustom()) {
1585 // f64 and vector types are split into multiple registers or
1586 // register/stack-slot combinations. The types will not match
1587 // the registers; give up on memory f64 refs until we figure
1588 // out what to do about this.
1591 if (!ArgLocs[++i].isRegLoc())
1593 if (RegVT == MVT::v2f64) {
1594 if (!ArgLocs[++i].isRegLoc())
1596 if (!ArgLocs[++i].isRegLoc())
1599 } else if (!VA.isRegLoc()) {
1600 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1612 ARMTargetLowering::LowerReturn(SDValue Chain,
1613 CallingConv::ID CallConv, bool isVarArg,
1614 const SmallVectorImpl<ISD::OutputArg> &Outs,
1615 const SmallVectorImpl<SDValue> &OutVals,
1616 DebugLoc dl, SelectionDAG &DAG) const {
1618 // CCValAssign - represent the assignment of the return value to a location.
1619 SmallVector<CCValAssign, 16> RVLocs;
1621 // CCState - Info about the registers and stack slots.
1622 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1625 // Analyze outgoing return values.
1626 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1629 // If this is the first return lowered for this function, add
1630 // the regs to the liveout set for the function.
1631 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1632 for (unsigned i = 0; i != RVLocs.size(); ++i)
1633 if (RVLocs[i].isRegLoc())
1634 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1639 // Copy the result values into the output registers.
1640 for (unsigned i = 0, realRVLocIdx = 0;
1642 ++i, ++realRVLocIdx) {
1643 CCValAssign &VA = RVLocs[i];
1644 assert(VA.isRegLoc() && "Can only return in registers!");
1646 SDValue Arg = OutVals[realRVLocIdx];
1648 switch (VA.getLocInfo()) {
1649 default: llvm_unreachable("Unknown loc info!");
1650 case CCValAssign::Full: break;
1651 case CCValAssign::BCvt:
1652 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1656 if (VA.needsCustom()) {
1657 if (VA.getLocVT() == MVT::v2f64) {
1658 // Extract the first half and return it in two registers.
1659 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1660 DAG.getConstant(0, MVT::i32));
1661 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1662 DAG.getVTList(MVT::i32, MVT::i32), Half);
1664 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1665 Flag = Chain.getValue(1);
1666 VA = RVLocs[++i]; // skip ahead to next loc
1667 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1668 HalfGPRs.getValue(1), Flag);
1669 Flag = Chain.getValue(1);
1670 VA = RVLocs[++i]; // skip ahead to next loc
1672 // Extract the 2nd half and fall through to handle it as an f64 value.
1673 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1674 DAG.getConstant(1, MVT::i32));
1676 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1678 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1679 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1680 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1681 Flag = Chain.getValue(1);
1682 VA = RVLocs[++i]; // skip ahead to next loc
1683 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1686 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1688 // Guarantee that all emitted copies are
1689 // stuck together, avoiding something bad.
1690 Flag = Chain.getValue(1);
1695 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1697 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1702 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1703 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1704 // one of the above mentioned nodes. It has to be wrapped because otherwise
1705 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1706 // be used to form addressing mode. These wrapped nodes will be selected
1708 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1709 EVT PtrVT = Op.getValueType();
1710 // FIXME there is no actual debug info here
1711 DebugLoc dl = Op.getDebugLoc();
1712 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1714 if (CP->isMachineConstantPoolEntry())
1715 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1716 CP->getAlignment());
1718 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1719 CP->getAlignment());
1720 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1723 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1724 return MachineJumpTableInfo::EK_Inline;
1727 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1728 SelectionDAG &DAG) const {
1729 MachineFunction &MF = DAG.getMachineFunction();
1730 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1731 unsigned ARMPCLabelIndex = 0;
1732 DebugLoc DL = Op.getDebugLoc();
1733 EVT PtrVT = getPointerTy();
1734 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1735 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1737 if (RelocM == Reloc::Static) {
1738 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1740 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1741 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1742 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1743 ARMCP::CPBlockAddress,
1745 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1747 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1748 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1749 MachinePointerInfo::getConstantPool(),
1751 if (RelocM == Reloc::Static)
1753 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1754 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1757 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1759 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1760 SelectionDAG &DAG) const {
1761 DebugLoc dl = GA->getDebugLoc();
1762 EVT PtrVT = getPointerTy();
1763 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1764 MachineFunction &MF = DAG.getMachineFunction();
1765 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1766 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1767 ARMConstantPoolValue *CPV =
1768 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1769 ARMCP::CPValue, PCAdj, "tlsgd", true);
1770 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1771 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1772 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1773 MachinePointerInfo::getConstantPool(),
1775 SDValue Chain = Argument.getValue(1);
1777 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1778 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1780 // call __tls_get_addr.
1783 Entry.Node = Argument;
1784 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1785 Args.push_back(Entry);
1786 // FIXME: is there useful debug info available here?
1787 std::pair<SDValue, SDValue> CallResult =
1788 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1789 false, false, false, false,
1790 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1791 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1792 return CallResult.first;
1795 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1796 // "local exec" model.
1798 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1799 SelectionDAG &DAG) const {
1800 const GlobalValue *GV = GA->getGlobal();
1801 DebugLoc dl = GA->getDebugLoc();
1803 SDValue Chain = DAG.getEntryNode();
1804 EVT PtrVT = getPointerTy();
1805 // Get the Thread Pointer
1806 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1808 if (GV->isDeclaration()) {
1809 MachineFunction &MF = DAG.getMachineFunction();
1810 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1811 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1812 // Initial exec model.
1813 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1814 ARMConstantPoolValue *CPV =
1815 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1816 ARMCP::CPValue, PCAdj, "gottpoff", true);
1817 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1818 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1819 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1820 MachinePointerInfo::getConstantPool(),
1822 Chain = Offset.getValue(1);
1824 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1825 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1827 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1828 MachinePointerInfo::getConstantPool(),
1832 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1833 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1834 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1835 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1836 MachinePointerInfo::getConstantPool(),
1840 // The address of the thread local variable is the add of the thread
1841 // pointer with the offset of the variable.
1842 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1846 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1847 // TODO: implement the "local dynamic" model
1848 assert(Subtarget->isTargetELF() &&
1849 "TLS not implemented for non-ELF targets");
1850 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1851 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1852 // otherwise use the "Local Exec" TLS Model
1853 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1854 return LowerToTLSGeneralDynamicModel(GA, DAG);
1856 return LowerToTLSExecModels(GA, DAG);
1859 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1860 SelectionDAG &DAG) const {
1861 EVT PtrVT = getPointerTy();
1862 DebugLoc dl = Op.getDebugLoc();
1863 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1864 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1865 if (RelocM == Reloc::PIC_) {
1866 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1867 ARMConstantPoolValue *CPV =
1868 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1869 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1870 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1871 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1873 MachinePointerInfo::getConstantPool(),
1875 SDValue Chain = Result.getValue(1);
1876 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1877 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1879 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1880 MachinePointerInfo::getGOT(), false, false, 0);
1883 // If we have T2 ops, we can materialize the address directly via movt/movw
1884 // pair. This is always cheaper.
1885 if (Subtarget->useMovt()) {
1886 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1887 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1889 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1890 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1891 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1892 MachinePointerInfo::getConstantPool(),
1898 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1899 SelectionDAG &DAG) const {
1900 MachineFunction &MF = DAG.getMachineFunction();
1901 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1902 unsigned ARMPCLabelIndex = 0;
1903 EVT PtrVT = getPointerTy();
1904 DebugLoc dl = Op.getDebugLoc();
1905 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1906 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1908 if (RelocM == Reloc::Static)
1909 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1911 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1912 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1913 ARMConstantPoolValue *CPV =
1914 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1915 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1917 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1919 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1920 MachinePointerInfo::getConstantPool(),
1922 SDValue Chain = Result.getValue(1);
1924 if (RelocM == Reloc::PIC_) {
1925 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1926 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1929 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1930 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1936 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1937 SelectionDAG &DAG) const {
1938 assert(Subtarget->isTargetELF() &&
1939 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1940 MachineFunction &MF = DAG.getMachineFunction();
1941 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1942 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1943 EVT PtrVT = getPointerTy();
1944 DebugLoc dl = Op.getDebugLoc();
1945 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1946 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1947 "_GLOBAL_OFFSET_TABLE_",
1948 ARMPCLabelIndex, PCAdj);
1949 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1950 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1951 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1952 MachinePointerInfo::getConstantPool(),
1954 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1955 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1959 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
1961 DebugLoc dl = Op.getDebugLoc();
1962 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
1963 Op.getOperand(0), Op.getOperand(1));
1967 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1968 DebugLoc dl = Op.getDebugLoc();
1969 SDValue Val = DAG.getConstant(0, MVT::i32);
1970 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1971 Op.getOperand(1), Val);
1975 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1976 DebugLoc dl = Op.getDebugLoc();
1977 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1978 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1982 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1983 const ARMSubtarget *Subtarget) const {
1984 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1985 DebugLoc dl = Op.getDebugLoc();
1987 default: return SDValue(); // Don't custom lower most intrinsics.
1988 case Intrinsic::arm_thread_pointer: {
1989 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1990 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1992 case Intrinsic::eh_sjlj_lsda: {
1993 MachineFunction &MF = DAG.getMachineFunction();
1994 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1995 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1996 EVT PtrVT = getPointerTy();
1997 DebugLoc dl = Op.getDebugLoc();
1998 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2000 unsigned PCAdj = (RelocM != Reloc::PIC_)
2001 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2002 ARMConstantPoolValue *CPV =
2003 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2004 ARMCP::CPLSDA, PCAdj);
2005 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2006 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2008 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2009 MachinePointerInfo::getConstantPool(),
2012 if (RelocM == Reloc::PIC_) {
2013 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2014 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2021 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2022 const ARMSubtarget *Subtarget) {
2023 DebugLoc dl = Op.getDebugLoc();
2024 SDValue Op5 = Op.getOperand(5);
2025 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
2026 // Some subtargets which have dmb and dsb instructions can handle barriers
2027 // directly. Some ARMv6 cpus can support them with the help of mcr
2028 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
2030 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
2031 if (Subtarget->hasDataBarrier())
2032 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2034 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2035 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2036 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2037 DAG.getConstant(0, MVT::i32));
2041 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2042 MachineFunction &MF = DAG.getMachineFunction();
2043 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2045 // vastart just stores the address of the VarArgsFrameIndex slot into the
2046 // memory location argument.
2047 DebugLoc dl = Op.getDebugLoc();
2048 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2049 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2050 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2051 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2052 MachinePointerInfo(SV), false, false, 0);
2056 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2057 SDValue &Root, SelectionDAG &DAG,
2058 DebugLoc dl) const {
2059 MachineFunction &MF = DAG.getMachineFunction();
2060 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2062 TargetRegisterClass *RC;
2063 if (AFI->isThumb1OnlyFunction())
2064 RC = ARM::tGPRRegisterClass;
2066 RC = ARM::GPRRegisterClass;
2068 // Transform the arguments stored in physical registers into virtual ones.
2069 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2070 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2073 if (NextVA.isMemLoc()) {
2074 MachineFrameInfo *MFI = MF.getFrameInfo();
2075 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2077 // Create load node to retrieve arguments from the stack.
2078 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2079 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2080 MachinePointerInfo::getFixedStack(FI),
2083 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2084 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2087 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2091 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2092 CallingConv::ID CallConv, bool isVarArg,
2093 const SmallVectorImpl<ISD::InputArg>
2095 DebugLoc dl, SelectionDAG &DAG,
2096 SmallVectorImpl<SDValue> &InVals)
2099 MachineFunction &MF = DAG.getMachineFunction();
2100 MachineFrameInfo *MFI = MF.getFrameInfo();
2102 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2104 // Assign locations to all of the incoming arguments.
2105 SmallVector<CCValAssign, 16> ArgLocs;
2106 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2108 CCInfo.AnalyzeFormalArguments(Ins,
2109 CCAssignFnForNode(CallConv, /* Return*/ false,
2112 SmallVector<SDValue, 16> ArgValues;
2114 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2115 CCValAssign &VA = ArgLocs[i];
2117 // Arguments stored in registers.
2118 if (VA.isRegLoc()) {
2119 EVT RegVT = VA.getLocVT();
2122 if (VA.needsCustom()) {
2123 // f64 and vector types are split up into multiple registers or
2124 // combinations of registers and stack slots.
2125 if (VA.getLocVT() == MVT::v2f64) {
2126 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2128 VA = ArgLocs[++i]; // skip ahead to next loc
2130 if (VA.isMemLoc()) {
2131 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2132 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2133 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2134 MachinePointerInfo::getFixedStack(FI),
2137 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2140 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2141 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2142 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2143 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2144 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2146 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2149 TargetRegisterClass *RC;
2151 if (RegVT == MVT::f32)
2152 RC = ARM::SPRRegisterClass;
2153 else if (RegVT == MVT::f64)
2154 RC = ARM::DPRRegisterClass;
2155 else if (RegVT == MVT::v2f64)
2156 RC = ARM::QPRRegisterClass;
2157 else if (RegVT == MVT::i32)
2158 RC = (AFI->isThumb1OnlyFunction() ?
2159 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2161 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2163 // Transform the arguments in physical registers into virtual ones.
2164 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2165 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2168 // If this is an 8 or 16-bit value, it is really passed promoted
2169 // to 32 bits. Insert an assert[sz]ext to capture this, then
2170 // truncate to the right size.
2171 switch (VA.getLocInfo()) {
2172 default: llvm_unreachable("Unknown loc info!");
2173 case CCValAssign::Full: break;
2174 case CCValAssign::BCvt:
2175 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2177 case CCValAssign::SExt:
2178 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2179 DAG.getValueType(VA.getValVT()));
2180 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2182 case CCValAssign::ZExt:
2183 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2184 DAG.getValueType(VA.getValVT()));
2185 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2189 InVals.push_back(ArgValue);
2191 } else { // VA.isRegLoc()
2194 assert(VA.isMemLoc());
2195 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2197 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2198 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2200 // Create load nodes to retrieve arguments from the stack.
2201 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2202 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2203 MachinePointerInfo::getFixedStack(FI),
2210 static const unsigned GPRArgRegs[] = {
2211 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2214 unsigned NumGPRs = CCInfo.getFirstUnallocated
2215 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2217 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2218 unsigned VARegSize = (4 - NumGPRs) * 4;
2219 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2220 unsigned ArgOffset = CCInfo.getNextStackOffset();
2221 if (VARegSaveSize) {
2222 // If this function is vararg, store any remaining integer argument regs
2223 // to their spots on the stack so that they may be loaded by deferencing
2224 // the result of va_next.
2225 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2226 AFI->setVarArgsFrameIndex(
2227 MFI->CreateFixedObject(VARegSaveSize,
2228 ArgOffset + VARegSaveSize - VARegSize,
2230 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2233 SmallVector<SDValue, 4> MemOps;
2234 for (; NumGPRs < 4; ++NumGPRs) {
2235 TargetRegisterClass *RC;
2236 if (AFI->isThumb1OnlyFunction())
2237 RC = ARM::tGPRRegisterClass;
2239 RC = ARM::GPRRegisterClass;
2241 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2242 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2244 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2245 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2247 MemOps.push_back(Store);
2248 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2249 DAG.getConstant(4, getPointerTy()));
2251 if (!MemOps.empty())
2252 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2253 &MemOps[0], MemOps.size());
2255 // This will point to the next argument passed via stack.
2256 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2262 /// isFloatingPointZero - Return true if this is +0.0.
2263 static bool isFloatingPointZero(SDValue Op) {
2264 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2265 return CFP->getValueAPF().isPosZero();
2266 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2267 // Maybe this has already been legalized into the constant pool?
2268 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2269 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2270 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2271 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2272 return CFP->getValueAPF().isPosZero();
2278 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2279 /// the given operands.
2281 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2282 SDValue &ARMcc, SelectionDAG &DAG,
2283 DebugLoc dl) const {
2284 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2285 unsigned C = RHSC->getZExtValue();
2286 if (!isLegalICmpImmediate(C)) {
2287 // Constant does not fit, try adjusting it by one?
2292 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2293 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2294 RHS = DAG.getConstant(C-1, MVT::i32);
2299 if (C != 0 && isLegalICmpImmediate(C-1)) {
2300 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2301 RHS = DAG.getConstant(C-1, MVT::i32);
2306 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2307 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2308 RHS = DAG.getConstant(C+1, MVT::i32);
2313 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2314 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2315 RHS = DAG.getConstant(C+1, MVT::i32);
2322 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2323 ARMISD::NodeType CompareType;
2326 CompareType = ARMISD::CMP;
2331 CompareType = ARMISD::CMPZ;
2334 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2335 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2338 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2340 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2341 DebugLoc dl) const {
2343 if (!isFloatingPointZero(RHS))
2344 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2346 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2347 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2350 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2351 SDValue Cond = Op.getOperand(0);
2352 SDValue SelectTrue = Op.getOperand(1);
2353 SDValue SelectFalse = Op.getOperand(2);
2354 DebugLoc dl = Op.getDebugLoc();
2358 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2359 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2361 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2362 const ConstantSDNode *CMOVTrue =
2363 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2364 const ConstantSDNode *CMOVFalse =
2365 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2367 if (CMOVTrue && CMOVFalse) {
2368 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2369 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2373 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2375 False = SelectFalse;
2376 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2381 if (True.getNode() && False.getNode()) {
2382 EVT VT = Cond.getValueType();
2383 SDValue ARMcc = Cond.getOperand(2);
2384 SDValue CCR = Cond.getOperand(3);
2385 SDValue Cmp = Cond.getOperand(4);
2386 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2391 return DAG.getSelectCC(dl, Cond,
2392 DAG.getConstant(0, Cond.getValueType()),
2393 SelectTrue, SelectFalse, ISD::SETNE);
2396 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2397 EVT VT = Op.getValueType();
2398 SDValue LHS = Op.getOperand(0);
2399 SDValue RHS = Op.getOperand(1);
2400 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2401 SDValue TrueVal = Op.getOperand(2);
2402 SDValue FalseVal = Op.getOperand(3);
2403 DebugLoc dl = Op.getDebugLoc();
2405 if (LHS.getValueType() == MVT::i32) {
2407 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2408 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2409 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2412 ARMCC::CondCodes CondCode, CondCode2;
2413 FPCCToARMCC(CC, CondCode, CondCode2);
2415 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2416 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2417 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2418 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2420 if (CondCode2 != ARMCC::AL) {
2421 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2422 // FIXME: Needs another CMP because flag can have but one use.
2423 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2424 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2425 Result, TrueVal, ARMcc2, CCR, Cmp2);
2430 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2431 /// to morph to an integer compare sequence.
2432 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2433 const ARMSubtarget *Subtarget) {
2434 SDNode *N = Op.getNode();
2435 if (!N->hasOneUse())
2436 // Otherwise it requires moving the value from fp to integer registers.
2438 if (!N->getNumValues())
2440 EVT VT = Op.getValueType();
2441 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2442 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2443 // vmrs are very slow, e.g. cortex-a8.
2446 if (isFloatingPointZero(Op)) {
2450 return ISD::isNormalLoad(N);
2453 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2454 if (isFloatingPointZero(Op))
2455 return DAG.getConstant(0, MVT::i32);
2457 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2458 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2459 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2460 Ld->isVolatile(), Ld->isNonTemporal(),
2461 Ld->getAlignment());
2463 llvm_unreachable("Unknown VFP cmp argument!");
2466 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2467 SDValue &RetVal1, SDValue &RetVal2) {
2468 if (isFloatingPointZero(Op)) {
2469 RetVal1 = DAG.getConstant(0, MVT::i32);
2470 RetVal2 = DAG.getConstant(0, MVT::i32);
2474 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2475 SDValue Ptr = Ld->getBasePtr();
2476 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2477 Ld->getChain(), Ptr,
2478 Ld->getPointerInfo(),
2479 Ld->isVolatile(), Ld->isNonTemporal(),
2480 Ld->getAlignment());
2482 EVT PtrType = Ptr.getValueType();
2483 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2484 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2485 PtrType, Ptr, DAG.getConstant(4, PtrType));
2486 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2487 Ld->getChain(), NewPtr,
2488 Ld->getPointerInfo().getWithOffset(4),
2489 Ld->isVolatile(), Ld->isNonTemporal(),
2494 llvm_unreachable("Unknown VFP cmp argument!");
2497 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2498 /// f32 and even f64 comparisons to integer ones.
2500 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2501 SDValue Chain = Op.getOperand(0);
2502 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2503 SDValue LHS = Op.getOperand(2);
2504 SDValue RHS = Op.getOperand(3);
2505 SDValue Dest = Op.getOperand(4);
2506 DebugLoc dl = Op.getDebugLoc();
2508 bool SeenZero = false;
2509 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2510 canChangeToInt(RHS, SeenZero, Subtarget) &&
2511 // If one of the operand is zero, it's safe to ignore the NaN case since
2512 // we only care about equality comparisons.
2513 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2514 // If unsafe fp math optimization is enabled and there are no othter uses of
2515 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2516 // to an integer comparison.
2517 if (CC == ISD::SETOEQ)
2519 else if (CC == ISD::SETUNE)
2523 if (LHS.getValueType() == MVT::f32) {
2524 LHS = bitcastf32Toi32(LHS, DAG);
2525 RHS = bitcastf32Toi32(RHS, DAG);
2526 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2527 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2528 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2529 Chain, Dest, ARMcc, CCR, Cmp);
2534 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2535 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2536 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2537 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2538 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2539 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2540 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2546 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2547 SDValue Chain = Op.getOperand(0);
2548 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2549 SDValue LHS = Op.getOperand(2);
2550 SDValue RHS = Op.getOperand(3);
2551 SDValue Dest = Op.getOperand(4);
2552 DebugLoc dl = Op.getDebugLoc();
2554 if (LHS.getValueType() == MVT::i32) {
2556 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2557 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2558 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2559 Chain, Dest, ARMcc, CCR, Cmp);
2562 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2565 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2566 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2567 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2568 if (Result.getNode())
2572 ARMCC::CondCodes CondCode, CondCode2;
2573 FPCCToARMCC(CC, CondCode, CondCode2);
2575 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2576 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2577 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2578 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2579 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2580 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2581 if (CondCode2 != ARMCC::AL) {
2582 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2583 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2584 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2589 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2590 SDValue Chain = Op.getOperand(0);
2591 SDValue Table = Op.getOperand(1);
2592 SDValue Index = Op.getOperand(2);
2593 DebugLoc dl = Op.getDebugLoc();
2595 EVT PTy = getPointerTy();
2596 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2597 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2598 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2599 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2600 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2601 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2602 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2603 if (Subtarget->isThumb2()) {
2604 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2605 // which does another jump to the destination. This also makes it easier
2606 // to translate it to TBB / TBH later.
2607 // FIXME: This might not work if the function is extremely large.
2608 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2609 Addr, Op.getOperand(2), JTI, UId);
2611 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2612 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2613 MachinePointerInfo::getJumpTable(),
2615 Chain = Addr.getValue(1);
2616 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2617 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2619 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2620 MachinePointerInfo::getJumpTable(), false, false, 0);
2621 Chain = Addr.getValue(1);
2622 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2626 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2627 DebugLoc dl = Op.getDebugLoc();
2630 switch (Op.getOpcode()) {
2632 assert(0 && "Invalid opcode!");
2633 case ISD::FP_TO_SINT:
2634 Opc = ARMISD::FTOSI;
2636 case ISD::FP_TO_UINT:
2637 Opc = ARMISD::FTOUI;
2640 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2641 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2644 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2645 EVT VT = Op.getValueType();
2646 DebugLoc dl = Op.getDebugLoc();
2649 switch (Op.getOpcode()) {
2651 assert(0 && "Invalid opcode!");
2652 case ISD::SINT_TO_FP:
2653 Opc = ARMISD::SITOF;
2655 case ISD::UINT_TO_FP:
2656 Opc = ARMISD::UITOF;
2660 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2661 return DAG.getNode(Opc, dl, VT, Op);
2664 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2665 // Implement fcopysign with a fabs and a conditional fneg.
2666 SDValue Tmp0 = Op.getOperand(0);
2667 SDValue Tmp1 = Op.getOperand(1);
2668 DebugLoc dl = Op.getDebugLoc();
2669 EVT VT = Op.getValueType();
2670 EVT SrcVT = Tmp1.getValueType();
2671 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2672 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2673 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2674 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2675 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2676 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2679 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2680 MachineFunction &MF = DAG.getMachineFunction();
2681 MachineFrameInfo *MFI = MF.getFrameInfo();
2682 MFI->setReturnAddressIsTaken(true);
2684 EVT VT = Op.getValueType();
2685 DebugLoc dl = Op.getDebugLoc();
2686 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2688 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2689 SDValue Offset = DAG.getConstant(4, MVT::i32);
2690 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2691 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2692 MachinePointerInfo(), false, false, 0);
2695 // Return LR, which contains the return address. Mark it an implicit live-in.
2696 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2697 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2700 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2701 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2702 MFI->setFrameAddressIsTaken(true);
2704 EVT VT = Op.getValueType();
2705 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2706 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2707 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2708 ? ARM::R7 : ARM::R11;
2709 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2711 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2712 MachinePointerInfo(),
2717 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2718 /// expand a bit convert where either the source or destination type is i64 to
2719 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2720 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2721 /// vectors), since the legalizer won't know what to do with that.
2722 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2723 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2724 DebugLoc dl = N->getDebugLoc();
2725 SDValue Op = N->getOperand(0);
2727 // This function is only supposed to be called for i64 types, either as the
2728 // source or destination of the bit convert.
2729 EVT SrcVT = Op.getValueType();
2730 EVT DstVT = N->getValueType(0);
2731 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2732 "ExpandBIT_CONVERT called for non-i64 type");
2734 // Turn i64->f64 into VMOVDRR.
2735 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2736 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2737 DAG.getConstant(0, MVT::i32));
2738 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2739 DAG.getConstant(1, MVT::i32));
2740 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2741 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2744 // Turn f64->i64 into VMOVRRD.
2745 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2746 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2747 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2748 // Merge the pieces into a single i64 value.
2749 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2755 /// getZeroVector - Returns a vector of specified type with all zero elements.
2756 /// Zero vectors are used to represent vector negation and in those cases
2757 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2758 /// not support i64 elements, so sometimes the zero vectors will need to be
2759 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2761 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2762 assert(VT.isVector() && "Expected a vector type");
2763 // The canonical modified immediate encoding of a zero vector is....0!
2764 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2765 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2766 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2767 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2770 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2771 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2772 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2773 SelectionDAG &DAG) const {
2774 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2775 EVT VT = Op.getValueType();
2776 unsigned VTBits = VT.getSizeInBits();
2777 DebugLoc dl = Op.getDebugLoc();
2778 SDValue ShOpLo = Op.getOperand(0);
2779 SDValue ShOpHi = Op.getOperand(1);
2780 SDValue ShAmt = Op.getOperand(2);
2782 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2784 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2786 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2787 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2788 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2789 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2790 DAG.getConstant(VTBits, MVT::i32));
2791 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2792 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2793 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2795 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2796 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2798 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2799 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2802 SDValue Ops[2] = { Lo, Hi };
2803 return DAG.getMergeValues(Ops, 2, dl);
2806 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2807 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2808 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2809 SelectionDAG &DAG) const {
2810 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2811 EVT VT = Op.getValueType();
2812 unsigned VTBits = VT.getSizeInBits();
2813 DebugLoc dl = Op.getDebugLoc();
2814 SDValue ShOpLo = Op.getOperand(0);
2815 SDValue ShOpHi = Op.getOperand(1);
2816 SDValue ShAmt = Op.getOperand(2);
2819 assert(Op.getOpcode() == ISD::SHL_PARTS);
2820 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2821 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2822 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2823 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2824 DAG.getConstant(VTBits, MVT::i32));
2825 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2826 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2828 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2829 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2830 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2832 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2833 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2836 SDValue Ops[2] = { Lo, Hi };
2837 return DAG.getMergeValues(Ops, 2, dl);
2840 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2841 SelectionDAG &DAG) const {
2842 // The rounding mode is in bits 23:22 of the FPSCR.
2843 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2844 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2845 // so that the shift + and get folded into a bitfield extract.
2846 DebugLoc dl = Op.getDebugLoc();
2847 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2848 DAG.getConstant(Intrinsic::arm_get_fpscr,
2850 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2851 DAG.getConstant(1U << 22, MVT::i32));
2852 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2853 DAG.getConstant(22, MVT::i32));
2854 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2855 DAG.getConstant(3, MVT::i32));
2858 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2859 const ARMSubtarget *ST) {
2860 EVT VT = N->getValueType(0);
2861 DebugLoc dl = N->getDebugLoc();
2863 if (!ST->hasV6T2Ops())
2866 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2867 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2870 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2871 const ARMSubtarget *ST) {
2872 EVT VT = N->getValueType(0);
2873 DebugLoc dl = N->getDebugLoc();
2875 // Lower vector shifts on NEON to use VSHL.
2876 if (VT.isVector()) {
2877 assert(ST->hasNEON() && "unexpected vector shift");
2879 // Left shifts translate directly to the vshiftu intrinsic.
2880 if (N->getOpcode() == ISD::SHL)
2881 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2882 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2883 N->getOperand(0), N->getOperand(1));
2885 assert((N->getOpcode() == ISD::SRA ||
2886 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2888 // NEON uses the same intrinsics for both left and right shifts. For
2889 // right shifts, the shift amounts are negative, so negate the vector of
2891 EVT ShiftVT = N->getOperand(1).getValueType();
2892 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2893 getZeroVector(ShiftVT, DAG, dl),
2895 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2896 Intrinsic::arm_neon_vshifts :
2897 Intrinsic::arm_neon_vshiftu);
2898 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2899 DAG.getConstant(vshiftInt, MVT::i32),
2900 N->getOperand(0), NegatedCount);
2903 // We can get here for a node like i32 = ISD::SHL i32, i64
2907 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2908 "Unknown shift to lower!");
2910 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2911 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2912 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2915 // If we are in thumb mode, we don't have RRX.
2916 if (ST->isThumb1Only()) return SDValue();
2918 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2919 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2920 DAG.getConstant(0, MVT::i32));
2921 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2922 DAG.getConstant(1, MVT::i32));
2924 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2925 // captures the result into a carry flag.
2926 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2927 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2929 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2930 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2932 // Merge the pieces into a single i64 value.
2933 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2936 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2937 SDValue TmpOp0, TmpOp1;
2938 bool Invert = false;
2942 SDValue Op0 = Op.getOperand(0);
2943 SDValue Op1 = Op.getOperand(1);
2944 SDValue CC = Op.getOperand(2);
2945 EVT VT = Op.getValueType();
2946 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2947 DebugLoc dl = Op.getDebugLoc();
2949 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2950 switch (SetCCOpcode) {
2951 default: llvm_unreachable("Illegal FP comparison"); break;
2953 case ISD::SETNE: Invert = true; // Fallthrough
2955 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2957 case ISD::SETLT: Swap = true; // Fallthrough
2959 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2961 case ISD::SETLE: Swap = true; // Fallthrough
2963 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2964 case ISD::SETUGE: Swap = true; // Fallthrough
2965 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2966 case ISD::SETUGT: Swap = true; // Fallthrough
2967 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2968 case ISD::SETUEQ: Invert = true; // Fallthrough
2970 // Expand this to (OLT | OGT).
2974 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2975 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2977 case ISD::SETUO: Invert = true; // Fallthrough
2979 // Expand this to (OLT | OGE).
2983 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2984 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2988 // Integer comparisons.
2989 switch (SetCCOpcode) {
2990 default: llvm_unreachable("Illegal integer comparison"); break;
2991 case ISD::SETNE: Invert = true;
2992 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2993 case ISD::SETLT: Swap = true;
2994 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2995 case ISD::SETLE: Swap = true;
2996 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2997 case ISD::SETULT: Swap = true;
2998 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2999 case ISD::SETULE: Swap = true;
3000 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3003 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3004 if (Opc == ARMISD::VCEQ) {
3007 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3009 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3012 // Ignore bitconvert.
3013 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
3014 AndOp = AndOp.getOperand(0);
3016 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3018 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
3019 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
3026 std::swap(Op0, Op1);
3028 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3031 Result = DAG.getNOT(dl, Result, VT);
3036 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3037 /// valid vector constant for a NEON instruction with a "modified immediate"
3038 /// operand (e.g., VMOV). If so, return the encoded value.
3039 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3040 unsigned SplatBitSize, SelectionDAG &DAG,
3041 EVT &VT, bool is128Bits, bool isVMOV) {
3042 unsigned OpCmode, Imm;
3044 // SplatBitSize is set to the smallest size that splats the vector, so a
3045 // zero vector will always have SplatBitSize == 8. However, NEON modified
3046 // immediate instructions others than VMOV do not support the 8-bit encoding
3047 // of a zero vector, and the default encoding of zero is supposed to be the
3052 switch (SplatBitSize) {
3056 // Any 1-byte value is OK. Op=0, Cmode=1110.
3057 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3060 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3064 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3065 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3066 if ((SplatBits & ~0xff) == 0) {
3067 // Value = 0x00nn: Op=x, Cmode=100x.
3072 if ((SplatBits & ~0xff00) == 0) {
3073 // Value = 0xnn00: Op=x, Cmode=101x.
3075 Imm = SplatBits >> 8;
3081 // NEON's 32-bit VMOV supports splat values where:
3082 // * only one byte is nonzero, or
3083 // * the least significant byte is 0xff and the second byte is nonzero, or
3084 // * the least significant 2 bytes are 0xff and the third is nonzero.
3085 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3086 if ((SplatBits & ~0xff) == 0) {
3087 // Value = 0x000000nn: Op=x, Cmode=000x.
3092 if ((SplatBits & ~0xff00) == 0) {
3093 // Value = 0x0000nn00: Op=x, Cmode=001x.
3095 Imm = SplatBits >> 8;
3098 if ((SplatBits & ~0xff0000) == 0) {
3099 // Value = 0x00nn0000: Op=x, Cmode=010x.
3101 Imm = SplatBits >> 16;
3104 if ((SplatBits & ~0xff000000) == 0) {
3105 // Value = 0xnn000000: Op=x, Cmode=011x.
3107 Imm = SplatBits >> 24;
3111 if ((SplatBits & ~0xffff) == 0 &&
3112 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3113 // Value = 0x0000nnff: Op=x, Cmode=1100.
3115 Imm = SplatBits >> 8;
3120 if ((SplatBits & ~0xffffff) == 0 &&
3121 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3122 // Value = 0x00nnffff: Op=x, Cmode=1101.
3124 Imm = SplatBits >> 16;
3125 SplatBits |= 0xffff;
3129 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3130 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3131 // VMOV.I32. A (very) minor optimization would be to replicate the value
3132 // and fall through here to test for a valid 64-bit splat. But, then the
3133 // caller would also need to check and handle the change in size.
3139 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3140 uint64_t BitMask = 0xff;
3142 unsigned ImmMask = 1;
3144 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3145 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3148 } else if ((SplatBits & BitMask) != 0) {
3154 // Op=1, Cmode=1110.
3157 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3162 llvm_unreachable("unexpected size for isNEONModifiedImm");
3166 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3167 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3170 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3171 bool &ReverseVEXT, unsigned &Imm) {
3172 unsigned NumElts = VT.getVectorNumElements();
3173 ReverseVEXT = false;
3175 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3181 // If this is a VEXT shuffle, the immediate value is the index of the first
3182 // element. The other shuffle indices must be the successive elements after
3184 unsigned ExpectedElt = Imm;
3185 for (unsigned i = 1; i < NumElts; ++i) {
3186 // Increment the expected index. If it wraps around, it may still be
3187 // a VEXT but the source vectors must be swapped.
3189 if (ExpectedElt == NumElts * 2) {
3194 if (M[i] < 0) continue; // ignore UNDEF indices
3195 if (ExpectedElt != static_cast<unsigned>(M[i]))
3199 // Adjust the index value if the source operands will be swapped.
3206 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3207 /// instruction with the specified blocksize. (The order of the elements
3208 /// within each block of the vector is reversed.)
3209 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3210 unsigned BlockSize) {
3211 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3212 "Only possible block sizes for VREV are: 16, 32, 64");
3214 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3218 unsigned NumElts = VT.getVectorNumElements();
3219 unsigned BlockElts = M[0] + 1;
3220 // If the first shuffle index is UNDEF, be optimistic.
3222 BlockElts = BlockSize / EltSz;
3224 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3227 for (unsigned i = 0; i < NumElts; ++i) {
3228 if (M[i] < 0) continue; // ignore UNDEF indices
3229 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3236 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3237 unsigned &WhichResult) {
3238 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3242 unsigned NumElts = VT.getVectorNumElements();
3243 WhichResult = (M[0] == 0 ? 0 : 1);
3244 for (unsigned i = 0; i < NumElts; i += 2) {
3245 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3246 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3252 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3253 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3254 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3255 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3256 unsigned &WhichResult) {
3257 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3261 unsigned NumElts = VT.getVectorNumElements();
3262 WhichResult = (M[0] == 0 ? 0 : 1);
3263 for (unsigned i = 0; i < NumElts; i += 2) {
3264 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3265 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3271 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3272 unsigned &WhichResult) {
3273 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3277 unsigned NumElts = VT.getVectorNumElements();
3278 WhichResult = (M[0] == 0 ? 0 : 1);
3279 for (unsigned i = 0; i != NumElts; ++i) {
3280 if (M[i] < 0) continue; // ignore UNDEF indices
3281 if ((unsigned) M[i] != 2 * i + WhichResult)
3285 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3286 if (VT.is64BitVector() && EltSz == 32)
3292 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3293 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3294 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3295 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3296 unsigned &WhichResult) {
3297 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3301 unsigned Half = VT.getVectorNumElements() / 2;
3302 WhichResult = (M[0] == 0 ? 0 : 1);
3303 for (unsigned j = 0; j != 2; ++j) {
3304 unsigned Idx = WhichResult;
3305 for (unsigned i = 0; i != Half; ++i) {
3306 int MIdx = M[i + j * Half];
3307 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3313 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3314 if (VT.is64BitVector() && EltSz == 32)
3320 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3321 unsigned &WhichResult) {
3322 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3326 unsigned NumElts = VT.getVectorNumElements();
3327 WhichResult = (M[0] == 0 ? 0 : 1);
3328 unsigned Idx = WhichResult * NumElts / 2;
3329 for (unsigned i = 0; i != NumElts; i += 2) {
3330 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3331 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3336 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3337 if (VT.is64BitVector() && EltSz == 32)
3343 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3344 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3345 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3346 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3347 unsigned &WhichResult) {
3348 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3352 unsigned NumElts = VT.getVectorNumElements();
3353 WhichResult = (M[0] == 0 ? 0 : 1);
3354 unsigned Idx = WhichResult * NumElts / 2;
3355 for (unsigned i = 0; i != NumElts; i += 2) {
3356 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3357 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3362 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3363 if (VT.is64BitVector() && EltSz == 32)
3369 // If N is an integer constant that can be moved into a register in one
3370 // instruction, return an SDValue of such a constant (will become a MOV
3371 // instruction). Otherwise return null.
3372 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3373 const ARMSubtarget *ST, DebugLoc dl) {
3375 if (!isa<ConstantSDNode>(N))
3377 Val = cast<ConstantSDNode>(N)->getZExtValue();
3379 if (ST->isThumb1Only()) {
3380 if (Val <= 255 || ~Val <= 255)
3381 return DAG.getConstant(Val, MVT::i32);
3383 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3384 return DAG.getConstant(Val, MVT::i32);
3389 // If this is a case we can't handle, return null and let the default
3390 // expansion code take care of it.
3391 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3392 const ARMSubtarget *ST) {
3393 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3394 DebugLoc dl = Op.getDebugLoc();
3395 EVT VT = Op.getValueType();
3397 APInt SplatBits, SplatUndef;
3398 unsigned SplatBitSize;
3400 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3401 if (SplatBitSize <= 64) {
3402 // Check if an immediate VMOV works.
3404 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3405 SplatUndef.getZExtValue(), SplatBitSize,
3406 DAG, VmovVT, VT.is128BitVector(), true);
3407 if (Val.getNode()) {
3408 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3409 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3412 // Try an immediate VMVN.
3413 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3414 ((1LL << SplatBitSize) - 1));
3415 Val = isNEONModifiedImm(NegatedImm,
3416 SplatUndef.getZExtValue(), SplatBitSize,
3417 DAG, VmovVT, VT.is128BitVector(), false);
3418 if (Val.getNode()) {
3419 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3420 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3425 // Scan through the operands to see if only one value is used.
3426 unsigned NumElts = VT.getVectorNumElements();
3427 bool isOnlyLowElement = true;
3428 bool usesOnlyOneValue = true;
3429 bool isConstant = true;
3431 for (unsigned i = 0; i < NumElts; ++i) {
3432 SDValue V = Op.getOperand(i);
3433 if (V.getOpcode() == ISD::UNDEF)
3436 isOnlyLowElement = false;
3437 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3440 if (!Value.getNode())
3442 else if (V != Value)
3443 usesOnlyOneValue = false;
3446 if (!Value.getNode())
3447 return DAG.getUNDEF(VT);
3449 if (isOnlyLowElement)
3450 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3452 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3454 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3455 // i32 and try again.
3456 if (usesOnlyOneValue && EltSize <= 32) {
3458 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3459 if (VT.getVectorElementType().isFloatingPoint()) {
3460 SmallVector<SDValue, 8> Ops;
3461 for (unsigned i = 0; i < NumElts; ++i)
3462 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3464 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3466 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3468 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3470 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3472 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3475 // If all elements are constants and the case above didn't get hit, fall back
3476 // to the default expansion, which will generate a load from the constant
3481 // Vectors with 32- or 64-bit elements can be built by directly assigning
3482 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3483 // will be legalized.
3484 if (EltSize >= 32) {
3485 // Do the expansion with floating-point types, since that is what the VFP
3486 // registers are defined to use, and since i64 is not legal.
3487 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3488 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3489 SmallVector<SDValue, 8> Ops;
3490 for (unsigned i = 0; i < NumElts; ++i)
3491 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3492 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3493 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3499 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3500 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3501 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3502 /// are assumed to be legal.
3504 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3506 if (VT.getVectorNumElements() == 4 &&
3507 (VT.is128BitVector() || VT.is64BitVector())) {
3508 unsigned PFIndexes[4];
3509 for (unsigned i = 0; i != 4; ++i) {
3513 PFIndexes[i] = M[i];
3516 // Compute the index in the perfect shuffle table.
3517 unsigned PFTableIndex =
3518 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3519 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3520 unsigned Cost = (PFEntry >> 30);
3527 unsigned Imm, WhichResult;
3529 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3530 return (EltSize >= 32 ||
3531 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3532 isVREVMask(M, VT, 64) ||
3533 isVREVMask(M, VT, 32) ||
3534 isVREVMask(M, VT, 16) ||
3535 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3536 isVTRNMask(M, VT, WhichResult) ||
3537 isVUZPMask(M, VT, WhichResult) ||
3538 isVZIPMask(M, VT, WhichResult) ||
3539 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3540 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3541 isVZIP_v_undef_Mask(M, VT, WhichResult));
3544 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3545 /// the specified operations to build the shuffle.
3546 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3547 SDValue RHS, SelectionDAG &DAG,
3549 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3550 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3551 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3554 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3563 OP_VUZPL, // VUZP, left result
3564 OP_VUZPR, // VUZP, right result
3565 OP_VZIPL, // VZIP, left result
3566 OP_VZIPR, // VZIP, right result
3567 OP_VTRNL, // VTRN, left result
3568 OP_VTRNR // VTRN, right result
3571 if (OpNum == OP_COPY) {
3572 if (LHSID == (1*9+2)*9+3) return LHS;
3573 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3577 SDValue OpLHS, OpRHS;
3578 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3579 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3580 EVT VT = OpLHS.getValueType();
3583 default: llvm_unreachable("Unknown shuffle opcode!");
3585 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3590 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3591 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3595 return DAG.getNode(ARMISD::VEXT, dl, VT,
3597 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3600 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3601 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3604 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3605 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3608 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3609 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3613 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3614 SDValue V1 = Op.getOperand(0);
3615 SDValue V2 = Op.getOperand(1);
3616 DebugLoc dl = Op.getDebugLoc();
3617 EVT VT = Op.getValueType();
3618 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3619 SmallVector<int, 8> ShuffleMask;
3621 // Convert shuffles that are directly supported on NEON to target-specific
3622 // DAG nodes, instead of keeping them as shuffles and matching them again
3623 // during code selection. This is more efficient and avoids the possibility
3624 // of inconsistencies between legalization and selection.
3625 // FIXME: floating-point vectors should be canonicalized to integer vectors
3626 // of the same time so that they get CSEd properly.
3627 SVN->getMask(ShuffleMask);
3629 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3630 if (EltSize <= 32) {
3631 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3632 int Lane = SVN->getSplatIndex();
3633 // If this is undef splat, generate it via "just" vdup, if possible.
3634 if (Lane == -1) Lane = 0;
3636 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3637 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3639 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3640 DAG.getConstant(Lane, MVT::i32));
3645 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3648 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3649 DAG.getConstant(Imm, MVT::i32));
3652 if (isVREVMask(ShuffleMask, VT, 64))
3653 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3654 if (isVREVMask(ShuffleMask, VT, 32))
3655 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3656 if (isVREVMask(ShuffleMask, VT, 16))
3657 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3659 // Check for Neon shuffles that modify both input vectors in place.
3660 // If both results are used, i.e., if there are two shuffles with the same
3661 // source operands and with masks corresponding to both results of one of
3662 // these operations, DAG memoization will ensure that a single node is
3663 // used for both shuffles.
3664 unsigned WhichResult;
3665 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3666 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3667 V1, V2).getValue(WhichResult);
3668 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3669 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3670 V1, V2).getValue(WhichResult);
3671 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3672 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3673 V1, V2).getValue(WhichResult);
3675 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3676 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3677 V1, V1).getValue(WhichResult);
3678 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3679 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3680 V1, V1).getValue(WhichResult);
3681 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3682 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3683 V1, V1).getValue(WhichResult);
3686 // If the shuffle is not directly supported and it has 4 elements, use
3687 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3688 unsigned NumElts = VT.getVectorNumElements();
3690 unsigned PFIndexes[4];
3691 for (unsigned i = 0; i != 4; ++i) {
3692 if (ShuffleMask[i] < 0)
3695 PFIndexes[i] = ShuffleMask[i];
3698 // Compute the index in the perfect shuffle table.
3699 unsigned PFTableIndex =
3700 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3701 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3702 unsigned Cost = (PFEntry >> 30);
3705 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3708 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3709 if (EltSize >= 32) {
3710 // Do the expansion with floating-point types, since that is what the VFP
3711 // registers are defined to use, and since i64 is not legal.
3712 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3713 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3714 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3715 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3716 SmallVector<SDValue, 8> Ops;
3717 for (unsigned i = 0; i < NumElts; ++i) {
3718 if (ShuffleMask[i] < 0)
3719 Ops.push_back(DAG.getUNDEF(EltVT));
3721 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3722 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3723 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3726 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3727 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3733 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3734 EVT VT = Op.getValueType();
3735 DebugLoc dl = Op.getDebugLoc();
3736 SDValue Vec = Op.getOperand(0);
3737 SDValue Lane = Op.getOperand(1);
3738 assert(VT == MVT::i32 &&
3739 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3740 "unexpected type for custom-lowering vector extract");
3741 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3744 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3745 // The only time a CONCAT_VECTORS operation can have legal types is when
3746 // two 64-bit vectors are concatenated to a 128-bit vector.
3747 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3748 "unexpected CONCAT_VECTORS");
3749 DebugLoc dl = Op.getDebugLoc();
3750 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3751 SDValue Op0 = Op.getOperand(0);
3752 SDValue Op1 = Op.getOperand(1);
3753 if (Op0.getOpcode() != ISD::UNDEF)
3754 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3755 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3756 DAG.getIntPtrConstant(0));
3757 if (Op1.getOpcode() != ISD::UNDEF)
3758 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3759 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3760 DAG.getIntPtrConstant(1));
3761 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3764 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3765 /// an extending load, return the unextended value.
3766 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3767 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3768 return N->getOperand(0);
3769 LoadSDNode *LD = cast<LoadSDNode>(N);
3770 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3771 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3772 LD->isNonTemporal(), LD->getAlignment());
3775 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3776 // Multiplications are only custom-lowered for 128-bit vectors so that
3777 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3778 EVT VT = Op.getValueType();
3779 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3780 SDNode *N0 = Op.getOperand(0).getNode();
3781 SDNode *N1 = Op.getOperand(1).getNode();
3782 unsigned NewOpc = 0;
3783 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3784 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3785 NewOpc = ARMISD::VMULLs;
3786 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3787 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3788 NewOpc = ARMISD::VMULLu;
3789 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3790 // Fall through to expand this. It is not legal.
3793 // Other vector multiplications are legal.
3797 // Legalize to a VMULL instruction.
3798 DebugLoc DL = Op.getDebugLoc();
3799 SDValue Op0 = SkipExtension(N0, DAG);
3800 SDValue Op1 = SkipExtension(N1, DAG);
3802 assert(Op0.getValueType().is64BitVector() &&
3803 Op1.getValueType().is64BitVector() &&
3804 "unexpected types for extended operands to VMULL");
3805 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3808 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3809 switch (Op.getOpcode()) {
3810 default: llvm_unreachable("Don't know how to custom lower this!");
3811 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3812 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3813 case ISD::GlobalAddress:
3814 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3815 LowerGlobalAddressELF(Op, DAG);
3816 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3817 case ISD::SELECT: return LowerSELECT(Op, DAG);
3818 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3819 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3820 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3821 case ISD::VASTART: return LowerVASTART(Op, DAG);
3822 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3823 case ISD::SINT_TO_FP:
3824 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3825 case ISD::FP_TO_SINT:
3826 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3827 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3828 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3829 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3830 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3831 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3832 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3833 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
3834 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3836 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3839 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3840 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3841 case ISD::SRL_PARTS:
3842 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3843 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3844 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3845 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3846 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3847 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3848 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3849 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3850 case ISD::MUL: return LowerMUL(Op, DAG);
3855 /// ReplaceNodeResults - Replace the results of node with an illegal result
3856 /// type with new values built out of custom code.
3857 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3858 SmallVectorImpl<SDValue>&Results,
3859 SelectionDAG &DAG) const {
3861 switch (N->getOpcode()) {
3863 llvm_unreachable("Don't know how to custom expand this!");
3865 case ISD::BIT_CONVERT:
3866 Res = ExpandBIT_CONVERT(N, DAG);
3870 Res = LowerShift(N, DAG, Subtarget);
3874 Results.push_back(Res);
3877 //===----------------------------------------------------------------------===//
3878 // ARM Scheduler Hooks
3879 //===----------------------------------------------------------------------===//
3882 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3883 MachineBasicBlock *BB,
3884 unsigned Size) const {
3885 unsigned dest = MI->getOperand(0).getReg();
3886 unsigned ptr = MI->getOperand(1).getReg();
3887 unsigned oldval = MI->getOperand(2).getReg();
3888 unsigned newval = MI->getOperand(3).getReg();
3889 unsigned scratch = BB->getParent()->getRegInfo()
3890 .createVirtualRegister(ARM::GPRRegisterClass);
3891 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3892 DebugLoc dl = MI->getDebugLoc();
3893 bool isThumb2 = Subtarget->isThumb2();
3895 unsigned ldrOpc, strOpc;
3897 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3899 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3900 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3903 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3904 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3907 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3908 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3912 MachineFunction *MF = BB->getParent();
3913 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3914 MachineFunction::iterator It = BB;
3915 ++It; // insert the new blocks after the current block
3917 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3918 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3919 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3920 MF->insert(It, loop1MBB);
3921 MF->insert(It, loop2MBB);
3922 MF->insert(It, exitMBB);
3924 // Transfer the remainder of BB and its successor edges to exitMBB.
3925 exitMBB->splice(exitMBB->begin(), BB,
3926 llvm::next(MachineBasicBlock::iterator(MI)),
3928 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3932 // fallthrough --> loop1MBB
3933 BB->addSuccessor(loop1MBB);
3936 // ldrex dest, [ptr]
3940 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3941 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3942 .addReg(dest).addReg(oldval));
3943 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3944 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3945 BB->addSuccessor(loop2MBB);
3946 BB->addSuccessor(exitMBB);
3949 // strex scratch, newval, [ptr]
3953 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3955 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3956 .addReg(scratch).addImm(0));
3957 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3958 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3959 BB->addSuccessor(loop1MBB);
3960 BB->addSuccessor(exitMBB);
3966 MI->eraseFromParent(); // The instruction is gone now.
3972 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3973 unsigned Size, unsigned BinOpcode) const {
3974 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3975 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3977 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3978 MachineFunction *MF = BB->getParent();
3979 MachineFunction::iterator It = BB;
3982 unsigned dest = MI->getOperand(0).getReg();
3983 unsigned ptr = MI->getOperand(1).getReg();
3984 unsigned incr = MI->getOperand(2).getReg();
3985 DebugLoc dl = MI->getDebugLoc();
3987 bool isThumb2 = Subtarget->isThumb2();
3988 unsigned ldrOpc, strOpc;
3990 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3992 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3993 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3996 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3997 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4000 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4001 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4005 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4006 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4007 MF->insert(It, loopMBB);
4008 MF->insert(It, exitMBB);
4010 // Transfer the remainder of BB and its successor edges to exitMBB.
4011 exitMBB->splice(exitMBB->begin(), BB,
4012 llvm::next(MachineBasicBlock::iterator(MI)),
4014 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4016 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4017 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4018 unsigned scratch2 = (!BinOpcode) ? incr :
4019 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4023 // fallthrough --> loopMBB
4024 BB->addSuccessor(loopMBB);
4028 // <binop> scratch2, dest, incr
4029 // strex scratch, scratch2, ptr
4032 // fallthrough --> exitMBB
4034 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4036 // operand order needs to go the other way for NAND
4037 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4038 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4039 addReg(incr).addReg(dest)).addReg(0);
4041 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4042 addReg(dest).addReg(incr)).addReg(0);
4045 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4047 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4048 .addReg(scratch).addImm(0));
4049 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4050 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4052 BB->addSuccessor(loopMBB);
4053 BB->addSuccessor(exitMBB);
4059 MI->eraseFromParent(); // The instruction is gone now.
4065 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4066 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4067 E = MBB->succ_end(); I != E; ++I)
4070 llvm_unreachable("Expecting a BB with two successors!");
4074 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4075 MachineBasicBlock *BB) const {
4076 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4077 DebugLoc dl = MI->getDebugLoc();
4078 bool isThumb2 = Subtarget->isThumb2();
4079 switch (MI->getOpcode()) {
4082 llvm_unreachable("Unexpected instr type to insert");
4084 case ARM::ATOMIC_LOAD_ADD_I8:
4085 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4086 case ARM::ATOMIC_LOAD_ADD_I16:
4087 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4088 case ARM::ATOMIC_LOAD_ADD_I32:
4089 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4091 case ARM::ATOMIC_LOAD_AND_I8:
4092 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4093 case ARM::ATOMIC_LOAD_AND_I16:
4094 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4095 case ARM::ATOMIC_LOAD_AND_I32:
4096 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4098 case ARM::ATOMIC_LOAD_OR_I8:
4099 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4100 case ARM::ATOMIC_LOAD_OR_I16:
4101 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4102 case ARM::ATOMIC_LOAD_OR_I32:
4103 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4105 case ARM::ATOMIC_LOAD_XOR_I8:
4106 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4107 case ARM::ATOMIC_LOAD_XOR_I16:
4108 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4109 case ARM::ATOMIC_LOAD_XOR_I32:
4110 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4112 case ARM::ATOMIC_LOAD_NAND_I8:
4113 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4114 case ARM::ATOMIC_LOAD_NAND_I16:
4115 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4116 case ARM::ATOMIC_LOAD_NAND_I32:
4117 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4119 case ARM::ATOMIC_LOAD_SUB_I8:
4120 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4121 case ARM::ATOMIC_LOAD_SUB_I16:
4122 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4123 case ARM::ATOMIC_LOAD_SUB_I32:
4124 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4126 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4127 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4128 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4130 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4131 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4132 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4134 case ARM::tMOVCCr_pseudo: {
4135 // To "insert" a SELECT_CC instruction, we actually have to insert the
4136 // diamond control-flow pattern. The incoming instruction knows the
4137 // destination vreg to set, the condition code register to branch on, the
4138 // true/false values to select between, and a branch opcode to use.
4139 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4140 MachineFunction::iterator It = BB;
4146 // cmpTY ccX, r1, r2
4148 // fallthrough --> copy0MBB
4149 MachineBasicBlock *thisMBB = BB;
4150 MachineFunction *F = BB->getParent();
4151 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4152 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4153 F->insert(It, copy0MBB);
4154 F->insert(It, sinkMBB);
4156 // Transfer the remainder of BB and its successor edges to sinkMBB.
4157 sinkMBB->splice(sinkMBB->begin(), BB,
4158 llvm::next(MachineBasicBlock::iterator(MI)),
4160 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4162 BB->addSuccessor(copy0MBB);
4163 BB->addSuccessor(sinkMBB);
4165 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4166 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4169 // %FalseValue = ...
4170 // # fallthrough to sinkMBB
4173 // Update machine-CFG edges
4174 BB->addSuccessor(sinkMBB);
4177 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4180 BuildMI(*BB, BB->begin(), dl,
4181 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4182 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4183 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4185 MI->eraseFromParent(); // The pseudo instruction is gone now.
4190 case ARM::BCCZi64: {
4191 // Compare both parts that make up the double comparison separately for
4193 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4195 unsigned LHS1 = MI->getOperand(1).getReg();
4196 unsigned LHS2 = MI->getOperand(2).getReg();
4198 AddDefaultPred(BuildMI(BB, dl,
4199 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4200 .addReg(LHS1).addImm(0));
4201 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4202 .addReg(LHS2).addImm(0)
4203 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4205 unsigned RHS1 = MI->getOperand(3).getReg();
4206 unsigned RHS2 = MI->getOperand(4).getReg();
4207 AddDefaultPred(BuildMI(BB, dl,
4208 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4209 .addReg(LHS1).addReg(RHS1));
4210 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4211 .addReg(LHS2).addReg(RHS2)
4212 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4215 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4216 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4217 if (MI->getOperand(0).getImm() == ARMCC::NE)
4218 std::swap(destMBB, exitMBB);
4220 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4221 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4222 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4225 MI->eraseFromParent(); // The pseudo instruction is gone now.
4231 //===----------------------------------------------------------------------===//
4232 // ARM Optimization Hooks
4233 //===----------------------------------------------------------------------===//
4236 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4237 TargetLowering::DAGCombinerInfo &DCI) {
4238 SelectionDAG &DAG = DCI.DAG;
4239 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4240 EVT VT = N->getValueType(0);
4241 unsigned Opc = N->getOpcode();
4242 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4243 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4244 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4245 ISD::CondCode CC = ISD::SETCC_INVALID;
4248 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4250 SDValue CCOp = Slct.getOperand(0);
4251 if (CCOp.getOpcode() == ISD::SETCC)
4252 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4255 bool DoXform = false;
4257 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4260 if (LHS.getOpcode() == ISD::Constant &&
4261 cast<ConstantSDNode>(LHS)->isNullValue()) {
4263 } else if (CC != ISD::SETCC_INVALID &&
4264 RHS.getOpcode() == ISD::Constant &&
4265 cast<ConstantSDNode>(RHS)->isNullValue()) {
4266 std::swap(LHS, RHS);
4267 SDValue Op0 = Slct.getOperand(0);
4268 EVT OpVT = isSlctCC ? Op0.getValueType() :
4269 Op0.getOperand(0).getValueType();
4270 bool isInt = OpVT.isInteger();
4271 CC = ISD::getSetCCInverse(CC, isInt);
4273 if (!TLI.isCondCodeLegal(CC, OpVT))
4274 return SDValue(); // Inverse operator isn't legal.
4281 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4283 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4284 Slct.getOperand(0), Slct.getOperand(1), CC);
4285 SDValue CCOp = Slct.getOperand(0);
4287 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4288 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4289 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4290 CCOp, OtherOp, Result);
4295 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4296 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4297 /// called with the default operands, and if that fails, with commuted
4299 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4300 TargetLowering::DAGCombinerInfo &DCI) {
4301 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4302 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4303 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4304 if (Result.getNode()) return Result;
4309 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4311 static SDValue PerformADDCombine(SDNode *N,
4312 TargetLowering::DAGCombinerInfo &DCI) {
4313 SDValue N0 = N->getOperand(0);
4314 SDValue N1 = N->getOperand(1);
4316 // First try with the default operand order.
4317 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4318 if (Result.getNode())
4321 // If that didn't work, try again with the operands commuted.
4322 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4325 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4327 static SDValue PerformSUBCombine(SDNode *N,
4328 TargetLowering::DAGCombinerInfo &DCI) {
4329 SDValue N0 = N->getOperand(0);
4330 SDValue N1 = N->getOperand(1);
4332 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4333 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4334 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4335 if (Result.getNode()) return Result;
4341 static SDValue PerformMULCombine(SDNode *N,
4342 TargetLowering::DAGCombinerInfo &DCI,
4343 const ARMSubtarget *Subtarget) {
4344 SelectionDAG &DAG = DCI.DAG;
4346 if (Subtarget->isThumb1Only())
4349 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4352 EVT VT = N->getValueType(0);
4356 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4360 uint64_t MulAmt = C->getZExtValue();
4361 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4362 ShiftAmt = ShiftAmt & (32 - 1);
4363 SDValue V = N->getOperand(0);
4364 DebugLoc DL = N->getDebugLoc();
4367 MulAmt >>= ShiftAmt;
4368 if (isPowerOf2_32(MulAmt - 1)) {
4369 // (mul x, 2^N + 1) => (add (shl x, N), x)
4370 Res = DAG.getNode(ISD::ADD, DL, VT,
4371 V, DAG.getNode(ISD::SHL, DL, VT,
4372 V, DAG.getConstant(Log2_32(MulAmt-1),
4374 } else if (isPowerOf2_32(MulAmt + 1)) {
4375 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4376 Res = DAG.getNode(ISD::SUB, DL, VT,
4377 DAG.getNode(ISD::SHL, DL, VT,
4378 V, DAG.getConstant(Log2_32(MulAmt+1),
4385 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4386 DAG.getConstant(ShiftAmt, MVT::i32));
4388 // Do not add new nodes to DAG combiner worklist.
4389 DCI.CombineTo(N, Res, false);
4393 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4394 static SDValue PerformORCombine(SDNode *N,
4395 TargetLowering::DAGCombinerInfo &DCI,
4396 const ARMSubtarget *Subtarget) {
4397 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4400 // BFI is only available on V6T2+
4401 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4404 SelectionDAG &DAG = DCI.DAG;
4405 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4406 DebugLoc DL = N->getDebugLoc();
4407 // 1) or (and A, mask), val => ARMbfi A, val, mask
4408 // iff (val & mask) == val
4410 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4411 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4412 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4413 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4414 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4415 // (i.e., copy a bitfield value into another bitfield of the same width)
4416 if (N0.getOpcode() != ISD::AND)
4419 EVT VT = N->getValueType(0);
4424 // The value and the mask need to be constants so we can verify this is
4425 // actually a bitfield set. If the mask is 0xffff, we can do better
4426 // via a movt instruction, so don't use BFI in that case.
4427 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4430 unsigned Mask = C->getZExtValue();
4434 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4435 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4436 unsigned Val = C->getZExtValue();
4437 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4439 Val >>= CountTrailingZeros_32(~Mask);
4441 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4442 DAG.getConstant(Val, MVT::i32),
4443 DAG.getConstant(Mask, MVT::i32));
4445 // Do not add new nodes to DAG combiner worklist.
4446 DCI.CombineTo(N, Res, false);
4447 } else if (N1.getOpcode() == ISD::AND) {
4448 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4449 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4452 unsigned Mask2 = C->getZExtValue();
4454 if (ARM::isBitFieldInvertedMask(Mask) &&
4455 ARM::isBitFieldInvertedMask(~Mask2) &&
4456 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4457 // The pack halfword instruction works better for masks that fit it,
4458 // so use that when it's available.
4459 if (Subtarget->hasT2ExtractPack() &&
4460 (Mask == 0xffff || Mask == 0xffff0000))
4463 unsigned lsb = CountTrailingZeros_32(Mask2);
4464 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4465 DAG.getConstant(lsb, MVT::i32));
4466 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4467 DAG.getConstant(Mask, MVT::i32));
4468 // Do not add new nodes to DAG combiner worklist.
4469 DCI.CombineTo(N, Res, false);
4470 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4471 ARM::isBitFieldInvertedMask(Mask2) &&
4472 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4473 // The pack halfword instruction works better for masks that fit it,
4474 // so use that when it's available.
4475 if (Subtarget->hasT2ExtractPack() &&
4476 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4479 unsigned lsb = CountTrailingZeros_32(Mask);
4480 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4481 DAG.getConstant(lsb, MVT::i32));
4482 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4483 DAG.getConstant(Mask2, MVT::i32));
4484 // Do not add new nodes to DAG combiner worklist.
4485 DCI.CombineTo(N, Res, false);
4492 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4493 /// ARMISD::VMOVRRD.
4494 static SDValue PerformVMOVRRDCombine(SDNode *N,
4495 TargetLowering::DAGCombinerInfo &DCI) {
4496 // vmovrrd(vmovdrr x, y) -> x,y
4497 SDValue InDouble = N->getOperand(0);
4498 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4499 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4503 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4504 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4505 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4506 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4507 SDValue Op0 = N->getOperand(0);
4508 SDValue Op1 = N->getOperand(1);
4509 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4510 Op0 = Op0.getOperand(0);
4511 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4512 Op1 = Op1.getOperand(0);
4513 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4514 Op0.getNode() == Op1.getNode() &&
4515 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4516 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4517 N->getValueType(0), Op0.getOperand(0));
4521 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4522 /// ISD::BUILD_VECTOR.
4523 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4524 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4525 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4526 // into a pair of GPRs, which is fine when the value is used as a scalar,
4527 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4528 if (N->getNumOperands() == 2)
4529 return PerformVMOVDRRCombine(N, DAG);
4534 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4535 /// ARMISD::VDUPLANE.
4536 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
4537 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4539 SDValue Op = N->getOperand(0);
4540 EVT VT = N->getValueType(0);
4542 // Ignore bit_converts.
4543 while (Op.getOpcode() == ISD::BIT_CONVERT)
4544 Op = Op.getOperand(0);
4545 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4548 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4549 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4550 // The canonical VMOV for a zero vector uses a 32-bit element size.
4551 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4553 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4555 if (EltSize > VT.getVectorElementType().getSizeInBits())
4558 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4561 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4562 /// operand of a vector shift operation, where all the elements of the
4563 /// build_vector must have the same constant integer value.
4564 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4565 // Ignore bit_converts.
4566 while (Op.getOpcode() == ISD::BIT_CONVERT)
4567 Op = Op.getOperand(0);
4568 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4569 APInt SplatBits, SplatUndef;
4570 unsigned SplatBitSize;
4572 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4573 HasAnyUndefs, ElementBits) ||
4574 SplatBitSize > ElementBits)
4576 Cnt = SplatBits.getSExtValue();
4580 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4581 /// operand of a vector shift left operation. That value must be in the range:
4582 /// 0 <= Value < ElementBits for a left shift; or
4583 /// 0 <= Value <= ElementBits for a long left shift.
4584 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4585 assert(VT.isVector() && "vector shift count is not a vector type");
4586 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4587 if (! getVShiftImm(Op, ElementBits, Cnt))
4589 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4592 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4593 /// operand of a vector shift right operation. For a shift opcode, the value
4594 /// is positive, but for an intrinsic the value count must be negative. The
4595 /// absolute value must be in the range:
4596 /// 1 <= |Value| <= ElementBits for a right shift; or
4597 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4598 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4600 assert(VT.isVector() && "vector shift count is not a vector type");
4601 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4602 if (! getVShiftImm(Op, ElementBits, Cnt))
4606 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4609 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4610 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4611 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4614 // Don't do anything for most intrinsics.
4617 // Vector shifts: check for immediate versions and lower them.
4618 // Note: This is done during DAG combining instead of DAG legalizing because
4619 // the build_vectors for 64-bit vector element shift counts are generally
4620 // not legal, and it is hard to see their values after they get legalized to
4621 // loads from a constant pool.
4622 case Intrinsic::arm_neon_vshifts:
4623 case Intrinsic::arm_neon_vshiftu:
4624 case Intrinsic::arm_neon_vshiftls:
4625 case Intrinsic::arm_neon_vshiftlu:
4626 case Intrinsic::arm_neon_vshiftn:
4627 case Intrinsic::arm_neon_vrshifts:
4628 case Intrinsic::arm_neon_vrshiftu:
4629 case Intrinsic::arm_neon_vrshiftn:
4630 case Intrinsic::arm_neon_vqshifts:
4631 case Intrinsic::arm_neon_vqshiftu:
4632 case Intrinsic::arm_neon_vqshiftsu:
4633 case Intrinsic::arm_neon_vqshiftns:
4634 case Intrinsic::arm_neon_vqshiftnu:
4635 case Intrinsic::arm_neon_vqshiftnsu:
4636 case Intrinsic::arm_neon_vqrshiftns:
4637 case Intrinsic::arm_neon_vqrshiftnu:
4638 case Intrinsic::arm_neon_vqrshiftnsu: {
4639 EVT VT = N->getOperand(1).getValueType();
4641 unsigned VShiftOpc = 0;
4644 case Intrinsic::arm_neon_vshifts:
4645 case Intrinsic::arm_neon_vshiftu:
4646 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4647 VShiftOpc = ARMISD::VSHL;
4650 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4651 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4652 ARMISD::VSHRs : ARMISD::VSHRu);
4657 case Intrinsic::arm_neon_vshiftls:
4658 case Intrinsic::arm_neon_vshiftlu:
4659 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4661 llvm_unreachable("invalid shift count for vshll intrinsic");
4663 case Intrinsic::arm_neon_vrshifts:
4664 case Intrinsic::arm_neon_vrshiftu:
4665 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4669 case Intrinsic::arm_neon_vqshifts:
4670 case Intrinsic::arm_neon_vqshiftu:
4671 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4675 case Intrinsic::arm_neon_vqshiftsu:
4676 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4678 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4680 case Intrinsic::arm_neon_vshiftn:
4681 case Intrinsic::arm_neon_vrshiftn:
4682 case Intrinsic::arm_neon_vqshiftns:
4683 case Intrinsic::arm_neon_vqshiftnu:
4684 case Intrinsic::arm_neon_vqshiftnsu:
4685 case Intrinsic::arm_neon_vqrshiftns:
4686 case Intrinsic::arm_neon_vqrshiftnu:
4687 case Intrinsic::arm_neon_vqrshiftnsu:
4688 // Narrowing shifts require an immediate right shift.
4689 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4691 llvm_unreachable("invalid shift count for narrowing vector shift "
4695 llvm_unreachable("unhandled vector shift");
4699 case Intrinsic::arm_neon_vshifts:
4700 case Intrinsic::arm_neon_vshiftu:
4701 // Opcode already set above.
4703 case Intrinsic::arm_neon_vshiftls:
4704 case Intrinsic::arm_neon_vshiftlu:
4705 if (Cnt == VT.getVectorElementType().getSizeInBits())
4706 VShiftOpc = ARMISD::VSHLLi;
4708 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4709 ARMISD::VSHLLs : ARMISD::VSHLLu);
4711 case Intrinsic::arm_neon_vshiftn:
4712 VShiftOpc = ARMISD::VSHRN; break;
4713 case Intrinsic::arm_neon_vrshifts:
4714 VShiftOpc = ARMISD::VRSHRs; break;
4715 case Intrinsic::arm_neon_vrshiftu:
4716 VShiftOpc = ARMISD::VRSHRu; break;
4717 case Intrinsic::arm_neon_vrshiftn:
4718 VShiftOpc = ARMISD::VRSHRN; break;
4719 case Intrinsic::arm_neon_vqshifts:
4720 VShiftOpc = ARMISD::VQSHLs; break;
4721 case Intrinsic::arm_neon_vqshiftu:
4722 VShiftOpc = ARMISD::VQSHLu; break;
4723 case Intrinsic::arm_neon_vqshiftsu:
4724 VShiftOpc = ARMISD::VQSHLsu; break;
4725 case Intrinsic::arm_neon_vqshiftns:
4726 VShiftOpc = ARMISD::VQSHRNs; break;
4727 case Intrinsic::arm_neon_vqshiftnu:
4728 VShiftOpc = ARMISD::VQSHRNu; break;
4729 case Intrinsic::arm_neon_vqshiftnsu:
4730 VShiftOpc = ARMISD::VQSHRNsu; break;
4731 case Intrinsic::arm_neon_vqrshiftns:
4732 VShiftOpc = ARMISD::VQRSHRNs; break;
4733 case Intrinsic::arm_neon_vqrshiftnu:
4734 VShiftOpc = ARMISD::VQRSHRNu; break;
4735 case Intrinsic::arm_neon_vqrshiftnsu:
4736 VShiftOpc = ARMISD::VQRSHRNsu; break;
4739 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4740 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4743 case Intrinsic::arm_neon_vshiftins: {
4744 EVT VT = N->getOperand(1).getValueType();
4746 unsigned VShiftOpc = 0;
4748 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4749 VShiftOpc = ARMISD::VSLI;
4750 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4751 VShiftOpc = ARMISD::VSRI;
4753 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4756 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4757 N->getOperand(1), N->getOperand(2),
4758 DAG.getConstant(Cnt, MVT::i32));
4761 case Intrinsic::arm_neon_vqrshifts:
4762 case Intrinsic::arm_neon_vqrshiftu:
4763 // No immediate versions of these to check for.
4770 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4771 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4772 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4773 /// vector element shift counts are generally not legal, and it is hard to see
4774 /// their values after they get legalized to loads from a constant pool.
4775 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4776 const ARMSubtarget *ST) {
4777 EVT VT = N->getValueType(0);
4779 // Nothing to be done for scalar shifts.
4780 if (! VT.isVector())
4783 assert(ST->hasNEON() && "unexpected vector shift");
4786 switch (N->getOpcode()) {
4787 default: llvm_unreachable("unexpected shift opcode");
4790 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4791 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4792 DAG.getConstant(Cnt, MVT::i32));
4797 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4798 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4799 ARMISD::VSHRs : ARMISD::VSHRu);
4800 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4801 DAG.getConstant(Cnt, MVT::i32));
4807 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4808 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4809 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4810 const ARMSubtarget *ST) {
4811 SDValue N0 = N->getOperand(0);
4813 // Check for sign- and zero-extensions of vector extract operations of 8-
4814 // and 16-bit vector elements. NEON supports these directly. They are
4815 // handled during DAG combining because type legalization will promote them
4816 // to 32-bit types and it is messy to recognize the operations after that.
4817 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4818 SDValue Vec = N0.getOperand(0);
4819 SDValue Lane = N0.getOperand(1);
4820 EVT VT = N->getValueType(0);
4821 EVT EltVT = N0.getValueType();
4822 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4824 if (VT == MVT::i32 &&
4825 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4826 TLI.isTypeLegal(Vec.getValueType())) {
4829 switch (N->getOpcode()) {
4830 default: llvm_unreachable("unexpected opcode");
4831 case ISD::SIGN_EXTEND:
4832 Opc = ARMISD::VGETLANEs;
4834 case ISD::ZERO_EXTEND:
4835 case ISD::ANY_EXTEND:
4836 Opc = ARMISD::VGETLANEu;
4839 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4846 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4847 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4848 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4849 const ARMSubtarget *ST) {
4850 // If the target supports NEON, try to use vmax/vmin instructions for f32
4851 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4852 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4853 // a NaN; only do the transformation when it matches that behavior.
4855 // For now only do this when using NEON for FP operations; if using VFP, it
4856 // is not obvious that the benefit outweighs the cost of switching to the
4858 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4859 N->getValueType(0) != MVT::f32)
4862 SDValue CondLHS = N->getOperand(0);
4863 SDValue CondRHS = N->getOperand(1);
4864 SDValue LHS = N->getOperand(2);
4865 SDValue RHS = N->getOperand(3);
4866 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4868 unsigned Opcode = 0;
4870 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4871 IsReversed = false; // x CC y ? x : y
4872 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4873 IsReversed = true ; // x CC y ? y : x
4887 // If LHS is NaN, an ordered comparison will be false and the result will
4888 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4889 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4890 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4891 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4893 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4894 // will return -0, so vmin can only be used for unsafe math or if one of
4895 // the operands is known to be nonzero.
4896 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4898 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4900 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4909 // If LHS is NaN, an ordered comparison will be false and the result will
4910 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4911 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4912 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4913 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4915 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4916 // will return +0, so vmax can only be used for unsafe math or if one of
4917 // the operands is known to be nonzero.
4918 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4920 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4922 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4928 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4931 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4932 DAGCombinerInfo &DCI) const {
4933 switch (N->getOpcode()) {
4935 case ISD::ADD: return PerformADDCombine(N, DCI);
4936 case ISD::SUB: return PerformSUBCombine(N, DCI);
4937 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4938 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4939 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4940 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
4941 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
4942 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
4943 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4946 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4947 case ISD::SIGN_EXTEND:
4948 case ISD::ZERO_EXTEND:
4949 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4950 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4955 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4956 if (!Subtarget->allowsUnalignedMem())
4959 switch (VT.getSimpleVT().SimpleTy) {
4966 // FIXME: VLD1 etc with standard alignment is legal.
4970 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4975 switch (VT.getSimpleVT().SimpleTy) {
4976 default: return false;
4991 if ((V & (Scale - 1)) != 0)
4994 return V == (V & ((1LL << 5) - 1));
4997 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4998 const ARMSubtarget *Subtarget) {
5005 switch (VT.getSimpleVT().SimpleTy) {
5006 default: return false;
5011 // + imm12 or - imm8
5013 return V == (V & ((1LL << 8) - 1));
5014 return V == (V & ((1LL << 12) - 1));
5017 // Same as ARM mode. FIXME: NEON?
5018 if (!Subtarget->hasVFP2())
5023 return V == (V & ((1LL << 8) - 1));
5027 /// isLegalAddressImmediate - Return true if the integer value can be used
5028 /// as the offset of the target addressing mode for load / store of the
5030 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5031 const ARMSubtarget *Subtarget) {
5038 if (Subtarget->isThumb1Only())
5039 return isLegalT1AddressImmediate(V, VT);
5040 else if (Subtarget->isThumb2())
5041 return isLegalT2AddressImmediate(V, VT, Subtarget);
5046 switch (VT.getSimpleVT().SimpleTy) {
5047 default: return false;
5052 return V == (V & ((1LL << 12) - 1));
5055 return V == (V & ((1LL << 8) - 1));
5058 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5063 return V == (V & ((1LL << 8) - 1));
5067 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5069 int Scale = AM.Scale;
5073 switch (VT.getSimpleVT().SimpleTy) {
5074 default: return false;
5083 return Scale == 2 || Scale == 4 || Scale == 8;
5086 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5090 // Note, we allow "void" uses (basically, uses that aren't loads or
5091 // stores), because arm allows folding a scale into many arithmetic
5092 // operations. This should be made more precise and revisited later.
5094 // Allow r << imm, but the imm has to be a multiple of two.
5095 if (Scale & 1) return false;
5096 return isPowerOf2_32(Scale);
5100 /// isLegalAddressingMode - Return true if the addressing mode represented
5101 /// by AM is legal for this target, for a load/store of the specified type.
5102 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5103 const Type *Ty) const {
5104 EVT VT = getValueType(Ty, true);
5105 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5108 // Can never fold addr of global into load/store.
5113 case 0: // no scale reg, must be "r+i" or "r", or "i".
5116 if (Subtarget->isThumb1Only())
5120 // ARM doesn't support any R+R*scale+imm addr modes.
5127 if (Subtarget->isThumb2())
5128 return isLegalT2ScaledAddressingMode(AM, VT);
5130 int Scale = AM.Scale;
5131 switch (VT.getSimpleVT().SimpleTy) {
5132 default: return false;
5136 if (Scale < 0) Scale = -Scale;
5140 return isPowerOf2_32(Scale & ~1);
5144 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5149 // Note, we allow "void" uses (basically, uses that aren't loads or
5150 // stores), because arm allows folding a scale into many arithmetic
5151 // operations. This should be made more precise and revisited later.
5153 // Allow r << imm, but the imm has to be a multiple of two.
5154 if (Scale & 1) return false;
5155 return isPowerOf2_32(Scale);
5162 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5163 /// icmp immediate, that is the target has icmp instructions which can compare
5164 /// a register against the immediate without having to materialize the
5165 /// immediate into a register.
5166 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5167 if (!Subtarget->isThumb())
5168 return ARM_AM::getSOImmVal(Imm) != -1;
5169 if (Subtarget->isThumb2())
5170 return ARM_AM::getT2SOImmVal(Imm) != -1;
5171 return Imm >= 0 && Imm <= 255;
5174 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5175 bool isSEXTLoad, SDValue &Base,
5176 SDValue &Offset, bool &isInc,
5177 SelectionDAG &DAG) {
5178 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5181 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5183 Base = Ptr->getOperand(0);
5184 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5185 int RHSC = (int)RHS->getZExtValue();
5186 if (RHSC < 0 && RHSC > -256) {
5187 assert(Ptr->getOpcode() == ISD::ADD);
5189 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5193 isInc = (Ptr->getOpcode() == ISD::ADD);
5194 Offset = Ptr->getOperand(1);
5196 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5198 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5199 int RHSC = (int)RHS->getZExtValue();
5200 if (RHSC < 0 && RHSC > -0x1000) {
5201 assert(Ptr->getOpcode() == ISD::ADD);
5203 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5204 Base = Ptr->getOperand(0);
5209 if (Ptr->getOpcode() == ISD::ADD) {
5211 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5212 if (ShOpcVal != ARM_AM::no_shift) {
5213 Base = Ptr->getOperand(1);
5214 Offset = Ptr->getOperand(0);
5216 Base = Ptr->getOperand(0);
5217 Offset = Ptr->getOperand(1);
5222 isInc = (Ptr->getOpcode() == ISD::ADD);
5223 Base = Ptr->getOperand(0);
5224 Offset = Ptr->getOperand(1);
5228 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5232 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5233 bool isSEXTLoad, SDValue &Base,
5234 SDValue &Offset, bool &isInc,
5235 SelectionDAG &DAG) {
5236 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5239 Base = Ptr->getOperand(0);
5240 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5241 int RHSC = (int)RHS->getZExtValue();
5242 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5243 assert(Ptr->getOpcode() == ISD::ADD);
5245 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5247 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5248 isInc = Ptr->getOpcode() == ISD::ADD;
5249 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5257 /// getPreIndexedAddressParts - returns true by value, base pointer and
5258 /// offset pointer and addressing mode by reference if the node's address
5259 /// can be legally represented as pre-indexed load / store address.
5261 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5263 ISD::MemIndexedMode &AM,
5264 SelectionDAG &DAG) const {
5265 if (Subtarget->isThumb1Only())
5270 bool isSEXTLoad = false;
5271 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5272 Ptr = LD->getBasePtr();
5273 VT = LD->getMemoryVT();
5274 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5275 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5276 Ptr = ST->getBasePtr();
5277 VT = ST->getMemoryVT();
5282 bool isLegal = false;
5283 if (Subtarget->isThumb2())
5284 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5285 Offset, isInc, DAG);
5287 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5288 Offset, isInc, DAG);
5292 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5296 /// getPostIndexedAddressParts - returns true by value, base pointer and
5297 /// offset pointer and addressing mode by reference if this node can be
5298 /// combined with a load / store to form a post-indexed load / store.
5299 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5302 ISD::MemIndexedMode &AM,
5303 SelectionDAG &DAG) const {
5304 if (Subtarget->isThumb1Only())
5309 bool isSEXTLoad = false;
5310 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5311 VT = LD->getMemoryVT();
5312 Ptr = LD->getBasePtr();
5313 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5314 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5315 VT = ST->getMemoryVT();
5316 Ptr = ST->getBasePtr();
5321 bool isLegal = false;
5322 if (Subtarget->isThumb2())
5323 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5326 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5332 // Swap base ptr and offset to catch more post-index load / store when
5333 // it's legal. In Thumb2 mode, offset must be an immediate.
5334 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5335 !Subtarget->isThumb2())
5336 std::swap(Base, Offset);
5338 // Post-indexed load / store update the base pointer.
5343 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5347 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5351 const SelectionDAG &DAG,
5352 unsigned Depth) const {
5353 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5354 switch (Op.getOpcode()) {
5356 case ARMISD::CMOV: {
5357 // Bits are known zero/one if known on the LHS and RHS.
5358 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5359 if (KnownZero == 0 && KnownOne == 0) return;
5361 APInt KnownZeroRHS, KnownOneRHS;
5362 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5363 KnownZeroRHS, KnownOneRHS, Depth+1);
5364 KnownZero &= KnownZeroRHS;
5365 KnownOne &= KnownOneRHS;
5371 //===----------------------------------------------------------------------===//
5372 // ARM Inline Assembly Support
5373 //===----------------------------------------------------------------------===//
5375 /// getConstraintType - Given a constraint letter, return the type of
5376 /// constraint it is for this target.
5377 ARMTargetLowering::ConstraintType
5378 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5379 if (Constraint.size() == 1) {
5380 switch (Constraint[0]) {
5382 case 'l': return C_RegisterClass;
5383 case 'w': return C_RegisterClass;
5386 return TargetLowering::getConstraintType(Constraint);
5389 std::pair<unsigned, const TargetRegisterClass*>
5390 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5392 if (Constraint.size() == 1) {
5393 // GCC ARM Constraint Letters
5394 switch (Constraint[0]) {
5396 if (Subtarget->isThumb())
5397 return std::make_pair(0U, ARM::tGPRRegisterClass);
5399 return std::make_pair(0U, ARM::GPRRegisterClass);
5401 return std::make_pair(0U, ARM::GPRRegisterClass);
5404 return std::make_pair(0U, ARM::SPRRegisterClass);
5405 if (VT.getSizeInBits() == 64)
5406 return std::make_pair(0U, ARM::DPRRegisterClass);
5407 if (VT.getSizeInBits() == 128)
5408 return std::make_pair(0U, ARM::QPRRegisterClass);
5412 if (StringRef("{cc}").equals_lower(Constraint))
5413 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5415 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5418 std::vector<unsigned> ARMTargetLowering::
5419 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5421 if (Constraint.size() != 1)
5422 return std::vector<unsigned>();
5424 switch (Constraint[0]) { // GCC ARM Constraint Letters
5427 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5428 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5431 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5432 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5433 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5434 ARM::R12, ARM::LR, 0);
5437 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5438 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5439 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5440 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5441 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5442 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5443 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5444 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5445 if (VT.getSizeInBits() == 64)
5446 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5447 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5448 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5449 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5450 if (VT.getSizeInBits() == 128)
5451 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5452 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5456 return std::vector<unsigned>();
5459 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5460 /// vector. If it is invalid, don't add anything to Ops.
5461 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5463 std::vector<SDValue>&Ops,
5464 SelectionDAG &DAG) const {
5465 SDValue Result(0, 0);
5467 switch (Constraint) {
5469 case 'I': case 'J': case 'K': case 'L':
5470 case 'M': case 'N': case 'O':
5471 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5475 int64_t CVal64 = C->getSExtValue();
5476 int CVal = (int) CVal64;
5477 // None of these constraints allow values larger than 32 bits. Check
5478 // that the value fits in an int.
5482 switch (Constraint) {
5484 if (Subtarget->isThumb1Only()) {
5485 // This must be a constant between 0 and 255, for ADD
5487 if (CVal >= 0 && CVal <= 255)
5489 } else if (Subtarget->isThumb2()) {
5490 // A constant that can be used as an immediate value in a
5491 // data-processing instruction.
5492 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5495 // A constant that can be used as an immediate value in a
5496 // data-processing instruction.
5497 if (ARM_AM::getSOImmVal(CVal) != -1)
5503 if (Subtarget->isThumb()) { // FIXME thumb2
5504 // This must be a constant between -255 and -1, for negated ADD
5505 // immediates. This can be used in GCC with an "n" modifier that
5506 // prints the negated value, for use with SUB instructions. It is
5507 // not useful otherwise but is implemented for compatibility.
5508 if (CVal >= -255 && CVal <= -1)
5511 // This must be a constant between -4095 and 4095. It is not clear
5512 // what this constraint is intended for. Implemented for
5513 // compatibility with GCC.
5514 if (CVal >= -4095 && CVal <= 4095)
5520 if (Subtarget->isThumb1Only()) {
5521 // A 32-bit value where only one byte has a nonzero value. Exclude
5522 // zero to match GCC. This constraint is used by GCC internally for
5523 // constants that can be loaded with a move/shift combination.
5524 // It is not useful otherwise but is implemented for compatibility.
5525 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5527 } else if (Subtarget->isThumb2()) {
5528 // A constant whose bitwise inverse can be used as an immediate
5529 // value in a data-processing instruction. This can be used in GCC
5530 // with a "B" modifier that prints the inverted value, for use with
5531 // BIC and MVN instructions. It is not useful otherwise but is
5532 // implemented for compatibility.
5533 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5536 // A constant whose bitwise inverse can be used as an immediate
5537 // value in a data-processing instruction. This can be used in GCC
5538 // with a "B" modifier that prints the inverted value, for use with
5539 // BIC and MVN instructions. It is not useful otherwise but is
5540 // implemented for compatibility.
5541 if (ARM_AM::getSOImmVal(~CVal) != -1)
5547 if (Subtarget->isThumb1Only()) {
5548 // This must be a constant between -7 and 7,
5549 // for 3-operand ADD/SUB immediate instructions.
5550 if (CVal >= -7 && CVal < 7)
5552 } else if (Subtarget->isThumb2()) {
5553 // A constant whose negation can be used as an immediate value in a
5554 // data-processing instruction. This can be used in GCC with an "n"
5555 // modifier that prints the negated value, for use with SUB
5556 // instructions. It is not useful otherwise but is implemented for
5558 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5561 // A constant whose negation can be used as an immediate value in a
5562 // data-processing instruction. This can be used in GCC with an "n"
5563 // modifier that prints the negated value, for use with SUB
5564 // instructions. It is not useful otherwise but is implemented for
5566 if (ARM_AM::getSOImmVal(-CVal) != -1)
5572 if (Subtarget->isThumb()) { // FIXME thumb2
5573 // This must be a multiple of 4 between 0 and 1020, for
5574 // ADD sp + immediate.
5575 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5578 // A power of two or a constant between 0 and 32. This is used in
5579 // GCC for the shift amount on shifted register operands, but it is
5580 // useful in general for any shift amounts.
5581 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5587 if (Subtarget->isThumb()) { // FIXME thumb2
5588 // This must be a constant between 0 and 31, for shift amounts.
5589 if (CVal >= 0 && CVal <= 31)
5595 if (Subtarget->isThumb()) { // FIXME thumb2
5596 // This must be a multiple of 4 between -508 and 508, for
5597 // ADD/SUB sp = sp + immediate.
5598 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5603 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5607 if (Result.getNode()) {
5608 Ops.push_back(Result);
5611 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5615 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5616 // The ARM target isn't yet aware of offsets.
5620 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5621 APInt Imm = FPImm.bitcastToAPInt();
5622 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5623 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5624 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5626 // We can handle 4 bits of mantissa.
5627 // mantissa = (16+UInt(e:f:g:h))/16.
5628 if (Mantissa & 0x7ffff)
5631 if ((Mantissa & 0xf) != Mantissa)
5634 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5635 if (Exp < -3 || Exp > 4)
5637 Exp = ((Exp+3) & 0x7) ^ 4;
5639 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5642 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5643 APInt Imm = FPImm.bitcastToAPInt();
5644 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5645 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5646 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5648 // We can handle 4 bits of mantissa.
5649 // mantissa = (16+UInt(e:f:g:h))/16.
5650 if (Mantissa & 0xffffffffffffLL)
5653 if ((Mantissa & 0xf) != Mantissa)
5656 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5657 if (Exp < -3 || Exp > 4)
5659 Exp = ((Exp+3) & 0x7) ^ 4;
5661 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5664 bool ARM::isBitFieldInvertedMask(unsigned v) {
5665 if (v == 0xffffffff)
5667 // there can be 1's on either or both "outsides", all the "inside"
5669 unsigned int lsb = 0, msb = 31;
5670 while (v & (1 << msb)) --msb;
5671 while (v & (1 << lsb)) ++lsb;
5672 for (unsigned int i = lsb; i <= msb; ++i) {
5679 /// isFPImmLegal - Returns true if the target can instruction select the
5680 /// specified FP immediate natively. If false, the legalizer will
5681 /// materialize the FP immediate as a load from a constant pool.
5682 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5683 if (!Subtarget->hasVFP3())
5686 return ARM::getVFPf32Imm(Imm) != -1;
5688 return ARM::getVFPf64Imm(Imm) != -1;
5692 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5693 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5694 /// specified in the intrinsic calls.
5695 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5697 unsigned Intrinsic) const {
5698 switch (Intrinsic) {
5699 case Intrinsic::arm_neon_vld1:
5700 case Intrinsic::arm_neon_vld2:
5701 case Intrinsic::arm_neon_vld3:
5702 case Intrinsic::arm_neon_vld4:
5703 case Intrinsic::arm_neon_vld2lane:
5704 case Intrinsic::arm_neon_vld3lane:
5705 case Intrinsic::arm_neon_vld4lane: {
5706 Info.opc = ISD::INTRINSIC_W_CHAIN;
5707 // Conservatively set memVT to the entire set of vectors loaded.
5708 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
5709 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5710 Info.ptrVal = I.getArgOperand(0);
5712 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5713 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5714 Info.vol = false; // volatile loads with NEON intrinsics not supported
5715 Info.readMem = true;
5716 Info.writeMem = false;
5719 case Intrinsic::arm_neon_vst1:
5720 case Intrinsic::arm_neon_vst2:
5721 case Intrinsic::arm_neon_vst3:
5722 case Intrinsic::arm_neon_vst4:
5723 case Intrinsic::arm_neon_vst2lane:
5724 case Intrinsic::arm_neon_vst3lane:
5725 case Intrinsic::arm_neon_vst4lane: {
5726 Info.opc = ISD::INTRINSIC_VOID;
5727 // Conservatively set memVT to the entire set of vectors stored.
5728 unsigned NumElts = 0;
5729 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5730 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
5731 if (!ArgTy->isVectorTy())
5733 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
5735 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5736 Info.ptrVal = I.getArgOperand(0);
5738 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5739 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5740 Info.vol = false; // volatile stores with NEON intrinsics not supported
5741 Info.readMem = false;
5742 Info.writeMem = true;