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 if (!Subtarget->isFPOnlySP())
276 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
278 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
281 if (Subtarget->hasNEON()) {
282 addDRTypeForNEON(MVT::v2f32);
283 addDRTypeForNEON(MVT::v8i8);
284 addDRTypeForNEON(MVT::v4i16);
285 addDRTypeForNEON(MVT::v2i32);
286 addDRTypeForNEON(MVT::v1i64);
288 addQRTypeForNEON(MVT::v4f32);
289 addQRTypeForNEON(MVT::v2f64);
290 addQRTypeForNEON(MVT::v16i8);
291 addQRTypeForNEON(MVT::v8i16);
292 addQRTypeForNEON(MVT::v4i32);
293 addQRTypeForNEON(MVT::v2i64);
295 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
296 // neither Neon nor VFP support any arithmetic operations on it.
297 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
298 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
299 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
300 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
301 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
302 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
303 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
304 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
305 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
306 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
307 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
308 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
309 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
310 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
311 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
312 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
313 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
314 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
315 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
316 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
317 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
318 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
319 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
320 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
322 // Neon does not support some operations on v1i64 and v2i64 types.
323 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
324 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
325 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
326 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
328 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
329 setTargetDAGCombine(ISD::SHL);
330 setTargetDAGCombine(ISD::SRL);
331 setTargetDAGCombine(ISD::SRA);
332 setTargetDAGCombine(ISD::SIGN_EXTEND);
333 setTargetDAGCombine(ISD::ZERO_EXTEND);
334 setTargetDAGCombine(ISD::ANY_EXTEND);
335 setTargetDAGCombine(ISD::SELECT_CC);
338 computeRegisterProperties();
340 // ARM does not have f32 extending load.
341 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
343 // ARM does not have i1 sign extending load.
344 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
346 // ARM supports all 4 flavors of integer indexed load / store.
347 if (!Subtarget->isThumb1Only()) {
348 for (unsigned im = (unsigned)ISD::PRE_INC;
349 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
350 setIndexedLoadAction(im, MVT::i1, Legal);
351 setIndexedLoadAction(im, MVT::i8, Legal);
352 setIndexedLoadAction(im, MVT::i16, Legal);
353 setIndexedLoadAction(im, MVT::i32, Legal);
354 setIndexedStoreAction(im, MVT::i1, Legal);
355 setIndexedStoreAction(im, MVT::i8, Legal);
356 setIndexedStoreAction(im, MVT::i16, Legal);
357 setIndexedStoreAction(im, MVT::i32, Legal);
361 // i64 operation support.
362 if (Subtarget->isThumb1Only()) {
363 setOperationAction(ISD::MUL, MVT::i64, Expand);
364 setOperationAction(ISD::MULHU, MVT::i32, Expand);
365 setOperationAction(ISD::MULHS, MVT::i32, Expand);
366 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
367 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
369 setOperationAction(ISD::MUL, MVT::i64, Expand);
370 setOperationAction(ISD::MULHU, MVT::i32, Expand);
371 if (!Subtarget->hasV6Ops())
372 setOperationAction(ISD::MULHS, MVT::i32, Expand);
374 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
375 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
376 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
377 setOperationAction(ISD::SRL, MVT::i64, Custom);
378 setOperationAction(ISD::SRA, MVT::i64, Custom);
380 // ARM does not have ROTL.
381 setOperationAction(ISD::ROTL, MVT::i32, Expand);
382 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
383 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
384 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
385 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
387 // Only ARMv6 has BSWAP.
388 if (!Subtarget->hasV6Ops())
389 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
391 // These are expanded into libcalls.
392 if (!Subtarget->hasDivide()) {
393 // v7M has a hardware divider
394 setOperationAction(ISD::SDIV, MVT::i32, Expand);
395 setOperationAction(ISD::UDIV, MVT::i32, Expand);
397 setOperationAction(ISD::SREM, MVT::i32, Expand);
398 setOperationAction(ISD::UREM, MVT::i32, Expand);
399 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
400 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
402 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
403 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
404 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
405 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
406 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
408 setOperationAction(ISD::TRAP, MVT::Other, Legal);
410 // Use the default implementation.
411 setOperationAction(ISD::VASTART, MVT::Other, Custom);
412 setOperationAction(ISD::VAARG, MVT::Other, Expand);
413 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
414 setOperationAction(ISD::VAEND, MVT::Other, Expand);
415 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
416 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
417 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
418 // FIXME: Shouldn't need this, since no register is used, but the legalizer
419 // doesn't yet know how to not do that for SjLj.
420 setExceptionSelectorRegister(ARM::R0);
421 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
422 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
423 // the default expansion.
424 if (Subtarget->hasDataBarrier() ||
425 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
426 // membarrier needs custom lowering; the rest are legal and handled
428 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
430 // Set them all for expansion, which will force libcalls.
431 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
432 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
433 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
434 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
435 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
436 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
437 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
449 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
450 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
451 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
452 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
453 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
454 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
455 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
456 // Since the libcalls include locking, fold in the fences
457 setShouldFoldAtomicFences(true);
459 // 64-bit versions are always libcalls (for now)
460 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
461 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
462 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
463 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
464 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
465 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
466 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
467 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
469 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
470 if (!Subtarget->hasV6Ops()) {
471 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
472 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
474 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
476 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
477 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
478 // iff target supports vfp2.
479 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
480 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
483 // We want to custom lower some of our intrinsics.
484 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
485 if (Subtarget->isTargetDarwin()) {
486 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
487 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
490 setOperationAction(ISD::SETCC, MVT::i32, Expand);
491 setOperationAction(ISD::SETCC, MVT::f32, Expand);
492 setOperationAction(ISD::SETCC, MVT::f64, Expand);
493 setOperationAction(ISD::SELECT, MVT::i32, Custom);
494 setOperationAction(ISD::SELECT, MVT::f32, Custom);
495 setOperationAction(ISD::SELECT, MVT::f64, Custom);
496 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
497 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
498 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
500 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
501 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
502 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
503 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
504 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
506 // We don't support sin/cos/fmod/copysign/pow
507 setOperationAction(ISD::FSIN, MVT::f64, Expand);
508 setOperationAction(ISD::FSIN, MVT::f32, Expand);
509 setOperationAction(ISD::FCOS, MVT::f32, Expand);
510 setOperationAction(ISD::FCOS, MVT::f64, Expand);
511 setOperationAction(ISD::FREM, MVT::f64, Expand);
512 setOperationAction(ISD::FREM, MVT::f32, Expand);
513 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
514 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
515 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
517 setOperationAction(ISD::FPOW, MVT::f64, Expand);
518 setOperationAction(ISD::FPOW, MVT::f32, Expand);
520 // Various VFP goodness
521 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
522 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
523 if (Subtarget->hasVFP2()) {
524 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
525 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
526 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
527 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
529 // Special handling for half-precision FP.
530 if (!Subtarget->hasFP16()) {
531 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
532 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
536 // We have target-specific dag combine patterns for the following nodes:
537 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
538 setTargetDAGCombine(ISD::ADD);
539 setTargetDAGCombine(ISD::SUB);
540 setTargetDAGCombine(ISD::MUL);
542 if (Subtarget->hasV6T2Ops())
543 setTargetDAGCombine(ISD::OR);
545 setStackPointerRegisterToSaveRestore(ARM::SP);
547 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
548 setSchedulingPreference(Sched::RegPressure);
550 setSchedulingPreference(Sched::Hybrid);
552 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
554 // On ARM arguments smaller than 4 bytes are extended, so all arguments
555 // are at least 4 bytes aligned.
556 setMinStackArgumentAlignment(4);
558 if (EnableARMCodePlacement)
559 benefitFromCodePlacementOpt = true;
562 std::pair<const TargetRegisterClass*, uint8_t>
563 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
564 const TargetRegisterClass *RRC = 0;
566 switch (VT.getSimpleVT().SimpleTy) {
568 return TargetLowering::findRepresentativeClass(VT);
569 // Use DPR as representative register class for all floating point
570 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
571 // the cost is 1 for both f32 and f64.
572 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
573 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
574 RRC = ARM::DPRRegisterClass;
576 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
577 case MVT::v4f32: case MVT::v2f64:
578 RRC = ARM::DPRRegisterClass;
582 RRC = ARM::DPRRegisterClass;
586 RRC = ARM::DPRRegisterClass;
590 return std::make_pair(RRC, Cost);
593 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
596 case ARMISD::Wrapper: return "ARMISD::Wrapper";
597 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
598 case ARMISD::CALL: return "ARMISD::CALL";
599 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
600 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
601 case ARMISD::tCALL: return "ARMISD::tCALL";
602 case ARMISD::BRCOND: return "ARMISD::BRCOND";
603 case ARMISD::BR_JT: return "ARMISD::BR_JT";
604 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
605 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
606 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
607 case ARMISD::CMP: return "ARMISD::CMP";
608 case ARMISD::CMPZ: return "ARMISD::CMPZ";
609 case ARMISD::CMPFP: return "ARMISD::CMPFP";
610 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
611 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
612 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
613 case ARMISD::CMOV: return "ARMISD::CMOV";
614 case ARMISD::CNEG: return "ARMISD::CNEG";
616 case ARMISD::RBIT: return "ARMISD::RBIT";
618 case ARMISD::FTOSI: return "ARMISD::FTOSI";
619 case ARMISD::FTOUI: return "ARMISD::FTOUI";
620 case ARMISD::SITOF: return "ARMISD::SITOF";
621 case ARMISD::UITOF: return "ARMISD::UITOF";
623 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
624 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
625 case ARMISD::RRX: return "ARMISD::RRX";
627 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
628 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
630 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
631 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
633 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
635 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
637 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
639 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
640 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
642 case ARMISD::VCEQ: return "ARMISD::VCEQ";
643 case ARMISD::VCGE: return "ARMISD::VCGE";
644 case ARMISD::VCGEU: return "ARMISD::VCGEU";
645 case ARMISD::VCGT: return "ARMISD::VCGT";
646 case ARMISD::VCGTU: return "ARMISD::VCGTU";
647 case ARMISD::VTST: return "ARMISD::VTST";
649 case ARMISD::VSHL: return "ARMISD::VSHL";
650 case ARMISD::VSHRs: return "ARMISD::VSHRs";
651 case ARMISD::VSHRu: return "ARMISD::VSHRu";
652 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
653 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
654 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
655 case ARMISD::VSHRN: return "ARMISD::VSHRN";
656 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
657 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
658 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
659 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
660 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
661 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
662 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
663 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
664 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
665 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
666 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
667 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
668 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
669 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
670 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
671 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
672 case ARMISD::VDUP: return "ARMISD::VDUP";
673 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
674 case ARMISD::VEXT: return "ARMISD::VEXT";
675 case ARMISD::VREV64: return "ARMISD::VREV64";
676 case ARMISD::VREV32: return "ARMISD::VREV32";
677 case ARMISD::VREV16: return "ARMISD::VREV16";
678 case ARMISD::VZIP: return "ARMISD::VZIP";
679 case ARMISD::VUZP: return "ARMISD::VUZP";
680 case ARMISD::VTRN: return "ARMISD::VTRN";
681 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
682 case ARMISD::FMAX: return "ARMISD::FMAX";
683 case ARMISD::FMIN: return "ARMISD::FMIN";
684 case ARMISD::BFI: return "ARMISD::BFI";
688 /// getRegClassFor - Return the register class that should be used for the
689 /// specified value type.
690 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
691 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
692 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
693 // load / store 4 to 8 consecutive D registers.
694 if (Subtarget->hasNEON()) {
695 if (VT == MVT::v4i64)
696 return ARM::QQPRRegisterClass;
697 else if (VT == MVT::v8i64)
698 return ARM::QQQQPRRegisterClass;
700 return TargetLowering::getRegClassFor(VT);
703 // Create a fast isel object.
705 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
706 return ARM::createFastISel(funcInfo);
709 /// getFunctionAlignment - Return the Log2 alignment of this function.
710 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
711 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
714 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
715 /// be used for loads / stores from the global.
716 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
717 return (Subtarget->isThumb1Only() ? 127 : 4095);
720 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
721 unsigned NumVals = N->getNumValues();
723 return Sched::RegPressure;
725 for (unsigned i = 0; i != NumVals; ++i) {
726 EVT VT = N->getValueType(i);
727 if (VT.isFloatingPoint() || VT.isVector())
728 return Sched::Latency;
731 if (!N->isMachineOpcode())
732 return Sched::RegPressure;
734 // Load are scheduled for latency even if there instruction itinerary
736 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
737 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
739 return Sched::Latency;
741 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
742 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
743 return Sched::Latency;
744 return Sched::RegPressure;
748 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
749 MachineFunction &MF) const {
750 switch (RC->getID()) {
753 case ARM::tGPRRegClassID:
754 return RegInfo->hasFP(MF) ? 4 : 5;
755 case ARM::GPRRegClassID: {
756 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
757 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
759 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
760 case ARM::DPRRegClassID:
765 //===----------------------------------------------------------------------===//
767 //===----------------------------------------------------------------------===//
769 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
770 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
772 default: llvm_unreachable("Unknown condition code!");
773 case ISD::SETNE: return ARMCC::NE;
774 case ISD::SETEQ: return ARMCC::EQ;
775 case ISD::SETGT: return ARMCC::GT;
776 case ISD::SETGE: return ARMCC::GE;
777 case ISD::SETLT: return ARMCC::LT;
778 case ISD::SETLE: return ARMCC::LE;
779 case ISD::SETUGT: return ARMCC::HI;
780 case ISD::SETUGE: return ARMCC::HS;
781 case ISD::SETULT: return ARMCC::LO;
782 case ISD::SETULE: return ARMCC::LS;
786 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
787 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
788 ARMCC::CondCodes &CondCode2) {
789 CondCode2 = ARMCC::AL;
791 default: llvm_unreachable("Unknown FP condition!");
793 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
795 case ISD::SETOGT: CondCode = ARMCC::GT; break;
797 case ISD::SETOGE: CondCode = ARMCC::GE; break;
798 case ISD::SETOLT: CondCode = ARMCC::MI; break;
799 case ISD::SETOLE: CondCode = ARMCC::LS; break;
800 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
801 case ISD::SETO: CondCode = ARMCC::VC; break;
802 case ISD::SETUO: CondCode = ARMCC::VS; break;
803 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
804 case ISD::SETUGT: CondCode = ARMCC::HI; break;
805 case ISD::SETUGE: CondCode = ARMCC::PL; break;
807 case ISD::SETULT: CondCode = ARMCC::LT; break;
809 case ISD::SETULE: CondCode = ARMCC::LE; break;
811 case ISD::SETUNE: CondCode = ARMCC::NE; break;
815 //===----------------------------------------------------------------------===//
816 // Calling Convention Implementation
817 //===----------------------------------------------------------------------===//
819 #include "ARMGenCallingConv.inc"
821 // APCS f64 is in register pairs, possibly split to stack
822 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
823 CCValAssign::LocInfo &LocInfo,
824 CCState &State, bool CanFail) {
825 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
827 // Try to get the first register.
828 if (unsigned Reg = State.AllocateReg(RegList, 4))
829 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
831 // For the 2nd half of a v2f64, do not fail.
835 // Put the whole thing on the stack.
836 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
837 State.AllocateStack(8, 4),
842 // Try to get the second register.
843 if (unsigned Reg = State.AllocateReg(RegList, 4))
844 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
846 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
847 State.AllocateStack(4, 4),
852 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
853 CCValAssign::LocInfo &LocInfo,
854 ISD::ArgFlagsTy &ArgFlags,
856 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
858 if (LocVT == MVT::v2f64 &&
859 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
861 return true; // we handled it
864 // AAPCS f64 is in aligned register pairs
865 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
866 CCValAssign::LocInfo &LocInfo,
867 CCState &State, bool CanFail) {
868 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
869 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
870 static const unsigned ShadowRegList[] = { ARM::R0, ARM::R1 };
872 unsigned Reg = State.AllocateReg(HiRegList, ShadowRegList, 2);
874 // For the 2nd half of a v2f64, do not just fail.
878 // Put the whole thing on the stack.
879 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
880 State.AllocateStack(8, 8),
886 for (i = 0; i < 2; ++i)
887 if (HiRegList[i] == Reg)
890 unsigned T = State.AllocateReg(LoRegList[i]);
892 assert(T == LoRegList[i] && "Could not allocate register");
894 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
895 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
900 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
901 CCValAssign::LocInfo &LocInfo,
902 ISD::ArgFlagsTy &ArgFlags,
904 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
906 if (LocVT == MVT::v2f64 &&
907 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
909 return true; // we handled it
912 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
913 CCValAssign::LocInfo &LocInfo, CCState &State) {
914 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
915 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
917 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
919 return false; // we didn't handle it
922 for (i = 0; i < 2; ++i)
923 if (HiRegList[i] == Reg)
926 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
927 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
932 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
933 CCValAssign::LocInfo &LocInfo,
934 ISD::ArgFlagsTy &ArgFlags,
936 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
938 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
940 return true; // we handled it
943 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
944 CCValAssign::LocInfo &LocInfo,
945 ISD::ArgFlagsTy &ArgFlags,
947 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
951 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
952 /// given CallingConvention value.
953 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
955 bool isVarArg) const {
958 llvm_unreachable("Unsupported calling convention");
960 case CallingConv::Fast:
961 // Use target triple & subtarget features to do actual dispatch.
962 if (Subtarget->isAAPCS_ABI()) {
963 if (Subtarget->hasVFP2() &&
964 FloatABIType == FloatABI::Hard && !isVarArg)
965 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
967 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
969 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
970 case CallingConv::ARM_AAPCS_VFP:
971 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
972 case CallingConv::ARM_AAPCS:
973 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
974 case CallingConv::ARM_APCS:
975 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
979 /// LowerCallResult - Lower the result values of a call into the
980 /// appropriate copies out of appropriate physical registers.
982 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
983 CallingConv::ID CallConv, bool isVarArg,
984 const SmallVectorImpl<ISD::InputArg> &Ins,
985 DebugLoc dl, SelectionDAG &DAG,
986 SmallVectorImpl<SDValue> &InVals) const {
988 // Assign locations to each value returned by this call.
989 SmallVector<CCValAssign, 16> RVLocs;
990 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
991 RVLocs, *DAG.getContext());
992 CCInfo.AnalyzeCallResult(Ins,
993 CCAssignFnForNode(CallConv, /* Return*/ true,
996 // Copy all of the result registers out of their specified physreg.
997 for (unsigned i = 0; i != RVLocs.size(); ++i) {
998 CCValAssign VA = RVLocs[i];
1001 if (VA.needsCustom()) {
1002 // Handle f64 or half of a v2f64.
1003 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1005 Chain = Lo.getValue(1);
1006 InFlag = Lo.getValue(2);
1007 VA = RVLocs[++i]; // skip ahead to next loc
1008 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1010 Chain = Hi.getValue(1);
1011 InFlag = Hi.getValue(2);
1012 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1014 if (VA.getLocVT() == MVT::v2f64) {
1015 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1016 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1017 DAG.getConstant(0, MVT::i32));
1019 VA = RVLocs[++i]; // skip ahead to next loc
1020 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1021 Chain = Lo.getValue(1);
1022 InFlag = Lo.getValue(2);
1023 VA = RVLocs[++i]; // skip ahead to next loc
1024 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1025 Chain = Hi.getValue(1);
1026 InFlag = Hi.getValue(2);
1027 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1028 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1029 DAG.getConstant(1, MVT::i32));
1032 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1034 Chain = Val.getValue(1);
1035 InFlag = Val.getValue(2);
1038 switch (VA.getLocInfo()) {
1039 default: llvm_unreachable("Unknown loc info!");
1040 case CCValAssign::Full: break;
1041 case CCValAssign::BCvt:
1042 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1046 InVals.push_back(Val);
1052 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1053 /// by "Src" to address "Dst" of size "Size". Alignment information is
1054 /// specified by the specific parameter attribute. The copy will be passed as
1055 /// a byval function parameter.
1056 /// Sometimes what we are copying is the end of a larger object, the part that
1057 /// does not fit in registers.
1059 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1060 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1062 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1063 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1064 /*isVolatile=*/false, /*AlwaysInline=*/false,
1068 /// LowerMemOpCallTo - Store the argument to the stack.
1070 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1071 SDValue StackPtr, SDValue Arg,
1072 DebugLoc dl, SelectionDAG &DAG,
1073 const CCValAssign &VA,
1074 ISD::ArgFlagsTy Flags) const {
1075 unsigned LocMemOffset = VA.getLocMemOffset();
1076 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1077 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1078 if (Flags.isByVal()) {
1079 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1081 return DAG.getStore(Chain, dl, Arg, PtrOff,
1082 PseudoSourceValue::getStack(), LocMemOffset,
1086 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1087 SDValue Chain, SDValue &Arg,
1088 RegsToPassVector &RegsToPass,
1089 CCValAssign &VA, CCValAssign &NextVA,
1091 SmallVector<SDValue, 8> &MemOpChains,
1092 ISD::ArgFlagsTy Flags) const {
1094 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1095 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1096 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1098 if (NextVA.isRegLoc())
1099 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1101 assert(NextVA.isMemLoc());
1102 if (StackPtr.getNode() == 0)
1103 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1105 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1111 /// LowerCall - Lowering a call into a callseq_start <-
1112 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1115 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1116 CallingConv::ID CallConv, bool isVarArg,
1118 const SmallVectorImpl<ISD::OutputArg> &Outs,
1119 const SmallVectorImpl<SDValue> &OutVals,
1120 const SmallVectorImpl<ISD::InputArg> &Ins,
1121 DebugLoc dl, SelectionDAG &DAG,
1122 SmallVectorImpl<SDValue> &InVals) const {
1123 MachineFunction &MF = DAG.getMachineFunction();
1124 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1125 bool IsSibCall = false;
1126 // Temporarily disable tail calls so things don't break.
1127 if (!EnableARMTailCalls)
1130 // Check if it's really possible to do a tail call.
1131 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1132 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1133 Outs, OutVals, Ins, DAG);
1134 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1135 // detected sibcalls.
1142 // Analyze operands of the call, assigning locations to each operand.
1143 SmallVector<CCValAssign, 16> ArgLocs;
1144 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1146 CCInfo.AnalyzeCallOperands(Outs,
1147 CCAssignFnForNode(CallConv, /* Return*/ false,
1150 // Get a count of how many bytes are to be pushed on the stack.
1151 unsigned NumBytes = CCInfo.getNextStackOffset();
1153 // For tail calls, memory operands are available in our caller's stack.
1157 // Adjust the stack pointer for the new arguments...
1158 // These operations are automatically eliminated by the prolog/epilog pass
1160 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1162 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1164 RegsToPassVector RegsToPass;
1165 SmallVector<SDValue, 8> MemOpChains;
1167 // Walk the register/memloc assignments, inserting copies/loads. In the case
1168 // of tail call optimization, arguments are handled later.
1169 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1171 ++i, ++realArgIdx) {
1172 CCValAssign &VA = ArgLocs[i];
1173 SDValue Arg = OutVals[realArgIdx];
1174 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1176 // Promote the value if needed.
1177 switch (VA.getLocInfo()) {
1178 default: llvm_unreachable("Unknown loc info!");
1179 case CCValAssign::Full: break;
1180 case CCValAssign::SExt:
1181 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1183 case CCValAssign::ZExt:
1184 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1186 case CCValAssign::AExt:
1187 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1189 case CCValAssign::BCvt:
1190 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1194 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1195 if (VA.needsCustom()) {
1196 if (VA.getLocVT() == MVT::v2f64) {
1197 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1198 DAG.getConstant(0, MVT::i32));
1199 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1200 DAG.getConstant(1, MVT::i32));
1202 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1203 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1205 VA = ArgLocs[++i]; // skip ahead to next loc
1206 if (VA.isRegLoc()) {
1207 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1208 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1210 assert(VA.isMemLoc());
1212 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1213 dl, DAG, VA, Flags));
1216 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1217 StackPtr, MemOpChains, Flags);
1219 } else if (VA.isRegLoc()) {
1220 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1221 } else if (!IsSibCall) {
1222 assert(VA.isMemLoc());
1224 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1225 dl, DAG, VA, Flags));
1229 if (!MemOpChains.empty())
1230 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1231 &MemOpChains[0], MemOpChains.size());
1233 // Build a sequence of copy-to-reg nodes chained together with token chain
1234 // and flag operands which copy the outgoing args into the appropriate regs.
1236 // Tail call byval lowering might overwrite argument registers so in case of
1237 // tail call optimization the copies to registers are lowered later.
1239 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1240 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1241 RegsToPass[i].second, InFlag);
1242 InFlag = Chain.getValue(1);
1245 // For tail calls lower the arguments to the 'real' stack slot.
1247 // Force all the incoming stack arguments to be loaded from the stack
1248 // before any new outgoing arguments are stored to the stack, because the
1249 // outgoing stack slots may alias the incoming argument stack slots, and
1250 // the alias isn't otherwise explicit. This is slightly more conservative
1251 // than necessary, because it means that each store effectively depends
1252 // on every argument instead of just those arguments it would clobber.
1254 // Do not flag preceeding copytoreg stuff together with the following stuff.
1256 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1257 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1258 RegsToPass[i].second, InFlag);
1259 InFlag = Chain.getValue(1);
1264 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1265 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1266 // node so that legalize doesn't hack it.
1267 bool isDirect = false;
1268 bool isARMFunc = false;
1269 bool isLocalARMFunc = false;
1270 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1272 if (EnableARMLongCalls) {
1273 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1274 && "long-calls with non-static relocation model!");
1275 // Handle a global address or an external symbol. If it's not one of
1276 // those, the target's already in a register, so we don't need to do
1278 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1279 const GlobalValue *GV = G->getGlobal();
1280 // Create a constant pool entry for the callee address
1281 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1282 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1285 // Get the address of the callee into a register
1286 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1287 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1288 Callee = DAG.getLoad(getPointerTy(), dl,
1289 DAG.getEntryNode(), CPAddr,
1290 PseudoSourceValue::getConstantPool(), 0,
1292 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1293 const char *Sym = S->getSymbol();
1295 // Create a constant pool entry for the callee address
1296 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1297 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1298 Sym, ARMPCLabelIndex, 0);
1299 // Get the address of the callee into a register
1300 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1301 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1302 Callee = DAG.getLoad(getPointerTy(), dl,
1303 DAG.getEntryNode(), CPAddr,
1304 PseudoSourceValue::getConstantPool(), 0,
1307 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1308 const GlobalValue *GV = G->getGlobal();
1310 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1311 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1312 getTargetMachine().getRelocationModel() != Reloc::Static;
1313 isARMFunc = !Subtarget->isThumb() || isStub;
1314 // ARM call to a local ARM function is predicable.
1315 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1316 // tBX takes a register source operand.
1317 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1318 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1319 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1322 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1323 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1324 Callee = DAG.getLoad(getPointerTy(), dl,
1325 DAG.getEntryNode(), CPAddr,
1326 PseudoSourceValue::getConstantPool(), 0,
1328 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1329 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1330 getPointerTy(), Callee, PICLabel);
1332 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1333 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1335 bool isStub = Subtarget->isTargetDarwin() &&
1336 getTargetMachine().getRelocationModel() != Reloc::Static;
1337 isARMFunc = !Subtarget->isThumb() || isStub;
1338 // tBX takes a register source operand.
1339 const char *Sym = S->getSymbol();
1340 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1341 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1342 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1343 Sym, ARMPCLabelIndex, 4);
1344 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1345 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1346 Callee = DAG.getLoad(getPointerTy(), dl,
1347 DAG.getEntryNode(), CPAddr,
1348 PseudoSourceValue::getConstantPool(), 0,
1350 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1351 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1352 getPointerTy(), Callee, PICLabel);
1354 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1357 // FIXME: handle tail calls differently.
1359 if (Subtarget->isThumb()) {
1360 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1361 CallOpc = ARMISD::CALL_NOLINK;
1363 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1365 CallOpc = (isDirect || Subtarget->hasV5TOps())
1366 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1367 : ARMISD::CALL_NOLINK;
1370 std::vector<SDValue> Ops;
1371 Ops.push_back(Chain);
1372 Ops.push_back(Callee);
1374 // Add argument registers to the end of the list so that they are known live
1376 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1377 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1378 RegsToPass[i].second.getValueType()));
1380 if (InFlag.getNode())
1381 Ops.push_back(InFlag);
1383 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1385 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1387 // Returns a chain and a flag for retval copy to use.
1388 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1389 InFlag = Chain.getValue(1);
1391 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1392 DAG.getIntPtrConstant(0, true), InFlag);
1394 InFlag = Chain.getValue(1);
1396 // Handle result values, copying them out of physregs into vregs that we
1398 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1402 /// MatchingStackOffset - Return true if the given stack call argument is
1403 /// already available in the same position (relatively) of the caller's
1404 /// incoming argument stack.
1406 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1407 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1408 const ARMInstrInfo *TII) {
1409 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1411 if (Arg.getOpcode() == ISD::CopyFromReg) {
1412 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1413 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1415 MachineInstr *Def = MRI->getVRegDef(VR);
1418 if (!Flags.isByVal()) {
1419 if (!TII->isLoadFromStackSlot(Def, FI))
1424 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1425 if (Flags.isByVal())
1426 // ByVal argument is passed in as a pointer but it's now being
1427 // dereferenced. e.g.
1428 // define @foo(%struct.X* %A) {
1429 // tail call @bar(%struct.X* byval %A)
1432 SDValue Ptr = Ld->getBasePtr();
1433 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1436 FI = FINode->getIndex();
1440 assert(FI != INT_MAX);
1441 if (!MFI->isFixedObjectIndex(FI))
1443 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1446 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1447 /// for tail call optimization. Targets which want to do tail call
1448 /// optimization should implement this function.
1450 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1451 CallingConv::ID CalleeCC,
1453 bool isCalleeStructRet,
1454 bool isCallerStructRet,
1455 const SmallVectorImpl<ISD::OutputArg> &Outs,
1456 const SmallVectorImpl<SDValue> &OutVals,
1457 const SmallVectorImpl<ISD::InputArg> &Ins,
1458 SelectionDAG& DAG) const {
1459 const Function *CallerF = DAG.getMachineFunction().getFunction();
1460 CallingConv::ID CallerCC = CallerF->getCallingConv();
1461 bool CCMatch = CallerCC == CalleeCC;
1463 // Look for obvious safe cases to perform tail call optimization that do not
1464 // require ABI changes. This is what gcc calls sibcall.
1466 // Do not sibcall optimize vararg calls unless the call site is not passing
1468 if (isVarArg && !Outs.empty())
1471 // Also avoid sibcall optimization if either caller or callee uses struct
1472 // return semantics.
1473 if (isCalleeStructRet || isCallerStructRet)
1476 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1477 // emitEpilogue is not ready for them.
1478 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1479 // LR. This means if we need to reload LR, it takes an extra instructions,
1480 // which outweighs the value of the tail call; but here we don't know yet
1481 // whether LR is going to be used. Probably the right approach is to
1482 // generate the tail call here and turn it back into CALL/RET in
1483 // emitEpilogue if LR is used.
1484 if (Subtarget->isThumb1Only())
1487 // For the moment, we can only do this to functions defined in this
1488 // compilation, or to indirect calls. A Thumb B to an ARM function,
1489 // or vice versa, is not easily fixed up in the linker unlike BL.
1490 // (We could do this by loading the address of the callee into a register;
1491 // that is an extra instruction over the direct call and burns a register
1492 // as well, so is not likely to be a win.)
1494 // It might be safe to remove this restriction on non-Darwin.
1496 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1497 // but we need to make sure there are enough registers; the only valid
1498 // registers are the 4 used for parameters. We don't currently do this
1500 if (isa<ExternalSymbolSDNode>(Callee))
1503 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1504 const GlobalValue *GV = G->getGlobal();
1505 if (GV->isDeclaration() || GV->isWeakForLinker())
1509 // If the calling conventions do not match, then we'd better make sure the
1510 // results are returned in the same way as what the caller expects.
1512 SmallVector<CCValAssign, 16> RVLocs1;
1513 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1514 RVLocs1, *DAG.getContext());
1515 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1517 SmallVector<CCValAssign, 16> RVLocs2;
1518 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1519 RVLocs2, *DAG.getContext());
1520 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1522 if (RVLocs1.size() != RVLocs2.size())
1524 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1525 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1527 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1529 if (RVLocs1[i].isRegLoc()) {
1530 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1533 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1539 // If the callee takes no arguments then go on to check the results of the
1541 if (!Outs.empty()) {
1542 // Check if stack adjustment is needed. For now, do not do this if any
1543 // argument is passed on the stack.
1544 SmallVector<CCValAssign, 16> ArgLocs;
1545 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1546 ArgLocs, *DAG.getContext());
1547 CCInfo.AnalyzeCallOperands(Outs,
1548 CCAssignFnForNode(CalleeCC, false, isVarArg));
1549 if (CCInfo.getNextStackOffset()) {
1550 MachineFunction &MF = DAG.getMachineFunction();
1552 // Check if the arguments are already laid out in the right way as
1553 // the caller's fixed stack objects.
1554 MachineFrameInfo *MFI = MF.getFrameInfo();
1555 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1556 const ARMInstrInfo *TII =
1557 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1558 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1560 ++i, ++realArgIdx) {
1561 CCValAssign &VA = ArgLocs[i];
1562 EVT RegVT = VA.getLocVT();
1563 SDValue Arg = OutVals[realArgIdx];
1564 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1565 if (VA.getLocInfo() == CCValAssign::Indirect)
1567 if (VA.needsCustom()) {
1568 // f64 and vector types are split into multiple registers or
1569 // register/stack-slot combinations. The types will not match
1570 // the registers; give up on memory f64 refs until we figure
1571 // out what to do about this.
1574 if (!ArgLocs[++i].isRegLoc())
1576 if (RegVT == MVT::v2f64) {
1577 if (!ArgLocs[++i].isRegLoc())
1579 if (!ArgLocs[++i].isRegLoc())
1582 } else if (!VA.isRegLoc()) {
1583 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1595 ARMTargetLowering::LowerReturn(SDValue Chain,
1596 CallingConv::ID CallConv, bool isVarArg,
1597 const SmallVectorImpl<ISD::OutputArg> &Outs,
1598 const SmallVectorImpl<SDValue> &OutVals,
1599 DebugLoc dl, SelectionDAG &DAG) const {
1601 // CCValAssign - represent the assignment of the return value to a location.
1602 SmallVector<CCValAssign, 16> RVLocs;
1604 // CCState - Info about the registers and stack slots.
1605 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1608 // Analyze outgoing return values.
1609 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1612 // If this is the first return lowered for this function, add
1613 // the regs to the liveout set for the function.
1614 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1615 for (unsigned i = 0; i != RVLocs.size(); ++i)
1616 if (RVLocs[i].isRegLoc())
1617 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1622 // Copy the result values into the output registers.
1623 for (unsigned i = 0, realRVLocIdx = 0;
1625 ++i, ++realRVLocIdx) {
1626 CCValAssign &VA = RVLocs[i];
1627 assert(VA.isRegLoc() && "Can only return in registers!");
1629 SDValue Arg = OutVals[realRVLocIdx];
1631 switch (VA.getLocInfo()) {
1632 default: llvm_unreachable("Unknown loc info!");
1633 case CCValAssign::Full: break;
1634 case CCValAssign::BCvt:
1635 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1639 if (VA.needsCustom()) {
1640 if (VA.getLocVT() == MVT::v2f64) {
1641 // Extract the first half and return it in two registers.
1642 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1643 DAG.getConstant(0, MVT::i32));
1644 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1645 DAG.getVTList(MVT::i32, MVT::i32), Half);
1647 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1648 Flag = Chain.getValue(1);
1649 VA = RVLocs[++i]; // skip ahead to next loc
1650 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1651 HalfGPRs.getValue(1), Flag);
1652 Flag = Chain.getValue(1);
1653 VA = RVLocs[++i]; // skip ahead to next loc
1655 // Extract the 2nd half and fall through to handle it as an f64 value.
1656 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1657 DAG.getConstant(1, MVT::i32));
1659 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1661 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1662 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1663 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1664 Flag = Chain.getValue(1);
1665 VA = RVLocs[++i]; // skip ahead to next loc
1666 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1669 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1671 // Guarantee that all emitted copies are
1672 // stuck together, avoiding something bad.
1673 Flag = Chain.getValue(1);
1678 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1680 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1685 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1686 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1687 // one of the above mentioned nodes. It has to be wrapped because otherwise
1688 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1689 // be used to form addressing mode. These wrapped nodes will be selected
1691 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1692 EVT PtrVT = Op.getValueType();
1693 // FIXME there is no actual debug info here
1694 DebugLoc dl = Op.getDebugLoc();
1695 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1697 if (CP->isMachineConstantPoolEntry())
1698 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1699 CP->getAlignment());
1701 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1702 CP->getAlignment());
1703 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1706 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1707 return MachineJumpTableInfo::EK_Inline;
1710 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1711 SelectionDAG &DAG) const {
1712 MachineFunction &MF = DAG.getMachineFunction();
1713 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1714 unsigned ARMPCLabelIndex = 0;
1715 DebugLoc DL = Op.getDebugLoc();
1716 EVT PtrVT = getPointerTy();
1717 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1718 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1720 if (RelocM == Reloc::Static) {
1721 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1723 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1724 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1725 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1726 ARMCP::CPBlockAddress,
1728 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1730 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1731 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1732 PseudoSourceValue::getConstantPool(), 0,
1734 if (RelocM == Reloc::Static)
1736 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1737 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1740 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1742 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1743 SelectionDAG &DAG) const {
1744 DebugLoc dl = GA->getDebugLoc();
1745 EVT PtrVT = getPointerTy();
1746 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1747 MachineFunction &MF = DAG.getMachineFunction();
1748 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1749 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1750 ARMConstantPoolValue *CPV =
1751 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1752 ARMCP::CPValue, PCAdj, "tlsgd", true);
1753 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1754 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1755 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1756 PseudoSourceValue::getConstantPool(), 0,
1758 SDValue Chain = Argument.getValue(1);
1760 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1761 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1763 // call __tls_get_addr.
1766 Entry.Node = Argument;
1767 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1768 Args.push_back(Entry);
1769 // FIXME: is there useful debug info available here?
1770 std::pair<SDValue, SDValue> CallResult =
1771 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1772 false, false, false, false,
1773 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1774 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1775 return CallResult.first;
1778 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1779 // "local exec" model.
1781 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1782 SelectionDAG &DAG) const {
1783 const GlobalValue *GV = GA->getGlobal();
1784 DebugLoc dl = GA->getDebugLoc();
1786 SDValue Chain = DAG.getEntryNode();
1787 EVT PtrVT = getPointerTy();
1788 // Get the Thread Pointer
1789 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1791 if (GV->isDeclaration()) {
1792 MachineFunction &MF = DAG.getMachineFunction();
1793 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1794 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1795 // Initial exec model.
1796 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1797 ARMConstantPoolValue *CPV =
1798 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1799 ARMCP::CPValue, PCAdj, "gottpoff", true);
1800 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1801 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1802 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1803 PseudoSourceValue::getConstantPool(), 0,
1805 Chain = Offset.getValue(1);
1807 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1808 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1810 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1811 PseudoSourceValue::getConstantPool(), 0,
1815 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1816 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1817 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1818 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1819 PseudoSourceValue::getConstantPool(), 0,
1823 // The address of the thread local variable is the add of the thread
1824 // pointer with the offset of the variable.
1825 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1829 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1830 // TODO: implement the "local dynamic" model
1831 assert(Subtarget->isTargetELF() &&
1832 "TLS not implemented for non-ELF targets");
1833 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1834 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1835 // otherwise use the "Local Exec" TLS Model
1836 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1837 return LowerToTLSGeneralDynamicModel(GA, DAG);
1839 return LowerToTLSExecModels(GA, DAG);
1842 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1843 SelectionDAG &DAG) const {
1844 EVT PtrVT = getPointerTy();
1845 DebugLoc dl = Op.getDebugLoc();
1846 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1847 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1848 if (RelocM == Reloc::PIC_) {
1849 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1850 ARMConstantPoolValue *CPV =
1851 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1852 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1853 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1854 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1856 PseudoSourceValue::getConstantPool(), 0,
1858 SDValue Chain = Result.getValue(1);
1859 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1860 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1862 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1863 PseudoSourceValue::getGOT(), 0,
1867 // If we have T2 ops, we can materialize the address directly via movt/movw
1868 // pair. This is always cheaper.
1869 if (Subtarget->useMovt()) {
1870 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1871 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1873 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1874 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1875 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1876 PseudoSourceValue::getConstantPool(), 0,
1882 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1883 SelectionDAG &DAG) const {
1884 MachineFunction &MF = DAG.getMachineFunction();
1885 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1886 unsigned ARMPCLabelIndex = 0;
1887 EVT PtrVT = getPointerTy();
1888 DebugLoc dl = Op.getDebugLoc();
1889 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1890 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1892 if (RelocM == Reloc::Static)
1893 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1895 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1896 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1897 ARMConstantPoolValue *CPV =
1898 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1899 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1901 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1903 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1904 PseudoSourceValue::getConstantPool(), 0,
1906 SDValue Chain = Result.getValue(1);
1908 if (RelocM == Reloc::PIC_) {
1909 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1910 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1913 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1914 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1915 PseudoSourceValue::getGOT(), 0,
1921 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1922 SelectionDAG &DAG) const {
1923 assert(Subtarget->isTargetELF() &&
1924 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1925 MachineFunction &MF = DAG.getMachineFunction();
1926 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1927 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1928 EVT PtrVT = getPointerTy();
1929 DebugLoc dl = Op.getDebugLoc();
1930 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1931 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1932 "_GLOBAL_OFFSET_TABLE_",
1933 ARMPCLabelIndex, PCAdj);
1934 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1935 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1936 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1937 PseudoSourceValue::getConstantPool(), 0,
1939 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1940 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1944 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1945 DebugLoc dl = Op.getDebugLoc();
1946 SDValue Val = DAG.getConstant(0, MVT::i32);
1947 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1948 Op.getOperand(1), Val);
1952 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1953 DebugLoc dl = Op.getDebugLoc();
1954 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1955 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1959 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1960 const ARMSubtarget *Subtarget) const {
1961 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1962 DebugLoc dl = Op.getDebugLoc();
1964 default: return SDValue(); // Don't custom lower most intrinsics.
1965 case Intrinsic::arm_thread_pointer: {
1966 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1967 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1969 case Intrinsic::eh_sjlj_lsda: {
1970 MachineFunction &MF = DAG.getMachineFunction();
1971 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1972 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1973 EVT PtrVT = getPointerTy();
1974 DebugLoc dl = Op.getDebugLoc();
1975 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1977 unsigned PCAdj = (RelocM != Reloc::PIC_)
1978 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1979 ARMConstantPoolValue *CPV =
1980 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1981 ARMCP::CPLSDA, PCAdj);
1982 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1983 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1985 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1986 PseudoSourceValue::getConstantPool(), 0,
1989 if (RelocM == Reloc::PIC_) {
1990 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1991 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1998 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1999 const ARMSubtarget *Subtarget) {
2000 DebugLoc dl = Op.getDebugLoc();
2001 SDValue Op5 = Op.getOperand(5);
2002 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
2003 // Some subtargets which have dmb and dsb instructions can handle barriers
2004 // directly. Some ARMv6 cpus can support them with the help of mcr
2005 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
2007 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
2008 if (Subtarget->hasDataBarrier())
2009 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2011 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2012 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2013 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2014 DAG.getConstant(0, MVT::i32));
2018 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2019 MachineFunction &MF = DAG.getMachineFunction();
2020 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2022 // vastart just stores the address of the VarArgsFrameIndex slot into the
2023 // memory location argument.
2024 DebugLoc dl = Op.getDebugLoc();
2025 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2026 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2027 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2028 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
2033 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2034 SDValue &Root, SelectionDAG &DAG,
2035 DebugLoc dl) const {
2036 MachineFunction &MF = DAG.getMachineFunction();
2037 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2039 TargetRegisterClass *RC;
2040 if (AFI->isThumb1OnlyFunction())
2041 RC = ARM::tGPRRegisterClass;
2043 RC = ARM::GPRRegisterClass;
2045 // Transform the arguments stored in physical registers into virtual ones.
2046 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2047 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2050 if (NextVA.isMemLoc()) {
2051 MachineFrameInfo *MFI = MF.getFrameInfo();
2052 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2054 // Create load node to retrieve arguments from the stack.
2055 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2056 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2057 PseudoSourceValue::getFixedStack(FI), 0,
2060 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2061 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2064 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2068 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2069 CallingConv::ID CallConv, bool isVarArg,
2070 const SmallVectorImpl<ISD::InputArg>
2072 DebugLoc dl, SelectionDAG &DAG,
2073 SmallVectorImpl<SDValue> &InVals)
2076 MachineFunction &MF = DAG.getMachineFunction();
2077 MachineFrameInfo *MFI = MF.getFrameInfo();
2079 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2081 // Assign locations to all of the incoming arguments.
2082 SmallVector<CCValAssign, 16> ArgLocs;
2083 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2085 CCInfo.AnalyzeFormalArguments(Ins,
2086 CCAssignFnForNode(CallConv, /* Return*/ false,
2089 SmallVector<SDValue, 16> ArgValues;
2091 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2092 CCValAssign &VA = ArgLocs[i];
2094 // Arguments stored in registers.
2095 if (VA.isRegLoc()) {
2096 EVT RegVT = VA.getLocVT();
2099 if (VA.needsCustom()) {
2100 // f64 and vector types are split up into multiple registers or
2101 // combinations of registers and stack slots.
2102 if (VA.getLocVT() == MVT::v2f64) {
2103 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2105 VA = ArgLocs[++i]; // skip ahead to next loc
2107 if (VA.isMemLoc()) {
2108 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2109 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2110 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2111 PseudoSourceValue::getFixedStack(FI), 0,
2114 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2117 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2118 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2119 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2120 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2121 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2123 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2126 TargetRegisterClass *RC;
2128 if (RegVT == MVT::f32)
2129 RC = ARM::SPRRegisterClass;
2130 else if (RegVT == MVT::f64)
2131 RC = ARM::DPRRegisterClass;
2132 else if (RegVT == MVT::v2f64)
2133 RC = ARM::QPRRegisterClass;
2134 else if (RegVT == MVT::i32)
2135 RC = (AFI->isThumb1OnlyFunction() ?
2136 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2138 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2140 // Transform the arguments in physical registers into virtual ones.
2141 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2142 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2145 // If this is an 8 or 16-bit value, it is really passed promoted
2146 // to 32 bits. Insert an assert[sz]ext to capture this, then
2147 // truncate to the right size.
2148 switch (VA.getLocInfo()) {
2149 default: llvm_unreachable("Unknown loc info!");
2150 case CCValAssign::Full: break;
2151 case CCValAssign::BCvt:
2152 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2154 case CCValAssign::SExt:
2155 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2156 DAG.getValueType(VA.getValVT()));
2157 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2159 case CCValAssign::ZExt:
2160 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2161 DAG.getValueType(VA.getValVT()));
2162 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2166 InVals.push_back(ArgValue);
2168 } else { // VA.isRegLoc()
2171 assert(VA.isMemLoc());
2172 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2174 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2175 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2177 // Create load nodes to retrieve arguments from the stack.
2178 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2179 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2180 PseudoSourceValue::getFixedStack(FI), 0,
2187 static const unsigned GPRArgRegs[] = {
2188 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2191 unsigned NumGPRs = CCInfo.getFirstUnallocated
2192 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2194 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2195 unsigned VARegSize = (4 - NumGPRs) * 4;
2196 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2197 unsigned ArgOffset = CCInfo.getNextStackOffset();
2198 if (VARegSaveSize) {
2199 // If this function is vararg, store any remaining integer argument regs
2200 // to their spots on the stack so that they may be loaded by deferencing
2201 // the result of va_next.
2202 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2203 AFI->setVarArgsFrameIndex(
2204 MFI->CreateFixedObject(VARegSaveSize,
2205 ArgOffset + VARegSaveSize - VARegSize,
2207 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2210 SmallVector<SDValue, 4> MemOps;
2211 for (; NumGPRs < 4; ++NumGPRs) {
2212 TargetRegisterClass *RC;
2213 if (AFI->isThumb1OnlyFunction())
2214 RC = ARM::tGPRRegisterClass;
2216 RC = ARM::GPRRegisterClass;
2218 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2219 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2221 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2222 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2223 0, false, false, 0);
2224 MemOps.push_back(Store);
2225 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2226 DAG.getConstant(4, getPointerTy()));
2228 if (!MemOps.empty())
2229 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2230 &MemOps[0], MemOps.size());
2232 // This will point to the next argument passed via stack.
2233 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2239 /// isFloatingPointZero - Return true if this is +0.0.
2240 static bool isFloatingPointZero(SDValue Op) {
2241 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2242 return CFP->getValueAPF().isPosZero();
2243 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2244 // Maybe this has already been legalized into the constant pool?
2245 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2246 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2247 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2248 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2249 return CFP->getValueAPF().isPosZero();
2255 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2256 /// the given operands.
2258 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2259 SDValue &ARMcc, SelectionDAG &DAG,
2260 DebugLoc dl) const {
2261 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2262 unsigned C = RHSC->getZExtValue();
2263 if (!isLegalICmpImmediate(C)) {
2264 // Constant does not fit, try adjusting it by one?
2269 if (isLegalICmpImmediate(C-1)) {
2270 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2271 RHS = DAG.getConstant(C-1, MVT::i32);
2276 if (C > 0 && isLegalICmpImmediate(C-1)) {
2277 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2278 RHS = DAG.getConstant(C-1, MVT::i32);
2283 if (isLegalICmpImmediate(C+1)) {
2284 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2285 RHS = DAG.getConstant(C+1, MVT::i32);
2290 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2291 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2292 RHS = DAG.getConstant(C+1, MVT::i32);
2299 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2300 ARMISD::NodeType CompareType;
2303 CompareType = ARMISD::CMP;
2308 CompareType = ARMISD::CMPZ;
2311 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2312 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2315 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2317 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2318 DebugLoc dl) const {
2320 if (!isFloatingPointZero(RHS))
2321 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2323 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2324 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2327 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2328 SDValue Cond = Op.getOperand(0);
2329 SDValue SelectTrue = Op.getOperand(1);
2330 SDValue SelectFalse = Op.getOperand(2);
2331 DebugLoc dl = Op.getDebugLoc();
2335 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2336 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2338 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2339 const ConstantSDNode *CMOVTrue =
2340 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2341 const ConstantSDNode *CMOVFalse =
2342 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2344 if (CMOVTrue && CMOVFalse) {
2345 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2346 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2350 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2352 False = SelectFalse;
2353 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2358 if (True.getNode() && False.getNode()) {
2359 EVT VT = Cond.getValueType();
2360 SDValue ARMcc = Cond.getOperand(2);
2361 SDValue CCR = Cond.getOperand(3);
2362 SDValue Cmp = Cond.getOperand(4);
2363 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2368 return DAG.getSelectCC(dl, Cond,
2369 DAG.getConstant(0, Cond.getValueType()),
2370 SelectTrue, SelectFalse, ISD::SETNE);
2373 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2374 EVT VT = Op.getValueType();
2375 SDValue LHS = Op.getOperand(0);
2376 SDValue RHS = Op.getOperand(1);
2377 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2378 SDValue TrueVal = Op.getOperand(2);
2379 SDValue FalseVal = Op.getOperand(3);
2380 DebugLoc dl = Op.getDebugLoc();
2382 if (LHS.getValueType() == MVT::i32) {
2384 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2385 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2386 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2389 ARMCC::CondCodes CondCode, CondCode2;
2390 FPCCToARMCC(CC, CondCode, CondCode2);
2392 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2393 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2394 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2395 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2397 if (CondCode2 != ARMCC::AL) {
2398 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2399 // FIXME: Needs another CMP because flag can have but one use.
2400 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2401 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2402 Result, TrueVal, ARMcc2, CCR, Cmp2);
2407 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2408 /// to morph to an integer compare sequence.
2409 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2410 const ARMSubtarget *Subtarget) {
2411 SDNode *N = Op.getNode();
2412 if (!N->hasOneUse())
2413 // Otherwise it requires moving the value from fp to integer registers.
2415 if (!N->getNumValues())
2417 EVT VT = Op.getValueType();
2418 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2419 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2420 // vmrs are very slow, e.g. cortex-a8.
2423 if (isFloatingPointZero(Op)) {
2427 return ISD::isNormalLoad(N);
2430 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2431 if (isFloatingPointZero(Op))
2432 return DAG.getConstant(0, MVT::i32);
2434 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2435 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2436 Ld->getChain(), Ld->getBasePtr(),
2437 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2438 Ld->isVolatile(), Ld->isNonTemporal(),
2439 Ld->getAlignment());
2441 llvm_unreachable("Unknown VFP cmp argument!");
2444 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2445 SDValue &RetVal1, SDValue &RetVal2) {
2446 if (isFloatingPointZero(Op)) {
2447 RetVal1 = DAG.getConstant(0, MVT::i32);
2448 RetVal2 = DAG.getConstant(0, MVT::i32);
2452 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2453 SDValue Ptr = Ld->getBasePtr();
2454 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2455 Ld->getChain(), Ptr,
2456 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2457 Ld->isVolatile(), Ld->isNonTemporal(),
2458 Ld->getAlignment());
2460 EVT PtrType = Ptr.getValueType();
2461 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2462 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2463 PtrType, Ptr, DAG.getConstant(4, PtrType));
2464 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2465 Ld->getChain(), NewPtr,
2466 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2467 Ld->isVolatile(), Ld->isNonTemporal(),
2472 llvm_unreachable("Unknown VFP cmp argument!");
2475 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2476 /// f32 and even f64 comparisons to integer ones.
2478 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2479 SDValue Chain = Op.getOperand(0);
2480 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2481 SDValue LHS = Op.getOperand(2);
2482 SDValue RHS = Op.getOperand(3);
2483 SDValue Dest = Op.getOperand(4);
2484 DebugLoc dl = Op.getDebugLoc();
2486 bool SeenZero = false;
2487 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2488 canChangeToInt(RHS, SeenZero, Subtarget) &&
2489 // If one of the operand is zero, it's safe to ignore the NaN case since
2490 // we only care about equality comparisons.
2491 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2492 // If unsafe fp math optimization is enabled and there are no othter uses of
2493 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2494 // to an integer comparison.
2495 if (CC == ISD::SETOEQ)
2497 else if (CC == ISD::SETUNE)
2501 if (LHS.getValueType() == MVT::f32) {
2502 LHS = bitcastf32Toi32(LHS, DAG);
2503 RHS = bitcastf32Toi32(RHS, DAG);
2504 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2505 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2506 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2507 Chain, Dest, ARMcc, CCR, Cmp);
2512 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2513 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2514 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2515 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2516 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2517 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2518 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2524 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2525 SDValue Chain = Op.getOperand(0);
2526 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2527 SDValue LHS = Op.getOperand(2);
2528 SDValue RHS = Op.getOperand(3);
2529 SDValue Dest = Op.getOperand(4);
2530 DebugLoc dl = Op.getDebugLoc();
2532 if (LHS.getValueType() == MVT::i32) {
2534 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2535 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2536 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2537 Chain, Dest, ARMcc, CCR, Cmp);
2540 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2543 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2544 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2545 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2546 if (Result.getNode())
2550 ARMCC::CondCodes CondCode, CondCode2;
2551 FPCCToARMCC(CC, CondCode, CondCode2);
2553 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2554 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2555 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2556 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2557 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2558 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2559 if (CondCode2 != ARMCC::AL) {
2560 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2561 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2562 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2567 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2568 SDValue Chain = Op.getOperand(0);
2569 SDValue Table = Op.getOperand(1);
2570 SDValue Index = Op.getOperand(2);
2571 DebugLoc dl = Op.getDebugLoc();
2573 EVT PTy = getPointerTy();
2574 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2575 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2576 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2577 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2578 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2579 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2580 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2581 if (Subtarget->isThumb2()) {
2582 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2583 // which does another jump to the destination. This also makes it easier
2584 // to translate it to TBB / TBH later.
2585 // FIXME: This might not work if the function is extremely large.
2586 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2587 Addr, Op.getOperand(2), JTI, UId);
2589 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2590 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2591 PseudoSourceValue::getJumpTable(), 0,
2593 Chain = Addr.getValue(1);
2594 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2595 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2597 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2598 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2599 Chain = Addr.getValue(1);
2600 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2604 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2605 DebugLoc dl = Op.getDebugLoc();
2608 switch (Op.getOpcode()) {
2610 assert(0 && "Invalid opcode!");
2611 case ISD::FP_TO_SINT:
2612 Opc = ARMISD::FTOSI;
2614 case ISD::FP_TO_UINT:
2615 Opc = ARMISD::FTOUI;
2618 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2619 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2622 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2623 EVT VT = Op.getValueType();
2624 DebugLoc dl = Op.getDebugLoc();
2627 switch (Op.getOpcode()) {
2629 assert(0 && "Invalid opcode!");
2630 case ISD::SINT_TO_FP:
2631 Opc = ARMISD::SITOF;
2633 case ISD::UINT_TO_FP:
2634 Opc = ARMISD::UITOF;
2638 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2639 return DAG.getNode(Opc, dl, VT, Op);
2642 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2643 // Implement fcopysign with a fabs and a conditional fneg.
2644 SDValue Tmp0 = Op.getOperand(0);
2645 SDValue Tmp1 = Op.getOperand(1);
2646 DebugLoc dl = Op.getDebugLoc();
2647 EVT VT = Op.getValueType();
2648 EVT SrcVT = Tmp1.getValueType();
2649 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2650 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2651 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2652 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2653 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2654 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2657 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2658 MachineFunction &MF = DAG.getMachineFunction();
2659 MachineFrameInfo *MFI = MF.getFrameInfo();
2660 MFI->setReturnAddressIsTaken(true);
2662 EVT VT = Op.getValueType();
2663 DebugLoc dl = Op.getDebugLoc();
2664 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2666 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2667 SDValue Offset = DAG.getConstant(4, MVT::i32);
2668 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2669 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2670 NULL, 0, false, false, 0);
2673 // Return LR, which contains the return address. Mark it an implicit live-in.
2674 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2675 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2678 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2679 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2680 MFI->setFrameAddressIsTaken(true);
2682 EVT VT = Op.getValueType();
2683 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2684 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2685 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2686 ? ARM::R7 : ARM::R11;
2687 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2689 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2694 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2695 /// expand a bit convert where either the source or destination type is i64 to
2696 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2697 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2698 /// vectors), since the legalizer won't know what to do with that.
2699 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2700 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2701 DebugLoc dl = N->getDebugLoc();
2702 SDValue Op = N->getOperand(0);
2704 // This function is only supposed to be called for i64 types, either as the
2705 // source or destination of the bit convert.
2706 EVT SrcVT = Op.getValueType();
2707 EVT DstVT = N->getValueType(0);
2708 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2709 "ExpandBIT_CONVERT called for non-i64 type");
2711 // Turn i64->f64 into VMOVDRR.
2712 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2713 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2714 DAG.getConstant(0, MVT::i32));
2715 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2716 DAG.getConstant(1, MVT::i32));
2717 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2718 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2721 // Turn f64->i64 into VMOVRRD.
2722 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2723 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2724 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2725 // Merge the pieces into a single i64 value.
2726 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2732 /// getZeroVector - Returns a vector of specified type with all zero elements.
2733 /// Zero vectors are used to represent vector negation and in those cases
2734 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2735 /// not support i64 elements, so sometimes the zero vectors will need to be
2736 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2738 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2739 assert(VT.isVector() && "Expected a vector type");
2740 // The canonical modified immediate encoding of a zero vector is....0!
2741 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2742 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2743 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2744 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2747 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2748 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2749 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2750 SelectionDAG &DAG) const {
2751 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2752 EVT VT = Op.getValueType();
2753 unsigned VTBits = VT.getSizeInBits();
2754 DebugLoc dl = Op.getDebugLoc();
2755 SDValue ShOpLo = Op.getOperand(0);
2756 SDValue ShOpHi = Op.getOperand(1);
2757 SDValue ShAmt = Op.getOperand(2);
2759 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2761 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2763 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2764 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2765 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2766 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2767 DAG.getConstant(VTBits, MVT::i32));
2768 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2769 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2770 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2772 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2773 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2775 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2776 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2779 SDValue Ops[2] = { Lo, Hi };
2780 return DAG.getMergeValues(Ops, 2, dl);
2783 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2784 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2785 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2786 SelectionDAG &DAG) const {
2787 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2788 EVT VT = Op.getValueType();
2789 unsigned VTBits = VT.getSizeInBits();
2790 DebugLoc dl = Op.getDebugLoc();
2791 SDValue ShOpLo = Op.getOperand(0);
2792 SDValue ShOpHi = Op.getOperand(1);
2793 SDValue ShAmt = Op.getOperand(2);
2796 assert(Op.getOpcode() == ISD::SHL_PARTS);
2797 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2798 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2799 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2800 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2801 DAG.getConstant(VTBits, MVT::i32));
2802 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2803 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2805 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2806 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2807 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2809 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2810 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2813 SDValue Ops[2] = { Lo, Hi };
2814 return DAG.getMergeValues(Ops, 2, dl);
2817 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2818 SelectionDAG &DAG) const {
2819 // The rounding mode is in bits 23:22 of the FPSCR.
2820 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2821 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2822 // so that the shift + and get folded into a bitfield extract.
2823 DebugLoc dl = Op.getDebugLoc();
2824 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2825 DAG.getConstant(Intrinsic::arm_get_fpscr,
2827 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2828 DAG.getConstant(1U << 22, MVT::i32));
2829 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2830 DAG.getConstant(22, MVT::i32));
2831 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2832 DAG.getConstant(3, MVT::i32));
2835 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2836 const ARMSubtarget *ST) {
2837 EVT VT = N->getValueType(0);
2838 DebugLoc dl = N->getDebugLoc();
2840 if (!ST->hasV6T2Ops())
2843 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2844 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2847 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2848 const ARMSubtarget *ST) {
2849 EVT VT = N->getValueType(0);
2850 DebugLoc dl = N->getDebugLoc();
2852 // Lower vector shifts on NEON to use VSHL.
2853 if (VT.isVector()) {
2854 assert(ST->hasNEON() && "unexpected vector shift");
2856 // Left shifts translate directly to the vshiftu intrinsic.
2857 if (N->getOpcode() == ISD::SHL)
2858 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2859 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2860 N->getOperand(0), N->getOperand(1));
2862 assert((N->getOpcode() == ISD::SRA ||
2863 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2865 // NEON uses the same intrinsics for both left and right shifts. For
2866 // right shifts, the shift amounts are negative, so negate the vector of
2868 EVT ShiftVT = N->getOperand(1).getValueType();
2869 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2870 getZeroVector(ShiftVT, DAG, dl),
2872 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2873 Intrinsic::arm_neon_vshifts :
2874 Intrinsic::arm_neon_vshiftu);
2875 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2876 DAG.getConstant(vshiftInt, MVT::i32),
2877 N->getOperand(0), NegatedCount);
2880 // We can get here for a node like i32 = ISD::SHL i32, i64
2884 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2885 "Unknown shift to lower!");
2887 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2888 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2889 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2892 // If we are in thumb mode, we don't have RRX.
2893 if (ST->isThumb1Only()) return SDValue();
2895 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2896 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2897 DAG.getConstant(0, MVT::i32));
2898 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2899 DAG.getConstant(1, MVT::i32));
2901 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2902 // captures the result into a carry flag.
2903 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2904 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2906 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2907 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2909 // Merge the pieces into a single i64 value.
2910 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2913 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2914 SDValue TmpOp0, TmpOp1;
2915 bool Invert = false;
2919 SDValue Op0 = Op.getOperand(0);
2920 SDValue Op1 = Op.getOperand(1);
2921 SDValue CC = Op.getOperand(2);
2922 EVT VT = Op.getValueType();
2923 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2924 DebugLoc dl = Op.getDebugLoc();
2926 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2927 switch (SetCCOpcode) {
2928 default: llvm_unreachable("Illegal FP comparison"); break;
2930 case ISD::SETNE: Invert = true; // Fallthrough
2932 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2934 case ISD::SETLT: Swap = true; // Fallthrough
2936 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2938 case ISD::SETLE: Swap = true; // Fallthrough
2940 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2941 case ISD::SETUGE: Swap = true; // Fallthrough
2942 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2943 case ISD::SETUGT: Swap = true; // Fallthrough
2944 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2945 case ISD::SETUEQ: Invert = true; // Fallthrough
2947 // Expand this to (OLT | OGT).
2951 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2952 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2954 case ISD::SETUO: Invert = true; // Fallthrough
2956 // Expand this to (OLT | OGE).
2960 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2961 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2965 // Integer comparisons.
2966 switch (SetCCOpcode) {
2967 default: llvm_unreachable("Illegal integer comparison"); break;
2968 case ISD::SETNE: Invert = true;
2969 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2970 case ISD::SETLT: Swap = true;
2971 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2972 case ISD::SETLE: Swap = true;
2973 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2974 case ISD::SETULT: Swap = true;
2975 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2976 case ISD::SETULE: Swap = true;
2977 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2980 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2981 if (Opc == ARMISD::VCEQ) {
2984 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2986 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2989 // Ignore bitconvert.
2990 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2991 AndOp = AndOp.getOperand(0);
2993 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2995 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2996 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
3003 std::swap(Op0, Op1);
3005 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3008 Result = DAG.getNOT(dl, Result, VT);
3013 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3014 /// valid vector constant for a NEON instruction with a "modified immediate"
3015 /// operand (e.g., VMOV). If so, return the encoded value.
3016 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3017 unsigned SplatBitSize, SelectionDAG &DAG,
3018 EVT &VT, bool is128Bits, bool isVMOV) {
3019 unsigned OpCmode, Imm;
3021 // SplatBitSize is set to the smallest size that splats the vector, so a
3022 // zero vector will always have SplatBitSize == 8. However, NEON modified
3023 // immediate instructions others than VMOV do not support the 8-bit encoding
3024 // of a zero vector, and the default encoding of zero is supposed to be the
3029 switch (SplatBitSize) {
3033 // Any 1-byte value is OK. Op=0, Cmode=1110.
3034 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3037 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3041 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3042 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3043 if ((SplatBits & ~0xff) == 0) {
3044 // Value = 0x00nn: Op=x, Cmode=100x.
3049 if ((SplatBits & ~0xff00) == 0) {
3050 // Value = 0xnn00: Op=x, Cmode=101x.
3052 Imm = SplatBits >> 8;
3058 // NEON's 32-bit VMOV supports splat values where:
3059 // * only one byte is nonzero, or
3060 // * the least significant byte is 0xff and the second byte is nonzero, or
3061 // * the least significant 2 bytes are 0xff and the third is nonzero.
3062 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3063 if ((SplatBits & ~0xff) == 0) {
3064 // Value = 0x000000nn: Op=x, Cmode=000x.
3069 if ((SplatBits & ~0xff00) == 0) {
3070 // Value = 0x0000nn00: Op=x, Cmode=001x.
3072 Imm = SplatBits >> 8;
3075 if ((SplatBits & ~0xff0000) == 0) {
3076 // Value = 0x00nn0000: Op=x, Cmode=010x.
3078 Imm = SplatBits >> 16;
3081 if ((SplatBits & ~0xff000000) == 0) {
3082 // Value = 0xnn000000: Op=x, Cmode=011x.
3084 Imm = SplatBits >> 24;
3088 if ((SplatBits & ~0xffff) == 0 &&
3089 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3090 // Value = 0x0000nnff: Op=x, Cmode=1100.
3092 Imm = SplatBits >> 8;
3097 if ((SplatBits & ~0xffffff) == 0 &&
3098 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3099 // Value = 0x00nnffff: Op=x, Cmode=1101.
3101 Imm = SplatBits >> 16;
3102 SplatBits |= 0xffff;
3106 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3107 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3108 // VMOV.I32. A (very) minor optimization would be to replicate the value
3109 // and fall through here to test for a valid 64-bit splat. But, then the
3110 // caller would also need to check and handle the change in size.
3116 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3117 uint64_t BitMask = 0xff;
3119 unsigned ImmMask = 1;
3121 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3122 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3125 } else if ((SplatBits & BitMask) != 0) {
3131 // Op=1, Cmode=1110.
3134 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3139 llvm_unreachable("unexpected size for isNEONModifiedImm");
3143 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3144 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3147 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3148 bool &ReverseVEXT, unsigned &Imm) {
3149 unsigned NumElts = VT.getVectorNumElements();
3150 ReverseVEXT = false;
3152 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3158 // If this is a VEXT shuffle, the immediate value is the index of the first
3159 // element. The other shuffle indices must be the successive elements after
3161 unsigned ExpectedElt = Imm;
3162 for (unsigned i = 1; i < NumElts; ++i) {
3163 // Increment the expected index. If it wraps around, it may still be
3164 // a VEXT but the source vectors must be swapped.
3166 if (ExpectedElt == NumElts * 2) {
3171 if (M[i] < 0) continue; // ignore UNDEF indices
3172 if (ExpectedElt != static_cast<unsigned>(M[i]))
3176 // Adjust the index value if the source operands will be swapped.
3183 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3184 /// instruction with the specified blocksize. (The order of the elements
3185 /// within each block of the vector is reversed.)
3186 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3187 unsigned BlockSize) {
3188 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3189 "Only possible block sizes for VREV are: 16, 32, 64");
3191 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3195 unsigned NumElts = VT.getVectorNumElements();
3196 unsigned BlockElts = M[0] + 1;
3197 // If the first shuffle index is UNDEF, be optimistic.
3199 BlockElts = BlockSize / EltSz;
3201 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3204 for (unsigned i = 0; i < NumElts; ++i) {
3205 if (M[i] < 0) continue; // ignore UNDEF indices
3206 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3213 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3214 unsigned &WhichResult) {
3215 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3219 unsigned NumElts = VT.getVectorNumElements();
3220 WhichResult = (M[0] == 0 ? 0 : 1);
3221 for (unsigned i = 0; i < NumElts; i += 2) {
3222 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3223 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3229 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3230 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3231 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3232 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3233 unsigned &WhichResult) {
3234 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3238 unsigned NumElts = VT.getVectorNumElements();
3239 WhichResult = (M[0] == 0 ? 0 : 1);
3240 for (unsigned i = 0; i < NumElts; i += 2) {
3241 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3242 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3248 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3249 unsigned &WhichResult) {
3250 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3254 unsigned NumElts = VT.getVectorNumElements();
3255 WhichResult = (M[0] == 0 ? 0 : 1);
3256 for (unsigned i = 0; i != NumElts; ++i) {
3257 if (M[i] < 0) continue; // ignore UNDEF indices
3258 if ((unsigned) M[i] != 2 * i + WhichResult)
3262 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3263 if (VT.is64BitVector() && EltSz == 32)
3269 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3270 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3271 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3272 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3273 unsigned &WhichResult) {
3274 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3278 unsigned Half = VT.getVectorNumElements() / 2;
3279 WhichResult = (M[0] == 0 ? 0 : 1);
3280 for (unsigned j = 0; j != 2; ++j) {
3281 unsigned Idx = WhichResult;
3282 for (unsigned i = 0; i != Half; ++i) {
3283 int MIdx = M[i + j * Half];
3284 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3290 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3291 if (VT.is64BitVector() && EltSz == 32)
3297 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3298 unsigned &WhichResult) {
3299 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3303 unsigned NumElts = VT.getVectorNumElements();
3304 WhichResult = (M[0] == 0 ? 0 : 1);
3305 unsigned Idx = WhichResult * NumElts / 2;
3306 for (unsigned i = 0; i != NumElts; i += 2) {
3307 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3308 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3313 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3314 if (VT.is64BitVector() && EltSz == 32)
3320 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3321 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3322 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3323 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3324 unsigned &WhichResult) {
3325 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3329 unsigned NumElts = VT.getVectorNumElements();
3330 WhichResult = (M[0] == 0 ? 0 : 1);
3331 unsigned Idx = WhichResult * NumElts / 2;
3332 for (unsigned i = 0; i != NumElts; i += 2) {
3333 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3334 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3339 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3340 if (VT.is64BitVector() && EltSz == 32)
3346 // If N is an integer constant that can be moved into a register in one
3347 // instruction, return an SDValue of such a constant (will become a MOV
3348 // instruction). Otherwise return null.
3349 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3350 const ARMSubtarget *ST, DebugLoc dl) {
3352 if (!isa<ConstantSDNode>(N))
3354 Val = cast<ConstantSDNode>(N)->getZExtValue();
3356 if (ST->isThumb1Only()) {
3357 if (Val <= 255 || ~Val <= 255)
3358 return DAG.getConstant(Val, MVT::i32);
3360 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3361 return DAG.getConstant(Val, MVT::i32);
3366 // If this is a case we can't handle, return null and let the default
3367 // expansion code take care of it.
3368 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3369 const ARMSubtarget *ST) {
3370 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3371 DebugLoc dl = Op.getDebugLoc();
3372 EVT VT = Op.getValueType();
3374 APInt SplatBits, SplatUndef;
3375 unsigned SplatBitSize;
3377 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3378 if (SplatBitSize <= 64) {
3379 // Check if an immediate VMOV works.
3381 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3382 SplatUndef.getZExtValue(), SplatBitSize,
3383 DAG, VmovVT, VT.is128BitVector(), true);
3384 if (Val.getNode()) {
3385 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3386 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3389 // Try an immediate VMVN.
3390 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3391 ((1LL << SplatBitSize) - 1));
3392 Val = isNEONModifiedImm(NegatedImm,
3393 SplatUndef.getZExtValue(), SplatBitSize,
3394 DAG, VmovVT, VT.is128BitVector(), false);
3395 if (Val.getNode()) {
3396 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3397 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3402 // Scan through the operands to see if only one value is used.
3403 unsigned NumElts = VT.getVectorNumElements();
3404 bool isOnlyLowElement = true;
3405 bool usesOnlyOneValue = true;
3406 bool isConstant = true;
3408 for (unsigned i = 0; i < NumElts; ++i) {
3409 SDValue V = Op.getOperand(i);
3410 if (V.getOpcode() == ISD::UNDEF)
3413 isOnlyLowElement = false;
3414 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3417 if (!Value.getNode())
3419 else if (V != Value)
3420 usesOnlyOneValue = false;
3423 if (!Value.getNode())
3424 return DAG.getUNDEF(VT);
3426 if (isOnlyLowElement)
3427 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3429 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3431 if (EnableARMVDUPsplat) {
3432 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3433 // i32 and try again.
3434 if (usesOnlyOneValue && EltSize <= 32) {
3436 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3437 if (VT.getVectorElementType().isFloatingPoint()) {
3438 SmallVector<SDValue, 8> Ops;
3439 for (unsigned i = 0; i < NumElts; ++i)
3440 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3442 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3444 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3445 LowerBUILD_VECTOR(Val, DAG, ST));
3447 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3449 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3453 // If all elements are constants and the case above didn't get hit, fall back
3454 // to the default expansion, which will generate a load from the constant
3459 if (!EnableARMVDUPsplat) {
3460 // Use VDUP for non-constant splats.
3461 if (usesOnlyOneValue && EltSize <= 32)
3462 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3465 // Vectors with 32- or 64-bit elements can be built by directly assigning
3466 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3467 // will be legalized.
3468 if (EltSize >= 32) {
3469 // Do the expansion with floating-point types, since that is what the VFP
3470 // registers are defined to use, and since i64 is not legal.
3471 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3472 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3473 SmallVector<SDValue, 8> Ops;
3474 for (unsigned i = 0; i < NumElts; ++i)
3475 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3476 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3477 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3483 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3484 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3485 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3486 /// are assumed to be legal.
3488 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3490 if (VT.getVectorNumElements() == 4 &&
3491 (VT.is128BitVector() || VT.is64BitVector())) {
3492 unsigned PFIndexes[4];
3493 for (unsigned i = 0; i != 4; ++i) {
3497 PFIndexes[i] = M[i];
3500 // Compute the index in the perfect shuffle table.
3501 unsigned PFTableIndex =
3502 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3503 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3504 unsigned Cost = (PFEntry >> 30);
3511 unsigned Imm, WhichResult;
3513 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3514 return (EltSize >= 32 ||
3515 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3516 isVREVMask(M, VT, 64) ||
3517 isVREVMask(M, VT, 32) ||
3518 isVREVMask(M, VT, 16) ||
3519 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3520 isVTRNMask(M, VT, WhichResult) ||
3521 isVUZPMask(M, VT, WhichResult) ||
3522 isVZIPMask(M, VT, WhichResult) ||
3523 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3524 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3525 isVZIP_v_undef_Mask(M, VT, WhichResult));
3528 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3529 /// the specified operations to build the shuffle.
3530 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3531 SDValue RHS, SelectionDAG &DAG,
3533 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3534 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3535 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3538 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3547 OP_VUZPL, // VUZP, left result
3548 OP_VUZPR, // VUZP, right result
3549 OP_VZIPL, // VZIP, left result
3550 OP_VZIPR, // VZIP, right result
3551 OP_VTRNL, // VTRN, left result
3552 OP_VTRNR // VTRN, right result
3555 if (OpNum == OP_COPY) {
3556 if (LHSID == (1*9+2)*9+3) return LHS;
3557 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3561 SDValue OpLHS, OpRHS;
3562 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3563 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3564 EVT VT = OpLHS.getValueType();
3567 default: llvm_unreachable("Unknown shuffle opcode!");
3569 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3574 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3575 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3579 return DAG.getNode(ARMISD::VEXT, dl, VT,
3581 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3584 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3585 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3588 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3589 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3592 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3593 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3597 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3598 SDValue V1 = Op.getOperand(0);
3599 SDValue V2 = Op.getOperand(1);
3600 DebugLoc dl = Op.getDebugLoc();
3601 EVT VT = Op.getValueType();
3602 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3603 SmallVector<int, 8> ShuffleMask;
3605 // Convert shuffles that are directly supported on NEON to target-specific
3606 // DAG nodes, instead of keeping them as shuffles and matching them again
3607 // during code selection. This is more efficient and avoids the possibility
3608 // of inconsistencies between legalization and selection.
3609 // FIXME: floating-point vectors should be canonicalized to integer vectors
3610 // of the same time so that they get CSEd properly.
3611 SVN->getMask(ShuffleMask);
3613 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3614 if (EltSize <= 32) {
3615 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3616 int Lane = SVN->getSplatIndex();
3617 // If this is undef splat, generate it via "just" vdup, if possible.
3618 if (Lane == -1) Lane = 0;
3620 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3621 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3623 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3624 DAG.getConstant(Lane, MVT::i32));
3629 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3632 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3633 DAG.getConstant(Imm, MVT::i32));
3636 if (isVREVMask(ShuffleMask, VT, 64))
3637 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3638 if (isVREVMask(ShuffleMask, VT, 32))
3639 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3640 if (isVREVMask(ShuffleMask, VT, 16))
3641 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3643 // Check for Neon shuffles that modify both input vectors in place.
3644 // If both results are used, i.e., if there are two shuffles with the same
3645 // source operands and with masks corresponding to both results of one of
3646 // these operations, DAG memoization will ensure that a single node is
3647 // used for both shuffles.
3648 unsigned WhichResult;
3649 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3650 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3651 V1, V2).getValue(WhichResult);
3652 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3653 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3654 V1, V2).getValue(WhichResult);
3655 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3656 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3657 V1, V2).getValue(WhichResult);
3659 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3660 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3661 V1, V1).getValue(WhichResult);
3662 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3663 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3664 V1, V1).getValue(WhichResult);
3665 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3666 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3667 V1, V1).getValue(WhichResult);
3670 // If the shuffle is not directly supported and it has 4 elements, use
3671 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3672 unsigned NumElts = VT.getVectorNumElements();
3674 unsigned PFIndexes[4];
3675 for (unsigned i = 0; i != 4; ++i) {
3676 if (ShuffleMask[i] < 0)
3679 PFIndexes[i] = ShuffleMask[i];
3682 // Compute the index in the perfect shuffle table.
3683 unsigned PFTableIndex =
3684 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3685 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3686 unsigned Cost = (PFEntry >> 30);
3689 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3692 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3693 if (EltSize >= 32) {
3694 // Do the expansion with floating-point types, since that is what the VFP
3695 // registers are defined to use, and since i64 is not legal.
3696 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3697 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3698 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3699 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3700 SmallVector<SDValue, 8> Ops;
3701 for (unsigned i = 0; i < NumElts; ++i) {
3702 if (ShuffleMask[i] < 0)
3703 Ops.push_back(DAG.getUNDEF(EltVT));
3705 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3706 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3707 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3710 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3711 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3717 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3718 EVT VT = Op.getValueType();
3719 DebugLoc dl = Op.getDebugLoc();
3720 SDValue Vec = Op.getOperand(0);
3721 SDValue Lane = Op.getOperand(1);
3722 assert(VT == MVT::i32 &&
3723 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3724 "unexpected type for custom-lowering vector extract");
3725 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3728 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3729 // The only time a CONCAT_VECTORS operation can have legal types is when
3730 // two 64-bit vectors are concatenated to a 128-bit vector.
3731 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3732 "unexpected CONCAT_VECTORS");
3733 DebugLoc dl = Op.getDebugLoc();
3734 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3735 SDValue Op0 = Op.getOperand(0);
3736 SDValue Op1 = Op.getOperand(1);
3737 if (Op0.getOpcode() != ISD::UNDEF)
3738 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3739 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3740 DAG.getIntPtrConstant(0));
3741 if (Op1.getOpcode() != ISD::UNDEF)
3742 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3743 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3744 DAG.getIntPtrConstant(1));
3745 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3748 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3749 switch (Op.getOpcode()) {
3750 default: llvm_unreachable("Don't know how to custom lower this!");
3751 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3752 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3753 case ISD::GlobalAddress:
3754 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3755 LowerGlobalAddressELF(Op, DAG);
3756 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3757 case ISD::SELECT: return LowerSELECT(Op, DAG);
3758 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3759 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3760 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3761 case ISD::VASTART: return LowerVASTART(Op, DAG);
3762 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3763 case ISD::SINT_TO_FP:
3764 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3765 case ISD::FP_TO_SINT:
3766 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3767 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3768 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3769 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3770 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3771 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3772 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3773 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3775 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3778 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3779 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3780 case ISD::SRL_PARTS:
3781 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3782 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3783 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3784 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3785 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3786 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3787 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3788 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3793 /// ReplaceNodeResults - Replace the results of node with an illegal result
3794 /// type with new values built out of custom code.
3795 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3796 SmallVectorImpl<SDValue>&Results,
3797 SelectionDAG &DAG) const {
3799 switch (N->getOpcode()) {
3801 llvm_unreachable("Don't know how to custom expand this!");
3803 case ISD::BIT_CONVERT:
3804 Res = ExpandBIT_CONVERT(N, DAG);
3808 Res = LowerShift(N, DAG, Subtarget);
3812 Results.push_back(Res);
3815 //===----------------------------------------------------------------------===//
3816 // ARM Scheduler Hooks
3817 //===----------------------------------------------------------------------===//
3820 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3821 MachineBasicBlock *BB,
3822 unsigned Size) const {
3823 unsigned dest = MI->getOperand(0).getReg();
3824 unsigned ptr = MI->getOperand(1).getReg();
3825 unsigned oldval = MI->getOperand(2).getReg();
3826 unsigned newval = MI->getOperand(3).getReg();
3827 unsigned scratch = BB->getParent()->getRegInfo()
3828 .createVirtualRegister(ARM::GPRRegisterClass);
3829 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3830 DebugLoc dl = MI->getDebugLoc();
3831 bool isThumb2 = Subtarget->isThumb2();
3833 unsigned ldrOpc, strOpc;
3835 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3837 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3838 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3841 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3842 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3845 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3846 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3850 MachineFunction *MF = BB->getParent();
3851 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3852 MachineFunction::iterator It = BB;
3853 ++It; // insert the new blocks after the current block
3855 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3856 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3857 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3858 MF->insert(It, loop1MBB);
3859 MF->insert(It, loop2MBB);
3860 MF->insert(It, exitMBB);
3862 // Transfer the remainder of BB and its successor edges to exitMBB.
3863 exitMBB->splice(exitMBB->begin(), BB,
3864 llvm::next(MachineBasicBlock::iterator(MI)),
3866 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3870 // fallthrough --> loop1MBB
3871 BB->addSuccessor(loop1MBB);
3874 // ldrex dest, [ptr]
3878 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3879 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3880 .addReg(dest).addReg(oldval));
3881 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3882 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3883 BB->addSuccessor(loop2MBB);
3884 BB->addSuccessor(exitMBB);
3887 // strex scratch, newval, [ptr]
3891 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3893 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3894 .addReg(scratch).addImm(0));
3895 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3896 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3897 BB->addSuccessor(loop1MBB);
3898 BB->addSuccessor(exitMBB);
3904 MI->eraseFromParent(); // The instruction is gone now.
3910 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3911 unsigned Size, unsigned BinOpcode) const {
3912 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3913 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3915 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3916 MachineFunction *MF = BB->getParent();
3917 MachineFunction::iterator It = BB;
3920 unsigned dest = MI->getOperand(0).getReg();
3921 unsigned ptr = MI->getOperand(1).getReg();
3922 unsigned incr = MI->getOperand(2).getReg();
3923 DebugLoc dl = MI->getDebugLoc();
3925 bool isThumb2 = Subtarget->isThumb2();
3926 unsigned ldrOpc, strOpc;
3928 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3930 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3931 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3934 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3935 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3938 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3939 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3943 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3944 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3945 MF->insert(It, loopMBB);
3946 MF->insert(It, exitMBB);
3948 // Transfer the remainder of BB and its successor edges to exitMBB.
3949 exitMBB->splice(exitMBB->begin(), BB,
3950 llvm::next(MachineBasicBlock::iterator(MI)),
3952 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3954 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3955 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3956 unsigned scratch2 = (!BinOpcode) ? incr :
3957 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3961 // fallthrough --> loopMBB
3962 BB->addSuccessor(loopMBB);
3966 // <binop> scratch2, dest, incr
3967 // strex scratch, scratch2, ptr
3970 // fallthrough --> exitMBB
3972 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3974 // operand order needs to go the other way for NAND
3975 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3976 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3977 addReg(incr).addReg(dest)).addReg(0);
3979 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3980 addReg(dest).addReg(incr)).addReg(0);
3983 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3985 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3986 .addReg(scratch).addImm(0));
3987 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3988 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3990 BB->addSuccessor(loopMBB);
3991 BB->addSuccessor(exitMBB);
3997 MI->eraseFromParent(); // The instruction is gone now.
4003 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
4004 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
4005 E = MBB->succ_end(); I != E; ++I)
4008 llvm_unreachable("Expecting a BB with two successors!");
4012 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4013 MachineBasicBlock *BB) const {
4014 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4015 DebugLoc dl = MI->getDebugLoc();
4016 bool isThumb2 = Subtarget->isThumb2();
4017 switch (MI->getOpcode()) {
4020 llvm_unreachable("Unexpected instr type to insert");
4022 case ARM::ATOMIC_LOAD_ADD_I8:
4023 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4024 case ARM::ATOMIC_LOAD_ADD_I16:
4025 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4026 case ARM::ATOMIC_LOAD_ADD_I32:
4027 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4029 case ARM::ATOMIC_LOAD_AND_I8:
4030 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4031 case ARM::ATOMIC_LOAD_AND_I16:
4032 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4033 case ARM::ATOMIC_LOAD_AND_I32:
4034 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4036 case ARM::ATOMIC_LOAD_OR_I8:
4037 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4038 case ARM::ATOMIC_LOAD_OR_I16:
4039 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4040 case ARM::ATOMIC_LOAD_OR_I32:
4041 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4043 case ARM::ATOMIC_LOAD_XOR_I8:
4044 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4045 case ARM::ATOMIC_LOAD_XOR_I16:
4046 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4047 case ARM::ATOMIC_LOAD_XOR_I32:
4048 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4050 case ARM::ATOMIC_LOAD_NAND_I8:
4051 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4052 case ARM::ATOMIC_LOAD_NAND_I16:
4053 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4054 case ARM::ATOMIC_LOAD_NAND_I32:
4055 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4057 case ARM::ATOMIC_LOAD_SUB_I8:
4058 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4059 case ARM::ATOMIC_LOAD_SUB_I16:
4060 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4061 case ARM::ATOMIC_LOAD_SUB_I32:
4062 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4064 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4065 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4066 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4068 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4069 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4070 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4072 case ARM::tMOVCCr_pseudo: {
4073 // To "insert" a SELECT_CC instruction, we actually have to insert the
4074 // diamond control-flow pattern. The incoming instruction knows the
4075 // destination vreg to set, the condition code register to branch on, the
4076 // true/false values to select between, and a branch opcode to use.
4077 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4078 MachineFunction::iterator It = BB;
4084 // cmpTY ccX, r1, r2
4086 // fallthrough --> copy0MBB
4087 MachineBasicBlock *thisMBB = BB;
4088 MachineFunction *F = BB->getParent();
4089 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4090 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4091 F->insert(It, copy0MBB);
4092 F->insert(It, sinkMBB);
4094 // Transfer the remainder of BB and its successor edges to sinkMBB.
4095 sinkMBB->splice(sinkMBB->begin(), BB,
4096 llvm::next(MachineBasicBlock::iterator(MI)),
4098 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4100 BB->addSuccessor(copy0MBB);
4101 BB->addSuccessor(sinkMBB);
4103 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4104 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4107 // %FalseValue = ...
4108 // # fallthrough to sinkMBB
4111 // Update machine-CFG edges
4112 BB->addSuccessor(sinkMBB);
4115 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4118 BuildMI(*BB, BB->begin(), dl,
4119 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4120 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4121 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4123 MI->eraseFromParent(); // The pseudo instruction is gone now.
4128 case ARM::BCCZi64: {
4129 // Compare both parts that make up the double comparison separately for
4131 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4133 unsigned LHS1 = MI->getOperand(1).getReg();
4134 unsigned LHS2 = MI->getOperand(2).getReg();
4136 AddDefaultPred(BuildMI(BB, dl,
4137 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4138 .addReg(LHS1).addImm(0));
4139 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4140 .addReg(LHS2).addImm(0)
4141 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4143 unsigned RHS1 = MI->getOperand(3).getReg();
4144 unsigned RHS2 = MI->getOperand(4).getReg();
4145 AddDefaultPred(BuildMI(BB, dl,
4146 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4147 .addReg(LHS1).addReg(RHS1));
4148 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4149 .addReg(LHS2).addReg(RHS2)
4150 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4153 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4154 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4155 if (MI->getOperand(0).getImm() == ARMCC::NE)
4156 std::swap(destMBB, exitMBB);
4158 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4159 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4160 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4163 MI->eraseFromParent(); // The pseudo instruction is gone now.
4169 //===----------------------------------------------------------------------===//
4170 // ARM Optimization Hooks
4171 //===----------------------------------------------------------------------===//
4174 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4175 TargetLowering::DAGCombinerInfo &DCI) {
4176 SelectionDAG &DAG = DCI.DAG;
4177 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4178 EVT VT = N->getValueType(0);
4179 unsigned Opc = N->getOpcode();
4180 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4181 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4182 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4183 ISD::CondCode CC = ISD::SETCC_INVALID;
4186 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4188 SDValue CCOp = Slct.getOperand(0);
4189 if (CCOp.getOpcode() == ISD::SETCC)
4190 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4193 bool DoXform = false;
4195 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4198 if (LHS.getOpcode() == ISD::Constant &&
4199 cast<ConstantSDNode>(LHS)->isNullValue()) {
4201 } else if (CC != ISD::SETCC_INVALID &&
4202 RHS.getOpcode() == ISD::Constant &&
4203 cast<ConstantSDNode>(RHS)->isNullValue()) {
4204 std::swap(LHS, RHS);
4205 SDValue Op0 = Slct.getOperand(0);
4206 EVT OpVT = isSlctCC ? Op0.getValueType() :
4207 Op0.getOperand(0).getValueType();
4208 bool isInt = OpVT.isInteger();
4209 CC = ISD::getSetCCInverse(CC, isInt);
4211 if (!TLI.isCondCodeLegal(CC, OpVT))
4212 return SDValue(); // Inverse operator isn't legal.
4219 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4221 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4222 Slct.getOperand(0), Slct.getOperand(1), CC);
4223 SDValue CCOp = Slct.getOperand(0);
4225 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4226 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4227 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4228 CCOp, OtherOp, Result);
4233 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4234 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4235 /// called with the default operands, and if that fails, with commuted
4237 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4238 TargetLowering::DAGCombinerInfo &DCI) {
4239 SelectionDAG &DAG = DCI.DAG;
4241 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4242 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4243 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4244 if (Result.getNode()) return Result;
4247 // fold (add (arm_neon_vabd a, b) c) -> (arm_neon_vaba c, a, b)
4248 EVT VT = N->getValueType(0);
4249 if (N0.getOpcode() == ISD::INTRINSIC_WO_CHAIN && VT.isInteger()) {
4250 unsigned IntNo = cast<ConstantSDNode>(N0.getOperand(0))->getZExtValue();
4251 if (IntNo == Intrinsic::arm_neon_vabds)
4252 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), VT,
4253 DAG.getConstant(Intrinsic::arm_neon_vabas, MVT::i32),
4254 N1, N0.getOperand(1), N0.getOperand(2));
4255 if (IntNo == Intrinsic::arm_neon_vabdu)
4256 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), VT,
4257 DAG.getConstant(Intrinsic::arm_neon_vabau, MVT::i32),
4258 N1, N0.getOperand(1), N0.getOperand(2));
4264 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4266 static SDValue PerformADDCombine(SDNode *N,
4267 TargetLowering::DAGCombinerInfo &DCI) {
4268 SDValue N0 = N->getOperand(0);
4269 SDValue N1 = N->getOperand(1);
4271 // First try with the default operand order.
4272 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4273 if (Result.getNode())
4276 // If that didn't work, try again with the operands commuted.
4277 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4280 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4282 static SDValue PerformSUBCombine(SDNode *N,
4283 TargetLowering::DAGCombinerInfo &DCI) {
4284 SDValue N0 = N->getOperand(0);
4285 SDValue N1 = N->getOperand(1);
4287 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4288 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4289 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4290 if (Result.getNode()) return Result;
4296 static SDValue PerformMULCombine(SDNode *N,
4297 TargetLowering::DAGCombinerInfo &DCI,
4298 const ARMSubtarget *Subtarget) {
4299 SelectionDAG &DAG = DCI.DAG;
4301 if (Subtarget->isThumb1Only())
4304 if (DAG.getMachineFunction().
4305 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4308 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4311 EVT VT = N->getValueType(0);
4315 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4319 uint64_t MulAmt = C->getZExtValue();
4320 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4321 ShiftAmt = ShiftAmt & (32 - 1);
4322 SDValue V = N->getOperand(0);
4323 DebugLoc DL = N->getDebugLoc();
4326 MulAmt >>= ShiftAmt;
4327 if (isPowerOf2_32(MulAmt - 1)) {
4328 // (mul x, 2^N + 1) => (add (shl x, N), x)
4329 Res = DAG.getNode(ISD::ADD, DL, VT,
4330 V, DAG.getNode(ISD::SHL, DL, VT,
4331 V, DAG.getConstant(Log2_32(MulAmt-1),
4333 } else if (isPowerOf2_32(MulAmt + 1)) {
4334 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4335 Res = DAG.getNode(ISD::SUB, DL, VT,
4336 DAG.getNode(ISD::SHL, DL, VT,
4337 V, DAG.getConstant(Log2_32(MulAmt+1),
4344 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4345 DAG.getConstant(ShiftAmt, MVT::i32));
4347 // Do not add new nodes to DAG combiner worklist.
4348 DCI.CombineTo(N, Res, false);
4352 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4353 static SDValue PerformORCombine(SDNode *N,
4354 TargetLowering::DAGCombinerInfo &DCI,
4355 const ARMSubtarget *Subtarget) {
4356 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4359 // BFI is only available on V6T2+
4360 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4363 SelectionDAG &DAG = DCI.DAG;
4364 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4365 DebugLoc DL = N->getDebugLoc();
4366 // 1) or (and A, mask), val => ARMbfi A, val, mask
4367 // iff (val & mask) == val
4369 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4370 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4371 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4372 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4373 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4374 // (i.e., copy a bitfield value into another bitfield of the same width)
4375 if (N0.getOpcode() != ISD::AND)
4378 EVT VT = N->getValueType(0);
4383 // The value and the mask need to be constants so we can verify this is
4384 // actually a bitfield set. If the mask is 0xffff, we can do better
4385 // via a movt instruction, so don't use BFI in that case.
4386 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4389 unsigned Mask = C->getZExtValue();
4393 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4394 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4395 unsigned Val = C->getZExtValue();
4396 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4398 Val >>= CountTrailingZeros_32(~Mask);
4400 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4401 DAG.getConstant(Val, MVT::i32),
4402 DAG.getConstant(Mask, MVT::i32));
4404 // Do not add new nodes to DAG combiner worklist.
4405 DCI.CombineTo(N, Res, false);
4406 } else if (N1.getOpcode() == ISD::AND) {
4407 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4408 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4411 unsigned Mask2 = C->getZExtValue();
4413 if (ARM::isBitFieldInvertedMask(Mask) &&
4414 ARM::isBitFieldInvertedMask(~Mask2) &&
4415 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4416 // The pack halfword instruction works better for masks that fit it,
4417 // so use that when it's available.
4418 if (Subtarget->hasT2ExtractPack() &&
4419 (Mask == 0xffff || Mask == 0xffff0000))
4422 unsigned lsb = CountTrailingZeros_32(Mask2);
4423 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4424 DAG.getConstant(lsb, MVT::i32));
4425 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4426 DAG.getConstant(Mask, MVT::i32));
4427 // Do not add new nodes to DAG combiner worklist.
4428 DCI.CombineTo(N, Res, false);
4429 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4430 ARM::isBitFieldInvertedMask(Mask2) &&
4431 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4432 // The pack halfword instruction works better for masks that fit it,
4433 // so use that when it's available.
4434 if (Subtarget->hasT2ExtractPack() &&
4435 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4438 unsigned lsb = CountTrailingZeros_32(Mask);
4439 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4440 DAG.getConstant(lsb, MVT::i32));
4441 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4442 DAG.getConstant(Mask2, MVT::i32));
4443 // Do not add new nodes to DAG combiner worklist.
4444 DCI.CombineTo(N, Res, false);
4451 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4452 /// ARMISD::VMOVRRD.
4453 static SDValue PerformVMOVRRDCombine(SDNode *N,
4454 TargetLowering::DAGCombinerInfo &DCI) {
4455 // fmrrd(fmdrr x, y) -> x,y
4456 SDValue InDouble = N->getOperand(0);
4457 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4458 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4462 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4463 /// ARMISD::VDUPLANE.
4464 static SDValue PerformVDUPLANECombine(SDNode *N,
4465 TargetLowering::DAGCombinerInfo &DCI) {
4466 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4468 SDValue Op = N->getOperand(0);
4469 EVT VT = N->getValueType(0);
4471 // Ignore bit_converts.
4472 while (Op.getOpcode() == ISD::BIT_CONVERT)
4473 Op = Op.getOperand(0);
4474 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4477 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4478 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4479 // The canonical VMOV for a zero vector uses a 32-bit element size.
4480 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4482 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4484 if (EltSize > VT.getVectorElementType().getSizeInBits())
4487 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4488 return DCI.CombineTo(N, Res, false);
4491 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4492 /// operand of a vector shift operation, where all the elements of the
4493 /// build_vector must have the same constant integer value.
4494 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4495 // Ignore bit_converts.
4496 while (Op.getOpcode() == ISD::BIT_CONVERT)
4497 Op = Op.getOperand(0);
4498 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4499 APInt SplatBits, SplatUndef;
4500 unsigned SplatBitSize;
4502 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4503 HasAnyUndefs, ElementBits) ||
4504 SplatBitSize > ElementBits)
4506 Cnt = SplatBits.getSExtValue();
4510 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4511 /// operand of a vector shift left operation. That value must be in the range:
4512 /// 0 <= Value < ElementBits for a left shift; or
4513 /// 0 <= Value <= ElementBits for a long left shift.
4514 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4515 assert(VT.isVector() && "vector shift count is not a vector type");
4516 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4517 if (! getVShiftImm(Op, ElementBits, Cnt))
4519 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4522 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4523 /// operand of a vector shift right operation. For a shift opcode, the value
4524 /// is positive, but for an intrinsic the value count must be negative. The
4525 /// absolute value must be in the range:
4526 /// 1 <= |Value| <= ElementBits for a right shift; or
4527 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4528 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4530 assert(VT.isVector() && "vector shift count is not a vector type");
4531 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4532 if (! getVShiftImm(Op, ElementBits, Cnt))
4536 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4539 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4540 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4541 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4544 // Don't do anything for most intrinsics.
4547 // Vector shifts: check for immediate versions and lower them.
4548 // Note: This is done during DAG combining instead of DAG legalizing because
4549 // the build_vectors for 64-bit vector element shift counts are generally
4550 // not legal, and it is hard to see their values after they get legalized to
4551 // loads from a constant pool.
4552 case Intrinsic::arm_neon_vshifts:
4553 case Intrinsic::arm_neon_vshiftu:
4554 case Intrinsic::arm_neon_vshiftls:
4555 case Intrinsic::arm_neon_vshiftlu:
4556 case Intrinsic::arm_neon_vshiftn:
4557 case Intrinsic::arm_neon_vrshifts:
4558 case Intrinsic::arm_neon_vrshiftu:
4559 case Intrinsic::arm_neon_vrshiftn:
4560 case Intrinsic::arm_neon_vqshifts:
4561 case Intrinsic::arm_neon_vqshiftu:
4562 case Intrinsic::arm_neon_vqshiftsu:
4563 case Intrinsic::arm_neon_vqshiftns:
4564 case Intrinsic::arm_neon_vqshiftnu:
4565 case Intrinsic::arm_neon_vqshiftnsu:
4566 case Intrinsic::arm_neon_vqrshiftns:
4567 case Intrinsic::arm_neon_vqrshiftnu:
4568 case Intrinsic::arm_neon_vqrshiftnsu: {
4569 EVT VT = N->getOperand(1).getValueType();
4571 unsigned VShiftOpc = 0;
4574 case Intrinsic::arm_neon_vshifts:
4575 case Intrinsic::arm_neon_vshiftu:
4576 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4577 VShiftOpc = ARMISD::VSHL;
4580 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4581 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4582 ARMISD::VSHRs : ARMISD::VSHRu);
4587 case Intrinsic::arm_neon_vshiftls:
4588 case Intrinsic::arm_neon_vshiftlu:
4589 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4591 llvm_unreachable("invalid shift count for vshll intrinsic");
4593 case Intrinsic::arm_neon_vrshifts:
4594 case Intrinsic::arm_neon_vrshiftu:
4595 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4599 case Intrinsic::arm_neon_vqshifts:
4600 case Intrinsic::arm_neon_vqshiftu:
4601 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4605 case Intrinsic::arm_neon_vqshiftsu:
4606 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4608 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4610 case Intrinsic::arm_neon_vshiftn:
4611 case Intrinsic::arm_neon_vrshiftn:
4612 case Intrinsic::arm_neon_vqshiftns:
4613 case Intrinsic::arm_neon_vqshiftnu:
4614 case Intrinsic::arm_neon_vqshiftnsu:
4615 case Intrinsic::arm_neon_vqrshiftns:
4616 case Intrinsic::arm_neon_vqrshiftnu:
4617 case Intrinsic::arm_neon_vqrshiftnsu:
4618 // Narrowing shifts require an immediate right shift.
4619 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4621 llvm_unreachable("invalid shift count for narrowing vector shift "
4625 llvm_unreachable("unhandled vector shift");
4629 case Intrinsic::arm_neon_vshifts:
4630 case Intrinsic::arm_neon_vshiftu:
4631 // Opcode already set above.
4633 case Intrinsic::arm_neon_vshiftls:
4634 case Intrinsic::arm_neon_vshiftlu:
4635 if (Cnt == VT.getVectorElementType().getSizeInBits())
4636 VShiftOpc = ARMISD::VSHLLi;
4638 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4639 ARMISD::VSHLLs : ARMISD::VSHLLu);
4641 case Intrinsic::arm_neon_vshiftn:
4642 VShiftOpc = ARMISD::VSHRN; break;
4643 case Intrinsic::arm_neon_vrshifts:
4644 VShiftOpc = ARMISD::VRSHRs; break;
4645 case Intrinsic::arm_neon_vrshiftu:
4646 VShiftOpc = ARMISD::VRSHRu; break;
4647 case Intrinsic::arm_neon_vrshiftn:
4648 VShiftOpc = ARMISD::VRSHRN; break;
4649 case Intrinsic::arm_neon_vqshifts:
4650 VShiftOpc = ARMISD::VQSHLs; break;
4651 case Intrinsic::arm_neon_vqshiftu:
4652 VShiftOpc = ARMISD::VQSHLu; break;
4653 case Intrinsic::arm_neon_vqshiftsu:
4654 VShiftOpc = ARMISD::VQSHLsu; break;
4655 case Intrinsic::arm_neon_vqshiftns:
4656 VShiftOpc = ARMISD::VQSHRNs; break;
4657 case Intrinsic::arm_neon_vqshiftnu:
4658 VShiftOpc = ARMISD::VQSHRNu; break;
4659 case Intrinsic::arm_neon_vqshiftnsu:
4660 VShiftOpc = ARMISD::VQSHRNsu; break;
4661 case Intrinsic::arm_neon_vqrshiftns:
4662 VShiftOpc = ARMISD::VQRSHRNs; break;
4663 case Intrinsic::arm_neon_vqrshiftnu:
4664 VShiftOpc = ARMISD::VQRSHRNu; break;
4665 case Intrinsic::arm_neon_vqrshiftnsu:
4666 VShiftOpc = ARMISD::VQRSHRNsu; break;
4669 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4670 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4673 case Intrinsic::arm_neon_vshiftins: {
4674 EVT VT = N->getOperand(1).getValueType();
4676 unsigned VShiftOpc = 0;
4678 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4679 VShiftOpc = ARMISD::VSLI;
4680 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4681 VShiftOpc = ARMISD::VSRI;
4683 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4686 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4687 N->getOperand(1), N->getOperand(2),
4688 DAG.getConstant(Cnt, MVT::i32));
4691 case Intrinsic::arm_neon_vqrshifts:
4692 case Intrinsic::arm_neon_vqrshiftu:
4693 // No immediate versions of these to check for.
4700 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4701 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4702 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4703 /// vector element shift counts are generally not legal, and it is hard to see
4704 /// their values after they get legalized to loads from a constant pool.
4705 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4706 const ARMSubtarget *ST) {
4707 EVT VT = N->getValueType(0);
4709 // Nothing to be done for scalar shifts.
4710 if (! VT.isVector())
4713 assert(ST->hasNEON() && "unexpected vector shift");
4716 switch (N->getOpcode()) {
4717 default: llvm_unreachable("unexpected shift opcode");
4720 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4721 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4722 DAG.getConstant(Cnt, MVT::i32));
4727 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4728 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4729 ARMISD::VSHRs : ARMISD::VSHRu);
4730 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4731 DAG.getConstant(Cnt, MVT::i32));
4737 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4738 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4739 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4740 const ARMSubtarget *ST) {
4741 SDValue N0 = N->getOperand(0);
4743 // Check for sign- and zero-extensions of vector extract operations of 8-
4744 // and 16-bit vector elements. NEON supports these directly. They are
4745 // handled during DAG combining because type legalization will promote them
4746 // to 32-bit types and it is messy to recognize the operations after that.
4747 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4748 SDValue Vec = N0.getOperand(0);
4749 SDValue Lane = N0.getOperand(1);
4750 EVT VT = N->getValueType(0);
4751 EVT EltVT = N0.getValueType();
4752 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4754 if (VT == MVT::i32 &&
4755 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4756 TLI.isTypeLegal(Vec.getValueType())) {
4759 switch (N->getOpcode()) {
4760 default: llvm_unreachable("unexpected opcode");
4761 case ISD::SIGN_EXTEND:
4762 Opc = ARMISD::VGETLANEs;
4764 case ISD::ZERO_EXTEND:
4765 case ISD::ANY_EXTEND:
4766 Opc = ARMISD::VGETLANEu;
4769 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4776 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4777 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4778 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4779 const ARMSubtarget *ST) {
4780 // If the target supports NEON, try to use vmax/vmin instructions for f32
4781 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4782 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4783 // a NaN; only do the transformation when it matches that behavior.
4785 // For now only do this when using NEON for FP operations; if using VFP, it
4786 // is not obvious that the benefit outweighs the cost of switching to the
4788 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4789 N->getValueType(0) != MVT::f32)
4792 SDValue CondLHS = N->getOperand(0);
4793 SDValue CondRHS = N->getOperand(1);
4794 SDValue LHS = N->getOperand(2);
4795 SDValue RHS = N->getOperand(3);
4796 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4798 unsigned Opcode = 0;
4800 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4801 IsReversed = false; // x CC y ? x : y
4802 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4803 IsReversed = true ; // x CC y ? y : x
4817 // If LHS is NaN, an ordered comparison will be false and the result will
4818 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4819 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4820 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4821 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4823 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4824 // will return -0, so vmin can only be used for unsafe math or if one of
4825 // the operands is known to be nonzero.
4826 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4828 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4830 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4839 // If LHS is NaN, an ordered comparison will be false and the result will
4840 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4841 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4842 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4843 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4845 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4846 // will return +0, so vmax can only be used for unsafe math or if one of
4847 // the operands is known to be nonzero.
4848 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4850 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4852 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4858 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4861 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4862 DAGCombinerInfo &DCI) const {
4863 switch (N->getOpcode()) {
4865 case ISD::ADD: return PerformADDCombine(N, DCI);
4866 case ISD::SUB: return PerformSUBCombine(N, DCI);
4867 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4868 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4869 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4870 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4871 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4874 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4875 case ISD::SIGN_EXTEND:
4876 case ISD::ZERO_EXTEND:
4877 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4878 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4883 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4884 if (!Subtarget->hasV6Ops())
4885 // Pre-v6 does not support unaligned mem access.
4888 // v6+ may or may not support unaligned mem access depending on the system
4890 // FIXME: This is pretty conservative. Should we provide cmdline option to
4891 // control the behaviour?
4892 if (!Subtarget->isTargetDarwin())
4895 switch (VT.getSimpleVT().SimpleTy) {
4902 // FIXME: VLD1 etc with standard alignment is legal.
4906 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4911 switch (VT.getSimpleVT().SimpleTy) {
4912 default: return false;
4927 if ((V & (Scale - 1)) != 0)
4930 return V == (V & ((1LL << 5) - 1));
4933 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4934 const ARMSubtarget *Subtarget) {
4941 switch (VT.getSimpleVT().SimpleTy) {
4942 default: return false;
4947 // + imm12 or - imm8
4949 return V == (V & ((1LL << 8) - 1));
4950 return V == (V & ((1LL << 12) - 1));
4953 // Same as ARM mode. FIXME: NEON?
4954 if (!Subtarget->hasVFP2())
4959 return V == (V & ((1LL << 8) - 1));
4963 /// isLegalAddressImmediate - Return true if the integer value can be used
4964 /// as the offset of the target addressing mode for load / store of the
4966 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4967 const ARMSubtarget *Subtarget) {
4974 if (Subtarget->isThumb1Only())
4975 return isLegalT1AddressImmediate(V, VT);
4976 else if (Subtarget->isThumb2())
4977 return isLegalT2AddressImmediate(V, VT, Subtarget);
4982 switch (VT.getSimpleVT().SimpleTy) {
4983 default: return false;
4988 return V == (V & ((1LL << 12) - 1));
4991 return V == (V & ((1LL << 8) - 1));
4994 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4999 return V == (V & ((1LL << 8) - 1));
5003 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5005 int Scale = AM.Scale;
5009 switch (VT.getSimpleVT().SimpleTy) {
5010 default: return false;
5019 return Scale == 2 || Scale == 4 || Scale == 8;
5022 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5026 // Note, we allow "void" uses (basically, uses that aren't loads or
5027 // stores), because arm allows folding a scale into many arithmetic
5028 // operations. This should be made more precise and revisited later.
5030 // Allow r << imm, but the imm has to be a multiple of two.
5031 if (Scale & 1) return false;
5032 return isPowerOf2_32(Scale);
5036 /// isLegalAddressingMode - Return true if the addressing mode represented
5037 /// by AM is legal for this target, for a load/store of the specified type.
5038 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5039 const Type *Ty) const {
5040 EVT VT = getValueType(Ty, true);
5041 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5044 // Can never fold addr of global into load/store.
5049 case 0: // no scale reg, must be "r+i" or "r", or "i".
5052 if (Subtarget->isThumb1Only())
5056 // ARM doesn't support any R+R*scale+imm addr modes.
5063 if (Subtarget->isThumb2())
5064 return isLegalT2ScaledAddressingMode(AM, VT);
5066 int Scale = AM.Scale;
5067 switch (VT.getSimpleVT().SimpleTy) {
5068 default: return false;
5072 if (Scale < 0) Scale = -Scale;
5076 return isPowerOf2_32(Scale & ~1);
5080 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5085 // Note, we allow "void" uses (basically, uses that aren't loads or
5086 // stores), because arm allows folding a scale into many arithmetic
5087 // operations. This should be made more precise and revisited later.
5089 // Allow r << imm, but the imm has to be a multiple of two.
5090 if (Scale & 1) return false;
5091 return isPowerOf2_32(Scale);
5098 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5099 /// icmp immediate, that is the target has icmp instructions which can compare
5100 /// a register against the immediate without having to materialize the
5101 /// immediate into a register.
5102 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5103 if (!Subtarget->isThumb())
5104 return ARM_AM::getSOImmVal(Imm) != -1;
5105 if (Subtarget->isThumb2())
5106 return ARM_AM::getT2SOImmVal(Imm) != -1;
5107 return Imm >= 0 && Imm <= 255;
5110 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5111 bool isSEXTLoad, SDValue &Base,
5112 SDValue &Offset, bool &isInc,
5113 SelectionDAG &DAG) {
5114 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5117 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5119 Base = Ptr->getOperand(0);
5120 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5121 int RHSC = (int)RHS->getZExtValue();
5122 if (RHSC < 0 && RHSC > -256) {
5123 assert(Ptr->getOpcode() == ISD::ADD);
5125 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5129 isInc = (Ptr->getOpcode() == ISD::ADD);
5130 Offset = Ptr->getOperand(1);
5132 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5134 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5135 int RHSC = (int)RHS->getZExtValue();
5136 if (RHSC < 0 && RHSC > -0x1000) {
5137 assert(Ptr->getOpcode() == ISD::ADD);
5139 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5140 Base = Ptr->getOperand(0);
5145 if (Ptr->getOpcode() == ISD::ADD) {
5147 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5148 if (ShOpcVal != ARM_AM::no_shift) {
5149 Base = Ptr->getOperand(1);
5150 Offset = Ptr->getOperand(0);
5152 Base = Ptr->getOperand(0);
5153 Offset = Ptr->getOperand(1);
5158 isInc = (Ptr->getOpcode() == ISD::ADD);
5159 Base = Ptr->getOperand(0);
5160 Offset = Ptr->getOperand(1);
5164 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5168 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5169 bool isSEXTLoad, SDValue &Base,
5170 SDValue &Offset, bool &isInc,
5171 SelectionDAG &DAG) {
5172 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5175 Base = Ptr->getOperand(0);
5176 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5177 int RHSC = (int)RHS->getZExtValue();
5178 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5179 assert(Ptr->getOpcode() == ISD::ADD);
5181 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5183 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5184 isInc = Ptr->getOpcode() == ISD::ADD;
5185 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5193 /// getPreIndexedAddressParts - returns true by value, base pointer and
5194 /// offset pointer and addressing mode by reference if the node's address
5195 /// can be legally represented as pre-indexed load / store address.
5197 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5199 ISD::MemIndexedMode &AM,
5200 SelectionDAG &DAG) const {
5201 if (Subtarget->isThumb1Only())
5206 bool isSEXTLoad = false;
5207 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5208 Ptr = LD->getBasePtr();
5209 VT = LD->getMemoryVT();
5210 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5211 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5212 Ptr = ST->getBasePtr();
5213 VT = ST->getMemoryVT();
5218 bool isLegal = false;
5219 if (Subtarget->isThumb2())
5220 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5221 Offset, isInc, DAG);
5223 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5224 Offset, isInc, DAG);
5228 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5232 /// getPostIndexedAddressParts - returns true by value, base pointer and
5233 /// offset pointer and addressing mode by reference if this node can be
5234 /// combined with a load / store to form a post-indexed load / store.
5235 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5238 ISD::MemIndexedMode &AM,
5239 SelectionDAG &DAG) const {
5240 if (Subtarget->isThumb1Only())
5245 bool isSEXTLoad = false;
5246 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5247 VT = LD->getMemoryVT();
5248 Ptr = LD->getBasePtr();
5249 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5250 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5251 VT = ST->getMemoryVT();
5252 Ptr = ST->getBasePtr();
5257 bool isLegal = false;
5258 if (Subtarget->isThumb2())
5259 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5262 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5268 // Swap base ptr and offset to catch more post-index load / store when
5269 // it's legal. In Thumb2 mode, offset must be an immediate.
5270 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5271 !Subtarget->isThumb2())
5272 std::swap(Base, Offset);
5274 // Post-indexed load / store update the base pointer.
5279 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5283 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5287 const SelectionDAG &DAG,
5288 unsigned Depth) const {
5289 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5290 switch (Op.getOpcode()) {
5292 case ARMISD::CMOV: {
5293 // Bits are known zero/one if known on the LHS and RHS.
5294 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5295 if (KnownZero == 0 && KnownOne == 0) return;
5297 APInt KnownZeroRHS, KnownOneRHS;
5298 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5299 KnownZeroRHS, KnownOneRHS, Depth+1);
5300 KnownZero &= KnownZeroRHS;
5301 KnownOne &= KnownOneRHS;
5307 //===----------------------------------------------------------------------===//
5308 // ARM Inline Assembly Support
5309 //===----------------------------------------------------------------------===//
5311 /// getConstraintType - Given a constraint letter, return the type of
5312 /// constraint it is for this target.
5313 ARMTargetLowering::ConstraintType
5314 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5315 if (Constraint.size() == 1) {
5316 switch (Constraint[0]) {
5318 case 'l': return C_RegisterClass;
5319 case 'w': return C_RegisterClass;
5322 return TargetLowering::getConstraintType(Constraint);
5325 std::pair<unsigned, const TargetRegisterClass*>
5326 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5328 if (Constraint.size() == 1) {
5329 // GCC ARM Constraint Letters
5330 switch (Constraint[0]) {
5332 if (Subtarget->isThumb())
5333 return std::make_pair(0U, ARM::tGPRRegisterClass);
5335 return std::make_pair(0U, ARM::GPRRegisterClass);
5337 return std::make_pair(0U, ARM::GPRRegisterClass);
5340 return std::make_pair(0U, ARM::SPRRegisterClass);
5341 if (VT.getSizeInBits() == 64)
5342 return std::make_pair(0U, ARM::DPRRegisterClass);
5343 if (VT.getSizeInBits() == 128)
5344 return std::make_pair(0U, ARM::QPRRegisterClass);
5348 if (StringRef("{cc}").equals_lower(Constraint))
5349 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5351 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5354 std::vector<unsigned> ARMTargetLowering::
5355 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5357 if (Constraint.size() != 1)
5358 return std::vector<unsigned>();
5360 switch (Constraint[0]) { // GCC ARM Constraint Letters
5363 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5364 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5367 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5368 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5369 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5370 ARM::R12, ARM::LR, 0);
5373 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5374 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5375 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5376 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5377 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5378 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5379 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5380 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5381 if (VT.getSizeInBits() == 64)
5382 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5383 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5384 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5385 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5386 if (VT.getSizeInBits() == 128)
5387 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5388 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5392 return std::vector<unsigned>();
5395 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5396 /// vector. If it is invalid, don't add anything to Ops.
5397 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5399 std::vector<SDValue>&Ops,
5400 SelectionDAG &DAG) const {
5401 SDValue Result(0, 0);
5403 switch (Constraint) {
5405 case 'I': case 'J': case 'K': case 'L':
5406 case 'M': case 'N': case 'O':
5407 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5411 int64_t CVal64 = C->getSExtValue();
5412 int CVal = (int) CVal64;
5413 // None of these constraints allow values larger than 32 bits. Check
5414 // that the value fits in an int.
5418 switch (Constraint) {
5420 if (Subtarget->isThumb1Only()) {
5421 // This must be a constant between 0 and 255, for ADD
5423 if (CVal >= 0 && CVal <= 255)
5425 } else if (Subtarget->isThumb2()) {
5426 // A constant that can be used as an immediate value in a
5427 // data-processing instruction.
5428 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5431 // A constant that can be used as an immediate value in a
5432 // data-processing instruction.
5433 if (ARM_AM::getSOImmVal(CVal) != -1)
5439 if (Subtarget->isThumb()) { // FIXME thumb2
5440 // This must be a constant between -255 and -1, for negated ADD
5441 // immediates. This can be used in GCC with an "n" modifier that
5442 // prints the negated value, for use with SUB instructions. It is
5443 // not useful otherwise but is implemented for compatibility.
5444 if (CVal >= -255 && CVal <= -1)
5447 // This must be a constant between -4095 and 4095. It is not clear
5448 // what this constraint is intended for. Implemented for
5449 // compatibility with GCC.
5450 if (CVal >= -4095 && CVal <= 4095)
5456 if (Subtarget->isThumb1Only()) {
5457 // A 32-bit value where only one byte has a nonzero value. Exclude
5458 // zero to match GCC. This constraint is used by GCC internally for
5459 // constants that can be loaded with a move/shift combination.
5460 // It is not useful otherwise but is implemented for compatibility.
5461 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5463 } else if (Subtarget->isThumb2()) {
5464 // A constant whose bitwise inverse can be used as an immediate
5465 // value in a data-processing instruction. This can be used in GCC
5466 // with a "B" modifier that prints the inverted value, for use with
5467 // BIC and MVN instructions. It is not useful otherwise but is
5468 // implemented for compatibility.
5469 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5472 // A constant whose bitwise inverse can be used as an immediate
5473 // value in a data-processing instruction. This can be used in GCC
5474 // with a "B" modifier that prints the inverted value, for use with
5475 // BIC and MVN instructions. It is not useful otherwise but is
5476 // implemented for compatibility.
5477 if (ARM_AM::getSOImmVal(~CVal) != -1)
5483 if (Subtarget->isThumb1Only()) {
5484 // This must be a constant between -7 and 7,
5485 // for 3-operand ADD/SUB immediate instructions.
5486 if (CVal >= -7 && CVal < 7)
5488 } else if (Subtarget->isThumb2()) {
5489 // A constant whose negation can be used as an immediate value in a
5490 // data-processing instruction. This can be used in GCC with an "n"
5491 // modifier that prints the negated value, for use with SUB
5492 // instructions. It is not useful otherwise but is implemented for
5494 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5497 // A constant whose negation can be used as an immediate value in a
5498 // data-processing instruction. This can be used in GCC with an "n"
5499 // modifier that prints the negated value, for use with SUB
5500 // instructions. It is not useful otherwise but is implemented for
5502 if (ARM_AM::getSOImmVal(-CVal) != -1)
5508 if (Subtarget->isThumb()) { // FIXME thumb2
5509 // This must be a multiple of 4 between 0 and 1020, for
5510 // ADD sp + immediate.
5511 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5514 // A power of two or a constant between 0 and 32. This is used in
5515 // GCC for the shift amount on shifted register operands, but it is
5516 // useful in general for any shift amounts.
5517 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5523 if (Subtarget->isThumb()) { // FIXME thumb2
5524 // This must be a constant between 0 and 31, for shift amounts.
5525 if (CVal >= 0 && CVal <= 31)
5531 if (Subtarget->isThumb()) { // FIXME thumb2
5532 // This must be a multiple of 4 between -508 and 508, for
5533 // ADD/SUB sp = sp + immediate.
5534 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5539 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5543 if (Result.getNode()) {
5544 Ops.push_back(Result);
5547 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5551 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5552 // The ARM target isn't yet aware of offsets.
5556 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5557 APInt Imm = FPImm.bitcastToAPInt();
5558 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5559 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5560 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5562 // We can handle 4 bits of mantissa.
5563 // mantissa = (16+UInt(e:f:g:h))/16.
5564 if (Mantissa & 0x7ffff)
5567 if ((Mantissa & 0xf) != Mantissa)
5570 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5571 if (Exp < -3 || Exp > 4)
5573 Exp = ((Exp+3) & 0x7) ^ 4;
5575 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5578 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5579 APInt Imm = FPImm.bitcastToAPInt();
5580 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5581 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5582 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5584 // We can handle 4 bits of mantissa.
5585 // mantissa = (16+UInt(e:f:g:h))/16.
5586 if (Mantissa & 0xffffffffffffLL)
5589 if ((Mantissa & 0xf) != Mantissa)
5592 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5593 if (Exp < -3 || Exp > 4)
5595 Exp = ((Exp+3) & 0x7) ^ 4;
5597 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5600 bool ARM::isBitFieldInvertedMask(unsigned v) {
5601 if (v == 0xffffffff)
5603 // there can be 1's on either or both "outsides", all the "inside"
5605 unsigned int lsb = 0, msb = 31;
5606 while (v & (1 << msb)) --msb;
5607 while (v & (1 << lsb)) ++lsb;
5608 for (unsigned int i = lsb; i <= msb; ++i) {
5615 /// isFPImmLegal - Returns true if the target can instruction select the
5616 /// specified FP immediate natively. If false, the legalizer will
5617 /// materialize the FP immediate as a load from a constant pool.
5618 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5619 if (!Subtarget->hasVFP3())
5622 return ARM::getVFPf32Imm(Imm) != -1;
5624 return ARM::getVFPf64Imm(Imm) != -1;