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 "ARMCallingConv.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "MCTargetDesc/ARMAddressingModes.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/IntrinsicLowering.h"
37 #include "llvm/CodeGen/MachineBasicBlock.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/PseudoSourceValue.h"
43 #include "llvm/CodeGen/SelectionDAG.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Target/TargetOptions.h"
46 #include "llvm/ADT/VectorExtras.h"
47 #include "llvm/ADT/StringExtras.h"
48 #include "llvm/ADT/Statistic.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
56 STATISTIC(NumTailCalls, "Number of tail calls");
57 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
59 // This option should go away when tail calls fully work.
61 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
62 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
66 EnableARMLongCalls("arm-long-calls", cl::Hidden,
67 cl::desc("Generate calls via indirect call instructions"),
71 ARMInterworking("arm-interworking", cl::Hidden,
72 cl::desc("Enable / disable ARM interworking (for debugging only)"),
76 class ARMCCState : public CCState {
78 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
79 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
80 LLVMContext &C, ParmContext PC)
81 : CCState(CC, isVarArg, MF, TM, locs, C) {
82 assert(((PC == Call) || (PC == Prologue)) &&
83 "ARMCCState users must specify whether their context is call"
84 "or prologue generation.");
90 // The APCS parameter registers.
91 static const unsigned GPRArgRegs[] = {
92 ARM::R0, ARM::R1, ARM::R2, ARM::R3
95 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
96 EVT PromotedBitwiseVT) {
97 if (VT != PromotedLdStVT) {
98 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
99 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
100 PromotedLdStVT.getSimpleVT());
102 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
103 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
104 PromotedLdStVT.getSimpleVT());
107 EVT ElemTy = VT.getVectorElementType();
108 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
109 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
110 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
111 if (ElemTy != MVT::i32) {
112 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
119 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
120 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
121 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
122 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
123 if (VT.isInteger()) {
124 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
125 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
126 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
127 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
128 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
129 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
130 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
131 setTruncStoreAction(VT.getSimpleVT(),
132 (MVT::SimpleValueType)InnerVT, Expand);
134 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
136 // Promote all bit-wise operations.
137 if (VT.isInteger() && VT != PromotedBitwiseVT) {
138 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
139 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
140 PromotedBitwiseVT.getSimpleVT());
141 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
142 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
143 PromotedBitwiseVT.getSimpleVT());
144 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
145 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
146 PromotedBitwiseVT.getSimpleVT());
149 // Neon does not support vector divide/remainder operations.
150 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
154 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
155 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
158 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
159 addRegisterClass(VT, ARM::DPRRegisterClass);
160 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
163 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
164 addRegisterClass(VT, ARM::QPRRegisterClass);
165 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
168 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
169 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
170 return new TargetLoweringObjectFileMachO();
172 return new ARMElfTargetObjectFile();
175 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
176 : TargetLowering(TM, createTLOF(TM)) {
177 Subtarget = &TM.getSubtarget<ARMSubtarget>();
178 RegInfo = TM.getRegisterInfo();
179 Itins = TM.getInstrItineraryData();
181 if (Subtarget->isTargetDarwin()) {
182 // Uses VFP for Thumb libfuncs if available.
183 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
184 // Single-precision floating-point arithmetic.
185 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
186 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
187 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
188 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
190 // Double-precision floating-point arithmetic.
191 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
192 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
193 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
194 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
196 // Single-precision comparisons.
197 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
198 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
199 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
200 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
201 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
202 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
203 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
204 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
206 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
215 // Double-precision comparisons.
216 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
217 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
218 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
219 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
220 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
221 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
222 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
223 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
225 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
232 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
234 // Floating-point to integer conversions.
235 // i64 conversions are done via library routines even when generating VFP
236 // instructions, so use the same ones.
237 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
238 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
239 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
240 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
242 // Conversions between floating types.
243 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
244 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
246 // Integer to floating-point conversions.
247 // i64 conversions are done via library routines even when generating VFP
248 // instructions, so use the same ones.
249 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
250 // e.g., __floatunsidf vs. __floatunssidfvfp.
251 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
252 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
253 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
254 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
258 // These libcalls are not available in 32-bit.
259 setLibcallName(RTLIB::SHL_I128, 0);
260 setLibcallName(RTLIB::SRL_I128, 0);
261 setLibcallName(RTLIB::SRA_I128, 0);
263 if (Subtarget->isAAPCS_ABI()) {
264 // Double-precision floating-point arithmetic helper functions
265 // RTABI chapter 4.1.2, Table 2
266 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
267 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
268 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
269 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
270 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
275 // Double-precision floating-point comparison helper functions
276 // RTABI chapter 4.1.2, Table 3
277 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
278 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
279 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
280 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
281 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
282 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
284 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
285 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
286 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
287 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
288 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
289 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
290 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
291 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
292 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
293 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
302 // Single-precision floating-point arithmetic helper functions
303 // RTABI chapter 4.1.2, Table 4
304 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
305 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
306 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
307 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
308 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
313 // Single-precision floating-point comparison helper functions
314 // RTABI chapter 4.1.2, Table 5
315 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
316 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
317 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
318 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
319 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
320 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
322 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
323 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
324 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
325 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
326 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
327 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
328 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
329 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
330 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
331 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
340 // Floating-point to integer conversions.
341 // RTABI chapter 4.1.2, Table 6
342 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
343 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
345 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
346 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
347 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
348 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
349 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
350 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
359 // Conversions between floating types.
360 // RTABI chapter 4.1.2, Table 7
361 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
362 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
363 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
364 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
366 // Integer to floating-point conversions.
367 // RTABI chapter 4.1.2, Table 8
368 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
369 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
370 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
371 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
372 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
373 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
374 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
375 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
385 // Long long helper functions
386 // RTABI chapter 4.2, Table 9
387 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
388 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
389 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
390 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
391 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
392 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
393 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
396 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
397 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
398 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
400 // Integer division functions
401 // RTABI chapter 4.3.1
402 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
403 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
404 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
405 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
406 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
407 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
408 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
416 // RTABI chapter 4.3.4
417 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
418 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
419 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
422 if (Subtarget->isThumb1Only())
423 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
425 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
426 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
427 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
428 if (!Subtarget->isFPOnlySP())
429 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
431 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
434 if (Subtarget->hasNEON()) {
435 addDRTypeForNEON(MVT::v2f32);
436 addDRTypeForNEON(MVT::v8i8);
437 addDRTypeForNEON(MVT::v4i16);
438 addDRTypeForNEON(MVT::v2i32);
439 addDRTypeForNEON(MVT::v1i64);
441 addQRTypeForNEON(MVT::v4f32);
442 addQRTypeForNEON(MVT::v2f64);
443 addQRTypeForNEON(MVT::v16i8);
444 addQRTypeForNEON(MVT::v8i16);
445 addQRTypeForNEON(MVT::v4i32);
446 addQRTypeForNEON(MVT::v2i64);
448 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
449 // neither Neon nor VFP support any arithmetic operations on it.
450 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
451 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
452 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
453 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
454 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
455 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
456 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
457 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
458 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
459 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
460 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
461 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
462 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
463 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
464 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
465 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
466 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
467 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
468 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
469 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
470 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
471 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
472 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
473 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
475 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
477 // Neon does not support some operations on v1i64 and v2i64 types.
478 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
479 // Custom handling for some quad-vector types to detect VMULL.
480 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
481 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
482 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
483 // Custom handling for some vector types to avoid expensive expansions
484 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
485 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
486 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
487 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
488 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
489 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
490 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
491 // a destination type that is wider than the source.
492 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
493 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
495 setTargetDAGCombine(ISD::INTRINSIC_VOID);
496 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
497 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
498 setTargetDAGCombine(ISD::SHL);
499 setTargetDAGCombine(ISD::SRL);
500 setTargetDAGCombine(ISD::SRA);
501 setTargetDAGCombine(ISD::SIGN_EXTEND);
502 setTargetDAGCombine(ISD::ZERO_EXTEND);
503 setTargetDAGCombine(ISD::ANY_EXTEND);
504 setTargetDAGCombine(ISD::SELECT_CC);
505 setTargetDAGCombine(ISD::BUILD_VECTOR);
506 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
507 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
508 setTargetDAGCombine(ISD::STORE);
509 setTargetDAGCombine(ISD::FP_TO_SINT);
510 setTargetDAGCombine(ISD::FP_TO_UINT);
511 setTargetDAGCombine(ISD::FDIV);
514 computeRegisterProperties();
516 // ARM does not have f32 extending load.
517 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
519 // ARM does not have i1 sign extending load.
520 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
522 // ARM supports all 4 flavors of integer indexed load / store.
523 if (!Subtarget->isThumb1Only()) {
524 for (unsigned im = (unsigned)ISD::PRE_INC;
525 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
526 setIndexedLoadAction(im, MVT::i1, Legal);
527 setIndexedLoadAction(im, MVT::i8, Legal);
528 setIndexedLoadAction(im, MVT::i16, Legal);
529 setIndexedLoadAction(im, MVT::i32, Legal);
530 setIndexedStoreAction(im, MVT::i1, Legal);
531 setIndexedStoreAction(im, MVT::i8, Legal);
532 setIndexedStoreAction(im, MVT::i16, Legal);
533 setIndexedStoreAction(im, MVT::i32, Legal);
537 // i64 operation support.
538 setOperationAction(ISD::MUL, MVT::i64, Expand);
539 setOperationAction(ISD::MULHU, MVT::i32, Expand);
540 if (Subtarget->isThumb1Only()) {
541 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
542 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
544 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
545 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
546 setOperationAction(ISD::MULHS, MVT::i32, Expand);
548 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
549 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
550 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
551 setOperationAction(ISD::SRL, MVT::i64, Custom);
552 setOperationAction(ISD::SRA, MVT::i64, Custom);
554 // ARM does not have ROTL.
555 setOperationAction(ISD::ROTL, MVT::i32, Expand);
556 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
557 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
558 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
559 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
561 // Only ARMv6 has BSWAP.
562 if (!Subtarget->hasV6Ops())
563 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
565 // These are expanded into libcalls.
566 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
567 // v7M has a hardware divider
568 setOperationAction(ISD::SDIV, MVT::i32, Expand);
569 setOperationAction(ISD::UDIV, MVT::i32, Expand);
571 setOperationAction(ISD::SREM, MVT::i32, Expand);
572 setOperationAction(ISD::UREM, MVT::i32, Expand);
573 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
574 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
576 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
577 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
578 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
579 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
580 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
582 setOperationAction(ISD::TRAP, MVT::Other, Legal);
584 // Use the default implementation.
585 setOperationAction(ISD::VASTART, MVT::Other, Custom);
586 setOperationAction(ISD::VAARG, MVT::Other, Expand);
587 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
588 setOperationAction(ISD::VAEND, MVT::Other, Expand);
589 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
590 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
591 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
592 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
593 setExceptionPointerRegister(ARM::R0);
594 setExceptionSelectorRegister(ARM::R1);
596 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
597 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
598 // the default expansion.
599 if (Subtarget->hasDataBarrier() ||
600 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
601 // membarrier needs custom lowering; the rest are legal and handled
603 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
604 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
605 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
606 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom);
607 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom);
608 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
609 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
610 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom);
611 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom);
612 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom);
613 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom);
614 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom);
615 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom);
616 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
618 // Set them all for expansion, which will force libcalls.
619 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
620 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
621 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
622 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
623 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
624 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
625 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
626 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
627 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
628 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
629 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
630 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
631 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
632 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
633 // Since the libcalls include locking, fold in the fences
634 setShouldFoldAtomicFences(true);
637 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
639 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
640 if (!Subtarget->hasV6Ops()) {
641 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
642 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
644 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
646 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
647 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
648 // iff target supports vfp2.
649 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
650 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
653 // We want to custom lower some of our intrinsics.
654 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
655 if (Subtarget->isTargetDarwin()) {
656 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
657 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
658 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
659 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
662 setOperationAction(ISD::SETCC, MVT::i32, Expand);
663 setOperationAction(ISD::SETCC, MVT::f32, Expand);
664 setOperationAction(ISD::SETCC, MVT::f64, Expand);
665 setOperationAction(ISD::SELECT, MVT::i32, Custom);
666 setOperationAction(ISD::SELECT, MVT::f32, Custom);
667 setOperationAction(ISD::SELECT, MVT::f64, Custom);
668 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
669 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
670 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
672 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
673 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
674 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
675 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
676 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
678 // We don't support sin/cos/fmod/copysign/pow
679 setOperationAction(ISD::FSIN, MVT::f64, Expand);
680 setOperationAction(ISD::FSIN, MVT::f32, Expand);
681 setOperationAction(ISD::FCOS, MVT::f32, Expand);
682 setOperationAction(ISD::FCOS, MVT::f64, Expand);
683 setOperationAction(ISD::FREM, MVT::f64, Expand);
684 setOperationAction(ISD::FREM, MVT::f32, Expand);
685 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
686 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
687 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
689 setOperationAction(ISD::FPOW, MVT::f64, Expand);
690 setOperationAction(ISD::FPOW, MVT::f32, Expand);
692 setOperationAction(ISD::FMA, MVT::f64, Expand);
693 setOperationAction(ISD::FMA, MVT::f32, Expand);
695 // Various VFP goodness
696 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
697 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
698 if (Subtarget->hasVFP2()) {
699 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
700 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
701 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
702 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
704 // Special handling for half-precision FP.
705 if (!Subtarget->hasFP16()) {
706 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
707 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
711 // We have target-specific dag combine patterns for the following nodes:
712 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
713 setTargetDAGCombine(ISD::ADD);
714 setTargetDAGCombine(ISD::SUB);
715 setTargetDAGCombine(ISD::MUL);
717 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
718 setTargetDAGCombine(ISD::OR);
719 if (Subtarget->hasNEON())
720 setTargetDAGCombine(ISD::AND);
722 setStackPointerRegisterToSaveRestore(ARM::SP);
724 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
725 setSchedulingPreference(Sched::RegPressure);
727 setSchedulingPreference(Sched::Hybrid);
729 //// temporary - rewrite interface to use type
730 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
732 // On ARM arguments smaller than 4 bytes are extended, so all arguments
733 // are at least 4 bytes aligned.
734 setMinStackArgumentAlignment(4);
736 benefitFromCodePlacementOpt = true;
738 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
741 // FIXME: It might make sense to define the representative register class as the
742 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
743 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
744 // SPR's representative would be DPR_VFP2. This should work well if register
745 // pressure tracking were modified such that a register use would increment the
746 // pressure of the register class's representative and all of it's super
747 // classes' representatives transitively. We have not implemented this because
748 // of the difficulty prior to coalescing of modeling operand register classes
749 // due to the common occurrence of cross class copies and subregister insertions
751 std::pair<const TargetRegisterClass*, uint8_t>
752 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
753 const TargetRegisterClass *RRC = 0;
755 switch (VT.getSimpleVT().SimpleTy) {
757 return TargetLowering::findRepresentativeClass(VT);
758 // Use DPR as representative register class for all floating point
759 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
760 // the cost is 1 for both f32 and f64.
761 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
762 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
763 RRC = ARM::DPRRegisterClass;
764 // When NEON is used for SP, only half of the register file is available
765 // because operations that define both SP and DP results will be constrained
766 // to the VFP2 class (D0-D15). We currently model this constraint prior to
767 // coalescing by double-counting the SP regs. See the FIXME above.
768 if (Subtarget->useNEONForSinglePrecisionFP())
771 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
772 case MVT::v4f32: case MVT::v2f64:
773 RRC = ARM::DPRRegisterClass;
777 RRC = ARM::DPRRegisterClass;
781 RRC = ARM::DPRRegisterClass;
785 return std::make_pair(RRC, Cost);
788 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
791 case ARMISD::Wrapper: return "ARMISD::Wrapper";
792 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
793 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
794 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
795 case ARMISD::CALL: return "ARMISD::CALL";
796 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
797 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
798 case ARMISD::tCALL: return "ARMISD::tCALL";
799 case ARMISD::BRCOND: return "ARMISD::BRCOND";
800 case ARMISD::BR_JT: return "ARMISD::BR_JT";
801 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
802 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
803 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
804 case ARMISD::CMP: return "ARMISD::CMP";
805 case ARMISD::CMPZ: return "ARMISD::CMPZ";
806 case ARMISD::CMPFP: return "ARMISD::CMPFP";
807 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
808 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
809 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
810 case ARMISD::CMOV: return "ARMISD::CMOV";
812 case ARMISD::RBIT: return "ARMISD::RBIT";
814 case ARMISD::FTOSI: return "ARMISD::FTOSI";
815 case ARMISD::FTOUI: return "ARMISD::FTOUI";
816 case ARMISD::SITOF: return "ARMISD::SITOF";
817 case ARMISD::UITOF: return "ARMISD::UITOF";
819 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
820 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
821 case ARMISD::RRX: return "ARMISD::RRX";
823 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
824 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
826 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
827 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
828 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
830 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
832 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
834 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
836 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
837 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
839 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
841 case ARMISD::VCEQ: return "ARMISD::VCEQ";
842 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
843 case ARMISD::VCGE: return "ARMISD::VCGE";
844 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
845 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
846 case ARMISD::VCGEU: return "ARMISD::VCGEU";
847 case ARMISD::VCGT: return "ARMISD::VCGT";
848 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
849 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
850 case ARMISD::VCGTU: return "ARMISD::VCGTU";
851 case ARMISD::VTST: return "ARMISD::VTST";
853 case ARMISD::VSHL: return "ARMISD::VSHL";
854 case ARMISD::VSHRs: return "ARMISD::VSHRs";
855 case ARMISD::VSHRu: return "ARMISD::VSHRu";
856 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
857 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
858 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
859 case ARMISD::VSHRN: return "ARMISD::VSHRN";
860 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
861 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
862 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
863 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
864 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
865 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
866 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
867 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
868 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
869 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
870 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
871 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
872 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
873 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
874 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
875 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
876 case ARMISD::VDUP: return "ARMISD::VDUP";
877 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
878 case ARMISD::VEXT: return "ARMISD::VEXT";
879 case ARMISD::VREV64: return "ARMISD::VREV64";
880 case ARMISD::VREV32: return "ARMISD::VREV32";
881 case ARMISD::VREV16: return "ARMISD::VREV16";
882 case ARMISD::VZIP: return "ARMISD::VZIP";
883 case ARMISD::VUZP: return "ARMISD::VUZP";
884 case ARMISD::VTRN: return "ARMISD::VTRN";
885 case ARMISD::VTBL1: return "ARMISD::VTBL1";
886 case ARMISD::VTBL2: return "ARMISD::VTBL2";
887 case ARMISD::VMULLs: return "ARMISD::VMULLs";
888 case ARMISD::VMULLu: return "ARMISD::VMULLu";
889 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
890 case ARMISD::FMAX: return "ARMISD::FMAX";
891 case ARMISD::FMIN: return "ARMISD::FMIN";
892 case ARMISD::BFI: return "ARMISD::BFI";
893 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
894 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
895 case ARMISD::VBSL: return "ARMISD::VBSL";
896 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
897 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
898 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
899 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
900 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
901 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
902 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
903 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
904 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
905 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
906 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
907 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
908 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
909 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
910 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
911 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
912 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
913 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
914 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
915 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
919 /// getRegClassFor - Return the register class that should be used for the
920 /// specified value type.
921 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
922 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
923 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
924 // load / store 4 to 8 consecutive D registers.
925 if (Subtarget->hasNEON()) {
926 if (VT == MVT::v4i64)
927 return ARM::QQPRRegisterClass;
928 else if (VT == MVT::v8i64)
929 return ARM::QQQQPRRegisterClass;
931 return TargetLowering::getRegClassFor(VT);
934 // Create a fast isel object.
936 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
937 return ARM::createFastISel(funcInfo);
940 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
941 /// be used for loads / stores from the global.
942 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
943 return (Subtarget->isThumb1Only() ? 127 : 4095);
946 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
947 unsigned NumVals = N->getNumValues();
949 return Sched::RegPressure;
951 for (unsigned i = 0; i != NumVals; ++i) {
952 EVT VT = N->getValueType(i);
953 if (VT == MVT::Glue || VT == MVT::Other)
955 if (VT.isFloatingPoint() || VT.isVector())
956 return Sched::Latency;
959 if (!N->isMachineOpcode())
960 return Sched::RegPressure;
962 // Load are scheduled for latency even if there instruction itinerary
964 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
965 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
967 if (MCID.getNumDefs() == 0)
968 return Sched::RegPressure;
969 if (!Itins->isEmpty() &&
970 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
971 return Sched::Latency;
973 return Sched::RegPressure;
976 //===----------------------------------------------------------------------===//
978 //===----------------------------------------------------------------------===//
980 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
981 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
983 default: llvm_unreachable("Unknown condition code!");
984 case ISD::SETNE: return ARMCC::NE;
985 case ISD::SETEQ: return ARMCC::EQ;
986 case ISD::SETGT: return ARMCC::GT;
987 case ISD::SETGE: return ARMCC::GE;
988 case ISD::SETLT: return ARMCC::LT;
989 case ISD::SETLE: return ARMCC::LE;
990 case ISD::SETUGT: return ARMCC::HI;
991 case ISD::SETUGE: return ARMCC::HS;
992 case ISD::SETULT: return ARMCC::LO;
993 case ISD::SETULE: return ARMCC::LS;
997 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
998 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
999 ARMCC::CondCodes &CondCode2) {
1000 CondCode2 = ARMCC::AL;
1002 default: llvm_unreachable("Unknown FP condition!");
1004 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1006 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1008 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1009 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1010 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1011 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1012 case ISD::SETO: CondCode = ARMCC::VC; break;
1013 case ISD::SETUO: CondCode = ARMCC::VS; break;
1014 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1015 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1016 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1018 case ISD::SETULT: CondCode = ARMCC::LT; break;
1020 case ISD::SETULE: CondCode = ARMCC::LE; break;
1022 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1026 //===----------------------------------------------------------------------===//
1027 // Calling Convention Implementation
1028 //===----------------------------------------------------------------------===//
1030 #include "ARMGenCallingConv.inc"
1032 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1033 /// given CallingConvention value.
1034 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1036 bool isVarArg) const {
1039 llvm_unreachable("Unsupported calling convention");
1040 case CallingConv::Fast:
1041 if (Subtarget->hasVFP2() && !isVarArg) {
1042 if (!Subtarget->isAAPCS_ABI())
1043 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1044 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1045 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1048 case CallingConv::C: {
1049 // Use target triple & subtarget features to do actual dispatch.
1050 if (!Subtarget->isAAPCS_ABI())
1051 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1052 else if (Subtarget->hasVFP2() &&
1053 FloatABIType == FloatABI::Hard && !isVarArg)
1054 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1055 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1057 case CallingConv::ARM_AAPCS_VFP:
1058 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1059 case CallingConv::ARM_AAPCS:
1060 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1061 case CallingConv::ARM_APCS:
1062 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1066 /// LowerCallResult - Lower the result values of a call into the
1067 /// appropriate copies out of appropriate physical registers.
1069 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1070 CallingConv::ID CallConv, bool isVarArg,
1071 const SmallVectorImpl<ISD::InputArg> &Ins,
1072 DebugLoc dl, SelectionDAG &DAG,
1073 SmallVectorImpl<SDValue> &InVals) const {
1075 // Assign locations to each value returned by this call.
1076 SmallVector<CCValAssign, 16> RVLocs;
1077 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1078 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1079 CCInfo.AnalyzeCallResult(Ins,
1080 CCAssignFnForNode(CallConv, /* Return*/ true,
1083 // Copy all of the result registers out of their specified physreg.
1084 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1085 CCValAssign VA = RVLocs[i];
1088 if (VA.needsCustom()) {
1089 // Handle f64 or half of a v2f64.
1090 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1092 Chain = Lo.getValue(1);
1093 InFlag = Lo.getValue(2);
1094 VA = RVLocs[++i]; // skip ahead to next loc
1095 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1097 Chain = Hi.getValue(1);
1098 InFlag = Hi.getValue(2);
1099 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1101 if (VA.getLocVT() == MVT::v2f64) {
1102 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1103 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1104 DAG.getConstant(0, MVT::i32));
1106 VA = RVLocs[++i]; // skip ahead to next loc
1107 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1108 Chain = Lo.getValue(1);
1109 InFlag = Lo.getValue(2);
1110 VA = RVLocs[++i]; // skip ahead to next loc
1111 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1112 Chain = Hi.getValue(1);
1113 InFlag = Hi.getValue(2);
1114 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1115 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1116 DAG.getConstant(1, MVT::i32));
1119 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1121 Chain = Val.getValue(1);
1122 InFlag = Val.getValue(2);
1125 switch (VA.getLocInfo()) {
1126 default: llvm_unreachable("Unknown loc info!");
1127 case CCValAssign::Full: break;
1128 case CCValAssign::BCvt:
1129 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1133 InVals.push_back(Val);
1139 /// LowerMemOpCallTo - Store the argument to the stack.
1141 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1142 SDValue StackPtr, SDValue Arg,
1143 DebugLoc dl, SelectionDAG &DAG,
1144 const CCValAssign &VA,
1145 ISD::ArgFlagsTy Flags) const {
1146 unsigned LocMemOffset = VA.getLocMemOffset();
1147 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1148 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1149 return DAG.getStore(Chain, dl, Arg, PtrOff,
1150 MachinePointerInfo::getStack(LocMemOffset),
1154 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1155 SDValue Chain, SDValue &Arg,
1156 RegsToPassVector &RegsToPass,
1157 CCValAssign &VA, CCValAssign &NextVA,
1159 SmallVector<SDValue, 8> &MemOpChains,
1160 ISD::ArgFlagsTy Flags) const {
1162 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1163 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1164 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1166 if (NextVA.isRegLoc())
1167 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1169 assert(NextVA.isMemLoc());
1170 if (StackPtr.getNode() == 0)
1171 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1173 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1179 /// LowerCall - Lowering a call into a callseq_start <-
1180 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1183 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1184 CallingConv::ID CallConv, bool isVarArg,
1186 const SmallVectorImpl<ISD::OutputArg> &Outs,
1187 const SmallVectorImpl<SDValue> &OutVals,
1188 const SmallVectorImpl<ISD::InputArg> &Ins,
1189 DebugLoc dl, SelectionDAG &DAG,
1190 SmallVectorImpl<SDValue> &InVals) const {
1191 MachineFunction &MF = DAG.getMachineFunction();
1192 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1193 bool IsSibCall = false;
1194 // Temporarily disable tail calls so things don't break.
1195 if (!EnableARMTailCalls)
1198 // Check if it's really possible to do a tail call.
1199 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1200 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1201 Outs, OutVals, Ins, DAG);
1202 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1203 // detected sibcalls.
1210 // Analyze operands of the call, assigning locations to each operand.
1211 SmallVector<CCValAssign, 16> ArgLocs;
1212 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1213 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1214 CCInfo.AnalyzeCallOperands(Outs,
1215 CCAssignFnForNode(CallConv, /* Return*/ false,
1218 // Get a count of how many bytes are to be pushed on the stack.
1219 unsigned NumBytes = CCInfo.getNextStackOffset();
1221 // For tail calls, memory operands are available in our caller's stack.
1225 // Adjust the stack pointer for the new arguments...
1226 // These operations are automatically eliminated by the prolog/epilog pass
1228 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1230 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1232 RegsToPassVector RegsToPass;
1233 SmallVector<SDValue, 8> MemOpChains;
1235 // Walk the register/memloc assignments, inserting copies/loads. In the case
1236 // of tail call optimization, arguments are handled later.
1237 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1239 ++i, ++realArgIdx) {
1240 CCValAssign &VA = ArgLocs[i];
1241 SDValue Arg = OutVals[realArgIdx];
1242 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1243 bool isByVal = Flags.isByVal();
1245 // Promote the value if needed.
1246 switch (VA.getLocInfo()) {
1247 default: llvm_unreachable("Unknown loc info!");
1248 case CCValAssign::Full: break;
1249 case CCValAssign::SExt:
1250 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1252 case CCValAssign::ZExt:
1253 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1255 case CCValAssign::AExt:
1256 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1258 case CCValAssign::BCvt:
1259 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1263 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1264 if (VA.needsCustom()) {
1265 if (VA.getLocVT() == MVT::v2f64) {
1266 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1267 DAG.getConstant(0, MVT::i32));
1268 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1269 DAG.getConstant(1, MVT::i32));
1271 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1272 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1274 VA = ArgLocs[++i]; // skip ahead to next loc
1275 if (VA.isRegLoc()) {
1276 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1277 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1279 assert(VA.isMemLoc());
1281 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1282 dl, DAG, VA, Flags));
1285 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1286 StackPtr, MemOpChains, Flags);
1288 } else if (VA.isRegLoc()) {
1289 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1290 } else if (isByVal) {
1291 assert(VA.isMemLoc());
1292 unsigned offset = 0;
1294 // True if this byval aggregate will be split between registers
1296 if (CCInfo.isFirstByValRegValid()) {
1297 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1299 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1300 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1301 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1302 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1303 MachinePointerInfo(),
1305 MemOpChains.push_back(Load.getValue(1));
1306 RegsToPass.push_back(std::make_pair(j, Load));
1308 offset = ARM::R4 - CCInfo.getFirstByValReg();
1309 CCInfo.clearFirstByValReg();
1312 unsigned LocMemOffset = VA.getLocMemOffset();
1313 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1314 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1316 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1317 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1318 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1320 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1321 Flags.getByValAlign(),
1322 /*isVolatile=*/false,
1323 /*AlwaysInline=*/false,
1324 MachinePointerInfo(0),
1325 MachinePointerInfo(0)));
1327 } else if (!IsSibCall) {
1328 assert(VA.isMemLoc());
1330 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1331 dl, DAG, VA, Flags));
1335 if (!MemOpChains.empty())
1336 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1337 &MemOpChains[0], MemOpChains.size());
1339 // Build a sequence of copy-to-reg nodes chained together with token chain
1340 // and flag operands which copy the outgoing args into the appropriate regs.
1342 // Tail call byval lowering might overwrite argument registers so in case of
1343 // tail call optimization the copies to registers are lowered later.
1345 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1346 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1347 RegsToPass[i].second, InFlag);
1348 InFlag = Chain.getValue(1);
1351 // For tail calls lower the arguments to the 'real' stack slot.
1353 // Force all the incoming stack arguments to be loaded from the stack
1354 // before any new outgoing arguments are stored to the stack, because the
1355 // outgoing stack slots may alias the incoming argument stack slots, and
1356 // the alias isn't otherwise explicit. This is slightly more conservative
1357 // than necessary, because it means that each store effectively depends
1358 // on every argument instead of just those arguments it would clobber.
1360 // Do not flag preceding copytoreg stuff together with the following stuff.
1362 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1363 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1364 RegsToPass[i].second, InFlag);
1365 InFlag = Chain.getValue(1);
1370 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1371 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1372 // node so that legalize doesn't hack it.
1373 bool isDirect = false;
1374 bool isARMFunc = false;
1375 bool isLocalARMFunc = false;
1376 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1378 if (EnableARMLongCalls) {
1379 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1380 && "long-calls with non-static relocation model!");
1381 // Handle a global address or an external symbol. If it's not one of
1382 // those, the target's already in a register, so we don't need to do
1384 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1385 const GlobalValue *GV = G->getGlobal();
1386 // Create a constant pool entry for the callee address
1387 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1388 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1391 // Get the address of the callee into a register
1392 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1393 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1394 Callee = DAG.getLoad(getPointerTy(), dl,
1395 DAG.getEntryNode(), CPAddr,
1396 MachinePointerInfo::getConstantPool(),
1398 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1399 const char *Sym = S->getSymbol();
1401 // Create a constant pool entry for the callee address
1402 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1403 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1404 Sym, ARMPCLabelIndex, 0);
1405 // Get the address of the callee into a register
1406 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1407 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1408 Callee = DAG.getLoad(getPointerTy(), dl,
1409 DAG.getEntryNode(), CPAddr,
1410 MachinePointerInfo::getConstantPool(),
1413 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1414 const GlobalValue *GV = G->getGlobal();
1416 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1417 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1418 getTargetMachine().getRelocationModel() != Reloc::Static;
1419 isARMFunc = !Subtarget->isThumb() || isStub;
1420 // ARM call to a local ARM function is predicable.
1421 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1422 // tBX takes a register source operand.
1423 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1424 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1425 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1428 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1429 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1430 Callee = DAG.getLoad(getPointerTy(), dl,
1431 DAG.getEntryNode(), CPAddr,
1432 MachinePointerInfo::getConstantPool(),
1434 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1435 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1436 getPointerTy(), Callee, PICLabel);
1438 // On ELF targets for PIC code, direct calls should go through the PLT
1439 unsigned OpFlags = 0;
1440 if (Subtarget->isTargetELF() &&
1441 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1442 OpFlags = ARMII::MO_PLT;
1443 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1445 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1447 bool isStub = Subtarget->isTargetDarwin() &&
1448 getTargetMachine().getRelocationModel() != Reloc::Static;
1449 isARMFunc = !Subtarget->isThumb() || isStub;
1450 // tBX takes a register source operand.
1451 const char *Sym = S->getSymbol();
1452 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1453 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1454 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1455 Sym, ARMPCLabelIndex, 4);
1456 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1457 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1458 Callee = DAG.getLoad(getPointerTy(), dl,
1459 DAG.getEntryNode(), CPAddr,
1460 MachinePointerInfo::getConstantPool(),
1462 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1463 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1464 getPointerTy(), Callee, PICLabel);
1466 unsigned OpFlags = 0;
1467 // On ELF targets for PIC code, direct calls should go through the PLT
1468 if (Subtarget->isTargetELF() &&
1469 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1470 OpFlags = ARMII::MO_PLT;
1471 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1475 // FIXME: handle tail calls differently.
1477 if (Subtarget->isThumb()) {
1478 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1479 CallOpc = ARMISD::CALL_NOLINK;
1481 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1483 CallOpc = (isDirect || Subtarget->hasV5TOps())
1484 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1485 : ARMISD::CALL_NOLINK;
1488 std::vector<SDValue> Ops;
1489 Ops.push_back(Chain);
1490 Ops.push_back(Callee);
1492 // Add argument registers to the end of the list so that they are known live
1494 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1495 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1496 RegsToPass[i].second.getValueType()));
1498 if (InFlag.getNode())
1499 Ops.push_back(InFlag);
1501 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1503 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1505 // Returns a chain and a flag for retval copy to use.
1506 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1507 InFlag = Chain.getValue(1);
1509 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1510 DAG.getIntPtrConstant(0, true), InFlag);
1512 InFlag = Chain.getValue(1);
1514 // Handle result values, copying them out of physregs into vregs that we
1516 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1520 /// HandleByVal - Every parameter *after* a byval parameter is passed
1521 /// on the stack. Remember the next parameter register to allocate,
1522 /// and then confiscate the rest of the parameter registers to insure
1525 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1526 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1527 assert((State->getCallOrPrologue() == Prologue ||
1528 State->getCallOrPrologue() == Call) &&
1529 "unhandled ParmContext");
1530 if ((!State->isFirstByValRegValid()) &&
1531 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1532 State->setFirstByValReg(reg);
1533 // At a call site, a byval parameter that is split between
1534 // registers and memory needs its size truncated here. In a
1535 // function prologue, such byval parameters are reassembled in
1536 // memory, and are not truncated.
1537 if (State->getCallOrPrologue() == Call) {
1538 unsigned excess = 4 * (ARM::R4 - reg);
1539 assert(size >= excess && "expected larger existing stack allocation");
1543 // Confiscate any remaining parameter registers to preclude their
1544 // assignment to subsequent parameters.
1545 while (State->AllocateReg(GPRArgRegs, 4))
1549 /// MatchingStackOffset - Return true if the given stack call argument is
1550 /// already available in the same position (relatively) of the caller's
1551 /// incoming argument stack.
1553 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1554 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1555 const ARMInstrInfo *TII) {
1556 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1558 if (Arg.getOpcode() == ISD::CopyFromReg) {
1559 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1560 if (!TargetRegisterInfo::isVirtualRegister(VR))
1562 MachineInstr *Def = MRI->getVRegDef(VR);
1565 if (!Flags.isByVal()) {
1566 if (!TII->isLoadFromStackSlot(Def, FI))
1571 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1572 if (Flags.isByVal())
1573 // ByVal argument is passed in as a pointer but it's now being
1574 // dereferenced. e.g.
1575 // define @foo(%struct.X* %A) {
1576 // tail call @bar(%struct.X* byval %A)
1579 SDValue Ptr = Ld->getBasePtr();
1580 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1583 FI = FINode->getIndex();
1587 assert(FI != INT_MAX);
1588 if (!MFI->isFixedObjectIndex(FI))
1590 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1593 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1594 /// for tail call optimization. Targets which want to do tail call
1595 /// optimization should implement this function.
1597 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1598 CallingConv::ID CalleeCC,
1600 bool isCalleeStructRet,
1601 bool isCallerStructRet,
1602 const SmallVectorImpl<ISD::OutputArg> &Outs,
1603 const SmallVectorImpl<SDValue> &OutVals,
1604 const SmallVectorImpl<ISD::InputArg> &Ins,
1605 SelectionDAG& DAG) const {
1606 const Function *CallerF = DAG.getMachineFunction().getFunction();
1607 CallingConv::ID CallerCC = CallerF->getCallingConv();
1608 bool CCMatch = CallerCC == CalleeCC;
1610 // Look for obvious safe cases to perform tail call optimization that do not
1611 // require ABI changes. This is what gcc calls sibcall.
1613 // Do not sibcall optimize vararg calls unless the call site is not passing
1615 if (isVarArg && !Outs.empty())
1618 // Also avoid sibcall optimization if either caller or callee uses struct
1619 // return semantics.
1620 if (isCalleeStructRet || isCallerStructRet)
1623 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1624 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1625 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1626 // support in the assembler and linker to be used. This would need to be
1627 // fixed to fully support tail calls in Thumb1.
1629 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1630 // LR. This means if we need to reload LR, it takes an extra instructions,
1631 // which outweighs the value of the tail call; but here we don't know yet
1632 // whether LR is going to be used. Probably the right approach is to
1633 // generate the tail call here and turn it back into CALL/RET in
1634 // emitEpilogue if LR is used.
1636 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1637 // but we need to make sure there are enough registers; the only valid
1638 // registers are the 4 used for parameters. We don't currently do this
1640 if (Subtarget->isThumb1Only())
1643 // If the calling conventions do not match, then we'd better make sure the
1644 // results are returned in the same way as what the caller expects.
1646 SmallVector<CCValAssign, 16> RVLocs1;
1647 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1648 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1649 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1651 SmallVector<CCValAssign, 16> RVLocs2;
1652 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1653 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1654 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1656 if (RVLocs1.size() != RVLocs2.size())
1658 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1659 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1661 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1663 if (RVLocs1[i].isRegLoc()) {
1664 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1667 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1673 // If the callee takes no arguments then go on to check the results of the
1675 if (!Outs.empty()) {
1676 // Check if stack adjustment is needed. For now, do not do this if any
1677 // argument is passed on the stack.
1678 SmallVector<CCValAssign, 16> ArgLocs;
1679 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1680 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1681 CCInfo.AnalyzeCallOperands(Outs,
1682 CCAssignFnForNode(CalleeCC, false, isVarArg));
1683 if (CCInfo.getNextStackOffset()) {
1684 MachineFunction &MF = DAG.getMachineFunction();
1686 // Check if the arguments are already laid out in the right way as
1687 // the caller's fixed stack objects.
1688 MachineFrameInfo *MFI = MF.getFrameInfo();
1689 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1690 const ARMInstrInfo *TII =
1691 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1692 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1694 ++i, ++realArgIdx) {
1695 CCValAssign &VA = ArgLocs[i];
1696 EVT RegVT = VA.getLocVT();
1697 SDValue Arg = OutVals[realArgIdx];
1698 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1699 if (VA.getLocInfo() == CCValAssign::Indirect)
1701 if (VA.needsCustom()) {
1702 // f64 and vector types are split into multiple registers or
1703 // register/stack-slot combinations. The types will not match
1704 // the registers; give up on memory f64 refs until we figure
1705 // out what to do about this.
1708 if (!ArgLocs[++i].isRegLoc())
1710 if (RegVT == MVT::v2f64) {
1711 if (!ArgLocs[++i].isRegLoc())
1713 if (!ArgLocs[++i].isRegLoc())
1716 } else if (!VA.isRegLoc()) {
1717 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1729 ARMTargetLowering::LowerReturn(SDValue Chain,
1730 CallingConv::ID CallConv, bool isVarArg,
1731 const SmallVectorImpl<ISD::OutputArg> &Outs,
1732 const SmallVectorImpl<SDValue> &OutVals,
1733 DebugLoc dl, SelectionDAG &DAG) const {
1735 // CCValAssign - represent the assignment of the return value to a location.
1736 SmallVector<CCValAssign, 16> RVLocs;
1738 // CCState - Info about the registers and stack slots.
1739 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1740 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1742 // Analyze outgoing return values.
1743 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1746 // If this is the first return lowered for this function, add
1747 // the regs to the liveout set for the function.
1748 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1749 for (unsigned i = 0; i != RVLocs.size(); ++i)
1750 if (RVLocs[i].isRegLoc())
1751 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1756 // Copy the result values into the output registers.
1757 for (unsigned i = 0, realRVLocIdx = 0;
1759 ++i, ++realRVLocIdx) {
1760 CCValAssign &VA = RVLocs[i];
1761 assert(VA.isRegLoc() && "Can only return in registers!");
1763 SDValue Arg = OutVals[realRVLocIdx];
1765 switch (VA.getLocInfo()) {
1766 default: llvm_unreachable("Unknown loc info!");
1767 case CCValAssign::Full: break;
1768 case CCValAssign::BCvt:
1769 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1773 if (VA.needsCustom()) {
1774 if (VA.getLocVT() == MVT::v2f64) {
1775 // Extract the first half and return it in two registers.
1776 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1777 DAG.getConstant(0, MVT::i32));
1778 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1779 DAG.getVTList(MVT::i32, MVT::i32), Half);
1781 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1782 Flag = Chain.getValue(1);
1783 VA = RVLocs[++i]; // skip ahead to next loc
1784 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1785 HalfGPRs.getValue(1), Flag);
1786 Flag = Chain.getValue(1);
1787 VA = RVLocs[++i]; // skip ahead to next loc
1789 // Extract the 2nd half and fall through to handle it as an f64 value.
1790 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1791 DAG.getConstant(1, MVT::i32));
1793 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1795 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1796 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1797 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1798 Flag = Chain.getValue(1);
1799 VA = RVLocs[++i]; // skip ahead to next loc
1800 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1803 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1805 // Guarantee that all emitted copies are
1806 // stuck together, avoiding something bad.
1807 Flag = Chain.getValue(1);
1812 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1814 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1819 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1820 if (N->getNumValues() != 1)
1822 if (!N->hasNUsesOfValue(1, 0))
1825 unsigned NumCopies = 0;
1827 SDNode *Use = *N->use_begin();
1828 if (Use->getOpcode() == ISD::CopyToReg) {
1829 Copies[NumCopies++] = Use;
1830 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1831 // f64 returned in a pair of GPRs.
1832 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1834 if (UI->getOpcode() != ISD::CopyToReg)
1836 Copies[UI.getUse().getResNo()] = *UI;
1839 } else if (Use->getOpcode() == ISD::BITCAST) {
1840 // f32 returned in a single GPR.
1841 if (!Use->hasNUsesOfValue(1, 0))
1843 Use = *Use->use_begin();
1844 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1846 Copies[NumCopies++] = Use;
1851 if (NumCopies != 1 && NumCopies != 2)
1854 bool HasRet = false;
1855 for (unsigned i = 0; i < NumCopies; ++i) {
1856 SDNode *Copy = Copies[i];
1857 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1859 if (UI->getOpcode() == ISD::CopyToReg) {
1861 if (Use == Copies[0] || Use == Copies[1])
1865 if (UI->getOpcode() != ARMISD::RET_FLAG)
1874 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1875 if (!EnableARMTailCalls)
1878 if (!CI->isTailCall())
1881 return !Subtarget->isThumb1Only();
1884 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1885 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1886 // one of the above mentioned nodes. It has to be wrapped because otherwise
1887 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1888 // be used to form addressing mode. These wrapped nodes will be selected
1890 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1891 EVT PtrVT = Op.getValueType();
1892 // FIXME there is no actual debug info here
1893 DebugLoc dl = Op.getDebugLoc();
1894 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1896 if (CP->isMachineConstantPoolEntry())
1897 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1898 CP->getAlignment());
1900 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1901 CP->getAlignment());
1902 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1905 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1906 return MachineJumpTableInfo::EK_Inline;
1909 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1910 SelectionDAG &DAG) const {
1911 MachineFunction &MF = DAG.getMachineFunction();
1912 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1913 unsigned ARMPCLabelIndex = 0;
1914 DebugLoc DL = Op.getDebugLoc();
1915 EVT PtrVT = getPointerTy();
1916 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1917 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1919 if (RelocM == Reloc::Static) {
1920 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1922 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1923 ARMPCLabelIndex = AFI->createPICLabelUId();
1924 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1925 ARMCP::CPBlockAddress,
1927 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1929 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1930 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1931 MachinePointerInfo::getConstantPool(),
1933 if (RelocM == Reloc::Static)
1935 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1936 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1939 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1941 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1942 SelectionDAG &DAG) const {
1943 DebugLoc dl = GA->getDebugLoc();
1944 EVT PtrVT = getPointerTy();
1945 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1946 MachineFunction &MF = DAG.getMachineFunction();
1947 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1948 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1949 ARMConstantPoolValue *CPV =
1950 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1951 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1952 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1953 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1954 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1955 MachinePointerInfo::getConstantPool(),
1957 SDValue Chain = Argument.getValue(1);
1959 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1960 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1962 // call __tls_get_addr.
1965 Entry.Node = Argument;
1966 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
1967 Args.push_back(Entry);
1968 // FIXME: is there useful debug info available here?
1969 std::pair<SDValue, SDValue> CallResult =
1970 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
1971 false, false, false, false,
1972 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1973 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1974 return CallResult.first;
1977 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1978 // "local exec" model.
1980 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1981 SelectionDAG &DAG) const {
1982 const GlobalValue *GV = GA->getGlobal();
1983 DebugLoc dl = GA->getDebugLoc();
1985 SDValue Chain = DAG.getEntryNode();
1986 EVT PtrVT = getPointerTy();
1987 // Get the Thread Pointer
1988 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1990 if (GV->isDeclaration()) {
1991 MachineFunction &MF = DAG.getMachineFunction();
1992 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1993 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1994 // Initial exec model.
1995 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1996 ARMConstantPoolValue *CPV =
1997 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1998 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
1999 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2000 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2001 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2002 MachinePointerInfo::getConstantPool(),
2004 Chain = Offset.getValue(1);
2006 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2007 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2009 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2010 MachinePointerInfo::getConstantPool(),
2014 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
2015 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2016 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2017 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2018 MachinePointerInfo::getConstantPool(),
2022 // The address of the thread local variable is the add of the thread
2023 // pointer with the offset of the variable.
2024 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2028 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2029 // TODO: implement the "local dynamic" model
2030 assert(Subtarget->isTargetELF() &&
2031 "TLS not implemented for non-ELF targets");
2032 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2033 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2034 // otherwise use the "Local Exec" TLS Model
2035 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2036 return LowerToTLSGeneralDynamicModel(GA, DAG);
2038 return LowerToTLSExecModels(GA, DAG);
2041 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2042 SelectionDAG &DAG) const {
2043 EVT PtrVT = getPointerTy();
2044 DebugLoc dl = Op.getDebugLoc();
2045 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2046 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2047 if (RelocM == Reloc::PIC_) {
2048 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2049 ARMConstantPoolValue *CPV =
2050 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2051 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2052 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2053 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2055 MachinePointerInfo::getConstantPool(),
2057 SDValue Chain = Result.getValue(1);
2058 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2059 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2061 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2062 MachinePointerInfo::getGOT(), false, false, 0);
2066 // If we have T2 ops, we can materialize the address directly via movt/movw
2067 // pair. This is always cheaper.
2068 if (Subtarget->useMovt()) {
2070 // FIXME: Once remat is capable of dealing with instructions with register
2071 // operands, expand this into two nodes.
2072 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2073 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2075 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2076 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2077 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2078 MachinePointerInfo::getConstantPool(),
2083 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2084 SelectionDAG &DAG) const {
2085 EVT PtrVT = getPointerTy();
2086 DebugLoc dl = Op.getDebugLoc();
2087 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2088 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2089 MachineFunction &MF = DAG.getMachineFunction();
2090 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2092 // FIXME: Enable this for static codegen when tool issues are fixed.
2093 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2095 // FIXME: Once remat is capable of dealing with instructions with register
2096 // operands, expand this into two nodes.
2097 if (RelocM == Reloc::Static)
2098 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2099 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2101 unsigned Wrapper = (RelocM == Reloc::PIC_)
2102 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2103 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2104 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2105 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2106 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2107 MachinePointerInfo::getGOT(), false, false, 0);
2111 unsigned ARMPCLabelIndex = 0;
2113 if (RelocM == Reloc::Static) {
2114 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2116 ARMPCLabelIndex = AFI->createPICLabelUId();
2117 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2118 ARMConstantPoolValue *CPV =
2119 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2120 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2122 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2124 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2125 MachinePointerInfo::getConstantPool(),
2127 SDValue Chain = Result.getValue(1);
2129 if (RelocM == Reloc::PIC_) {
2130 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2131 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2134 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2135 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2141 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2142 SelectionDAG &DAG) const {
2143 assert(Subtarget->isTargetELF() &&
2144 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2145 MachineFunction &MF = DAG.getMachineFunction();
2146 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2147 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2148 EVT PtrVT = getPointerTy();
2149 DebugLoc dl = Op.getDebugLoc();
2150 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2151 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2152 "_GLOBAL_OFFSET_TABLE_",
2153 ARMPCLabelIndex, PCAdj);
2154 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2155 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2156 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2157 MachinePointerInfo::getConstantPool(),
2159 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2160 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2164 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2166 DebugLoc dl = Op.getDebugLoc();
2167 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2168 Op.getOperand(0), Op.getOperand(1));
2172 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2173 DebugLoc dl = Op.getDebugLoc();
2174 SDValue Val = DAG.getConstant(0, MVT::i32);
2175 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2176 Op.getOperand(1), Val);
2180 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2181 DebugLoc dl = Op.getDebugLoc();
2182 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2183 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2187 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2188 const ARMSubtarget *Subtarget) const {
2189 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2190 DebugLoc dl = Op.getDebugLoc();
2192 default: return SDValue(); // Don't custom lower most intrinsics.
2193 case Intrinsic::arm_thread_pointer: {
2194 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2195 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2197 case Intrinsic::eh_sjlj_lsda: {
2198 MachineFunction &MF = DAG.getMachineFunction();
2199 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2200 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2201 EVT PtrVT = getPointerTy();
2202 DebugLoc dl = Op.getDebugLoc();
2203 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2205 unsigned PCAdj = (RelocM != Reloc::PIC_)
2206 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2207 ARMConstantPoolValue *CPV =
2208 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2209 ARMCP::CPLSDA, PCAdj);
2210 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2211 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2213 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2214 MachinePointerInfo::getConstantPool(),
2217 if (RelocM == Reloc::PIC_) {
2218 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2219 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2223 case Intrinsic::arm_neon_vmulls:
2224 case Intrinsic::arm_neon_vmullu: {
2225 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2226 ? ARMISD::VMULLs : ARMISD::VMULLu;
2227 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2228 Op.getOperand(1), Op.getOperand(2));
2233 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2234 const ARMSubtarget *Subtarget) {
2235 DebugLoc dl = Op.getDebugLoc();
2236 if (!Subtarget->hasDataBarrier()) {
2237 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2238 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2240 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2241 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2242 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2243 DAG.getConstant(0, MVT::i32));
2246 SDValue Op5 = Op.getOperand(5);
2247 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2248 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2249 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2250 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2252 ARM_MB::MemBOpt DMBOpt;
2253 if (isDeviceBarrier)
2254 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2256 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2257 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2258 DAG.getConstant(DMBOpt, MVT::i32));
2261 static SDValue getFence(SDValue InChain, DebugLoc dl, SelectionDAG &DAG,
2262 const ARMSubtarget *Subtarget) {
2263 if (!Subtarget->hasDataBarrier()) {
2264 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2265 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2267 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2268 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2269 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, InChain,
2270 DAG.getConstant(0, MVT::i32));
2273 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, InChain,
2274 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2277 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2278 const ARMSubtarget *Subtarget) {
2279 // FIXME: handle "fence singlethread" more efficiently.
2280 DebugLoc dl = Op.getDebugLoc();
2281 return getFence(Op.getOperand(0), dl, DAG, Subtarget);
2284 static SDValue LowerAtomicMemOp(SDValue Op, SelectionDAG &DAG,
2285 const ARMSubtarget *Subtarget) {
2286 DebugLoc dl = Op.getDebugLoc();
2287 int Order = cast<AtomicSDNode>(Op)->getOrdering();
2288 if (Order <= Monotonic)
2291 SDValue InChain = Op.getOperand(0);
2293 // Fence, if necessary
2294 if (Order == Release || Order >= AcquireRelease)
2295 InChain = getFence(InChain, dl, DAG, Subtarget);
2297 // Rather than mess with target-specific nodes, use the target-indepedent
2298 // node, and assume the DAGCombiner will not touch it post-legalize.
2300 if (Op.getOpcode() == ISD::ATOMIC_CMP_SWAP)
2301 OutVal = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
2302 cast<AtomicSDNode>(Op)->getMemoryVT(),
2303 InChain, Op.getOperand(1), Op.getOperand(2),
2305 cast<AtomicSDNode>(Op)->getMemOperand(),
2307 cast<AtomicSDNode>(Op)->getSynchScope());
2309 OutVal = DAG.getAtomic(Op.getOpcode(), dl,
2310 cast<AtomicSDNode>(Op)->getMemoryVT(),
2311 InChain, Op.getOperand(1), Op.getOperand(2),
2312 cast<AtomicSDNode>(Op)->getMemOperand(),
2314 cast<AtomicSDNode>(Op)->getSynchScope());
2316 SDValue OutChain = OutVal.getValue(1);
2318 // Fence, if necessary
2319 if (Order == Acquire || Order >= AcquireRelease)
2320 OutChain = getFence(OutChain, dl, DAG, Subtarget);
2322 SDValue Ops[2] = { OutVal, OutChain };
2323 return DAG.getMergeValues(Ops, 2, dl);
2327 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2328 const ARMSubtarget *Subtarget) {
2329 // ARM pre v5TE and Thumb1 does not have preload instructions.
2330 if (!(Subtarget->isThumb2() ||
2331 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2332 // Just preserve the chain.
2333 return Op.getOperand(0);
2335 DebugLoc dl = Op.getDebugLoc();
2336 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2338 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2339 // ARMv7 with MP extension has PLDW.
2340 return Op.getOperand(0);
2342 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2343 if (Subtarget->isThumb()) {
2345 isRead = ~isRead & 1;
2346 isData = ~isData & 1;
2349 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2350 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2351 DAG.getConstant(isData, MVT::i32));
2354 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2355 MachineFunction &MF = DAG.getMachineFunction();
2356 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2358 // vastart just stores the address of the VarArgsFrameIndex slot into the
2359 // memory location argument.
2360 DebugLoc dl = Op.getDebugLoc();
2361 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2362 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2363 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2364 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2365 MachinePointerInfo(SV), false, false, 0);
2369 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2370 SDValue &Root, SelectionDAG &DAG,
2371 DebugLoc dl) const {
2372 MachineFunction &MF = DAG.getMachineFunction();
2373 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2375 TargetRegisterClass *RC;
2376 if (AFI->isThumb1OnlyFunction())
2377 RC = ARM::tGPRRegisterClass;
2379 RC = ARM::GPRRegisterClass;
2381 // Transform the arguments stored in physical registers into virtual ones.
2382 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2383 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2386 if (NextVA.isMemLoc()) {
2387 MachineFrameInfo *MFI = MF.getFrameInfo();
2388 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2390 // Create load node to retrieve arguments from the stack.
2391 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2392 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2393 MachinePointerInfo::getFixedStack(FI),
2396 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2397 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2400 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2404 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2405 unsigned &VARegSize, unsigned &VARegSaveSize)
2408 if (CCInfo.isFirstByValRegValid())
2409 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2411 unsigned int firstUnalloced;
2412 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2413 sizeof(GPRArgRegs) /
2414 sizeof(GPRArgRegs[0]));
2415 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2418 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2419 VARegSize = NumGPRs * 4;
2420 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2423 // The remaining GPRs hold either the beginning of variable-argument
2424 // data, or the beginning of an aggregate passed by value (usuall
2425 // byval). Either way, we allocate stack slots adjacent to the data
2426 // provided by our caller, and store the unallocated registers there.
2427 // If this is a variadic function, the va_list pointer will begin with
2428 // these values; otherwise, this reassembles a (byval) structure that
2429 // was split between registers and memory.
2431 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2432 DebugLoc dl, SDValue &Chain,
2433 unsigned ArgOffset) const {
2434 MachineFunction &MF = DAG.getMachineFunction();
2435 MachineFrameInfo *MFI = MF.getFrameInfo();
2436 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2437 unsigned firstRegToSaveIndex;
2438 if (CCInfo.isFirstByValRegValid())
2439 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2441 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2442 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2445 unsigned VARegSize, VARegSaveSize;
2446 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2447 if (VARegSaveSize) {
2448 // If this function is vararg, store any remaining integer argument regs
2449 // to their spots on the stack so that they may be loaded by deferencing
2450 // the result of va_next.
2451 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2452 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2453 ArgOffset + VARegSaveSize
2456 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2459 SmallVector<SDValue, 4> MemOps;
2460 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2461 TargetRegisterClass *RC;
2462 if (AFI->isThumb1OnlyFunction())
2463 RC = ARM::tGPRRegisterClass;
2465 RC = ARM::GPRRegisterClass;
2467 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2468 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2470 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2471 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2473 MemOps.push_back(Store);
2474 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2475 DAG.getConstant(4, getPointerTy()));
2477 if (!MemOps.empty())
2478 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2479 &MemOps[0], MemOps.size());
2481 // This will point to the next argument passed via stack.
2482 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2486 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2487 CallingConv::ID CallConv, bool isVarArg,
2488 const SmallVectorImpl<ISD::InputArg>
2490 DebugLoc dl, SelectionDAG &DAG,
2491 SmallVectorImpl<SDValue> &InVals)
2493 MachineFunction &MF = DAG.getMachineFunction();
2494 MachineFrameInfo *MFI = MF.getFrameInfo();
2496 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2498 // Assign locations to all of the incoming arguments.
2499 SmallVector<CCValAssign, 16> ArgLocs;
2500 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2501 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2502 CCInfo.AnalyzeFormalArguments(Ins,
2503 CCAssignFnForNode(CallConv, /* Return*/ false,
2506 SmallVector<SDValue, 16> ArgValues;
2507 int lastInsIndex = -1;
2510 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2511 CCValAssign &VA = ArgLocs[i];
2513 // Arguments stored in registers.
2514 if (VA.isRegLoc()) {
2515 EVT RegVT = VA.getLocVT();
2517 if (VA.needsCustom()) {
2518 // f64 and vector types are split up into multiple registers or
2519 // combinations of registers and stack slots.
2520 if (VA.getLocVT() == MVT::v2f64) {
2521 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2523 VA = ArgLocs[++i]; // skip ahead to next loc
2525 if (VA.isMemLoc()) {
2526 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2527 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2528 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2529 MachinePointerInfo::getFixedStack(FI),
2532 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2535 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2536 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2537 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2538 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2539 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2541 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2544 TargetRegisterClass *RC;
2546 if (RegVT == MVT::f32)
2547 RC = ARM::SPRRegisterClass;
2548 else if (RegVT == MVT::f64)
2549 RC = ARM::DPRRegisterClass;
2550 else if (RegVT == MVT::v2f64)
2551 RC = ARM::QPRRegisterClass;
2552 else if (RegVT == MVT::i32)
2553 RC = (AFI->isThumb1OnlyFunction() ?
2554 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2556 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2558 // Transform the arguments in physical registers into virtual ones.
2559 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2560 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2563 // If this is an 8 or 16-bit value, it is really passed promoted
2564 // to 32 bits. Insert an assert[sz]ext to capture this, then
2565 // truncate to the right size.
2566 switch (VA.getLocInfo()) {
2567 default: llvm_unreachable("Unknown loc info!");
2568 case CCValAssign::Full: break;
2569 case CCValAssign::BCvt:
2570 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2572 case CCValAssign::SExt:
2573 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2574 DAG.getValueType(VA.getValVT()));
2575 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2577 case CCValAssign::ZExt:
2578 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2579 DAG.getValueType(VA.getValVT()));
2580 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2584 InVals.push_back(ArgValue);
2586 } else { // VA.isRegLoc()
2589 assert(VA.isMemLoc());
2590 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2592 int index = ArgLocs[i].getValNo();
2594 // Some Ins[] entries become multiple ArgLoc[] entries.
2595 // Process them only once.
2596 if (index != lastInsIndex)
2598 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2599 // FIXME: For now, all byval parameter objects are marked mutable.
2600 // This can be changed with more analysis.
2601 // In case of tail call optimization mark all arguments mutable.
2602 // Since they could be overwritten by lowering of arguments in case of
2604 if (Flags.isByVal()) {
2605 unsigned VARegSize, VARegSaveSize;
2606 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2607 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2608 unsigned Bytes = Flags.getByValSize() - VARegSize;
2609 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2610 int FI = MFI->CreateFixedObject(Bytes,
2611 VA.getLocMemOffset(), false);
2612 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2614 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2615 VA.getLocMemOffset(), true);
2617 // Create load nodes to retrieve arguments from the stack.
2618 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2619 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2620 MachinePointerInfo::getFixedStack(FI),
2623 lastInsIndex = index;
2630 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2635 /// isFloatingPointZero - Return true if this is +0.0.
2636 static bool isFloatingPointZero(SDValue Op) {
2637 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2638 return CFP->getValueAPF().isPosZero();
2639 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2640 // Maybe this has already been legalized into the constant pool?
2641 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2642 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2643 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2644 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2645 return CFP->getValueAPF().isPosZero();
2651 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2652 /// the given operands.
2654 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2655 SDValue &ARMcc, SelectionDAG &DAG,
2656 DebugLoc dl) const {
2657 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2658 unsigned C = RHSC->getZExtValue();
2659 if (!isLegalICmpImmediate(C)) {
2660 // Constant does not fit, try adjusting it by one?
2665 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2666 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2667 RHS = DAG.getConstant(C-1, MVT::i32);
2672 if (C != 0 && isLegalICmpImmediate(C-1)) {
2673 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2674 RHS = DAG.getConstant(C-1, MVT::i32);
2679 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2680 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2681 RHS = DAG.getConstant(C+1, MVT::i32);
2686 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2687 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2688 RHS = DAG.getConstant(C+1, MVT::i32);
2695 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2696 ARMISD::NodeType CompareType;
2699 CompareType = ARMISD::CMP;
2704 CompareType = ARMISD::CMPZ;
2707 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2708 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2711 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2713 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2714 DebugLoc dl) const {
2716 if (!isFloatingPointZero(RHS))
2717 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2719 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2720 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2723 /// duplicateCmp - Glue values can have only one use, so this function
2724 /// duplicates a comparison node.
2726 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2727 unsigned Opc = Cmp.getOpcode();
2728 DebugLoc DL = Cmp.getDebugLoc();
2729 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2730 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2732 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2733 Cmp = Cmp.getOperand(0);
2734 Opc = Cmp.getOpcode();
2735 if (Opc == ARMISD::CMPFP)
2736 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2738 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2739 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2741 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2744 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2745 SDValue Cond = Op.getOperand(0);
2746 SDValue SelectTrue = Op.getOperand(1);
2747 SDValue SelectFalse = Op.getOperand(2);
2748 DebugLoc dl = Op.getDebugLoc();
2752 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2753 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2755 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2756 const ConstantSDNode *CMOVTrue =
2757 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2758 const ConstantSDNode *CMOVFalse =
2759 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2761 if (CMOVTrue && CMOVFalse) {
2762 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2763 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2767 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2769 False = SelectFalse;
2770 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2775 if (True.getNode() && False.getNode()) {
2776 EVT VT = Op.getValueType();
2777 SDValue ARMcc = Cond.getOperand(2);
2778 SDValue CCR = Cond.getOperand(3);
2779 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2780 assert(True.getValueType() == VT);
2781 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2786 return DAG.getSelectCC(dl, Cond,
2787 DAG.getConstant(0, Cond.getValueType()),
2788 SelectTrue, SelectFalse, ISD::SETNE);
2791 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2792 EVT VT = Op.getValueType();
2793 SDValue LHS = Op.getOperand(0);
2794 SDValue RHS = Op.getOperand(1);
2795 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2796 SDValue TrueVal = Op.getOperand(2);
2797 SDValue FalseVal = Op.getOperand(3);
2798 DebugLoc dl = Op.getDebugLoc();
2800 if (LHS.getValueType() == MVT::i32) {
2802 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2803 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2804 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp);
2807 ARMCC::CondCodes CondCode, CondCode2;
2808 FPCCToARMCC(CC, CondCode, CondCode2);
2810 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2811 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2812 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2813 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2815 if (CondCode2 != ARMCC::AL) {
2816 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2817 // FIXME: Needs another CMP because flag can have but one use.
2818 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2819 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2820 Result, TrueVal, ARMcc2, CCR, Cmp2);
2825 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2826 /// to morph to an integer compare sequence.
2827 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2828 const ARMSubtarget *Subtarget) {
2829 SDNode *N = Op.getNode();
2830 if (!N->hasOneUse())
2831 // Otherwise it requires moving the value from fp to integer registers.
2833 if (!N->getNumValues())
2835 EVT VT = Op.getValueType();
2836 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2837 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2838 // vmrs are very slow, e.g. cortex-a8.
2841 if (isFloatingPointZero(Op)) {
2845 return ISD::isNormalLoad(N);
2848 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2849 if (isFloatingPointZero(Op))
2850 return DAG.getConstant(0, MVT::i32);
2852 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2853 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2854 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2855 Ld->isVolatile(), Ld->isNonTemporal(),
2856 Ld->getAlignment());
2858 llvm_unreachable("Unknown VFP cmp argument!");
2861 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2862 SDValue &RetVal1, SDValue &RetVal2) {
2863 if (isFloatingPointZero(Op)) {
2864 RetVal1 = DAG.getConstant(0, MVT::i32);
2865 RetVal2 = DAG.getConstant(0, MVT::i32);
2869 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2870 SDValue Ptr = Ld->getBasePtr();
2871 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2872 Ld->getChain(), Ptr,
2873 Ld->getPointerInfo(),
2874 Ld->isVolatile(), Ld->isNonTemporal(),
2875 Ld->getAlignment());
2877 EVT PtrType = Ptr.getValueType();
2878 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2879 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2880 PtrType, Ptr, DAG.getConstant(4, PtrType));
2881 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2882 Ld->getChain(), NewPtr,
2883 Ld->getPointerInfo().getWithOffset(4),
2884 Ld->isVolatile(), Ld->isNonTemporal(),
2889 llvm_unreachable("Unknown VFP cmp argument!");
2892 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2893 /// f32 and even f64 comparisons to integer ones.
2895 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2896 SDValue Chain = Op.getOperand(0);
2897 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2898 SDValue LHS = Op.getOperand(2);
2899 SDValue RHS = Op.getOperand(3);
2900 SDValue Dest = Op.getOperand(4);
2901 DebugLoc dl = Op.getDebugLoc();
2903 bool SeenZero = false;
2904 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2905 canChangeToInt(RHS, SeenZero, Subtarget) &&
2906 // If one of the operand is zero, it's safe to ignore the NaN case since
2907 // we only care about equality comparisons.
2908 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2909 // If unsafe fp math optimization is enabled and there are no other uses of
2910 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2911 // to an integer comparison.
2912 if (CC == ISD::SETOEQ)
2914 else if (CC == ISD::SETUNE)
2918 if (LHS.getValueType() == MVT::f32) {
2919 LHS = bitcastf32Toi32(LHS, DAG);
2920 RHS = bitcastf32Toi32(RHS, DAG);
2921 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2922 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2923 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2924 Chain, Dest, ARMcc, CCR, Cmp);
2929 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2930 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2931 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2932 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2933 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2934 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2935 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2941 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2942 SDValue Chain = Op.getOperand(0);
2943 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2944 SDValue LHS = Op.getOperand(2);
2945 SDValue RHS = Op.getOperand(3);
2946 SDValue Dest = Op.getOperand(4);
2947 DebugLoc dl = Op.getDebugLoc();
2949 if (LHS.getValueType() == MVT::i32) {
2951 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2952 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2953 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2954 Chain, Dest, ARMcc, CCR, Cmp);
2957 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2960 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2961 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2962 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2963 if (Result.getNode())
2967 ARMCC::CondCodes CondCode, CondCode2;
2968 FPCCToARMCC(CC, CondCode, CondCode2);
2970 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2971 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2972 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2973 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2974 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2975 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2976 if (CondCode2 != ARMCC::AL) {
2977 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2978 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2979 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2984 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2985 SDValue Chain = Op.getOperand(0);
2986 SDValue Table = Op.getOperand(1);
2987 SDValue Index = Op.getOperand(2);
2988 DebugLoc dl = Op.getDebugLoc();
2990 EVT PTy = getPointerTy();
2991 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2992 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2993 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2994 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2995 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2996 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2997 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2998 if (Subtarget->isThumb2()) {
2999 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3000 // which does another jump to the destination. This also makes it easier
3001 // to translate it to TBB / TBH later.
3002 // FIXME: This might not work if the function is extremely large.
3003 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3004 Addr, Op.getOperand(2), JTI, UId);
3006 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3007 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3008 MachinePointerInfo::getJumpTable(),
3010 Chain = Addr.getValue(1);
3011 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3012 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3014 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3015 MachinePointerInfo::getJumpTable(), false, false, 0);
3016 Chain = Addr.getValue(1);
3017 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3021 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3022 DebugLoc dl = Op.getDebugLoc();
3025 switch (Op.getOpcode()) {
3027 assert(0 && "Invalid opcode!");
3028 case ISD::FP_TO_SINT:
3029 Opc = ARMISD::FTOSI;
3031 case ISD::FP_TO_UINT:
3032 Opc = ARMISD::FTOUI;
3035 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3036 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3039 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3040 EVT VT = Op.getValueType();
3041 DebugLoc dl = Op.getDebugLoc();
3043 EVT OperandVT = Op.getOperand(0).getValueType();
3044 assert(OperandVT == MVT::v4i16 && "Invalid type for custom lowering!");
3045 if (VT != MVT::v4f32)
3046 return DAG.UnrollVectorOp(Op.getNode());
3050 switch (Op.getOpcode()) {
3052 assert(0 && "Invalid opcode!");
3053 case ISD::SINT_TO_FP:
3054 CastOpc = ISD::SIGN_EXTEND;
3055 Opc = ISD::SINT_TO_FP;
3057 case ISD::UINT_TO_FP:
3058 CastOpc = ISD::ZERO_EXTEND;
3059 Opc = ISD::UINT_TO_FP;
3063 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3064 return DAG.getNode(Opc, dl, VT, Op);
3067 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3068 EVT VT = Op.getValueType();
3070 return LowerVectorINT_TO_FP(Op, DAG);
3072 DebugLoc dl = Op.getDebugLoc();
3075 switch (Op.getOpcode()) {
3077 assert(0 && "Invalid opcode!");
3078 case ISD::SINT_TO_FP:
3079 Opc = ARMISD::SITOF;
3081 case ISD::UINT_TO_FP:
3082 Opc = ARMISD::UITOF;
3086 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3087 return DAG.getNode(Opc, dl, VT, Op);
3090 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3091 // Implement fcopysign with a fabs and a conditional fneg.
3092 SDValue Tmp0 = Op.getOperand(0);
3093 SDValue Tmp1 = Op.getOperand(1);
3094 DebugLoc dl = Op.getDebugLoc();
3095 EVT VT = Op.getValueType();
3096 EVT SrcVT = Tmp1.getValueType();
3097 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3098 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3099 bool UseNEON = !InGPR && Subtarget->hasNEON();
3102 // Use VBSL to copy the sign bit.
3103 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3104 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3105 DAG.getTargetConstant(EncodedVal, MVT::i32));
3106 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3108 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3109 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3110 DAG.getConstant(32, MVT::i32));
3111 else /*if (VT == MVT::f32)*/
3112 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3113 if (SrcVT == MVT::f32) {
3114 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3116 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3117 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3118 DAG.getConstant(32, MVT::i32));
3119 } else if (VT == MVT::f32)
3120 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3121 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3122 DAG.getConstant(32, MVT::i32));
3123 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3124 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3126 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3128 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3129 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3130 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3132 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3133 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3134 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3135 if (VT == MVT::f32) {
3136 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3137 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3138 DAG.getConstant(0, MVT::i32));
3140 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3146 // Bitcast operand 1 to i32.
3147 if (SrcVT == MVT::f64)
3148 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3149 &Tmp1, 1).getValue(1);
3150 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3152 // Or in the signbit with integer operations.
3153 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3154 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3155 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3156 if (VT == MVT::f32) {
3157 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3158 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3159 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3160 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3163 // f64: Or the high part with signbit and then combine two parts.
3164 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3166 SDValue Lo = Tmp0.getValue(0);
3167 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3168 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3169 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3172 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3173 MachineFunction &MF = DAG.getMachineFunction();
3174 MachineFrameInfo *MFI = MF.getFrameInfo();
3175 MFI->setReturnAddressIsTaken(true);
3177 EVT VT = Op.getValueType();
3178 DebugLoc dl = Op.getDebugLoc();
3179 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3181 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3182 SDValue Offset = DAG.getConstant(4, MVT::i32);
3183 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3184 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3185 MachinePointerInfo(), false, false, 0);
3188 // Return LR, which contains the return address. Mark it an implicit live-in.
3189 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3190 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3193 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3194 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3195 MFI->setFrameAddressIsTaken(true);
3197 EVT VT = Op.getValueType();
3198 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3199 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3200 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3201 ? ARM::R7 : ARM::R11;
3202 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3204 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3205 MachinePointerInfo(),
3210 /// ExpandBITCAST - If the target supports VFP, this function is called to
3211 /// expand a bit convert where either the source or destination type is i64 to
3212 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3213 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3214 /// vectors), since the legalizer won't know what to do with that.
3215 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3216 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3217 DebugLoc dl = N->getDebugLoc();
3218 SDValue Op = N->getOperand(0);
3220 // This function is only supposed to be called for i64 types, either as the
3221 // source or destination of the bit convert.
3222 EVT SrcVT = Op.getValueType();
3223 EVT DstVT = N->getValueType(0);
3224 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3225 "ExpandBITCAST called for non-i64 type");
3227 // Turn i64->f64 into VMOVDRR.
3228 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3229 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3230 DAG.getConstant(0, MVT::i32));
3231 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3232 DAG.getConstant(1, MVT::i32));
3233 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3234 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3237 // Turn f64->i64 into VMOVRRD.
3238 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3239 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3240 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3241 // Merge the pieces into a single i64 value.
3242 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3248 /// getZeroVector - Returns a vector of specified type with all zero elements.
3249 /// Zero vectors are used to represent vector negation and in those cases
3250 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3251 /// not support i64 elements, so sometimes the zero vectors will need to be
3252 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3254 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3255 assert(VT.isVector() && "Expected a vector type");
3256 // The canonical modified immediate encoding of a zero vector is....0!
3257 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3258 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3259 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3260 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3263 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3264 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3265 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3266 SelectionDAG &DAG) const {
3267 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3268 EVT VT = Op.getValueType();
3269 unsigned VTBits = VT.getSizeInBits();
3270 DebugLoc dl = Op.getDebugLoc();
3271 SDValue ShOpLo = Op.getOperand(0);
3272 SDValue ShOpHi = Op.getOperand(1);
3273 SDValue ShAmt = Op.getOperand(2);
3275 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3277 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3279 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3280 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3281 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3282 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3283 DAG.getConstant(VTBits, MVT::i32));
3284 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3285 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3286 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3288 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3289 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3291 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3292 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3295 SDValue Ops[2] = { Lo, Hi };
3296 return DAG.getMergeValues(Ops, 2, dl);
3299 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3300 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3301 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3302 SelectionDAG &DAG) const {
3303 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3304 EVT VT = Op.getValueType();
3305 unsigned VTBits = VT.getSizeInBits();
3306 DebugLoc dl = Op.getDebugLoc();
3307 SDValue ShOpLo = Op.getOperand(0);
3308 SDValue ShOpHi = Op.getOperand(1);
3309 SDValue ShAmt = Op.getOperand(2);
3312 assert(Op.getOpcode() == ISD::SHL_PARTS);
3313 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3314 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3315 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3316 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3317 DAG.getConstant(VTBits, MVT::i32));
3318 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3319 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3321 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3322 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3323 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3325 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3326 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3329 SDValue Ops[2] = { Lo, Hi };
3330 return DAG.getMergeValues(Ops, 2, dl);
3333 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3334 SelectionDAG &DAG) const {
3335 // The rounding mode is in bits 23:22 of the FPSCR.
3336 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3337 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3338 // so that the shift + and get folded into a bitfield extract.
3339 DebugLoc dl = Op.getDebugLoc();
3340 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3341 DAG.getConstant(Intrinsic::arm_get_fpscr,
3343 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3344 DAG.getConstant(1U << 22, MVT::i32));
3345 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3346 DAG.getConstant(22, MVT::i32));
3347 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3348 DAG.getConstant(3, MVT::i32));
3351 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3352 const ARMSubtarget *ST) {
3353 EVT VT = N->getValueType(0);
3354 DebugLoc dl = N->getDebugLoc();
3356 if (!ST->hasV6T2Ops())
3359 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3360 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3363 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3364 const ARMSubtarget *ST) {
3365 EVT VT = N->getValueType(0);
3366 DebugLoc dl = N->getDebugLoc();
3371 // Lower vector shifts on NEON to use VSHL.
3372 assert(ST->hasNEON() && "unexpected vector shift");
3374 // Left shifts translate directly to the vshiftu intrinsic.
3375 if (N->getOpcode() == ISD::SHL)
3376 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3377 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3378 N->getOperand(0), N->getOperand(1));
3380 assert((N->getOpcode() == ISD::SRA ||
3381 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3383 // NEON uses the same intrinsics for both left and right shifts. For
3384 // right shifts, the shift amounts are negative, so negate the vector of
3386 EVT ShiftVT = N->getOperand(1).getValueType();
3387 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3388 getZeroVector(ShiftVT, DAG, dl),
3390 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3391 Intrinsic::arm_neon_vshifts :
3392 Intrinsic::arm_neon_vshiftu);
3393 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3394 DAG.getConstant(vshiftInt, MVT::i32),
3395 N->getOperand(0), NegatedCount);
3398 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3399 const ARMSubtarget *ST) {
3400 EVT VT = N->getValueType(0);
3401 DebugLoc dl = N->getDebugLoc();
3403 // We can get here for a node like i32 = ISD::SHL i32, i64
3407 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3408 "Unknown shift to lower!");
3410 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3411 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3412 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3415 // If we are in thumb mode, we don't have RRX.
3416 if (ST->isThumb1Only()) return SDValue();
3418 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3419 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3420 DAG.getConstant(0, MVT::i32));
3421 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3422 DAG.getConstant(1, MVT::i32));
3424 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3425 // captures the result into a carry flag.
3426 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3427 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3429 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3430 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3432 // Merge the pieces into a single i64 value.
3433 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3436 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3437 SDValue TmpOp0, TmpOp1;
3438 bool Invert = false;
3442 SDValue Op0 = Op.getOperand(0);
3443 SDValue Op1 = Op.getOperand(1);
3444 SDValue CC = Op.getOperand(2);
3445 EVT VT = Op.getValueType();
3446 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3447 DebugLoc dl = Op.getDebugLoc();
3449 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3450 switch (SetCCOpcode) {
3451 default: llvm_unreachable("Illegal FP comparison"); break;
3453 case ISD::SETNE: Invert = true; // Fallthrough
3455 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3457 case ISD::SETLT: Swap = true; // Fallthrough
3459 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3461 case ISD::SETLE: Swap = true; // Fallthrough
3463 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3464 case ISD::SETUGE: Swap = true; // Fallthrough
3465 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3466 case ISD::SETUGT: Swap = true; // Fallthrough
3467 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3468 case ISD::SETUEQ: Invert = true; // Fallthrough
3470 // Expand this to (OLT | OGT).
3474 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3475 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3477 case ISD::SETUO: Invert = true; // Fallthrough
3479 // Expand this to (OLT | OGE).
3483 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3484 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3488 // Integer comparisons.
3489 switch (SetCCOpcode) {
3490 default: llvm_unreachable("Illegal integer comparison"); break;
3491 case ISD::SETNE: Invert = true;
3492 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3493 case ISD::SETLT: Swap = true;
3494 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3495 case ISD::SETLE: Swap = true;
3496 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3497 case ISD::SETULT: Swap = true;
3498 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3499 case ISD::SETULE: Swap = true;
3500 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3503 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3504 if (Opc == ARMISD::VCEQ) {
3507 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3509 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3512 // Ignore bitconvert.
3513 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3514 AndOp = AndOp.getOperand(0);
3516 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3518 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3519 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3526 std::swap(Op0, Op1);
3528 // If one of the operands is a constant vector zero, attempt to fold the
3529 // comparison to a specialized compare-against-zero form.
3531 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3533 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3534 if (Opc == ARMISD::VCGE)
3535 Opc = ARMISD::VCLEZ;
3536 else if (Opc == ARMISD::VCGT)
3537 Opc = ARMISD::VCLTZ;
3542 if (SingleOp.getNode()) {
3545 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3547 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3549 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3551 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3553 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3555 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3558 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3562 Result = DAG.getNOT(dl, Result, VT);
3567 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3568 /// valid vector constant for a NEON instruction with a "modified immediate"
3569 /// operand (e.g., VMOV). If so, return the encoded value.
3570 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3571 unsigned SplatBitSize, SelectionDAG &DAG,
3572 EVT &VT, bool is128Bits, NEONModImmType type) {
3573 unsigned OpCmode, Imm;
3575 // SplatBitSize is set to the smallest size that splats the vector, so a
3576 // zero vector will always have SplatBitSize == 8. However, NEON modified
3577 // immediate instructions others than VMOV do not support the 8-bit encoding
3578 // of a zero vector, and the default encoding of zero is supposed to be the
3583 switch (SplatBitSize) {
3585 if (type != VMOVModImm)
3587 // Any 1-byte value is OK. Op=0, Cmode=1110.
3588 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3591 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3595 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3596 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3597 if ((SplatBits & ~0xff) == 0) {
3598 // Value = 0x00nn: Op=x, Cmode=100x.
3603 if ((SplatBits & ~0xff00) == 0) {
3604 // Value = 0xnn00: Op=x, Cmode=101x.
3606 Imm = SplatBits >> 8;
3612 // NEON's 32-bit VMOV supports splat values where:
3613 // * only one byte is nonzero, or
3614 // * the least significant byte is 0xff and the second byte is nonzero, or
3615 // * the least significant 2 bytes are 0xff and the third is nonzero.
3616 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3617 if ((SplatBits & ~0xff) == 0) {
3618 // Value = 0x000000nn: Op=x, Cmode=000x.
3623 if ((SplatBits & ~0xff00) == 0) {
3624 // Value = 0x0000nn00: Op=x, Cmode=001x.
3626 Imm = SplatBits >> 8;
3629 if ((SplatBits & ~0xff0000) == 0) {
3630 // Value = 0x00nn0000: Op=x, Cmode=010x.
3632 Imm = SplatBits >> 16;
3635 if ((SplatBits & ~0xff000000) == 0) {
3636 // Value = 0xnn000000: Op=x, Cmode=011x.
3638 Imm = SplatBits >> 24;
3642 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3643 if (type == OtherModImm) return SDValue();
3645 if ((SplatBits & ~0xffff) == 0 &&
3646 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3647 // Value = 0x0000nnff: Op=x, Cmode=1100.
3649 Imm = SplatBits >> 8;
3654 if ((SplatBits & ~0xffffff) == 0 &&
3655 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3656 // Value = 0x00nnffff: Op=x, Cmode=1101.
3658 Imm = SplatBits >> 16;
3659 SplatBits |= 0xffff;
3663 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3664 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3665 // VMOV.I32. A (very) minor optimization would be to replicate the value
3666 // and fall through here to test for a valid 64-bit splat. But, then the
3667 // caller would also need to check and handle the change in size.
3671 if (type != VMOVModImm)
3673 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3674 uint64_t BitMask = 0xff;
3676 unsigned ImmMask = 1;
3678 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3679 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3682 } else if ((SplatBits & BitMask) != 0) {
3688 // Op=1, Cmode=1110.
3691 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3696 llvm_unreachable("unexpected size for isNEONModifiedImm");
3700 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3701 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3704 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3705 bool &ReverseVEXT, unsigned &Imm) {
3706 unsigned NumElts = VT.getVectorNumElements();
3707 ReverseVEXT = false;
3709 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3715 // If this is a VEXT shuffle, the immediate value is the index of the first
3716 // element. The other shuffle indices must be the successive elements after
3718 unsigned ExpectedElt = Imm;
3719 for (unsigned i = 1; i < NumElts; ++i) {
3720 // Increment the expected index. If it wraps around, it may still be
3721 // a VEXT but the source vectors must be swapped.
3723 if (ExpectedElt == NumElts * 2) {
3728 if (M[i] < 0) continue; // ignore UNDEF indices
3729 if (ExpectedElt != static_cast<unsigned>(M[i]))
3733 // Adjust the index value if the source operands will be swapped.
3740 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3741 /// instruction with the specified blocksize. (The order of the elements
3742 /// within each block of the vector is reversed.)
3743 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3744 unsigned BlockSize) {
3745 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3746 "Only possible block sizes for VREV are: 16, 32, 64");
3748 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3752 unsigned NumElts = VT.getVectorNumElements();
3753 unsigned BlockElts = M[0] + 1;
3754 // If the first shuffle index is UNDEF, be optimistic.
3756 BlockElts = BlockSize / EltSz;
3758 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3761 for (unsigned i = 0; i < NumElts; ++i) {
3762 if (M[i] < 0) continue; // ignore UNDEF indices
3763 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3770 static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3771 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3772 // range, then 0 is placed into the resulting vector. So pretty much any mask
3773 // of 8 elements can work here.
3774 return VT == MVT::v8i8 && M.size() == 8;
3777 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3778 unsigned &WhichResult) {
3779 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3783 unsigned NumElts = VT.getVectorNumElements();
3784 WhichResult = (M[0] == 0 ? 0 : 1);
3785 for (unsigned i = 0; i < NumElts; i += 2) {
3786 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3787 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3793 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3794 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3795 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3796 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3797 unsigned &WhichResult) {
3798 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3802 unsigned NumElts = VT.getVectorNumElements();
3803 WhichResult = (M[0] == 0 ? 0 : 1);
3804 for (unsigned i = 0; i < NumElts; i += 2) {
3805 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3806 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3812 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3813 unsigned &WhichResult) {
3814 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3818 unsigned NumElts = VT.getVectorNumElements();
3819 WhichResult = (M[0] == 0 ? 0 : 1);
3820 for (unsigned i = 0; i != NumElts; ++i) {
3821 if (M[i] < 0) continue; // ignore UNDEF indices
3822 if ((unsigned) M[i] != 2 * i + WhichResult)
3826 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3827 if (VT.is64BitVector() && EltSz == 32)
3833 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3834 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3835 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3836 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3837 unsigned &WhichResult) {
3838 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3842 unsigned Half = VT.getVectorNumElements() / 2;
3843 WhichResult = (M[0] == 0 ? 0 : 1);
3844 for (unsigned j = 0; j != 2; ++j) {
3845 unsigned Idx = WhichResult;
3846 for (unsigned i = 0; i != Half; ++i) {
3847 int MIdx = M[i + j * Half];
3848 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3854 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3855 if (VT.is64BitVector() && EltSz == 32)
3861 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3862 unsigned &WhichResult) {
3863 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3867 unsigned NumElts = VT.getVectorNumElements();
3868 WhichResult = (M[0] == 0 ? 0 : 1);
3869 unsigned Idx = WhichResult * NumElts / 2;
3870 for (unsigned i = 0; i != NumElts; i += 2) {
3871 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3872 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3877 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3878 if (VT.is64BitVector() && EltSz == 32)
3884 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3885 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3886 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3887 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3888 unsigned &WhichResult) {
3889 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3893 unsigned NumElts = VT.getVectorNumElements();
3894 WhichResult = (M[0] == 0 ? 0 : 1);
3895 unsigned Idx = WhichResult * NumElts / 2;
3896 for (unsigned i = 0; i != NumElts; i += 2) {
3897 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3898 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3903 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3904 if (VT.is64BitVector() && EltSz == 32)
3910 // If N is an integer constant that can be moved into a register in one
3911 // instruction, return an SDValue of such a constant (will become a MOV
3912 // instruction). Otherwise return null.
3913 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3914 const ARMSubtarget *ST, DebugLoc dl) {
3916 if (!isa<ConstantSDNode>(N))
3918 Val = cast<ConstantSDNode>(N)->getZExtValue();
3920 if (ST->isThumb1Only()) {
3921 if (Val <= 255 || ~Val <= 255)
3922 return DAG.getConstant(Val, MVT::i32);
3924 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3925 return DAG.getConstant(Val, MVT::i32);
3930 // If this is a case we can't handle, return null and let the default
3931 // expansion code take care of it.
3932 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3933 const ARMSubtarget *ST) const {
3934 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3935 DebugLoc dl = Op.getDebugLoc();
3936 EVT VT = Op.getValueType();
3938 APInt SplatBits, SplatUndef;
3939 unsigned SplatBitSize;
3941 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3942 if (SplatBitSize <= 64) {
3943 // Check if an immediate VMOV works.
3945 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3946 SplatUndef.getZExtValue(), SplatBitSize,
3947 DAG, VmovVT, VT.is128BitVector(),
3949 if (Val.getNode()) {
3950 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3951 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3954 // Try an immediate VMVN.
3955 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3956 ((1LL << SplatBitSize) - 1));
3957 Val = isNEONModifiedImm(NegatedImm,
3958 SplatUndef.getZExtValue(), SplatBitSize,
3959 DAG, VmovVT, VT.is128BitVector(),
3961 if (Val.getNode()) {
3962 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3963 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3968 // Scan through the operands to see if only one value is used.
3969 unsigned NumElts = VT.getVectorNumElements();
3970 bool isOnlyLowElement = true;
3971 bool usesOnlyOneValue = true;
3972 bool isConstant = true;
3974 for (unsigned i = 0; i < NumElts; ++i) {
3975 SDValue V = Op.getOperand(i);
3976 if (V.getOpcode() == ISD::UNDEF)
3979 isOnlyLowElement = false;
3980 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3983 if (!Value.getNode())
3985 else if (V != Value)
3986 usesOnlyOneValue = false;
3989 if (!Value.getNode())
3990 return DAG.getUNDEF(VT);
3992 if (isOnlyLowElement)
3993 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3995 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3997 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3998 // i32 and try again.
3999 if (usesOnlyOneValue && EltSize <= 32) {
4001 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4002 if (VT.getVectorElementType().isFloatingPoint()) {
4003 SmallVector<SDValue, 8> Ops;
4004 for (unsigned i = 0; i < NumElts; ++i)
4005 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4007 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4008 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4009 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4011 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4013 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4015 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4018 // If all elements are constants and the case above didn't get hit, fall back
4019 // to the default expansion, which will generate a load from the constant
4024 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4026 SDValue shuffle = ReconstructShuffle(Op, DAG);
4027 if (shuffle != SDValue())
4031 // Vectors with 32- or 64-bit elements can be built by directly assigning
4032 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4033 // will be legalized.
4034 if (EltSize >= 32) {
4035 // Do the expansion with floating-point types, since that is what the VFP
4036 // registers are defined to use, and since i64 is not legal.
4037 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4038 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4039 SmallVector<SDValue, 8> Ops;
4040 for (unsigned i = 0; i < NumElts; ++i)
4041 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4042 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4043 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4049 // Gather data to see if the operation can be modelled as a
4050 // shuffle in combination with VEXTs.
4051 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4052 SelectionDAG &DAG) const {
4053 DebugLoc dl = Op.getDebugLoc();
4054 EVT VT = Op.getValueType();
4055 unsigned NumElts = VT.getVectorNumElements();
4057 SmallVector<SDValue, 2> SourceVecs;
4058 SmallVector<unsigned, 2> MinElts;
4059 SmallVector<unsigned, 2> MaxElts;
4061 for (unsigned i = 0; i < NumElts; ++i) {
4062 SDValue V = Op.getOperand(i);
4063 if (V.getOpcode() == ISD::UNDEF)
4065 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4066 // A shuffle can only come from building a vector from various
4067 // elements of other vectors.
4071 // Record this extraction against the appropriate vector if possible...
4072 SDValue SourceVec = V.getOperand(0);
4073 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4074 bool FoundSource = false;
4075 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4076 if (SourceVecs[j] == SourceVec) {
4077 if (MinElts[j] > EltNo)
4079 if (MaxElts[j] < EltNo)
4086 // Or record a new source if not...
4088 SourceVecs.push_back(SourceVec);
4089 MinElts.push_back(EltNo);
4090 MaxElts.push_back(EltNo);
4094 // Currently only do something sane when at most two source vectors
4096 if (SourceVecs.size() > 2)
4099 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4100 int VEXTOffsets[2] = {0, 0};
4102 // This loop extracts the usage patterns of the source vectors
4103 // and prepares appropriate SDValues for a shuffle if possible.
4104 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4105 if (SourceVecs[i].getValueType() == VT) {
4106 // No VEXT necessary
4107 ShuffleSrcs[i] = SourceVecs[i];
4110 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4111 // It probably isn't worth padding out a smaller vector just to
4112 // break it down again in a shuffle.
4116 // Since only 64-bit and 128-bit vectors are legal on ARM and
4117 // we've eliminated the other cases...
4118 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4119 "unexpected vector sizes in ReconstructShuffle");
4121 if (MaxElts[i] - MinElts[i] >= NumElts) {
4122 // Span too large for a VEXT to cope
4126 if (MinElts[i] >= NumElts) {
4127 // The extraction can just take the second half
4128 VEXTOffsets[i] = NumElts;
4129 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4131 DAG.getIntPtrConstant(NumElts));
4132 } else if (MaxElts[i] < NumElts) {
4133 // The extraction can just take the first half
4135 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4137 DAG.getIntPtrConstant(0));
4139 // An actual VEXT is needed
4140 VEXTOffsets[i] = MinElts[i];
4141 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4143 DAG.getIntPtrConstant(0));
4144 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4146 DAG.getIntPtrConstant(NumElts));
4147 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4148 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4152 SmallVector<int, 8> Mask;
4154 for (unsigned i = 0; i < NumElts; ++i) {
4155 SDValue Entry = Op.getOperand(i);
4156 if (Entry.getOpcode() == ISD::UNDEF) {
4161 SDValue ExtractVec = Entry.getOperand(0);
4162 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4163 .getOperand(1))->getSExtValue();
4164 if (ExtractVec == SourceVecs[0]) {
4165 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4167 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4171 // Final check before we try to produce nonsense...
4172 if (isShuffleMaskLegal(Mask, VT))
4173 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4179 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4180 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4181 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4182 /// are assumed to be legal.
4184 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4186 if (VT.getVectorNumElements() == 4 &&
4187 (VT.is128BitVector() || VT.is64BitVector())) {
4188 unsigned PFIndexes[4];
4189 for (unsigned i = 0; i != 4; ++i) {
4193 PFIndexes[i] = M[i];
4196 // Compute the index in the perfect shuffle table.
4197 unsigned PFTableIndex =
4198 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4199 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4200 unsigned Cost = (PFEntry >> 30);
4207 unsigned Imm, WhichResult;
4209 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4210 return (EltSize >= 32 ||
4211 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4212 isVREVMask(M, VT, 64) ||
4213 isVREVMask(M, VT, 32) ||
4214 isVREVMask(M, VT, 16) ||
4215 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4216 isVTBLMask(M, VT) ||
4217 isVTRNMask(M, VT, WhichResult) ||
4218 isVUZPMask(M, VT, WhichResult) ||
4219 isVZIPMask(M, VT, WhichResult) ||
4220 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4221 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4222 isVZIP_v_undef_Mask(M, VT, WhichResult));
4225 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4226 /// the specified operations to build the shuffle.
4227 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4228 SDValue RHS, SelectionDAG &DAG,
4230 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4231 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4232 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4235 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4244 OP_VUZPL, // VUZP, left result
4245 OP_VUZPR, // VUZP, right result
4246 OP_VZIPL, // VZIP, left result
4247 OP_VZIPR, // VZIP, right result
4248 OP_VTRNL, // VTRN, left result
4249 OP_VTRNR // VTRN, right result
4252 if (OpNum == OP_COPY) {
4253 if (LHSID == (1*9+2)*9+3) return LHS;
4254 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4258 SDValue OpLHS, OpRHS;
4259 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4260 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4261 EVT VT = OpLHS.getValueType();
4264 default: llvm_unreachable("Unknown shuffle opcode!");
4266 // VREV divides the vector in half and swaps within the half.
4267 if (VT.getVectorElementType() == MVT::i32 ||
4268 VT.getVectorElementType() == MVT::f32)
4269 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4270 // vrev <4 x i16> -> VREV32
4271 if (VT.getVectorElementType() == MVT::i16)
4272 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4273 // vrev <4 x i8> -> VREV16
4274 assert(VT.getVectorElementType() == MVT::i8);
4275 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4280 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4281 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4285 return DAG.getNode(ARMISD::VEXT, dl, VT,
4287 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4290 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4291 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4294 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4295 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4298 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4299 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4303 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4304 SmallVectorImpl<int> &ShuffleMask,
4305 SelectionDAG &DAG) {
4306 // Check to see if we can use the VTBL instruction.
4307 SDValue V1 = Op.getOperand(0);
4308 SDValue V2 = Op.getOperand(1);
4309 DebugLoc DL = Op.getDebugLoc();
4311 SmallVector<SDValue, 8> VTBLMask;
4312 for (SmallVectorImpl<int>::iterator
4313 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4314 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4316 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4317 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4318 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4321 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4322 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4326 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4327 SDValue V1 = Op.getOperand(0);
4328 SDValue V2 = Op.getOperand(1);
4329 DebugLoc dl = Op.getDebugLoc();
4330 EVT VT = Op.getValueType();
4331 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4332 SmallVector<int, 8> ShuffleMask;
4334 // Convert shuffles that are directly supported on NEON to target-specific
4335 // DAG nodes, instead of keeping them as shuffles and matching them again
4336 // during code selection. This is more efficient and avoids the possibility
4337 // of inconsistencies between legalization and selection.
4338 // FIXME: floating-point vectors should be canonicalized to integer vectors
4339 // of the same time so that they get CSEd properly.
4340 SVN->getMask(ShuffleMask);
4342 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4343 if (EltSize <= 32) {
4344 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4345 int Lane = SVN->getSplatIndex();
4346 // If this is undef splat, generate it via "just" vdup, if possible.
4347 if (Lane == -1) Lane = 0;
4349 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4350 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4352 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4353 DAG.getConstant(Lane, MVT::i32));
4358 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4361 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4362 DAG.getConstant(Imm, MVT::i32));
4365 if (isVREVMask(ShuffleMask, VT, 64))
4366 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4367 if (isVREVMask(ShuffleMask, VT, 32))
4368 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4369 if (isVREVMask(ShuffleMask, VT, 16))
4370 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4372 // Check for Neon shuffles that modify both input vectors in place.
4373 // If both results are used, i.e., if there are two shuffles with the same
4374 // source operands and with masks corresponding to both results of one of
4375 // these operations, DAG memoization will ensure that a single node is
4376 // used for both shuffles.
4377 unsigned WhichResult;
4378 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4379 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4380 V1, V2).getValue(WhichResult);
4381 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4382 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4383 V1, V2).getValue(WhichResult);
4384 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4385 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4386 V1, V2).getValue(WhichResult);
4388 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4389 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4390 V1, V1).getValue(WhichResult);
4391 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4392 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4393 V1, V1).getValue(WhichResult);
4394 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4395 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4396 V1, V1).getValue(WhichResult);
4399 // If the shuffle is not directly supported and it has 4 elements, use
4400 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4401 unsigned NumElts = VT.getVectorNumElements();
4403 unsigned PFIndexes[4];
4404 for (unsigned i = 0; i != 4; ++i) {
4405 if (ShuffleMask[i] < 0)
4408 PFIndexes[i] = ShuffleMask[i];
4411 // Compute the index in the perfect shuffle table.
4412 unsigned PFTableIndex =
4413 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4414 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4415 unsigned Cost = (PFEntry >> 30);
4418 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4421 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4422 if (EltSize >= 32) {
4423 // Do the expansion with floating-point types, since that is what the VFP
4424 // registers are defined to use, and since i64 is not legal.
4425 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4426 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4427 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4428 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4429 SmallVector<SDValue, 8> Ops;
4430 for (unsigned i = 0; i < NumElts; ++i) {
4431 if (ShuffleMask[i] < 0)
4432 Ops.push_back(DAG.getUNDEF(EltVT));
4434 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4435 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4436 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4439 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4440 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4443 if (VT == MVT::v8i8) {
4444 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4445 if (NewOp.getNode())
4452 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4453 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4454 SDValue Lane = Op.getOperand(1);
4455 if (!isa<ConstantSDNode>(Lane))
4458 SDValue Vec = Op.getOperand(0);
4459 if (Op.getValueType() == MVT::i32 &&
4460 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4461 DebugLoc dl = Op.getDebugLoc();
4462 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4468 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4469 // The only time a CONCAT_VECTORS operation can have legal types is when
4470 // two 64-bit vectors are concatenated to a 128-bit vector.
4471 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4472 "unexpected CONCAT_VECTORS");
4473 DebugLoc dl = Op.getDebugLoc();
4474 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4475 SDValue Op0 = Op.getOperand(0);
4476 SDValue Op1 = Op.getOperand(1);
4477 if (Op0.getOpcode() != ISD::UNDEF)
4478 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4479 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4480 DAG.getIntPtrConstant(0));
4481 if (Op1.getOpcode() != ISD::UNDEF)
4482 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4483 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4484 DAG.getIntPtrConstant(1));
4485 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4488 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4489 /// element has been zero/sign-extended, depending on the isSigned parameter,
4490 /// from an integer type half its size.
4491 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4493 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4494 EVT VT = N->getValueType(0);
4495 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4496 SDNode *BVN = N->getOperand(0).getNode();
4497 if (BVN->getValueType(0) != MVT::v4i32 ||
4498 BVN->getOpcode() != ISD::BUILD_VECTOR)
4500 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4501 unsigned HiElt = 1 - LoElt;
4502 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4503 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4504 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4505 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4506 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4509 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4510 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4513 if (Hi0->isNullValue() && Hi1->isNullValue())
4519 if (N->getOpcode() != ISD::BUILD_VECTOR)
4522 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4523 SDNode *Elt = N->getOperand(i).getNode();
4524 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4525 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4526 unsigned HalfSize = EltSize / 2;
4528 int64_t SExtVal = C->getSExtValue();
4529 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4532 if ((C->getZExtValue() >> HalfSize) != 0)
4543 /// isSignExtended - Check if a node is a vector value that is sign-extended
4544 /// or a constant BUILD_VECTOR with sign-extended elements.
4545 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4546 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4548 if (isExtendedBUILD_VECTOR(N, DAG, true))
4553 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4554 /// or a constant BUILD_VECTOR with zero-extended elements.
4555 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4556 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4558 if (isExtendedBUILD_VECTOR(N, DAG, false))
4563 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4564 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4565 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4566 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4567 return N->getOperand(0);
4568 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4569 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4570 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4571 LD->isNonTemporal(), LD->getAlignment());
4572 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4573 // have been legalized as a BITCAST from v4i32.
4574 if (N->getOpcode() == ISD::BITCAST) {
4575 SDNode *BVN = N->getOperand(0).getNode();
4576 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4577 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4578 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4579 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4580 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4582 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4583 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4584 EVT VT = N->getValueType(0);
4585 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4586 unsigned NumElts = VT.getVectorNumElements();
4587 MVT TruncVT = MVT::getIntegerVT(EltSize);
4588 SmallVector<SDValue, 8> Ops;
4589 for (unsigned i = 0; i != NumElts; ++i) {
4590 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4591 const APInt &CInt = C->getAPIntValue();
4592 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4594 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4595 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4598 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4599 unsigned Opcode = N->getOpcode();
4600 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4601 SDNode *N0 = N->getOperand(0).getNode();
4602 SDNode *N1 = N->getOperand(1).getNode();
4603 return N0->hasOneUse() && N1->hasOneUse() &&
4604 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4609 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4610 unsigned Opcode = N->getOpcode();
4611 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4612 SDNode *N0 = N->getOperand(0).getNode();
4613 SDNode *N1 = N->getOperand(1).getNode();
4614 return N0->hasOneUse() && N1->hasOneUse() &&
4615 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4620 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4621 // Multiplications are only custom-lowered for 128-bit vectors so that
4622 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4623 EVT VT = Op.getValueType();
4624 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4625 SDNode *N0 = Op.getOperand(0).getNode();
4626 SDNode *N1 = Op.getOperand(1).getNode();
4627 unsigned NewOpc = 0;
4629 bool isN0SExt = isSignExtended(N0, DAG);
4630 bool isN1SExt = isSignExtended(N1, DAG);
4631 if (isN0SExt && isN1SExt)
4632 NewOpc = ARMISD::VMULLs;
4634 bool isN0ZExt = isZeroExtended(N0, DAG);
4635 bool isN1ZExt = isZeroExtended(N1, DAG);
4636 if (isN0ZExt && isN1ZExt)
4637 NewOpc = ARMISD::VMULLu;
4638 else if (isN1SExt || isN1ZExt) {
4639 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4640 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4641 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4642 NewOpc = ARMISD::VMULLs;
4644 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4645 NewOpc = ARMISD::VMULLu;
4647 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4649 NewOpc = ARMISD::VMULLu;
4655 if (VT == MVT::v2i64)
4656 // Fall through to expand this. It is not legal.
4659 // Other vector multiplications are legal.
4664 // Legalize to a VMULL instruction.
4665 DebugLoc DL = Op.getDebugLoc();
4667 SDValue Op1 = SkipExtension(N1, DAG);
4669 Op0 = SkipExtension(N0, DAG);
4670 assert(Op0.getValueType().is64BitVector() &&
4671 Op1.getValueType().is64BitVector() &&
4672 "unexpected types for extended operands to VMULL");
4673 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4676 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4677 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4684 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4685 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4686 EVT Op1VT = Op1.getValueType();
4687 return DAG.getNode(N0->getOpcode(), DL, VT,
4688 DAG.getNode(NewOpc, DL, VT,
4689 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4690 DAG.getNode(NewOpc, DL, VT,
4691 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4695 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4697 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4698 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4699 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4700 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4701 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4702 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4703 // Get reciprocal estimate.
4704 // float4 recip = vrecpeq_f32(yf);
4705 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4706 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4707 // Because char has a smaller range than uchar, we can actually get away
4708 // without any newton steps. This requires that we use a weird bias
4709 // of 0xb000, however (again, this has been exhaustively tested).
4710 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4711 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4712 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4713 Y = DAG.getConstant(0xb000, MVT::i32);
4714 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4715 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4716 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4717 // Convert back to short.
4718 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4719 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4724 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4726 // Convert to float.
4727 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4728 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4729 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4730 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4731 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4732 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4734 // Use reciprocal estimate and one refinement step.
4735 // float4 recip = vrecpeq_f32(yf);
4736 // recip *= vrecpsq_f32(yf, recip);
4737 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4738 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4739 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4740 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4742 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4743 // Because short has a smaller range than ushort, we can actually get away
4744 // with only a single newton step. This requires that we use a weird bias
4745 // of 89, however (again, this has been exhaustively tested).
4746 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4747 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4748 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4749 N1 = DAG.getConstant(0x89, MVT::i32);
4750 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4751 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4752 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4753 // Convert back to integer and return.
4754 // return vmovn_s32(vcvt_s32_f32(result));
4755 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4756 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4760 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4761 EVT VT = Op.getValueType();
4762 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4763 "unexpected type for custom-lowering ISD::SDIV");
4765 DebugLoc dl = Op.getDebugLoc();
4766 SDValue N0 = Op.getOperand(0);
4767 SDValue N1 = Op.getOperand(1);
4770 if (VT == MVT::v8i8) {
4771 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4772 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4774 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4775 DAG.getIntPtrConstant(4));
4776 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4777 DAG.getIntPtrConstant(4));
4778 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4779 DAG.getIntPtrConstant(0));
4780 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4781 DAG.getIntPtrConstant(0));
4783 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4784 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4786 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4787 N0 = LowerCONCAT_VECTORS(N0, DAG);
4789 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4792 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4795 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4796 EVT VT = Op.getValueType();
4797 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4798 "unexpected type for custom-lowering ISD::UDIV");
4800 DebugLoc dl = Op.getDebugLoc();
4801 SDValue N0 = Op.getOperand(0);
4802 SDValue N1 = Op.getOperand(1);
4805 if (VT == MVT::v8i8) {
4806 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4807 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4809 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4810 DAG.getIntPtrConstant(4));
4811 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4812 DAG.getIntPtrConstant(4));
4813 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4814 DAG.getIntPtrConstant(0));
4815 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4816 DAG.getIntPtrConstant(0));
4818 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4819 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4821 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4822 N0 = LowerCONCAT_VECTORS(N0, DAG);
4824 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4825 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4830 // v4i16 sdiv ... Convert to float.
4831 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4832 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4833 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4834 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4835 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4836 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4838 // Use reciprocal estimate and two refinement steps.
4839 // float4 recip = vrecpeq_f32(yf);
4840 // recip *= vrecpsq_f32(yf, recip);
4841 // recip *= vrecpsq_f32(yf, recip);
4842 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4843 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4844 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4845 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4847 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4848 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4849 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4851 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4852 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4853 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4854 // and that it will never cause us to return an answer too large).
4855 // float4 result = as_float4(as_int4(xf*recip) + 2);
4856 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4857 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4858 N1 = DAG.getConstant(2, MVT::i32);
4859 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4860 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4861 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4862 // Convert back to integer and return.
4863 // return vmovn_u32(vcvt_s32_f32(result));
4864 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4865 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4869 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4870 switch (Op.getOpcode()) {
4871 default: llvm_unreachable("Don't know how to custom lower this!");
4872 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4873 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4874 case ISD::GlobalAddress:
4875 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4876 LowerGlobalAddressELF(Op, DAG);
4877 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4878 case ISD::SELECT: return LowerSELECT(Op, DAG);
4879 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4880 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4881 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4882 case ISD::VASTART: return LowerVASTART(Op, DAG);
4883 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4884 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
4885 case ISD::ATOMIC_CMP_SWAP:
4886 case ISD::ATOMIC_SWAP:
4887 case ISD::ATOMIC_LOAD_ADD:
4888 case ISD::ATOMIC_LOAD_SUB:
4889 case ISD::ATOMIC_LOAD_AND:
4890 case ISD::ATOMIC_LOAD_OR:
4891 case ISD::ATOMIC_LOAD_XOR:
4892 case ISD::ATOMIC_LOAD_NAND:
4893 case ISD::ATOMIC_LOAD_MIN:
4894 case ISD::ATOMIC_LOAD_MAX:
4895 case ISD::ATOMIC_LOAD_UMIN:
4896 case ISD::ATOMIC_LOAD_UMAX: return LowerAtomicMemOp(Op, DAG, Subtarget);
4897 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4898 case ISD::SINT_TO_FP:
4899 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4900 case ISD::FP_TO_SINT:
4901 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4902 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4903 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4904 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4905 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4906 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4907 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4908 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4909 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4911 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4914 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4915 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4916 case ISD::SRL_PARTS:
4917 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4918 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4919 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4920 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4921 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4922 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4923 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4924 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4925 case ISD::MUL: return LowerMUL(Op, DAG);
4926 case ISD::SDIV: return LowerSDIV(Op, DAG);
4927 case ISD::UDIV: return LowerUDIV(Op, DAG);
4932 /// ReplaceNodeResults - Replace the results of node with an illegal result
4933 /// type with new values built out of custom code.
4934 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4935 SmallVectorImpl<SDValue>&Results,
4936 SelectionDAG &DAG) const {
4938 switch (N->getOpcode()) {
4940 llvm_unreachable("Don't know how to custom expand this!");
4943 Res = ExpandBITCAST(N, DAG);
4947 Res = Expand64BitShift(N, DAG, Subtarget);
4951 Results.push_back(Res);
4954 //===----------------------------------------------------------------------===//
4955 // ARM Scheduler Hooks
4956 //===----------------------------------------------------------------------===//
4959 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4960 MachineBasicBlock *BB,
4961 unsigned Size) const {
4962 unsigned dest = MI->getOperand(0).getReg();
4963 unsigned ptr = MI->getOperand(1).getReg();
4964 unsigned oldval = MI->getOperand(2).getReg();
4965 unsigned newval = MI->getOperand(3).getReg();
4966 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4967 DebugLoc dl = MI->getDebugLoc();
4968 bool isThumb2 = Subtarget->isThumb2();
4970 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
4972 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
4973 : ARM::GPRRegisterClass);
4976 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
4977 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
4978 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
4981 unsigned ldrOpc, strOpc;
4983 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
4985 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
4986 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
4989 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
4990 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
4993 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
4994 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
4998 MachineFunction *MF = BB->getParent();
4999 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5000 MachineFunction::iterator It = BB;
5001 ++It; // insert the new blocks after the current block
5003 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5004 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5005 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5006 MF->insert(It, loop1MBB);
5007 MF->insert(It, loop2MBB);
5008 MF->insert(It, exitMBB);
5010 // Transfer the remainder of BB and its successor edges to exitMBB.
5011 exitMBB->splice(exitMBB->begin(), BB,
5012 llvm::next(MachineBasicBlock::iterator(MI)),
5014 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5018 // fallthrough --> loop1MBB
5019 BB->addSuccessor(loop1MBB);
5022 // ldrex dest, [ptr]
5026 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5027 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5028 .addReg(dest).addReg(oldval));
5029 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5030 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5031 BB->addSuccessor(loop2MBB);
5032 BB->addSuccessor(exitMBB);
5035 // strex scratch, newval, [ptr]
5039 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
5041 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5042 .addReg(scratch).addImm(0));
5043 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5044 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5045 BB->addSuccessor(loop1MBB);
5046 BB->addSuccessor(exitMBB);
5052 MI->eraseFromParent(); // The instruction is gone now.
5058 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5059 unsigned Size, unsigned BinOpcode) const {
5060 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5061 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5063 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5064 MachineFunction *MF = BB->getParent();
5065 MachineFunction::iterator It = BB;
5068 unsigned dest = MI->getOperand(0).getReg();
5069 unsigned ptr = MI->getOperand(1).getReg();
5070 unsigned incr = MI->getOperand(2).getReg();
5071 DebugLoc dl = MI->getDebugLoc();
5072 bool isThumb2 = Subtarget->isThumb2();
5074 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5076 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5077 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5080 unsigned ldrOpc, strOpc;
5082 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5084 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5085 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5088 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5089 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5092 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5093 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5097 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5098 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5099 MF->insert(It, loopMBB);
5100 MF->insert(It, exitMBB);
5102 // Transfer the remainder of BB and its successor edges to exitMBB.
5103 exitMBB->splice(exitMBB->begin(), BB,
5104 llvm::next(MachineBasicBlock::iterator(MI)),
5106 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5108 TargetRegisterClass *TRC =
5109 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5110 unsigned scratch = MRI.createVirtualRegister(TRC);
5111 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5115 // fallthrough --> loopMBB
5116 BB->addSuccessor(loopMBB);
5120 // <binop> scratch2, dest, incr
5121 // strex scratch, scratch2, ptr
5124 // fallthrough --> exitMBB
5126 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5128 // operand order needs to go the other way for NAND
5129 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5130 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5131 addReg(incr).addReg(dest)).addReg(0);
5133 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5134 addReg(dest).addReg(incr)).addReg(0);
5137 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5139 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5140 .addReg(scratch).addImm(0));
5141 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5142 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5144 BB->addSuccessor(loopMBB);
5145 BB->addSuccessor(exitMBB);
5151 MI->eraseFromParent(); // The instruction is gone now.
5157 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5158 MachineBasicBlock *BB,
5161 ARMCC::CondCodes Cond) const {
5162 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5164 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5165 MachineFunction *MF = BB->getParent();
5166 MachineFunction::iterator It = BB;
5169 unsigned dest = MI->getOperand(0).getReg();
5170 unsigned ptr = MI->getOperand(1).getReg();
5171 unsigned incr = MI->getOperand(2).getReg();
5172 unsigned oldval = dest;
5173 DebugLoc dl = MI->getDebugLoc();
5174 bool isThumb2 = Subtarget->isThumb2();
5176 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5178 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5179 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5182 unsigned ldrOpc, strOpc, extendOpc;
5184 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5186 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5187 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5188 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5191 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5192 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5193 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5196 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5197 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5202 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5203 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5204 MF->insert(It, loopMBB);
5205 MF->insert(It, exitMBB);
5207 // Transfer the remainder of BB and its successor edges to exitMBB.
5208 exitMBB->splice(exitMBB->begin(), BB,
5209 llvm::next(MachineBasicBlock::iterator(MI)),
5211 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5213 TargetRegisterClass *TRC =
5214 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5215 unsigned scratch = MRI.createVirtualRegister(TRC);
5216 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5220 // fallthrough --> loopMBB
5221 BB->addSuccessor(loopMBB);
5225 // (sign extend dest, if required)
5227 // cmov.cond scratch2, dest, incr
5228 // strex scratch, scratch2, ptr
5231 // fallthrough --> exitMBB
5233 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5235 // Sign extend the value, if necessary.
5236 if (signExtend && extendOpc) {
5237 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5238 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5243 // Build compare and cmov instructions.
5244 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5245 .addReg(oldval).addReg(incr));
5246 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5247 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5249 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5251 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5252 .addReg(scratch).addImm(0));
5253 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5254 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5256 BB->addSuccessor(loopMBB);
5257 BB->addSuccessor(exitMBB);
5263 MI->eraseFromParent(); // The instruction is gone now.
5269 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
5270 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
5271 E = MBB->succ_end(); I != E; ++I)
5274 llvm_unreachable("Expecting a BB with two successors!");
5277 // FIXME: This opcode table should obviously be expressed in the target
5278 // description. We probably just need a "machine opcode" value in the pseudo
5279 // instruction. But the ideal solution maybe to simply remove the "S" version
5280 // of the opcode altogether.
5281 struct AddSubFlagsOpcodePair {
5283 unsigned MachineOpc;
5286 static AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = {
5287 {ARM::ADCSri, ARM::ADCri},
5288 {ARM::ADCSrr, ARM::ADCrr},
5289 {ARM::ADCSrsi, ARM::ADCrsi},
5290 {ARM::ADCSrsr, ARM::ADCrsr},
5291 {ARM::SBCSri, ARM::SBCri},
5292 {ARM::SBCSrr, ARM::SBCrr},
5293 {ARM::SBCSrsi, ARM::SBCrsi},
5294 {ARM::SBCSrsr, ARM::SBCrsr},
5295 {ARM::RSBSri, ARM::RSBri},
5296 {ARM::RSBSrr, ARM::RSBrr},
5297 {ARM::RSBSrsi, ARM::RSBrsi},
5298 {ARM::RSBSrsr, ARM::RSBrsr},
5299 {ARM::RSCSri, ARM::RSCri},
5300 {ARM::RSCSrsi, ARM::RSCrsi},
5301 {ARM::RSCSrsr, ARM::RSCrsr},
5302 {ARM::t2ADCSri, ARM::t2ADCri},
5303 {ARM::t2ADCSrr, ARM::t2ADCrr},
5304 {ARM::t2ADCSrs, ARM::t2ADCrs},
5305 {ARM::t2SBCSri, ARM::t2SBCri},
5306 {ARM::t2SBCSrr, ARM::t2SBCrr},
5307 {ARM::t2SBCSrs, ARM::t2SBCrs},
5308 {ARM::t2RSBSri, ARM::t2RSBri},
5309 {ARM::t2RSBSrs, ARM::t2RSBrs},
5312 // Convert and Add or Subtract with Carry and Flags to a generic opcode with
5313 // CPSR<def> operand. e.g. ADCS (...) -> ADC (... CPSR<def>).
5315 // FIXME: Somewhere we should assert that CPSR<def> is in the correct
5316 // position to be recognized by the target descrition as the 'S' bit.
5317 bool ARMTargetLowering::RemapAddSubWithFlags(MachineInstr *MI,
5318 MachineBasicBlock *BB) const {
5319 unsigned OldOpc = MI->getOpcode();
5320 unsigned NewOpc = 0;
5322 // This is only called for instructions that need remapping, so iterating over
5323 // the tiny opcode table is not costly.
5324 static const int NPairs =
5325 sizeof(AddSubFlagsOpcodeMap) / sizeof(AddSubFlagsOpcodePair);
5326 for (AddSubFlagsOpcodePair *Pair = &AddSubFlagsOpcodeMap[0],
5327 *End = &AddSubFlagsOpcodeMap[NPairs]; Pair != End; ++Pair) {
5328 if (OldOpc == Pair->PseudoOpc) {
5329 NewOpc = Pair->MachineOpc;
5336 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5337 DebugLoc dl = MI->getDebugLoc();
5338 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
5339 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
5340 MIB.addOperand(MI->getOperand(i));
5341 AddDefaultPred(MIB);
5342 MIB.addReg(ARM::CPSR, RegState::Define); // S bit
5343 MI->eraseFromParent();
5348 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5349 MachineBasicBlock *BB) const {
5350 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5351 DebugLoc dl = MI->getDebugLoc();
5352 bool isThumb2 = Subtarget->isThumb2();
5353 switch (MI->getOpcode()) {
5355 if (RemapAddSubWithFlags(MI, BB))
5359 llvm_unreachable("Unexpected instr type to insert");
5361 case ARM::ATOMIC_LOAD_ADD_I8:
5362 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5363 case ARM::ATOMIC_LOAD_ADD_I16:
5364 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5365 case ARM::ATOMIC_LOAD_ADD_I32:
5366 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5368 case ARM::ATOMIC_LOAD_AND_I8:
5369 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5370 case ARM::ATOMIC_LOAD_AND_I16:
5371 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5372 case ARM::ATOMIC_LOAD_AND_I32:
5373 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5375 case ARM::ATOMIC_LOAD_OR_I8:
5376 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5377 case ARM::ATOMIC_LOAD_OR_I16:
5378 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5379 case ARM::ATOMIC_LOAD_OR_I32:
5380 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5382 case ARM::ATOMIC_LOAD_XOR_I8:
5383 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5384 case ARM::ATOMIC_LOAD_XOR_I16:
5385 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5386 case ARM::ATOMIC_LOAD_XOR_I32:
5387 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5389 case ARM::ATOMIC_LOAD_NAND_I8:
5390 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5391 case ARM::ATOMIC_LOAD_NAND_I16:
5392 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5393 case ARM::ATOMIC_LOAD_NAND_I32:
5394 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5396 case ARM::ATOMIC_LOAD_SUB_I8:
5397 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5398 case ARM::ATOMIC_LOAD_SUB_I16:
5399 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5400 case ARM::ATOMIC_LOAD_SUB_I32:
5401 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5403 case ARM::ATOMIC_LOAD_MIN_I8:
5404 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
5405 case ARM::ATOMIC_LOAD_MIN_I16:
5406 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
5407 case ARM::ATOMIC_LOAD_MIN_I32:
5408 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
5410 case ARM::ATOMIC_LOAD_MAX_I8:
5411 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
5412 case ARM::ATOMIC_LOAD_MAX_I16:
5413 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
5414 case ARM::ATOMIC_LOAD_MAX_I32:
5415 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
5417 case ARM::ATOMIC_LOAD_UMIN_I8:
5418 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
5419 case ARM::ATOMIC_LOAD_UMIN_I16:
5420 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
5421 case ARM::ATOMIC_LOAD_UMIN_I32:
5422 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
5424 case ARM::ATOMIC_LOAD_UMAX_I8:
5425 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
5426 case ARM::ATOMIC_LOAD_UMAX_I16:
5427 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
5428 case ARM::ATOMIC_LOAD_UMAX_I32:
5429 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
5431 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
5432 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
5433 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
5435 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
5436 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
5437 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
5439 case ARM::tMOVCCr_pseudo: {
5440 // To "insert" a SELECT_CC instruction, we actually have to insert the
5441 // diamond control-flow pattern. The incoming instruction knows the
5442 // destination vreg to set, the condition code register to branch on, the
5443 // true/false values to select between, and a branch opcode to use.
5444 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5445 MachineFunction::iterator It = BB;
5451 // cmpTY ccX, r1, r2
5453 // fallthrough --> copy0MBB
5454 MachineBasicBlock *thisMBB = BB;
5455 MachineFunction *F = BB->getParent();
5456 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5457 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5458 F->insert(It, copy0MBB);
5459 F->insert(It, sinkMBB);
5461 // Transfer the remainder of BB and its successor edges to sinkMBB.
5462 sinkMBB->splice(sinkMBB->begin(), BB,
5463 llvm::next(MachineBasicBlock::iterator(MI)),
5465 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5467 BB->addSuccessor(copy0MBB);
5468 BB->addSuccessor(sinkMBB);
5470 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
5471 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
5474 // %FalseValue = ...
5475 // # fallthrough to sinkMBB
5478 // Update machine-CFG edges
5479 BB->addSuccessor(sinkMBB);
5482 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5485 BuildMI(*BB, BB->begin(), dl,
5486 TII->get(ARM::PHI), MI->getOperand(0).getReg())
5487 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5488 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5490 MI->eraseFromParent(); // The pseudo instruction is gone now.
5495 case ARM::BCCZi64: {
5496 // If there is an unconditional branch to the other successor, remove it.
5497 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
5499 // Compare both parts that make up the double comparison separately for
5501 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
5503 unsigned LHS1 = MI->getOperand(1).getReg();
5504 unsigned LHS2 = MI->getOperand(2).getReg();
5506 AddDefaultPred(BuildMI(BB, dl,
5507 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5508 .addReg(LHS1).addImm(0));
5509 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5510 .addReg(LHS2).addImm(0)
5511 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5513 unsigned RHS1 = MI->getOperand(3).getReg();
5514 unsigned RHS2 = MI->getOperand(4).getReg();
5515 AddDefaultPred(BuildMI(BB, dl,
5516 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5517 .addReg(LHS1).addReg(RHS1));
5518 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5519 .addReg(LHS2).addReg(RHS2)
5520 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5523 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
5524 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
5525 if (MI->getOperand(0).getImm() == ARMCC::NE)
5526 std::swap(destMBB, exitMBB);
5528 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5529 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
5530 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
5533 MI->eraseFromParent(); // The pseudo instruction is gone now.
5539 //===----------------------------------------------------------------------===//
5540 // ARM Optimization Hooks
5541 //===----------------------------------------------------------------------===//
5544 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
5545 TargetLowering::DAGCombinerInfo &DCI) {
5546 SelectionDAG &DAG = DCI.DAG;
5547 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5548 EVT VT = N->getValueType(0);
5549 unsigned Opc = N->getOpcode();
5550 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
5551 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
5552 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
5553 ISD::CondCode CC = ISD::SETCC_INVALID;
5556 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
5558 SDValue CCOp = Slct.getOperand(0);
5559 if (CCOp.getOpcode() == ISD::SETCC)
5560 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
5563 bool DoXform = false;
5565 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
5568 if (LHS.getOpcode() == ISD::Constant &&
5569 cast<ConstantSDNode>(LHS)->isNullValue()) {
5571 } else if (CC != ISD::SETCC_INVALID &&
5572 RHS.getOpcode() == ISD::Constant &&
5573 cast<ConstantSDNode>(RHS)->isNullValue()) {
5574 std::swap(LHS, RHS);
5575 SDValue Op0 = Slct.getOperand(0);
5576 EVT OpVT = isSlctCC ? Op0.getValueType() :
5577 Op0.getOperand(0).getValueType();
5578 bool isInt = OpVT.isInteger();
5579 CC = ISD::getSetCCInverse(CC, isInt);
5581 if (!TLI.isCondCodeLegal(CC, OpVT))
5582 return SDValue(); // Inverse operator isn't legal.
5589 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
5591 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
5592 Slct.getOperand(0), Slct.getOperand(1), CC);
5593 SDValue CCOp = Slct.getOperand(0);
5595 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
5596 CCOp.getOperand(0), CCOp.getOperand(1), CC);
5597 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
5598 CCOp, OtherOp, Result);
5603 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
5604 // (only after legalization).
5605 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
5606 TargetLowering::DAGCombinerInfo &DCI,
5607 const ARMSubtarget *Subtarget) {
5609 // Only perform optimization if after legalize, and if NEON is available. We
5610 // also expected both operands to be BUILD_VECTORs.
5611 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
5612 || N0.getOpcode() != ISD::BUILD_VECTOR
5613 || N1.getOpcode() != ISD::BUILD_VECTOR)
5616 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
5617 EVT VT = N->getValueType(0);
5618 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
5621 // Check that the vector operands are of the right form.
5622 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
5623 // operands, where N is the size of the formed vector.
5624 // Each EXTRACT_VECTOR should have the same input vector and odd or even
5625 // index such that we have a pair wise add pattern.
5627 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
5628 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5630 SDValue Vec = N0->getOperand(0)->getOperand(0);
5631 SDNode *V = Vec.getNode();
5632 unsigned nextIndex = 0;
5634 // For each operands to the ADD which are BUILD_VECTORs,
5635 // check to see if each of their operands are an EXTRACT_VECTOR with
5636 // the same vector and appropriate index.
5637 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
5638 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
5639 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
5641 SDValue ExtVec0 = N0->getOperand(i);
5642 SDValue ExtVec1 = N1->getOperand(i);
5644 // First operand is the vector, verify its the same.
5645 if (V != ExtVec0->getOperand(0).getNode() ||
5646 V != ExtVec1->getOperand(0).getNode())
5649 // Second is the constant, verify its correct.
5650 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
5651 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
5653 // For the constant, we want to see all the even or all the odd.
5654 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
5655 || C1->getZExtValue() != nextIndex+1)
5664 // Create VPADDL node.
5665 SelectionDAG &DAG = DCI.DAG;
5666 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5668 // Build operand list.
5669 SmallVector<SDValue, 8> Ops;
5670 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
5671 TLI.getPointerTy()));
5673 // Input is the vector.
5676 // Get widened type and narrowed type.
5678 unsigned numElem = VT.getVectorNumElements();
5679 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
5680 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
5681 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
5682 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
5684 assert(0 && "Invalid vector element type for padd optimization.");
5687 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
5688 widenType, &Ops[0], Ops.size());
5689 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
5692 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
5693 /// operands N0 and N1. This is a helper for PerformADDCombine that is
5694 /// called with the default operands, and if that fails, with commuted
5696 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
5697 TargetLowering::DAGCombinerInfo &DCI,
5698 const ARMSubtarget *Subtarget){
5700 // Attempt to create vpaddl for this add.
5701 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
5702 if (Result.getNode())
5705 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
5706 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
5707 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
5708 if (Result.getNode()) return Result;
5713 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
5715 static SDValue PerformADDCombine(SDNode *N,
5716 TargetLowering::DAGCombinerInfo &DCI,
5717 const ARMSubtarget *Subtarget) {
5718 SDValue N0 = N->getOperand(0);
5719 SDValue N1 = N->getOperand(1);
5721 // First try with the default operand order.
5722 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
5723 if (Result.getNode())
5726 // If that didn't work, try again with the operands commuted.
5727 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
5730 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
5732 static SDValue PerformSUBCombine(SDNode *N,
5733 TargetLowering::DAGCombinerInfo &DCI) {
5734 SDValue N0 = N->getOperand(0);
5735 SDValue N1 = N->getOperand(1);
5737 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
5738 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
5739 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
5740 if (Result.getNode()) return Result;
5746 /// PerformVMULCombine
5747 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
5748 /// special multiplier accumulator forwarding.
5754 static SDValue PerformVMULCombine(SDNode *N,
5755 TargetLowering::DAGCombinerInfo &DCI,
5756 const ARMSubtarget *Subtarget) {
5757 if (!Subtarget->hasVMLxForwarding())
5760 SelectionDAG &DAG = DCI.DAG;
5761 SDValue N0 = N->getOperand(0);
5762 SDValue N1 = N->getOperand(1);
5763 unsigned Opcode = N0.getOpcode();
5764 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5765 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
5766 Opcode = N1.getOpcode();
5767 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5768 Opcode != ISD::FADD && Opcode != ISD::FSUB)
5773 EVT VT = N->getValueType(0);
5774 DebugLoc DL = N->getDebugLoc();
5775 SDValue N00 = N0->getOperand(0);
5776 SDValue N01 = N0->getOperand(1);
5777 return DAG.getNode(Opcode, DL, VT,
5778 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
5779 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
5782 static SDValue PerformMULCombine(SDNode *N,
5783 TargetLowering::DAGCombinerInfo &DCI,
5784 const ARMSubtarget *Subtarget) {
5785 SelectionDAG &DAG = DCI.DAG;
5787 if (Subtarget->isThumb1Only())
5790 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5793 EVT VT = N->getValueType(0);
5794 if (VT.is64BitVector() || VT.is128BitVector())
5795 return PerformVMULCombine(N, DCI, Subtarget);
5799 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
5803 uint64_t MulAmt = C->getZExtValue();
5804 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
5805 ShiftAmt = ShiftAmt & (32 - 1);
5806 SDValue V = N->getOperand(0);
5807 DebugLoc DL = N->getDebugLoc();
5810 MulAmt >>= ShiftAmt;
5811 if (isPowerOf2_32(MulAmt - 1)) {
5812 // (mul x, 2^N + 1) => (add (shl x, N), x)
5813 Res = DAG.getNode(ISD::ADD, DL, VT,
5814 V, DAG.getNode(ISD::SHL, DL, VT,
5815 V, DAG.getConstant(Log2_32(MulAmt-1),
5817 } else if (isPowerOf2_32(MulAmt + 1)) {
5818 // (mul x, 2^N - 1) => (sub (shl x, N), x)
5819 Res = DAG.getNode(ISD::SUB, DL, VT,
5820 DAG.getNode(ISD::SHL, DL, VT,
5821 V, DAG.getConstant(Log2_32(MulAmt+1),
5828 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
5829 DAG.getConstant(ShiftAmt, MVT::i32));
5831 // Do not add new nodes to DAG combiner worklist.
5832 DCI.CombineTo(N, Res, false);
5836 static SDValue PerformANDCombine(SDNode *N,
5837 TargetLowering::DAGCombinerInfo &DCI) {
5839 // Attempt to use immediate-form VBIC
5840 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5841 DebugLoc dl = N->getDebugLoc();
5842 EVT VT = N->getValueType(0);
5843 SelectionDAG &DAG = DCI.DAG;
5845 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5848 APInt SplatBits, SplatUndef;
5849 unsigned SplatBitSize;
5852 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5853 if (SplatBitSize <= 64) {
5855 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
5856 SplatUndef.getZExtValue(), SplatBitSize,
5857 DAG, VbicVT, VT.is128BitVector(),
5859 if (Val.getNode()) {
5861 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
5862 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
5863 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
5871 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
5872 static SDValue PerformORCombine(SDNode *N,
5873 TargetLowering::DAGCombinerInfo &DCI,
5874 const ARMSubtarget *Subtarget) {
5875 // Attempt to use immediate-form VORR
5876 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5877 DebugLoc dl = N->getDebugLoc();
5878 EVT VT = N->getValueType(0);
5879 SelectionDAG &DAG = DCI.DAG;
5881 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5884 APInt SplatBits, SplatUndef;
5885 unsigned SplatBitSize;
5887 if (BVN && Subtarget->hasNEON() &&
5888 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5889 if (SplatBitSize <= 64) {
5891 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5892 SplatUndef.getZExtValue(), SplatBitSize,
5893 DAG, VorrVT, VT.is128BitVector(),
5895 if (Val.getNode()) {
5897 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
5898 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
5899 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
5904 SDValue N0 = N->getOperand(0);
5905 if (N0.getOpcode() != ISD::AND)
5907 SDValue N1 = N->getOperand(1);
5909 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
5910 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
5911 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
5913 unsigned SplatBitSize;
5916 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
5918 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
5919 HasAnyUndefs) && !HasAnyUndefs) {
5920 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
5922 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
5923 HasAnyUndefs) && !HasAnyUndefs &&
5924 SplatBits0 == ~SplatBits1) {
5925 // Canonicalize the vector type to make instruction selection simpler.
5926 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
5927 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
5928 N0->getOperand(1), N0->getOperand(0),
5930 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
5935 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
5938 // BFI is only available on V6T2+
5939 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
5942 DebugLoc DL = N->getDebugLoc();
5943 // 1) or (and A, mask), val => ARMbfi A, val, mask
5944 // iff (val & mask) == val
5946 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5947 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
5948 // && mask == ~mask2
5949 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
5950 // && ~mask == mask2
5951 // (i.e., copy a bitfield value into another bitfield of the same width)
5956 SDValue N00 = N0.getOperand(0);
5958 // The value and the mask need to be constants so we can verify this is
5959 // actually a bitfield set. If the mask is 0xffff, we can do better
5960 // via a movt instruction, so don't use BFI in that case.
5961 SDValue MaskOp = N0.getOperand(1);
5962 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
5965 unsigned Mask = MaskC->getZExtValue();
5969 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
5970 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
5972 unsigned Val = N1C->getZExtValue();
5973 if ((Val & ~Mask) != Val)
5976 if (ARM::isBitFieldInvertedMask(Mask)) {
5977 Val >>= CountTrailingZeros_32(~Mask);
5979 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
5980 DAG.getConstant(Val, MVT::i32),
5981 DAG.getConstant(Mask, MVT::i32));
5983 // Do not add new nodes to DAG combiner worklist.
5984 DCI.CombineTo(N, Res, false);
5987 } else if (N1.getOpcode() == ISD::AND) {
5988 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
5989 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
5992 unsigned Mask2 = N11C->getZExtValue();
5994 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
5996 if (ARM::isBitFieldInvertedMask(Mask) &&
5998 // The pack halfword instruction works better for masks that fit it,
5999 // so use that when it's available.
6000 if (Subtarget->hasT2ExtractPack() &&
6001 (Mask == 0xffff || Mask == 0xffff0000))
6004 unsigned amt = CountTrailingZeros_32(Mask2);
6005 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
6006 DAG.getConstant(amt, MVT::i32));
6007 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
6008 DAG.getConstant(Mask, MVT::i32));
6009 // Do not add new nodes to DAG combiner worklist.
6010 DCI.CombineTo(N, Res, false);
6012 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
6014 // The pack halfword instruction works better for masks that fit it,
6015 // so use that when it's available.
6016 if (Subtarget->hasT2ExtractPack() &&
6017 (Mask2 == 0xffff || Mask2 == 0xffff0000))
6020 unsigned lsb = CountTrailingZeros_32(Mask);
6021 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
6022 DAG.getConstant(lsb, MVT::i32));
6023 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
6024 DAG.getConstant(Mask2, MVT::i32));
6025 // Do not add new nodes to DAG combiner worklist.
6026 DCI.CombineTo(N, Res, false);
6031 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
6032 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
6033 ARM::isBitFieldInvertedMask(~Mask)) {
6034 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
6035 // where lsb(mask) == #shamt and masked bits of B are known zero.
6036 SDValue ShAmt = N00.getOperand(1);
6037 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6038 unsigned LSB = CountTrailingZeros_32(Mask);
6042 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
6043 DAG.getConstant(~Mask, MVT::i32));
6045 // Do not add new nodes to DAG combiner worklist.
6046 DCI.CombineTo(N, Res, false);
6052 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
6053 /// the bits being cleared by the AND are not demanded by the BFI.
6054 static SDValue PerformBFICombine(SDNode *N,
6055 TargetLowering::DAGCombinerInfo &DCI) {
6056 SDValue N1 = N->getOperand(1);
6057 if (N1.getOpcode() == ISD::AND) {
6058 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6061 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
6062 unsigned LSB = CountTrailingZeros_32(~InvMask);
6063 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
6064 unsigned Mask = (1 << Width)-1;
6065 unsigned Mask2 = N11C->getZExtValue();
6066 if ((Mask & (~Mask2)) == 0)
6067 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
6068 N->getOperand(0), N1.getOperand(0),
6074 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
6075 /// ARMISD::VMOVRRD.
6076 static SDValue PerformVMOVRRDCombine(SDNode *N,
6077 TargetLowering::DAGCombinerInfo &DCI) {
6078 // vmovrrd(vmovdrr x, y) -> x,y
6079 SDValue InDouble = N->getOperand(0);
6080 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
6081 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
6083 // vmovrrd(load f64) -> (load i32), (load i32)
6084 SDNode *InNode = InDouble.getNode();
6085 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
6086 InNode->getValueType(0) == MVT::f64 &&
6087 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
6088 !cast<LoadSDNode>(InNode)->isVolatile()) {
6089 // TODO: Should this be done for non-FrameIndex operands?
6090 LoadSDNode *LD = cast<LoadSDNode>(InNode);
6092 SelectionDAG &DAG = DCI.DAG;
6093 DebugLoc DL = LD->getDebugLoc();
6094 SDValue BasePtr = LD->getBasePtr();
6095 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
6096 LD->getPointerInfo(), LD->isVolatile(),
6097 LD->isNonTemporal(), LD->getAlignment());
6099 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
6100 DAG.getConstant(4, MVT::i32));
6101 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
6102 LD->getPointerInfo(), LD->isVolatile(),
6103 LD->isNonTemporal(),
6104 std::min(4U, LD->getAlignment() / 2));
6106 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
6107 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
6108 DCI.RemoveFromWorklist(LD);
6116 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
6117 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
6118 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
6119 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
6120 SDValue Op0 = N->getOperand(0);
6121 SDValue Op1 = N->getOperand(1);
6122 if (Op0.getOpcode() == ISD::BITCAST)
6123 Op0 = Op0.getOperand(0);
6124 if (Op1.getOpcode() == ISD::BITCAST)
6125 Op1 = Op1.getOperand(0);
6126 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
6127 Op0.getNode() == Op1.getNode() &&
6128 Op0.getResNo() == 0 && Op1.getResNo() == 1)
6129 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6130 N->getValueType(0), Op0.getOperand(0));
6134 /// PerformSTORECombine - Target-specific dag combine xforms for
6136 static SDValue PerformSTORECombine(SDNode *N,
6137 TargetLowering::DAGCombinerInfo &DCI) {
6138 // Bitcast an i64 store extracted from a vector to f64.
6139 // Otherwise, the i64 value will be legalized to a pair of i32 values.
6140 StoreSDNode *St = cast<StoreSDNode>(N);
6141 SDValue StVal = St->getValue();
6142 if (!ISD::isNormalStore(St) || St->isVolatile())
6145 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
6146 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
6147 SelectionDAG &DAG = DCI.DAG;
6148 DebugLoc DL = St->getDebugLoc();
6149 SDValue BasePtr = St->getBasePtr();
6150 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
6151 StVal.getNode()->getOperand(0), BasePtr,
6152 St->getPointerInfo(), St->isVolatile(),
6153 St->isNonTemporal(), St->getAlignment());
6155 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
6156 DAG.getConstant(4, MVT::i32));
6157 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
6158 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
6159 St->isNonTemporal(),
6160 std::min(4U, St->getAlignment() / 2));
6163 if (StVal.getValueType() != MVT::i64 ||
6164 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6167 SelectionDAG &DAG = DCI.DAG;
6168 DebugLoc dl = StVal.getDebugLoc();
6169 SDValue IntVec = StVal.getOperand(0);
6170 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
6171 IntVec.getValueType().getVectorNumElements());
6172 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
6173 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6174 Vec, StVal.getOperand(1));
6175 dl = N->getDebugLoc();
6176 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
6177 // Make the DAGCombiner fold the bitcasts.
6178 DCI.AddToWorklist(Vec.getNode());
6179 DCI.AddToWorklist(ExtElt.getNode());
6180 DCI.AddToWorklist(V.getNode());
6181 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
6182 St->getPointerInfo(), St->isVolatile(),
6183 St->isNonTemporal(), St->getAlignment(),
6187 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
6188 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
6189 /// i64 vector to have f64 elements, since the value can then be loaded
6190 /// directly into a VFP register.
6191 static bool hasNormalLoadOperand(SDNode *N) {
6192 unsigned NumElts = N->getValueType(0).getVectorNumElements();
6193 for (unsigned i = 0; i < NumElts; ++i) {
6194 SDNode *Elt = N->getOperand(i).getNode();
6195 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
6201 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
6202 /// ISD::BUILD_VECTOR.
6203 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
6204 TargetLowering::DAGCombinerInfo &DCI){
6205 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
6206 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
6207 // into a pair of GPRs, which is fine when the value is used as a scalar,
6208 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
6209 SelectionDAG &DAG = DCI.DAG;
6210 if (N->getNumOperands() == 2) {
6211 SDValue RV = PerformVMOVDRRCombine(N, DAG);
6216 // Load i64 elements as f64 values so that type legalization does not split
6217 // them up into i32 values.
6218 EVT VT = N->getValueType(0);
6219 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
6221 DebugLoc dl = N->getDebugLoc();
6222 SmallVector<SDValue, 8> Ops;
6223 unsigned NumElts = VT.getVectorNumElements();
6224 for (unsigned i = 0; i < NumElts; ++i) {
6225 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
6227 // Make the DAGCombiner fold the bitcast.
6228 DCI.AddToWorklist(V.getNode());
6230 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
6231 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
6232 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
6235 /// PerformInsertEltCombine - Target-specific dag combine xforms for
6236 /// ISD::INSERT_VECTOR_ELT.
6237 static SDValue PerformInsertEltCombine(SDNode *N,
6238 TargetLowering::DAGCombinerInfo &DCI) {
6239 // Bitcast an i64 load inserted into a vector to f64.
6240 // Otherwise, the i64 value will be legalized to a pair of i32 values.
6241 EVT VT = N->getValueType(0);
6242 SDNode *Elt = N->getOperand(1).getNode();
6243 if (VT.getVectorElementType() != MVT::i64 ||
6244 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
6247 SelectionDAG &DAG = DCI.DAG;
6248 DebugLoc dl = N->getDebugLoc();
6249 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
6250 VT.getVectorNumElements());
6251 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
6252 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
6253 // Make the DAGCombiner fold the bitcasts.
6254 DCI.AddToWorklist(Vec.getNode());
6255 DCI.AddToWorklist(V.getNode());
6256 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
6257 Vec, V, N->getOperand(2));
6258 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
6261 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
6262 /// ISD::VECTOR_SHUFFLE.
6263 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
6264 // The LLVM shufflevector instruction does not require the shuffle mask
6265 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
6266 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
6267 // operands do not match the mask length, they are extended by concatenating
6268 // them with undef vectors. That is probably the right thing for other
6269 // targets, but for NEON it is better to concatenate two double-register
6270 // size vector operands into a single quad-register size vector. Do that
6271 // transformation here:
6272 // shuffle(concat(v1, undef), concat(v2, undef)) ->
6273 // shuffle(concat(v1, v2), undef)
6274 SDValue Op0 = N->getOperand(0);
6275 SDValue Op1 = N->getOperand(1);
6276 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
6277 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
6278 Op0.getNumOperands() != 2 ||
6279 Op1.getNumOperands() != 2)
6281 SDValue Concat0Op1 = Op0.getOperand(1);
6282 SDValue Concat1Op1 = Op1.getOperand(1);
6283 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
6284 Concat1Op1.getOpcode() != ISD::UNDEF)
6286 // Skip the transformation if any of the types are illegal.
6287 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6288 EVT VT = N->getValueType(0);
6289 if (!TLI.isTypeLegal(VT) ||
6290 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
6291 !TLI.isTypeLegal(Concat1Op1.getValueType()))
6294 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
6295 Op0.getOperand(0), Op1.getOperand(0));
6296 // Translate the shuffle mask.
6297 SmallVector<int, 16> NewMask;
6298 unsigned NumElts = VT.getVectorNumElements();
6299 unsigned HalfElts = NumElts/2;
6300 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
6301 for (unsigned n = 0; n < NumElts; ++n) {
6302 int MaskElt = SVN->getMaskElt(n);
6304 if (MaskElt < (int)HalfElts)
6306 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
6307 NewElt = HalfElts + MaskElt - NumElts;
6308 NewMask.push_back(NewElt);
6310 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
6311 DAG.getUNDEF(VT), NewMask.data());
6314 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
6315 /// NEON load/store intrinsics to merge base address updates.
6316 static SDValue CombineBaseUpdate(SDNode *N,
6317 TargetLowering::DAGCombinerInfo &DCI) {
6318 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6321 SelectionDAG &DAG = DCI.DAG;
6322 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
6323 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
6324 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
6325 SDValue Addr = N->getOperand(AddrOpIdx);
6327 // Search for a use of the address operand that is an increment.
6328 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
6329 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
6331 if (User->getOpcode() != ISD::ADD ||
6332 UI.getUse().getResNo() != Addr.getResNo())
6335 // Check that the add is independent of the load/store. Otherwise, folding
6336 // it would create a cycle.
6337 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
6340 // Find the new opcode for the updating load/store.
6342 bool isLaneOp = false;
6343 unsigned NewOpc = 0;
6344 unsigned NumVecs = 0;
6346 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
6348 default: assert(0 && "unexpected intrinsic for Neon base update");
6349 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
6351 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
6353 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
6355 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
6357 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
6358 NumVecs = 2; isLaneOp = true; break;
6359 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
6360 NumVecs = 3; isLaneOp = true; break;
6361 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
6362 NumVecs = 4; isLaneOp = true; break;
6363 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
6364 NumVecs = 1; isLoad = false; break;
6365 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
6366 NumVecs = 2; isLoad = false; break;
6367 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
6368 NumVecs = 3; isLoad = false; break;
6369 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
6370 NumVecs = 4; isLoad = false; break;
6371 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
6372 NumVecs = 2; isLoad = false; isLaneOp = true; break;
6373 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
6374 NumVecs = 3; isLoad = false; isLaneOp = true; break;
6375 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
6376 NumVecs = 4; isLoad = false; isLaneOp = true; break;
6380 switch (N->getOpcode()) {
6381 default: assert(0 && "unexpected opcode for Neon base update");
6382 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
6383 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
6384 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
6388 // Find the size of memory referenced by the load/store.
6391 VecTy = N->getValueType(0);
6393 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
6394 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
6396 NumBytes /= VecTy.getVectorNumElements();
6398 // If the increment is a constant, it must match the memory ref size.
6399 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
6400 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
6401 uint64_t IncVal = CInc->getZExtValue();
6402 if (IncVal != NumBytes)
6404 } else if (NumBytes >= 3 * 16) {
6405 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
6406 // separate instructions that make it harder to use a non-constant update.
6410 // Create the new updating load/store node.
6412 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
6414 for (n = 0; n < NumResultVecs; ++n)
6416 Tys[n++] = MVT::i32;
6417 Tys[n] = MVT::Other;
6418 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
6419 SmallVector<SDValue, 8> Ops;
6420 Ops.push_back(N->getOperand(0)); // incoming chain
6421 Ops.push_back(N->getOperand(AddrOpIdx));
6423 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
6424 Ops.push_back(N->getOperand(i));
6426 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
6427 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
6428 Ops.data(), Ops.size(),
6429 MemInt->getMemoryVT(),
6430 MemInt->getMemOperand());
6433 std::vector<SDValue> NewResults;
6434 for (unsigned i = 0; i < NumResultVecs; ++i) {
6435 NewResults.push_back(SDValue(UpdN.getNode(), i));
6437 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
6438 DCI.CombineTo(N, NewResults);
6439 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
6446 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
6447 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
6448 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
6450 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
6451 SelectionDAG &DAG = DCI.DAG;
6452 EVT VT = N->getValueType(0);
6453 // vldN-dup instructions only support 64-bit vectors for N > 1.
6454 if (!VT.is64BitVector())
6457 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
6458 SDNode *VLD = N->getOperand(0).getNode();
6459 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
6461 unsigned NumVecs = 0;
6462 unsigned NewOpc = 0;
6463 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
6464 if (IntNo == Intrinsic::arm_neon_vld2lane) {
6466 NewOpc = ARMISD::VLD2DUP;
6467 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
6469 NewOpc = ARMISD::VLD3DUP;
6470 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
6472 NewOpc = ARMISD::VLD4DUP;
6477 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
6478 // numbers match the load.
6479 unsigned VLDLaneNo =
6480 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
6481 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6483 // Ignore uses of the chain result.
6484 if (UI.getUse().getResNo() == NumVecs)
6487 if (User->getOpcode() != ARMISD::VDUPLANE ||
6488 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
6492 // Create the vldN-dup node.
6495 for (n = 0; n < NumVecs; ++n)
6497 Tys[n] = MVT::Other;
6498 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
6499 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
6500 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
6501 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
6502 Ops, 2, VLDMemInt->getMemoryVT(),
6503 VLDMemInt->getMemOperand());
6506 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6508 unsigned ResNo = UI.getUse().getResNo();
6509 // Ignore uses of the chain result.
6510 if (ResNo == NumVecs)
6513 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
6516 // Now the vldN-lane intrinsic is dead except for its chain result.
6517 // Update uses of the chain.
6518 std::vector<SDValue> VLDDupResults;
6519 for (unsigned n = 0; n < NumVecs; ++n)
6520 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
6521 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
6522 DCI.CombineTo(VLD, VLDDupResults);
6527 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
6528 /// ARMISD::VDUPLANE.
6529 static SDValue PerformVDUPLANECombine(SDNode *N,
6530 TargetLowering::DAGCombinerInfo &DCI) {
6531 SDValue Op = N->getOperand(0);
6533 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
6534 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
6535 if (CombineVLDDUP(N, DCI))
6536 return SDValue(N, 0);
6538 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
6539 // redundant. Ignore bit_converts for now; element sizes are checked below.
6540 while (Op.getOpcode() == ISD::BITCAST)
6541 Op = Op.getOperand(0);
6542 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
6545 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
6546 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
6547 // The canonical VMOV for a zero vector uses a 32-bit element size.
6548 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6550 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
6552 EVT VT = N->getValueType(0);
6553 if (EltSize > VT.getVectorElementType().getSizeInBits())
6556 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
6559 // isConstVecPow2 - Return true if each vector element is a power of 2, all
6560 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
6561 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
6565 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
6567 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
6572 APFloat APF = C->getValueAPF();
6573 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
6574 != APFloat::opOK || !isExact)
6577 c0 = (I == 0) ? cN : c0;
6578 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
6585 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
6586 /// can replace combinations of VMUL and VCVT (floating-point to integer)
6587 /// when the VMUL has a constant operand that is a power of 2.
6589 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
6590 /// vmul.f32 d16, d17, d16
6591 /// vcvt.s32.f32 d16, d16
6593 /// vcvt.s32.f32 d16, d16, #3
6594 static SDValue PerformVCVTCombine(SDNode *N,
6595 TargetLowering::DAGCombinerInfo &DCI,
6596 const ARMSubtarget *Subtarget) {
6597 SelectionDAG &DAG = DCI.DAG;
6598 SDValue Op = N->getOperand(0);
6600 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
6601 Op.getOpcode() != ISD::FMUL)
6605 SDValue N0 = Op->getOperand(0);
6606 SDValue ConstVec = Op->getOperand(1);
6607 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
6609 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
6610 !isConstVecPow2(ConstVec, isSigned, C))
6613 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
6614 Intrinsic::arm_neon_vcvtfp2fxu;
6615 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6617 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
6618 DAG.getConstant(Log2_64(C), MVT::i32));
6621 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
6622 /// can replace combinations of VCVT (integer to floating-point) and VDIV
6623 /// when the VDIV has a constant operand that is a power of 2.
6625 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
6626 /// vcvt.f32.s32 d16, d16
6627 /// vdiv.f32 d16, d17, d16
6629 /// vcvt.f32.s32 d16, d16, #3
6630 static SDValue PerformVDIVCombine(SDNode *N,
6631 TargetLowering::DAGCombinerInfo &DCI,
6632 const ARMSubtarget *Subtarget) {
6633 SelectionDAG &DAG = DCI.DAG;
6634 SDValue Op = N->getOperand(0);
6635 unsigned OpOpcode = Op.getNode()->getOpcode();
6637 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
6638 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
6642 SDValue ConstVec = N->getOperand(1);
6643 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
6645 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
6646 !isConstVecPow2(ConstVec, isSigned, C))
6649 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
6650 Intrinsic::arm_neon_vcvtfxu2fp;
6651 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6653 DAG.getConstant(IntrinsicOpcode, MVT::i32),
6654 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
6657 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
6658 /// operand of a vector shift operation, where all the elements of the
6659 /// build_vector must have the same constant integer value.
6660 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6661 // Ignore bit_converts.
6662 while (Op.getOpcode() == ISD::BITCAST)
6663 Op = Op.getOperand(0);
6664 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6665 APInt SplatBits, SplatUndef;
6666 unsigned SplatBitSize;
6668 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6669 HasAnyUndefs, ElementBits) ||
6670 SplatBitSize > ElementBits)
6672 Cnt = SplatBits.getSExtValue();
6676 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6677 /// operand of a vector shift left operation. That value must be in the range:
6678 /// 0 <= Value < ElementBits for a left shift; or
6679 /// 0 <= Value <= ElementBits for a long left shift.
6680 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6681 assert(VT.isVector() && "vector shift count is not a vector type");
6682 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6683 if (! getVShiftImm(Op, ElementBits, Cnt))
6685 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
6688 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6689 /// operand of a vector shift right operation. For a shift opcode, the value
6690 /// is positive, but for an intrinsic the value count must be negative. The
6691 /// absolute value must be in the range:
6692 /// 1 <= |Value| <= ElementBits for a right shift; or
6693 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6694 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6696 assert(VT.isVector() && "vector shift count is not a vector type");
6697 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6698 if (! getVShiftImm(Op, ElementBits, Cnt))
6702 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
6705 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
6706 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
6707 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
6710 // Don't do anything for most intrinsics.
6713 // Vector shifts: check for immediate versions and lower them.
6714 // Note: This is done during DAG combining instead of DAG legalizing because
6715 // the build_vectors for 64-bit vector element shift counts are generally
6716 // not legal, and it is hard to see their values after they get legalized to
6717 // loads from a constant pool.
6718 case Intrinsic::arm_neon_vshifts:
6719 case Intrinsic::arm_neon_vshiftu:
6720 case Intrinsic::arm_neon_vshiftls:
6721 case Intrinsic::arm_neon_vshiftlu:
6722 case Intrinsic::arm_neon_vshiftn:
6723 case Intrinsic::arm_neon_vrshifts:
6724 case Intrinsic::arm_neon_vrshiftu:
6725 case Intrinsic::arm_neon_vrshiftn:
6726 case Intrinsic::arm_neon_vqshifts:
6727 case Intrinsic::arm_neon_vqshiftu:
6728 case Intrinsic::arm_neon_vqshiftsu:
6729 case Intrinsic::arm_neon_vqshiftns:
6730 case Intrinsic::arm_neon_vqshiftnu:
6731 case Intrinsic::arm_neon_vqshiftnsu:
6732 case Intrinsic::arm_neon_vqrshiftns:
6733 case Intrinsic::arm_neon_vqrshiftnu:
6734 case Intrinsic::arm_neon_vqrshiftnsu: {
6735 EVT VT = N->getOperand(1).getValueType();
6737 unsigned VShiftOpc = 0;
6740 case Intrinsic::arm_neon_vshifts:
6741 case Intrinsic::arm_neon_vshiftu:
6742 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
6743 VShiftOpc = ARMISD::VSHL;
6746 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
6747 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
6748 ARMISD::VSHRs : ARMISD::VSHRu);
6753 case Intrinsic::arm_neon_vshiftls:
6754 case Intrinsic::arm_neon_vshiftlu:
6755 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
6757 llvm_unreachable("invalid shift count for vshll intrinsic");
6759 case Intrinsic::arm_neon_vrshifts:
6760 case Intrinsic::arm_neon_vrshiftu:
6761 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
6765 case Intrinsic::arm_neon_vqshifts:
6766 case Intrinsic::arm_neon_vqshiftu:
6767 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6771 case Intrinsic::arm_neon_vqshiftsu:
6772 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6774 llvm_unreachable("invalid shift count for vqshlu intrinsic");
6776 case Intrinsic::arm_neon_vshiftn:
6777 case Intrinsic::arm_neon_vrshiftn:
6778 case Intrinsic::arm_neon_vqshiftns:
6779 case Intrinsic::arm_neon_vqshiftnu:
6780 case Intrinsic::arm_neon_vqshiftnsu:
6781 case Intrinsic::arm_neon_vqrshiftns:
6782 case Intrinsic::arm_neon_vqrshiftnu:
6783 case Intrinsic::arm_neon_vqrshiftnsu:
6784 // Narrowing shifts require an immediate right shift.
6785 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
6787 llvm_unreachable("invalid shift count for narrowing vector shift "
6791 llvm_unreachable("unhandled vector shift");
6795 case Intrinsic::arm_neon_vshifts:
6796 case Intrinsic::arm_neon_vshiftu:
6797 // Opcode already set above.
6799 case Intrinsic::arm_neon_vshiftls:
6800 case Intrinsic::arm_neon_vshiftlu:
6801 if (Cnt == VT.getVectorElementType().getSizeInBits())
6802 VShiftOpc = ARMISD::VSHLLi;
6804 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
6805 ARMISD::VSHLLs : ARMISD::VSHLLu);
6807 case Intrinsic::arm_neon_vshiftn:
6808 VShiftOpc = ARMISD::VSHRN; break;
6809 case Intrinsic::arm_neon_vrshifts:
6810 VShiftOpc = ARMISD::VRSHRs; break;
6811 case Intrinsic::arm_neon_vrshiftu:
6812 VShiftOpc = ARMISD::VRSHRu; break;
6813 case Intrinsic::arm_neon_vrshiftn:
6814 VShiftOpc = ARMISD::VRSHRN; break;
6815 case Intrinsic::arm_neon_vqshifts:
6816 VShiftOpc = ARMISD::VQSHLs; break;
6817 case Intrinsic::arm_neon_vqshiftu:
6818 VShiftOpc = ARMISD::VQSHLu; break;
6819 case Intrinsic::arm_neon_vqshiftsu:
6820 VShiftOpc = ARMISD::VQSHLsu; break;
6821 case Intrinsic::arm_neon_vqshiftns:
6822 VShiftOpc = ARMISD::VQSHRNs; break;
6823 case Intrinsic::arm_neon_vqshiftnu:
6824 VShiftOpc = ARMISD::VQSHRNu; break;
6825 case Intrinsic::arm_neon_vqshiftnsu:
6826 VShiftOpc = ARMISD::VQSHRNsu; break;
6827 case Intrinsic::arm_neon_vqrshiftns:
6828 VShiftOpc = ARMISD::VQRSHRNs; break;
6829 case Intrinsic::arm_neon_vqrshiftnu:
6830 VShiftOpc = ARMISD::VQRSHRNu; break;
6831 case Intrinsic::arm_neon_vqrshiftnsu:
6832 VShiftOpc = ARMISD::VQRSHRNsu; break;
6835 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6836 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
6839 case Intrinsic::arm_neon_vshiftins: {
6840 EVT VT = N->getOperand(1).getValueType();
6842 unsigned VShiftOpc = 0;
6844 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
6845 VShiftOpc = ARMISD::VSLI;
6846 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
6847 VShiftOpc = ARMISD::VSRI;
6849 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
6852 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6853 N->getOperand(1), N->getOperand(2),
6854 DAG.getConstant(Cnt, MVT::i32));
6857 case Intrinsic::arm_neon_vqrshifts:
6858 case Intrinsic::arm_neon_vqrshiftu:
6859 // No immediate versions of these to check for.
6866 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
6867 /// lowers them. As with the vector shift intrinsics, this is done during DAG
6868 /// combining instead of DAG legalizing because the build_vectors for 64-bit
6869 /// vector element shift counts are generally not legal, and it is hard to see
6870 /// their values after they get legalized to loads from a constant pool.
6871 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
6872 const ARMSubtarget *ST) {
6873 EVT VT = N->getValueType(0);
6875 // Nothing to be done for scalar shifts.
6876 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6877 if (!VT.isVector() || !TLI.isTypeLegal(VT))
6880 assert(ST->hasNEON() && "unexpected vector shift");
6883 switch (N->getOpcode()) {
6884 default: llvm_unreachable("unexpected shift opcode");
6887 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
6888 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
6889 DAG.getConstant(Cnt, MVT::i32));
6894 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
6895 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
6896 ARMISD::VSHRs : ARMISD::VSHRu);
6897 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
6898 DAG.getConstant(Cnt, MVT::i32));
6904 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
6905 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
6906 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
6907 const ARMSubtarget *ST) {
6908 SDValue N0 = N->getOperand(0);
6910 // Check for sign- and zero-extensions of vector extract operations of 8-
6911 // and 16-bit vector elements. NEON supports these directly. They are
6912 // handled during DAG combining because type legalization will promote them
6913 // to 32-bit types and it is messy to recognize the operations after that.
6914 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6915 SDValue Vec = N0.getOperand(0);
6916 SDValue Lane = N0.getOperand(1);
6917 EVT VT = N->getValueType(0);
6918 EVT EltVT = N0.getValueType();
6919 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6921 if (VT == MVT::i32 &&
6922 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
6923 TLI.isTypeLegal(Vec.getValueType()) &&
6924 isa<ConstantSDNode>(Lane)) {
6927 switch (N->getOpcode()) {
6928 default: llvm_unreachable("unexpected opcode");
6929 case ISD::SIGN_EXTEND:
6930 Opc = ARMISD::VGETLANEs;
6932 case ISD::ZERO_EXTEND:
6933 case ISD::ANY_EXTEND:
6934 Opc = ARMISD::VGETLANEu;
6937 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
6944 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
6945 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
6946 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
6947 const ARMSubtarget *ST) {
6948 // If the target supports NEON, try to use vmax/vmin instructions for f32
6949 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
6950 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
6951 // a NaN; only do the transformation when it matches that behavior.
6953 // For now only do this when using NEON for FP operations; if using VFP, it
6954 // is not obvious that the benefit outweighs the cost of switching to the
6956 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
6957 N->getValueType(0) != MVT::f32)
6960 SDValue CondLHS = N->getOperand(0);
6961 SDValue CondRHS = N->getOperand(1);
6962 SDValue LHS = N->getOperand(2);
6963 SDValue RHS = N->getOperand(3);
6964 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
6966 unsigned Opcode = 0;
6968 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
6969 IsReversed = false; // x CC y ? x : y
6970 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
6971 IsReversed = true ; // x CC y ? y : x
6985 // If LHS is NaN, an ordered comparison will be false and the result will
6986 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
6987 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
6988 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
6989 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
6991 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
6992 // will return -0, so vmin can only be used for unsafe math or if one of
6993 // the operands is known to be nonzero.
6994 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
6996 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
6998 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
7007 // If LHS is NaN, an ordered comparison will be false and the result will
7008 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
7009 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7010 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
7011 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7013 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
7014 // will return +0, so vmax can only be used for unsafe math or if one of
7015 // the operands is known to be nonzero.
7016 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
7018 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7020 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
7026 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
7029 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
7031 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
7032 SDValue Cmp = N->getOperand(4);
7033 if (Cmp.getOpcode() != ARMISD::CMPZ)
7034 // Only looking at EQ and NE cases.
7037 EVT VT = N->getValueType(0);
7038 DebugLoc dl = N->getDebugLoc();
7039 SDValue LHS = Cmp.getOperand(0);
7040 SDValue RHS = Cmp.getOperand(1);
7041 SDValue FalseVal = N->getOperand(0);
7042 SDValue TrueVal = N->getOperand(1);
7043 SDValue ARMcc = N->getOperand(2);
7044 ARMCC::CondCodes CC = (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
7062 /// FIXME: Turn this into a target neutral optimization?
7064 if (CC == ARMCC::NE && FalseVal == RHS) {
7065 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
7066 N->getOperand(3), Cmp);
7067 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
7069 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
7070 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
7071 N->getOperand(3), NewCmp);
7074 if (Res.getNode()) {
7075 APInt KnownZero, KnownOne;
7076 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
7077 DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
7078 // Capture demanded bits information that would be otherwise lost.
7079 if (KnownZero == 0xfffffffe)
7080 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
7081 DAG.getValueType(MVT::i1));
7082 else if (KnownZero == 0xffffff00)
7083 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
7084 DAG.getValueType(MVT::i8));
7085 else if (KnownZero == 0xffff0000)
7086 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
7087 DAG.getValueType(MVT::i16));
7093 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
7094 DAGCombinerInfo &DCI) const {
7095 switch (N->getOpcode()) {
7097 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
7098 case ISD::SUB: return PerformSUBCombine(N, DCI);
7099 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
7100 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
7101 case ISD::AND: return PerformANDCombine(N, DCI);
7102 case ARMISD::BFI: return PerformBFICombine(N, DCI);
7103 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
7104 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
7105 case ISD::STORE: return PerformSTORECombine(N, DCI);
7106 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
7107 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
7108 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
7109 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
7110 case ISD::FP_TO_SINT:
7111 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
7112 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
7113 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
7116 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
7117 case ISD::SIGN_EXTEND:
7118 case ISD::ZERO_EXTEND:
7119 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
7120 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
7121 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
7122 case ARMISD::VLD2DUP:
7123 case ARMISD::VLD3DUP:
7124 case ARMISD::VLD4DUP:
7125 return CombineBaseUpdate(N, DCI);
7126 case ISD::INTRINSIC_VOID:
7127 case ISD::INTRINSIC_W_CHAIN:
7128 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
7129 case Intrinsic::arm_neon_vld1:
7130 case Intrinsic::arm_neon_vld2:
7131 case Intrinsic::arm_neon_vld3:
7132 case Intrinsic::arm_neon_vld4:
7133 case Intrinsic::arm_neon_vld2lane:
7134 case Intrinsic::arm_neon_vld3lane:
7135 case Intrinsic::arm_neon_vld4lane:
7136 case Intrinsic::arm_neon_vst1:
7137 case Intrinsic::arm_neon_vst2:
7138 case Intrinsic::arm_neon_vst3:
7139 case Intrinsic::arm_neon_vst4:
7140 case Intrinsic::arm_neon_vst2lane:
7141 case Intrinsic::arm_neon_vst3lane:
7142 case Intrinsic::arm_neon_vst4lane:
7143 return CombineBaseUpdate(N, DCI);
7151 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
7153 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
7156 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
7157 if (!Subtarget->allowsUnalignedMem())
7160 switch (VT.getSimpleVT().SimpleTy) {
7167 // FIXME: VLD1 etc with standard alignment is legal.
7171 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
7176 switch (VT.getSimpleVT().SimpleTy) {
7177 default: return false;
7192 if ((V & (Scale - 1)) != 0)
7195 return V == (V & ((1LL << 5) - 1));
7198 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
7199 const ARMSubtarget *Subtarget) {
7206 switch (VT.getSimpleVT().SimpleTy) {
7207 default: return false;
7212 // + imm12 or - imm8
7214 return V == (V & ((1LL << 8) - 1));
7215 return V == (V & ((1LL << 12) - 1));
7218 // Same as ARM mode. FIXME: NEON?
7219 if (!Subtarget->hasVFP2())
7224 return V == (V & ((1LL << 8) - 1));
7228 /// isLegalAddressImmediate - Return true if the integer value can be used
7229 /// as the offset of the target addressing mode for load / store of the
7231 static bool isLegalAddressImmediate(int64_t V, EVT VT,
7232 const ARMSubtarget *Subtarget) {
7239 if (Subtarget->isThumb1Only())
7240 return isLegalT1AddressImmediate(V, VT);
7241 else if (Subtarget->isThumb2())
7242 return isLegalT2AddressImmediate(V, VT, Subtarget);
7247 switch (VT.getSimpleVT().SimpleTy) {
7248 default: return false;
7253 return V == (V & ((1LL << 12) - 1));
7256 return V == (V & ((1LL << 8) - 1));
7259 if (!Subtarget->hasVFP2()) // FIXME: NEON?
7264 return V == (V & ((1LL << 8) - 1));
7268 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
7270 int Scale = AM.Scale;
7274 switch (VT.getSimpleVT().SimpleTy) {
7275 default: return false;
7284 return Scale == 2 || Scale == 4 || Scale == 8;
7287 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
7291 // Note, we allow "void" uses (basically, uses that aren't loads or
7292 // stores), because arm allows folding a scale into many arithmetic
7293 // operations. This should be made more precise and revisited later.
7295 // Allow r << imm, but the imm has to be a multiple of two.
7296 if (Scale & 1) return false;
7297 return isPowerOf2_32(Scale);
7301 /// isLegalAddressingMode - Return true if the addressing mode represented
7302 /// by AM is legal for this target, for a load/store of the specified type.
7303 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
7305 EVT VT = getValueType(Ty, true);
7306 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
7309 // Can never fold addr of global into load/store.
7314 case 0: // no scale reg, must be "r+i" or "r", or "i".
7317 if (Subtarget->isThumb1Only())
7321 // ARM doesn't support any R+R*scale+imm addr modes.
7328 if (Subtarget->isThumb2())
7329 return isLegalT2ScaledAddressingMode(AM, VT);
7331 int Scale = AM.Scale;
7332 switch (VT.getSimpleVT().SimpleTy) {
7333 default: return false;
7337 if (Scale < 0) Scale = -Scale;
7341 return isPowerOf2_32(Scale & ~1);
7345 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
7350 // Note, we allow "void" uses (basically, uses that aren't loads or
7351 // stores), because arm allows folding a scale into many arithmetic
7352 // operations. This should be made more precise and revisited later.
7354 // Allow r << imm, but the imm has to be a multiple of two.
7355 if (Scale & 1) return false;
7356 return isPowerOf2_32(Scale);
7363 /// isLegalICmpImmediate - Return true if the specified immediate is legal
7364 /// icmp immediate, that is the target has icmp instructions which can compare
7365 /// a register against the immediate without having to materialize the
7366 /// immediate into a register.
7367 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
7368 if (!Subtarget->isThumb())
7369 return ARM_AM::getSOImmVal(Imm) != -1;
7370 if (Subtarget->isThumb2())
7371 return ARM_AM::getT2SOImmVal(Imm) != -1;
7372 return Imm >= 0 && Imm <= 255;
7375 /// isLegalAddImmediate - Return true if the specified immediate is legal
7376 /// add immediate, that is the target has add instructions which can add
7377 /// a register with the immediate without having to materialize the
7378 /// immediate into a register.
7379 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
7380 return ARM_AM::getSOImmVal(Imm) != -1;
7383 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
7384 bool isSEXTLoad, SDValue &Base,
7385 SDValue &Offset, bool &isInc,
7386 SelectionDAG &DAG) {
7387 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
7390 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
7392 Base = Ptr->getOperand(0);
7393 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7394 int RHSC = (int)RHS->getZExtValue();
7395 if (RHSC < 0 && RHSC > -256) {
7396 assert(Ptr->getOpcode() == ISD::ADD);
7398 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7402 isInc = (Ptr->getOpcode() == ISD::ADD);
7403 Offset = Ptr->getOperand(1);
7405 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
7407 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7408 int RHSC = (int)RHS->getZExtValue();
7409 if (RHSC < 0 && RHSC > -0x1000) {
7410 assert(Ptr->getOpcode() == ISD::ADD);
7412 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7413 Base = Ptr->getOperand(0);
7418 if (Ptr->getOpcode() == ISD::ADD) {
7420 ARM_AM::ShiftOpc ShOpcVal=
7421 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
7422 if (ShOpcVal != ARM_AM::no_shift) {
7423 Base = Ptr->getOperand(1);
7424 Offset = Ptr->getOperand(0);
7426 Base = Ptr->getOperand(0);
7427 Offset = Ptr->getOperand(1);
7432 isInc = (Ptr->getOpcode() == ISD::ADD);
7433 Base = Ptr->getOperand(0);
7434 Offset = Ptr->getOperand(1);
7438 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
7442 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
7443 bool isSEXTLoad, SDValue &Base,
7444 SDValue &Offset, bool &isInc,
7445 SelectionDAG &DAG) {
7446 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
7449 Base = Ptr->getOperand(0);
7450 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7451 int RHSC = (int)RHS->getZExtValue();
7452 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
7453 assert(Ptr->getOpcode() == ISD::ADD);
7455 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7457 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
7458 isInc = Ptr->getOpcode() == ISD::ADD;
7459 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
7467 /// getPreIndexedAddressParts - returns true by value, base pointer and
7468 /// offset pointer and addressing mode by reference if the node's address
7469 /// can be legally represented as pre-indexed load / store address.
7471 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7473 ISD::MemIndexedMode &AM,
7474 SelectionDAG &DAG) const {
7475 if (Subtarget->isThumb1Only())
7480 bool isSEXTLoad = false;
7481 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7482 Ptr = LD->getBasePtr();
7483 VT = LD->getMemoryVT();
7484 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7485 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7486 Ptr = ST->getBasePtr();
7487 VT = ST->getMemoryVT();
7492 bool isLegal = false;
7493 if (Subtarget->isThumb2())
7494 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7495 Offset, isInc, DAG);
7497 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7498 Offset, isInc, DAG);
7502 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
7506 /// getPostIndexedAddressParts - returns true by value, base pointer and
7507 /// offset pointer and addressing mode by reference if this node can be
7508 /// combined with a load / store to form a post-indexed load / store.
7509 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7512 ISD::MemIndexedMode &AM,
7513 SelectionDAG &DAG) const {
7514 if (Subtarget->isThumb1Only())
7519 bool isSEXTLoad = false;
7520 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7521 VT = LD->getMemoryVT();
7522 Ptr = LD->getBasePtr();
7523 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7524 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7525 VT = ST->getMemoryVT();
7526 Ptr = ST->getBasePtr();
7531 bool isLegal = false;
7532 if (Subtarget->isThumb2())
7533 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7536 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7542 // Swap base ptr and offset to catch more post-index load / store when
7543 // it's legal. In Thumb2 mode, offset must be an immediate.
7544 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
7545 !Subtarget->isThumb2())
7546 std::swap(Base, Offset);
7548 // Post-indexed load / store update the base pointer.
7553 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
7557 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7561 const SelectionDAG &DAG,
7562 unsigned Depth) const {
7563 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
7564 switch (Op.getOpcode()) {
7566 case ARMISD::CMOV: {
7567 // Bits are known zero/one if known on the LHS and RHS.
7568 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
7569 if (KnownZero == 0 && KnownOne == 0) return;
7571 APInt KnownZeroRHS, KnownOneRHS;
7572 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
7573 KnownZeroRHS, KnownOneRHS, Depth+1);
7574 KnownZero &= KnownZeroRHS;
7575 KnownOne &= KnownOneRHS;
7581 //===----------------------------------------------------------------------===//
7582 // ARM Inline Assembly Support
7583 //===----------------------------------------------------------------------===//
7585 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
7586 // Looking for "rev" which is V6+.
7587 if (!Subtarget->hasV6Ops())
7590 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
7591 std::string AsmStr = IA->getAsmString();
7592 SmallVector<StringRef, 4> AsmPieces;
7593 SplitString(AsmStr, AsmPieces, ";\n");
7595 switch (AsmPieces.size()) {
7596 default: return false;
7598 AsmStr = AsmPieces[0];
7600 SplitString(AsmStr, AsmPieces, " \t,");
7603 if (AsmPieces.size() == 3 &&
7604 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
7605 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
7606 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
7607 if (Ty && Ty->getBitWidth() == 32)
7608 return IntrinsicLowering::LowerToByteSwap(CI);
7616 /// getConstraintType - Given a constraint letter, return the type of
7617 /// constraint it is for this target.
7618 ARMTargetLowering::ConstraintType
7619 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
7620 if (Constraint.size() == 1) {
7621 switch (Constraint[0]) {
7623 case 'l': return C_RegisterClass;
7624 case 'w': return C_RegisterClass;
7625 case 'h': return C_RegisterClass;
7626 case 'x': return C_RegisterClass;
7627 case 't': return C_RegisterClass;
7628 case 'j': return C_Other; // Constant for movw.
7629 // An address with a single base register. Due to the way we
7630 // currently handle addresses it is the same as an 'r' memory constraint.
7631 case 'Q': return C_Memory;
7633 } else if (Constraint.size() == 2) {
7634 switch (Constraint[0]) {
7636 // All 'U+' constraints are addresses.
7637 case 'U': return C_Memory;
7640 return TargetLowering::getConstraintType(Constraint);
7643 /// Examine constraint type and operand type and determine a weight value.
7644 /// This object must already have been set up with the operand type
7645 /// and the current alternative constraint selected.
7646 TargetLowering::ConstraintWeight
7647 ARMTargetLowering::getSingleConstraintMatchWeight(
7648 AsmOperandInfo &info, const char *constraint) const {
7649 ConstraintWeight weight = CW_Invalid;
7650 Value *CallOperandVal = info.CallOperandVal;
7651 // If we don't have a value, we can't do a match,
7652 // but allow it at the lowest weight.
7653 if (CallOperandVal == NULL)
7655 Type *type = CallOperandVal->getType();
7656 // Look at the constraint type.
7657 switch (*constraint) {
7659 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
7662 if (type->isIntegerTy()) {
7663 if (Subtarget->isThumb())
7664 weight = CW_SpecificReg;
7666 weight = CW_Register;
7670 if (type->isFloatingPointTy())
7671 weight = CW_Register;
7677 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
7679 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7681 if (Constraint.size() == 1) {
7682 // GCC ARM Constraint Letters
7683 switch (Constraint[0]) {
7684 case 'l': // Low regs or general regs.
7685 if (Subtarget->isThumb())
7686 return RCPair(0U, ARM::tGPRRegisterClass);
7688 return RCPair(0U, ARM::GPRRegisterClass);
7689 case 'h': // High regs or no regs.
7690 if (Subtarget->isThumb())
7691 return RCPair(0U, ARM::hGPRRegisterClass);
7694 return RCPair(0U, ARM::GPRRegisterClass);
7697 return RCPair(0U, ARM::SPRRegisterClass);
7698 if (VT.getSizeInBits() == 64)
7699 return RCPair(0U, ARM::DPRRegisterClass);
7700 if (VT.getSizeInBits() == 128)
7701 return RCPair(0U, ARM::QPRRegisterClass);
7705 return RCPair(0U, ARM::SPR_8RegisterClass);
7706 if (VT.getSizeInBits() == 64)
7707 return RCPair(0U, ARM::DPR_8RegisterClass);
7708 if (VT.getSizeInBits() == 128)
7709 return RCPair(0U, ARM::QPR_8RegisterClass);
7713 return RCPair(0U, ARM::SPRRegisterClass);
7717 if (StringRef("{cc}").equals_lower(Constraint))
7718 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
7720 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7723 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7724 /// vector. If it is invalid, don't add anything to Ops.
7725 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7726 std::string &Constraint,
7727 std::vector<SDValue>&Ops,
7728 SelectionDAG &DAG) const {
7729 SDValue Result(0, 0);
7731 // Currently only support length 1 constraints.
7732 if (Constraint.length() != 1) return;
7734 char ConstraintLetter = Constraint[0];
7735 switch (ConstraintLetter) {
7738 case 'I': case 'J': case 'K': case 'L':
7739 case 'M': case 'N': case 'O':
7740 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
7744 int64_t CVal64 = C->getSExtValue();
7745 int CVal = (int) CVal64;
7746 // None of these constraints allow values larger than 32 bits. Check
7747 // that the value fits in an int.
7751 switch (ConstraintLetter) {
7753 // Constant suitable for movw, must be between 0 and
7755 if (Subtarget->hasV6T2Ops())
7756 if (CVal >= 0 && CVal <= 65535)
7760 if (Subtarget->isThumb1Only()) {
7761 // This must be a constant between 0 and 255, for ADD
7763 if (CVal >= 0 && CVal <= 255)
7765 } else if (Subtarget->isThumb2()) {
7766 // A constant that can be used as an immediate value in a
7767 // data-processing instruction.
7768 if (ARM_AM::getT2SOImmVal(CVal) != -1)
7771 // A constant that can be used as an immediate value in a
7772 // data-processing instruction.
7773 if (ARM_AM::getSOImmVal(CVal) != -1)
7779 if (Subtarget->isThumb()) { // FIXME thumb2
7780 // This must be a constant between -255 and -1, for negated ADD
7781 // immediates. This can be used in GCC with an "n" modifier that
7782 // prints the negated value, for use with SUB instructions. It is
7783 // not useful otherwise but is implemented for compatibility.
7784 if (CVal >= -255 && CVal <= -1)
7787 // This must be a constant between -4095 and 4095. It is not clear
7788 // what this constraint is intended for. Implemented for
7789 // compatibility with GCC.
7790 if (CVal >= -4095 && CVal <= 4095)
7796 if (Subtarget->isThumb1Only()) {
7797 // A 32-bit value where only one byte has a nonzero value. Exclude
7798 // zero to match GCC. This constraint is used by GCC internally for
7799 // constants that can be loaded with a move/shift combination.
7800 // It is not useful otherwise but is implemented for compatibility.
7801 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
7803 } else if (Subtarget->isThumb2()) {
7804 // A constant whose bitwise inverse can be used as an immediate
7805 // value in a data-processing instruction. This can be used in GCC
7806 // with a "B" modifier that prints the inverted value, for use with
7807 // BIC and MVN instructions. It is not useful otherwise but is
7808 // implemented for compatibility.
7809 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
7812 // A constant whose bitwise inverse can be used as an immediate
7813 // value in a data-processing instruction. This can be used in GCC
7814 // with a "B" modifier that prints the inverted value, for use with
7815 // BIC and MVN instructions. It is not useful otherwise but is
7816 // implemented for compatibility.
7817 if (ARM_AM::getSOImmVal(~CVal) != -1)
7823 if (Subtarget->isThumb1Only()) {
7824 // This must be a constant between -7 and 7,
7825 // for 3-operand ADD/SUB immediate instructions.
7826 if (CVal >= -7 && CVal < 7)
7828 } else if (Subtarget->isThumb2()) {
7829 // A constant whose negation can be used as an immediate value in a
7830 // data-processing instruction. This can be used in GCC with an "n"
7831 // modifier that prints the negated value, for use with SUB
7832 // instructions. It is not useful otherwise but is implemented for
7834 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
7837 // A constant whose negation can be used as an immediate value in a
7838 // data-processing instruction. This can be used in GCC with an "n"
7839 // modifier that prints the negated value, for use with SUB
7840 // instructions. It is not useful otherwise but is implemented for
7842 if (ARM_AM::getSOImmVal(-CVal) != -1)
7848 if (Subtarget->isThumb()) { // FIXME thumb2
7849 // This must be a multiple of 4 between 0 and 1020, for
7850 // ADD sp + immediate.
7851 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
7854 // A power of two or a constant between 0 and 32. This is used in
7855 // GCC for the shift amount on shifted register operands, but it is
7856 // useful in general for any shift amounts.
7857 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
7863 if (Subtarget->isThumb()) { // FIXME thumb2
7864 // This must be a constant between 0 and 31, for shift amounts.
7865 if (CVal >= 0 && CVal <= 31)
7871 if (Subtarget->isThumb()) { // FIXME thumb2
7872 // This must be a multiple of 4 between -508 and 508, for
7873 // ADD/SUB sp = sp + immediate.
7874 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
7879 Result = DAG.getTargetConstant(CVal, Op.getValueType());
7883 if (Result.getNode()) {
7884 Ops.push_back(Result);
7887 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
7891 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
7892 // The ARM target isn't yet aware of offsets.
7896 int ARM::getVFPf32Imm(const APFloat &FPImm) {
7897 APInt Imm = FPImm.bitcastToAPInt();
7898 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
7899 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
7900 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
7902 // We can handle 4 bits of mantissa.
7903 // mantissa = (16+UInt(e:f:g:h))/16.
7904 if (Mantissa & 0x7ffff)
7907 if ((Mantissa & 0xf) != Mantissa)
7910 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
7911 if (Exp < -3 || Exp > 4)
7913 Exp = ((Exp+3) & 0x7) ^ 4;
7915 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
7918 int ARM::getVFPf64Imm(const APFloat &FPImm) {
7919 APInt Imm = FPImm.bitcastToAPInt();
7920 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
7921 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
7922 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
7924 // We can handle 4 bits of mantissa.
7925 // mantissa = (16+UInt(e:f:g:h))/16.
7926 if (Mantissa & 0xffffffffffffLL)
7929 if ((Mantissa & 0xf) != Mantissa)
7932 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
7933 if (Exp < -3 || Exp > 4)
7935 Exp = ((Exp+3) & 0x7) ^ 4;
7937 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
7940 bool ARM::isBitFieldInvertedMask(unsigned v) {
7941 if (v == 0xffffffff)
7943 // there can be 1's on either or both "outsides", all the "inside"
7945 unsigned int lsb = 0, msb = 31;
7946 while (v & (1 << msb)) --msb;
7947 while (v & (1 << lsb)) ++lsb;
7948 for (unsigned int i = lsb; i <= msb; ++i) {
7955 /// isFPImmLegal - Returns true if the target can instruction select the
7956 /// specified FP immediate natively. If false, the legalizer will
7957 /// materialize the FP immediate as a load from a constant pool.
7958 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
7959 if (!Subtarget->hasVFP3())
7962 return ARM::getVFPf32Imm(Imm) != -1;
7964 return ARM::getVFPf64Imm(Imm) != -1;
7968 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
7969 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
7970 /// specified in the intrinsic calls.
7971 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
7973 unsigned Intrinsic) const {
7974 switch (Intrinsic) {
7975 case Intrinsic::arm_neon_vld1:
7976 case Intrinsic::arm_neon_vld2:
7977 case Intrinsic::arm_neon_vld3:
7978 case Intrinsic::arm_neon_vld4:
7979 case Intrinsic::arm_neon_vld2lane:
7980 case Intrinsic::arm_neon_vld3lane:
7981 case Intrinsic::arm_neon_vld4lane: {
7982 Info.opc = ISD::INTRINSIC_W_CHAIN;
7983 // Conservatively set memVT to the entire set of vectors loaded.
7984 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
7985 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
7986 Info.ptrVal = I.getArgOperand(0);
7988 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
7989 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
7990 Info.vol = false; // volatile loads with NEON intrinsics not supported
7991 Info.readMem = true;
7992 Info.writeMem = false;
7995 case Intrinsic::arm_neon_vst1:
7996 case Intrinsic::arm_neon_vst2:
7997 case Intrinsic::arm_neon_vst3:
7998 case Intrinsic::arm_neon_vst4:
7999 case Intrinsic::arm_neon_vst2lane:
8000 case Intrinsic::arm_neon_vst3lane:
8001 case Intrinsic::arm_neon_vst4lane: {
8002 Info.opc = ISD::INTRINSIC_VOID;
8003 // Conservatively set memVT to the entire set of vectors stored.
8004 unsigned NumElts = 0;
8005 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
8006 Type *ArgTy = I.getArgOperand(ArgI)->getType();
8007 if (!ArgTy->isVectorTy())
8009 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
8011 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8012 Info.ptrVal = I.getArgOperand(0);
8014 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8015 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8016 Info.vol = false; // volatile stores with NEON intrinsics not supported
8017 Info.readMem = false;
8018 Info.writeMem = true;
8021 case Intrinsic::arm_strexd: {
8022 Info.opc = ISD::INTRINSIC_W_CHAIN;
8023 Info.memVT = MVT::i64;
8024 Info.ptrVal = I.getArgOperand(2);
8028 Info.readMem = false;
8029 Info.writeMem = true;
8032 case Intrinsic::arm_ldrexd: {
8033 Info.opc = ISD::INTRINSIC_W_CHAIN;
8034 Info.memVT = MVT::i64;
8035 Info.ptrVal = I.getArgOperand(0);
8039 Info.readMem = true;
8040 Info.writeMem = false;