1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMAddressingModes.h"
18 #include "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 "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Type.h"
33 #include "llvm/CodeGen/CallingConvLower.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineInstrBuilder.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/CodeGen/PseudoSourceValue.h"
40 #include "llvm/CodeGen/SelectionDAG.h"
41 #include "llvm/MC/MCSectionMachO.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/ADT/VectorExtras.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
52 STATISTIC(NumTailCalls, "Number of tail calls");
54 // This option should go away when tail calls fully work.
56 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
57 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
60 // This option should go away when Machine LICM is smart enough to hoist a
63 EnableARMVDUPsplat("arm-vdup-splat", cl::Hidden,
64 cl::desc("Generate VDUP for integer constant splats (TEMPORARY OPTION)."),
68 EnableARMLongCalls("arm-long-calls", cl::Hidden,
69 cl::desc("Generate calls via indirect call instructions"),
73 ARMInterworking("arm-interworking", cl::Hidden,
74 cl::desc("Enable / disable ARM interworking (for debugging only)"),
78 EnableARMCodePlacement("arm-code-placement", cl::Hidden,
79 cl::desc("Enable code placement pass for ARM"),
82 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
83 CCValAssign::LocInfo &LocInfo,
84 ISD::ArgFlagsTy &ArgFlags,
86 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
87 CCValAssign::LocInfo &LocInfo,
88 ISD::ArgFlagsTy &ArgFlags,
90 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
91 CCValAssign::LocInfo &LocInfo,
92 ISD::ArgFlagsTy &ArgFlags,
94 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
95 CCValAssign::LocInfo &LocInfo,
96 ISD::ArgFlagsTy &ArgFlags,
99 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
100 EVT PromotedBitwiseVT) {
101 if (VT != PromotedLdStVT) {
102 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
103 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
104 PromotedLdStVT.getSimpleVT());
106 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
107 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
108 PromotedLdStVT.getSimpleVT());
111 EVT ElemTy = VT.getVectorElementType();
112 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
113 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
114 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
115 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
116 if (ElemTy != MVT::i32) {
117 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
119 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
122 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
123 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
124 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
125 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
128 if (VT.isInteger()) {
129 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
130 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
131 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
134 // Promote all bit-wise operations.
135 if (VT.isInteger() && VT != PromotedBitwiseVT) {
136 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
137 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
138 PromotedBitwiseVT.getSimpleVT());
139 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
140 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
141 PromotedBitwiseVT.getSimpleVT());
142 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
143 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
144 PromotedBitwiseVT.getSimpleVT());
147 // Neon does not support vector divide/remainder operations.
148 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
149 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
150 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
156 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
157 addRegisterClass(VT, ARM::DPRRegisterClass);
158 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
161 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
162 addRegisterClass(VT, ARM::QPRRegisterClass);
163 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
166 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
167 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
168 return new TargetLoweringObjectFileMachO();
170 return new ARMElfTargetObjectFile();
173 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
174 : TargetLowering(TM, createTLOF(TM)) {
175 Subtarget = &TM.getSubtarget<ARMSubtarget>();
176 RegInfo = TM.getRegisterInfo();
178 if (Subtarget->isTargetDarwin()) {
179 // Uses VFP for Thumb libfuncs if available.
180 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
181 // Single-precision floating-point arithmetic.
182 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
183 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
184 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
185 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
187 // Double-precision floating-point arithmetic.
188 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
189 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
190 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
191 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
193 // Single-precision comparisons.
194 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
195 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
196 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
197 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
198 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
199 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
200 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
201 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
203 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
212 // Double-precision comparisons.
213 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
214 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
215 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
216 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
217 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
218 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
219 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
220 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
222 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
231 // Floating-point to integer conversions.
232 // i64 conversions are done via library routines even when generating VFP
233 // instructions, so use the same ones.
234 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
235 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
236 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
237 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
239 // Conversions between floating types.
240 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
241 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
243 // Integer to floating-point conversions.
244 // i64 conversions are done via library routines even when generating VFP
245 // instructions, so use the same ones.
246 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
247 // e.g., __floatunsidf vs. __floatunssidfvfp.
248 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
249 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
250 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
251 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
255 // These libcalls are not available in 32-bit.
256 setLibcallName(RTLIB::SHL_I128, 0);
257 setLibcallName(RTLIB::SRL_I128, 0);
258 setLibcallName(RTLIB::SRA_I128, 0);
260 // Libcalls should use the AAPCS base standard ABI, even if hard float
261 // is in effect, as per the ARM RTABI specification, section 4.1.2.
262 if (Subtarget->isAAPCS_ABI()) {
263 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
264 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
265 CallingConv::ARM_AAPCS);
269 if (Subtarget->isThumb1Only())
270 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
272 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
273 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
274 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
275 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
277 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
280 if (Subtarget->hasNEON()) {
281 addDRTypeForNEON(MVT::v2f32);
282 addDRTypeForNEON(MVT::v8i8);
283 addDRTypeForNEON(MVT::v4i16);
284 addDRTypeForNEON(MVT::v2i32);
285 addDRTypeForNEON(MVT::v1i64);
287 addQRTypeForNEON(MVT::v4f32);
288 addQRTypeForNEON(MVT::v2f64);
289 addQRTypeForNEON(MVT::v16i8);
290 addQRTypeForNEON(MVT::v8i16);
291 addQRTypeForNEON(MVT::v4i32);
292 addQRTypeForNEON(MVT::v2i64);
294 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
295 // neither Neon nor VFP support any arithmetic operations on it.
296 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
297 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
298 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
299 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
300 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
301 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
302 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
303 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
304 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
305 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
306 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
307 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
308 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
309 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
310 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
311 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
312 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
313 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
314 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
315 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
316 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
317 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
318 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
319 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
321 // Neon does not support some operations on v1i64 and v2i64 types.
322 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
323 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
324 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
325 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
327 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
328 setTargetDAGCombine(ISD::SHL);
329 setTargetDAGCombine(ISD::SRL);
330 setTargetDAGCombine(ISD::SRA);
331 setTargetDAGCombine(ISD::SIGN_EXTEND);
332 setTargetDAGCombine(ISD::ZERO_EXTEND);
333 setTargetDAGCombine(ISD::ANY_EXTEND);
334 setTargetDAGCombine(ISD::SELECT_CC);
337 computeRegisterProperties();
339 // ARM does not have f32 extending load.
340 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
342 // ARM does not have i1 sign extending load.
343 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
345 // ARM supports all 4 flavors of integer indexed load / store.
346 if (!Subtarget->isThumb1Only()) {
347 for (unsigned im = (unsigned)ISD::PRE_INC;
348 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
349 setIndexedLoadAction(im, MVT::i1, Legal);
350 setIndexedLoadAction(im, MVT::i8, Legal);
351 setIndexedLoadAction(im, MVT::i16, Legal);
352 setIndexedLoadAction(im, MVT::i32, Legal);
353 setIndexedStoreAction(im, MVT::i1, Legal);
354 setIndexedStoreAction(im, MVT::i8, Legal);
355 setIndexedStoreAction(im, MVT::i16, Legal);
356 setIndexedStoreAction(im, MVT::i32, Legal);
360 // i64 operation support.
361 if (Subtarget->isThumb1Only()) {
362 setOperationAction(ISD::MUL, MVT::i64, Expand);
363 setOperationAction(ISD::MULHU, MVT::i32, Expand);
364 setOperationAction(ISD::MULHS, MVT::i32, Expand);
365 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
366 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
368 setOperationAction(ISD::MUL, MVT::i64, Expand);
369 setOperationAction(ISD::MULHU, MVT::i32, Expand);
370 if (!Subtarget->hasV6Ops())
371 setOperationAction(ISD::MULHS, MVT::i32, Expand);
373 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
374 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
375 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
376 setOperationAction(ISD::SRL, MVT::i64, Custom);
377 setOperationAction(ISD::SRA, MVT::i64, Custom);
379 // ARM does not have ROTL.
380 setOperationAction(ISD::ROTL, MVT::i32, Expand);
381 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
382 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
383 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
384 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
386 // Only ARMv6 has BSWAP.
387 if (!Subtarget->hasV6Ops())
388 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
390 // These are expanded into libcalls.
391 if (!Subtarget->hasDivide()) {
392 // v7M has a hardware divider
393 setOperationAction(ISD::SDIV, MVT::i32, Expand);
394 setOperationAction(ISD::UDIV, MVT::i32, Expand);
396 setOperationAction(ISD::SREM, MVT::i32, Expand);
397 setOperationAction(ISD::UREM, MVT::i32, Expand);
398 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
399 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
401 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
402 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
403 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
404 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
405 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
407 setOperationAction(ISD::TRAP, MVT::Other, Legal);
409 // Use the default implementation.
410 setOperationAction(ISD::VASTART, MVT::Other, Custom);
411 setOperationAction(ISD::VAARG, MVT::Other, Expand);
412 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
413 setOperationAction(ISD::VAEND, MVT::Other, Expand);
414 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
415 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
416 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
417 // FIXME: Shouldn't need this, since no register is used, but the legalizer
418 // doesn't yet know how to not do that for SjLj.
419 setExceptionSelectorRegister(ARM::R0);
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
421 // Handle atomics directly for ARMv[67] (except for Thumb1), otherwise
422 // use the default expansion.
423 bool canHandleAtomics =
424 (Subtarget->hasV7Ops() ||
425 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only()));
426 if (canHandleAtomics) {
427 // membarrier needs custom lowering; the rest are legal and handled
429 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
431 // Set them all for expansion, which will force libcalls.
432 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
433 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
434 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
435 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
436 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
437 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
438 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
449 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
450 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
451 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
452 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
453 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
454 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
455 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
456 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
457 // Since the libcalls include locking, fold in the fences
458 setShouldFoldAtomicFences(true);
460 // 64-bit versions are always libcalls (for now)
461 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
462 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
463 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
464 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
465 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
466 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
467 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
468 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
470 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
471 if (!Subtarget->hasV6Ops()) {
472 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
473 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
475 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
477 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
478 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
479 // iff target supports vfp2.
480 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
482 // We want to custom lower some of our intrinsics.
483 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
484 if (Subtarget->isTargetDarwin()) {
485 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
486 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
489 setOperationAction(ISD::SETCC, MVT::i32, Expand);
490 setOperationAction(ISD::SETCC, MVT::f32, Expand);
491 setOperationAction(ISD::SETCC, MVT::f64, Expand);
492 setOperationAction(ISD::SELECT, MVT::i32, Expand);
493 setOperationAction(ISD::SELECT, MVT::f32, Expand);
494 setOperationAction(ISD::SELECT, MVT::f64, Expand);
495 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
496 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
497 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
499 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
500 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
501 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
502 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
503 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
505 // We don't support sin/cos/fmod/copysign/pow
506 setOperationAction(ISD::FSIN, MVT::f64, Expand);
507 setOperationAction(ISD::FSIN, MVT::f32, Expand);
508 setOperationAction(ISD::FCOS, MVT::f32, Expand);
509 setOperationAction(ISD::FCOS, MVT::f64, Expand);
510 setOperationAction(ISD::FREM, MVT::f64, Expand);
511 setOperationAction(ISD::FREM, MVT::f32, Expand);
512 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
513 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
514 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
516 setOperationAction(ISD::FPOW, MVT::f64, Expand);
517 setOperationAction(ISD::FPOW, MVT::f32, Expand);
519 // Various VFP goodness
520 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
521 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
522 if (Subtarget->hasVFP2()) {
523 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
524 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
525 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
526 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
528 // Special handling for half-precision FP.
529 if (!Subtarget->hasFP16()) {
530 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
531 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
535 // We have target-specific dag combine patterns for the following nodes:
536 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
537 setTargetDAGCombine(ISD::ADD);
538 setTargetDAGCombine(ISD::SUB);
539 setTargetDAGCombine(ISD::MUL);
541 if (Subtarget->hasV6T2Ops())
542 setTargetDAGCombine(ISD::OR);
544 setStackPointerRegisterToSaveRestore(ARM::SP);
546 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
547 setSchedulingPreference(Sched::RegPressure);
549 setSchedulingPreference(Sched::Hybrid);
551 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
553 // On ARM arguments smaller than 4 bytes are extended, so all arguments
554 // are at least 4 bytes aligned.
555 setMinStackArgumentAlignment(4);
557 if (EnableARMCodePlacement)
558 benefitFromCodePlacementOpt = true;
561 std::pair<const TargetRegisterClass*, uint8_t>
562 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
563 const TargetRegisterClass *RRC = 0;
565 switch (VT.getSimpleVT().SimpleTy) {
567 return TargetLowering::findRepresentativeClass(VT);
568 // Use DPR as representative register class for all floating point
569 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
570 // the cost is 1 for both f32 and f64.
571 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
572 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
573 RRC = ARM::DPRRegisterClass;
575 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
576 case MVT::v4f32: case MVT::v2f64:
577 RRC = ARM::DPRRegisterClass;
581 RRC = ARM::DPRRegisterClass;
585 RRC = ARM::DPRRegisterClass;
589 return std::make_pair(RRC, Cost);
592 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
595 case ARMISD::Wrapper: return "ARMISD::Wrapper";
596 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
597 case ARMISD::CALL: return "ARMISD::CALL";
598 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
599 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
600 case ARMISD::tCALL: return "ARMISD::tCALL";
601 case ARMISD::BRCOND: return "ARMISD::BRCOND";
602 case ARMISD::BR_JT: return "ARMISD::BR_JT";
603 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
604 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
605 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
606 case ARMISD::CMP: return "ARMISD::CMP";
607 case ARMISD::CMPZ: return "ARMISD::CMPZ";
608 case ARMISD::CMPFP: return "ARMISD::CMPFP";
609 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
610 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
611 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
612 case ARMISD::CMOV: return "ARMISD::CMOV";
613 case ARMISD::CNEG: return "ARMISD::CNEG";
615 case ARMISD::RBIT: return "ARMISD::RBIT";
617 case ARMISD::FTOSI: return "ARMISD::FTOSI";
618 case ARMISD::FTOUI: return "ARMISD::FTOUI";
619 case ARMISD::SITOF: return "ARMISD::SITOF";
620 case ARMISD::UITOF: return "ARMISD::UITOF";
622 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
623 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
624 case ARMISD::RRX: return "ARMISD::RRX";
626 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
627 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
629 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
630 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
632 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
634 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
636 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
638 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
639 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
641 case ARMISD::VCEQ: return "ARMISD::VCEQ";
642 case ARMISD::VCGE: return "ARMISD::VCGE";
643 case ARMISD::VCGEU: return "ARMISD::VCGEU";
644 case ARMISD::VCGT: return "ARMISD::VCGT";
645 case ARMISD::VCGTU: return "ARMISD::VCGTU";
646 case ARMISD::VTST: return "ARMISD::VTST";
648 case ARMISD::VSHL: return "ARMISD::VSHL";
649 case ARMISD::VSHRs: return "ARMISD::VSHRs";
650 case ARMISD::VSHRu: return "ARMISD::VSHRu";
651 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
652 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
653 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
654 case ARMISD::VSHRN: return "ARMISD::VSHRN";
655 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
656 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
657 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
658 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
659 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
660 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
661 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
662 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
663 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
664 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
665 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
666 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
667 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
668 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
669 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
670 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
671 case ARMISD::VDUP: return "ARMISD::VDUP";
672 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
673 case ARMISD::VEXT: return "ARMISD::VEXT";
674 case ARMISD::VREV64: return "ARMISD::VREV64";
675 case ARMISD::VREV32: return "ARMISD::VREV32";
676 case ARMISD::VREV16: return "ARMISD::VREV16";
677 case ARMISD::VZIP: return "ARMISD::VZIP";
678 case ARMISD::VUZP: return "ARMISD::VUZP";
679 case ARMISD::VTRN: return "ARMISD::VTRN";
680 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
681 case ARMISD::FMAX: return "ARMISD::FMAX";
682 case ARMISD::FMIN: return "ARMISD::FMIN";
683 case ARMISD::BFI: return "ARMISD::BFI";
687 /// getRegClassFor - Return the register class that should be used for the
688 /// specified value type.
689 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
690 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
691 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
692 // load / store 4 to 8 consecutive D registers.
693 if (Subtarget->hasNEON()) {
694 if (VT == MVT::v4i64)
695 return ARM::QQPRRegisterClass;
696 else if (VT == MVT::v8i64)
697 return ARM::QQQQPRRegisterClass;
699 return TargetLowering::getRegClassFor(VT);
702 // Create a fast isel object.
704 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
705 return ARM::createFastISel(funcInfo);
708 /// getFunctionAlignment - Return the Log2 alignment of this function.
709 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
710 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
713 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
714 /// be used for loads / stores from the global.
715 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
716 return (Subtarget->isThumb1Only() ? 127 : 4095);
719 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
720 unsigned NumVals = N->getNumValues();
722 return Sched::RegPressure;
724 for (unsigned i = 0; i != NumVals; ++i) {
725 EVT VT = N->getValueType(i);
726 if (VT.isFloatingPoint() || VT.isVector())
727 return Sched::Latency;
730 if (!N->isMachineOpcode())
731 return Sched::RegPressure;
733 // Load are scheduled for latency even if there instruction itinerary
735 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
736 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
738 return Sched::Latency;
740 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
741 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
742 return Sched::Latency;
743 return Sched::RegPressure;
747 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
748 MachineFunction &MF) const {
749 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
750 switch (RC->getID()) {
753 case ARM::tGPRRegClassID:
755 case ARM::GPRRegClassID:
756 return 10 - FPDiff - (Subtarget->isR9Reserved() ? 1 : 0);
757 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
758 case ARM::DPRRegClassID:
763 //===----------------------------------------------------------------------===//
765 //===----------------------------------------------------------------------===//
767 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
768 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
770 default: llvm_unreachable("Unknown condition code!");
771 case ISD::SETNE: return ARMCC::NE;
772 case ISD::SETEQ: return ARMCC::EQ;
773 case ISD::SETGT: return ARMCC::GT;
774 case ISD::SETGE: return ARMCC::GE;
775 case ISD::SETLT: return ARMCC::LT;
776 case ISD::SETLE: return ARMCC::LE;
777 case ISD::SETUGT: return ARMCC::HI;
778 case ISD::SETUGE: return ARMCC::HS;
779 case ISD::SETULT: return ARMCC::LO;
780 case ISD::SETULE: return ARMCC::LS;
784 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
785 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
786 ARMCC::CondCodes &CondCode2) {
787 CondCode2 = ARMCC::AL;
789 default: llvm_unreachable("Unknown FP condition!");
791 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
793 case ISD::SETOGT: CondCode = ARMCC::GT; break;
795 case ISD::SETOGE: CondCode = ARMCC::GE; break;
796 case ISD::SETOLT: CondCode = ARMCC::MI; break;
797 case ISD::SETOLE: CondCode = ARMCC::LS; break;
798 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
799 case ISD::SETO: CondCode = ARMCC::VC; break;
800 case ISD::SETUO: CondCode = ARMCC::VS; break;
801 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
802 case ISD::SETUGT: CondCode = ARMCC::HI; break;
803 case ISD::SETUGE: CondCode = ARMCC::PL; break;
805 case ISD::SETULT: CondCode = ARMCC::LT; break;
807 case ISD::SETULE: CondCode = ARMCC::LE; break;
809 case ISD::SETUNE: CondCode = ARMCC::NE; break;
813 //===----------------------------------------------------------------------===//
814 // Calling Convention Implementation
815 //===----------------------------------------------------------------------===//
817 #include "ARMGenCallingConv.inc"
819 // APCS f64 is in register pairs, possibly split to stack
820 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
821 CCValAssign::LocInfo &LocInfo,
822 CCState &State, bool CanFail) {
823 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
825 // Try to get the first register.
826 if (unsigned Reg = State.AllocateReg(RegList, 4))
827 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
829 // For the 2nd half of a v2f64, do not fail.
833 // Put the whole thing on the stack.
834 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
835 State.AllocateStack(8, 4),
840 // Try to get the second register.
841 if (unsigned Reg = State.AllocateReg(RegList, 4))
842 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
844 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
845 State.AllocateStack(4, 4),
850 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
851 CCValAssign::LocInfo &LocInfo,
852 ISD::ArgFlagsTy &ArgFlags,
854 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
856 if (LocVT == MVT::v2f64 &&
857 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
859 return true; // we handled it
862 // AAPCS f64 is in aligned register pairs
863 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
864 CCValAssign::LocInfo &LocInfo,
865 CCState &State, bool CanFail) {
866 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
867 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
868 static const unsigned ShadowRegList[] = { ARM::R0, ARM::R1 };
870 unsigned Reg = State.AllocateReg(HiRegList, ShadowRegList, 2);
872 // For the 2nd half of a v2f64, do not just fail.
876 // Put the whole thing on the stack.
877 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
878 State.AllocateStack(8, 8),
884 for (i = 0; i < 2; ++i)
885 if (HiRegList[i] == Reg)
888 unsigned T = State.AllocateReg(LoRegList[i]);
890 assert(T == LoRegList[i] && "Could not allocate register");
892 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
893 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
898 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
899 CCValAssign::LocInfo &LocInfo,
900 ISD::ArgFlagsTy &ArgFlags,
902 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
904 if (LocVT == MVT::v2f64 &&
905 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
907 return true; // we handled it
910 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
911 CCValAssign::LocInfo &LocInfo, CCState &State) {
912 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
913 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
915 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
917 return false; // we didn't handle it
920 for (i = 0; i < 2; ++i)
921 if (HiRegList[i] == Reg)
924 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
925 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
930 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
931 CCValAssign::LocInfo &LocInfo,
932 ISD::ArgFlagsTy &ArgFlags,
934 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
936 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
938 return true; // we handled it
941 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
942 CCValAssign::LocInfo &LocInfo,
943 ISD::ArgFlagsTy &ArgFlags,
945 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
949 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
950 /// given CallingConvention value.
951 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
953 bool isVarArg) const {
956 llvm_unreachable("Unsupported calling convention");
958 case CallingConv::Fast:
959 // Use target triple & subtarget features to do actual dispatch.
960 if (Subtarget->isAAPCS_ABI()) {
961 if (Subtarget->hasVFP2() &&
962 FloatABIType == FloatABI::Hard && !isVarArg)
963 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
965 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
967 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
968 case CallingConv::ARM_AAPCS_VFP:
969 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
970 case CallingConv::ARM_AAPCS:
971 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
972 case CallingConv::ARM_APCS:
973 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
977 /// LowerCallResult - Lower the result values of a call into the
978 /// appropriate copies out of appropriate physical registers.
980 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
981 CallingConv::ID CallConv, bool isVarArg,
982 const SmallVectorImpl<ISD::InputArg> &Ins,
983 DebugLoc dl, SelectionDAG &DAG,
984 SmallVectorImpl<SDValue> &InVals) const {
986 // Assign locations to each value returned by this call.
987 SmallVector<CCValAssign, 16> RVLocs;
988 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
989 RVLocs, *DAG.getContext());
990 CCInfo.AnalyzeCallResult(Ins,
991 CCAssignFnForNode(CallConv, /* Return*/ true,
994 // Copy all of the result registers out of their specified physreg.
995 for (unsigned i = 0; i != RVLocs.size(); ++i) {
996 CCValAssign VA = RVLocs[i];
999 if (VA.needsCustom()) {
1000 // Handle f64 or half of a v2f64.
1001 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1003 Chain = Lo.getValue(1);
1004 InFlag = Lo.getValue(2);
1005 VA = RVLocs[++i]; // skip ahead to next loc
1006 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1008 Chain = Hi.getValue(1);
1009 InFlag = Hi.getValue(2);
1010 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1012 if (VA.getLocVT() == MVT::v2f64) {
1013 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1014 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1015 DAG.getConstant(0, MVT::i32));
1017 VA = RVLocs[++i]; // skip ahead to next loc
1018 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1019 Chain = Lo.getValue(1);
1020 InFlag = Lo.getValue(2);
1021 VA = RVLocs[++i]; // skip ahead to next loc
1022 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1023 Chain = Hi.getValue(1);
1024 InFlag = Hi.getValue(2);
1025 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1026 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1027 DAG.getConstant(1, MVT::i32));
1030 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1032 Chain = Val.getValue(1);
1033 InFlag = Val.getValue(2);
1036 switch (VA.getLocInfo()) {
1037 default: llvm_unreachable("Unknown loc info!");
1038 case CCValAssign::Full: break;
1039 case CCValAssign::BCvt:
1040 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1044 InVals.push_back(Val);
1050 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1051 /// by "Src" to address "Dst" of size "Size". Alignment information is
1052 /// specified by the specific parameter attribute. The copy will be passed as
1053 /// a byval function parameter.
1054 /// Sometimes what we are copying is the end of a larger object, the part that
1055 /// does not fit in registers.
1057 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1058 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1060 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1061 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1062 /*isVolatile=*/false, /*AlwaysInline=*/false,
1066 /// LowerMemOpCallTo - Store the argument to the stack.
1068 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1069 SDValue StackPtr, SDValue Arg,
1070 DebugLoc dl, SelectionDAG &DAG,
1071 const CCValAssign &VA,
1072 ISD::ArgFlagsTy Flags) const {
1073 unsigned LocMemOffset = VA.getLocMemOffset();
1074 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1075 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1076 if (Flags.isByVal()) {
1077 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1079 return DAG.getStore(Chain, dl, Arg, PtrOff,
1080 PseudoSourceValue::getStack(), LocMemOffset,
1084 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1085 SDValue Chain, SDValue &Arg,
1086 RegsToPassVector &RegsToPass,
1087 CCValAssign &VA, CCValAssign &NextVA,
1089 SmallVector<SDValue, 8> &MemOpChains,
1090 ISD::ArgFlagsTy Flags) const {
1092 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1093 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1094 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1096 if (NextVA.isRegLoc())
1097 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1099 assert(NextVA.isMemLoc());
1100 if (StackPtr.getNode() == 0)
1101 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1103 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1109 /// LowerCall - Lowering a call into a callseq_start <-
1110 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1113 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1114 CallingConv::ID CallConv, bool isVarArg,
1116 const SmallVectorImpl<ISD::OutputArg> &Outs,
1117 const SmallVectorImpl<SDValue> &OutVals,
1118 const SmallVectorImpl<ISD::InputArg> &Ins,
1119 DebugLoc dl, SelectionDAG &DAG,
1120 SmallVectorImpl<SDValue> &InVals) const {
1121 MachineFunction &MF = DAG.getMachineFunction();
1122 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1123 bool IsSibCall = false;
1124 // Temporarily disable tail calls so things don't break.
1125 if (!EnableARMTailCalls)
1128 // Check if it's really possible to do a tail call.
1129 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1130 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1131 Outs, OutVals, Ins, DAG);
1132 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1133 // detected sibcalls.
1140 // Analyze operands of the call, assigning locations to each operand.
1141 SmallVector<CCValAssign, 16> ArgLocs;
1142 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1144 CCInfo.AnalyzeCallOperands(Outs,
1145 CCAssignFnForNode(CallConv, /* Return*/ false,
1148 // Get a count of how many bytes are to be pushed on the stack.
1149 unsigned NumBytes = CCInfo.getNextStackOffset();
1151 // For tail calls, memory operands are available in our caller's stack.
1155 // Adjust the stack pointer for the new arguments...
1156 // These operations are automatically eliminated by the prolog/epilog pass
1158 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1160 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1162 RegsToPassVector RegsToPass;
1163 SmallVector<SDValue, 8> MemOpChains;
1165 // Walk the register/memloc assignments, inserting copies/loads. In the case
1166 // of tail call optimization, arguments are handled later.
1167 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1169 ++i, ++realArgIdx) {
1170 CCValAssign &VA = ArgLocs[i];
1171 SDValue Arg = OutVals[realArgIdx];
1172 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1174 // Promote the value if needed.
1175 switch (VA.getLocInfo()) {
1176 default: llvm_unreachable("Unknown loc info!");
1177 case CCValAssign::Full: break;
1178 case CCValAssign::SExt:
1179 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1181 case CCValAssign::ZExt:
1182 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1184 case CCValAssign::AExt:
1185 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1187 case CCValAssign::BCvt:
1188 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1192 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1193 if (VA.needsCustom()) {
1194 if (VA.getLocVT() == MVT::v2f64) {
1195 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1196 DAG.getConstant(0, MVT::i32));
1197 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1198 DAG.getConstant(1, MVT::i32));
1200 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1201 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1203 VA = ArgLocs[++i]; // skip ahead to next loc
1204 if (VA.isRegLoc()) {
1205 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1206 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1208 assert(VA.isMemLoc());
1210 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1211 dl, DAG, VA, Flags));
1214 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1215 StackPtr, MemOpChains, Flags);
1217 } else if (VA.isRegLoc()) {
1218 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1219 } else if (!IsSibCall) {
1220 assert(VA.isMemLoc());
1222 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1223 dl, DAG, VA, Flags));
1227 if (!MemOpChains.empty())
1228 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1229 &MemOpChains[0], MemOpChains.size());
1231 // Build a sequence of copy-to-reg nodes chained together with token chain
1232 // and flag operands which copy the outgoing args into the appropriate regs.
1234 // Tail call byval lowering might overwrite argument registers so in case of
1235 // tail call optimization the copies to registers are lowered later.
1237 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1238 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1239 RegsToPass[i].second, InFlag);
1240 InFlag = Chain.getValue(1);
1243 // For tail calls lower the arguments to the 'real' stack slot.
1245 // Force all the incoming stack arguments to be loaded from the stack
1246 // before any new outgoing arguments are stored to the stack, because the
1247 // outgoing stack slots may alias the incoming argument stack slots, and
1248 // the alias isn't otherwise explicit. This is slightly more conservative
1249 // than necessary, because it means that each store effectively depends
1250 // on every argument instead of just those arguments it would clobber.
1252 // Do not flag preceeding copytoreg stuff together with the following stuff.
1254 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1255 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1256 RegsToPass[i].second, InFlag);
1257 InFlag = Chain.getValue(1);
1262 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1263 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1264 // node so that legalize doesn't hack it.
1265 bool isDirect = false;
1266 bool isARMFunc = false;
1267 bool isLocalARMFunc = false;
1268 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1270 if (EnableARMLongCalls) {
1271 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1272 && "long-calls with non-static relocation model!");
1273 // Handle a global address or an external symbol. If it's not one of
1274 // those, the target's already in a register, so we don't need to do
1276 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1277 const GlobalValue *GV = G->getGlobal();
1278 // Create a constant pool entry for the callee address
1279 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1280 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1283 // Get the address of the callee into a register
1284 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1285 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1286 Callee = DAG.getLoad(getPointerTy(), dl,
1287 DAG.getEntryNode(), CPAddr,
1288 PseudoSourceValue::getConstantPool(), 0,
1290 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1291 const char *Sym = S->getSymbol();
1293 // Create a constant pool entry for the callee address
1294 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1295 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1296 Sym, ARMPCLabelIndex, 0);
1297 // Get the address of the callee into a register
1298 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1299 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1300 Callee = DAG.getLoad(getPointerTy(), dl,
1301 DAG.getEntryNode(), CPAddr,
1302 PseudoSourceValue::getConstantPool(), 0,
1305 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1306 const GlobalValue *GV = G->getGlobal();
1308 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1309 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1310 getTargetMachine().getRelocationModel() != Reloc::Static;
1311 isARMFunc = !Subtarget->isThumb() || isStub;
1312 // ARM call to a local ARM function is predicable.
1313 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1314 // tBX takes a register source operand.
1315 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1316 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1317 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1320 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1321 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1322 Callee = DAG.getLoad(getPointerTy(), dl,
1323 DAG.getEntryNode(), CPAddr,
1324 PseudoSourceValue::getConstantPool(), 0,
1326 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1327 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1328 getPointerTy(), Callee, PICLabel);
1330 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1331 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1333 bool isStub = Subtarget->isTargetDarwin() &&
1334 getTargetMachine().getRelocationModel() != Reloc::Static;
1335 isARMFunc = !Subtarget->isThumb() || isStub;
1336 // tBX takes a register source operand.
1337 const char *Sym = S->getSymbol();
1338 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1339 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1340 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1341 Sym, ARMPCLabelIndex, 4);
1342 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1343 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1344 Callee = DAG.getLoad(getPointerTy(), dl,
1345 DAG.getEntryNode(), CPAddr,
1346 PseudoSourceValue::getConstantPool(), 0,
1348 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1349 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1350 getPointerTy(), Callee, PICLabel);
1352 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1355 // FIXME: handle tail calls differently.
1357 if (Subtarget->isThumb()) {
1358 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1359 CallOpc = ARMISD::CALL_NOLINK;
1361 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1363 CallOpc = (isDirect || Subtarget->hasV5TOps())
1364 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1365 : ARMISD::CALL_NOLINK;
1368 std::vector<SDValue> Ops;
1369 Ops.push_back(Chain);
1370 Ops.push_back(Callee);
1372 // Add argument registers to the end of the list so that they are known live
1374 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1375 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1376 RegsToPass[i].second.getValueType()));
1378 if (InFlag.getNode())
1379 Ops.push_back(InFlag);
1381 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1383 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1385 // Returns a chain and a flag for retval copy to use.
1386 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1387 InFlag = Chain.getValue(1);
1389 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1390 DAG.getIntPtrConstant(0, true), InFlag);
1392 InFlag = Chain.getValue(1);
1394 // Handle result values, copying them out of physregs into vregs that we
1396 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1400 /// MatchingStackOffset - Return true if the given stack call argument is
1401 /// already available in the same position (relatively) of the caller's
1402 /// incoming argument stack.
1404 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1405 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1406 const ARMInstrInfo *TII) {
1407 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1409 if (Arg.getOpcode() == ISD::CopyFromReg) {
1410 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1411 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1413 MachineInstr *Def = MRI->getVRegDef(VR);
1416 if (!Flags.isByVal()) {
1417 if (!TII->isLoadFromStackSlot(Def, FI))
1422 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1423 if (Flags.isByVal())
1424 // ByVal argument is passed in as a pointer but it's now being
1425 // dereferenced. e.g.
1426 // define @foo(%struct.X* %A) {
1427 // tail call @bar(%struct.X* byval %A)
1430 SDValue Ptr = Ld->getBasePtr();
1431 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1434 FI = FINode->getIndex();
1438 assert(FI != INT_MAX);
1439 if (!MFI->isFixedObjectIndex(FI))
1441 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1444 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1445 /// for tail call optimization. Targets which want to do tail call
1446 /// optimization should implement this function.
1448 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1449 CallingConv::ID CalleeCC,
1451 bool isCalleeStructRet,
1452 bool isCallerStructRet,
1453 const SmallVectorImpl<ISD::OutputArg> &Outs,
1454 const SmallVectorImpl<SDValue> &OutVals,
1455 const SmallVectorImpl<ISD::InputArg> &Ins,
1456 SelectionDAG& DAG) const {
1457 const Function *CallerF = DAG.getMachineFunction().getFunction();
1458 CallingConv::ID CallerCC = CallerF->getCallingConv();
1459 bool CCMatch = CallerCC == CalleeCC;
1461 // Look for obvious safe cases to perform tail call optimization that do not
1462 // require ABI changes. This is what gcc calls sibcall.
1464 // Do not sibcall optimize vararg calls unless the call site is not passing
1466 if (isVarArg && !Outs.empty())
1469 // Also avoid sibcall optimization if either caller or callee uses struct
1470 // return semantics.
1471 if (isCalleeStructRet || isCallerStructRet)
1474 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1475 // emitEpilogue is not ready for them.
1476 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1477 // LR. This means if we need to reload LR, it takes an extra instructions,
1478 // which outweighs the value of the tail call; but here we don't know yet
1479 // whether LR is going to be used. Probably the right approach is to
1480 // generate the tail call here and turn it back into CALL/RET in
1481 // emitEpilogue if LR is used.
1482 if (Subtarget->isThumb1Only())
1485 // For the moment, we can only do this to functions defined in this
1486 // compilation, or to indirect calls. A Thumb B to an ARM function,
1487 // or vice versa, is not easily fixed up in the linker unlike BL.
1488 // (We could do this by loading the address of the callee into a register;
1489 // that is an extra instruction over the direct call and burns a register
1490 // as well, so is not likely to be a win.)
1492 // It might be safe to remove this restriction on non-Darwin.
1494 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1495 // but we need to make sure there are enough registers; the only valid
1496 // registers are the 4 used for parameters. We don't currently do this
1498 if (isa<ExternalSymbolSDNode>(Callee))
1501 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1502 const GlobalValue *GV = G->getGlobal();
1503 if (GV->isDeclaration() || GV->isWeakForLinker())
1507 // If the calling conventions do not match, then we'd better make sure the
1508 // results are returned in the same way as what the caller expects.
1510 SmallVector<CCValAssign, 16> RVLocs1;
1511 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1512 RVLocs1, *DAG.getContext());
1513 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1515 SmallVector<CCValAssign, 16> RVLocs2;
1516 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1517 RVLocs2, *DAG.getContext());
1518 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1520 if (RVLocs1.size() != RVLocs2.size())
1522 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1523 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1525 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1527 if (RVLocs1[i].isRegLoc()) {
1528 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1531 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1537 // If the callee takes no arguments then go on to check the results of the
1539 if (!Outs.empty()) {
1540 // Check if stack adjustment is needed. For now, do not do this if any
1541 // argument is passed on the stack.
1542 SmallVector<CCValAssign, 16> ArgLocs;
1543 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1544 ArgLocs, *DAG.getContext());
1545 CCInfo.AnalyzeCallOperands(Outs,
1546 CCAssignFnForNode(CalleeCC, false, isVarArg));
1547 if (CCInfo.getNextStackOffset()) {
1548 MachineFunction &MF = DAG.getMachineFunction();
1550 // Check if the arguments are already laid out in the right way as
1551 // the caller's fixed stack objects.
1552 MachineFrameInfo *MFI = MF.getFrameInfo();
1553 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1554 const ARMInstrInfo *TII =
1555 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1556 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1558 ++i, ++realArgIdx) {
1559 CCValAssign &VA = ArgLocs[i];
1560 EVT RegVT = VA.getLocVT();
1561 SDValue Arg = OutVals[realArgIdx];
1562 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1563 if (VA.getLocInfo() == CCValAssign::Indirect)
1565 if (VA.needsCustom()) {
1566 // f64 and vector types are split into multiple registers or
1567 // register/stack-slot combinations. The types will not match
1568 // the registers; give up on memory f64 refs until we figure
1569 // out what to do about this.
1572 if (!ArgLocs[++i].isRegLoc())
1574 if (RegVT == MVT::v2f64) {
1575 if (!ArgLocs[++i].isRegLoc())
1577 if (!ArgLocs[++i].isRegLoc())
1580 } else if (!VA.isRegLoc()) {
1581 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1593 ARMTargetLowering::LowerReturn(SDValue Chain,
1594 CallingConv::ID CallConv, bool isVarArg,
1595 const SmallVectorImpl<ISD::OutputArg> &Outs,
1596 const SmallVectorImpl<SDValue> &OutVals,
1597 DebugLoc dl, SelectionDAG &DAG) const {
1599 // CCValAssign - represent the assignment of the return value to a location.
1600 SmallVector<CCValAssign, 16> RVLocs;
1602 // CCState - Info about the registers and stack slots.
1603 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1606 // Analyze outgoing return values.
1607 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1610 // If this is the first return lowered for this function, add
1611 // the regs to the liveout set for the function.
1612 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1613 for (unsigned i = 0; i != RVLocs.size(); ++i)
1614 if (RVLocs[i].isRegLoc())
1615 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1620 // Copy the result values into the output registers.
1621 for (unsigned i = 0, realRVLocIdx = 0;
1623 ++i, ++realRVLocIdx) {
1624 CCValAssign &VA = RVLocs[i];
1625 assert(VA.isRegLoc() && "Can only return in registers!");
1627 SDValue Arg = OutVals[realRVLocIdx];
1629 switch (VA.getLocInfo()) {
1630 default: llvm_unreachable("Unknown loc info!");
1631 case CCValAssign::Full: break;
1632 case CCValAssign::BCvt:
1633 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1637 if (VA.needsCustom()) {
1638 if (VA.getLocVT() == MVT::v2f64) {
1639 // Extract the first half and return it in two registers.
1640 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1641 DAG.getConstant(0, MVT::i32));
1642 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1643 DAG.getVTList(MVT::i32, MVT::i32), Half);
1645 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1646 Flag = Chain.getValue(1);
1647 VA = RVLocs[++i]; // skip ahead to next loc
1648 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1649 HalfGPRs.getValue(1), Flag);
1650 Flag = Chain.getValue(1);
1651 VA = RVLocs[++i]; // skip ahead to next loc
1653 // Extract the 2nd half and fall through to handle it as an f64 value.
1654 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1655 DAG.getConstant(1, MVT::i32));
1657 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1659 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1660 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1661 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1662 Flag = Chain.getValue(1);
1663 VA = RVLocs[++i]; // skip ahead to next loc
1664 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1667 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1669 // Guarantee that all emitted copies are
1670 // stuck together, avoiding something bad.
1671 Flag = Chain.getValue(1);
1676 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1678 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1683 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1684 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1685 // one of the above mentioned nodes. It has to be wrapped because otherwise
1686 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1687 // be used to form addressing mode. These wrapped nodes will be selected
1689 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1690 EVT PtrVT = Op.getValueType();
1691 // FIXME there is no actual debug info here
1692 DebugLoc dl = Op.getDebugLoc();
1693 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1695 if (CP->isMachineConstantPoolEntry())
1696 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1697 CP->getAlignment());
1699 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1700 CP->getAlignment());
1701 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1704 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1705 return MachineJumpTableInfo::EK_Inline;
1708 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1709 SelectionDAG &DAG) const {
1710 MachineFunction &MF = DAG.getMachineFunction();
1711 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1712 unsigned ARMPCLabelIndex = 0;
1713 DebugLoc DL = Op.getDebugLoc();
1714 EVT PtrVT = getPointerTy();
1715 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1716 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1718 if (RelocM == Reloc::Static) {
1719 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1721 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1722 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1723 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1724 ARMCP::CPBlockAddress,
1726 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1728 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1729 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1730 PseudoSourceValue::getConstantPool(), 0,
1732 if (RelocM == Reloc::Static)
1734 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1735 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1738 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1740 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1741 SelectionDAG &DAG) const {
1742 DebugLoc dl = GA->getDebugLoc();
1743 EVT PtrVT = getPointerTy();
1744 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1745 MachineFunction &MF = DAG.getMachineFunction();
1746 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1747 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1748 ARMConstantPoolValue *CPV =
1749 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1750 ARMCP::CPValue, PCAdj, "tlsgd", true);
1751 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1752 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1753 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1754 PseudoSourceValue::getConstantPool(), 0,
1756 SDValue Chain = Argument.getValue(1);
1758 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1759 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1761 // call __tls_get_addr.
1764 Entry.Node = Argument;
1765 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1766 Args.push_back(Entry);
1767 // FIXME: is there useful debug info available here?
1768 std::pair<SDValue, SDValue> CallResult =
1769 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1770 false, false, false, false,
1771 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1772 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1773 return CallResult.first;
1776 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1777 // "local exec" model.
1779 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1780 SelectionDAG &DAG) const {
1781 const GlobalValue *GV = GA->getGlobal();
1782 DebugLoc dl = GA->getDebugLoc();
1784 SDValue Chain = DAG.getEntryNode();
1785 EVT PtrVT = getPointerTy();
1786 // Get the Thread Pointer
1787 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1789 if (GV->isDeclaration()) {
1790 MachineFunction &MF = DAG.getMachineFunction();
1791 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1792 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1793 // Initial exec model.
1794 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1795 ARMConstantPoolValue *CPV =
1796 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1797 ARMCP::CPValue, PCAdj, "gottpoff", true);
1798 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1799 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1800 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1801 PseudoSourceValue::getConstantPool(), 0,
1803 Chain = Offset.getValue(1);
1805 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1806 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1808 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1809 PseudoSourceValue::getConstantPool(), 0,
1813 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1814 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1815 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1816 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1817 PseudoSourceValue::getConstantPool(), 0,
1821 // The address of the thread local variable is the add of the thread
1822 // pointer with the offset of the variable.
1823 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1827 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1828 // TODO: implement the "local dynamic" model
1829 assert(Subtarget->isTargetELF() &&
1830 "TLS not implemented for non-ELF targets");
1831 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1832 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1833 // otherwise use the "Local Exec" TLS Model
1834 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1835 return LowerToTLSGeneralDynamicModel(GA, DAG);
1837 return LowerToTLSExecModels(GA, DAG);
1840 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1841 SelectionDAG &DAG) const {
1842 EVT PtrVT = getPointerTy();
1843 DebugLoc dl = Op.getDebugLoc();
1844 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1845 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1846 if (RelocM == Reloc::PIC_) {
1847 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1848 ARMConstantPoolValue *CPV =
1849 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1850 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1851 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1852 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1854 PseudoSourceValue::getConstantPool(), 0,
1856 SDValue Chain = Result.getValue(1);
1857 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1858 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1860 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1861 PseudoSourceValue::getGOT(), 0,
1865 // If we have T2 ops, we can materialize the address directly via movt/movw
1866 // pair. This is always cheaper.
1867 if (Subtarget->useMovt()) {
1868 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1869 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1871 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1872 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1873 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1874 PseudoSourceValue::getConstantPool(), 0,
1880 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1881 SelectionDAG &DAG) const {
1882 MachineFunction &MF = DAG.getMachineFunction();
1883 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1884 unsigned ARMPCLabelIndex = 0;
1885 EVT PtrVT = getPointerTy();
1886 DebugLoc dl = Op.getDebugLoc();
1887 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1888 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1890 if (RelocM == Reloc::Static)
1891 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1893 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1894 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1895 ARMConstantPoolValue *CPV =
1896 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1897 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1899 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1901 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1902 PseudoSourceValue::getConstantPool(), 0,
1904 SDValue Chain = Result.getValue(1);
1906 if (RelocM == Reloc::PIC_) {
1907 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1908 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1911 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1912 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1913 PseudoSourceValue::getGOT(), 0,
1919 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1920 SelectionDAG &DAG) const {
1921 assert(Subtarget->isTargetELF() &&
1922 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1923 MachineFunction &MF = DAG.getMachineFunction();
1924 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1925 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1926 EVT PtrVT = getPointerTy();
1927 DebugLoc dl = Op.getDebugLoc();
1928 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1929 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1930 "_GLOBAL_OFFSET_TABLE_",
1931 ARMPCLabelIndex, PCAdj);
1932 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1933 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1934 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1935 PseudoSourceValue::getConstantPool(), 0,
1937 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1938 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1942 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1943 DebugLoc dl = Op.getDebugLoc();
1944 SDValue Val = DAG.getConstant(0, MVT::i32);
1945 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1946 Op.getOperand(1), Val);
1950 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1951 DebugLoc dl = Op.getDebugLoc();
1952 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1953 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1957 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1958 const ARMSubtarget *Subtarget) const {
1959 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1960 DebugLoc dl = Op.getDebugLoc();
1962 default: return SDValue(); // Don't custom lower most intrinsics.
1963 case Intrinsic::arm_thread_pointer: {
1964 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1965 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1967 case Intrinsic::eh_sjlj_lsda: {
1968 MachineFunction &MF = DAG.getMachineFunction();
1969 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1970 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1971 EVT PtrVT = getPointerTy();
1972 DebugLoc dl = Op.getDebugLoc();
1973 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1975 unsigned PCAdj = (RelocM != Reloc::PIC_)
1976 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1977 ARMConstantPoolValue *CPV =
1978 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1979 ARMCP::CPLSDA, PCAdj);
1980 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1981 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1983 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1984 PseudoSourceValue::getConstantPool(), 0,
1987 if (RelocM == Reloc::PIC_) {
1988 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1989 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1996 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1997 const ARMSubtarget *Subtarget) {
1998 DebugLoc dl = Op.getDebugLoc();
1999 SDValue Op5 = Op.getOperand(5);
2000 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
2001 // v6 and v7 can both handle barriers directly, but need handled a bit
2002 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
2004 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
2005 if (Subtarget->hasV7Ops())
2006 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2007 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
2008 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2009 DAG.getConstant(0, MVT::i32));
2010 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2014 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2015 MachineFunction &MF = DAG.getMachineFunction();
2016 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2018 // vastart just stores the address of the VarArgsFrameIndex slot into the
2019 // memory location argument.
2020 DebugLoc dl = Op.getDebugLoc();
2021 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2022 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2023 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2024 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
2029 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2030 SDValue &Root, SelectionDAG &DAG,
2031 DebugLoc dl) const {
2032 MachineFunction &MF = DAG.getMachineFunction();
2033 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2035 TargetRegisterClass *RC;
2036 if (AFI->isThumb1OnlyFunction())
2037 RC = ARM::tGPRRegisterClass;
2039 RC = ARM::GPRRegisterClass;
2041 // Transform the arguments stored in physical registers into virtual ones.
2042 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2043 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2046 if (NextVA.isMemLoc()) {
2047 MachineFrameInfo *MFI = MF.getFrameInfo();
2048 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2050 // Create load node to retrieve arguments from the stack.
2051 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2052 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2053 PseudoSourceValue::getFixedStack(FI), 0,
2056 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2057 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2060 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2064 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2065 CallingConv::ID CallConv, bool isVarArg,
2066 const SmallVectorImpl<ISD::InputArg>
2068 DebugLoc dl, SelectionDAG &DAG,
2069 SmallVectorImpl<SDValue> &InVals)
2072 MachineFunction &MF = DAG.getMachineFunction();
2073 MachineFrameInfo *MFI = MF.getFrameInfo();
2075 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2077 // Assign locations to all of the incoming arguments.
2078 SmallVector<CCValAssign, 16> ArgLocs;
2079 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2081 CCInfo.AnalyzeFormalArguments(Ins,
2082 CCAssignFnForNode(CallConv, /* Return*/ false,
2085 SmallVector<SDValue, 16> ArgValues;
2087 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2088 CCValAssign &VA = ArgLocs[i];
2090 // Arguments stored in registers.
2091 if (VA.isRegLoc()) {
2092 EVT RegVT = VA.getLocVT();
2095 if (VA.needsCustom()) {
2096 // f64 and vector types are split up into multiple registers or
2097 // combinations of registers and stack slots.
2098 if (VA.getLocVT() == MVT::v2f64) {
2099 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2101 VA = ArgLocs[++i]; // skip ahead to next loc
2103 if (VA.isMemLoc()) {
2104 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2105 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2106 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2107 PseudoSourceValue::getFixedStack(FI), 0,
2110 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2113 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2114 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2115 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2116 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2117 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2119 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2122 TargetRegisterClass *RC;
2124 if (RegVT == MVT::f32)
2125 RC = ARM::SPRRegisterClass;
2126 else if (RegVT == MVT::f64)
2127 RC = ARM::DPRRegisterClass;
2128 else if (RegVT == MVT::v2f64)
2129 RC = ARM::QPRRegisterClass;
2130 else if (RegVT == MVT::i32)
2131 RC = (AFI->isThumb1OnlyFunction() ?
2132 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2134 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2136 // Transform the arguments in physical registers into virtual ones.
2137 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2138 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2141 // If this is an 8 or 16-bit value, it is really passed promoted
2142 // to 32 bits. Insert an assert[sz]ext to capture this, then
2143 // truncate to the right size.
2144 switch (VA.getLocInfo()) {
2145 default: llvm_unreachable("Unknown loc info!");
2146 case CCValAssign::Full: break;
2147 case CCValAssign::BCvt:
2148 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2150 case CCValAssign::SExt:
2151 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2152 DAG.getValueType(VA.getValVT()));
2153 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2155 case CCValAssign::ZExt:
2156 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2157 DAG.getValueType(VA.getValVT()));
2158 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2162 InVals.push_back(ArgValue);
2164 } else { // VA.isRegLoc()
2167 assert(VA.isMemLoc());
2168 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2170 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2171 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2173 // Create load nodes to retrieve arguments from the stack.
2174 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2175 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2176 PseudoSourceValue::getFixedStack(FI), 0,
2183 static const unsigned GPRArgRegs[] = {
2184 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2187 unsigned NumGPRs = CCInfo.getFirstUnallocated
2188 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2190 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2191 unsigned VARegSize = (4 - NumGPRs) * 4;
2192 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2193 unsigned ArgOffset = CCInfo.getNextStackOffset();
2194 if (VARegSaveSize) {
2195 // If this function is vararg, store any remaining integer argument regs
2196 // to their spots on the stack so that they may be loaded by deferencing
2197 // the result of va_next.
2198 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2199 AFI->setVarArgsFrameIndex(
2200 MFI->CreateFixedObject(VARegSaveSize,
2201 ArgOffset + VARegSaveSize - VARegSize,
2203 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2206 SmallVector<SDValue, 4> MemOps;
2207 for (; NumGPRs < 4; ++NumGPRs) {
2208 TargetRegisterClass *RC;
2209 if (AFI->isThumb1OnlyFunction())
2210 RC = ARM::tGPRRegisterClass;
2212 RC = ARM::GPRRegisterClass;
2214 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2215 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2217 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2218 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2219 0, false, false, 0);
2220 MemOps.push_back(Store);
2221 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2222 DAG.getConstant(4, getPointerTy()));
2224 if (!MemOps.empty())
2225 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2226 &MemOps[0], MemOps.size());
2228 // This will point to the next argument passed via stack.
2229 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2235 /// isFloatingPointZero - Return true if this is +0.0.
2236 static bool isFloatingPointZero(SDValue Op) {
2237 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2238 return CFP->getValueAPF().isPosZero();
2239 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2240 // Maybe this has already been legalized into the constant pool?
2241 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2242 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2243 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2244 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2245 return CFP->getValueAPF().isPosZero();
2251 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2252 /// the given operands.
2254 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2255 SDValue &ARMcc, SelectionDAG &DAG,
2256 DebugLoc dl) const {
2257 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2258 unsigned C = RHSC->getZExtValue();
2259 if (!isLegalICmpImmediate(C)) {
2260 // Constant does not fit, try adjusting it by one?
2265 if (isLegalICmpImmediate(C-1)) {
2266 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2267 RHS = DAG.getConstant(C-1, MVT::i32);
2272 if (C > 0 && isLegalICmpImmediate(C-1)) {
2273 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2274 RHS = DAG.getConstant(C-1, MVT::i32);
2279 if (isLegalICmpImmediate(C+1)) {
2280 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2281 RHS = DAG.getConstant(C+1, MVT::i32);
2286 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2287 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2288 RHS = DAG.getConstant(C+1, MVT::i32);
2295 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2296 ARMISD::NodeType CompareType;
2299 CompareType = ARMISD::CMP;
2304 CompareType = ARMISD::CMPZ;
2307 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2308 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2311 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2313 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2314 DebugLoc dl) const {
2316 if (!isFloatingPointZero(RHS))
2317 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2319 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2320 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2323 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2324 EVT VT = Op.getValueType();
2325 SDValue LHS = Op.getOperand(0);
2326 SDValue RHS = Op.getOperand(1);
2327 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2328 SDValue TrueVal = Op.getOperand(2);
2329 SDValue FalseVal = Op.getOperand(3);
2330 DebugLoc dl = Op.getDebugLoc();
2332 if (LHS.getValueType() == MVT::i32) {
2334 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2335 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2336 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2339 ARMCC::CondCodes CondCode, CondCode2;
2340 FPCCToARMCC(CC, CondCode, CondCode2);
2342 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2343 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2344 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2345 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2347 if (CondCode2 != ARMCC::AL) {
2348 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2349 // FIXME: Needs another CMP because flag can have but one use.
2350 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2351 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2352 Result, TrueVal, ARMcc2, CCR, Cmp2);
2357 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2358 /// to morph to an integer compare sequence.
2359 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2360 const ARMSubtarget *Subtarget) {
2361 SDNode *N = Op.getNode();
2362 if (!N->hasOneUse())
2363 // Otherwise it requires moving the value from fp to integer registers.
2365 if (!N->getNumValues())
2367 EVT VT = Op.getValueType();
2368 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2369 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2370 // vmrs are very slow, e.g. cortex-a8.
2373 if (isFloatingPointZero(Op)) {
2377 return ISD::isNormalLoad(N);
2380 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2381 if (isFloatingPointZero(Op))
2382 return DAG.getConstant(0, MVT::i32);
2384 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2385 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2386 Ld->getChain(), Ld->getBasePtr(),
2387 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2388 Ld->isVolatile(), Ld->isNonTemporal(),
2389 Ld->getAlignment());
2391 llvm_unreachable("Unknown VFP cmp argument!");
2394 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2395 SDValue &RetVal1, SDValue &RetVal2) {
2396 if (isFloatingPointZero(Op)) {
2397 RetVal1 = DAG.getConstant(0, MVT::i32);
2398 RetVal2 = DAG.getConstant(0, MVT::i32);
2402 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2403 SDValue Ptr = Ld->getBasePtr();
2404 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2405 Ld->getChain(), Ptr,
2406 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2407 Ld->isVolatile(), Ld->isNonTemporal(),
2408 Ld->getAlignment());
2410 EVT PtrType = Ptr.getValueType();
2411 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2412 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2413 PtrType, Ptr, DAG.getConstant(4, PtrType));
2414 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2415 Ld->getChain(), NewPtr,
2416 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2417 Ld->isVolatile(), Ld->isNonTemporal(),
2422 llvm_unreachable("Unknown VFP cmp argument!");
2425 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2426 /// f32 and even f64 comparisons to integer ones.
2428 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2429 SDValue Chain = Op.getOperand(0);
2430 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2431 SDValue LHS = Op.getOperand(2);
2432 SDValue RHS = Op.getOperand(3);
2433 SDValue Dest = Op.getOperand(4);
2434 DebugLoc dl = Op.getDebugLoc();
2436 bool SeenZero = false;
2437 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2438 canChangeToInt(RHS, SeenZero, Subtarget) &&
2439 // If one of the operand is zero, it's safe to ignore the NaN case since
2440 // we only care about equality comparisons.
2441 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2442 // If unsafe fp math optimization is enabled and there are no othter uses of
2443 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2444 // to an integer comparison.
2445 if (CC == ISD::SETOEQ)
2447 else if (CC == ISD::SETUNE)
2451 if (LHS.getValueType() == MVT::f32) {
2452 LHS = bitcastf32Toi32(LHS, DAG);
2453 RHS = bitcastf32Toi32(RHS, DAG);
2454 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2455 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2456 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2457 Chain, Dest, ARMcc, CCR, Cmp);
2462 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2463 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2464 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2465 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2466 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2467 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2468 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2474 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2475 SDValue Chain = Op.getOperand(0);
2476 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2477 SDValue LHS = Op.getOperand(2);
2478 SDValue RHS = Op.getOperand(3);
2479 SDValue Dest = Op.getOperand(4);
2480 DebugLoc dl = Op.getDebugLoc();
2482 if (LHS.getValueType() == MVT::i32) {
2484 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2485 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2486 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2487 Chain, Dest, ARMcc, CCR, Cmp);
2490 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2493 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2494 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2495 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2496 if (Result.getNode())
2500 ARMCC::CondCodes CondCode, CondCode2;
2501 FPCCToARMCC(CC, CondCode, CondCode2);
2503 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2504 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2505 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2506 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2507 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2508 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2509 if (CondCode2 != ARMCC::AL) {
2510 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2511 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2512 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2517 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2518 SDValue Chain = Op.getOperand(0);
2519 SDValue Table = Op.getOperand(1);
2520 SDValue Index = Op.getOperand(2);
2521 DebugLoc dl = Op.getDebugLoc();
2523 EVT PTy = getPointerTy();
2524 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2525 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2526 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2527 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2528 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2529 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2530 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2531 if (Subtarget->isThumb2()) {
2532 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2533 // which does another jump to the destination. This also makes it easier
2534 // to translate it to TBB / TBH later.
2535 // FIXME: This might not work if the function is extremely large.
2536 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2537 Addr, Op.getOperand(2), JTI, UId);
2539 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2540 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2541 PseudoSourceValue::getJumpTable(), 0,
2543 Chain = Addr.getValue(1);
2544 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2545 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2547 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2548 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2549 Chain = Addr.getValue(1);
2550 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2554 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2555 DebugLoc dl = Op.getDebugLoc();
2558 switch (Op.getOpcode()) {
2560 assert(0 && "Invalid opcode!");
2561 case ISD::FP_TO_SINT:
2562 Opc = ARMISD::FTOSI;
2564 case ISD::FP_TO_UINT:
2565 Opc = ARMISD::FTOUI;
2568 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2569 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2572 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2573 EVT VT = Op.getValueType();
2574 DebugLoc dl = Op.getDebugLoc();
2577 switch (Op.getOpcode()) {
2579 assert(0 && "Invalid opcode!");
2580 case ISD::SINT_TO_FP:
2581 Opc = ARMISD::SITOF;
2583 case ISD::UINT_TO_FP:
2584 Opc = ARMISD::UITOF;
2588 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2589 return DAG.getNode(Opc, dl, VT, Op);
2592 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2593 // Implement fcopysign with a fabs and a conditional fneg.
2594 SDValue Tmp0 = Op.getOperand(0);
2595 SDValue Tmp1 = Op.getOperand(1);
2596 DebugLoc dl = Op.getDebugLoc();
2597 EVT VT = Op.getValueType();
2598 EVT SrcVT = Tmp1.getValueType();
2599 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2600 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2601 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2602 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2603 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2604 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2607 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2608 MachineFunction &MF = DAG.getMachineFunction();
2609 MachineFrameInfo *MFI = MF.getFrameInfo();
2610 MFI->setReturnAddressIsTaken(true);
2612 EVT VT = Op.getValueType();
2613 DebugLoc dl = Op.getDebugLoc();
2614 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2616 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2617 SDValue Offset = DAG.getConstant(4, MVT::i32);
2618 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2619 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2620 NULL, 0, false, false, 0);
2623 // Return LR, which contains the return address. Mark it an implicit live-in.
2624 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2625 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2628 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2629 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2630 MFI->setFrameAddressIsTaken(true);
2632 EVT VT = Op.getValueType();
2633 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2634 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2635 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2636 ? ARM::R7 : ARM::R11;
2637 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2639 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2644 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2645 /// expand a bit convert where either the source or destination type is i64 to
2646 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2647 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2648 /// vectors), since the legalizer won't know what to do with that.
2649 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2650 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2651 DebugLoc dl = N->getDebugLoc();
2652 SDValue Op = N->getOperand(0);
2654 // This function is only supposed to be called for i64 types, either as the
2655 // source or destination of the bit convert.
2656 EVT SrcVT = Op.getValueType();
2657 EVT DstVT = N->getValueType(0);
2658 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2659 "ExpandBIT_CONVERT called for non-i64 type");
2661 // Turn i64->f64 into VMOVDRR.
2662 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2663 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2664 DAG.getConstant(0, MVT::i32));
2665 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2666 DAG.getConstant(1, MVT::i32));
2667 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2668 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2671 // Turn f64->i64 into VMOVRRD.
2672 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2673 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2674 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2675 // Merge the pieces into a single i64 value.
2676 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2682 /// getZeroVector - Returns a vector of specified type with all zero elements.
2683 /// Zero vectors are used to represent vector negation and in those cases
2684 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2685 /// not support i64 elements, so sometimes the zero vectors will need to be
2686 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2688 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2689 assert(VT.isVector() && "Expected a vector type");
2690 // The canonical modified immediate encoding of a zero vector is....0!
2691 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2692 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2693 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2694 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2697 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2698 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2699 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2700 SelectionDAG &DAG) const {
2701 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2702 EVT VT = Op.getValueType();
2703 unsigned VTBits = VT.getSizeInBits();
2704 DebugLoc dl = Op.getDebugLoc();
2705 SDValue ShOpLo = Op.getOperand(0);
2706 SDValue ShOpHi = Op.getOperand(1);
2707 SDValue ShAmt = Op.getOperand(2);
2709 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2711 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2713 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2714 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2715 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2716 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2717 DAG.getConstant(VTBits, MVT::i32));
2718 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2719 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2720 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2722 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2723 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2725 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2726 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2729 SDValue Ops[2] = { Lo, Hi };
2730 return DAG.getMergeValues(Ops, 2, dl);
2733 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2734 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2735 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2736 SelectionDAG &DAG) const {
2737 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2738 EVT VT = Op.getValueType();
2739 unsigned VTBits = VT.getSizeInBits();
2740 DebugLoc dl = Op.getDebugLoc();
2741 SDValue ShOpLo = Op.getOperand(0);
2742 SDValue ShOpHi = Op.getOperand(1);
2743 SDValue ShAmt = Op.getOperand(2);
2746 assert(Op.getOpcode() == ISD::SHL_PARTS);
2747 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2748 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2749 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2750 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2751 DAG.getConstant(VTBits, MVT::i32));
2752 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2753 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2755 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2756 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2757 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2759 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2760 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2763 SDValue Ops[2] = { Lo, Hi };
2764 return DAG.getMergeValues(Ops, 2, dl);
2767 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2768 const ARMSubtarget *ST) {
2769 EVT VT = N->getValueType(0);
2770 DebugLoc dl = N->getDebugLoc();
2772 if (!ST->hasV6T2Ops())
2775 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2776 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2779 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2780 const ARMSubtarget *ST) {
2781 EVT VT = N->getValueType(0);
2782 DebugLoc dl = N->getDebugLoc();
2784 // Lower vector shifts on NEON to use VSHL.
2785 if (VT.isVector()) {
2786 assert(ST->hasNEON() && "unexpected vector shift");
2788 // Left shifts translate directly to the vshiftu intrinsic.
2789 if (N->getOpcode() == ISD::SHL)
2790 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2791 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2792 N->getOperand(0), N->getOperand(1));
2794 assert((N->getOpcode() == ISD::SRA ||
2795 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2797 // NEON uses the same intrinsics for both left and right shifts. For
2798 // right shifts, the shift amounts are negative, so negate the vector of
2800 EVT ShiftVT = N->getOperand(1).getValueType();
2801 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2802 getZeroVector(ShiftVT, DAG, dl),
2804 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2805 Intrinsic::arm_neon_vshifts :
2806 Intrinsic::arm_neon_vshiftu);
2807 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2808 DAG.getConstant(vshiftInt, MVT::i32),
2809 N->getOperand(0), NegatedCount);
2812 // We can get here for a node like i32 = ISD::SHL i32, i64
2816 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2817 "Unknown shift to lower!");
2819 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2820 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2821 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2824 // If we are in thumb mode, we don't have RRX.
2825 if (ST->isThumb1Only()) return SDValue();
2827 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2828 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2829 DAG.getConstant(0, MVT::i32));
2830 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2831 DAG.getConstant(1, MVT::i32));
2833 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2834 // captures the result into a carry flag.
2835 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2836 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2838 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2839 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2841 // Merge the pieces into a single i64 value.
2842 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2845 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2846 SDValue TmpOp0, TmpOp1;
2847 bool Invert = false;
2851 SDValue Op0 = Op.getOperand(0);
2852 SDValue Op1 = Op.getOperand(1);
2853 SDValue CC = Op.getOperand(2);
2854 EVT VT = Op.getValueType();
2855 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2856 DebugLoc dl = Op.getDebugLoc();
2858 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2859 switch (SetCCOpcode) {
2860 default: llvm_unreachable("Illegal FP comparison"); break;
2862 case ISD::SETNE: Invert = true; // Fallthrough
2864 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2866 case ISD::SETLT: Swap = true; // Fallthrough
2868 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2870 case ISD::SETLE: Swap = true; // Fallthrough
2872 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2873 case ISD::SETUGE: Swap = true; // Fallthrough
2874 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2875 case ISD::SETUGT: Swap = true; // Fallthrough
2876 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2877 case ISD::SETUEQ: Invert = true; // Fallthrough
2879 // Expand this to (OLT | OGT).
2883 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2884 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2886 case ISD::SETUO: Invert = true; // Fallthrough
2888 // Expand this to (OLT | OGE).
2892 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2893 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2897 // Integer comparisons.
2898 switch (SetCCOpcode) {
2899 default: llvm_unreachable("Illegal integer comparison"); break;
2900 case ISD::SETNE: Invert = true;
2901 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2902 case ISD::SETLT: Swap = true;
2903 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2904 case ISD::SETLE: Swap = true;
2905 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2906 case ISD::SETULT: Swap = true;
2907 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2908 case ISD::SETULE: Swap = true;
2909 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2912 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2913 if (Opc == ARMISD::VCEQ) {
2916 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2918 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2921 // Ignore bitconvert.
2922 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2923 AndOp = AndOp.getOperand(0);
2925 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2927 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2928 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2935 std::swap(Op0, Op1);
2937 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2940 Result = DAG.getNOT(dl, Result, VT);
2945 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2946 /// valid vector constant for a NEON instruction with a "modified immediate"
2947 /// operand (e.g., VMOV). If so, return the encoded value.
2948 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2949 unsigned SplatBitSize, SelectionDAG &DAG,
2950 EVT &VT, bool is128Bits, bool isVMOV) {
2951 unsigned OpCmode, Imm;
2953 // SplatBitSize is set to the smallest size that splats the vector, so a
2954 // zero vector will always have SplatBitSize == 8. However, NEON modified
2955 // immediate instructions others than VMOV do not support the 8-bit encoding
2956 // of a zero vector, and the default encoding of zero is supposed to be the
2961 switch (SplatBitSize) {
2965 // Any 1-byte value is OK. Op=0, Cmode=1110.
2966 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2969 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2973 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2974 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2975 if ((SplatBits & ~0xff) == 0) {
2976 // Value = 0x00nn: Op=x, Cmode=100x.
2981 if ((SplatBits & ~0xff00) == 0) {
2982 // Value = 0xnn00: Op=x, Cmode=101x.
2984 Imm = SplatBits >> 8;
2990 // NEON's 32-bit VMOV supports splat values where:
2991 // * only one byte is nonzero, or
2992 // * the least significant byte is 0xff and the second byte is nonzero, or
2993 // * the least significant 2 bytes are 0xff and the third is nonzero.
2994 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
2995 if ((SplatBits & ~0xff) == 0) {
2996 // Value = 0x000000nn: Op=x, Cmode=000x.
3001 if ((SplatBits & ~0xff00) == 0) {
3002 // Value = 0x0000nn00: Op=x, Cmode=001x.
3004 Imm = SplatBits >> 8;
3007 if ((SplatBits & ~0xff0000) == 0) {
3008 // Value = 0x00nn0000: Op=x, Cmode=010x.
3010 Imm = SplatBits >> 16;
3013 if ((SplatBits & ~0xff000000) == 0) {
3014 // Value = 0xnn000000: Op=x, Cmode=011x.
3016 Imm = SplatBits >> 24;
3020 if ((SplatBits & ~0xffff) == 0 &&
3021 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3022 // Value = 0x0000nnff: Op=x, Cmode=1100.
3024 Imm = SplatBits >> 8;
3029 if ((SplatBits & ~0xffffff) == 0 &&
3030 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3031 // Value = 0x00nnffff: Op=x, Cmode=1101.
3033 Imm = SplatBits >> 16;
3034 SplatBits |= 0xffff;
3038 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3039 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3040 // VMOV.I32. A (very) minor optimization would be to replicate the value
3041 // and fall through here to test for a valid 64-bit splat. But, then the
3042 // caller would also need to check and handle the change in size.
3048 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3049 uint64_t BitMask = 0xff;
3051 unsigned ImmMask = 1;
3053 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3054 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3057 } else if ((SplatBits & BitMask) != 0) {
3063 // Op=1, Cmode=1110.
3066 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3071 llvm_unreachable("unexpected size for isNEONModifiedImm");
3075 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3076 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3079 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3080 bool &ReverseVEXT, unsigned &Imm) {
3081 unsigned NumElts = VT.getVectorNumElements();
3082 ReverseVEXT = false;
3085 // If this is a VEXT shuffle, the immediate value is the index of the first
3086 // element. The other shuffle indices must be the successive elements after
3088 unsigned ExpectedElt = Imm;
3089 for (unsigned i = 1; i < NumElts; ++i) {
3090 // Increment the expected index. If it wraps around, it may still be
3091 // a VEXT but the source vectors must be swapped.
3093 if (ExpectedElt == NumElts * 2) {
3098 if (ExpectedElt != static_cast<unsigned>(M[i]))
3102 // Adjust the index value if the source operands will be swapped.
3109 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3110 /// instruction with the specified blocksize. (The order of the elements
3111 /// within each block of the vector is reversed.)
3112 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3113 unsigned BlockSize) {
3114 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3115 "Only possible block sizes for VREV are: 16, 32, 64");
3117 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3121 unsigned NumElts = VT.getVectorNumElements();
3122 unsigned BlockElts = M[0] + 1;
3124 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3127 for (unsigned i = 0; i < NumElts; ++i) {
3128 if ((unsigned) M[i] !=
3129 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3136 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3137 unsigned &WhichResult) {
3138 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3142 unsigned NumElts = VT.getVectorNumElements();
3143 WhichResult = (M[0] == 0 ? 0 : 1);
3144 for (unsigned i = 0; i < NumElts; i += 2) {
3145 if ((unsigned) M[i] != i + WhichResult ||
3146 (unsigned) M[i+1] != i + NumElts + WhichResult)
3152 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3153 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3154 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3155 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3156 unsigned &WhichResult) {
3157 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3161 unsigned NumElts = VT.getVectorNumElements();
3162 WhichResult = (M[0] == 0 ? 0 : 1);
3163 for (unsigned i = 0; i < NumElts; i += 2) {
3164 if ((unsigned) M[i] != i + WhichResult ||
3165 (unsigned) M[i+1] != i + WhichResult)
3171 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3172 unsigned &WhichResult) {
3173 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3177 unsigned NumElts = VT.getVectorNumElements();
3178 WhichResult = (M[0] == 0 ? 0 : 1);
3179 for (unsigned i = 0; i != NumElts; ++i) {
3180 if ((unsigned) M[i] != 2 * i + WhichResult)
3184 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3185 if (VT.is64BitVector() && EltSz == 32)
3191 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3192 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3193 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3194 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3195 unsigned &WhichResult) {
3196 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3200 unsigned Half = VT.getVectorNumElements() / 2;
3201 WhichResult = (M[0] == 0 ? 0 : 1);
3202 for (unsigned j = 0; j != 2; ++j) {
3203 unsigned Idx = WhichResult;
3204 for (unsigned i = 0; i != Half; ++i) {
3205 if ((unsigned) M[i + j * Half] != Idx)
3211 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3212 if (VT.is64BitVector() && EltSz == 32)
3218 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3219 unsigned &WhichResult) {
3220 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3224 unsigned NumElts = VT.getVectorNumElements();
3225 WhichResult = (M[0] == 0 ? 0 : 1);
3226 unsigned Idx = WhichResult * NumElts / 2;
3227 for (unsigned i = 0; i != NumElts; i += 2) {
3228 if ((unsigned) M[i] != Idx ||
3229 (unsigned) M[i+1] != Idx + NumElts)
3234 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3235 if (VT.is64BitVector() && EltSz == 32)
3241 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3242 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3243 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3244 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3245 unsigned &WhichResult) {
3246 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3250 unsigned NumElts = VT.getVectorNumElements();
3251 WhichResult = (M[0] == 0 ? 0 : 1);
3252 unsigned Idx = WhichResult * NumElts / 2;
3253 for (unsigned i = 0; i != NumElts; i += 2) {
3254 if ((unsigned) M[i] != Idx ||
3255 (unsigned) M[i+1] != Idx)
3260 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3261 if (VT.is64BitVector() && EltSz == 32)
3267 // If N is an integer constant that can be moved into a register in one
3268 // instruction, return an SDValue of such a constant (will become a MOV
3269 // instruction). Otherwise return null.
3270 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3271 const ARMSubtarget *ST, DebugLoc dl) {
3273 if (!isa<ConstantSDNode>(N))
3275 Val = cast<ConstantSDNode>(N)->getZExtValue();
3277 if (ST->isThumb1Only()) {
3278 if (Val <= 255 || ~Val <= 255)
3279 return DAG.getConstant(Val, MVT::i32);
3281 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3282 return DAG.getConstant(Val, MVT::i32);
3287 // If this is a case we can't handle, return null and let the default
3288 // expansion code take care of it.
3289 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3290 const ARMSubtarget *ST) {
3291 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3292 DebugLoc dl = Op.getDebugLoc();
3293 EVT VT = Op.getValueType();
3295 APInt SplatBits, SplatUndef;
3296 unsigned SplatBitSize;
3298 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3299 if (SplatBitSize <= 64) {
3300 // Check if an immediate VMOV works.
3302 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3303 SplatUndef.getZExtValue(), SplatBitSize,
3304 DAG, VmovVT, VT.is128BitVector(), true);
3305 if (Val.getNode()) {
3306 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3307 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3310 // Try an immediate VMVN.
3311 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3312 ((1LL << SplatBitSize) - 1));
3313 Val = isNEONModifiedImm(NegatedImm,
3314 SplatUndef.getZExtValue(), SplatBitSize,
3315 DAG, VmovVT, VT.is128BitVector(), false);
3316 if (Val.getNode()) {
3317 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3318 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3323 // Scan through the operands to see if only one value is used.
3324 unsigned NumElts = VT.getVectorNumElements();
3325 bool isOnlyLowElement = true;
3326 bool usesOnlyOneValue = true;
3327 bool isConstant = true;
3329 for (unsigned i = 0; i < NumElts; ++i) {
3330 SDValue V = Op.getOperand(i);
3331 if (V.getOpcode() == ISD::UNDEF)
3334 isOnlyLowElement = false;
3335 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3338 if (!Value.getNode())
3340 else if (V != Value)
3341 usesOnlyOneValue = false;
3344 if (!Value.getNode())
3345 return DAG.getUNDEF(VT);
3347 if (isOnlyLowElement)
3348 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3350 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3352 if (EnableARMVDUPsplat) {
3353 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3354 // i32 and try again.
3355 if (usesOnlyOneValue && EltSize <= 32) {
3357 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3358 if (VT.getVectorElementType().isFloatingPoint()) {
3359 SmallVector<SDValue, 8> Ops;
3360 for (unsigned i = 0; i < NumElts; ++i)
3361 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3363 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3365 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3366 LowerBUILD_VECTOR(Val, DAG, ST));
3368 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3370 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3374 // If all elements are constants and the case above didn't get hit, fall back
3375 // to the default expansion, which will generate a load from the constant
3380 if (!EnableARMVDUPsplat) {
3381 // Use VDUP for non-constant splats.
3382 if (usesOnlyOneValue && EltSize <= 32)
3383 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3386 // Vectors with 32- or 64-bit elements can be built by directly assigning
3387 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3388 // will be legalized.
3389 if (EltSize >= 32) {
3390 // Do the expansion with floating-point types, since that is what the VFP
3391 // registers are defined to use, and since i64 is not legal.
3392 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3393 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3394 SmallVector<SDValue, 8> Ops;
3395 for (unsigned i = 0; i < NumElts; ++i)
3396 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3397 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3398 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3404 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3405 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3406 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3407 /// are assumed to be legal.
3409 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3411 if (VT.getVectorNumElements() == 4 &&
3412 (VT.is128BitVector() || VT.is64BitVector())) {
3413 unsigned PFIndexes[4];
3414 for (unsigned i = 0; i != 4; ++i) {
3418 PFIndexes[i] = M[i];
3421 // Compute the index in the perfect shuffle table.
3422 unsigned PFTableIndex =
3423 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3424 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3425 unsigned Cost = (PFEntry >> 30);
3432 unsigned Imm, WhichResult;
3434 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3435 return (EltSize >= 32 ||
3436 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3437 isVREVMask(M, VT, 64) ||
3438 isVREVMask(M, VT, 32) ||
3439 isVREVMask(M, VT, 16) ||
3440 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3441 isVTRNMask(M, VT, WhichResult) ||
3442 isVUZPMask(M, VT, WhichResult) ||
3443 isVZIPMask(M, VT, WhichResult) ||
3444 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3445 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3446 isVZIP_v_undef_Mask(M, VT, WhichResult));
3449 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3450 /// the specified operations to build the shuffle.
3451 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3452 SDValue RHS, SelectionDAG &DAG,
3454 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3455 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3456 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3459 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3468 OP_VUZPL, // VUZP, left result
3469 OP_VUZPR, // VUZP, right result
3470 OP_VZIPL, // VZIP, left result
3471 OP_VZIPR, // VZIP, right result
3472 OP_VTRNL, // VTRN, left result
3473 OP_VTRNR // VTRN, right result
3476 if (OpNum == OP_COPY) {
3477 if (LHSID == (1*9+2)*9+3) return LHS;
3478 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3482 SDValue OpLHS, OpRHS;
3483 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3484 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3485 EVT VT = OpLHS.getValueType();
3488 default: llvm_unreachable("Unknown shuffle opcode!");
3490 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3495 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3496 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3500 return DAG.getNode(ARMISD::VEXT, dl, VT,
3502 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3505 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3506 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3509 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3510 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3513 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3514 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3518 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3519 SDValue V1 = Op.getOperand(0);
3520 SDValue V2 = Op.getOperand(1);
3521 DebugLoc dl = Op.getDebugLoc();
3522 EVT VT = Op.getValueType();
3523 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3524 SmallVector<int, 8> ShuffleMask;
3526 // Convert shuffles that are directly supported on NEON to target-specific
3527 // DAG nodes, instead of keeping them as shuffles and matching them again
3528 // during code selection. This is more efficient and avoids the possibility
3529 // of inconsistencies between legalization and selection.
3530 // FIXME: floating-point vectors should be canonicalized to integer vectors
3531 // of the same time so that they get CSEd properly.
3532 SVN->getMask(ShuffleMask);
3534 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3535 if (EltSize <= 32) {
3536 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3537 int Lane = SVN->getSplatIndex();
3538 // If this is undef splat, generate it via "just" vdup, if possible.
3539 if (Lane == -1) Lane = 0;
3541 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3542 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3544 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3545 DAG.getConstant(Lane, MVT::i32));
3550 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3553 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3554 DAG.getConstant(Imm, MVT::i32));
3557 if (isVREVMask(ShuffleMask, VT, 64))
3558 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3559 if (isVREVMask(ShuffleMask, VT, 32))
3560 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3561 if (isVREVMask(ShuffleMask, VT, 16))
3562 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3564 // Check for Neon shuffles that modify both input vectors in place.
3565 // If both results are used, i.e., if there are two shuffles with the same
3566 // source operands and with masks corresponding to both results of one of
3567 // these operations, DAG memoization will ensure that a single node is
3568 // used for both shuffles.
3569 unsigned WhichResult;
3570 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3571 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3572 V1, V2).getValue(WhichResult);
3573 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3574 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3575 V1, V2).getValue(WhichResult);
3576 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3577 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3578 V1, V2).getValue(WhichResult);
3580 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3581 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3582 V1, V1).getValue(WhichResult);
3583 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3584 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3585 V1, V1).getValue(WhichResult);
3586 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3587 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3588 V1, V1).getValue(WhichResult);
3591 // If the shuffle is not directly supported and it has 4 elements, use
3592 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3593 unsigned NumElts = VT.getVectorNumElements();
3595 unsigned PFIndexes[4];
3596 for (unsigned i = 0; i != 4; ++i) {
3597 if (ShuffleMask[i] < 0)
3600 PFIndexes[i] = ShuffleMask[i];
3603 // Compute the index in the perfect shuffle table.
3604 unsigned PFTableIndex =
3605 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3606 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3607 unsigned Cost = (PFEntry >> 30);
3610 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3613 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3614 if (EltSize >= 32) {
3615 // Do the expansion with floating-point types, since that is what the VFP
3616 // registers are defined to use, and since i64 is not legal.
3617 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3618 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3619 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3620 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3621 SmallVector<SDValue, 8> Ops;
3622 for (unsigned i = 0; i < NumElts; ++i) {
3623 if (ShuffleMask[i] < 0)
3624 Ops.push_back(DAG.getUNDEF(EltVT));
3626 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3627 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3628 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3631 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3632 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3638 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3639 EVT VT = Op.getValueType();
3640 DebugLoc dl = Op.getDebugLoc();
3641 SDValue Vec = Op.getOperand(0);
3642 SDValue Lane = Op.getOperand(1);
3643 assert(VT == MVT::i32 &&
3644 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3645 "unexpected type for custom-lowering vector extract");
3646 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3649 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3650 // The only time a CONCAT_VECTORS operation can have legal types is when
3651 // two 64-bit vectors are concatenated to a 128-bit vector.
3652 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3653 "unexpected CONCAT_VECTORS");
3654 DebugLoc dl = Op.getDebugLoc();
3655 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3656 SDValue Op0 = Op.getOperand(0);
3657 SDValue Op1 = Op.getOperand(1);
3658 if (Op0.getOpcode() != ISD::UNDEF)
3659 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3660 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3661 DAG.getIntPtrConstant(0));
3662 if (Op1.getOpcode() != ISD::UNDEF)
3663 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3664 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3665 DAG.getIntPtrConstant(1));
3666 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3669 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3670 switch (Op.getOpcode()) {
3671 default: llvm_unreachable("Don't know how to custom lower this!");
3672 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3673 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3674 case ISD::GlobalAddress:
3675 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3676 LowerGlobalAddressELF(Op, DAG);
3677 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3678 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3679 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3680 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3681 case ISD::VASTART: return LowerVASTART(Op, DAG);
3682 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3683 case ISD::SINT_TO_FP:
3684 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3685 case ISD::FP_TO_SINT:
3686 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3687 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3688 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3689 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3690 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3691 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3692 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3693 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3695 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3698 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3699 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3700 case ISD::SRL_PARTS:
3701 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3702 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3703 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3704 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3705 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3706 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3707 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3712 /// ReplaceNodeResults - Replace the results of node with an illegal result
3713 /// type with new values built out of custom code.
3714 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3715 SmallVectorImpl<SDValue>&Results,
3716 SelectionDAG &DAG) const {
3718 switch (N->getOpcode()) {
3720 llvm_unreachable("Don't know how to custom expand this!");
3722 case ISD::BIT_CONVERT:
3723 Res = ExpandBIT_CONVERT(N, DAG);
3727 Res = LowerShift(N, DAG, Subtarget);
3731 Results.push_back(Res);
3734 //===----------------------------------------------------------------------===//
3735 // ARM Scheduler Hooks
3736 //===----------------------------------------------------------------------===//
3739 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3740 MachineBasicBlock *BB,
3741 unsigned Size) const {
3742 unsigned dest = MI->getOperand(0).getReg();
3743 unsigned ptr = MI->getOperand(1).getReg();
3744 unsigned oldval = MI->getOperand(2).getReg();
3745 unsigned newval = MI->getOperand(3).getReg();
3746 unsigned scratch = BB->getParent()->getRegInfo()
3747 .createVirtualRegister(ARM::GPRRegisterClass);
3748 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3749 DebugLoc dl = MI->getDebugLoc();
3750 bool isThumb2 = Subtarget->isThumb2();
3752 unsigned ldrOpc, strOpc;
3754 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3756 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3757 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3760 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3761 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3764 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3765 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3769 MachineFunction *MF = BB->getParent();
3770 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3771 MachineFunction::iterator It = BB;
3772 ++It; // insert the new blocks after the current block
3774 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3775 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3776 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3777 MF->insert(It, loop1MBB);
3778 MF->insert(It, loop2MBB);
3779 MF->insert(It, exitMBB);
3781 // Transfer the remainder of BB and its successor edges to exitMBB.
3782 exitMBB->splice(exitMBB->begin(), BB,
3783 llvm::next(MachineBasicBlock::iterator(MI)),
3785 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3789 // fallthrough --> loop1MBB
3790 BB->addSuccessor(loop1MBB);
3793 // ldrex dest, [ptr]
3797 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3798 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3799 .addReg(dest).addReg(oldval));
3800 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3801 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3802 BB->addSuccessor(loop2MBB);
3803 BB->addSuccessor(exitMBB);
3806 // strex scratch, newval, [ptr]
3810 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3812 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3813 .addReg(scratch).addImm(0));
3814 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3815 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3816 BB->addSuccessor(loop1MBB);
3817 BB->addSuccessor(exitMBB);
3823 MI->eraseFromParent(); // The instruction is gone now.
3829 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3830 unsigned Size, unsigned BinOpcode) const {
3831 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3832 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3834 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3835 MachineFunction *MF = BB->getParent();
3836 MachineFunction::iterator It = BB;
3839 unsigned dest = MI->getOperand(0).getReg();
3840 unsigned ptr = MI->getOperand(1).getReg();
3841 unsigned incr = MI->getOperand(2).getReg();
3842 DebugLoc dl = MI->getDebugLoc();
3844 bool isThumb2 = Subtarget->isThumb2();
3845 unsigned ldrOpc, strOpc;
3847 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3849 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3850 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3853 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3854 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3857 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3858 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3862 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3863 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3864 MF->insert(It, loopMBB);
3865 MF->insert(It, exitMBB);
3867 // Transfer the remainder of BB and its successor edges to exitMBB.
3868 exitMBB->splice(exitMBB->begin(), BB,
3869 llvm::next(MachineBasicBlock::iterator(MI)),
3871 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3873 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3874 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3875 unsigned scratch2 = (!BinOpcode) ? incr :
3876 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3880 // fallthrough --> loopMBB
3881 BB->addSuccessor(loopMBB);
3885 // <binop> scratch2, dest, incr
3886 // strex scratch, scratch2, ptr
3889 // fallthrough --> exitMBB
3891 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3893 // operand order needs to go the other way for NAND
3894 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3895 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3896 addReg(incr).addReg(dest)).addReg(0);
3898 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3899 addReg(dest).addReg(incr)).addReg(0);
3902 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3904 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3905 .addReg(scratch).addImm(0));
3906 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3907 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3909 BB->addSuccessor(loopMBB);
3910 BB->addSuccessor(exitMBB);
3916 MI->eraseFromParent(); // The instruction is gone now.
3922 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3923 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3924 E = MBB->succ_end(); I != E; ++I)
3927 llvm_unreachable("Expecting a BB with two successors!");
3931 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3932 MachineBasicBlock *BB) const {
3933 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3934 DebugLoc dl = MI->getDebugLoc();
3935 bool isThumb2 = Subtarget->isThumb2();
3936 switch (MI->getOpcode()) {
3939 llvm_unreachable("Unexpected instr type to insert");
3941 case ARM::ATOMIC_LOAD_ADD_I8:
3942 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3943 case ARM::ATOMIC_LOAD_ADD_I16:
3944 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3945 case ARM::ATOMIC_LOAD_ADD_I32:
3946 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3948 case ARM::ATOMIC_LOAD_AND_I8:
3949 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3950 case ARM::ATOMIC_LOAD_AND_I16:
3951 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3952 case ARM::ATOMIC_LOAD_AND_I32:
3953 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3955 case ARM::ATOMIC_LOAD_OR_I8:
3956 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3957 case ARM::ATOMIC_LOAD_OR_I16:
3958 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3959 case ARM::ATOMIC_LOAD_OR_I32:
3960 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3962 case ARM::ATOMIC_LOAD_XOR_I8:
3963 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3964 case ARM::ATOMIC_LOAD_XOR_I16:
3965 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3966 case ARM::ATOMIC_LOAD_XOR_I32:
3967 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3969 case ARM::ATOMIC_LOAD_NAND_I8:
3970 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3971 case ARM::ATOMIC_LOAD_NAND_I16:
3972 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3973 case ARM::ATOMIC_LOAD_NAND_I32:
3974 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3976 case ARM::ATOMIC_LOAD_SUB_I8:
3977 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3978 case ARM::ATOMIC_LOAD_SUB_I16:
3979 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3980 case ARM::ATOMIC_LOAD_SUB_I32:
3981 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3983 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3984 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3985 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3987 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3988 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3989 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3991 case ARM::tMOVCCr_pseudo: {
3992 // To "insert" a SELECT_CC instruction, we actually have to insert the
3993 // diamond control-flow pattern. The incoming instruction knows the
3994 // destination vreg to set, the condition code register to branch on, the
3995 // true/false values to select between, and a branch opcode to use.
3996 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3997 MachineFunction::iterator It = BB;
4003 // cmpTY ccX, r1, r2
4005 // fallthrough --> copy0MBB
4006 MachineBasicBlock *thisMBB = BB;
4007 MachineFunction *F = BB->getParent();
4008 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4009 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4010 F->insert(It, copy0MBB);
4011 F->insert(It, sinkMBB);
4013 // Transfer the remainder of BB and its successor edges to sinkMBB.
4014 sinkMBB->splice(sinkMBB->begin(), BB,
4015 llvm::next(MachineBasicBlock::iterator(MI)),
4017 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4019 BB->addSuccessor(copy0MBB);
4020 BB->addSuccessor(sinkMBB);
4022 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4023 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4026 // %FalseValue = ...
4027 // # fallthrough to sinkMBB
4030 // Update machine-CFG edges
4031 BB->addSuccessor(sinkMBB);
4034 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4037 BuildMI(*BB, BB->begin(), dl,
4038 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4039 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4040 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4042 MI->eraseFromParent(); // The pseudo instruction is gone now.
4047 case ARM::BCCZi64: {
4048 // Compare both parts that make up the double comparison separately for
4050 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4052 unsigned LHS1 = MI->getOperand(1).getReg();
4053 unsigned LHS2 = MI->getOperand(2).getReg();
4055 AddDefaultPred(BuildMI(BB, dl,
4056 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4057 .addReg(LHS1).addImm(0));
4058 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4059 .addReg(LHS2).addImm(0)
4060 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4062 unsigned RHS1 = MI->getOperand(3).getReg();
4063 unsigned RHS2 = MI->getOperand(4).getReg();
4064 AddDefaultPred(BuildMI(BB, dl,
4065 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4066 .addReg(LHS1).addReg(RHS1));
4067 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4068 .addReg(LHS2).addReg(RHS2)
4069 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4072 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4073 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4074 if (MI->getOperand(0).getImm() == ARMCC::NE)
4075 std::swap(destMBB, exitMBB);
4077 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4078 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4079 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4082 MI->eraseFromParent(); // The pseudo instruction is gone now.
4089 case ARM::t2SUBrSPi_:
4090 case ARM::t2SUBrSPi12_:
4091 case ARM::t2SUBrSPs_: {
4092 MachineFunction *MF = BB->getParent();
4093 unsigned DstReg = MI->getOperand(0).getReg();
4094 unsigned SrcReg = MI->getOperand(1).getReg();
4095 bool DstIsDead = MI->getOperand(0).isDead();
4096 bool SrcIsKill = MI->getOperand(1).isKill();
4098 if (SrcReg != ARM::SP) {
4099 // Copy the source to SP from virtual register.
4100 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
4101 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4102 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
4103 BuildMI(*BB, MI, dl, TII->get(CopyOpc), ARM::SP)
4104 .addReg(SrcReg, getKillRegState(SrcIsKill));
4108 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
4109 switch (MI->getOpcode()) {
4111 llvm_unreachable("Unexpected pseudo instruction!");
4117 OpOpc = ARM::tADDspr;
4120 OpOpc = ARM::tSUBspi;
4122 case ARM::t2SUBrSPi_:
4123 OpOpc = ARM::t2SUBrSPi;
4124 NeedPred = true; NeedCC = true;
4126 case ARM::t2SUBrSPi12_:
4127 OpOpc = ARM::t2SUBrSPi12;
4130 case ARM::t2SUBrSPs_:
4131 OpOpc = ARM::t2SUBrSPs;
4132 NeedPred = true; NeedCC = true; NeedOp3 = true;
4135 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(OpOpc), ARM::SP);
4136 if (OpOpc == ARM::tAND)
4137 AddDefaultT1CC(MIB);
4138 MIB.addReg(ARM::SP);
4139 MIB.addOperand(MI->getOperand(2));
4141 MIB.addOperand(MI->getOperand(3));
4143 AddDefaultPred(MIB);
4147 // Copy the result from SP to virtual register.
4148 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
4149 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4150 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
4151 BuildMI(*BB, MI, dl, TII->get(CopyOpc))
4152 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
4154 MI->eraseFromParent(); // The pseudo instruction is gone now.
4160 //===----------------------------------------------------------------------===//
4161 // ARM Optimization Hooks
4162 //===----------------------------------------------------------------------===//
4165 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4166 TargetLowering::DAGCombinerInfo &DCI) {
4167 SelectionDAG &DAG = DCI.DAG;
4168 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4169 EVT VT = N->getValueType(0);
4170 unsigned Opc = N->getOpcode();
4171 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4172 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4173 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4174 ISD::CondCode CC = ISD::SETCC_INVALID;
4177 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4179 SDValue CCOp = Slct.getOperand(0);
4180 if (CCOp.getOpcode() == ISD::SETCC)
4181 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4184 bool DoXform = false;
4186 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4189 if (LHS.getOpcode() == ISD::Constant &&
4190 cast<ConstantSDNode>(LHS)->isNullValue()) {
4192 } else if (CC != ISD::SETCC_INVALID &&
4193 RHS.getOpcode() == ISD::Constant &&
4194 cast<ConstantSDNode>(RHS)->isNullValue()) {
4195 std::swap(LHS, RHS);
4196 SDValue Op0 = Slct.getOperand(0);
4197 EVT OpVT = isSlctCC ? Op0.getValueType() :
4198 Op0.getOperand(0).getValueType();
4199 bool isInt = OpVT.isInteger();
4200 CC = ISD::getSetCCInverse(CC, isInt);
4202 if (!TLI.isCondCodeLegal(CC, OpVT))
4203 return SDValue(); // Inverse operator isn't legal.
4210 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4212 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4213 Slct.getOperand(0), Slct.getOperand(1), CC);
4214 SDValue CCOp = Slct.getOperand(0);
4216 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4217 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4218 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4219 CCOp, OtherOp, Result);
4224 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4225 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4226 /// called with the default operands, and if that fails, with commuted
4228 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4229 TargetLowering::DAGCombinerInfo &DCI) {
4230 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4231 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4232 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4233 if (Result.getNode()) return Result;
4239 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4241 static SDValue PerformADDCombine(SDNode *N,
4242 TargetLowering::DAGCombinerInfo &DCI) {
4243 SDValue N0 = N->getOperand(0);
4244 SDValue N1 = N->getOperand(1);
4246 // First try with the default operand order.
4247 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4248 if (Result.getNode())
4251 // If that didn't work, try again with the operands commuted.
4252 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4255 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4257 static SDValue PerformSUBCombine(SDNode *N,
4258 TargetLowering::DAGCombinerInfo &DCI) {
4259 SDValue N0 = N->getOperand(0);
4260 SDValue N1 = N->getOperand(1);
4262 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4263 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4264 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4265 if (Result.getNode()) return Result;
4271 static SDValue PerformMULCombine(SDNode *N,
4272 TargetLowering::DAGCombinerInfo &DCI,
4273 const ARMSubtarget *Subtarget) {
4274 SelectionDAG &DAG = DCI.DAG;
4276 if (Subtarget->isThumb1Only())
4279 if (DAG.getMachineFunction().
4280 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4283 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4286 EVT VT = N->getValueType(0);
4290 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4294 uint64_t MulAmt = C->getZExtValue();
4295 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4296 ShiftAmt = ShiftAmt & (32 - 1);
4297 SDValue V = N->getOperand(0);
4298 DebugLoc DL = N->getDebugLoc();
4301 MulAmt >>= ShiftAmt;
4302 if (isPowerOf2_32(MulAmt - 1)) {
4303 // (mul x, 2^N + 1) => (add (shl x, N), x)
4304 Res = DAG.getNode(ISD::ADD, DL, VT,
4305 V, DAG.getNode(ISD::SHL, DL, VT,
4306 V, DAG.getConstant(Log2_32(MulAmt-1),
4308 } else if (isPowerOf2_32(MulAmt + 1)) {
4309 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4310 Res = DAG.getNode(ISD::SUB, DL, VT,
4311 DAG.getNode(ISD::SHL, DL, VT,
4312 V, DAG.getConstant(Log2_32(MulAmt+1),
4319 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4320 DAG.getConstant(ShiftAmt, MVT::i32));
4322 // Do not add new nodes to DAG combiner worklist.
4323 DCI.CombineTo(N, Res, false);
4327 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4328 static SDValue PerformORCombine(SDNode *N,
4329 TargetLowering::DAGCombinerInfo &DCI,
4330 const ARMSubtarget *Subtarget) {
4331 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4334 // BFI is only available on V6T2+
4335 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4338 SelectionDAG &DAG = DCI.DAG;
4339 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4340 DebugLoc DL = N->getDebugLoc();
4341 // 1) or (and A, mask), val => ARMbfi A, val, mask
4342 // iff (val & mask) == val
4344 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4345 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4346 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4347 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4348 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4349 // (i.e., copy a bitfield value into another bitfield of the same width)
4350 if (N0.getOpcode() != ISD::AND)
4353 EVT VT = N->getValueType(0);
4358 // The value and the mask need to be constants so we can verify this is
4359 // actually a bitfield set. If the mask is 0xffff, we can do better
4360 // via a movt instruction, so don't use BFI in that case.
4361 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4364 unsigned Mask = C->getZExtValue();
4368 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4369 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4370 unsigned Val = C->getZExtValue();
4371 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4373 Val >>= CountTrailingZeros_32(~Mask);
4375 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4376 DAG.getConstant(Val, MVT::i32),
4377 DAG.getConstant(Mask, MVT::i32));
4379 // Do not add new nodes to DAG combiner worklist.
4380 DCI.CombineTo(N, Res, false);
4381 } else if (N1.getOpcode() == ISD::AND) {
4382 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4383 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4386 unsigned Mask2 = C->getZExtValue();
4388 if (ARM::isBitFieldInvertedMask(Mask) &&
4389 ARM::isBitFieldInvertedMask(~Mask2) &&
4390 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4391 // The pack halfword instruction works better for masks that fit it,
4392 // so use that when it's available.
4393 if (Subtarget->hasT2ExtractPack() &&
4394 (Mask == 0xffff || Mask == 0xffff0000))
4397 unsigned lsb = CountTrailingZeros_32(Mask2);
4398 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4399 DAG.getConstant(lsb, MVT::i32));
4400 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4401 DAG.getConstant(Mask, MVT::i32));
4402 // Do not add new nodes to DAG combiner worklist.
4403 DCI.CombineTo(N, Res, false);
4404 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4405 ARM::isBitFieldInvertedMask(Mask2) &&
4406 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4407 // The pack halfword instruction works better for masks that fit it,
4408 // so use that when it's available.
4409 if (Subtarget->hasT2ExtractPack() &&
4410 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4413 unsigned lsb = CountTrailingZeros_32(Mask);
4414 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4415 DAG.getConstant(lsb, MVT::i32));
4416 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4417 DAG.getConstant(Mask2, MVT::i32));
4418 // Do not add new nodes to DAG combiner worklist.
4419 DCI.CombineTo(N, Res, false);
4426 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4427 /// ARMISD::VMOVRRD.
4428 static SDValue PerformVMOVRRDCombine(SDNode *N,
4429 TargetLowering::DAGCombinerInfo &DCI) {
4430 // fmrrd(fmdrr x, y) -> x,y
4431 SDValue InDouble = N->getOperand(0);
4432 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4433 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4437 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4438 /// ARMISD::VDUPLANE.
4439 static SDValue PerformVDUPLANECombine(SDNode *N,
4440 TargetLowering::DAGCombinerInfo &DCI) {
4441 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4443 SDValue Op = N->getOperand(0);
4444 EVT VT = N->getValueType(0);
4446 // Ignore bit_converts.
4447 while (Op.getOpcode() == ISD::BIT_CONVERT)
4448 Op = Op.getOperand(0);
4449 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4452 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4453 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4454 // The canonical VMOV for a zero vector uses a 32-bit element size.
4455 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4457 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4459 if (EltSize > VT.getVectorElementType().getSizeInBits())
4462 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4463 return DCI.CombineTo(N, Res, false);
4466 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4467 /// operand of a vector shift operation, where all the elements of the
4468 /// build_vector must have the same constant integer value.
4469 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4470 // Ignore bit_converts.
4471 while (Op.getOpcode() == ISD::BIT_CONVERT)
4472 Op = Op.getOperand(0);
4473 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4474 APInt SplatBits, SplatUndef;
4475 unsigned SplatBitSize;
4477 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4478 HasAnyUndefs, ElementBits) ||
4479 SplatBitSize > ElementBits)
4481 Cnt = SplatBits.getSExtValue();
4485 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4486 /// operand of a vector shift left operation. That value must be in the range:
4487 /// 0 <= Value < ElementBits for a left shift; or
4488 /// 0 <= Value <= ElementBits for a long left shift.
4489 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4490 assert(VT.isVector() && "vector shift count is not a vector type");
4491 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4492 if (! getVShiftImm(Op, ElementBits, Cnt))
4494 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4497 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4498 /// operand of a vector shift right operation. For a shift opcode, the value
4499 /// is positive, but for an intrinsic the value count must be negative. The
4500 /// absolute value must be in the range:
4501 /// 1 <= |Value| <= ElementBits for a right shift; or
4502 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4503 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4505 assert(VT.isVector() && "vector shift count is not a vector type");
4506 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4507 if (! getVShiftImm(Op, ElementBits, Cnt))
4511 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4514 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4515 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4516 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4519 // Don't do anything for most intrinsics.
4522 // Vector shifts: check for immediate versions and lower them.
4523 // Note: This is done during DAG combining instead of DAG legalizing because
4524 // the build_vectors for 64-bit vector element shift counts are generally
4525 // not legal, and it is hard to see their values after they get legalized to
4526 // loads from a constant pool.
4527 case Intrinsic::arm_neon_vshifts:
4528 case Intrinsic::arm_neon_vshiftu:
4529 case Intrinsic::arm_neon_vshiftls:
4530 case Intrinsic::arm_neon_vshiftlu:
4531 case Intrinsic::arm_neon_vshiftn:
4532 case Intrinsic::arm_neon_vrshifts:
4533 case Intrinsic::arm_neon_vrshiftu:
4534 case Intrinsic::arm_neon_vrshiftn:
4535 case Intrinsic::arm_neon_vqshifts:
4536 case Intrinsic::arm_neon_vqshiftu:
4537 case Intrinsic::arm_neon_vqshiftsu:
4538 case Intrinsic::arm_neon_vqshiftns:
4539 case Intrinsic::arm_neon_vqshiftnu:
4540 case Intrinsic::arm_neon_vqshiftnsu:
4541 case Intrinsic::arm_neon_vqrshiftns:
4542 case Intrinsic::arm_neon_vqrshiftnu:
4543 case Intrinsic::arm_neon_vqrshiftnsu: {
4544 EVT VT = N->getOperand(1).getValueType();
4546 unsigned VShiftOpc = 0;
4549 case Intrinsic::arm_neon_vshifts:
4550 case Intrinsic::arm_neon_vshiftu:
4551 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4552 VShiftOpc = ARMISD::VSHL;
4555 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4556 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4557 ARMISD::VSHRs : ARMISD::VSHRu);
4562 case Intrinsic::arm_neon_vshiftls:
4563 case Intrinsic::arm_neon_vshiftlu:
4564 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4566 llvm_unreachable("invalid shift count for vshll intrinsic");
4568 case Intrinsic::arm_neon_vrshifts:
4569 case Intrinsic::arm_neon_vrshiftu:
4570 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4574 case Intrinsic::arm_neon_vqshifts:
4575 case Intrinsic::arm_neon_vqshiftu:
4576 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4580 case Intrinsic::arm_neon_vqshiftsu:
4581 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4583 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4585 case Intrinsic::arm_neon_vshiftn:
4586 case Intrinsic::arm_neon_vrshiftn:
4587 case Intrinsic::arm_neon_vqshiftns:
4588 case Intrinsic::arm_neon_vqshiftnu:
4589 case Intrinsic::arm_neon_vqshiftnsu:
4590 case Intrinsic::arm_neon_vqrshiftns:
4591 case Intrinsic::arm_neon_vqrshiftnu:
4592 case Intrinsic::arm_neon_vqrshiftnsu:
4593 // Narrowing shifts require an immediate right shift.
4594 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4596 llvm_unreachable("invalid shift count for narrowing vector shift "
4600 llvm_unreachable("unhandled vector shift");
4604 case Intrinsic::arm_neon_vshifts:
4605 case Intrinsic::arm_neon_vshiftu:
4606 // Opcode already set above.
4608 case Intrinsic::arm_neon_vshiftls:
4609 case Intrinsic::arm_neon_vshiftlu:
4610 if (Cnt == VT.getVectorElementType().getSizeInBits())
4611 VShiftOpc = ARMISD::VSHLLi;
4613 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4614 ARMISD::VSHLLs : ARMISD::VSHLLu);
4616 case Intrinsic::arm_neon_vshiftn:
4617 VShiftOpc = ARMISD::VSHRN; break;
4618 case Intrinsic::arm_neon_vrshifts:
4619 VShiftOpc = ARMISD::VRSHRs; break;
4620 case Intrinsic::arm_neon_vrshiftu:
4621 VShiftOpc = ARMISD::VRSHRu; break;
4622 case Intrinsic::arm_neon_vrshiftn:
4623 VShiftOpc = ARMISD::VRSHRN; break;
4624 case Intrinsic::arm_neon_vqshifts:
4625 VShiftOpc = ARMISD::VQSHLs; break;
4626 case Intrinsic::arm_neon_vqshiftu:
4627 VShiftOpc = ARMISD::VQSHLu; break;
4628 case Intrinsic::arm_neon_vqshiftsu:
4629 VShiftOpc = ARMISD::VQSHLsu; break;
4630 case Intrinsic::arm_neon_vqshiftns:
4631 VShiftOpc = ARMISD::VQSHRNs; break;
4632 case Intrinsic::arm_neon_vqshiftnu:
4633 VShiftOpc = ARMISD::VQSHRNu; break;
4634 case Intrinsic::arm_neon_vqshiftnsu:
4635 VShiftOpc = ARMISD::VQSHRNsu; break;
4636 case Intrinsic::arm_neon_vqrshiftns:
4637 VShiftOpc = ARMISD::VQRSHRNs; break;
4638 case Intrinsic::arm_neon_vqrshiftnu:
4639 VShiftOpc = ARMISD::VQRSHRNu; break;
4640 case Intrinsic::arm_neon_vqrshiftnsu:
4641 VShiftOpc = ARMISD::VQRSHRNsu; break;
4644 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4645 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4648 case Intrinsic::arm_neon_vshiftins: {
4649 EVT VT = N->getOperand(1).getValueType();
4651 unsigned VShiftOpc = 0;
4653 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4654 VShiftOpc = ARMISD::VSLI;
4655 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4656 VShiftOpc = ARMISD::VSRI;
4658 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4661 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4662 N->getOperand(1), N->getOperand(2),
4663 DAG.getConstant(Cnt, MVT::i32));
4666 case Intrinsic::arm_neon_vqrshifts:
4667 case Intrinsic::arm_neon_vqrshiftu:
4668 // No immediate versions of these to check for.
4675 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4676 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4677 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4678 /// vector element shift counts are generally not legal, and it is hard to see
4679 /// their values after they get legalized to loads from a constant pool.
4680 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4681 const ARMSubtarget *ST) {
4682 EVT VT = N->getValueType(0);
4684 // Nothing to be done for scalar shifts.
4685 if (! VT.isVector())
4688 assert(ST->hasNEON() && "unexpected vector shift");
4691 switch (N->getOpcode()) {
4692 default: llvm_unreachable("unexpected shift opcode");
4695 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4696 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4697 DAG.getConstant(Cnt, MVT::i32));
4702 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4703 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4704 ARMISD::VSHRs : ARMISD::VSHRu);
4705 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4706 DAG.getConstant(Cnt, MVT::i32));
4712 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4713 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4714 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4715 const ARMSubtarget *ST) {
4716 SDValue N0 = N->getOperand(0);
4718 // Check for sign- and zero-extensions of vector extract operations of 8-
4719 // and 16-bit vector elements. NEON supports these directly. They are
4720 // handled during DAG combining because type legalization will promote them
4721 // to 32-bit types and it is messy to recognize the operations after that.
4722 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4723 SDValue Vec = N0.getOperand(0);
4724 SDValue Lane = N0.getOperand(1);
4725 EVT VT = N->getValueType(0);
4726 EVT EltVT = N0.getValueType();
4727 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4729 if (VT == MVT::i32 &&
4730 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4731 TLI.isTypeLegal(Vec.getValueType())) {
4734 switch (N->getOpcode()) {
4735 default: llvm_unreachable("unexpected opcode");
4736 case ISD::SIGN_EXTEND:
4737 Opc = ARMISD::VGETLANEs;
4739 case ISD::ZERO_EXTEND:
4740 case ISD::ANY_EXTEND:
4741 Opc = ARMISD::VGETLANEu;
4744 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4751 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4752 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4753 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4754 const ARMSubtarget *ST) {
4755 // If the target supports NEON, try to use vmax/vmin instructions for f32
4756 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4757 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4758 // a NaN; only do the transformation when it matches that behavior.
4760 // For now only do this when using NEON for FP operations; if using VFP, it
4761 // is not obvious that the benefit outweighs the cost of switching to the
4763 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4764 N->getValueType(0) != MVT::f32)
4767 SDValue CondLHS = N->getOperand(0);
4768 SDValue CondRHS = N->getOperand(1);
4769 SDValue LHS = N->getOperand(2);
4770 SDValue RHS = N->getOperand(3);
4771 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4773 unsigned Opcode = 0;
4775 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4776 IsReversed = false; // x CC y ? x : y
4777 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4778 IsReversed = true ; // x CC y ? y : x
4792 // If LHS is NaN, an ordered comparison will be false and the result will
4793 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4794 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4795 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4796 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4798 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4799 // will return -0, so vmin can only be used for unsafe math or if one of
4800 // the operands is known to be nonzero.
4801 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4803 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4805 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4814 // If LHS is NaN, an ordered comparison will be false and the result will
4815 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4816 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4817 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4818 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4820 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4821 // will return +0, so vmax can only be used for unsafe math or if one of
4822 // the operands is known to be nonzero.
4823 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4825 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4827 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4833 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4836 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4837 DAGCombinerInfo &DCI) const {
4838 switch (N->getOpcode()) {
4840 case ISD::ADD: return PerformADDCombine(N, DCI);
4841 case ISD::SUB: return PerformSUBCombine(N, DCI);
4842 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4843 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4844 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4845 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4846 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4849 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4850 case ISD::SIGN_EXTEND:
4851 case ISD::ZERO_EXTEND:
4852 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4853 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4858 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4859 if (!Subtarget->hasV6Ops())
4860 // Pre-v6 does not support unaligned mem access.
4863 // v6+ may or may not support unaligned mem access depending on the system
4865 // FIXME: This is pretty conservative. Should we provide cmdline option to
4866 // control the behaviour?
4867 if (!Subtarget->isTargetDarwin())
4870 switch (VT.getSimpleVT().SimpleTy) {
4877 // FIXME: VLD1 etc with standard alignment is legal.
4881 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4886 switch (VT.getSimpleVT().SimpleTy) {
4887 default: return false;
4902 if ((V & (Scale - 1)) != 0)
4905 return V == (V & ((1LL << 5) - 1));
4908 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4909 const ARMSubtarget *Subtarget) {
4916 switch (VT.getSimpleVT().SimpleTy) {
4917 default: return false;
4922 // + imm12 or - imm8
4924 return V == (V & ((1LL << 8) - 1));
4925 return V == (V & ((1LL << 12) - 1));
4928 // Same as ARM mode. FIXME: NEON?
4929 if (!Subtarget->hasVFP2())
4934 return V == (V & ((1LL << 8) - 1));
4938 /// isLegalAddressImmediate - Return true if the integer value can be used
4939 /// as the offset of the target addressing mode for load / store of the
4941 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4942 const ARMSubtarget *Subtarget) {
4949 if (Subtarget->isThumb1Only())
4950 return isLegalT1AddressImmediate(V, VT);
4951 else if (Subtarget->isThumb2())
4952 return isLegalT2AddressImmediate(V, VT, Subtarget);
4957 switch (VT.getSimpleVT().SimpleTy) {
4958 default: return false;
4963 return V == (V & ((1LL << 12) - 1));
4966 return V == (V & ((1LL << 8) - 1));
4969 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4974 return V == (V & ((1LL << 8) - 1));
4978 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4980 int Scale = AM.Scale;
4984 switch (VT.getSimpleVT().SimpleTy) {
4985 default: return false;
4994 return Scale == 2 || Scale == 4 || Scale == 8;
4997 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5001 // Note, we allow "void" uses (basically, uses that aren't loads or
5002 // stores), because arm allows folding a scale into many arithmetic
5003 // operations. This should be made more precise and revisited later.
5005 // Allow r << imm, but the imm has to be a multiple of two.
5006 if (Scale & 1) return false;
5007 return isPowerOf2_32(Scale);
5011 /// isLegalAddressingMode - Return true if the addressing mode represented
5012 /// by AM is legal for this target, for a load/store of the specified type.
5013 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5014 const Type *Ty) const {
5015 EVT VT = getValueType(Ty, true);
5016 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5019 // Can never fold addr of global into load/store.
5024 case 0: // no scale reg, must be "r+i" or "r", or "i".
5027 if (Subtarget->isThumb1Only())
5031 // ARM doesn't support any R+R*scale+imm addr modes.
5038 if (Subtarget->isThumb2())
5039 return isLegalT2ScaledAddressingMode(AM, VT);
5041 int Scale = AM.Scale;
5042 switch (VT.getSimpleVT().SimpleTy) {
5043 default: return false;
5047 if (Scale < 0) Scale = -Scale;
5051 return isPowerOf2_32(Scale & ~1);
5055 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5060 // Note, we allow "void" uses (basically, uses that aren't loads or
5061 // stores), because arm allows folding a scale into many arithmetic
5062 // operations. This should be made more precise and revisited later.
5064 // Allow r << imm, but the imm has to be a multiple of two.
5065 if (Scale & 1) return false;
5066 return isPowerOf2_32(Scale);
5073 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5074 /// icmp immediate, that is the target has icmp instructions which can compare
5075 /// a register against the immediate without having to materialize the
5076 /// immediate into a register.
5077 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5078 if (!Subtarget->isThumb())
5079 return ARM_AM::getSOImmVal(Imm) != -1;
5080 if (Subtarget->isThumb2())
5081 return ARM_AM::getT2SOImmVal(Imm) != -1;
5082 return Imm >= 0 && Imm <= 255;
5085 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5086 bool isSEXTLoad, SDValue &Base,
5087 SDValue &Offset, bool &isInc,
5088 SelectionDAG &DAG) {
5089 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5092 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5094 Base = Ptr->getOperand(0);
5095 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5096 int RHSC = (int)RHS->getZExtValue();
5097 if (RHSC < 0 && RHSC > -256) {
5098 assert(Ptr->getOpcode() == ISD::ADD);
5100 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5104 isInc = (Ptr->getOpcode() == ISD::ADD);
5105 Offset = Ptr->getOperand(1);
5107 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5109 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5110 int RHSC = (int)RHS->getZExtValue();
5111 if (RHSC < 0 && RHSC > -0x1000) {
5112 assert(Ptr->getOpcode() == ISD::ADD);
5114 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5115 Base = Ptr->getOperand(0);
5120 if (Ptr->getOpcode() == ISD::ADD) {
5122 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5123 if (ShOpcVal != ARM_AM::no_shift) {
5124 Base = Ptr->getOperand(1);
5125 Offset = Ptr->getOperand(0);
5127 Base = Ptr->getOperand(0);
5128 Offset = Ptr->getOperand(1);
5133 isInc = (Ptr->getOpcode() == ISD::ADD);
5134 Base = Ptr->getOperand(0);
5135 Offset = Ptr->getOperand(1);
5139 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5143 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5144 bool isSEXTLoad, SDValue &Base,
5145 SDValue &Offset, bool &isInc,
5146 SelectionDAG &DAG) {
5147 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5150 Base = Ptr->getOperand(0);
5151 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5152 int RHSC = (int)RHS->getZExtValue();
5153 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5154 assert(Ptr->getOpcode() == ISD::ADD);
5156 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5158 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5159 isInc = Ptr->getOpcode() == ISD::ADD;
5160 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5168 /// getPreIndexedAddressParts - returns true by value, base pointer and
5169 /// offset pointer and addressing mode by reference if the node's address
5170 /// can be legally represented as pre-indexed load / store address.
5172 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5174 ISD::MemIndexedMode &AM,
5175 SelectionDAG &DAG) const {
5176 if (Subtarget->isThumb1Only())
5181 bool isSEXTLoad = false;
5182 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5183 Ptr = LD->getBasePtr();
5184 VT = LD->getMemoryVT();
5185 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5186 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5187 Ptr = ST->getBasePtr();
5188 VT = ST->getMemoryVT();
5193 bool isLegal = false;
5194 if (Subtarget->isThumb2())
5195 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5196 Offset, isInc, DAG);
5198 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5199 Offset, isInc, DAG);
5203 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5207 /// getPostIndexedAddressParts - returns true by value, base pointer and
5208 /// offset pointer and addressing mode by reference if this node can be
5209 /// combined with a load / store to form a post-indexed load / store.
5210 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5213 ISD::MemIndexedMode &AM,
5214 SelectionDAG &DAG) const {
5215 if (Subtarget->isThumb1Only())
5220 bool isSEXTLoad = false;
5221 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5222 VT = LD->getMemoryVT();
5223 Ptr = LD->getBasePtr();
5224 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5225 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5226 VT = ST->getMemoryVT();
5227 Ptr = ST->getBasePtr();
5232 bool isLegal = false;
5233 if (Subtarget->isThumb2())
5234 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5237 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5243 // Swap base ptr and offset to catch more post-index load / store when
5244 // it's legal. In Thumb2 mode, offset must be an immediate.
5245 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5246 !Subtarget->isThumb2())
5247 std::swap(Base, Offset);
5249 // Post-indexed load / store update the base pointer.
5254 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5258 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5262 const SelectionDAG &DAG,
5263 unsigned Depth) const {
5264 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5265 switch (Op.getOpcode()) {
5267 case ARMISD::CMOV: {
5268 // Bits are known zero/one if known on the LHS and RHS.
5269 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5270 if (KnownZero == 0 && KnownOne == 0) return;
5272 APInt KnownZeroRHS, KnownOneRHS;
5273 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5274 KnownZeroRHS, KnownOneRHS, Depth+1);
5275 KnownZero &= KnownZeroRHS;
5276 KnownOne &= KnownOneRHS;
5282 //===----------------------------------------------------------------------===//
5283 // ARM Inline Assembly Support
5284 //===----------------------------------------------------------------------===//
5286 /// getConstraintType - Given a constraint letter, return the type of
5287 /// constraint it is for this target.
5288 ARMTargetLowering::ConstraintType
5289 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5290 if (Constraint.size() == 1) {
5291 switch (Constraint[0]) {
5293 case 'l': return C_RegisterClass;
5294 case 'w': return C_RegisterClass;
5297 return TargetLowering::getConstraintType(Constraint);
5300 std::pair<unsigned, const TargetRegisterClass*>
5301 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5303 if (Constraint.size() == 1) {
5304 // GCC ARM Constraint Letters
5305 switch (Constraint[0]) {
5307 if (Subtarget->isThumb())
5308 return std::make_pair(0U, ARM::tGPRRegisterClass);
5310 return std::make_pair(0U, ARM::GPRRegisterClass);
5312 return std::make_pair(0U, ARM::GPRRegisterClass);
5315 return std::make_pair(0U, ARM::SPRRegisterClass);
5316 if (VT.getSizeInBits() == 64)
5317 return std::make_pair(0U, ARM::DPRRegisterClass);
5318 if (VT.getSizeInBits() == 128)
5319 return std::make_pair(0U, ARM::QPRRegisterClass);
5323 if (StringRef("{cc}").equals_lower(Constraint))
5324 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5326 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5329 std::vector<unsigned> ARMTargetLowering::
5330 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5332 if (Constraint.size() != 1)
5333 return std::vector<unsigned>();
5335 switch (Constraint[0]) { // GCC ARM Constraint Letters
5338 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5339 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5342 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5343 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5344 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5345 ARM::R12, ARM::LR, 0);
5348 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5349 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5350 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5351 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5352 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5353 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5354 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5355 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5356 if (VT.getSizeInBits() == 64)
5357 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5358 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5359 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5360 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5361 if (VT.getSizeInBits() == 128)
5362 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5363 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5367 return std::vector<unsigned>();
5370 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5371 /// vector. If it is invalid, don't add anything to Ops.
5372 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5374 std::vector<SDValue>&Ops,
5375 SelectionDAG &DAG) const {
5376 SDValue Result(0, 0);
5378 switch (Constraint) {
5380 case 'I': case 'J': case 'K': case 'L':
5381 case 'M': case 'N': case 'O':
5382 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5386 int64_t CVal64 = C->getSExtValue();
5387 int CVal = (int) CVal64;
5388 // None of these constraints allow values larger than 32 bits. Check
5389 // that the value fits in an int.
5393 switch (Constraint) {
5395 if (Subtarget->isThumb1Only()) {
5396 // This must be a constant between 0 and 255, for ADD
5398 if (CVal >= 0 && CVal <= 255)
5400 } else if (Subtarget->isThumb2()) {
5401 // A constant that can be used as an immediate value in a
5402 // data-processing instruction.
5403 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5406 // A constant that can be used as an immediate value in a
5407 // data-processing instruction.
5408 if (ARM_AM::getSOImmVal(CVal) != -1)
5414 if (Subtarget->isThumb()) { // FIXME thumb2
5415 // This must be a constant between -255 and -1, for negated ADD
5416 // immediates. This can be used in GCC with an "n" modifier that
5417 // prints the negated value, for use with SUB instructions. It is
5418 // not useful otherwise but is implemented for compatibility.
5419 if (CVal >= -255 && CVal <= -1)
5422 // This must be a constant between -4095 and 4095. It is not clear
5423 // what this constraint is intended for. Implemented for
5424 // compatibility with GCC.
5425 if (CVal >= -4095 && CVal <= 4095)
5431 if (Subtarget->isThumb1Only()) {
5432 // A 32-bit value where only one byte has a nonzero value. Exclude
5433 // zero to match GCC. This constraint is used by GCC internally for
5434 // constants that can be loaded with a move/shift combination.
5435 // It is not useful otherwise but is implemented for compatibility.
5436 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5438 } else if (Subtarget->isThumb2()) {
5439 // A constant whose bitwise inverse can be used as an immediate
5440 // value in a data-processing instruction. This can be used in GCC
5441 // with a "B" modifier that prints the inverted value, for use with
5442 // BIC and MVN instructions. It is not useful otherwise but is
5443 // implemented for compatibility.
5444 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5447 // A constant whose bitwise inverse can be used as an immediate
5448 // value in a data-processing instruction. This can be used in GCC
5449 // with a "B" modifier that prints the inverted value, for use with
5450 // BIC and MVN instructions. It is not useful otherwise but is
5451 // implemented for compatibility.
5452 if (ARM_AM::getSOImmVal(~CVal) != -1)
5458 if (Subtarget->isThumb1Only()) {
5459 // This must be a constant between -7 and 7,
5460 // for 3-operand ADD/SUB immediate instructions.
5461 if (CVal >= -7 && CVal < 7)
5463 } else if (Subtarget->isThumb2()) {
5464 // A constant whose negation can be used as an immediate value in a
5465 // data-processing instruction. This can be used in GCC with an "n"
5466 // modifier that prints the negated value, for use with SUB
5467 // instructions. It is not useful otherwise but is implemented for
5469 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5472 // A constant whose negation can be used as an immediate value in a
5473 // data-processing instruction. This can be used in GCC with an "n"
5474 // modifier that prints the negated value, for use with SUB
5475 // instructions. It is not useful otherwise but is implemented for
5477 if (ARM_AM::getSOImmVal(-CVal) != -1)
5483 if (Subtarget->isThumb()) { // FIXME thumb2
5484 // This must be a multiple of 4 between 0 and 1020, for
5485 // ADD sp + immediate.
5486 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5489 // A power of two or a constant between 0 and 32. This is used in
5490 // GCC for the shift amount on shifted register operands, but it is
5491 // useful in general for any shift amounts.
5492 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5498 if (Subtarget->isThumb()) { // FIXME thumb2
5499 // This must be a constant between 0 and 31, for shift amounts.
5500 if (CVal >= 0 && CVal <= 31)
5506 if (Subtarget->isThumb()) { // FIXME thumb2
5507 // This must be a multiple of 4 between -508 and 508, for
5508 // ADD/SUB sp = sp + immediate.
5509 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5514 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5518 if (Result.getNode()) {
5519 Ops.push_back(Result);
5522 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5526 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5527 // The ARM target isn't yet aware of offsets.
5531 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5532 APInt Imm = FPImm.bitcastToAPInt();
5533 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5534 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5535 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5537 // We can handle 4 bits of mantissa.
5538 // mantissa = (16+UInt(e:f:g:h))/16.
5539 if (Mantissa & 0x7ffff)
5542 if ((Mantissa & 0xf) != Mantissa)
5545 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5546 if (Exp < -3 || Exp > 4)
5548 Exp = ((Exp+3) & 0x7) ^ 4;
5550 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5553 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5554 APInt Imm = FPImm.bitcastToAPInt();
5555 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5556 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5557 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5559 // We can handle 4 bits of mantissa.
5560 // mantissa = (16+UInt(e:f:g:h))/16.
5561 if (Mantissa & 0xffffffffffffLL)
5564 if ((Mantissa & 0xf) != Mantissa)
5567 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5568 if (Exp < -3 || Exp > 4)
5570 Exp = ((Exp+3) & 0x7) ^ 4;
5572 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5575 bool ARM::isBitFieldInvertedMask(unsigned v) {
5576 if (v == 0xffffffff)
5578 // there can be 1's on either or both "outsides", all the "inside"
5580 unsigned int lsb = 0, msb = 31;
5581 while (v & (1 << msb)) --msb;
5582 while (v & (1 << lsb)) ++lsb;
5583 for (unsigned int i = lsb; i <= msb; ++i) {
5590 /// isFPImmLegal - Returns true if the target can instruction select the
5591 /// specified FP immediate natively. If false, the legalizer will
5592 /// materialize the FP immediate as a load from a constant pool.
5593 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5594 if (!Subtarget->hasVFP3())
5597 return ARM::getVFPf32Imm(Imm) != -1;
5599 return ARM::getVFPf64Imm(Imm) != -1;