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)."),
62 // This option should go away when Machine LICM is smart enough to hoist a
65 EnableARMVDUPsplat("arm-vdup-splat", cl::Hidden,
66 cl::desc("Generate VDUP for integer constant splats (TEMPORARY OPTION)."),
70 EnableARMLongCalls("arm-long-calls", cl::Hidden,
71 cl::desc("Generate calls via indirect call instructions"),
75 ARMInterworking("arm-interworking", cl::Hidden,
76 cl::desc("Enable / disable ARM interworking (for debugging only)"),
79 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
80 EVT PromotedBitwiseVT) {
81 if (VT != PromotedLdStVT) {
82 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
83 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
84 PromotedLdStVT.getSimpleVT());
86 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
87 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
88 PromotedLdStVT.getSimpleVT());
91 EVT ElemTy = VT.getVectorElementType();
92 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
93 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
94 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
95 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
96 if (ElemTy != MVT::i32) {
97 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
98 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
99 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
100 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
102 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
103 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
104 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
105 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
106 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
108 if (VT.isInteger()) {
109 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
110 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
112 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
113 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
115 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
117 // Promote all bit-wise operations.
118 if (VT.isInteger() && VT != PromotedBitwiseVT) {
119 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
120 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
121 PromotedBitwiseVT.getSimpleVT());
122 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
123 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
124 PromotedBitwiseVT.getSimpleVT());
125 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
126 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
127 PromotedBitwiseVT.getSimpleVT());
130 // Neon does not support vector divide/remainder operations.
131 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
132 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
133 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
134 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
135 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
136 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
139 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
140 addRegisterClass(VT, ARM::DPRRegisterClass);
141 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
144 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
145 addRegisterClass(VT, ARM::QPRRegisterClass);
146 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
149 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
150 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
151 return new TargetLoweringObjectFileMachO();
153 return new ARMElfTargetObjectFile();
156 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
157 : TargetLowering(TM, createTLOF(TM)) {
158 Subtarget = &TM.getSubtarget<ARMSubtarget>();
159 RegInfo = TM.getRegisterInfo();
160 Itins = TM.getInstrItineraryData();
162 if (Subtarget->isTargetDarwin()) {
163 // Uses VFP for Thumb libfuncs if available.
164 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
165 // Single-precision floating-point arithmetic.
166 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
167 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
168 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
169 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
171 // Double-precision floating-point arithmetic.
172 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
173 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
174 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
175 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
177 // Single-precision comparisons.
178 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
179 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
180 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
181 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
182 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
183 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
184 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
185 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
187 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
196 // Double-precision comparisons.
197 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
198 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
199 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
200 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
201 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
202 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
203 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
204 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
206 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
215 // Floating-point to integer conversions.
216 // i64 conversions are done via library routines even when generating VFP
217 // instructions, so use the same ones.
218 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
219 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
220 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
221 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
223 // Conversions between floating types.
224 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
225 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
227 // Integer to floating-point conversions.
228 // i64 conversions are done via library routines even when generating VFP
229 // instructions, so use the same ones.
230 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
231 // e.g., __floatunsidf vs. __floatunssidfvfp.
232 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
233 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
234 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
235 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
239 // These libcalls are not available in 32-bit.
240 setLibcallName(RTLIB::SHL_I128, 0);
241 setLibcallName(RTLIB::SRL_I128, 0);
242 setLibcallName(RTLIB::SRA_I128, 0);
244 if (Subtarget->isAAPCS_ABI()) {
245 // Double-precision floating-point arithmetic helper functions
246 // RTABI chapter 4.1.2, Table 2
247 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
248 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
249 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
250 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
251 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
252 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
253 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
254 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
256 // Double-precision floating-point comparison helper functions
257 // RTABI chapter 4.1.2, Table 3
258 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
259 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
260 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
261 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
262 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
263 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
264 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
265 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
266 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
267 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
268 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
269 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
270 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
271 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
272 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
273 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
274 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
275 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
276 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
277 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
278 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
279 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
280 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
281 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
283 // Single-precision floating-point arithmetic helper functions
284 // RTABI chapter 4.1.2, Table 4
285 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
286 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
287 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
288 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
289 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
294 // Single-precision floating-point comparison helper functions
295 // RTABI chapter 4.1.2, Table 5
296 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
297 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
298 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
299 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
300 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
301 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
302 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
303 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
304 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
305 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
306 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
307 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
308 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
309 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
310 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
311 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
312 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
313 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
314 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
315 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
316 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
317 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
318 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
319 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
321 // Floating-point to integer conversions.
322 // RTABI chapter 4.1.2, Table 6
323 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
324 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
325 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
326 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
327 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
328 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
329 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
330 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
331 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
340 // Conversions between floating types.
341 // RTABI chapter 4.1.2, Table 7
342 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
343 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
344 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
345 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
347 // Integer to floating-point conversions.
348 // RTABI chapter 4.1.2, Table 8
349 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
350 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
351 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
352 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
353 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
354 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
355 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
356 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
357 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
361 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
362 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
363 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
364 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
366 // Long long helper functions
367 // RTABI chapter 4.2, Table 9
368 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
369 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
370 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
371 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
372 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
373 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
374 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
381 // Integer division functions
382 // RTABI chapter 4.3.1
383 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
384 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
385 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
386 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
387 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
388 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
389 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
397 if (Subtarget->isThumb1Only())
398 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
400 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
401 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
402 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
403 if (!Subtarget->isFPOnlySP())
404 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
406 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
409 if (Subtarget->hasNEON()) {
410 addDRTypeForNEON(MVT::v2f32);
411 addDRTypeForNEON(MVT::v8i8);
412 addDRTypeForNEON(MVT::v4i16);
413 addDRTypeForNEON(MVT::v2i32);
414 addDRTypeForNEON(MVT::v1i64);
416 addQRTypeForNEON(MVT::v4f32);
417 addQRTypeForNEON(MVT::v2f64);
418 addQRTypeForNEON(MVT::v16i8);
419 addQRTypeForNEON(MVT::v8i16);
420 addQRTypeForNEON(MVT::v4i32);
421 addQRTypeForNEON(MVT::v2i64);
423 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
424 // neither Neon nor VFP support any arithmetic operations on it.
425 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
426 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
427 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
428 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
429 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
430 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
431 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
432 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
433 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
434 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
435 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
436 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
437 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
438 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
439 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
440 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
441 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
442 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
443 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
444 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
445 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
446 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
447 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
448 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
450 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
452 // Neon does not support some operations on v1i64 and v2i64 types.
453 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
454 // Custom handling for some quad-vector types to detect VMULL.
455 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
456 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
457 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
458 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
459 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
461 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
462 setTargetDAGCombine(ISD::SHL);
463 setTargetDAGCombine(ISD::SRL);
464 setTargetDAGCombine(ISD::SRA);
465 setTargetDAGCombine(ISD::SIGN_EXTEND);
466 setTargetDAGCombine(ISD::ZERO_EXTEND);
467 setTargetDAGCombine(ISD::ANY_EXTEND);
468 setTargetDAGCombine(ISD::SELECT_CC);
469 setTargetDAGCombine(ISD::BUILD_VECTOR);
472 computeRegisterProperties();
474 // ARM does not have f32 extending load.
475 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
477 // ARM does not have i1 sign extending load.
478 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
480 // ARM supports all 4 flavors of integer indexed load / store.
481 if (!Subtarget->isThumb1Only()) {
482 for (unsigned im = (unsigned)ISD::PRE_INC;
483 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
484 setIndexedLoadAction(im, MVT::i1, Legal);
485 setIndexedLoadAction(im, MVT::i8, Legal);
486 setIndexedLoadAction(im, MVT::i16, Legal);
487 setIndexedLoadAction(im, MVT::i32, Legal);
488 setIndexedStoreAction(im, MVT::i1, Legal);
489 setIndexedStoreAction(im, MVT::i8, Legal);
490 setIndexedStoreAction(im, MVT::i16, Legal);
491 setIndexedStoreAction(im, MVT::i32, Legal);
495 // i64 operation support.
496 if (Subtarget->isThumb1Only()) {
497 setOperationAction(ISD::MUL, MVT::i64, Expand);
498 setOperationAction(ISD::MULHU, MVT::i32, Expand);
499 setOperationAction(ISD::MULHS, MVT::i32, Expand);
500 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
501 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
503 setOperationAction(ISD::MUL, MVT::i64, Expand);
504 setOperationAction(ISD::MULHU, MVT::i32, Expand);
505 if (!Subtarget->hasV6Ops())
506 setOperationAction(ISD::MULHS, MVT::i32, Expand);
508 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
509 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
510 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
511 setOperationAction(ISD::SRL, MVT::i64, Custom);
512 setOperationAction(ISD::SRA, MVT::i64, Custom);
514 // ARM does not have ROTL.
515 setOperationAction(ISD::ROTL, MVT::i32, Expand);
516 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
517 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
518 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
519 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
521 // Only ARMv6 has BSWAP.
522 if (!Subtarget->hasV6Ops())
523 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
525 // These are expanded into libcalls.
526 if (!Subtarget->hasDivide()) {
527 // v7M has a hardware divider
528 setOperationAction(ISD::SDIV, MVT::i32, Expand);
529 setOperationAction(ISD::UDIV, MVT::i32, Expand);
531 setOperationAction(ISD::SREM, MVT::i32, Expand);
532 setOperationAction(ISD::UREM, MVT::i32, Expand);
533 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
534 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
536 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
537 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
538 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
539 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
540 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
542 setOperationAction(ISD::TRAP, MVT::Other, Legal);
544 // Use the default implementation.
545 setOperationAction(ISD::VASTART, MVT::Other, Custom);
546 setOperationAction(ISD::VAARG, MVT::Other, Expand);
547 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
548 setOperationAction(ISD::VAEND, MVT::Other, Expand);
549 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
550 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
551 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
552 // FIXME: Shouldn't need this, since no register is used, but the legalizer
553 // doesn't yet know how to not do that for SjLj.
554 setExceptionSelectorRegister(ARM::R0);
555 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
556 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
557 // the default expansion.
558 if (Subtarget->hasDataBarrier() ||
559 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
560 // membarrier needs custom lowering; the rest are legal and handled
562 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
564 // Set them all for expansion, which will force libcalls.
565 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
566 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
567 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
568 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
569 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
570 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
571 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
572 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
573 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
574 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
575 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
576 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
577 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
578 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
579 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
580 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
581 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
582 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
583 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
584 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
585 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
586 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
587 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
588 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
589 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
590 // Since the libcalls include locking, fold in the fences
591 setShouldFoldAtomicFences(true);
593 // 64-bit versions are always libcalls (for now)
594 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
595 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
596 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
597 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
598 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
599 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
600 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
601 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
603 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
604 if (!Subtarget->hasV6Ops()) {
605 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
606 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
608 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
610 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
611 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
612 // iff target supports vfp2.
613 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
614 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
617 // We want to custom lower some of our intrinsics.
618 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
619 if (Subtarget->isTargetDarwin()) {
620 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
621 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
624 setOperationAction(ISD::SETCC, MVT::i32, Expand);
625 setOperationAction(ISD::SETCC, MVT::f32, Expand);
626 setOperationAction(ISD::SETCC, MVT::f64, Expand);
627 setOperationAction(ISD::SELECT, MVT::i32, Custom);
628 setOperationAction(ISD::SELECT, MVT::f32, Custom);
629 setOperationAction(ISD::SELECT, MVT::f64, Custom);
630 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
631 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
632 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
634 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
635 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
636 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
637 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
638 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
640 // We don't support sin/cos/fmod/copysign/pow
641 setOperationAction(ISD::FSIN, MVT::f64, Expand);
642 setOperationAction(ISD::FSIN, MVT::f32, Expand);
643 setOperationAction(ISD::FCOS, MVT::f32, Expand);
644 setOperationAction(ISD::FCOS, MVT::f64, Expand);
645 setOperationAction(ISD::FREM, MVT::f64, Expand);
646 setOperationAction(ISD::FREM, MVT::f32, Expand);
647 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
648 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
649 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
651 setOperationAction(ISD::FPOW, MVT::f64, Expand);
652 setOperationAction(ISD::FPOW, MVT::f32, Expand);
654 // Various VFP goodness
655 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
656 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
657 if (Subtarget->hasVFP2()) {
658 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
659 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
660 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
661 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
663 // Special handling for half-precision FP.
664 if (!Subtarget->hasFP16()) {
665 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
666 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
670 // We have target-specific dag combine patterns for the following nodes:
671 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
672 setTargetDAGCombine(ISD::ADD);
673 setTargetDAGCombine(ISD::SUB);
674 setTargetDAGCombine(ISD::MUL);
676 if (Subtarget->hasV6T2Ops())
677 setTargetDAGCombine(ISD::OR);
679 setStackPointerRegisterToSaveRestore(ARM::SP);
681 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
682 setSchedulingPreference(Sched::RegPressure);
684 setSchedulingPreference(Sched::Hybrid);
686 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
688 // On ARM arguments smaller than 4 bytes are extended, so all arguments
689 // are at least 4 bytes aligned.
690 setMinStackArgumentAlignment(4);
692 benefitFromCodePlacementOpt = true;
695 std::pair<const TargetRegisterClass*, uint8_t>
696 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
697 const TargetRegisterClass *RRC = 0;
699 switch (VT.getSimpleVT().SimpleTy) {
701 return TargetLowering::findRepresentativeClass(VT);
702 // Use DPR as representative register class for all floating point
703 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
704 // the cost is 1 for both f32 and f64.
705 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
706 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
707 RRC = ARM::DPRRegisterClass;
709 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
710 case MVT::v4f32: case MVT::v2f64:
711 RRC = ARM::DPRRegisterClass;
715 RRC = ARM::DPRRegisterClass;
719 RRC = ARM::DPRRegisterClass;
723 return std::make_pair(RRC, Cost);
726 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
729 case ARMISD::Wrapper: return "ARMISD::Wrapper";
730 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
731 case ARMISD::CALL: return "ARMISD::CALL";
732 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
733 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
734 case ARMISD::tCALL: return "ARMISD::tCALL";
735 case ARMISD::BRCOND: return "ARMISD::BRCOND";
736 case ARMISD::BR_JT: return "ARMISD::BR_JT";
737 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
738 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
739 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
740 case ARMISD::CMP: return "ARMISD::CMP";
741 case ARMISD::CMPZ: return "ARMISD::CMPZ";
742 case ARMISD::CMPFP: return "ARMISD::CMPFP";
743 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
744 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
745 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
746 case ARMISD::CMOV: return "ARMISD::CMOV";
747 case ARMISD::CNEG: return "ARMISD::CNEG";
749 case ARMISD::RBIT: return "ARMISD::RBIT";
751 case ARMISD::FTOSI: return "ARMISD::FTOSI";
752 case ARMISD::FTOUI: return "ARMISD::FTOUI";
753 case ARMISD::SITOF: return "ARMISD::SITOF";
754 case ARMISD::UITOF: return "ARMISD::UITOF";
756 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
757 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
758 case ARMISD::RRX: return "ARMISD::RRX";
760 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
761 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
763 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
764 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
766 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
768 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
770 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
772 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
773 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
775 case ARMISD::VCEQ: return "ARMISD::VCEQ";
776 case ARMISD::VCGE: return "ARMISD::VCGE";
777 case ARMISD::VCGEU: return "ARMISD::VCGEU";
778 case ARMISD::VCGT: return "ARMISD::VCGT";
779 case ARMISD::VCGTU: return "ARMISD::VCGTU";
780 case ARMISD::VTST: return "ARMISD::VTST";
782 case ARMISD::VSHL: return "ARMISD::VSHL";
783 case ARMISD::VSHRs: return "ARMISD::VSHRs";
784 case ARMISD::VSHRu: return "ARMISD::VSHRu";
785 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
786 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
787 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
788 case ARMISD::VSHRN: return "ARMISD::VSHRN";
789 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
790 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
791 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
792 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
793 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
794 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
795 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
796 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
797 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
798 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
799 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
800 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
801 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
802 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
803 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
804 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
805 case ARMISD::VDUP: return "ARMISD::VDUP";
806 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
807 case ARMISD::VEXT: return "ARMISD::VEXT";
808 case ARMISD::VREV64: return "ARMISD::VREV64";
809 case ARMISD::VREV32: return "ARMISD::VREV32";
810 case ARMISD::VREV16: return "ARMISD::VREV16";
811 case ARMISD::VZIP: return "ARMISD::VZIP";
812 case ARMISD::VUZP: return "ARMISD::VUZP";
813 case ARMISD::VTRN: return "ARMISD::VTRN";
814 case ARMISD::VMULLs: return "ARMISD::VMULLs";
815 case ARMISD::VMULLu: return "ARMISD::VMULLu";
816 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
817 case ARMISD::FMAX: return "ARMISD::FMAX";
818 case ARMISD::FMIN: return "ARMISD::FMIN";
819 case ARMISD::BFI: return "ARMISD::BFI";
823 /// getRegClassFor - Return the register class that should be used for the
824 /// specified value type.
825 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
826 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
827 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
828 // load / store 4 to 8 consecutive D registers.
829 if (Subtarget->hasNEON()) {
830 if (VT == MVT::v4i64)
831 return ARM::QQPRRegisterClass;
832 else if (VT == MVT::v8i64)
833 return ARM::QQQQPRRegisterClass;
835 return TargetLowering::getRegClassFor(VT);
838 // Create a fast isel object.
840 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
841 return ARM::createFastISel(funcInfo);
844 /// getFunctionAlignment - Return the Log2 alignment of this function.
845 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
846 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
849 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
850 /// be used for loads / stores from the global.
851 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
852 return (Subtarget->isThumb1Only() ? 127 : 4095);
855 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
856 unsigned NumVals = N->getNumValues();
858 return Sched::RegPressure;
860 for (unsigned i = 0; i != NumVals; ++i) {
861 EVT VT = N->getValueType(i);
862 if (VT.isFloatingPoint() || VT.isVector())
863 return Sched::Latency;
866 if (!N->isMachineOpcode())
867 return Sched::RegPressure;
869 // Load are scheduled for latency even if there instruction itinerary
871 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
872 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
874 return Sched::Latency;
876 if (!Itins->isEmpty() && Itins->getStageLatency(TID.getSchedClass()) > 2)
877 return Sched::Latency;
878 return Sched::RegPressure;
882 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
883 MachineFunction &MF) const {
884 switch (RC->getID()) {
887 case ARM::tGPRRegClassID:
888 return RegInfo->hasFP(MF) ? 4 : 5;
889 case ARM::GPRRegClassID: {
890 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
891 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
893 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
894 case ARM::DPRRegClassID:
899 //===----------------------------------------------------------------------===//
901 //===----------------------------------------------------------------------===//
903 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
904 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
906 default: llvm_unreachable("Unknown condition code!");
907 case ISD::SETNE: return ARMCC::NE;
908 case ISD::SETEQ: return ARMCC::EQ;
909 case ISD::SETGT: return ARMCC::GT;
910 case ISD::SETGE: return ARMCC::GE;
911 case ISD::SETLT: return ARMCC::LT;
912 case ISD::SETLE: return ARMCC::LE;
913 case ISD::SETUGT: return ARMCC::HI;
914 case ISD::SETUGE: return ARMCC::HS;
915 case ISD::SETULT: return ARMCC::LO;
916 case ISD::SETULE: return ARMCC::LS;
920 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
921 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
922 ARMCC::CondCodes &CondCode2) {
923 CondCode2 = ARMCC::AL;
925 default: llvm_unreachable("Unknown FP condition!");
927 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
929 case ISD::SETOGT: CondCode = ARMCC::GT; break;
931 case ISD::SETOGE: CondCode = ARMCC::GE; break;
932 case ISD::SETOLT: CondCode = ARMCC::MI; break;
933 case ISD::SETOLE: CondCode = ARMCC::LS; break;
934 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
935 case ISD::SETO: CondCode = ARMCC::VC; break;
936 case ISD::SETUO: CondCode = ARMCC::VS; break;
937 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
938 case ISD::SETUGT: CondCode = ARMCC::HI; break;
939 case ISD::SETUGE: CondCode = ARMCC::PL; break;
941 case ISD::SETULT: CondCode = ARMCC::LT; break;
943 case ISD::SETULE: CondCode = ARMCC::LE; break;
945 case ISD::SETUNE: CondCode = ARMCC::NE; break;
949 //===----------------------------------------------------------------------===//
950 // Calling Convention Implementation
951 //===----------------------------------------------------------------------===//
953 #include "ARMGenCallingConv.inc"
955 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
956 /// given CallingConvention value.
957 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
959 bool isVarArg) const {
962 llvm_unreachable("Unsupported calling convention");
964 case CallingConv::Fast:
965 // Use target triple & subtarget features to do actual dispatch.
966 if (Subtarget->isAAPCS_ABI()) {
967 if (Subtarget->hasVFP2() &&
968 FloatABIType == FloatABI::Hard && !isVarArg)
969 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
971 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
973 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
974 case CallingConv::ARM_AAPCS_VFP:
975 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
976 case CallingConv::ARM_AAPCS:
977 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
978 case CallingConv::ARM_APCS:
979 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
983 /// LowerCallResult - Lower the result values of a call into the
984 /// appropriate copies out of appropriate physical registers.
986 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
987 CallingConv::ID CallConv, bool isVarArg,
988 const SmallVectorImpl<ISD::InputArg> &Ins,
989 DebugLoc dl, SelectionDAG &DAG,
990 SmallVectorImpl<SDValue> &InVals) const {
992 // Assign locations to each value returned by this call.
993 SmallVector<CCValAssign, 16> RVLocs;
994 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
995 RVLocs, *DAG.getContext());
996 CCInfo.AnalyzeCallResult(Ins,
997 CCAssignFnForNode(CallConv, /* Return*/ true,
1000 // Copy all of the result registers out of their specified physreg.
1001 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1002 CCValAssign VA = RVLocs[i];
1005 if (VA.needsCustom()) {
1006 // Handle f64 or half of a v2f64.
1007 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1009 Chain = Lo.getValue(1);
1010 InFlag = Lo.getValue(2);
1011 VA = RVLocs[++i]; // skip ahead to next loc
1012 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1014 Chain = Hi.getValue(1);
1015 InFlag = Hi.getValue(2);
1016 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1018 if (VA.getLocVT() == MVT::v2f64) {
1019 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1020 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1021 DAG.getConstant(0, MVT::i32));
1023 VA = RVLocs[++i]; // skip ahead to next loc
1024 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1025 Chain = Lo.getValue(1);
1026 InFlag = Lo.getValue(2);
1027 VA = RVLocs[++i]; // skip ahead to next loc
1028 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1029 Chain = Hi.getValue(1);
1030 InFlag = Hi.getValue(2);
1031 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1032 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1033 DAG.getConstant(1, MVT::i32));
1036 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1038 Chain = Val.getValue(1);
1039 InFlag = Val.getValue(2);
1042 switch (VA.getLocInfo()) {
1043 default: llvm_unreachable("Unknown loc info!");
1044 case CCValAssign::Full: break;
1045 case CCValAssign::BCvt:
1046 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1050 InVals.push_back(Val);
1056 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1057 /// by "Src" to address "Dst" of size "Size". Alignment information is
1058 /// specified by the specific parameter attribute. The copy will be passed as
1059 /// a byval function parameter.
1060 /// Sometimes what we are copying is the end of a larger object, the part that
1061 /// does not fit in registers.
1063 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1064 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1066 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1067 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1068 /*isVolatile=*/false, /*AlwaysInline=*/false,
1069 MachinePointerInfo(0), MachinePointerInfo(0));
1072 /// LowerMemOpCallTo - Store the argument to the stack.
1074 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1075 SDValue StackPtr, SDValue Arg,
1076 DebugLoc dl, SelectionDAG &DAG,
1077 const CCValAssign &VA,
1078 ISD::ArgFlagsTy Flags) const {
1079 unsigned LocMemOffset = VA.getLocMemOffset();
1080 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1081 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1082 if (Flags.isByVal())
1083 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1085 return DAG.getStore(Chain, dl, Arg, PtrOff,
1086 MachinePointerInfo::getStack(LocMemOffset),
1090 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1091 SDValue Chain, SDValue &Arg,
1092 RegsToPassVector &RegsToPass,
1093 CCValAssign &VA, CCValAssign &NextVA,
1095 SmallVector<SDValue, 8> &MemOpChains,
1096 ISD::ArgFlagsTy Flags) const {
1098 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1099 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1100 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1102 if (NextVA.isRegLoc())
1103 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1105 assert(NextVA.isMemLoc());
1106 if (StackPtr.getNode() == 0)
1107 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1109 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1115 /// LowerCall - Lowering a call into a callseq_start <-
1116 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1119 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1120 CallingConv::ID CallConv, bool isVarArg,
1122 const SmallVectorImpl<ISD::OutputArg> &Outs,
1123 const SmallVectorImpl<SDValue> &OutVals,
1124 const SmallVectorImpl<ISD::InputArg> &Ins,
1125 DebugLoc dl, SelectionDAG &DAG,
1126 SmallVectorImpl<SDValue> &InVals) const {
1127 MachineFunction &MF = DAG.getMachineFunction();
1128 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1129 bool IsSibCall = false;
1130 // Temporarily disable tail calls so things don't break.
1131 if (!EnableARMTailCalls)
1134 // Check if it's really possible to do a tail call.
1135 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1136 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1137 Outs, OutVals, Ins, DAG);
1138 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1139 // detected sibcalls.
1146 // Analyze operands of the call, assigning locations to each operand.
1147 SmallVector<CCValAssign, 16> ArgLocs;
1148 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1150 CCInfo.AnalyzeCallOperands(Outs,
1151 CCAssignFnForNode(CallConv, /* Return*/ false,
1154 // Get a count of how many bytes are to be pushed on the stack.
1155 unsigned NumBytes = CCInfo.getNextStackOffset();
1157 // For tail calls, memory operands are available in our caller's stack.
1161 // Adjust the stack pointer for the new arguments...
1162 // These operations are automatically eliminated by the prolog/epilog pass
1164 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1166 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1168 RegsToPassVector RegsToPass;
1169 SmallVector<SDValue, 8> MemOpChains;
1171 // Walk the register/memloc assignments, inserting copies/loads. In the case
1172 // of tail call optimization, arguments are handled later.
1173 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1175 ++i, ++realArgIdx) {
1176 CCValAssign &VA = ArgLocs[i];
1177 SDValue Arg = OutVals[realArgIdx];
1178 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1180 // Promote the value if needed.
1181 switch (VA.getLocInfo()) {
1182 default: llvm_unreachable("Unknown loc info!");
1183 case CCValAssign::Full: break;
1184 case CCValAssign::SExt:
1185 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1187 case CCValAssign::ZExt:
1188 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1190 case CCValAssign::AExt:
1191 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1193 case CCValAssign::BCvt:
1194 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1198 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1199 if (VA.needsCustom()) {
1200 if (VA.getLocVT() == MVT::v2f64) {
1201 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1202 DAG.getConstant(0, MVT::i32));
1203 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1204 DAG.getConstant(1, MVT::i32));
1206 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1207 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1209 VA = ArgLocs[++i]; // skip ahead to next loc
1210 if (VA.isRegLoc()) {
1211 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1212 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1214 assert(VA.isMemLoc());
1216 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1217 dl, DAG, VA, Flags));
1220 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1221 StackPtr, MemOpChains, Flags);
1223 } else if (VA.isRegLoc()) {
1224 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1225 } else if (!IsSibCall) {
1226 assert(VA.isMemLoc());
1228 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1229 dl, DAG, VA, Flags));
1233 if (!MemOpChains.empty())
1234 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1235 &MemOpChains[0], MemOpChains.size());
1237 // Build a sequence of copy-to-reg nodes chained together with token chain
1238 // and flag operands which copy the outgoing args into the appropriate regs.
1240 // Tail call byval lowering might overwrite argument registers so in case of
1241 // tail call optimization the copies to registers are lowered later.
1243 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1244 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1245 RegsToPass[i].second, InFlag);
1246 InFlag = Chain.getValue(1);
1249 // For tail calls lower the arguments to the 'real' stack slot.
1251 // Force all the incoming stack arguments to be loaded from the stack
1252 // before any new outgoing arguments are stored to the stack, because the
1253 // outgoing stack slots may alias the incoming argument stack slots, and
1254 // the alias isn't otherwise explicit. This is slightly more conservative
1255 // than necessary, because it means that each store effectively depends
1256 // on every argument instead of just those arguments it would clobber.
1258 // Do not flag preceeding copytoreg stuff together with the following stuff.
1260 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1261 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1262 RegsToPass[i].second, InFlag);
1263 InFlag = Chain.getValue(1);
1268 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1269 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1270 // node so that legalize doesn't hack it.
1271 bool isDirect = false;
1272 bool isARMFunc = false;
1273 bool isLocalARMFunc = false;
1274 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1276 if (EnableARMLongCalls) {
1277 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1278 && "long-calls with non-static relocation model!");
1279 // Handle a global address or an external symbol. If it's not one of
1280 // those, the target's already in a register, so we don't need to do
1282 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1283 const GlobalValue *GV = G->getGlobal();
1284 // Create a constant pool entry for the callee address
1285 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1286 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1289 // Get the address of the callee into a register
1290 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1291 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1292 Callee = DAG.getLoad(getPointerTy(), dl,
1293 DAG.getEntryNode(), CPAddr,
1294 MachinePointerInfo::getConstantPool(),
1296 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1297 const char *Sym = S->getSymbol();
1299 // Create a constant pool entry for the callee address
1300 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1301 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1302 Sym, ARMPCLabelIndex, 0);
1303 // Get the address of the callee into a register
1304 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1305 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1306 Callee = DAG.getLoad(getPointerTy(), dl,
1307 DAG.getEntryNode(), CPAddr,
1308 MachinePointerInfo::getConstantPool(),
1311 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1312 const GlobalValue *GV = G->getGlobal();
1314 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1315 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1316 getTargetMachine().getRelocationModel() != Reloc::Static;
1317 isARMFunc = !Subtarget->isThumb() || isStub;
1318 // ARM call to a local ARM function is predicable.
1319 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1320 // tBX takes a register source operand.
1321 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1322 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1323 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1326 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1327 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1328 Callee = DAG.getLoad(getPointerTy(), dl,
1329 DAG.getEntryNode(), CPAddr,
1330 MachinePointerInfo::getConstantPool(),
1332 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1333 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1334 getPointerTy(), Callee, PICLabel);
1336 // On ELF targets for PIC code, direct calls should go through the PLT
1337 unsigned OpFlags = 0;
1338 if (Subtarget->isTargetELF() &&
1339 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1340 OpFlags = ARMII::MO_PLT;
1341 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1343 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1345 bool isStub = Subtarget->isTargetDarwin() &&
1346 getTargetMachine().getRelocationModel() != Reloc::Static;
1347 isARMFunc = !Subtarget->isThumb() || isStub;
1348 // tBX takes a register source operand.
1349 const char *Sym = S->getSymbol();
1350 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1351 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1352 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1353 Sym, ARMPCLabelIndex, 4);
1354 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1355 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1356 Callee = DAG.getLoad(getPointerTy(), dl,
1357 DAG.getEntryNode(), CPAddr,
1358 MachinePointerInfo::getConstantPool(),
1360 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1361 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1362 getPointerTy(), Callee, PICLabel);
1364 unsigned OpFlags = 0;
1365 // On ELF targets for PIC code, direct calls should go through the PLT
1366 if (Subtarget->isTargetELF() &&
1367 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1368 OpFlags = ARMII::MO_PLT;
1369 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1373 // FIXME: handle tail calls differently.
1375 if (Subtarget->isThumb()) {
1376 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1377 CallOpc = ARMISD::CALL_NOLINK;
1379 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1381 CallOpc = (isDirect || Subtarget->hasV5TOps())
1382 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1383 : ARMISD::CALL_NOLINK;
1386 std::vector<SDValue> Ops;
1387 Ops.push_back(Chain);
1388 Ops.push_back(Callee);
1390 // Add argument registers to the end of the list so that they are known live
1392 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1393 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1394 RegsToPass[i].second.getValueType()));
1396 if (InFlag.getNode())
1397 Ops.push_back(InFlag);
1399 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1401 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1403 // Returns a chain and a flag for retval copy to use.
1404 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1405 InFlag = Chain.getValue(1);
1407 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1408 DAG.getIntPtrConstant(0, true), InFlag);
1410 InFlag = Chain.getValue(1);
1412 // Handle result values, copying them out of physregs into vregs that we
1414 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1418 /// MatchingStackOffset - Return true if the given stack call argument is
1419 /// already available in the same position (relatively) of the caller's
1420 /// incoming argument stack.
1422 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1423 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1424 const ARMInstrInfo *TII) {
1425 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1427 if (Arg.getOpcode() == ISD::CopyFromReg) {
1428 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1429 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1431 MachineInstr *Def = MRI->getVRegDef(VR);
1434 if (!Flags.isByVal()) {
1435 if (!TII->isLoadFromStackSlot(Def, FI))
1440 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1441 if (Flags.isByVal())
1442 // ByVal argument is passed in as a pointer but it's now being
1443 // dereferenced. e.g.
1444 // define @foo(%struct.X* %A) {
1445 // tail call @bar(%struct.X* byval %A)
1448 SDValue Ptr = Ld->getBasePtr();
1449 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1452 FI = FINode->getIndex();
1456 assert(FI != INT_MAX);
1457 if (!MFI->isFixedObjectIndex(FI))
1459 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1462 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1463 /// for tail call optimization. Targets which want to do tail call
1464 /// optimization should implement this function.
1466 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1467 CallingConv::ID CalleeCC,
1469 bool isCalleeStructRet,
1470 bool isCallerStructRet,
1471 const SmallVectorImpl<ISD::OutputArg> &Outs,
1472 const SmallVectorImpl<SDValue> &OutVals,
1473 const SmallVectorImpl<ISD::InputArg> &Ins,
1474 SelectionDAG& DAG) const {
1475 const Function *CallerF = DAG.getMachineFunction().getFunction();
1476 CallingConv::ID CallerCC = CallerF->getCallingConv();
1477 bool CCMatch = CallerCC == CalleeCC;
1479 // Look for obvious safe cases to perform tail call optimization that do not
1480 // require ABI changes. This is what gcc calls sibcall.
1482 // Do not sibcall optimize vararg calls unless the call site is not passing
1484 if (isVarArg && !Outs.empty())
1487 // Also avoid sibcall optimization if either caller or callee uses struct
1488 // return semantics.
1489 if (isCalleeStructRet || isCallerStructRet)
1492 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1493 // emitEpilogue is not ready for them.
1494 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1495 // LR. This means if we need to reload LR, it takes an extra instructions,
1496 // which outweighs the value of the tail call; but here we don't know yet
1497 // whether LR is going to be used. Probably the right approach is to
1498 // generate the tail call here and turn it back into CALL/RET in
1499 // emitEpilogue if LR is used.
1500 if (Subtarget->isThumb1Only())
1503 // For the moment, we can only do this to functions defined in this
1504 // compilation, or to indirect calls. A Thumb B to an ARM function,
1505 // or vice versa, is not easily fixed up in the linker unlike BL.
1506 // (We could do this by loading the address of the callee into a register;
1507 // that is an extra instruction over the direct call and burns a register
1508 // as well, so is not likely to be a win.)
1510 // It might be safe to remove this restriction on non-Darwin.
1512 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1513 // but we need to make sure there are enough registers; the only valid
1514 // registers are the 4 used for parameters. We don't currently do this
1516 if (isa<ExternalSymbolSDNode>(Callee))
1519 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1520 const GlobalValue *GV = G->getGlobal();
1521 if (GV->isDeclaration() || GV->isWeakForLinker())
1525 // If the calling conventions do not match, then we'd better make sure the
1526 // results are returned in the same way as what the caller expects.
1528 SmallVector<CCValAssign, 16> RVLocs1;
1529 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1530 RVLocs1, *DAG.getContext());
1531 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1533 SmallVector<CCValAssign, 16> RVLocs2;
1534 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1535 RVLocs2, *DAG.getContext());
1536 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1538 if (RVLocs1.size() != RVLocs2.size())
1540 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1541 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1543 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1545 if (RVLocs1[i].isRegLoc()) {
1546 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1549 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1555 // If the callee takes no arguments then go on to check the results of the
1557 if (!Outs.empty()) {
1558 // Check if stack adjustment is needed. For now, do not do this if any
1559 // argument is passed on the stack.
1560 SmallVector<CCValAssign, 16> ArgLocs;
1561 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1562 ArgLocs, *DAG.getContext());
1563 CCInfo.AnalyzeCallOperands(Outs,
1564 CCAssignFnForNode(CalleeCC, false, isVarArg));
1565 if (CCInfo.getNextStackOffset()) {
1566 MachineFunction &MF = DAG.getMachineFunction();
1568 // Check if the arguments are already laid out in the right way as
1569 // the caller's fixed stack objects.
1570 MachineFrameInfo *MFI = MF.getFrameInfo();
1571 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1572 const ARMInstrInfo *TII =
1573 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1574 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1576 ++i, ++realArgIdx) {
1577 CCValAssign &VA = ArgLocs[i];
1578 EVT RegVT = VA.getLocVT();
1579 SDValue Arg = OutVals[realArgIdx];
1580 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1581 if (VA.getLocInfo() == CCValAssign::Indirect)
1583 if (VA.needsCustom()) {
1584 // f64 and vector types are split into multiple registers or
1585 // register/stack-slot combinations. The types will not match
1586 // the registers; give up on memory f64 refs until we figure
1587 // out what to do about this.
1590 if (!ArgLocs[++i].isRegLoc())
1592 if (RegVT == MVT::v2f64) {
1593 if (!ArgLocs[++i].isRegLoc())
1595 if (!ArgLocs[++i].isRegLoc())
1598 } else if (!VA.isRegLoc()) {
1599 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1611 ARMTargetLowering::LowerReturn(SDValue Chain,
1612 CallingConv::ID CallConv, bool isVarArg,
1613 const SmallVectorImpl<ISD::OutputArg> &Outs,
1614 const SmallVectorImpl<SDValue> &OutVals,
1615 DebugLoc dl, SelectionDAG &DAG) const {
1617 // CCValAssign - represent the assignment of the return value to a location.
1618 SmallVector<CCValAssign, 16> RVLocs;
1620 // CCState - Info about the registers and stack slots.
1621 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1624 // Analyze outgoing return values.
1625 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1628 // If this is the first return lowered for this function, add
1629 // the regs to the liveout set for the function.
1630 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1631 for (unsigned i = 0; i != RVLocs.size(); ++i)
1632 if (RVLocs[i].isRegLoc())
1633 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1638 // Copy the result values into the output registers.
1639 for (unsigned i = 0, realRVLocIdx = 0;
1641 ++i, ++realRVLocIdx) {
1642 CCValAssign &VA = RVLocs[i];
1643 assert(VA.isRegLoc() && "Can only return in registers!");
1645 SDValue Arg = OutVals[realRVLocIdx];
1647 switch (VA.getLocInfo()) {
1648 default: llvm_unreachable("Unknown loc info!");
1649 case CCValAssign::Full: break;
1650 case CCValAssign::BCvt:
1651 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1655 if (VA.needsCustom()) {
1656 if (VA.getLocVT() == MVT::v2f64) {
1657 // Extract the first half and return it in two registers.
1658 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1659 DAG.getConstant(0, MVT::i32));
1660 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1661 DAG.getVTList(MVT::i32, MVT::i32), Half);
1663 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1664 Flag = Chain.getValue(1);
1665 VA = RVLocs[++i]; // skip ahead to next loc
1666 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1667 HalfGPRs.getValue(1), Flag);
1668 Flag = Chain.getValue(1);
1669 VA = RVLocs[++i]; // skip ahead to next loc
1671 // Extract the 2nd half and fall through to handle it as an f64 value.
1672 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1673 DAG.getConstant(1, MVT::i32));
1675 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1677 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1678 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1679 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1680 Flag = Chain.getValue(1);
1681 VA = RVLocs[++i]; // skip ahead to next loc
1682 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1685 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1687 // Guarantee that all emitted copies are
1688 // stuck together, avoiding something bad.
1689 Flag = Chain.getValue(1);
1694 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1696 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1701 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1702 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1703 // one of the above mentioned nodes. It has to be wrapped because otherwise
1704 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1705 // be used to form addressing mode. These wrapped nodes will be selected
1707 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1708 EVT PtrVT = Op.getValueType();
1709 // FIXME there is no actual debug info here
1710 DebugLoc dl = Op.getDebugLoc();
1711 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1713 if (CP->isMachineConstantPoolEntry())
1714 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1715 CP->getAlignment());
1717 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1718 CP->getAlignment());
1719 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1722 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1723 return MachineJumpTableInfo::EK_Inline;
1726 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1727 SelectionDAG &DAG) const {
1728 MachineFunction &MF = DAG.getMachineFunction();
1729 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1730 unsigned ARMPCLabelIndex = 0;
1731 DebugLoc DL = Op.getDebugLoc();
1732 EVT PtrVT = getPointerTy();
1733 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1734 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1736 if (RelocM == Reloc::Static) {
1737 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1739 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1740 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1741 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1742 ARMCP::CPBlockAddress,
1744 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1746 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1747 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1748 MachinePointerInfo::getConstantPool(),
1750 if (RelocM == Reloc::Static)
1752 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1753 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1756 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1758 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1759 SelectionDAG &DAG) const {
1760 DebugLoc dl = GA->getDebugLoc();
1761 EVT PtrVT = getPointerTy();
1762 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1763 MachineFunction &MF = DAG.getMachineFunction();
1764 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1765 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1766 ARMConstantPoolValue *CPV =
1767 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1768 ARMCP::CPValue, PCAdj, "tlsgd", true);
1769 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1770 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1771 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1772 MachinePointerInfo::getConstantPool(),
1774 SDValue Chain = Argument.getValue(1);
1776 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1777 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1779 // call __tls_get_addr.
1782 Entry.Node = Argument;
1783 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1784 Args.push_back(Entry);
1785 // FIXME: is there useful debug info available here?
1786 std::pair<SDValue, SDValue> CallResult =
1787 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1788 false, false, false, false,
1789 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1790 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1791 return CallResult.first;
1794 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1795 // "local exec" model.
1797 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1798 SelectionDAG &DAG) const {
1799 const GlobalValue *GV = GA->getGlobal();
1800 DebugLoc dl = GA->getDebugLoc();
1802 SDValue Chain = DAG.getEntryNode();
1803 EVT PtrVT = getPointerTy();
1804 // Get the Thread Pointer
1805 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1807 if (GV->isDeclaration()) {
1808 MachineFunction &MF = DAG.getMachineFunction();
1809 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1810 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1811 // Initial exec model.
1812 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1813 ARMConstantPoolValue *CPV =
1814 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1815 ARMCP::CPValue, PCAdj, "gottpoff", true);
1816 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1817 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1818 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1819 MachinePointerInfo::getConstantPool(),
1821 Chain = Offset.getValue(1);
1823 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1824 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1826 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1827 MachinePointerInfo::getConstantPool(),
1831 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1832 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1833 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1834 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1835 MachinePointerInfo::getConstantPool(),
1839 // The address of the thread local variable is the add of the thread
1840 // pointer with the offset of the variable.
1841 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1845 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1846 // TODO: implement the "local dynamic" model
1847 assert(Subtarget->isTargetELF() &&
1848 "TLS not implemented for non-ELF targets");
1849 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1850 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1851 // otherwise use the "Local Exec" TLS Model
1852 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1853 return LowerToTLSGeneralDynamicModel(GA, DAG);
1855 return LowerToTLSExecModels(GA, DAG);
1858 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1859 SelectionDAG &DAG) const {
1860 EVT PtrVT = getPointerTy();
1861 DebugLoc dl = Op.getDebugLoc();
1862 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1863 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1864 if (RelocM == Reloc::PIC_) {
1865 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1866 ARMConstantPoolValue *CPV =
1867 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1868 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1869 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1870 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1872 MachinePointerInfo::getConstantPool(),
1874 SDValue Chain = Result.getValue(1);
1875 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1876 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1878 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1879 MachinePointerInfo::getGOT(), false, false, 0);
1882 // If we have T2 ops, we can materialize the address directly via movt/movw
1883 // pair. This is always cheaper.
1884 if (Subtarget->useMovt()) {
1885 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1886 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1888 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1889 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1890 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1891 MachinePointerInfo::getConstantPool(),
1897 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1898 SelectionDAG &DAG) const {
1899 MachineFunction &MF = DAG.getMachineFunction();
1900 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1901 unsigned ARMPCLabelIndex = 0;
1902 EVT PtrVT = getPointerTy();
1903 DebugLoc dl = Op.getDebugLoc();
1904 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1905 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1907 if (RelocM == Reloc::Static)
1908 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1910 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1911 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1912 ARMConstantPoolValue *CPV =
1913 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1914 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1916 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1918 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1919 MachinePointerInfo::getConstantPool(),
1921 SDValue Chain = Result.getValue(1);
1923 if (RelocM == Reloc::PIC_) {
1924 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1925 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1928 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1929 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1935 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1936 SelectionDAG &DAG) const {
1937 assert(Subtarget->isTargetELF() &&
1938 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1939 MachineFunction &MF = DAG.getMachineFunction();
1940 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1941 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1942 EVT PtrVT = getPointerTy();
1943 DebugLoc dl = Op.getDebugLoc();
1944 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1945 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1946 "_GLOBAL_OFFSET_TABLE_",
1947 ARMPCLabelIndex, PCAdj);
1948 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1949 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1950 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1951 MachinePointerInfo::getConstantPool(),
1953 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1954 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1958 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1959 DebugLoc dl = Op.getDebugLoc();
1960 SDValue Val = DAG.getConstant(0, MVT::i32);
1961 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1962 Op.getOperand(1), Val);
1966 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1967 DebugLoc dl = Op.getDebugLoc();
1968 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1969 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1973 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1974 const ARMSubtarget *Subtarget) const {
1975 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1976 DebugLoc dl = Op.getDebugLoc();
1978 default: return SDValue(); // Don't custom lower most intrinsics.
1979 case Intrinsic::arm_thread_pointer: {
1980 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1981 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1983 case Intrinsic::eh_sjlj_lsda: {
1984 MachineFunction &MF = DAG.getMachineFunction();
1985 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1986 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1987 EVT PtrVT = getPointerTy();
1988 DebugLoc dl = Op.getDebugLoc();
1989 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1991 unsigned PCAdj = (RelocM != Reloc::PIC_)
1992 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1993 ARMConstantPoolValue *CPV =
1994 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1995 ARMCP::CPLSDA, PCAdj);
1996 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1997 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1999 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2000 MachinePointerInfo::getConstantPool(),
2003 if (RelocM == Reloc::PIC_) {
2004 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2005 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2012 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2013 const ARMSubtarget *Subtarget) {
2014 DebugLoc dl = Op.getDebugLoc();
2015 SDValue Op5 = Op.getOperand(5);
2016 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
2017 // Some subtargets which have dmb and dsb instructions can handle barriers
2018 // directly. Some ARMv6 cpus can support them with the help of mcr
2019 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
2021 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
2022 if (Subtarget->hasDataBarrier())
2023 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2025 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2026 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2027 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2028 DAG.getConstant(0, MVT::i32));
2032 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2033 MachineFunction &MF = DAG.getMachineFunction();
2034 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2036 // vastart just stores the address of the VarArgsFrameIndex slot into the
2037 // memory location argument.
2038 DebugLoc dl = Op.getDebugLoc();
2039 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2040 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2041 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2042 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2043 MachinePointerInfo(SV), false, false, 0);
2047 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2048 SDValue &Root, SelectionDAG &DAG,
2049 DebugLoc dl) const {
2050 MachineFunction &MF = DAG.getMachineFunction();
2051 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2053 TargetRegisterClass *RC;
2054 if (AFI->isThumb1OnlyFunction())
2055 RC = ARM::tGPRRegisterClass;
2057 RC = ARM::GPRRegisterClass;
2059 // Transform the arguments stored in physical registers into virtual ones.
2060 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2061 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2064 if (NextVA.isMemLoc()) {
2065 MachineFrameInfo *MFI = MF.getFrameInfo();
2066 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2068 // Create load node to retrieve arguments from the stack.
2069 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2070 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2071 MachinePointerInfo::getFixedStack(FI),
2074 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2075 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2078 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2082 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2083 CallingConv::ID CallConv, bool isVarArg,
2084 const SmallVectorImpl<ISD::InputArg>
2086 DebugLoc dl, SelectionDAG &DAG,
2087 SmallVectorImpl<SDValue> &InVals)
2090 MachineFunction &MF = DAG.getMachineFunction();
2091 MachineFrameInfo *MFI = MF.getFrameInfo();
2093 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2095 // Assign locations to all of the incoming arguments.
2096 SmallVector<CCValAssign, 16> ArgLocs;
2097 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2099 CCInfo.AnalyzeFormalArguments(Ins,
2100 CCAssignFnForNode(CallConv, /* Return*/ false,
2103 SmallVector<SDValue, 16> ArgValues;
2105 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2106 CCValAssign &VA = ArgLocs[i];
2108 // Arguments stored in registers.
2109 if (VA.isRegLoc()) {
2110 EVT RegVT = VA.getLocVT();
2113 if (VA.needsCustom()) {
2114 // f64 and vector types are split up into multiple registers or
2115 // combinations of registers and stack slots.
2116 if (VA.getLocVT() == MVT::v2f64) {
2117 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2119 VA = ArgLocs[++i]; // skip ahead to next loc
2121 if (VA.isMemLoc()) {
2122 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2123 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2124 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2125 MachinePointerInfo::getFixedStack(FI),
2128 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2131 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2132 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2133 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2134 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2135 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2137 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2140 TargetRegisterClass *RC;
2142 if (RegVT == MVT::f32)
2143 RC = ARM::SPRRegisterClass;
2144 else if (RegVT == MVT::f64)
2145 RC = ARM::DPRRegisterClass;
2146 else if (RegVT == MVT::v2f64)
2147 RC = ARM::QPRRegisterClass;
2148 else if (RegVT == MVT::i32)
2149 RC = (AFI->isThumb1OnlyFunction() ?
2150 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2152 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2154 // Transform the arguments in physical registers into virtual ones.
2155 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2156 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2159 // If this is an 8 or 16-bit value, it is really passed promoted
2160 // to 32 bits. Insert an assert[sz]ext to capture this, then
2161 // truncate to the right size.
2162 switch (VA.getLocInfo()) {
2163 default: llvm_unreachable("Unknown loc info!");
2164 case CCValAssign::Full: break;
2165 case CCValAssign::BCvt:
2166 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2168 case CCValAssign::SExt:
2169 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2170 DAG.getValueType(VA.getValVT()));
2171 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2173 case CCValAssign::ZExt:
2174 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2175 DAG.getValueType(VA.getValVT()));
2176 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2180 InVals.push_back(ArgValue);
2182 } else { // VA.isRegLoc()
2185 assert(VA.isMemLoc());
2186 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2188 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2189 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2191 // Create load nodes to retrieve arguments from the stack.
2192 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2193 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2194 MachinePointerInfo::getFixedStack(FI),
2201 static const unsigned GPRArgRegs[] = {
2202 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2205 unsigned NumGPRs = CCInfo.getFirstUnallocated
2206 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2208 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2209 unsigned VARegSize = (4 - NumGPRs) * 4;
2210 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2211 unsigned ArgOffset = CCInfo.getNextStackOffset();
2212 if (VARegSaveSize) {
2213 // If this function is vararg, store any remaining integer argument regs
2214 // to their spots on the stack so that they may be loaded by deferencing
2215 // the result of va_next.
2216 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2217 AFI->setVarArgsFrameIndex(
2218 MFI->CreateFixedObject(VARegSaveSize,
2219 ArgOffset + VARegSaveSize - VARegSize,
2221 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2224 SmallVector<SDValue, 4> MemOps;
2225 for (; NumGPRs < 4; ++NumGPRs) {
2226 TargetRegisterClass *RC;
2227 if (AFI->isThumb1OnlyFunction())
2228 RC = ARM::tGPRRegisterClass;
2230 RC = ARM::GPRRegisterClass;
2232 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2233 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2235 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2236 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2238 MemOps.push_back(Store);
2239 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2240 DAG.getConstant(4, getPointerTy()));
2242 if (!MemOps.empty())
2243 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2244 &MemOps[0], MemOps.size());
2246 // This will point to the next argument passed via stack.
2247 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2253 /// isFloatingPointZero - Return true if this is +0.0.
2254 static bool isFloatingPointZero(SDValue Op) {
2255 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2256 return CFP->getValueAPF().isPosZero();
2257 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2258 // Maybe this has already been legalized into the constant pool?
2259 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2260 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2261 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2262 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2263 return CFP->getValueAPF().isPosZero();
2269 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2270 /// the given operands.
2272 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2273 SDValue &ARMcc, SelectionDAG &DAG,
2274 DebugLoc dl) const {
2275 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2276 unsigned C = RHSC->getZExtValue();
2277 if (!isLegalICmpImmediate(C)) {
2278 // Constant does not fit, try adjusting it by one?
2283 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2284 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2285 RHS = DAG.getConstant(C-1, MVT::i32);
2290 if (C != 0 && isLegalICmpImmediate(C-1)) {
2291 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2292 RHS = DAG.getConstant(C-1, MVT::i32);
2297 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2298 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2299 RHS = DAG.getConstant(C+1, MVT::i32);
2304 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2305 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2306 RHS = DAG.getConstant(C+1, MVT::i32);
2313 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2314 ARMISD::NodeType CompareType;
2317 CompareType = ARMISD::CMP;
2322 CompareType = ARMISD::CMPZ;
2325 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2326 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2329 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2331 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2332 DebugLoc dl) const {
2334 if (!isFloatingPointZero(RHS))
2335 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2337 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2338 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2341 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2342 SDValue Cond = Op.getOperand(0);
2343 SDValue SelectTrue = Op.getOperand(1);
2344 SDValue SelectFalse = Op.getOperand(2);
2345 DebugLoc dl = Op.getDebugLoc();
2349 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2350 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2352 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2353 const ConstantSDNode *CMOVTrue =
2354 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2355 const ConstantSDNode *CMOVFalse =
2356 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2358 if (CMOVTrue && CMOVFalse) {
2359 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2360 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2364 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2366 False = SelectFalse;
2367 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2372 if (True.getNode() && False.getNode()) {
2373 EVT VT = Cond.getValueType();
2374 SDValue ARMcc = Cond.getOperand(2);
2375 SDValue CCR = Cond.getOperand(3);
2376 SDValue Cmp = Cond.getOperand(4);
2377 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2382 return DAG.getSelectCC(dl, Cond,
2383 DAG.getConstant(0, Cond.getValueType()),
2384 SelectTrue, SelectFalse, ISD::SETNE);
2387 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2388 EVT VT = Op.getValueType();
2389 SDValue LHS = Op.getOperand(0);
2390 SDValue RHS = Op.getOperand(1);
2391 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2392 SDValue TrueVal = Op.getOperand(2);
2393 SDValue FalseVal = Op.getOperand(3);
2394 DebugLoc dl = Op.getDebugLoc();
2396 if (LHS.getValueType() == MVT::i32) {
2398 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2399 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2400 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2403 ARMCC::CondCodes CondCode, CondCode2;
2404 FPCCToARMCC(CC, CondCode, CondCode2);
2406 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2407 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2408 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2409 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2411 if (CondCode2 != ARMCC::AL) {
2412 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2413 // FIXME: Needs another CMP because flag can have but one use.
2414 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2415 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2416 Result, TrueVal, ARMcc2, CCR, Cmp2);
2421 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2422 /// to morph to an integer compare sequence.
2423 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2424 const ARMSubtarget *Subtarget) {
2425 SDNode *N = Op.getNode();
2426 if (!N->hasOneUse())
2427 // Otherwise it requires moving the value from fp to integer registers.
2429 if (!N->getNumValues())
2431 EVT VT = Op.getValueType();
2432 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2433 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2434 // vmrs are very slow, e.g. cortex-a8.
2437 if (isFloatingPointZero(Op)) {
2441 return ISD::isNormalLoad(N);
2444 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2445 if (isFloatingPointZero(Op))
2446 return DAG.getConstant(0, MVT::i32);
2448 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2449 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2450 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2451 Ld->isVolatile(), Ld->isNonTemporal(),
2452 Ld->getAlignment());
2454 llvm_unreachable("Unknown VFP cmp argument!");
2457 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2458 SDValue &RetVal1, SDValue &RetVal2) {
2459 if (isFloatingPointZero(Op)) {
2460 RetVal1 = DAG.getConstant(0, MVT::i32);
2461 RetVal2 = DAG.getConstant(0, MVT::i32);
2465 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2466 SDValue Ptr = Ld->getBasePtr();
2467 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2468 Ld->getChain(), Ptr,
2469 Ld->getPointerInfo(),
2470 Ld->isVolatile(), Ld->isNonTemporal(),
2471 Ld->getAlignment());
2473 EVT PtrType = Ptr.getValueType();
2474 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2475 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2476 PtrType, Ptr, DAG.getConstant(4, PtrType));
2477 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2478 Ld->getChain(), NewPtr,
2479 Ld->getPointerInfo().getWithOffset(4),
2480 Ld->isVolatile(), Ld->isNonTemporal(),
2485 llvm_unreachable("Unknown VFP cmp argument!");
2488 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2489 /// f32 and even f64 comparisons to integer ones.
2491 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2492 SDValue Chain = Op.getOperand(0);
2493 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2494 SDValue LHS = Op.getOperand(2);
2495 SDValue RHS = Op.getOperand(3);
2496 SDValue Dest = Op.getOperand(4);
2497 DebugLoc dl = Op.getDebugLoc();
2499 bool SeenZero = false;
2500 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2501 canChangeToInt(RHS, SeenZero, Subtarget) &&
2502 // If one of the operand is zero, it's safe to ignore the NaN case since
2503 // we only care about equality comparisons.
2504 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2505 // If unsafe fp math optimization is enabled and there are no othter uses of
2506 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2507 // to an integer comparison.
2508 if (CC == ISD::SETOEQ)
2510 else if (CC == ISD::SETUNE)
2514 if (LHS.getValueType() == MVT::f32) {
2515 LHS = bitcastf32Toi32(LHS, DAG);
2516 RHS = bitcastf32Toi32(RHS, DAG);
2517 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2518 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2519 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2520 Chain, Dest, ARMcc, CCR, Cmp);
2525 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2526 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2527 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2528 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2529 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2530 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2531 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2537 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2538 SDValue Chain = Op.getOperand(0);
2539 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2540 SDValue LHS = Op.getOperand(2);
2541 SDValue RHS = Op.getOperand(3);
2542 SDValue Dest = Op.getOperand(4);
2543 DebugLoc dl = Op.getDebugLoc();
2545 if (LHS.getValueType() == MVT::i32) {
2547 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2548 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2549 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2550 Chain, Dest, ARMcc, CCR, Cmp);
2553 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2556 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2557 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2558 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2559 if (Result.getNode())
2563 ARMCC::CondCodes CondCode, CondCode2;
2564 FPCCToARMCC(CC, CondCode, CondCode2);
2566 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2567 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2568 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2569 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2570 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2571 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2572 if (CondCode2 != ARMCC::AL) {
2573 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2574 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2575 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2580 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2581 SDValue Chain = Op.getOperand(0);
2582 SDValue Table = Op.getOperand(1);
2583 SDValue Index = Op.getOperand(2);
2584 DebugLoc dl = Op.getDebugLoc();
2586 EVT PTy = getPointerTy();
2587 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2588 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2589 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2590 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2591 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2592 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2593 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2594 if (Subtarget->isThumb2()) {
2595 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2596 // which does another jump to the destination. This also makes it easier
2597 // to translate it to TBB / TBH later.
2598 // FIXME: This might not work if the function is extremely large.
2599 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2600 Addr, Op.getOperand(2), JTI, UId);
2602 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2603 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2604 MachinePointerInfo::getJumpTable(),
2606 Chain = Addr.getValue(1);
2607 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2608 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2610 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2611 MachinePointerInfo::getJumpTable(), false, false, 0);
2612 Chain = Addr.getValue(1);
2613 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2617 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2618 DebugLoc dl = Op.getDebugLoc();
2621 switch (Op.getOpcode()) {
2623 assert(0 && "Invalid opcode!");
2624 case ISD::FP_TO_SINT:
2625 Opc = ARMISD::FTOSI;
2627 case ISD::FP_TO_UINT:
2628 Opc = ARMISD::FTOUI;
2631 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2632 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2635 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2636 EVT VT = Op.getValueType();
2637 DebugLoc dl = Op.getDebugLoc();
2640 switch (Op.getOpcode()) {
2642 assert(0 && "Invalid opcode!");
2643 case ISD::SINT_TO_FP:
2644 Opc = ARMISD::SITOF;
2646 case ISD::UINT_TO_FP:
2647 Opc = ARMISD::UITOF;
2651 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2652 return DAG.getNode(Opc, dl, VT, Op);
2655 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2656 // Implement fcopysign with a fabs and a conditional fneg.
2657 SDValue Tmp0 = Op.getOperand(0);
2658 SDValue Tmp1 = Op.getOperand(1);
2659 DebugLoc dl = Op.getDebugLoc();
2660 EVT VT = Op.getValueType();
2661 EVT SrcVT = Tmp1.getValueType();
2662 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2663 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2664 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2665 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2666 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2667 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2670 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2671 MachineFunction &MF = DAG.getMachineFunction();
2672 MachineFrameInfo *MFI = MF.getFrameInfo();
2673 MFI->setReturnAddressIsTaken(true);
2675 EVT VT = Op.getValueType();
2676 DebugLoc dl = Op.getDebugLoc();
2677 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2679 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2680 SDValue Offset = DAG.getConstant(4, MVT::i32);
2681 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2682 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2683 MachinePointerInfo(), false, false, 0);
2686 // Return LR, which contains the return address. Mark it an implicit live-in.
2687 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2688 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2691 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2692 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2693 MFI->setFrameAddressIsTaken(true);
2695 EVT VT = Op.getValueType();
2696 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2697 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2698 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2699 ? ARM::R7 : ARM::R11;
2700 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2702 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2703 MachinePointerInfo(),
2708 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2709 /// expand a bit convert where either the source or destination type is i64 to
2710 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2711 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2712 /// vectors), since the legalizer won't know what to do with that.
2713 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2714 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2715 DebugLoc dl = N->getDebugLoc();
2716 SDValue Op = N->getOperand(0);
2718 // This function is only supposed to be called for i64 types, either as the
2719 // source or destination of the bit convert.
2720 EVT SrcVT = Op.getValueType();
2721 EVT DstVT = N->getValueType(0);
2722 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2723 "ExpandBIT_CONVERT called for non-i64 type");
2725 // Turn i64->f64 into VMOVDRR.
2726 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2727 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2728 DAG.getConstant(0, MVT::i32));
2729 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2730 DAG.getConstant(1, MVT::i32));
2731 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2732 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2735 // Turn f64->i64 into VMOVRRD.
2736 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2737 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2738 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2739 // Merge the pieces into a single i64 value.
2740 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2746 /// getZeroVector - Returns a vector of specified type with all zero elements.
2747 /// Zero vectors are used to represent vector negation and in those cases
2748 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2749 /// not support i64 elements, so sometimes the zero vectors will need to be
2750 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2752 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2753 assert(VT.isVector() && "Expected a vector type");
2754 // The canonical modified immediate encoding of a zero vector is....0!
2755 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2756 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2757 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2758 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2761 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2762 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2763 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2764 SelectionDAG &DAG) const {
2765 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2766 EVT VT = Op.getValueType();
2767 unsigned VTBits = VT.getSizeInBits();
2768 DebugLoc dl = Op.getDebugLoc();
2769 SDValue ShOpLo = Op.getOperand(0);
2770 SDValue ShOpHi = Op.getOperand(1);
2771 SDValue ShAmt = Op.getOperand(2);
2773 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2775 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2777 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2778 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2779 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2780 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2781 DAG.getConstant(VTBits, MVT::i32));
2782 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2783 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2784 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2786 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2787 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2789 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2790 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2793 SDValue Ops[2] = { Lo, Hi };
2794 return DAG.getMergeValues(Ops, 2, dl);
2797 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2798 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2799 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2800 SelectionDAG &DAG) const {
2801 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2802 EVT VT = Op.getValueType();
2803 unsigned VTBits = VT.getSizeInBits();
2804 DebugLoc dl = Op.getDebugLoc();
2805 SDValue ShOpLo = Op.getOperand(0);
2806 SDValue ShOpHi = Op.getOperand(1);
2807 SDValue ShAmt = Op.getOperand(2);
2810 assert(Op.getOpcode() == ISD::SHL_PARTS);
2811 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2812 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2813 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2814 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2815 DAG.getConstant(VTBits, MVT::i32));
2816 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2817 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2819 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2820 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2821 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2823 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2824 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2827 SDValue Ops[2] = { Lo, Hi };
2828 return DAG.getMergeValues(Ops, 2, dl);
2831 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2832 SelectionDAG &DAG) const {
2833 // The rounding mode is in bits 23:22 of the FPSCR.
2834 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2835 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2836 // so that the shift + and get folded into a bitfield extract.
2837 DebugLoc dl = Op.getDebugLoc();
2838 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2839 DAG.getConstant(Intrinsic::arm_get_fpscr,
2841 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2842 DAG.getConstant(1U << 22, MVT::i32));
2843 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2844 DAG.getConstant(22, MVT::i32));
2845 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2846 DAG.getConstant(3, MVT::i32));
2849 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2850 const ARMSubtarget *ST) {
2851 EVT VT = N->getValueType(0);
2852 DebugLoc dl = N->getDebugLoc();
2854 if (!ST->hasV6T2Ops())
2857 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2858 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2861 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2862 const ARMSubtarget *ST) {
2863 EVT VT = N->getValueType(0);
2864 DebugLoc dl = N->getDebugLoc();
2866 // Lower vector shifts on NEON to use VSHL.
2867 if (VT.isVector()) {
2868 assert(ST->hasNEON() && "unexpected vector shift");
2870 // Left shifts translate directly to the vshiftu intrinsic.
2871 if (N->getOpcode() == ISD::SHL)
2872 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2873 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2874 N->getOperand(0), N->getOperand(1));
2876 assert((N->getOpcode() == ISD::SRA ||
2877 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2879 // NEON uses the same intrinsics for both left and right shifts. For
2880 // right shifts, the shift amounts are negative, so negate the vector of
2882 EVT ShiftVT = N->getOperand(1).getValueType();
2883 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2884 getZeroVector(ShiftVT, DAG, dl),
2886 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2887 Intrinsic::arm_neon_vshifts :
2888 Intrinsic::arm_neon_vshiftu);
2889 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2890 DAG.getConstant(vshiftInt, MVT::i32),
2891 N->getOperand(0), NegatedCount);
2894 // We can get here for a node like i32 = ISD::SHL i32, i64
2898 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2899 "Unknown shift to lower!");
2901 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2902 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2903 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2906 // If we are in thumb mode, we don't have RRX.
2907 if (ST->isThumb1Only()) return SDValue();
2909 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2910 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2911 DAG.getConstant(0, MVT::i32));
2912 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2913 DAG.getConstant(1, MVT::i32));
2915 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2916 // captures the result into a carry flag.
2917 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2918 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2920 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2921 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2923 // Merge the pieces into a single i64 value.
2924 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2927 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2928 SDValue TmpOp0, TmpOp1;
2929 bool Invert = false;
2933 SDValue Op0 = Op.getOperand(0);
2934 SDValue Op1 = Op.getOperand(1);
2935 SDValue CC = Op.getOperand(2);
2936 EVT VT = Op.getValueType();
2937 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2938 DebugLoc dl = Op.getDebugLoc();
2940 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2941 switch (SetCCOpcode) {
2942 default: llvm_unreachable("Illegal FP comparison"); break;
2944 case ISD::SETNE: Invert = true; // Fallthrough
2946 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2948 case ISD::SETLT: Swap = true; // Fallthrough
2950 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2952 case ISD::SETLE: Swap = true; // Fallthrough
2954 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2955 case ISD::SETUGE: Swap = true; // Fallthrough
2956 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2957 case ISD::SETUGT: Swap = true; // Fallthrough
2958 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2959 case ISD::SETUEQ: Invert = true; // Fallthrough
2961 // Expand this to (OLT | OGT).
2965 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2966 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2968 case ISD::SETUO: Invert = true; // Fallthrough
2970 // Expand this to (OLT | OGE).
2974 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2975 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2979 // Integer comparisons.
2980 switch (SetCCOpcode) {
2981 default: llvm_unreachable("Illegal integer comparison"); break;
2982 case ISD::SETNE: Invert = true;
2983 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2984 case ISD::SETLT: Swap = true;
2985 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2986 case ISD::SETLE: Swap = true;
2987 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2988 case ISD::SETULT: Swap = true;
2989 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2990 case ISD::SETULE: Swap = true;
2991 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2994 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2995 if (Opc == ARMISD::VCEQ) {
2998 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3000 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3003 // Ignore bitconvert.
3004 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
3005 AndOp = AndOp.getOperand(0);
3007 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3009 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
3010 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
3017 std::swap(Op0, Op1);
3019 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3022 Result = DAG.getNOT(dl, Result, VT);
3027 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3028 /// valid vector constant for a NEON instruction with a "modified immediate"
3029 /// operand (e.g., VMOV). If so, return the encoded value.
3030 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3031 unsigned SplatBitSize, SelectionDAG &DAG,
3032 EVT &VT, bool is128Bits, bool isVMOV) {
3033 unsigned OpCmode, Imm;
3035 // SplatBitSize is set to the smallest size that splats the vector, so a
3036 // zero vector will always have SplatBitSize == 8. However, NEON modified
3037 // immediate instructions others than VMOV do not support the 8-bit encoding
3038 // of a zero vector, and the default encoding of zero is supposed to be the
3043 switch (SplatBitSize) {
3047 // Any 1-byte value is OK. Op=0, Cmode=1110.
3048 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3051 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3055 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3056 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3057 if ((SplatBits & ~0xff) == 0) {
3058 // Value = 0x00nn: Op=x, Cmode=100x.
3063 if ((SplatBits & ~0xff00) == 0) {
3064 // Value = 0xnn00: Op=x, Cmode=101x.
3066 Imm = SplatBits >> 8;
3072 // NEON's 32-bit VMOV supports splat values where:
3073 // * only one byte is nonzero, or
3074 // * the least significant byte is 0xff and the second byte is nonzero, or
3075 // * the least significant 2 bytes are 0xff and the third is nonzero.
3076 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3077 if ((SplatBits & ~0xff) == 0) {
3078 // Value = 0x000000nn: Op=x, Cmode=000x.
3083 if ((SplatBits & ~0xff00) == 0) {
3084 // Value = 0x0000nn00: Op=x, Cmode=001x.
3086 Imm = SplatBits >> 8;
3089 if ((SplatBits & ~0xff0000) == 0) {
3090 // Value = 0x00nn0000: Op=x, Cmode=010x.
3092 Imm = SplatBits >> 16;
3095 if ((SplatBits & ~0xff000000) == 0) {
3096 // Value = 0xnn000000: Op=x, Cmode=011x.
3098 Imm = SplatBits >> 24;
3102 if ((SplatBits & ~0xffff) == 0 &&
3103 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3104 // Value = 0x0000nnff: Op=x, Cmode=1100.
3106 Imm = SplatBits >> 8;
3111 if ((SplatBits & ~0xffffff) == 0 &&
3112 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3113 // Value = 0x00nnffff: Op=x, Cmode=1101.
3115 Imm = SplatBits >> 16;
3116 SplatBits |= 0xffff;
3120 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3121 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3122 // VMOV.I32. A (very) minor optimization would be to replicate the value
3123 // and fall through here to test for a valid 64-bit splat. But, then the
3124 // caller would also need to check and handle the change in size.
3130 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3131 uint64_t BitMask = 0xff;
3133 unsigned ImmMask = 1;
3135 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3136 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3139 } else if ((SplatBits & BitMask) != 0) {
3145 // Op=1, Cmode=1110.
3148 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3153 llvm_unreachable("unexpected size for isNEONModifiedImm");
3157 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3158 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3161 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3162 bool &ReverseVEXT, unsigned &Imm) {
3163 unsigned NumElts = VT.getVectorNumElements();
3164 ReverseVEXT = false;
3166 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3172 // If this is a VEXT shuffle, the immediate value is the index of the first
3173 // element. The other shuffle indices must be the successive elements after
3175 unsigned ExpectedElt = Imm;
3176 for (unsigned i = 1; i < NumElts; ++i) {
3177 // Increment the expected index. If it wraps around, it may still be
3178 // a VEXT but the source vectors must be swapped.
3180 if (ExpectedElt == NumElts * 2) {
3185 if (M[i] < 0) continue; // ignore UNDEF indices
3186 if (ExpectedElt != static_cast<unsigned>(M[i]))
3190 // Adjust the index value if the source operands will be swapped.
3197 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3198 /// instruction with the specified blocksize. (The order of the elements
3199 /// within each block of the vector is reversed.)
3200 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3201 unsigned BlockSize) {
3202 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3203 "Only possible block sizes for VREV are: 16, 32, 64");
3205 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3209 unsigned NumElts = VT.getVectorNumElements();
3210 unsigned BlockElts = M[0] + 1;
3211 // If the first shuffle index is UNDEF, be optimistic.
3213 BlockElts = BlockSize / EltSz;
3215 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3218 for (unsigned i = 0; i < NumElts; ++i) {
3219 if (M[i] < 0) continue; // ignore UNDEF indices
3220 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3227 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3228 unsigned &WhichResult) {
3229 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3233 unsigned NumElts = VT.getVectorNumElements();
3234 WhichResult = (M[0] == 0 ? 0 : 1);
3235 for (unsigned i = 0; i < NumElts; i += 2) {
3236 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3237 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3243 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3244 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3245 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3246 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3247 unsigned &WhichResult) {
3248 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3252 unsigned NumElts = VT.getVectorNumElements();
3253 WhichResult = (M[0] == 0 ? 0 : 1);
3254 for (unsigned i = 0; i < NumElts; i += 2) {
3255 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3256 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3262 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3263 unsigned &WhichResult) {
3264 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3268 unsigned NumElts = VT.getVectorNumElements();
3269 WhichResult = (M[0] == 0 ? 0 : 1);
3270 for (unsigned i = 0; i != NumElts; ++i) {
3271 if (M[i] < 0) continue; // ignore UNDEF indices
3272 if ((unsigned) M[i] != 2 * i + WhichResult)
3276 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3277 if (VT.is64BitVector() && EltSz == 32)
3283 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3284 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3285 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3286 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3287 unsigned &WhichResult) {
3288 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3292 unsigned Half = VT.getVectorNumElements() / 2;
3293 WhichResult = (M[0] == 0 ? 0 : 1);
3294 for (unsigned j = 0; j != 2; ++j) {
3295 unsigned Idx = WhichResult;
3296 for (unsigned i = 0; i != Half; ++i) {
3297 int MIdx = M[i + j * Half];
3298 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3304 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3305 if (VT.is64BitVector() && EltSz == 32)
3311 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3312 unsigned &WhichResult) {
3313 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3317 unsigned NumElts = VT.getVectorNumElements();
3318 WhichResult = (M[0] == 0 ? 0 : 1);
3319 unsigned Idx = WhichResult * NumElts / 2;
3320 for (unsigned i = 0; i != NumElts; i += 2) {
3321 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3322 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3327 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3328 if (VT.is64BitVector() && EltSz == 32)
3334 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3335 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3336 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3337 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3338 unsigned &WhichResult) {
3339 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3343 unsigned NumElts = VT.getVectorNumElements();
3344 WhichResult = (M[0] == 0 ? 0 : 1);
3345 unsigned Idx = WhichResult * NumElts / 2;
3346 for (unsigned i = 0; i != NumElts; i += 2) {
3347 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3348 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3353 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3354 if (VT.is64BitVector() && EltSz == 32)
3360 // If N is an integer constant that can be moved into a register in one
3361 // instruction, return an SDValue of such a constant (will become a MOV
3362 // instruction). Otherwise return null.
3363 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3364 const ARMSubtarget *ST, DebugLoc dl) {
3366 if (!isa<ConstantSDNode>(N))
3368 Val = cast<ConstantSDNode>(N)->getZExtValue();
3370 if (ST->isThumb1Only()) {
3371 if (Val <= 255 || ~Val <= 255)
3372 return DAG.getConstant(Val, MVT::i32);
3374 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3375 return DAG.getConstant(Val, MVT::i32);
3380 // If this is a case we can't handle, return null and let the default
3381 // expansion code take care of it.
3382 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3383 const ARMSubtarget *ST) {
3384 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3385 DebugLoc dl = Op.getDebugLoc();
3386 EVT VT = Op.getValueType();
3388 APInt SplatBits, SplatUndef;
3389 unsigned SplatBitSize;
3391 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3392 if (SplatBitSize <= 64) {
3393 // Check if an immediate VMOV works.
3395 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3396 SplatUndef.getZExtValue(), SplatBitSize,
3397 DAG, VmovVT, VT.is128BitVector(), true);
3398 if (Val.getNode()) {
3399 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3400 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3403 // Try an immediate VMVN.
3404 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3405 ((1LL << SplatBitSize) - 1));
3406 Val = isNEONModifiedImm(NegatedImm,
3407 SplatUndef.getZExtValue(), SplatBitSize,
3408 DAG, VmovVT, VT.is128BitVector(), false);
3409 if (Val.getNode()) {
3410 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3411 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3416 // Scan through the operands to see if only one value is used.
3417 unsigned NumElts = VT.getVectorNumElements();
3418 bool isOnlyLowElement = true;
3419 bool usesOnlyOneValue = true;
3420 bool isConstant = true;
3422 for (unsigned i = 0; i < NumElts; ++i) {
3423 SDValue V = Op.getOperand(i);
3424 if (V.getOpcode() == ISD::UNDEF)
3427 isOnlyLowElement = false;
3428 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3431 if (!Value.getNode())
3433 else if (V != Value)
3434 usesOnlyOneValue = false;
3437 if (!Value.getNode())
3438 return DAG.getUNDEF(VT);
3440 if (isOnlyLowElement)
3441 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3443 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3445 if (EnableARMVDUPsplat) {
3446 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3447 // i32 and try again.
3448 if (usesOnlyOneValue && EltSize <= 32) {
3450 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3451 if (VT.getVectorElementType().isFloatingPoint()) {
3452 SmallVector<SDValue, 8> Ops;
3453 for (unsigned i = 0; i < NumElts; ++i)
3454 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3456 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3458 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3459 LowerBUILD_VECTOR(Val, DAG, ST));
3461 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3463 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3467 // If all elements are constants and the case above didn't get hit, fall back
3468 // to the default expansion, which will generate a load from the constant
3473 if (!EnableARMVDUPsplat) {
3474 // Use VDUP for non-constant splats.
3475 if (usesOnlyOneValue && EltSize <= 32)
3476 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3479 // Vectors with 32- or 64-bit elements can be built by directly assigning
3480 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3481 // will be legalized.
3482 if (EltSize >= 32) {
3483 // Do the expansion with floating-point types, since that is what the VFP
3484 // registers are defined to use, and since i64 is not legal.
3485 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3486 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3487 SmallVector<SDValue, 8> Ops;
3488 for (unsigned i = 0; i < NumElts; ++i)
3489 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3490 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3491 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3497 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3498 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3499 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3500 /// are assumed to be legal.
3502 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3504 if (VT.getVectorNumElements() == 4 &&
3505 (VT.is128BitVector() || VT.is64BitVector())) {
3506 unsigned PFIndexes[4];
3507 for (unsigned i = 0; i != 4; ++i) {
3511 PFIndexes[i] = M[i];
3514 // Compute the index in the perfect shuffle table.
3515 unsigned PFTableIndex =
3516 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3517 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3518 unsigned Cost = (PFEntry >> 30);
3525 unsigned Imm, WhichResult;
3527 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3528 return (EltSize >= 32 ||
3529 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3530 isVREVMask(M, VT, 64) ||
3531 isVREVMask(M, VT, 32) ||
3532 isVREVMask(M, VT, 16) ||
3533 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3534 isVTRNMask(M, VT, WhichResult) ||
3535 isVUZPMask(M, VT, WhichResult) ||
3536 isVZIPMask(M, VT, WhichResult) ||
3537 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3538 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3539 isVZIP_v_undef_Mask(M, VT, WhichResult));
3542 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3543 /// the specified operations to build the shuffle.
3544 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3545 SDValue RHS, SelectionDAG &DAG,
3547 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3548 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3549 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3552 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3561 OP_VUZPL, // VUZP, left result
3562 OP_VUZPR, // VUZP, right result
3563 OP_VZIPL, // VZIP, left result
3564 OP_VZIPR, // VZIP, right result
3565 OP_VTRNL, // VTRN, left result
3566 OP_VTRNR // VTRN, right result
3569 if (OpNum == OP_COPY) {
3570 if (LHSID == (1*9+2)*9+3) return LHS;
3571 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3575 SDValue OpLHS, OpRHS;
3576 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3577 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3578 EVT VT = OpLHS.getValueType();
3581 default: llvm_unreachable("Unknown shuffle opcode!");
3583 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3588 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3589 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3593 return DAG.getNode(ARMISD::VEXT, dl, VT,
3595 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3598 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3599 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3602 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3603 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3606 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3607 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3611 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3612 SDValue V1 = Op.getOperand(0);
3613 SDValue V2 = Op.getOperand(1);
3614 DebugLoc dl = Op.getDebugLoc();
3615 EVT VT = Op.getValueType();
3616 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3617 SmallVector<int, 8> ShuffleMask;
3619 // Convert shuffles that are directly supported on NEON to target-specific
3620 // DAG nodes, instead of keeping them as shuffles and matching them again
3621 // during code selection. This is more efficient and avoids the possibility
3622 // of inconsistencies between legalization and selection.
3623 // FIXME: floating-point vectors should be canonicalized to integer vectors
3624 // of the same time so that they get CSEd properly.
3625 SVN->getMask(ShuffleMask);
3627 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3628 if (EltSize <= 32) {
3629 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3630 int Lane = SVN->getSplatIndex();
3631 // If this is undef splat, generate it via "just" vdup, if possible.
3632 if (Lane == -1) Lane = 0;
3634 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3635 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3637 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3638 DAG.getConstant(Lane, MVT::i32));
3643 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3646 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3647 DAG.getConstant(Imm, MVT::i32));
3650 if (isVREVMask(ShuffleMask, VT, 64))
3651 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3652 if (isVREVMask(ShuffleMask, VT, 32))
3653 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3654 if (isVREVMask(ShuffleMask, VT, 16))
3655 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3657 // Check for Neon shuffles that modify both input vectors in place.
3658 // If both results are used, i.e., if there are two shuffles with the same
3659 // source operands and with masks corresponding to both results of one of
3660 // these operations, DAG memoization will ensure that a single node is
3661 // used for both shuffles.
3662 unsigned WhichResult;
3663 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3664 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3665 V1, V2).getValue(WhichResult);
3666 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3667 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3668 V1, V2).getValue(WhichResult);
3669 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3670 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3671 V1, V2).getValue(WhichResult);
3673 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3674 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3675 V1, V1).getValue(WhichResult);
3676 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3677 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3678 V1, V1).getValue(WhichResult);
3679 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3680 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3681 V1, V1).getValue(WhichResult);
3684 // If the shuffle is not directly supported and it has 4 elements, use
3685 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3686 unsigned NumElts = VT.getVectorNumElements();
3688 unsigned PFIndexes[4];
3689 for (unsigned i = 0; i != 4; ++i) {
3690 if (ShuffleMask[i] < 0)
3693 PFIndexes[i] = ShuffleMask[i];
3696 // Compute the index in the perfect shuffle table.
3697 unsigned PFTableIndex =
3698 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3699 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3700 unsigned Cost = (PFEntry >> 30);
3703 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3706 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3707 if (EltSize >= 32) {
3708 // Do the expansion with floating-point types, since that is what the VFP
3709 // registers are defined to use, and since i64 is not legal.
3710 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3711 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3712 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3713 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3714 SmallVector<SDValue, 8> Ops;
3715 for (unsigned i = 0; i < NumElts; ++i) {
3716 if (ShuffleMask[i] < 0)
3717 Ops.push_back(DAG.getUNDEF(EltVT));
3719 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3720 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3721 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3724 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3725 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3731 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3732 EVT VT = Op.getValueType();
3733 DebugLoc dl = Op.getDebugLoc();
3734 SDValue Vec = Op.getOperand(0);
3735 SDValue Lane = Op.getOperand(1);
3736 assert(VT == MVT::i32 &&
3737 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3738 "unexpected type for custom-lowering vector extract");
3739 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3742 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3743 // The only time a CONCAT_VECTORS operation can have legal types is when
3744 // two 64-bit vectors are concatenated to a 128-bit vector.
3745 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3746 "unexpected CONCAT_VECTORS");
3747 DebugLoc dl = Op.getDebugLoc();
3748 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3749 SDValue Op0 = Op.getOperand(0);
3750 SDValue Op1 = Op.getOperand(1);
3751 if (Op0.getOpcode() != ISD::UNDEF)
3752 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3753 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3754 DAG.getIntPtrConstant(0));
3755 if (Op1.getOpcode() != ISD::UNDEF)
3756 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3757 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3758 DAG.getIntPtrConstant(1));
3759 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3762 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3763 /// an extending load, return the unextended value.
3764 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3765 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3766 return N->getOperand(0);
3767 LoadSDNode *LD = cast<LoadSDNode>(N);
3768 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3769 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3770 LD->isNonTemporal(), LD->getAlignment());
3773 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3774 // Multiplications are only custom-lowered for 128-bit vectors so that
3775 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3776 EVT VT = Op.getValueType();
3777 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3778 SDNode *N0 = Op.getOperand(0).getNode();
3779 SDNode *N1 = Op.getOperand(1).getNode();
3780 unsigned NewOpc = 0;
3781 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3782 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3783 NewOpc = ARMISD::VMULLs;
3784 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3785 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3786 NewOpc = ARMISD::VMULLu;
3787 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3788 // Fall through to expand this. It is not legal.
3791 // Other vector multiplications are legal.
3795 // Legalize to a VMULL instruction.
3796 DebugLoc DL = Op.getDebugLoc();
3797 SDValue Op0 = SkipExtension(N0, DAG);
3798 SDValue Op1 = SkipExtension(N1, DAG);
3800 assert(Op0.getValueType().is64BitVector() &&
3801 Op1.getValueType().is64BitVector() &&
3802 "unexpected types for extended operands to VMULL");
3803 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3806 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3807 switch (Op.getOpcode()) {
3808 default: llvm_unreachable("Don't know how to custom lower this!");
3809 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3810 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3811 case ISD::GlobalAddress:
3812 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3813 LowerGlobalAddressELF(Op, DAG);
3814 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3815 case ISD::SELECT: return LowerSELECT(Op, DAG);
3816 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3817 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3818 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3819 case ISD::VASTART: return LowerVASTART(Op, DAG);
3820 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3821 case ISD::SINT_TO_FP:
3822 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3823 case ISD::FP_TO_SINT:
3824 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3825 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3826 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3827 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3828 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3829 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3830 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3831 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3833 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3836 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3837 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3838 case ISD::SRL_PARTS:
3839 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3840 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3841 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3842 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3843 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3844 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3845 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3846 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3847 case ISD::MUL: return LowerMUL(Op, DAG);
3852 /// ReplaceNodeResults - Replace the results of node with an illegal result
3853 /// type with new values built out of custom code.
3854 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3855 SmallVectorImpl<SDValue>&Results,
3856 SelectionDAG &DAG) const {
3858 switch (N->getOpcode()) {
3860 llvm_unreachable("Don't know how to custom expand this!");
3862 case ISD::BIT_CONVERT:
3863 Res = ExpandBIT_CONVERT(N, DAG);
3867 Res = LowerShift(N, DAG, Subtarget);
3871 Results.push_back(Res);
3874 //===----------------------------------------------------------------------===//
3875 // ARM Scheduler Hooks
3876 //===----------------------------------------------------------------------===//
3879 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3880 MachineBasicBlock *BB,
3881 unsigned Size) const {
3882 unsigned dest = MI->getOperand(0).getReg();
3883 unsigned ptr = MI->getOperand(1).getReg();
3884 unsigned oldval = MI->getOperand(2).getReg();
3885 unsigned newval = MI->getOperand(3).getReg();
3886 unsigned scratch = BB->getParent()->getRegInfo()
3887 .createVirtualRegister(ARM::GPRRegisterClass);
3888 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3889 DebugLoc dl = MI->getDebugLoc();
3890 bool isThumb2 = Subtarget->isThumb2();
3892 unsigned ldrOpc, strOpc;
3894 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3896 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3897 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3900 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3901 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3904 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3905 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3909 MachineFunction *MF = BB->getParent();
3910 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3911 MachineFunction::iterator It = BB;
3912 ++It; // insert the new blocks after the current block
3914 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3915 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3916 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3917 MF->insert(It, loop1MBB);
3918 MF->insert(It, loop2MBB);
3919 MF->insert(It, exitMBB);
3921 // Transfer the remainder of BB and its successor edges to exitMBB.
3922 exitMBB->splice(exitMBB->begin(), BB,
3923 llvm::next(MachineBasicBlock::iterator(MI)),
3925 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3929 // fallthrough --> loop1MBB
3930 BB->addSuccessor(loop1MBB);
3933 // ldrex dest, [ptr]
3937 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3938 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3939 .addReg(dest).addReg(oldval));
3940 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3941 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3942 BB->addSuccessor(loop2MBB);
3943 BB->addSuccessor(exitMBB);
3946 // strex scratch, newval, [ptr]
3950 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3952 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3953 .addReg(scratch).addImm(0));
3954 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3955 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3956 BB->addSuccessor(loop1MBB);
3957 BB->addSuccessor(exitMBB);
3963 MI->eraseFromParent(); // The instruction is gone now.
3969 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3970 unsigned Size, unsigned BinOpcode) const {
3971 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3972 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3974 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3975 MachineFunction *MF = BB->getParent();
3976 MachineFunction::iterator It = BB;
3979 unsigned dest = MI->getOperand(0).getReg();
3980 unsigned ptr = MI->getOperand(1).getReg();
3981 unsigned incr = MI->getOperand(2).getReg();
3982 DebugLoc dl = MI->getDebugLoc();
3984 bool isThumb2 = Subtarget->isThumb2();
3985 unsigned ldrOpc, strOpc;
3987 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3989 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3990 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3993 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3994 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3997 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3998 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4002 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4003 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
4004 MF->insert(It, loopMBB);
4005 MF->insert(It, exitMBB);
4007 // Transfer the remainder of BB and its successor edges to exitMBB.
4008 exitMBB->splice(exitMBB->begin(), BB,
4009 llvm::next(MachineBasicBlock::iterator(MI)),
4011 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4013 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4014 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4015 unsigned scratch2 = (!BinOpcode) ? incr :
4016 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
4020 // fallthrough --> loopMBB
4021 BB->addSuccessor(loopMBB);
4025 // <binop> scratch2, dest, incr
4026 // strex scratch, scratch2, ptr
4029 // fallthrough --> exitMBB
4031 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
4033 // operand order needs to go the other way for NAND
4034 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
4035 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4036 addReg(incr).addReg(dest)).addReg(0);
4038 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
4039 addReg(dest).addReg(incr)).addReg(0);
4042 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
4044 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4045 .addReg(scratch).addImm(0));
4046 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4047 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
4049 BB->addSuccessor(loopMBB);
4050 BB->addSuccessor(exitMBB);
4056 MI->eraseFromParent(); // The instruction is gone now.
4062 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4063 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4064 E = MBB->succ_end(); I != E; ++I)
4067 llvm_unreachable("Expecting a BB with two successors!");
4071 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4072 MachineBasicBlock *BB) const {
4073 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4074 DebugLoc dl = MI->getDebugLoc();
4075 bool isThumb2 = Subtarget->isThumb2();
4076 switch (MI->getOpcode()) {
4079 llvm_unreachable("Unexpected instr type to insert");
4081 case ARM::ATOMIC_LOAD_ADD_I8:
4082 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4083 case ARM::ATOMIC_LOAD_ADD_I16:
4084 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4085 case ARM::ATOMIC_LOAD_ADD_I32:
4086 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4088 case ARM::ATOMIC_LOAD_AND_I8:
4089 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4090 case ARM::ATOMIC_LOAD_AND_I16:
4091 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4092 case ARM::ATOMIC_LOAD_AND_I32:
4093 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4095 case ARM::ATOMIC_LOAD_OR_I8:
4096 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4097 case ARM::ATOMIC_LOAD_OR_I16:
4098 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4099 case ARM::ATOMIC_LOAD_OR_I32:
4100 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4102 case ARM::ATOMIC_LOAD_XOR_I8:
4103 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4104 case ARM::ATOMIC_LOAD_XOR_I16:
4105 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4106 case ARM::ATOMIC_LOAD_XOR_I32:
4107 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4109 case ARM::ATOMIC_LOAD_NAND_I8:
4110 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4111 case ARM::ATOMIC_LOAD_NAND_I16:
4112 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4113 case ARM::ATOMIC_LOAD_NAND_I32:
4114 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4116 case ARM::ATOMIC_LOAD_SUB_I8:
4117 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4118 case ARM::ATOMIC_LOAD_SUB_I16:
4119 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4120 case ARM::ATOMIC_LOAD_SUB_I32:
4121 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4123 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4124 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4125 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4127 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4128 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4129 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4131 case ARM::tMOVCCr_pseudo: {
4132 // To "insert" a SELECT_CC instruction, we actually have to insert the
4133 // diamond control-flow pattern. The incoming instruction knows the
4134 // destination vreg to set, the condition code register to branch on, the
4135 // true/false values to select between, and a branch opcode to use.
4136 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4137 MachineFunction::iterator It = BB;
4143 // cmpTY ccX, r1, r2
4145 // fallthrough --> copy0MBB
4146 MachineBasicBlock *thisMBB = BB;
4147 MachineFunction *F = BB->getParent();
4148 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4149 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4150 F->insert(It, copy0MBB);
4151 F->insert(It, sinkMBB);
4153 // Transfer the remainder of BB and its successor edges to sinkMBB.
4154 sinkMBB->splice(sinkMBB->begin(), BB,
4155 llvm::next(MachineBasicBlock::iterator(MI)),
4157 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4159 BB->addSuccessor(copy0MBB);
4160 BB->addSuccessor(sinkMBB);
4162 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4163 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4166 // %FalseValue = ...
4167 // # fallthrough to sinkMBB
4170 // Update machine-CFG edges
4171 BB->addSuccessor(sinkMBB);
4174 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4177 BuildMI(*BB, BB->begin(), dl,
4178 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4179 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4180 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4182 MI->eraseFromParent(); // The pseudo instruction is gone now.
4187 case ARM::BCCZi64: {
4188 // Compare both parts that make up the double comparison separately for
4190 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4192 unsigned LHS1 = MI->getOperand(1).getReg();
4193 unsigned LHS2 = MI->getOperand(2).getReg();
4195 AddDefaultPred(BuildMI(BB, dl,
4196 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4197 .addReg(LHS1).addImm(0));
4198 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4199 .addReg(LHS2).addImm(0)
4200 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4202 unsigned RHS1 = MI->getOperand(3).getReg();
4203 unsigned RHS2 = MI->getOperand(4).getReg();
4204 AddDefaultPred(BuildMI(BB, dl,
4205 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4206 .addReg(LHS1).addReg(RHS1));
4207 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4208 .addReg(LHS2).addReg(RHS2)
4209 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4212 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4213 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4214 if (MI->getOperand(0).getImm() == ARMCC::NE)
4215 std::swap(destMBB, exitMBB);
4217 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4218 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4219 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4222 MI->eraseFromParent(); // The pseudo instruction is gone now.
4228 //===----------------------------------------------------------------------===//
4229 // ARM Optimization Hooks
4230 //===----------------------------------------------------------------------===//
4233 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4234 TargetLowering::DAGCombinerInfo &DCI) {
4235 SelectionDAG &DAG = DCI.DAG;
4236 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4237 EVT VT = N->getValueType(0);
4238 unsigned Opc = N->getOpcode();
4239 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4240 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4241 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4242 ISD::CondCode CC = ISD::SETCC_INVALID;
4245 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4247 SDValue CCOp = Slct.getOperand(0);
4248 if (CCOp.getOpcode() == ISD::SETCC)
4249 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4252 bool DoXform = false;
4254 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4257 if (LHS.getOpcode() == ISD::Constant &&
4258 cast<ConstantSDNode>(LHS)->isNullValue()) {
4260 } else if (CC != ISD::SETCC_INVALID &&
4261 RHS.getOpcode() == ISD::Constant &&
4262 cast<ConstantSDNode>(RHS)->isNullValue()) {
4263 std::swap(LHS, RHS);
4264 SDValue Op0 = Slct.getOperand(0);
4265 EVT OpVT = isSlctCC ? Op0.getValueType() :
4266 Op0.getOperand(0).getValueType();
4267 bool isInt = OpVT.isInteger();
4268 CC = ISD::getSetCCInverse(CC, isInt);
4270 if (!TLI.isCondCodeLegal(CC, OpVT))
4271 return SDValue(); // Inverse operator isn't legal.
4278 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4280 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4281 Slct.getOperand(0), Slct.getOperand(1), CC);
4282 SDValue CCOp = Slct.getOperand(0);
4284 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4285 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4286 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4287 CCOp, OtherOp, Result);
4292 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4293 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4294 /// called with the default operands, and if that fails, with commuted
4296 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4297 TargetLowering::DAGCombinerInfo &DCI) {
4298 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4299 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4300 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4301 if (Result.getNode()) return Result;
4306 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4308 static SDValue PerformADDCombine(SDNode *N,
4309 TargetLowering::DAGCombinerInfo &DCI) {
4310 SDValue N0 = N->getOperand(0);
4311 SDValue N1 = N->getOperand(1);
4313 // First try with the default operand order.
4314 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4315 if (Result.getNode())
4318 // If that didn't work, try again with the operands commuted.
4319 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4322 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4324 static SDValue PerformSUBCombine(SDNode *N,
4325 TargetLowering::DAGCombinerInfo &DCI) {
4326 SDValue N0 = N->getOperand(0);
4327 SDValue N1 = N->getOperand(1);
4329 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4330 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4331 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4332 if (Result.getNode()) return Result;
4338 static SDValue PerformMULCombine(SDNode *N,
4339 TargetLowering::DAGCombinerInfo &DCI,
4340 const ARMSubtarget *Subtarget) {
4341 SelectionDAG &DAG = DCI.DAG;
4343 if (Subtarget->isThumb1Only())
4346 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4349 EVT VT = N->getValueType(0);
4353 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4357 uint64_t MulAmt = C->getZExtValue();
4358 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4359 ShiftAmt = ShiftAmt & (32 - 1);
4360 SDValue V = N->getOperand(0);
4361 DebugLoc DL = N->getDebugLoc();
4364 MulAmt >>= ShiftAmt;
4365 if (isPowerOf2_32(MulAmt - 1)) {
4366 // (mul x, 2^N + 1) => (add (shl x, N), x)
4367 Res = DAG.getNode(ISD::ADD, DL, VT,
4368 V, DAG.getNode(ISD::SHL, DL, VT,
4369 V, DAG.getConstant(Log2_32(MulAmt-1),
4371 } else if (isPowerOf2_32(MulAmt + 1)) {
4372 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4373 Res = DAG.getNode(ISD::SUB, DL, VT,
4374 DAG.getNode(ISD::SHL, DL, VT,
4375 V, DAG.getConstant(Log2_32(MulAmt+1),
4382 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4383 DAG.getConstant(ShiftAmt, MVT::i32));
4385 // Do not add new nodes to DAG combiner worklist.
4386 DCI.CombineTo(N, Res, false);
4390 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4391 static SDValue PerformORCombine(SDNode *N,
4392 TargetLowering::DAGCombinerInfo &DCI,
4393 const ARMSubtarget *Subtarget) {
4394 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4397 // BFI is only available on V6T2+
4398 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4401 SelectionDAG &DAG = DCI.DAG;
4402 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4403 DebugLoc DL = N->getDebugLoc();
4404 // 1) or (and A, mask), val => ARMbfi A, val, mask
4405 // iff (val & mask) == val
4407 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4408 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4409 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4410 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4411 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4412 // (i.e., copy a bitfield value into another bitfield of the same width)
4413 if (N0.getOpcode() != ISD::AND)
4416 EVT VT = N->getValueType(0);
4421 // The value and the mask need to be constants so we can verify this is
4422 // actually a bitfield set. If the mask is 0xffff, we can do better
4423 // via a movt instruction, so don't use BFI in that case.
4424 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4427 unsigned Mask = C->getZExtValue();
4431 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4432 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4433 unsigned Val = C->getZExtValue();
4434 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4436 Val >>= CountTrailingZeros_32(~Mask);
4438 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4439 DAG.getConstant(Val, MVT::i32),
4440 DAG.getConstant(Mask, MVT::i32));
4442 // Do not add new nodes to DAG combiner worklist.
4443 DCI.CombineTo(N, Res, false);
4444 } else if (N1.getOpcode() == ISD::AND) {
4445 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4446 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4449 unsigned Mask2 = C->getZExtValue();
4451 if (ARM::isBitFieldInvertedMask(Mask) &&
4452 ARM::isBitFieldInvertedMask(~Mask2) &&
4453 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4454 // The pack halfword instruction works better for masks that fit it,
4455 // so use that when it's available.
4456 if (Subtarget->hasT2ExtractPack() &&
4457 (Mask == 0xffff || Mask == 0xffff0000))
4460 unsigned lsb = CountTrailingZeros_32(Mask2);
4461 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4462 DAG.getConstant(lsb, MVT::i32));
4463 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4464 DAG.getConstant(Mask, MVT::i32));
4465 // Do not add new nodes to DAG combiner worklist.
4466 DCI.CombineTo(N, Res, false);
4467 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4468 ARM::isBitFieldInvertedMask(Mask2) &&
4469 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4470 // The pack halfword instruction works better for masks that fit it,
4471 // so use that when it's available.
4472 if (Subtarget->hasT2ExtractPack() &&
4473 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4476 unsigned lsb = CountTrailingZeros_32(Mask);
4477 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4478 DAG.getConstant(lsb, MVT::i32));
4479 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4480 DAG.getConstant(Mask2, MVT::i32));
4481 // Do not add new nodes to DAG combiner worklist.
4482 DCI.CombineTo(N, Res, false);
4489 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4490 /// ARMISD::VMOVRRD.
4491 static SDValue PerformVMOVRRDCombine(SDNode *N,
4492 TargetLowering::DAGCombinerInfo &DCI) {
4493 // vmovrrd(vmovdrr x, y) -> x,y
4494 SDValue InDouble = N->getOperand(0);
4495 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4496 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4500 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4501 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4502 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4503 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4504 SDValue Op0 = N->getOperand(0);
4505 SDValue Op1 = N->getOperand(1);
4506 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4507 Op0 = Op0.getOperand(0);
4508 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4509 Op1 = Op1.getOperand(0);
4510 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4511 Op0.getNode() == Op1.getNode() &&
4512 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4513 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4514 N->getValueType(0), Op0.getOperand(0));
4518 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4519 /// ISD::BUILD_VECTOR.
4520 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4521 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4522 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4523 // into a pair of GPRs, which is fine when the value is used as a scalar,
4524 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4525 if (N->getNumOperands() == 2)
4526 return PerformVMOVDRRCombine(N, DAG);
4531 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4532 /// ARMISD::VDUPLANE.
4533 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
4534 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4536 SDValue Op = N->getOperand(0);
4537 EVT VT = N->getValueType(0);
4539 // Ignore bit_converts.
4540 while (Op.getOpcode() == ISD::BIT_CONVERT)
4541 Op = Op.getOperand(0);
4542 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4545 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4546 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4547 // The canonical VMOV for a zero vector uses a 32-bit element size.
4548 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4550 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4552 if (EltSize > VT.getVectorElementType().getSizeInBits())
4555 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4558 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4559 /// operand of a vector shift operation, where all the elements of the
4560 /// build_vector must have the same constant integer value.
4561 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4562 // Ignore bit_converts.
4563 while (Op.getOpcode() == ISD::BIT_CONVERT)
4564 Op = Op.getOperand(0);
4565 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4566 APInt SplatBits, SplatUndef;
4567 unsigned SplatBitSize;
4569 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4570 HasAnyUndefs, ElementBits) ||
4571 SplatBitSize > ElementBits)
4573 Cnt = SplatBits.getSExtValue();
4577 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4578 /// operand of a vector shift left operation. That value must be in the range:
4579 /// 0 <= Value < ElementBits for a left shift; or
4580 /// 0 <= Value <= ElementBits for a long left shift.
4581 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4582 assert(VT.isVector() && "vector shift count is not a vector type");
4583 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4584 if (! getVShiftImm(Op, ElementBits, Cnt))
4586 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4589 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4590 /// operand of a vector shift right operation. For a shift opcode, the value
4591 /// is positive, but for an intrinsic the value count must be negative. The
4592 /// absolute value must be in the range:
4593 /// 1 <= |Value| <= ElementBits for a right shift; or
4594 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4595 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4597 assert(VT.isVector() && "vector shift count is not a vector type");
4598 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4599 if (! getVShiftImm(Op, ElementBits, Cnt))
4603 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4606 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4607 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4608 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4611 // Don't do anything for most intrinsics.
4614 // Vector shifts: check for immediate versions and lower them.
4615 // Note: This is done during DAG combining instead of DAG legalizing because
4616 // the build_vectors for 64-bit vector element shift counts are generally
4617 // not legal, and it is hard to see their values after they get legalized to
4618 // loads from a constant pool.
4619 case Intrinsic::arm_neon_vshifts:
4620 case Intrinsic::arm_neon_vshiftu:
4621 case Intrinsic::arm_neon_vshiftls:
4622 case Intrinsic::arm_neon_vshiftlu:
4623 case Intrinsic::arm_neon_vshiftn:
4624 case Intrinsic::arm_neon_vrshifts:
4625 case Intrinsic::arm_neon_vrshiftu:
4626 case Intrinsic::arm_neon_vrshiftn:
4627 case Intrinsic::arm_neon_vqshifts:
4628 case Intrinsic::arm_neon_vqshiftu:
4629 case Intrinsic::arm_neon_vqshiftsu:
4630 case Intrinsic::arm_neon_vqshiftns:
4631 case Intrinsic::arm_neon_vqshiftnu:
4632 case Intrinsic::arm_neon_vqshiftnsu:
4633 case Intrinsic::arm_neon_vqrshiftns:
4634 case Intrinsic::arm_neon_vqrshiftnu:
4635 case Intrinsic::arm_neon_vqrshiftnsu: {
4636 EVT VT = N->getOperand(1).getValueType();
4638 unsigned VShiftOpc = 0;
4641 case Intrinsic::arm_neon_vshifts:
4642 case Intrinsic::arm_neon_vshiftu:
4643 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4644 VShiftOpc = ARMISD::VSHL;
4647 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4648 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4649 ARMISD::VSHRs : ARMISD::VSHRu);
4654 case Intrinsic::arm_neon_vshiftls:
4655 case Intrinsic::arm_neon_vshiftlu:
4656 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4658 llvm_unreachable("invalid shift count for vshll intrinsic");
4660 case Intrinsic::arm_neon_vrshifts:
4661 case Intrinsic::arm_neon_vrshiftu:
4662 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4666 case Intrinsic::arm_neon_vqshifts:
4667 case Intrinsic::arm_neon_vqshiftu:
4668 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4672 case Intrinsic::arm_neon_vqshiftsu:
4673 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4675 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4677 case Intrinsic::arm_neon_vshiftn:
4678 case Intrinsic::arm_neon_vrshiftn:
4679 case Intrinsic::arm_neon_vqshiftns:
4680 case Intrinsic::arm_neon_vqshiftnu:
4681 case Intrinsic::arm_neon_vqshiftnsu:
4682 case Intrinsic::arm_neon_vqrshiftns:
4683 case Intrinsic::arm_neon_vqrshiftnu:
4684 case Intrinsic::arm_neon_vqrshiftnsu:
4685 // Narrowing shifts require an immediate right shift.
4686 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4688 llvm_unreachable("invalid shift count for narrowing vector shift "
4692 llvm_unreachable("unhandled vector shift");
4696 case Intrinsic::arm_neon_vshifts:
4697 case Intrinsic::arm_neon_vshiftu:
4698 // Opcode already set above.
4700 case Intrinsic::arm_neon_vshiftls:
4701 case Intrinsic::arm_neon_vshiftlu:
4702 if (Cnt == VT.getVectorElementType().getSizeInBits())
4703 VShiftOpc = ARMISD::VSHLLi;
4705 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4706 ARMISD::VSHLLs : ARMISD::VSHLLu);
4708 case Intrinsic::arm_neon_vshiftn:
4709 VShiftOpc = ARMISD::VSHRN; break;
4710 case Intrinsic::arm_neon_vrshifts:
4711 VShiftOpc = ARMISD::VRSHRs; break;
4712 case Intrinsic::arm_neon_vrshiftu:
4713 VShiftOpc = ARMISD::VRSHRu; break;
4714 case Intrinsic::arm_neon_vrshiftn:
4715 VShiftOpc = ARMISD::VRSHRN; break;
4716 case Intrinsic::arm_neon_vqshifts:
4717 VShiftOpc = ARMISD::VQSHLs; break;
4718 case Intrinsic::arm_neon_vqshiftu:
4719 VShiftOpc = ARMISD::VQSHLu; break;
4720 case Intrinsic::arm_neon_vqshiftsu:
4721 VShiftOpc = ARMISD::VQSHLsu; break;
4722 case Intrinsic::arm_neon_vqshiftns:
4723 VShiftOpc = ARMISD::VQSHRNs; break;
4724 case Intrinsic::arm_neon_vqshiftnu:
4725 VShiftOpc = ARMISD::VQSHRNu; break;
4726 case Intrinsic::arm_neon_vqshiftnsu:
4727 VShiftOpc = ARMISD::VQSHRNsu; break;
4728 case Intrinsic::arm_neon_vqrshiftns:
4729 VShiftOpc = ARMISD::VQRSHRNs; break;
4730 case Intrinsic::arm_neon_vqrshiftnu:
4731 VShiftOpc = ARMISD::VQRSHRNu; break;
4732 case Intrinsic::arm_neon_vqrshiftnsu:
4733 VShiftOpc = ARMISD::VQRSHRNsu; break;
4736 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4737 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4740 case Intrinsic::arm_neon_vshiftins: {
4741 EVT VT = N->getOperand(1).getValueType();
4743 unsigned VShiftOpc = 0;
4745 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4746 VShiftOpc = ARMISD::VSLI;
4747 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4748 VShiftOpc = ARMISD::VSRI;
4750 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4753 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4754 N->getOperand(1), N->getOperand(2),
4755 DAG.getConstant(Cnt, MVT::i32));
4758 case Intrinsic::arm_neon_vqrshifts:
4759 case Intrinsic::arm_neon_vqrshiftu:
4760 // No immediate versions of these to check for.
4767 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4768 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4769 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4770 /// vector element shift counts are generally not legal, and it is hard to see
4771 /// their values after they get legalized to loads from a constant pool.
4772 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4773 const ARMSubtarget *ST) {
4774 EVT VT = N->getValueType(0);
4776 // Nothing to be done for scalar shifts.
4777 if (! VT.isVector())
4780 assert(ST->hasNEON() && "unexpected vector shift");
4783 switch (N->getOpcode()) {
4784 default: llvm_unreachable("unexpected shift opcode");
4787 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4788 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4789 DAG.getConstant(Cnt, MVT::i32));
4794 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4795 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4796 ARMISD::VSHRs : ARMISD::VSHRu);
4797 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4798 DAG.getConstant(Cnt, MVT::i32));
4804 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4805 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4806 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4807 const ARMSubtarget *ST) {
4808 SDValue N0 = N->getOperand(0);
4810 // Check for sign- and zero-extensions of vector extract operations of 8-
4811 // and 16-bit vector elements. NEON supports these directly. They are
4812 // handled during DAG combining because type legalization will promote them
4813 // to 32-bit types and it is messy to recognize the operations after that.
4814 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4815 SDValue Vec = N0.getOperand(0);
4816 SDValue Lane = N0.getOperand(1);
4817 EVT VT = N->getValueType(0);
4818 EVT EltVT = N0.getValueType();
4819 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4821 if (VT == MVT::i32 &&
4822 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4823 TLI.isTypeLegal(Vec.getValueType())) {
4826 switch (N->getOpcode()) {
4827 default: llvm_unreachable("unexpected opcode");
4828 case ISD::SIGN_EXTEND:
4829 Opc = ARMISD::VGETLANEs;
4831 case ISD::ZERO_EXTEND:
4832 case ISD::ANY_EXTEND:
4833 Opc = ARMISD::VGETLANEu;
4836 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4843 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4844 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4845 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4846 const ARMSubtarget *ST) {
4847 // If the target supports NEON, try to use vmax/vmin instructions for f32
4848 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4849 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4850 // a NaN; only do the transformation when it matches that behavior.
4852 // For now only do this when using NEON for FP operations; if using VFP, it
4853 // is not obvious that the benefit outweighs the cost of switching to the
4855 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4856 N->getValueType(0) != MVT::f32)
4859 SDValue CondLHS = N->getOperand(0);
4860 SDValue CondRHS = N->getOperand(1);
4861 SDValue LHS = N->getOperand(2);
4862 SDValue RHS = N->getOperand(3);
4863 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4865 unsigned Opcode = 0;
4867 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4868 IsReversed = false; // x CC y ? x : y
4869 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4870 IsReversed = true ; // x CC y ? y : x
4884 // If LHS is NaN, an ordered comparison will be false and the result will
4885 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4886 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4887 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4888 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4890 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4891 // will return -0, so vmin can only be used for unsafe math or if one of
4892 // the operands is known to be nonzero.
4893 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4895 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4897 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4906 // If LHS is NaN, an ordered comparison will be false and the result will
4907 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4908 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4909 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4910 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4912 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4913 // will return +0, so vmax can only be used for unsafe math or if one of
4914 // the operands is known to be nonzero.
4915 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4917 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4919 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4925 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4928 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4929 DAGCombinerInfo &DCI) const {
4930 switch (N->getOpcode()) {
4932 case ISD::ADD: return PerformADDCombine(N, DCI);
4933 case ISD::SUB: return PerformSUBCombine(N, DCI);
4934 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4935 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4936 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4937 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
4938 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
4939 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
4940 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4943 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4944 case ISD::SIGN_EXTEND:
4945 case ISD::ZERO_EXTEND:
4946 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4947 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4952 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4953 if (!Subtarget->allowsUnalignedMem())
4956 switch (VT.getSimpleVT().SimpleTy) {
4963 // FIXME: VLD1 etc with standard alignment is legal.
4967 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4972 switch (VT.getSimpleVT().SimpleTy) {
4973 default: return false;
4988 if ((V & (Scale - 1)) != 0)
4991 return V == (V & ((1LL << 5) - 1));
4994 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4995 const ARMSubtarget *Subtarget) {
5002 switch (VT.getSimpleVT().SimpleTy) {
5003 default: return false;
5008 // + imm12 or - imm8
5010 return V == (V & ((1LL << 8) - 1));
5011 return V == (V & ((1LL << 12) - 1));
5014 // Same as ARM mode. FIXME: NEON?
5015 if (!Subtarget->hasVFP2())
5020 return V == (V & ((1LL << 8) - 1));
5024 /// isLegalAddressImmediate - Return true if the integer value can be used
5025 /// as the offset of the target addressing mode for load / store of the
5027 static bool isLegalAddressImmediate(int64_t V, EVT VT,
5028 const ARMSubtarget *Subtarget) {
5035 if (Subtarget->isThumb1Only())
5036 return isLegalT1AddressImmediate(V, VT);
5037 else if (Subtarget->isThumb2())
5038 return isLegalT2AddressImmediate(V, VT, Subtarget);
5043 switch (VT.getSimpleVT().SimpleTy) {
5044 default: return false;
5049 return V == (V & ((1LL << 12) - 1));
5052 return V == (V & ((1LL << 8) - 1));
5055 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5060 return V == (V & ((1LL << 8) - 1));
5064 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5066 int Scale = AM.Scale;
5070 switch (VT.getSimpleVT().SimpleTy) {
5071 default: return false;
5080 return Scale == 2 || Scale == 4 || Scale == 8;
5083 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5087 // Note, we allow "void" uses (basically, uses that aren't loads or
5088 // stores), because arm allows folding a scale into many arithmetic
5089 // operations. This should be made more precise and revisited later.
5091 // Allow r << imm, but the imm has to be a multiple of two.
5092 if (Scale & 1) return false;
5093 return isPowerOf2_32(Scale);
5097 /// isLegalAddressingMode - Return true if the addressing mode represented
5098 /// by AM is legal for this target, for a load/store of the specified type.
5099 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5100 const Type *Ty) const {
5101 EVT VT = getValueType(Ty, true);
5102 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5105 // Can never fold addr of global into load/store.
5110 case 0: // no scale reg, must be "r+i" or "r", or "i".
5113 if (Subtarget->isThumb1Only())
5117 // ARM doesn't support any R+R*scale+imm addr modes.
5124 if (Subtarget->isThumb2())
5125 return isLegalT2ScaledAddressingMode(AM, VT);
5127 int Scale = AM.Scale;
5128 switch (VT.getSimpleVT().SimpleTy) {
5129 default: return false;
5133 if (Scale < 0) Scale = -Scale;
5137 return isPowerOf2_32(Scale & ~1);
5141 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5146 // Note, we allow "void" uses (basically, uses that aren't loads or
5147 // stores), because arm allows folding a scale into many arithmetic
5148 // operations. This should be made more precise and revisited later.
5150 // Allow r << imm, but the imm has to be a multiple of two.
5151 if (Scale & 1) return false;
5152 return isPowerOf2_32(Scale);
5159 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5160 /// icmp immediate, that is the target has icmp instructions which can compare
5161 /// a register against the immediate without having to materialize the
5162 /// immediate into a register.
5163 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5164 if (!Subtarget->isThumb())
5165 return ARM_AM::getSOImmVal(Imm) != -1;
5166 if (Subtarget->isThumb2())
5167 return ARM_AM::getT2SOImmVal(Imm) != -1;
5168 return Imm >= 0 && Imm <= 255;
5171 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5172 bool isSEXTLoad, SDValue &Base,
5173 SDValue &Offset, bool &isInc,
5174 SelectionDAG &DAG) {
5175 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5178 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5180 Base = Ptr->getOperand(0);
5181 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5182 int RHSC = (int)RHS->getZExtValue();
5183 if (RHSC < 0 && RHSC > -256) {
5184 assert(Ptr->getOpcode() == ISD::ADD);
5186 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5190 isInc = (Ptr->getOpcode() == ISD::ADD);
5191 Offset = Ptr->getOperand(1);
5193 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5195 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5196 int RHSC = (int)RHS->getZExtValue();
5197 if (RHSC < 0 && RHSC > -0x1000) {
5198 assert(Ptr->getOpcode() == ISD::ADD);
5200 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5201 Base = Ptr->getOperand(0);
5206 if (Ptr->getOpcode() == ISD::ADD) {
5208 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5209 if (ShOpcVal != ARM_AM::no_shift) {
5210 Base = Ptr->getOperand(1);
5211 Offset = Ptr->getOperand(0);
5213 Base = Ptr->getOperand(0);
5214 Offset = Ptr->getOperand(1);
5219 isInc = (Ptr->getOpcode() == ISD::ADD);
5220 Base = Ptr->getOperand(0);
5221 Offset = Ptr->getOperand(1);
5225 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5229 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5230 bool isSEXTLoad, SDValue &Base,
5231 SDValue &Offset, bool &isInc,
5232 SelectionDAG &DAG) {
5233 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5236 Base = Ptr->getOperand(0);
5237 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5238 int RHSC = (int)RHS->getZExtValue();
5239 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5240 assert(Ptr->getOpcode() == ISD::ADD);
5242 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5244 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5245 isInc = Ptr->getOpcode() == ISD::ADD;
5246 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5254 /// getPreIndexedAddressParts - returns true by value, base pointer and
5255 /// offset pointer and addressing mode by reference if the node's address
5256 /// can be legally represented as pre-indexed load / store address.
5258 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5260 ISD::MemIndexedMode &AM,
5261 SelectionDAG &DAG) const {
5262 if (Subtarget->isThumb1Only())
5267 bool isSEXTLoad = false;
5268 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5269 Ptr = LD->getBasePtr();
5270 VT = LD->getMemoryVT();
5271 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5272 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5273 Ptr = ST->getBasePtr();
5274 VT = ST->getMemoryVT();
5279 bool isLegal = false;
5280 if (Subtarget->isThumb2())
5281 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5282 Offset, isInc, DAG);
5284 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5285 Offset, isInc, DAG);
5289 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5293 /// getPostIndexedAddressParts - returns true by value, base pointer and
5294 /// offset pointer and addressing mode by reference if this node can be
5295 /// combined with a load / store to form a post-indexed load / store.
5296 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5299 ISD::MemIndexedMode &AM,
5300 SelectionDAG &DAG) const {
5301 if (Subtarget->isThumb1Only())
5306 bool isSEXTLoad = false;
5307 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5308 VT = LD->getMemoryVT();
5309 Ptr = LD->getBasePtr();
5310 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5311 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5312 VT = ST->getMemoryVT();
5313 Ptr = ST->getBasePtr();
5318 bool isLegal = false;
5319 if (Subtarget->isThumb2())
5320 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5323 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5329 // Swap base ptr and offset to catch more post-index load / store when
5330 // it's legal. In Thumb2 mode, offset must be an immediate.
5331 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5332 !Subtarget->isThumb2())
5333 std::swap(Base, Offset);
5335 // Post-indexed load / store update the base pointer.
5340 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5344 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5348 const SelectionDAG &DAG,
5349 unsigned Depth) const {
5350 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5351 switch (Op.getOpcode()) {
5353 case ARMISD::CMOV: {
5354 // Bits are known zero/one if known on the LHS and RHS.
5355 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5356 if (KnownZero == 0 && KnownOne == 0) return;
5358 APInt KnownZeroRHS, KnownOneRHS;
5359 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5360 KnownZeroRHS, KnownOneRHS, Depth+1);
5361 KnownZero &= KnownZeroRHS;
5362 KnownOne &= KnownOneRHS;
5368 //===----------------------------------------------------------------------===//
5369 // ARM Inline Assembly Support
5370 //===----------------------------------------------------------------------===//
5372 /// getConstraintType - Given a constraint letter, return the type of
5373 /// constraint it is for this target.
5374 ARMTargetLowering::ConstraintType
5375 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5376 if (Constraint.size() == 1) {
5377 switch (Constraint[0]) {
5379 case 'l': return C_RegisterClass;
5380 case 'w': return C_RegisterClass;
5383 return TargetLowering::getConstraintType(Constraint);
5386 std::pair<unsigned, const TargetRegisterClass*>
5387 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5389 if (Constraint.size() == 1) {
5390 // GCC ARM Constraint Letters
5391 switch (Constraint[0]) {
5393 if (Subtarget->isThumb())
5394 return std::make_pair(0U, ARM::tGPRRegisterClass);
5396 return std::make_pair(0U, ARM::GPRRegisterClass);
5398 return std::make_pair(0U, ARM::GPRRegisterClass);
5401 return std::make_pair(0U, ARM::SPRRegisterClass);
5402 if (VT.getSizeInBits() == 64)
5403 return std::make_pair(0U, ARM::DPRRegisterClass);
5404 if (VT.getSizeInBits() == 128)
5405 return std::make_pair(0U, ARM::QPRRegisterClass);
5409 if (StringRef("{cc}").equals_lower(Constraint))
5410 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5412 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5415 std::vector<unsigned> ARMTargetLowering::
5416 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5418 if (Constraint.size() != 1)
5419 return std::vector<unsigned>();
5421 switch (Constraint[0]) { // GCC ARM Constraint Letters
5424 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5425 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5428 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5429 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5430 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5431 ARM::R12, ARM::LR, 0);
5434 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5435 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5436 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5437 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5438 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5439 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5440 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5441 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5442 if (VT.getSizeInBits() == 64)
5443 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5444 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5445 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5446 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5447 if (VT.getSizeInBits() == 128)
5448 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5449 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5453 return std::vector<unsigned>();
5456 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5457 /// vector. If it is invalid, don't add anything to Ops.
5458 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5460 std::vector<SDValue>&Ops,
5461 SelectionDAG &DAG) const {
5462 SDValue Result(0, 0);
5464 switch (Constraint) {
5466 case 'I': case 'J': case 'K': case 'L':
5467 case 'M': case 'N': case 'O':
5468 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5472 int64_t CVal64 = C->getSExtValue();
5473 int CVal = (int) CVal64;
5474 // None of these constraints allow values larger than 32 bits. Check
5475 // that the value fits in an int.
5479 switch (Constraint) {
5481 if (Subtarget->isThumb1Only()) {
5482 // This must be a constant between 0 and 255, for ADD
5484 if (CVal >= 0 && CVal <= 255)
5486 } else if (Subtarget->isThumb2()) {
5487 // A constant that can be used as an immediate value in a
5488 // data-processing instruction.
5489 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5492 // A constant that can be used as an immediate value in a
5493 // data-processing instruction.
5494 if (ARM_AM::getSOImmVal(CVal) != -1)
5500 if (Subtarget->isThumb()) { // FIXME thumb2
5501 // This must be a constant between -255 and -1, for negated ADD
5502 // immediates. This can be used in GCC with an "n" modifier that
5503 // prints the negated value, for use with SUB instructions. It is
5504 // not useful otherwise but is implemented for compatibility.
5505 if (CVal >= -255 && CVal <= -1)
5508 // This must be a constant between -4095 and 4095. It is not clear
5509 // what this constraint is intended for. Implemented for
5510 // compatibility with GCC.
5511 if (CVal >= -4095 && CVal <= 4095)
5517 if (Subtarget->isThumb1Only()) {
5518 // A 32-bit value where only one byte has a nonzero value. Exclude
5519 // zero to match GCC. This constraint is used by GCC internally for
5520 // constants that can be loaded with a move/shift combination.
5521 // It is not useful otherwise but is implemented for compatibility.
5522 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5524 } else if (Subtarget->isThumb2()) {
5525 // A constant whose bitwise inverse can be used as an immediate
5526 // value in a data-processing instruction. This can be used in GCC
5527 // with a "B" modifier that prints the inverted value, for use with
5528 // BIC and MVN instructions. It is not useful otherwise but is
5529 // implemented for compatibility.
5530 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5533 // A constant whose bitwise inverse can be used as an immediate
5534 // value in a data-processing instruction. This can be used in GCC
5535 // with a "B" modifier that prints the inverted value, for use with
5536 // BIC and MVN instructions. It is not useful otherwise but is
5537 // implemented for compatibility.
5538 if (ARM_AM::getSOImmVal(~CVal) != -1)
5544 if (Subtarget->isThumb1Only()) {
5545 // This must be a constant between -7 and 7,
5546 // for 3-operand ADD/SUB immediate instructions.
5547 if (CVal >= -7 && CVal < 7)
5549 } else if (Subtarget->isThumb2()) {
5550 // A constant whose negation can be used as an immediate value in a
5551 // data-processing instruction. This can be used in GCC with an "n"
5552 // modifier that prints the negated value, for use with SUB
5553 // instructions. It is not useful otherwise but is implemented for
5555 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5558 // A constant whose negation can be used as an immediate value in a
5559 // data-processing instruction. This can be used in GCC with an "n"
5560 // modifier that prints the negated value, for use with SUB
5561 // instructions. It is not useful otherwise but is implemented for
5563 if (ARM_AM::getSOImmVal(-CVal) != -1)
5569 if (Subtarget->isThumb()) { // FIXME thumb2
5570 // This must be a multiple of 4 between 0 and 1020, for
5571 // ADD sp + immediate.
5572 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5575 // A power of two or a constant between 0 and 32. This is used in
5576 // GCC for the shift amount on shifted register operands, but it is
5577 // useful in general for any shift amounts.
5578 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5584 if (Subtarget->isThumb()) { // FIXME thumb2
5585 // This must be a constant between 0 and 31, for shift amounts.
5586 if (CVal >= 0 && CVal <= 31)
5592 if (Subtarget->isThumb()) { // FIXME thumb2
5593 // This must be a multiple of 4 between -508 and 508, for
5594 // ADD/SUB sp = sp + immediate.
5595 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5600 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5604 if (Result.getNode()) {
5605 Ops.push_back(Result);
5608 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5612 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5613 // The ARM target isn't yet aware of offsets.
5617 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5618 APInt Imm = FPImm.bitcastToAPInt();
5619 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5620 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5621 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5623 // We can handle 4 bits of mantissa.
5624 // mantissa = (16+UInt(e:f:g:h))/16.
5625 if (Mantissa & 0x7ffff)
5628 if ((Mantissa & 0xf) != Mantissa)
5631 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5632 if (Exp < -3 || Exp > 4)
5634 Exp = ((Exp+3) & 0x7) ^ 4;
5636 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5639 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5640 APInt Imm = FPImm.bitcastToAPInt();
5641 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5642 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5643 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5645 // We can handle 4 bits of mantissa.
5646 // mantissa = (16+UInt(e:f:g:h))/16.
5647 if (Mantissa & 0xffffffffffffLL)
5650 if ((Mantissa & 0xf) != Mantissa)
5653 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5654 if (Exp < -3 || Exp > 4)
5656 Exp = ((Exp+3) & 0x7) ^ 4;
5658 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5661 bool ARM::isBitFieldInvertedMask(unsigned v) {
5662 if (v == 0xffffffff)
5664 // there can be 1's on either or both "outsides", all the "inside"
5666 unsigned int lsb = 0, msb = 31;
5667 while (v & (1 << msb)) --msb;
5668 while (v & (1 << lsb)) ++lsb;
5669 for (unsigned int i = lsb; i <= msb; ++i) {
5676 /// isFPImmLegal - Returns true if the target can instruction select the
5677 /// specified FP immediate natively. If false, the legalizer will
5678 /// materialize the FP immediate as a load from a constant pool.
5679 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5680 if (!Subtarget->hasVFP3())
5683 return ARM::getVFPf32Imm(Imm) != -1;
5685 return ARM::getVFPf64Imm(Imm) != -1;
5689 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5690 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5691 /// specified in the intrinsic calls.
5692 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5694 unsigned Intrinsic) const {
5695 switch (Intrinsic) {
5696 case Intrinsic::arm_neon_vld1:
5697 case Intrinsic::arm_neon_vld2:
5698 case Intrinsic::arm_neon_vld3:
5699 case Intrinsic::arm_neon_vld4:
5700 case Intrinsic::arm_neon_vld2lane:
5701 case Intrinsic::arm_neon_vld3lane:
5702 case Intrinsic::arm_neon_vld4lane: {
5703 Info.opc = ISD::INTRINSIC_W_CHAIN;
5704 // Conservatively set memVT to the entire set of vectors loaded.
5705 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
5706 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5707 Info.ptrVal = I.getArgOperand(0);
5709 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5710 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5711 Info.vol = false; // volatile loads with NEON intrinsics not supported
5712 Info.readMem = true;
5713 Info.writeMem = false;
5716 case Intrinsic::arm_neon_vst1:
5717 case Intrinsic::arm_neon_vst2:
5718 case Intrinsic::arm_neon_vst3:
5719 case Intrinsic::arm_neon_vst4:
5720 case Intrinsic::arm_neon_vst2lane:
5721 case Intrinsic::arm_neon_vst3lane:
5722 case Intrinsic::arm_neon_vst4lane: {
5723 Info.opc = ISD::INTRINSIC_VOID;
5724 // Conservatively set memVT to the entire set of vectors stored.
5725 unsigned NumElts = 0;
5726 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5727 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
5728 if (!ArgTy->isVectorTy())
5730 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
5732 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5733 Info.ptrVal = I.getArgOperand(0);
5735 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5736 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5737 Info.vol = false; // volatile stores with NEON intrinsics not supported
5738 Info.readMem = false;
5739 Info.writeMem = true;