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 Machine LICM is smart enough to hoist a
57 EnableARMVDUPsplat("arm-vdup-splat", cl::Hidden,
58 cl::desc("Generate VDUP for integer constant splats (TEMPORARY OPTION)."),
62 EnableARMLongCalls("arm-long-calls", cl::Hidden,
63 cl::desc("Generate calls via indirect call instructions"),
67 ARMInterworking("arm-interworking", cl::Hidden,
68 cl::desc("Enable / disable ARM interworking (for debugging only)"),
72 EnableARMCodePlacement("arm-code-placement", cl::Hidden,
73 cl::desc("Enable code placement pass for ARM"),
76 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
77 CCValAssign::LocInfo &LocInfo,
78 ISD::ArgFlagsTy &ArgFlags,
80 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
81 CCValAssign::LocInfo &LocInfo,
82 ISD::ArgFlagsTy &ArgFlags,
84 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
85 CCValAssign::LocInfo &LocInfo,
86 ISD::ArgFlagsTy &ArgFlags,
88 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
89 CCValAssign::LocInfo &LocInfo,
90 ISD::ArgFlagsTy &ArgFlags,
93 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
94 EVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
97 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
98 PromotedLdStVT.getSimpleVT());
100 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
101 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
102 PromotedLdStVT.getSimpleVT());
105 EVT ElemTy = VT.getVectorElementType();
106 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
107 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
108 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
109 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
110 if (ElemTy != MVT::i32) {
111 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
112 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
117 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
119 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
121 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
122 if (VT.isInteger()) {
123 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
124 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
125 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
128 // Promote all bit-wise operations.
129 if (VT.isInteger() && VT != PromotedBitwiseVT) {
130 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
131 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
132 PromotedBitwiseVT.getSimpleVT());
133 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
134 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
135 PromotedBitwiseVT.getSimpleVT());
136 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
137 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
138 PromotedBitwiseVT.getSimpleVT());
141 // Neon does not support vector divide/remainder operations.
142 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
143 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
144 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
145 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
146 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
147 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
150 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
151 addRegisterClass(VT, ARM::DPRRegisterClass);
152 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
155 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
156 addRegisterClass(VT, ARM::QPRRegisterClass);
157 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
160 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
161 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
162 return new TargetLoweringObjectFileMachO();
164 return new ARMElfTargetObjectFile();
167 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
168 : TargetLowering(TM, createTLOF(TM)) {
169 Subtarget = &TM.getSubtarget<ARMSubtarget>();
170 RegInfo = TM.getRegisterInfo();
172 if (Subtarget->isTargetDarwin()) {
173 // Uses VFP for Thumb libfuncs if available.
174 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
175 // Single-precision floating-point arithmetic.
176 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
177 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
178 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
179 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
181 // Double-precision floating-point arithmetic.
182 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
183 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
184 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
185 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
187 // Single-precision comparisons.
188 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
189 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
190 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
191 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
192 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
193 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
194 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
195 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
197 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
206 // Double-precision comparisons.
207 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
208 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
209 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
210 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
211 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
212 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
213 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
214 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
216 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
225 // Floating-point to integer conversions.
226 // i64 conversions are done via library routines even when generating VFP
227 // instructions, so use the same ones.
228 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
229 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
230 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
231 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
233 // Conversions between floating types.
234 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
235 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
237 // Integer to floating-point conversions.
238 // i64 conversions are done via library routines even when generating VFP
239 // instructions, so use the same ones.
240 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
241 // e.g., __floatunsidf vs. __floatunssidfvfp.
242 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
243 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
244 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
245 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
249 // These libcalls are not available in 32-bit.
250 setLibcallName(RTLIB::SHL_I128, 0);
251 setLibcallName(RTLIB::SRL_I128, 0);
252 setLibcallName(RTLIB::SRA_I128, 0);
254 // Libcalls should use the AAPCS base standard ABI, even if hard float
255 // is in effect, as per the ARM RTABI specification, section 4.1.2.
256 if (Subtarget->isAAPCS_ABI()) {
257 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
258 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
259 CallingConv::ARM_AAPCS);
263 if (Subtarget->isThumb1Only())
264 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
266 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
267 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
268 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
269 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
271 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
274 if (Subtarget->hasNEON()) {
275 addDRTypeForNEON(MVT::v2f32);
276 addDRTypeForNEON(MVT::v8i8);
277 addDRTypeForNEON(MVT::v4i16);
278 addDRTypeForNEON(MVT::v2i32);
279 addDRTypeForNEON(MVT::v1i64);
281 addQRTypeForNEON(MVT::v4f32);
282 addQRTypeForNEON(MVT::v2f64);
283 addQRTypeForNEON(MVT::v16i8);
284 addQRTypeForNEON(MVT::v8i16);
285 addQRTypeForNEON(MVT::v4i32);
286 addQRTypeForNEON(MVT::v2i64);
288 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
289 // neither Neon nor VFP support any arithmetic operations on it.
290 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
291 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
292 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
293 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
294 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
295 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
296 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
297 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
298 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
299 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
300 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
301 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
302 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
303 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
304 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
305 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
306 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
307 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
308 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
309 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
310 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
311 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
312 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
313 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
315 // Neon does not support some operations on v1i64 and v2i64 types.
316 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
317 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
318 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
319 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
321 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
322 setTargetDAGCombine(ISD::SHL);
323 setTargetDAGCombine(ISD::SRL);
324 setTargetDAGCombine(ISD::SRA);
325 setTargetDAGCombine(ISD::SIGN_EXTEND);
326 setTargetDAGCombine(ISD::ZERO_EXTEND);
327 setTargetDAGCombine(ISD::ANY_EXTEND);
328 setTargetDAGCombine(ISD::SELECT_CC);
331 computeRegisterProperties();
333 // ARM does not have f32 extending load.
334 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
336 // ARM does not have i1 sign extending load.
337 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
339 // ARM supports all 4 flavors of integer indexed load / store.
340 if (!Subtarget->isThumb1Only()) {
341 for (unsigned im = (unsigned)ISD::PRE_INC;
342 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
343 setIndexedLoadAction(im, MVT::i1, Legal);
344 setIndexedLoadAction(im, MVT::i8, Legal);
345 setIndexedLoadAction(im, MVT::i16, Legal);
346 setIndexedLoadAction(im, MVT::i32, Legal);
347 setIndexedStoreAction(im, MVT::i1, Legal);
348 setIndexedStoreAction(im, MVT::i8, Legal);
349 setIndexedStoreAction(im, MVT::i16, Legal);
350 setIndexedStoreAction(im, MVT::i32, Legal);
354 // i64 operation support.
355 if (Subtarget->isThumb1Only()) {
356 setOperationAction(ISD::MUL, MVT::i64, Expand);
357 setOperationAction(ISD::MULHU, MVT::i32, Expand);
358 setOperationAction(ISD::MULHS, MVT::i32, Expand);
359 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
360 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
362 setOperationAction(ISD::MUL, MVT::i64, Expand);
363 setOperationAction(ISD::MULHU, MVT::i32, Expand);
364 if (!Subtarget->hasV6Ops())
365 setOperationAction(ISD::MULHS, MVT::i32, Expand);
367 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
368 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
369 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
370 setOperationAction(ISD::SRL, MVT::i64, Custom);
371 setOperationAction(ISD::SRA, MVT::i64, Custom);
373 // ARM does not have ROTL.
374 setOperationAction(ISD::ROTL, MVT::i32, Expand);
375 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
376 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
377 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
378 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
380 // Only ARMv6 has BSWAP.
381 if (!Subtarget->hasV6Ops())
382 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
384 // These are expanded into libcalls.
385 if (!Subtarget->hasDivide()) {
386 // v7M has a hardware divider
387 setOperationAction(ISD::SDIV, MVT::i32, Expand);
388 setOperationAction(ISD::UDIV, MVT::i32, Expand);
390 setOperationAction(ISD::SREM, MVT::i32, Expand);
391 setOperationAction(ISD::UREM, MVT::i32, Expand);
392 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
393 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
395 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
396 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
397 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
398 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
399 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
401 setOperationAction(ISD::TRAP, MVT::Other, Legal);
403 // Use the default implementation.
404 setOperationAction(ISD::VASTART, MVT::Other, Custom);
405 setOperationAction(ISD::VAARG, MVT::Other, Expand);
406 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
407 setOperationAction(ISD::VAEND, MVT::Other, Expand);
408 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
409 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
410 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
411 // FIXME: Shouldn't need this, since no register is used, but the legalizer
412 // doesn't yet know how to not do that for SjLj.
413 setExceptionSelectorRegister(ARM::R0);
414 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
415 // Handle atomics directly for ARMv[67] (except for Thumb1), otherwise
416 // use the default expansion.
417 bool canHandleAtomics =
418 (Subtarget->hasV7Ops() ||
419 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only()));
420 if (canHandleAtomics) {
421 // membarrier needs custom lowering; the rest are legal and handled
423 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
425 // Set them all for expansion, which will force libcalls.
426 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
427 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
428 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
429 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
430 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
431 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
432 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
449 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
450 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
451 // Since the libcalls include locking, fold in the fences
452 setShouldFoldAtomicFences(true);
454 // 64-bit versions are always libcalls (for now)
455 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
456 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
457 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
458 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
459 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
460 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
461 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
462 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
464 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
465 if (!Subtarget->hasV6Ops()) {
466 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
467 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
469 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
471 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
472 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
473 // iff target supports vfp2.
474 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
475 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
478 // We want to custom lower some of our intrinsics.
479 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
480 if (Subtarget->isTargetDarwin()) {
481 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
482 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
485 setOperationAction(ISD::SETCC, MVT::i32, Expand);
486 setOperationAction(ISD::SETCC, MVT::f32, Expand);
487 setOperationAction(ISD::SETCC, MVT::f64, Expand);
488 setOperationAction(ISD::SELECT, MVT::i32, Expand);
489 setOperationAction(ISD::SELECT, MVT::f32, Expand);
490 setOperationAction(ISD::SELECT, MVT::f64, Expand);
491 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
492 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
493 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
495 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
496 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
497 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
498 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
499 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
501 // We don't support sin/cos/fmod/copysign/pow
502 setOperationAction(ISD::FSIN, MVT::f64, Expand);
503 setOperationAction(ISD::FSIN, MVT::f32, Expand);
504 setOperationAction(ISD::FCOS, MVT::f32, Expand);
505 setOperationAction(ISD::FCOS, MVT::f64, Expand);
506 setOperationAction(ISD::FREM, MVT::f64, Expand);
507 setOperationAction(ISD::FREM, MVT::f32, Expand);
508 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
509 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
510 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
512 setOperationAction(ISD::FPOW, MVT::f64, Expand);
513 setOperationAction(ISD::FPOW, MVT::f32, Expand);
515 // Various VFP goodness
516 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
517 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
518 if (Subtarget->hasVFP2()) {
519 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
520 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
521 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
522 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
524 // Special handling for half-precision FP.
525 if (!Subtarget->hasFP16()) {
526 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
527 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
531 // We have target-specific dag combine patterns for the following nodes:
532 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
533 setTargetDAGCombine(ISD::ADD);
534 setTargetDAGCombine(ISD::SUB);
535 setTargetDAGCombine(ISD::MUL);
537 if (Subtarget->hasV6T2Ops())
538 setTargetDAGCombine(ISD::OR);
540 setStackPointerRegisterToSaveRestore(ARM::SP);
542 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
543 setSchedulingPreference(Sched::RegPressure);
545 setSchedulingPreference(Sched::Hybrid);
547 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
549 // On ARM arguments smaller than 4 bytes are extended, so all arguments
550 // are at least 4 bytes aligned.
551 setMinStackArgumentAlignment(4);
553 if (EnableARMCodePlacement)
554 benefitFromCodePlacementOpt = true;
557 std::pair<const TargetRegisterClass*, uint8_t>
558 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
559 const TargetRegisterClass *RRC = 0;
561 switch (VT.getSimpleVT().SimpleTy) {
563 return TargetLowering::findRepresentativeClass(VT);
564 // Use DPR as representative register class for all floating point
565 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
566 // the cost is 1 for both f32 and f64.
567 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
568 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
569 RRC = ARM::DPRRegisterClass;
571 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
572 case MVT::v4f32: case MVT::v2f64:
573 RRC = ARM::DPRRegisterClass;
577 RRC = ARM::DPRRegisterClass;
581 RRC = ARM::DPRRegisterClass;
585 return std::make_pair(RRC, Cost);
588 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
591 case ARMISD::Wrapper: return "ARMISD::Wrapper";
592 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
593 case ARMISD::CALL: return "ARMISD::CALL";
594 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
595 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
596 case ARMISD::tCALL: return "ARMISD::tCALL";
597 case ARMISD::BRCOND: return "ARMISD::BRCOND";
598 case ARMISD::BR_JT: return "ARMISD::BR_JT";
599 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
600 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
601 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
602 case ARMISD::CMP: return "ARMISD::CMP";
603 case ARMISD::CMPZ: return "ARMISD::CMPZ";
604 case ARMISD::CMPFP: return "ARMISD::CMPFP";
605 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
606 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
607 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
608 case ARMISD::CMOV: return "ARMISD::CMOV";
609 case ARMISD::CNEG: return "ARMISD::CNEG";
611 case ARMISD::RBIT: return "ARMISD::RBIT";
613 case ARMISD::FTOSI: return "ARMISD::FTOSI";
614 case ARMISD::FTOUI: return "ARMISD::FTOUI";
615 case ARMISD::SITOF: return "ARMISD::SITOF";
616 case ARMISD::UITOF: return "ARMISD::UITOF";
618 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
619 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
620 case ARMISD::RRX: return "ARMISD::RRX";
622 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
623 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
625 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
626 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
628 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
630 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
632 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
634 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
635 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
637 case ARMISD::VCEQ: return "ARMISD::VCEQ";
638 case ARMISD::VCGE: return "ARMISD::VCGE";
639 case ARMISD::VCGEU: return "ARMISD::VCGEU";
640 case ARMISD::VCGT: return "ARMISD::VCGT";
641 case ARMISD::VCGTU: return "ARMISD::VCGTU";
642 case ARMISD::VTST: return "ARMISD::VTST";
644 case ARMISD::VSHL: return "ARMISD::VSHL";
645 case ARMISD::VSHRs: return "ARMISD::VSHRs";
646 case ARMISD::VSHRu: return "ARMISD::VSHRu";
647 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
648 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
649 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
650 case ARMISD::VSHRN: return "ARMISD::VSHRN";
651 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
652 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
653 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
654 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
655 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
656 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
657 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
658 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
659 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
660 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
661 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
662 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
663 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
664 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
665 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
666 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
667 case ARMISD::VDUP: return "ARMISD::VDUP";
668 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
669 case ARMISD::VEXT: return "ARMISD::VEXT";
670 case ARMISD::VREV64: return "ARMISD::VREV64";
671 case ARMISD::VREV32: return "ARMISD::VREV32";
672 case ARMISD::VREV16: return "ARMISD::VREV16";
673 case ARMISD::VZIP: return "ARMISD::VZIP";
674 case ARMISD::VUZP: return "ARMISD::VUZP";
675 case ARMISD::VTRN: return "ARMISD::VTRN";
676 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
677 case ARMISD::FMAX: return "ARMISD::FMAX";
678 case ARMISD::FMIN: return "ARMISD::FMIN";
679 case ARMISD::BFI: return "ARMISD::BFI";
683 /// getRegClassFor - Return the register class that should be used for the
684 /// specified value type.
685 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
686 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
687 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
688 // load / store 4 to 8 consecutive D registers.
689 if (Subtarget->hasNEON()) {
690 if (VT == MVT::v4i64)
691 return ARM::QQPRRegisterClass;
692 else if (VT == MVT::v8i64)
693 return ARM::QQQQPRRegisterClass;
695 return TargetLowering::getRegClassFor(VT);
698 // Create a fast isel object.
700 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
701 return ARM::createFastISel(funcInfo);
704 /// getFunctionAlignment - Return the Log2 alignment of this function.
705 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
706 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
709 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
710 /// be used for loads / stores from the global.
711 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
712 return (Subtarget->isThumb1Only() ? 127 : 4095);
715 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
716 unsigned NumVals = N->getNumValues();
718 return Sched::RegPressure;
720 for (unsigned i = 0; i != NumVals; ++i) {
721 EVT VT = N->getValueType(i);
722 if (VT.isFloatingPoint() || VT.isVector())
723 return Sched::Latency;
726 if (!N->isMachineOpcode())
727 return Sched::RegPressure;
729 // Load are scheduled for latency even if there instruction itinerary
731 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
732 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
734 return Sched::Latency;
736 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
737 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
738 return Sched::Latency;
739 return Sched::RegPressure;
743 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
744 MachineFunction &MF) const {
745 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
746 switch (RC->getID()) {
749 case ARM::tGPRRegClassID:
751 case ARM::GPRRegClassID:
752 return 10 - FPDiff - (Subtarget->isR9Reserved() ? 1 : 0);
753 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
754 case ARM::DPRRegClassID:
759 //===----------------------------------------------------------------------===//
761 //===----------------------------------------------------------------------===//
763 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
764 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
766 default: llvm_unreachable("Unknown condition code!");
767 case ISD::SETNE: return ARMCC::NE;
768 case ISD::SETEQ: return ARMCC::EQ;
769 case ISD::SETGT: return ARMCC::GT;
770 case ISD::SETGE: return ARMCC::GE;
771 case ISD::SETLT: return ARMCC::LT;
772 case ISD::SETLE: return ARMCC::LE;
773 case ISD::SETUGT: return ARMCC::HI;
774 case ISD::SETUGE: return ARMCC::HS;
775 case ISD::SETULT: return ARMCC::LO;
776 case ISD::SETULE: return ARMCC::LS;
780 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
781 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
782 ARMCC::CondCodes &CondCode2) {
783 CondCode2 = ARMCC::AL;
785 default: llvm_unreachable("Unknown FP condition!");
787 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
789 case ISD::SETOGT: CondCode = ARMCC::GT; break;
791 case ISD::SETOGE: CondCode = ARMCC::GE; break;
792 case ISD::SETOLT: CondCode = ARMCC::MI; break;
793 case ISD::SETOLE: CondCode = ARMCC::LS; break;
794 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
795 case ISD::SETO: CondCode = ARMCC::VC; break;
796 case ISD::SETUO: CondCode = ARMCC::VS; break;
797 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
798 case ISD::SETUGT: CondCode = ARMCC::HI; break;
799 case ISD::SETUGE: CondCode = ARMCC::PL; break;
801 case ISD::SETULT: CondCode = ARMCC::LT; break;
803 case ISD::SETULE: CondCode = ARMCC::LE; break;
805 case ISD::SETUNE: CondCode = ARMCC::NE; break;
809 //===----------------------------------------------------------------------===//
810 // Calling Convention Implementation
811 //===----------------------------------------------------------------------===//
813 #include "ARMGenCallingConv.inc"
815 // APCS f64 is in register pairs, possibly split to stack
816 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
817 CCValAssign::LocInfo &LocInfo,
818 CCState &State, bool CanFail) {
819 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
821 // Try to get the first register.
822 if (unsigned Reg = State.AllocateReg(RegList, 4))
823 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
825 // For the 2nd half of a v2f64, do not fail.
829 // Put the whole thing on the stack.
830 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
831 State.AllocateStack(8, 4),
836 // Try to get the second register.
837 if (unsigned Reg = State.AllocateReg(RegList, 4))
838 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
840 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
841 State.AllocateStack(4, 4),
846 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
847 CCValAssign::LocInfo &LocInfo,
848 ISD::ArgFlagsTy &ArgFlags,
850 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
852 if (LocVT == MVT::v2f64 &&
853 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
855 return true; // we handled it
858 // AAPCS f64 is in aligned register pairs
859 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
860 CCValAssign::LocInfo &LocInfo,
861 CCState &State, bool CanFail) {
862 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
863 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
864 static const unsigned ShadowRegList[] = { ARM::R0, ARM::R1 };
866 unsigned Reg = State.AllocateReg(HiRegList, ShadowRegList, 2);
868 // For the 2nd half of a v2f64, do not just fail.
872 // Put the whole thing on the stack.
873 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
874 State.AllocateStack(8, 8),
880 for (i = 0; i < 2; ++i)
881 if (HiRegList[i] == Reg)
884 unsigned T = State.AllocateReg(LoRegList[i]);
886 assert(T == LoRegList[i] && "Could not allocate register");
888 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
889 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
894 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
895 CCValAssign::LocInfo &LocInfo,
896 ISD::ArgFlagsTy &ArgFlags,
898 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
900 if (LocVT == MVT::v2f64 &&
901 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
903 return true; // we handled it
906 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
907 CCValAssign::LocInfo &LocInfo, CCState &State) {
908 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
909 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
911 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
913 return false; // we didn't handle it
916 for (i = 0; i < 2; ++i)
917 if (HiRegList[i] == Reg)
920 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
921 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
926 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
927 CCValAssign::LocInfo &LocInfo,
928 ISD::ArgFlagsTy &ArgFlags,
930 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
932 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
934 return true; // we handled it
937 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
938 CCValAssign::LocInfo &LocInfo,
939 ISD::ArgFlagsTy &ArgFlags,
941 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
945 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
946 /// given CallingConvention value.
947 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
949 bool isVarArg) const {
952 llvm_unreachable("Unsupported calling convention");
954 case CallingConv::Fast:
955 // Use target triple & subtarget features to do actual dispatch.
956 if (Subtarget->isAAPCS_ABI()) {
957 if (Subtarget->hasVFP2() &&
958 FloatABIType == FloatABI::Hard && !isVarArg)
959 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
961 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
963 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
964 case CallingConv::ARM_AAPCS_VFP:
965 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
966 case CallingConv::ARM_AAPCS:
967 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
968 case CallingConv::ARM_APCS:
969 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
973 /// LowerCallResult - Lower the result values of a call into the
974 /// appropriate copies out of appropriate physical registers.
976 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
977 CallingConv::ID CallConv, bool isVarArg,
978 const SmallVectorImpl<ISD::InputArg> &Ins,
979 DebugLoc dl, SelectionDAG &DAG,
980 SmallVectorImpl<SDValue> &InVals) const {
982 // Assign locations to each value returned by this call.
983 SmallVector<CCValAssign, 16> RVLocs;
984 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
985 RVLocs, *DAG.getContext());
986 CCInfo.AnalyzeCallResult(Ins,
987 CCAssignFnForNode(CallConv, /* Return*/ true,
990 // Copy all of the result registers out of their specified physreg.
991 for (unsigned i = 0; i != RVLocs.size(); ++i) {
992 CCValAssign VA = RVLocs[i];
995 if (VA.needsCustom()) {
996 // Handle f64 or half of a v2f64.
997 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
999 Chain = Lo.getValue(1);
1000 InFlag = Lo.getValue(2);
1001 VA = RVLocs[++i]; // skip ahead to next loc
1002 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1004 Chain = Hi.getValue(1);
1005 InFlag = Hi.getValue(2);
1006 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1008 if (VA.getLocVT() == MVT::v2f64) {
1009 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1010 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1011 DAG.getConstant(0, MVT::i32));
1013 VA = RVLocs[++i]; // skip ahead to next loc
1014 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1015 Chain = Lo.getValue(1);
1016 InFlag = Lo.getValue(2);
1017 VA = RVLocs[++i]; // skip ahead to next loc
1018 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1019 Chain = Hi.getValue(1);
1020 InFlag = Hi.getValue(2);
1021 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1022 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1023 DAG.getConstant(1, MVT::i32));
1026 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1028 Chain = Val.getValue(1);
1029 InFlag = Val.getValue(2);
1032 switch (VA.getLocInfo()) {
1033 default: llvm_unreachable("Unknown loc info!");
1034 case CCValAssign::Full: break;
1035 case CCValAssign::BCvt:
1036 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1040 InVals.push_back(Val);
1046 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1047 /// by "Src" to address "Dst" of size "Size". Alignment information is
1048 /// specified by the specific parameter attribute. The copy will be passed as
1049 /// a byval function parameter.
1050 /// Sometimes what we are copying is the end of a larger object, the part that
1051 /// does not fit in registers.
1053 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1054 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1056 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1057 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1058 /*isVolatile=*/false, /*AlwaysInline=*/false,
1062 /// LowerMemOpCallTo - Store the argument to the stack.
1064 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1065 SDValue StackPtr, SDValue Arg,
1066 DebugLoc dl, SelectionDAG &DAG,
1067 const CCValAssign &VA,
1068 ISD::ArgFlagsTy Flags) const {
1069 unsigned LocMemOffset = VA.getLocMemOffset();
1070 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1071 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1072 if (Flags.isByVal()) {
1073 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1075 return DAG.getStore(Chain, dl, Arg, PtrOff,
1076 PseudoSourceValue::getStack(), LocMemOffset,
1080 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1081 SDValue Chain, SDValue &Arg,
1082 RegsToPassVector &RegsToPass,
1083 CCValAssign &VA, CCValAssign &NextVA,
1085 SmallVector<SDValue, 8> &MemOpChains,
1086 ISD::ArgFlagsTy Flags) const {
1088 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1089 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1090 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1092 if (NextVA.isRegLoc())
1093 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1095 assert(NextVA.isMemLoc());
1096 if (StackPtr.getNode() == 0)
1097 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1099 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1105 /// LowerCall - Lowering a call into a callseq_start <-
1106 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1109 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1110 CallingConv::ID CallConv, bool isVarArg,
1112 const SmallVectorImpl<ISD::OutputArg> &Outs,
1113 const SmallVectorImpl<SDValue> &OutVals,
1114 const SmallVectorImpl<ISD::InputArg> &Ins,
1115 DebugLoc dl, SelectionDAG &DAG,
1116 SmallVectorImpl<SDValue> &InVals) const {
1117 MachineFunction &MF = DAG.getMachineFunction();
1118 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1119 bool IsSibCall = false;
1121 // Check if it's really possible to do a tail call.
1122 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1123 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1124 Outs, OutVals, Ins, DAG);
1125 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1126 // detected sibcalls.
1133 // Analyze operands of the call, assigning locations to each operand.
1134 SmallVector<CCValAssign, 16> ArgLocs;
1135 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1137 CCInfo.AnalyzeCallOperands(Outs,
1138 CCAssignFnForNode(CallConv, /* Return*/ false,
1141 // Get a count of how many bytes are to be pushed on the stack.
1142 unsigned NumBytes = CCInfo.getNextStackOffset();
1144 // For tail calls, memory operands are available in our caller's stack.
1148 // Adjust the stack pointer for the new arguments...
1149 // These operations are automatically eliminated by the prolog/epilog pass
1151 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1153 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1155 RegsToPassVector RegsToPass;
1156 SmallVector<SDValue, 8> MemOpChains;
1158 // Walk the register/memloc assignments, inserting copies/loads. In the case
1159 // of tail call optimization, arguments are handled later.
1160 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1162 ++i, ++realArgIdx) {
1163 CCValAssign &VA = ArgLocs[i];
1164 SDValue Arg = OutVals[realArgIdx];
1165 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1167 // Promote the value if needed.
1168 switch (VA.getLocInfo()) {
1169 default: llvm_unreachable("Unknown loc info!");
1170 case CCValAssign::Full: break;
1171 case CCValAssign::SExt:
1172 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1174 case CCValAssign::ZExt:
1175 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1177 case CCValAssign::AExt:
1178 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1180 case CCValAssign::BCvt:
1181 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1185 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1186 if (VA.needsCustom()) {
1187 if (VA.getLocVT() == MVT::v2f64) {
1188 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1189 DAG.getConstant(0, MVT::i32));
1190 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1191 DAG.getConstant(1, MVT::i32));
1193 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1194 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1196 VA = ArgLocs[++i]; // skip ahead to next loc
1197 if (VA.isRegLoc()) {
1198 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1199 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1201 assert(VA.isMemLoc());
1203 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1204 dl, DAG, VA, Flags));
1207 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1208 StackPtr, MemOpChains, Flags);
1210 } else if (VA.isRegLoc()) {
1211 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1212 } else if (!IsSibCall) {
1213 assert(VA.isMemLoc());
1215 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1216 dl, DAG, VA, Flags));
1220 if (!MemOpChains.empty())
1221 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1222 &MemOpChains[0], MemOpChains.size());
1224 // Build a sequence of copy-to-reg nodes chained together with token chain
1225 // and flag operands which copy the outgoing args into the appropriate regs.
1227 // Tail call byval lowering might overwrite argument registers so in case of
1228 // tail call optimization the copies to registers are lowered later.
1230 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1231 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1232 RegsToPass[i].second, InFlag);
1233 InFlag = Chain.getValue(1);
1236 // For tail calls lower the arguments to the 'real' stack slot.
1238 // Force all the incoming stack arguments to be loaded from the stack
1239 // before any new outgoing arguments are stored to the stack, because the
1240 // outgoing stack slots may alias the incoming argument stack slots, and
1241 // the alias isn't otherwise explicit. This is slightly more conservative
1242 // than necessary, because it means that each store effectively depends
1243 // on every argument instead of just those arguments it would clobber.
1245 // Do not flag preceeding copytoreg stuff together with the following stuff.
1247 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1248 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1249 RegsToPass[i].second, InFlag);
1250 InFlag = Chain.getValue(1);
1255 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1256 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1257 // node so that legalize doesn't hack it.
1258 bool isDirect = false;
1259 bool isARMFunc = false;
1260 bool isLocalARMFunc = false;
1261 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1263 if (EnableARMLongCalls) {
1264 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1265 && "long-calls with non-static relocation model!");
1266 // Handle a global address or an external symbol. If it's not one of
1267 // those, the target's already in a register, so we don't need to do
1269 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1270 const GlobalValue *GV = G->getGlobal();
1271 // Create a constant pool entry for the callee address
1272 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1273 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1276 // Get the address of the callee into a register
1277 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1278 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1279 Callee = DAG.getLoad(getPointerTy(), dl,
1280 DAG.getEntryNode(), CPAddr,
1281 PseudoSourceValue::getConstantPool(), 0,
1283 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1284 const char *Sym = S->getSymbol();
1286 // Create a constant pool entry for the callee address
1287 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1288 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1289 Sym, ARMPCLabelIndex, 0);
1290 // Get the address of the callee into a register
1291 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1292 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1293 Callee = DAG.getLoad(getPointerTy(), dl,
1294 DAG.getEntryNode(), CPAddr,
1295 PseudoSourceValue::getConstantPool(), 0,
1298 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1299 const GlobalValue *GV = G->getGlobal();
1301 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1302 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1303 getTargetMachine().getRelocationModel() != Reloc::Static;
1304 isARMFunc = !Subtarget->isThumb() || isStub;
1305 // ARM call to a local ARM function is predicable.
1306 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1307 // tBX takes a register source operand.
1308 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1309 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1310 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1313 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1314 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1315 Callee = DAG.getLoad(getPointerTy(), dl,
1316 DAG.getEntryNode(), CPAddr,
1317 PseudoSourceValue::getConstantPool(), 0,
1319 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1320 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1321 getPointerTy(), Callee, PICLabel);
1323 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1324 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1326 bool isStub = Subtarget->isTargetDarwin() &&
1327 getTargetMachine().getRelocationModel() != Reloc::Static;
1328 isARMFunc = !Subtarget->isThumb() || isStub;
1329 // tBX takes a register source operand.
1330 const char *Sym = S->getSymbol();
1331 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1332 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1333 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1334 Sym, ARMPCLabelIndex, 4);
1335 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1336 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1337 Callee = DAG.getLoad(getPointerTy(), dl,
1338 DAG.getEntryNode(), CPAddr,
1339 PseudoSourceValue::getConstantPool(), 0,
1341 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1342 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1343 getPointerTy(), Callee, PICLabel);
1345 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1348 // FIXME: handle tail calls differently.
1350 if (Subtarget->isThumb()) {
1351 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1352 CallOpc = ARMISD::CALL_NOLINK;
1354 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1356 CallOpc = (isDirect || Subtarget->hasV5TOps())
1357 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1358 : ARMISD::CALL_NOLINK;
1361 std::vector<SDValue> Ops;
1362 Ops.push_back(Chain);
1363 Ops.push_back(Callee);
1365 // Add argument registers to the end of the list so that they are known live
1367 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1368 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1369 RegsToPass[i].second.getValueType()));
1371 if (InFlag.getNode())
1372 Ops.push_back(InFlag);
1374 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1376 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1378 // Returns a chain and a flag for retval copy to use.
1379 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1380 InFlag = Chain.getValue(1);
1382 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1383 DAG.getIntPtrConstant(0, true), InFlag);
1385 InFlag = Chain.getValue(1);
1387 // Handle result values, copying them out of physregs into vregs that we
1389 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1393 /// MatchingStackOffset - Return true if the given stack call argument is
1394 /// already available in the same position (relatively) of the caller's
1395 /// incoming argument stack.
1397 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1398 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1399 const ARMInstrInfo *TII) {
1400 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1402 if (Arg.getOpcode() == ISD::CopyFromReg) {
1403 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1404 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1406 MachineInstr *Def = MRI->getVRegDef(VR);
1409 if (!Flags.isByVal()) {
1410 if (!TII->isLoadFromStackSlot(Def, FI))
1415 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1416 if (Flags.isByVal())
1417 // ByVal argument is passed in as a pointer but it's now being
1418 // dereferenced. e.g.
1419 // define @foo(%struct.X* %A) {
1420 // tail call @bar(%struct.X* byval %A)
1423 SDValue Ptr = Ld->getBasePtr();
1424 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1427 FI = FINode->getIndex();
1431 assert(FI != INT_MAX);
1432 if (!MFI->isFixedObjectIndex(FI))
1434 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1437 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1438 /// for tail call optimization. Targets which want to do tail call
1439 /// optimization should implement this function.
1441 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1442 CallingConv::ID CalleeCC,
1444 bool isCalleeStructRet,
1445 bool isCallerStructRet,
1446 const SmallVectorImpl<ISD::OutputArg> &Outs,
1447 const SmallVectorImpl<SDValue> &OutVals,
1448 const SmallVectorImpl<ISD::InputArg> &Ins,
1449 SelectionDAG& DAG) const {
1450 const Function *CallerF = DAG.getMachineFunction().getFunction();
1451 CallingConv::ID CallerCC = CallerF->getCallingConv();
1452 bool CCMatch = CallerCC == CalleeCC;
1454 // Look for obvious safe cases to perform tail call optimization that do not
1455 // require ABI changes. This is what gcc calls sibcall.
1457 // Do not sibcall optimize vararg calls unless the call site is not passing
1459 if (isVarArg && !Outs.empty())
1462 // Also avoid sibcall optimization if either caller or callee uses struct
1463 // return semantics.
1464 if (isCalleeStructRet || isCallerStructRet)
1467 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1468 // emitEpilogue is not ready for them.
1469 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1470 // LR. This means if we need to reload LR, it takes an extra instructions,
1471 // which outweighs the value of the tail call; but here we don't know yet
1472 // whether LR is going to be used. Probably the right approach is to
1473 // generate the tail call here and turn it back into CALL/RET in
1474 // emitEpilogue if LR is used.
1475 if (Subtarget->isThumb1Only())
1478 // For the moment, we can only do this to functions defined in this
1479 // compilation, or to indirect calls. A Thumb B to an ARM function,
1480 // or vice versa, is not easily fixed up in the linker unlike BL.
1481 // (We could do this by loading the address of the callee into a register;
1482 // that is an extra instruction over the direct call and burns a register
1483 // as well, so is not likely to be a win.)
1485 // It might be safe to remove this restriction on non-Darwin.
1487 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1488 // but we need to make sure there are enough registers; the only valid
1489 // registers are the 4 used for parameters. We don't currently do this
1491 if (isa<ExternalSymbolSDNode>(Callee))
1494 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1495 const GlobalValue *GV = G->getGlobal();
1496 if (GV->isDeclaration() || GV->isWeakForLinker())
1500 // If the calling conventions do not match, then we'd better make sure the
1501 // results are returned in the same way as what the caller expects.
1503 SmallVector<CCValAssign, 16> RVLocs1;
1504 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1505 RVLocs1, *DAG.getContext());
1506 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1508 SmallVector<CCValAssign, 16> RVLocs2;
1509 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1510 RVLocs2, *DAG.getContext());
1511 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1513 if (RVLocs1.size() != RVLocs2.size())
1515 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1516 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1518 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1520 if (RVLocs1[i].isRegLoc()) {
1521 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1524 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1530 // If the callee takes no arguments then go on to check the results of the
1532 if (!Outs.empty()) {
1533 // Check if stack adjustment is needed. For now, do not do this if any
1534 // argument is passed on the stack.
1535 SmallVector<CCValAssign, 16> ArgLocs;
1536 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1537 ArgLocs, *DAG.getContext());
1538 CCInfo.AnalyzeCallOperands(Outs,
1539 CCAssignFnForNode(CalleeCC, false, isVarArg));
1540 if (CCInfo.getNextStackOffset()) {
1541 MachineFunction &MF = DAG.getMachineFunction();
1543 // Check if the arguments are already laid out in the right way as
1544 // the caller's fixed stack objects.
1545 MachineFrameInfo *MFI = MF.getFrameInfo();
1546 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1547 const ARMInstrInfo *TII =
1548 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1549 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1551 ++i, ++realArgIdx) {
1552 CCValAssign &VA = ArgLocs[i];
1553 EVT RegVT = VA.getLocVT();
1554 SDValue Arg = OutVals[realArgIdx];
1555 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1556 if (VA.getLocInfo() == CCValAssign::Indirect)
1558 if (VA.needsCustom()) {
1559 // f64 and vector types are split into multiple registers or
1560 // register/stack-slot combinations. The types will not match
1561 // the registers; give up on memory f64 refs until we figure
1562 // out what to do about this.
1565 if (!ArgLocs[++i].isRegLoc())
1567 if (RegVT == MVT::v2f64) {
1568 if (!ArgLocs[++i].isRegLoc())
1570 if (!ArgLocs[++i].isRegLoc())
1573 } else if (!VA.isRegLoc()) {
1574 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1586 ARMTargetLowering::LowerReturn(SDValue Chain,
1587 CallingConv::ID CallConv, bool isVarArg,
1588 const SmallVectorImpl<ISD::OutputArg> &Outs,
1589 const SmallVectorImpl<SDValue> &OutVals,
1590 DebugLoc dl, SelectionDAG &DAG) const {
1592 // CCValAssign - represent the assignment of the return value to a location.
1593 SmallVector<CCValAssign, 16> RVLocs;
1595 // CCState - Info about the registers and stack slots.
1596 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1599 // Analyze outgoing return values.
1600 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1603 // If this is the first return lowered for this function, add
1604 // the regs to the liveout set for the function.
1605 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1606 for (unsigned i = 0; i != RVLocs.size(); ++i)
1607 if (RVLocs[i].isRegLoc())
1608 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1613 // Copy the result values into the output registers.
1614 for (unsigned i = 0, realRVLocIdx = 0;
1616 ++i, ++realRVLocIdx) {
1617 CCValAssign &VA = RVLocs[i];
1618 assert(VA.isRegLoc() && "Can only return in registers!");
1620 SDValue Arg = OutVals[realRVLocIdx];
1622 switch (VA.getLocInfo()) {
1623 default: llvm_unreachable("Unknown loc info!");
1624 case CCValAssign::Full: break;
1625 case CCValAssign::BCvt:
1626 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1630 if (VA.needsCustom()) {
1631 if (VA.getLocVT() == MVT::v2f64) {
1632 // Extract the first half and return it in two registers.
1633 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1634 DAG.getConstant(0, MVT::i32));
1635 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1636 DAG.getVTList(MVT::i32, MVT::i32), Half);
1638 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1639 Flag = Chain.getValue(1);
1640 VA = RVLocs[++i]; // skip ahead to next loc
1641 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1642 HalfGPRs.getValue(1), Flag);
1643 Flag = Chain.getValue(1);
1644 VA = RVLocs[++i]; // skip ahead to next loc
1646 // Extract the 2nd half and fall through to handle it as an f64 value.
1647 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1648 DAG.getConstant(1, MVT::i32));
1650 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1652 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1653 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1654 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1655 Flag = Chain.getValue(1);
1656 VA = RVLocs[++i]; // skip ahead to next loc
1657 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1660 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1662 // Guarantee that all emitted copies are
1663 // stuck together, avoiding something bad.
1664 Flag = Chain.getValue(1);
1669 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1671 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1676 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1677 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1678 // one of the above mentioned nodes. It has to be wrapped because otherwise
1679 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1680 // be used to form addressing mode. These wrapped nodes will be selected
1682 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1683 EVT PtrVT = Op.getValueType();
1684 // FIXME there is no actual debug info here
1685 DebugLoc dl = Op.getDebugLoc();
1686 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1688 if (CP->isMachineConstantPoolEntry())
1689 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1690 CP->getAlignment());
1692 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1693 CP->getAlignment());
1694 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1697 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1698 return MachineJumpTableInfo::EK_Inline;
1701 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1702 SelectionDAG &DAG) const {
1703 MachineFunction &MF = DAG.getMachineFunction();
1704 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1705 unsigned ARMPCLabelIndex = 0;
1706 DebugLoc DL = Op.getDebugLoc();
1707 EVT PtrVT = getPointerTy();
1708 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1709 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1711 if (RelocM == Reloc::Static) {
1712 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1714 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1715 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1716 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1717 ARMCP::CPBlockAddress,
1719 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1721 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1722 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1723 PseudoSourceValue::getConstantPool(), 0,
1725 if (RelocM == Reloc::Static)
1727 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1728 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1731 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1733 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1734 SelectionDAG &DAG) const {
1735 DebugLoc dl = GA->getDebugLoc();
1736 EVT PtrVT = getPointerTy();
1737 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1738 MachineFunction &MF = DAG.getMachineFunction();
1739 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1740 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1741 ARMConstantPoolValue *CPV =
1742 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1743 ARMCP::CPValue, PCAdj, "tlsgd", true);
1744 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1745 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1746 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1747 PseudoSourceValue::getConstantPool(), 0,
1749 SDValue Chain = Argument.getValue(1);
1751 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1752 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1754 // call __tls_get_addr.
1757 Entry.Node = Argument;
1758 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1759 Args.push_back(Entry);
1760 // FIXME: is there useful debug info available here?
1761 std::pair<SDValue, SDValue> CallResult =
1762 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1763 false, false, false, false,
1764 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1765 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1766 return CallResult.first;
1769 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1770 // "local exec" model.
1772 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1773 SelectionDAG &DAG) const {
1774 const GlobalValue *GV = GA->getGlobal();
1775 DebugLoc dl = GA->getDebugLoc();
1777 SDValue Chain = DAG.getEntryNode();
1778 EVT PtrVT = getPointerTy();
1779 // Get the Thread Pointer
1780 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1782 if (GV->isDeclaration()) {
1783 MachineFunction &MF = DAG.getMachineFunction();
1784 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1785 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1786 // Initial exec model.
1787 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1788 ARMConstantPoolValue *CPV =
1789 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1790 ARMCP::CPValue, PCAdj, "gottpoff", true);
1791 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1792 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1793 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1794 PseudoSourceValue::getConstantPool(), 0,
1796 Chain = Offset.getValue(1);
1798 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1799 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1801 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1802 PseudoSourceValue::getConstantPool(), 0,
1806 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1807 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1808 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1809 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1810 PseudoSourceValue::getConstantPool(), 0,
1814 // The address of the thread local variable is the add of the thread
1815 // pointer with the offset of the variable.
1816 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1820 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1821 // TODO: implement the "local dynamic" model
1822 assert(Subtarget->isTargetELF() &&
1823 "TLS not implemented for non-ELF targets");
1824 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1825 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1826 // otherwise use the "Local Exec" TLS Model
1827 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1828 return LowerToTLSGeneralDynamicModel(GA, DAG);
1830 return LowerToTLSExecModels(GA, DAG);
1833 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1834 SelectionDAG &DAG) const {
1835 EVT PtrVT = getPointerTy();
1836 DebugLoc dl = Op.getDebugLoc();
1837 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1838 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1839 if (RelocM == Reloc::PIC_) {
1840 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1841 ARMConstantPoolValue *CPV =
1842 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1843 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1844 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1845 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1847 PseudoSourceValue::getConstantPool(), 0,
1849 SDValue Chain = Result.getValue(1);
1850 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1851 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1853 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1854 PseudoSourceValue::getGOT(), 0,
1858 // If we have T2 ops, we can materialize the address directly via movt/movw
1859 // pair. This is always cheaper.
1860 if (Subtarget->useMovt()) {
1861 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1862 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1864 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1865 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1866 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1867 PseudoSourceValue::getConstantPool(), 0,
1873 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1874 SelectionDAG &DAG) const {
1875 MachineFunction &MF = DAG.getMachineFunction();
1876 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1877 unsigned ARMPCLabelIndex = 0;
1878 EVT PtrVT = getPointerTy();
1879 DebugLoc dl = Op.getDebugLoc();
1880 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1881 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1883 if (RelocM == Reloc::Static)
1884 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1886 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1887 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1888 ARMConstantPoolValue *CPV =
1889 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1890 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1892 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1894 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1895 PseudoSourceValue::getConstantPool(), 0,
1897 SDValue Chain = Result.getValue(1);
1899 if (RelocM == Reloc::PIC_) {
1900 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1901 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1904 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1905 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1906 PseudoSourceValue::getGOT(), 0,
1912 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1913 SelectionDAG &DAG) const {
1914 assert(Subtarget->isTargetELF() &&
1915 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1916 MachineFunction &MF = DAG.getMachineFunction();
1917 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1918 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1919 EVT PtrVT = getPointerTy();
1920 DebugLoc dl = Op.getDebugLoc();
1921 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1922 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1923 "_GLOBAL_OFFSET_TABLE_",
1924 ARMPCLabelIndex, PCAdj);
1925 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1926 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1927 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1928 PseudoSourceValue::getConstantPool(), 0,
1930 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1931 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1935 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1936 DebugLoc dl = Op.getDebugLoc();
1937 SDValue Val = DAG.getConstant(0, MVT::i32);
1938 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1939 Op.getOperand(1), Val);
1943 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1944 DebugLoc dl = Op.getDebugLoc();
1945 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1946 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1950 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1951 const ARMSubtarget *Subtarget) const {
1952 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1953 DebugLoc dl = Op.getDebugLoc();
1955 default: return SDValue(); // Don't custom lower most intrinsics.
1956 case Intrinsic::arm_thread_pointer: {
1957 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1958 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1960 case Intrinsic::eh_sjlj_lsda: {
1961 MachineFunction &MF = DAG.getMachineFunction();
1962 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1963 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1964 EVT PtrVT = getPointerTy();
1965 DebugLoc dl = Op.getDebugLoc();
1966 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1968 unsigned PCAdj = (RelocM != Reloc::PIC_)
1969 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1970 ARMConstantPoolValue *CPV =
1971 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1972 ARMCP::CPLSDA, PCAdj);
1973 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1974 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1976 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1977 PseudoSourceValue::getConstantPool(), 0,
1980 if (RelocM == Reloc::PIC_) {
1981 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1982 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1989 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1990 const ARMSubtarget *Subtarget) {
1991 DebugLoc dl = Op.getDebugLoc();
1992 SDValue Op5 = Op.getOperand(5);
1993 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1994 // v6 and v7 can both handle barriers directly, but need handled a bit
1995 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1997 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1998 if (Subtarget->hasV7Ops())
1999 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2000 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
2001 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2002 DAG.getConstant(0, MVT::i32));
2003 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2007 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2008 MachineFunction &MF = DAG.getMachineFunction();
2009 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2011 // vastart just stores the address of the VarArgsFrameIndex slot into the
2012 // memory location argument.
2013 DebugLoc dl = Op.getDebugLoc();
2014 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2015 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2016 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2017 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
2022 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2023 SDValue &Root, SelectionDAG &DAG,
2024 DebugLoc dl) const {
2025 MachineFunction &MF = DAG.getMachineFunction();
2026 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2028 TargetRegisterClass *RC;
2029 if (AFI->isThumb1OnlyFunction())
2030 RC = ARM::tGPRRegisterClass;
2032 RC = ARM::GPRRegisterClass;
2034 // Transform the arguments stored in physical registers into virtual ones.
2035 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2036 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2039 if (NextVA.isMemLoc()) {
2040 MachineFrameInfo *MFI = MF.getFrameInfo();
2041 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2043 // Create load node to retrieve arguments from the stack.
2044 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2045 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2046 PseudoSourceValue::getFixedStack(FI), 0,
2049 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2050 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2053 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2057 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2058 CallingConv::ID CallConv, bool isVarArg,
2059 const SmallVectorImpl<ISD::InputArg>
2061 DebugLoc dl, SelectionDAG &DAG,
2062 SmallVectorImpl<SDValue> &InVals)
2065 MachineFunction &MF = DAG.getMachineFunction();
2066 MachineFrameInfo *MFI = MF.getFrameInfo();
2068 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2070 // Assign locations to all of the incoming arguments.
2071 SmallVector<CCValAssign, 16> ArgLocs;
2072 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2074 CCInfo.AnalyzeFormalArguments(Ins,
2075 CCAssignFnForNode(CallConv, /* Return*/ false,
2078 SmallVector<SDValue, 16> ArgValues;
2080 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2081 CCValAssign &VA = ArgLocs[i];
2083 // Arguments stored in registers.
2084 if (VA.isRegLoc()) {
2085 EVT RegVT = VA.getLocVT();
2088 if (VA.needsCustom()) {
2089 // f64 and vector types are split up into multiple registers or
2090 // combinations of registers and stack slots.
2091 if (VA.getLocVT() == MVT::v2f64) {
2092 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2094 VA = ArgLocs[++i]; // skip ahead to next loc
2096 if (VA.isMemLoc()) {
2097 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2098 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2099 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2100 PseudoSourceValue::getFixedStack(FI), 0,
2103 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2106 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2107 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2108 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2109 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2110 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2112 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2115 TargetRegisterClass *RC;
2117 if (RegVT == MVT::f32)
2118 RC = ARM::SPRRegisterClass;
2119 else if (RegVT == MVT::f64)
2120 RC = ARM::DPRRegisterClass;
2121 else if (RegVT == MVT::v2f64)
2122 RC = ARM::QPRRegisterClass;
2123 else if (RegVT == MVT::i32)
2124 RC = (AFI->isThumb1OnlyFunction() ?
2125 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2127 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2129 // Transform the arguments in physical registers into virtual ones.
2130 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2131 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2134 // If this is an 8 or 16-bit value, it is really passed promoted
2135 // to 32 bits. Insert an assert[sz]ext to capture this, then
2136 // truncate to the right size.
2137 switch (VA.getLocInfo()) {
2138 default: llvm_unreachable("Unknown loc info!");
2139 case CCValAssign::Full: break;
2140 case CCValAssign::BCvt:
2141 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2143 case CCValAssign::SExt:
2144 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2145 DAG.getValueType(VA.getValVT()));
2146 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2148 case CCValAssign::ZExt:
2149 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2150 DAG.getValueType(VA.getValVT()));
2151 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2155 InVals.push_back(ArgValue);
2157 } else { // VA.isRegLoc()
2160 assert(VA.isMemLoc());
2161 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2163 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2164 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2166 // Create load nodes to retrieve arguments from the stack.
2167 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2168 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2169 PseudoSourceValue::getFixedStack(FI), 0,
2176 static const unsigned GPRArgRegs[] = {
2177 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2180 unsigned NumGPRs = CCInfo.getFirstUnallocated
2181 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2183 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2184 unsigned VARegSize = (4 - NumGPRs) * 4;
2185 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2186 unsigned ArgOffset = CCInfo.getNextStackOffset();
2187 if (VARegSaveSize) {
2188 // If this function is vararg, store any remaining integer argument regs
2189 // to their spots on the stack so that they may be loaded by deferencing
2190 // the result of va_next.
2191 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2192 AFI->setVarArgsFrameIndex(
2193 MFI->CreateFixedObject(VARegSaveSize,
2194 ArgOffset + VARegSaveSize - VARegSize,
2196 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2199 SmallVector<SDValue, 4> MemOps;
2200 for (; NumGPRs < 4; ++NumGPRs) {
2201 TargetRegisterClass *RC;
2202 if (AFI->isThumb1OnlyFunction())
2203 RC = ARM::tGPRRegisterClass;
2205 RC = ARM::GPRRegisterClass;
2207 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2208 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2210 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2211 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2212 0, false, false, 0);
2213 MemOps.push_back(Store);
2214 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2215 DAG.getConstant(4, getPointerTy()));
2217 if (!MemOps.empty())
2218 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2219 &MemOps[0], MemOps.size());
2221 // This will point to the next argument passed via stack.
2222 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2228 /// isFloatingPointZero - Return true if this is +0.0.
2229 static bool isFloatingPointZero(SDValue Op) {
2230 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2231 return CFP->getValueAPF().isPosZero();
2232 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2233 // Maybe this has already been legalized into the constant pool?
2234 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2235 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2236 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2237 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2238 return CFP->getValueAPF().isPosZero();
2244 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2245 /// the given operands.
2247 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2248 SDValue &ARMcc, SelectionDAG &DAG,
2249 DebugLoc dl) const {
2250 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2251 unsigned C = RHSC->getZExtValue();
2252 if (!isLegalICmpImmediate(C)) {
2253 // Constant does not fit, try adjusting it by one?
2258 if (isLegalICmpImmediate(C-1)) {
2259 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2260 RHS = DAG.getConstant(C-1, MVT::i32);
2265 if (C > 0 && isLegalICmpImmediate(C-1)) {
2266 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2267 RHS = DAG.getConstant(C-1, MVT::i32);
2272 if (isLegalICmpImmediate(C+1)) {
2273 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2274 RHS = DAG.getConstant(C+1, MVT::i32);
2279 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2280 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2281 RHS = DAG.getConstant(C+1, MVT::i32);
2288 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2289 ARMISD::NodeType CompareType;
2292 CompareType = ARMISD::CMP;
2297 CompareType = ARMISD::CMPZ;
2300 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2301 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2304 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2306 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2307 DebugLoc dl) const {
2309 if (!isFloatingPointZero(RHS))
2310 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2312 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2313 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2316 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2317 EVT VT = Op.getValueType();
2318 SDValue LHS = Op.getOperand(0);
2319 SDValue RHS = Op.getOperand(1);
2320 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2321 SDValue TrueVal = Op.getOperand(2);
2322 SDValue FalseVal = Op.getOperand(3);
2323 DebugLoc dl = Op.getDebugLoc();
2325 if (LHS.getValueType() == MVT::i32) {
2327 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2328 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2329 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2332 ARMCC::CondCodes CondCode, CondCode2;
2333 FPCCToARMCC(CC, CondCode, CondCode2);
2335 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2336 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2337 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2338 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2340 if (CondCode2 != ARMCC::AL) {
2341 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2342 // FIXME: Needs another CMP because flag can have but one use.
2343 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2344 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2345 Result, TrueVal, ARMcc2, CCR, Cmp2);
2350 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2351 /// to morph to an integer compare sequence.
2352 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2353 const ARMSubtarget *Subtarget) {
2354 SDNode *N = Op.getNode();
2355 if (!N->hasOneUse())
2356 // Otherwise it requires moving the value from fp to integer registers.
2358 if (!N->getNumValues())
2360 EVT VT = Op.getValueType();
2361 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2362 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2363 // vmrs are very slow, e.g. cortex-a8.
2366 if (isFloatingPointZero(Op)) {
2370 return ISD::isNormalLoad(N);
2373 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2374 if (isFloatingPointZero(Op))
2375 return DAG.getConstant(0, MVT::i32);
2377 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2378 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2379 Ld->getChain(), Ld->getBasePtr(),
2380 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2381 Ld->isVolatile(), Ld->isNonTemporal(),
2382 Ld->getAlignment());
2384 llvm_unreachable("Unknown VFP cmp argument!");
2387 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2388 SDValue &RetVal1, SDValue &RetVal2) {
2389 if (isFloatingPointZero(Op)) {
2390 RetVal1 = DAG.getConstant(0, MVT::i32);
2391 RetVal2 = DAG.getConstant(0, MVT::i32);
2395 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2396 SDValue Ptr = Ld->getBasePtr();
2397 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2398 Ld->getChain(), Ptr,
2399 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2400 Ld->isVolatile(), Ld->isNonTemporal(),
2401 Ld->getAlignment());
2403 EVT PtrType = Ptr.getValueType();
2404 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2405 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2406 PtrType, Ptr, DAG.getConstant(4, PtrType));
2407 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2408 Ld->getChain(), NewPtr,
2409 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2410 Ld->isVolatile(), Ld->isNonTemporal(),
2415 llvm_unreachable("Unknown VFP cmp argument!");
2418 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2419 /// f32 and even f64 comparisons to integer ones.
2421 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2422 SDValue Chain = Op.getOperand(0);
2423 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2424 SDValue LHS = Op.getOperand(2);
2425 SDValue RHS = Op.getOperand(3);
2426 SDValue Dest = Op.getOperand(4);
2427 DebugLoc dl = Op.getDebugLoc();
2429 bool SeenZero = false;
2430 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2431 canChangeToInt(RHS, SeenZero, Subtarget) &&
2432 // If one of the operand is zero, it's safe to ignore the NaN case since
2433 // we only care about equality comparisons.
2434 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2435 // If unsafe fp math optimization is enabled and there are no othter uses of
2436 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2437 // to an integer comparison.
2438 if (CC == ISD::SETOEQ)
2440 else if (CC == ISD::SETUNE)
2444 if (LHS.getValueType() == MVT::f32) {
2445 LHS = bitcastf32Toi32(LHS, DAG);
2446 RHS = bitcastf32Toi32(RHS, DAG);
2447 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2448 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2449 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2450 Chain, Dest, ARMcc, CCR, Cmp);
2455 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2456 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2457 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2458 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2459 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2460 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2461 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2467 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2468 SDValue Chain = Op.getOperand(0);
2469 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2470 SDValue LHS = Op.getOperand(2);
2471 SDValue RHS = Op.getOperand(3);
2472 SDValue Dest = Op.getOperand(4);
2473 DebugLoc dl = Op.getDebugLoc();
2475 if (LHS.getValueType() == MVT::i32) {
2477 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2478 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2479 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2480 Chain, Dest, ARMcc, CCR, Cmp);
2483 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2486 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2487 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2488 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2489 if (Result.getNode())
2493 ARMCC::CondCodes CondCode, CondCode2;
2494 FPCCToARMCC(CC, CondCode, CondCode2);
2496 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2497 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2498 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2499 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2500 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2501 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2502 if (CondCode2 != ARMCC::AL) {
2503 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2504 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2505 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2510 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2511 SDValue Chain = Op.getOperand(0);
2512 SDValue Table = Op.getOperand(1);
2513 SDValue Index = Op.getOperand(2);
2514 DebugLoc dl = Op.getDebugLoc();
2516 EVT PTy = getPointerTy();
2517 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2518 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2519 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2520 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2521 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2522 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2523 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2524 if (Subtarget->isThumb2()) {
2525 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2526 // which does another jump to the destination. This also makes it easier
2527 // to translate it to TBB / TBH later.
2528 // FIXME: This might not work if the function is extremely large.
2529 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2530 Addr, Op.getOperand(2), JTI, UId);
2532 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2533 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2534 PseudoSourceValue::getJumpTable(), 0,
2536 Chain = Addr.getValue(1);
2537 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2538 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2540 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2541 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2542 Chain = Addr.getValue(1);
2543 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2547 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2548 DebugLoc dl = Op.getDebugLoc();
2551 switch (Op.getOpcode()) {
2553 assert(0 && "Invalid opcode!");
2554 case ISD::FP_TO_SINT:
2555 Opc = ARMISD::FTOSI;
2557 case ISD::FP_TO_UINT:
2558 Opc = ARMISD::FTOUI;
2561 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2562 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2565 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2566 EVT VT = Op.getValueType();
2567 DebugLoc dl = Op.getDebugLoc();
2570 switch (Op.getOpcode()) {
2572 assert(0 && "Invalid opcode!");
2573 case ISD::SINT_TO_FP:
2574 Opc = ARMISD::SITOF;
2576 case ISD::UINT_TO_FP:
2577 Opc = ARMISD::UITOF;
2581 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2582 return DAG.getNode(Opc, dl, VT, Op);
2585 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2586 // Implement fcopysign with a fabs and a conditional fneg.
2587 SDValue Tmp0 = Op.getOperand(0);
2588 SDValue Tmp1 = Op.getOperand(1);
2589 DebugLoc dl = Op.getDebugLoc();
2590 EVT VT = Op.getValueType();
2591 EVT SrcVT = Tmp1.getValueType();
2592 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2593 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2594 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2595 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2596 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2597 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2600 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2601 MachineFunction &MF = DAG.getMachineFunction();
2602 MachineFrameInfo *MFI = MF.getFrameInfo();
2603 MFI->setReturnAddressIsTaken(true);
2605 EVT VT = Op.getValueType();
2606 DebugLoc dl = Op.getDebugLoc();
2607 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2609 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2610 SDValue Offset = DAG.getConstant(4, MVT::i32);
2611 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2612 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2613 NULL, 0, false, false, 0);
2616 // Return LR, which contains the return address. Mark it an implicit live-in.
2617 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2618 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2621 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2622 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2623 MFI->setFrameAddressIsTaken(true);
2625 EVT VT = Op.getValueType();
2626 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2627 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2628 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2629 ? ARM::R7 : ARM::R11;
2630 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2632 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2637 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2638 /// expand a bit convert where either the source or destination type is i64 to
2639 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2640 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2641 /// vectors), since the legalizer won't know what to do with that.
2642 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2643 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2644 DebugLoc dl = N->getDebugLoc();
2645 SDValue Op = N->getOperand(0);
2647 // This function is only supposed to be called for i64 types, either as the
2648 // source or destination of the bit convert.
2649 EVT SrcVT = Op.getValueType();
2650 EVT DstVT = N->getValueType(0);
2651 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2652 "ExpandBIT_CONVERT called for non-i64 type");
2654 // Turn i64->f64 into VMOVDRR.
2655 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2656 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2657 DAG.getConstant(0, MVT::i32));
2658 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2659 DAG.getConstant(1, MVT::i32));
2660 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2661 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2664 // Turn f64->i64 into VMOVRRD.
2665 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2666 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2667 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2668 // Merge the pieces into a single i64 value.
2669 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2675 /// getZeroVector - Returns a vector of specified type with all zero elements.
2676 /// Zero vectors are used to represent vector negation and in those cases
2677 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2678 /// not support i64 elements, so sometimes the zero vectors will need to be
2679 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2681 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2682 assert(VT.isVector() && "Expected a vector type");
2683 // The canonical modified immediate encoding of a zero vector is....0!
2684 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2685 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2686 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2687 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2690 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2691 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2692 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2693 SelectionDAG &DAG) const {
2694 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2695 EVT VT = Op.getValueType();
2696 unsigned VTBits = VT.getSizeInBits();
2697 DebugLoc dl = Op.getDebugLoc();
2698 SDValue ShOpLo = Op.getOperand(0);
2699 SDValue ShOpHi = Op.getOperand(1);
2700 SDValue ShAmt = Op.getOperand(2);
2702 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2704 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2706 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2707 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2708 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2709 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2710 DAG.getConstant(VTBits, MVT::i32));
2711 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2712 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2713 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2715 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2716 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2718 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2719 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2722 SDValue Ops[2] = { Lo, Hi };
2723 return DAG.getMergeValues(Ops, 2, dl);
2726 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2727 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2728 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2729 SelectionDAG &DAG) const {
2730 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2731 EVT VT = Op.getValueType();
2732 unsigned VTBits = VT.getSizeInBits();
2733 DebugLoc dl = Op.getDebugLoc();
2734 SDValue ShOpLo = Op.getOperand(0);
2735 SDValue ShOpHi = Op.getOperand(1);
2736 SDValue ShAmt = Op.getOperand(2);
2739 assert(Op.getOpcode() == ISD::SHL_PARTS);
2740 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2741 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2742 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2743 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2744 DAG.getConstant(VTBits, MVT::i32));
2745 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2746 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2748 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2749 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2750 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2752 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2753 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2756 SDValue Ops[2] = { Lo, Hi };
2757 return DAG.getMergeValues(Ops, 2, dl);
2760 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2761 SelectionDAG &DAG) const {
2762 // The rounding mode is in bits 23:22 of the FPSCR.
2763 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2764 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2765 // so that the shift + and get folded into a bitfield extract.
2766 DebugLoc dl = Op.getDebugLoc();
2767 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2768 DAG.getConstant(Intrinsic::arm_get_fpscr,
2770 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2771 DAG.getConstant(1U << 22, MVT::i32));
2772 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2773 DAG.getConstant(22, MVT::i32));
2774 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2775 DAG.getConstant(3, MVT::i32));
2778 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2779 const ARMSubtarget *ST) {
2780 EVT VT = N->getValueType(0);
2781 DebugLoc dl = N->getDebugLoc();
2783 if (!ST->hasV6T2Ops())
2786 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2787 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2790 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2791 const ARMSubtarget *ST) {
2792 EVT VT = N->getValueType(0);
2793 DebugLoc dl = N->getDebugLoc();
2795 // Lower vector shifts on NEON to use VSHL.
2796 if (VT.isVector()) {
2797 assert(ST->hasNEON() && "unexpected vector shift");
2799 // Left shifts translate directly to the vshiftu intrinsic.
2800 if (N->getOpcode() == ISD::SHL)
2801 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2802 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2803 N->getOperand(0), N->getOperand(1));
2805 assert((N->getOpcode() == ISD::SRA ||
2806 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2808 // NEON uses the same intrinsics for both left and right shifts. For
2809 // right shifts, the shift amounts are negative, so negate the vector of
2811 EVT ShiftVT = N->getOperand(1).getValueType();
2812 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2813 getZeroVector(ShiftVT, DAG, dl),
2815 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2816 Intrinsic::arm_neon_vshifts :
2817 Intrinsic::arm_neon_vshiftu);
2818 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2819 DAG.getConstant(vshiftInt, MVT::i32),
2820 N->getOperand(0), NegatedCount);
2823 // We can get here for a node like i32 = ISD::SHL i32, i64
2827 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2828 "Unknown shift to lower!");
2830 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2831 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2832 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2835 // If we are in thumb mode, we don't have RRX.
2836 if (ST->isThumb1Only()) return SDValue();
2838 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2839 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2840 DAG.getConstant(0, MVT::i32));
2841 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2842 DAG.getConstant(1, MVT::i32));
2844 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2845 // captures the result into a carry flag.
2846 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2847 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2849 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2850 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2852 // Merge the pieces into a single i64 value.
2853 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2856 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2857 SDValue TmpOp0, TmpOp1;
2858 bool Invert = false;
2862 SDValue Op0 = Op.getOperand(0);
2863 SDValue Op1 = Op.getOperand(1);
2864 SDValue CC = Op.getOperand(2);
2865 EVT VT = Op.getValueType();
2866 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2867 DebugLoc dl = Op.getDebugLoc();
2869 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2870 switch (SetCCOpcode) {
2871 default: llvm_unreachable("Illegal FP comparison"); break;
2873 case ISD::SETNE: Invert = true; // Fallthrough
2875 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2877 case ISD::SETLT: Swap = true; // Fallthrough
2879 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2881 case ISD::SETLE: Swap = true; // Fallthrough
2883 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2884 case ISD::SETUGE: Swap = true; // Fallthrough
2885 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2886 case ISD::SETUGT: Swap = true; // Fallthrough
2887 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2888 case ISD::SETUEQ: Invert = true; // Fallthrough
2890 // Expand this to (OLT | OGT).
2894 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2895 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2897 case ISD::SETUO: Invert = true; // Fallthrough
2899 // Expand this to (OLT | OGE).
2903 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2904 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2908 // Integer comparisons.
2909 switch (SetCCOpcode) {
2910 default: llvm_unreachable("Illegal integer comparison"); break;
2911 case ISD::SETNE: Invert = true;
2912 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2913 case ISD::SETLT: Swap = true;
2914 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2915 case ISD::SETLE: Swap = true;
2916 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2917 case ISD::SETULT: Swap = true;
2918 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2919 case ISD::SETULE: Swap = true;
2920 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2923 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2924 if (Opc == ARMISD::VCEQ) {
2927 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2929 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2932 // Ignore bitconvert.
2933 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2934 AndOp = AndOp.getOperand(0);
2936 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2938 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2939 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2946 std::swap(Op0, Op1);
2948 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2951 Result = DAG.getNOT(dl, Result, VT);
2956 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2957 /// valid vector constant for a NEON instruction with a "modified immediate"
2958 /// operand (e.g., VMOV). If so, return the encoded value.
2959 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2960 unsigned SplatBitSize, SelectionDAG &DAG,
2961 EVT &VT, bool is128Bits, bool isVMOV) {
2962 unsigned OpCmode, Imm;
2964 // SplatBitSize is set to the smallest size that splats the vector, so a
2965 // zero vector will always have SplatBitSize == 8. However, NEON modified
2966 // immediate instructions others than VMOV do not support the 8-bit encoding
2967 // of a zero vector, and the default encoding of zero is supposed to be the
2972 switch (SplatBitSize) {
2976 // Any 1-byte value is OK. Op=0, Cmode=1110.
2977 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2980 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2984 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2985 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2986 if ((SplatBits & ~0xff) == 0) {
2987 // Value = 0x00nn: Op=x, Cmode=100x.
2992 if ((SplatBits & ~0xff00) == 0) {
2993 // Value = 0xnn00: Op=x, Cmode=101x.
2995 Imm = SplatBits >> 8;
3001 // NEON's 32-bit VMOV supports splat values where:
3002 // * only one byte is nonzero, or
3003 // * the least significant byte is 0xff and the second byte is nonzero, or
3004 // * the least significant 2 bytes are 0xff and the third is nonzero.
3005 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3006 if ((SplatBits & ~0xff) == 0) {
3007 // Value = 0x000000nn: Op=x, Cmode=000x.
3012 if ((SplatBits & ~0xff00) == 0) {
3013 // Value = 0x0000nn00: Op=x, Cmode=001x.
3015 Imm = SplatBits >> 8;
3018 if ((SplatBits & ~0xff0000) == 0) {
3019 // Value = 0x00nn0000: Op=x, Cmode=010x.
3021 Imm = SplatBits >> 16;
3024 if ((SplatBits & ~0xff000000) == 0) {
3025 // Value = 0xnn000000: Op=x, Cmode=011x.
3027 Imm = SplatBits >> 24;
3031 if ((SplatBits & ~0xffff) == 0 &&
3032 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3033 // Value = 0x0000nnff: Op=x, Cmode=1100.
3035 Imm = SplatBits >> 8;
3040 if ((SplatBits & ~0xffffff) == 0 &&
3041 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3042 // Value = 0x00nnffff: Op=x, Cmode=1101.
3044 Imm = SplatBits >> 16;
3045 SplatBits |= 0xffff;
3049 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3050 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3051 // VMOV.I32. A (very) minor optimization would be to replicate the value
3052 // and fall through here to test for a valid 64-bit splat. But, then the
3053 // caller would also need to check and handle the change in size.
3059 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3060 uint64_t BitMask = 0xff;
3062 unsigned ImmMask = 1;
3064 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3065 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3068 } else if ((SplatBits & BitMask) != 0) {
3074 // Op=1, Cmode=1110.
3077 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3082 llvm_unreachable("unexpected size for isNEONModifiedImm");
3086 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3087 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3090 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3091 bool &ReverseVEXT, unsigned &Imm) {
3092 unsigned NumElts = VT.getVectorNumElements();
3093 ReverseVEXT = false;
3096 // If this is a VEXT shuffle, the immediate value is the index of the first
3097 // element. The other shuffle indices must be the successive elements after
3099 unsigned ExpectedElt = Imm;
3100 for (unsigned i = 1; i < NumElts; ++i) {
3101 // Increment the expected index. If it wraps around, it may still be
3102 // a VEXT but the source vectors must be swapped.
3104 if (ExpectedElt == NumElts * 2) {
3109 if (ExpectedElt != static_cast<unsigned>(M[i]))
3113 // Adjust the index value if the source operands will be swapped.
3120 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3121 /// instruction with the specified blocksize. (The order of the elements
3122 /// within each block of the vector is reversed.)
3123 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3124 unsigned BlockSize) {
3125 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3126 "Only possible block sizes for VREV are: 16, 32, 64");
3128 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3132 unsigned NumElts = VT.getVectorNumElements();
3133 unsigned BlockElts = M[0] + 1;
3135 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3138 for (unsigned i = 0; i < NumElts; ++i) {
3139 if ((unsigned) M[i] !=
3140 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3147 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3148 unsigned &WhichResult) {
3149 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3153 unsigned NumElts = VT.getVectorNumElements();
3154 WhichResult = (M[0] == 0 ? 0 : 1);
3155 for (unsigned i = 0; i < NumElts; i += 2) {
3156 if ((unsigned) M[i] != i + WhichResult ||
3157 (unsigned) M[i+1] != i + NumElts + WhichResult)
3163 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3164 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3165 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3166 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3167 unsigned &WhichResult) {
3168 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3172 unsigned NumElts = VT.getVectorNumElements();
3173 WhichResult = (M[0] == 0 ? 0 : 1);
3174 for (unsigned i = 0; i < NumElts; i += 2) {
3175 if ((unsigned) M[i] != i + WhichResult ||
3176 (unsigned) M[i+1] != i + WhichResult)
3182 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3183 unsigned &WhichResult) {
3184 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3188 unsigned NumElts = VT.getVectorNumElements();
3189 WhichResult = (M[0] == 0 ? 0 : 1);
3190 for (unsigned i = 0; i != NumElts; ++i) {
3191 if ((unsigned) M[i] != 2 * i + WhichResult)
3195 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3196 if (VT.is64BitVector() && EltSz == 32)
3202 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3203 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3204 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3205 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3206 unsigned &WhichResult) {
3207 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3211 unsigned Half = VT.getVectorNumElements() / 2;
3212 WhichResult = (M[0] == 0 ? 0 : 1);
3213 for (unsigned j = 0; j != 2; ++j) {
3214 unsigned Idx = WhichResult;
3215 for (unsigned i = 0; i != Half; ++i) {
3216 if ((unsigned) M[i + j * Half] != Idx)
3222 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3223 if (VT.is64BitVector() && EltSz == 32)
3229 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3230 unsigned &WhichResult) {
3231 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3235 unsigned NumElts = VT.getVectorNumElements();
3236 WhichResult = (M[0] == 0 ? 0 : 1);
3237 unsigned Idx = WhichResult * NumElts / 2;
3238 for (unsigned i = 0; i != NumElts; i += 2) {
3239 if ((unsigned) M[i] != Idx ||
3240 (unsigned) M[i+1] != Idx + NumElts)
3245 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3246 if (VT.is64BitVector() && EltSz == 32)
3252 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3253 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3254 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3255 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3256 unsigned &WhichResult) {
3257 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3261 unsigned NumElts = VT.getVectorNumElements();
3262 WhichResult = (M[0] == 0 ? 0 : 1);
3263 unsigned Idx = WhichResult * NumElts / 2;
3264 for (unsigned i = 0; i != NumElts; i += 2) {
3265 if ((unsigned) M[i] != Idx ||
3266 (unsigned) M[i+1] != Idx)
3271 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3272 if (VT.is64BitVector() && EltSz == 32)
3278 // If N is an integer constant that can be moved into a register in one
3279 // instruction, return an SDValue of such a constant (will become a MOV
3280 // instruction). Otherwise return null.
3281 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3282 const ARMSubtarget *ST, DebugLoc dl) {
3284 if (!isa<ConstantSDNode>(N))
3286 Val = cast<ConstantSDNode>(N)->getZExtValue();
3288 if (ST->isThumb1Only()) {
3289 if (Val <= 255 || ~Val <= 255)
3290 return DAG.getConstant(Val, MVT::i32);
3292 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3293 return DAG.getConstant(Val, MVT::i32);
3298 // If this is a case we can't handle, return null and let the default
3299 // expansion code take care of it.
3300 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3301 const ARMSubtarget *ST) {
3302 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3303 DebugLoc dl = Op.getDebugLoc();
3304 EVT VT = Op.getValueType();
3306 APInt SplatBits, SplatUndef;
3307 unsigned SplatBitSize;
3309 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3310 if (SplatBitSize <= 64) {
3311 // Check if an immediate VMOV works.
3313 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3314 SplatUndef.getZExtValue(), SplatBitSize,
3315 DAG, VmovVT, VT.is128BitVector(), true);
3316 if (Val.getNode()) {
3317 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3318 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3321 // Try an immediate VMVN.
3322 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3323 ((1LL << SplatBitSize) - 1));
3324 Val = isNEONModifiedImm(NegatedImm,
3325 SplatUndef.getZExtValue(), SplatBitSize,
3326 DAG, VmovVT, VT.is128BitVector(), false);
3327 if (Val.getNode()) {
3328 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3329 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3334 // Scan through the operands to see if only one value is used.
3335 unsigned NumElts = VT.getVectorNumElements();
3336 bool isOnlyLowElement = true;
3337 bool usesOnlyOneValue = true;
3338 bool isConstant = true;
3340 for (unsigned i = 0; i < NumElts; ++i) {
3341 SDValue V = Op.getOperand(i);
3342 if (V.getOpcode() == ISD::UNDEF)
3345 isOnlyLowElement = false;
3346 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3349 if (!Value.getNode())
3351 else if (V != Value)
3352 usesOnlyOneValue = false;
3355 if (!Value.getNode())
3356 return DAG.getUNDEF(VT);
3358 if (isOnlyLowElement)
3359 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3361 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3363 if (EnableARMVDUPsplat) {
3364 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3365 // i32 and try again.
3366 if (usesOnlyOneValue && EltSize <= 32) {
3368 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3369 if (VT.getVectorElementType().isFloatingPoint()) {
3370 SmallVector<SDValue, 8> Ops;
3371 for (unsigned i = 0; i < NumElts; ++i)
3372 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3374 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3376 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3377 LowerBUILD_VECTOR(Val, DAG, ST));
3379 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3381 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3385 // If all elements are constants and the case above didn't get hit, fall back
3386 // to the default expansion, which will generate a load from the constant
3391 if (!EnableARMVDUPsplat) {
3392 // Use VDUP for non-constant splats.
3393 if (usesOnlyOneValue && EltSize <= 32)
3394 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3397 // Vectors with 32- or 64-bit elements can be built by directly assigning
3398 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3399 // will be legalized.
3400 if (EltSize >= 32) {
3401 // Do the expansion with floating-point types, since that is what the VFP
3402 // registers are defined to use, and since i64 is not legal.
3403 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3404 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3405 SmallVector<SDValue, 8> Ops;
3406 for (unsigned i = 0; i < NumElts; ++i)
3407 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3408 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3409 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3415 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3416 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3417 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3418 /// are assumed to be legal.
3420 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3422 if (VT.getVectorNumElements() == 4 &&
3423 (VT.is128BitVector() || VT.is64BitVector())) {
3424 unsigned PFIndexes[4];
3425 for (unsigned i = 0; i != 4; ++i) {
3429 PFIndexes[i] = M[i];
3432 // Compute the index in the perfect shuffle table.
3433 unsigned PFTableIndex =
3434 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3435 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3436 unsigned Cost = (PFEntry >> 30);
3443 unsigned Imm, WhichResult;
3445 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3446 return (EltSize >= 32 ||
3447 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3448 isVREVMask(M, VT, 64) ||
3449 isVREVMask(M, VT, 32) ||
3450 isVREVMask(M, VT, 16) ||
3451 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3452 isVTRNMask(M, VT, WhichResult) ||
3453 isVUZPMask(M, VT, WhichResult) ||
3454 isVZIPMask(M, VT, WhichResult) ||
3455 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3456 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3457 isVZIP_v_undef_Mask(M, VT, WhichResult));
3460 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3461 /// the specified operations to build the shuffle.
3462 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3463 SDValue RHS, SelectionDAG &DAG,
3465 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3466 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3467 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3470 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3479 OP_VUZPL, // VUZP, left result
3480 OP_VUZPR, // VUZP, right result
3481 OP_VZIPL, // VZIP, left result
3482 OP_VZIPR, // VZIP, right result
3483 OP_VTRNL, // VTRN, left result
3484 OP_VTRNR // VTRN, right result
3487 if (OpNum == OP_COPY) {
3488 if (LHSID == (1*9+2)*9+3) return LHS;
3489 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3493 SDValue OpLHS, OpRHS;
3494 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3495 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3496 EVT VT = OpLHS.getValueType();
3499 default: llvm_unreachable("Unknown shuffle opcode!");
3501 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3506 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3507 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3511 return DAG.getNode(ARMISD::VEXT, dl, VT,
3513 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3516 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3517 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3520 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3521 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3524 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3525 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3529 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3530 SDValue V1 = Op.getOperand(0);
3531 SDValue V2 = Op.getOperand(1);
3532 DebugLoc dl = Op.getDebugLoc();
3533 EVT VT = Op.getValueType();
3534 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3535 SmallVector<int, 8> ShuffleMask;
3537 // Convert shuffles that are directly supported on NEON to target-specific
3538 // DAG nodes, instead of keeping them as shuffles and matching them again
3539 // during code selection. This is more efficient and avoids the possibility
3540 // of inconsistencies between legalization and selection.
3541 // FIXME: floating-point vectors should be canonicalized to integer vectors
3542 // of the same time so that they get CSEd properly.
3543 SVN->getMask(ShuffleMask);
3545 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3546 if (EltSize <= 32) {
3547 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3548 int Lane = SVN->getSplatIndex();
3549 // If this is undef splat, generate it via "just" vdup, if possible.
3550 if (Lane == -1) Lane = 0;
3552 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3553 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3555 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3556 DAG.getConstant(Lane, MVT::i32));
3561 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3564 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3565 DAG.getConstant(Imm, MVT::i32));
3568 if (isVREVMask(ShuffleMask, VT, 64))
3569 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3570 if (isVREVMask(ShuffleMask, VT, 32))
3571 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3572 if (isVREVMask(ShuffleMask, VT, 16))
3573 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3575 // Check for Neon shuffles that modify both input vectors in place.
3576 // If both results are used, i.e., if there are two shuffles with the same
3577 // source operands and with masks corresponding to both results of one of
3578 // these operations, DAG memoization will ensure that a single node is
3579 // used for both shuffles.
3580 unsigned WhichResult;
3581 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3582 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3583 V1, V2).getValue(WhichResult);
3584 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3585 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3586 V1, V2).getValue(WhichResult);
3587 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3588 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3589 V1, V2).getValue(WhichResult);
3591 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3592 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3593 V1, V1).getValue(WhichResult);
3594 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3595 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3596 V1, V1).getValue(WhichResult);
3597 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3598 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3599 V1, V1).getValue(WhichResult);
3602 // If the shuffle is not directly supported and it has 4 elements, use
3603 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3604 unsigned NumElts = VT.getVectorNumElements();
3606 unsigned PFIndexes[4];
3607 for (unsigned i = 0; i != 4; ++i) {
3608 if (ShuffleMask[i] < 0)
3611 PFIndexes[i] = ShuffleMask[i];
3614 // Compute the index in the perfect shuffle table.
3615 unsigned PFTableIndex =
3616 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3617 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3618 unsigned Cost = (PFEntry >> 30);
3621 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3624 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3625 if (EltSize >= 32) {
3626 // Do the expansion with floating-point types, since that is what the VFP
3627 // registers are defined to use, and since i64 is not legal.
3628 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3629 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3630 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3631 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3632 SmallVector<SDValue, 8> Ops;
3633 for (unsigned i = 0; i < NumElts; ++i) {
3634 if (ShuffleMask[i] < 0)
3635 Ops.push_back(DAG.getUNDEF(EltVT));
3637 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3638 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3639 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3642 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3643 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3649 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3650 EVT VT = Op.getValueType();
3651 DebugLoc dl = Op.getDebugLoc();
3652 SDValue Vec = Op.getOperand(0);
3653 SDValue Lane = Op.getOperand(1);
3654 assert(VT == MVT::i32 &&
3655 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3656 "unexpected type for custom-lowering vector extract");
3657 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3660 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3661 // The only time a CONCAT_VECTORS operation can have legal types is when
3662 // two 64-bit vectors are concatenated to a 128-bit vector.
3663 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3664 "unexpected CONCAT_VECTORS");
3665 DebugLoc dl = Op.getDebugLoc();
3666 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3667 SDValue Op0 = Op.getOperand(0);
3668 SDValue Op1 = Op.getOperand(1);
3669 if (Op0.getOpcode() != ISD::UNDEF)
3670 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3671 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3672 DAG.getIntPtrConstant(0));
3673 if (Op1.getOpcode() != ISD::UNDEF)
3674 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3675 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3676 DAG.getIntPtrConstant(1));
3677 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3680 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3681 switch (Op.getOpcode()) {
3682 default: llvm_unreachable("Don't know how to custom lower this!");
3683 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3684 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3685 case ISD::GlobalAddress:
3686 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3687 LowerGlobalAddressELF(Op, DAG);
3688 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3689 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3690 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3691 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3692 case ISD::VASTART: return LowerVASTART(Op, DAG);
3693 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3694 case ISD::SINT_TO_FP:
3695 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3696 case ISD::FP_TO_SINT:
3697 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3698 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3699 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3700 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3701 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3702 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3703 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3704 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3706 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3709 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3710 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3711 case ISD::SRL_PARTS:
3712 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3713 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3714 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3715 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3716 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3717 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3718 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3719 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3724 /// ReplaceNodeResults - Replace the results of node with an illegal result
3725 /// type with new values built out of custom code.
3726 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3727 SmallVectorImpl<SDValue>&Results,
3728 SelectionDAG &DAG) const {
3730 switch (N->getOpcode()) {
3732 llvm_unreachable("Don't know how to custom expand this!");
3734 case ISD::BIT_CONVERT:
3735 Res = ExpandBIT_CONVERT(N, DAG);
3739 Res = LowerShift(N, DAG, Subtarget);
3743 Results.push_back(Res);
3746 //===----------------------------------------------------------------------===//
3747 // ARM Scheduler Hooks
3748 //===----------------------------------------------------------------------===//
3751 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3752 MachineBasicBlock *BB,
3753 unsigned Size) const {
3754 unsigned dest = MI->getOperand(0).getReg();
3755 unsigned ptr = MI->getOperand(1).getReg();
3756 unsigned oldval = MI->getOperand(2).getReg();
3757 unsigned newval = MI->getOperand(3).getReg();
3758 unsigned scratch = BB->getParent()->getRegInfo()
3759 .createVirtualRegister(ARM::GPRRegisterClass);
3760 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3761 DebugLoc dl = MI->getDebugLoc();
3762 bool isThumb2 = Subtarget->isThumb2();
3764 unsigned ldrOpc, strOpc;
3766 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3768 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3769 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3772 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3773 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3776 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3777 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3781 MachineFunction *MF = BB->getParent();
3782 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3783 MachineFunction::iterator It = BB;
3784 ++It; // insert the new blocks after the current block
3786 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3787 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3788 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3789 MF->insert(It, loop1MBB);
3790 MF->insert(It, loop2MBB);
3791 MF->insert(It, exitMBB);
3793 // Transfer the remainder of BB and its successor edges to exitMBB.
3794 exitMBB->splice(exitMBB->begin(), BB,
3795 llvm::next(MachineBasicBlock::iterator(MI)),
3797 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3801 // fallthrough --> loop1MBB
3802 BB->addSuccessor(loop1MBB);
3805 // ldrex dest, [ptr]
3809 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3810 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3811 .addReg(dest).addReg(oldval));
3812 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3813 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3814 BB->addSuccessor(loop2MBB);
3815 BB->addSuccessor(exitMBB);
3818 // strex scratch, newval, [ptr]
3822 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3824 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3825 .addReg(scratch).addImm(0));
3826 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3827 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3828 BB->addSuccessor(loop1MBB);
3829 BB->addSuccessor(exitMBB);
3835 MI->eraseFromParent(); // The instruction is gone now.
3841 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3842 unsigned Size, unsigned BinOpcode) const {
3843 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3844 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3846 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3847 MachineFunction *MF = BB->getParent();
3848 MachineFunction::iterator It = BB;
3851 unsigned dest = MI->getOperand(0).getReg();
3852 unsigned ptr = MI->getOperand(1).getReg();
3853 unsigned incr = MI->getOperand(2).getReg();
3854 DebugLoc dl = MI->getDebugLoc();
3856 bool isThumb2 = Subtarget->isThumb2();
3857 unsigned ldrOpc, strOpc;
3859 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3861 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3862 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3865 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3866 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3869 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3870 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3874 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3875 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3876 MF->insert(It, loopMBB);
3877 MF->insert(It, exitMBB);
3879 // Transfer the remainder of BB and its successor edges to exitMBB.
3880 exitMBB->splice(exitMBB->begin(), BB,
3881 llvm::next(MachineBasicBlock::iterator(MI)),
3883 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3885 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3886 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3887 unsigned scratch2 = (!BinOpcode) ? incr :
3888 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3892 // fallthrough --> loopMBB
3893 BB->addSuccessor(loopMBB);
3897 // <binop> scratch2, dest, incr
3898 // strex scratch, scratch2, ptr
3901 // fallthrough --> exitMBB
3903 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3905 // operand order needs to go the other way for NAND
3906 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3907 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3908 addReg(incr).addReg(dest)).addReg(0);
3910 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3911 addReg(dest).addReg(incr)).addReg(0);
3914 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3916 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3917 .addReg(scratch).addImm(0));
3918 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3919 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3921 BB->addSuccessor(loopMBB);
3922 BB->addSuccessor(exitMBB);
3928 MI->eraseFromParent(); // The instruction is gone now.
3934 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3935 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3936 E = MBB->succ_end(); I != E; ++I)
3939 llvm_unreachable("Expecting a BB with two successors!");
3943 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3944 MachineBasicBlock *BB) const {
3945 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3946 DebugLoc dl = MI->getDebugLoc();
3947 bool isThumb2 = Subtarget->isThumb2();
3948 switch (MI->getOpcode()) {
3951 llvm_unreachable("Unexpected instr type to insert");
3953 case ARM::ATOMIC_LOAD_ADD_I8:
3954 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3955 case ARM::ATOMIC_LOAD_ADD_I16:
3956 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3957 case ARM::ATOMIC_LOAD_ADD_I32:
3958 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3960 case ARM::ATOMIC_LOAD_AND_I8:
3961 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3962 case ARM::ATOMIC_LOAD_AND_I16:
3963 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3964 case ARM::ATOMIC_LOAD_AND_I32:
3965 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3967 case ARM::ATOMIC_LOAD_OR_I8:
3968 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3969 case ARM::ATOMIC_LOAD_OR_I16:
3970 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3971 case ARM::ATOMIC_LOAD_OR_I32:
3972 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3974 case ARM::ATOMIC_LOAD_XOR_I8:
3975 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3976 case ARM::ATOMIC_LOAD_XOR_I16:
3977 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3978 case ARM::ATOMIC_LOAD_XOR_I32:
3979 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3981 case ARM::ATOMIC_LOAD_NAND_I8:
3982 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3983 case ARM::ATOMIC_LOAD_NAND_I16:
3984 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3985 case ARM::ATOMIC_LOAD_NAND_I32:
3986 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3988 case ARM::ATOMIC_LOAD_SUB_I8:
3989 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3990 case ARM::ATOMIC_LOAD_SUB_I16:
3991 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3992 case ARM::ATOMIC_LOAD_SUB_I32:
3993 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3995 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3996 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3997 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3999 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4000 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4001 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4003 case ARM::tMOVCCr_pseudo: {
4004 // To "insert" a SELECT_CC instruction, we actually have to insert the
4005 // diamond control-flow pattern. The incoming instruction knows the
4006 // destination vreg to set, the condition code register to branch on, the
4007 // true/false values to select between, and a branch opcode to use.
4008 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4009 MachineFunction::iterator It = BB;
4015 // cmpTY ccX, r1, r2
4017 // fallthrough --> copy0MBB
4018 MachineBasicBlock *thisMBB = BB;
4019 MachineFunction *F = BB->getParent();
4020 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4021 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4022 F->insert(It, copy0MBB);
4023 F->insert(It, sinkMBB);
4025 // Transfer the remainder of BB and its successor edges to sinkMBB.
4026 sinkMBB->splice(sinkMBB->begin(), BB,
4027 llvm::next(MachineBasicBlock::iterator(MI)),
4029 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4031 BB->addSuccessor(copy0MBB);
4032 BB->addSuccessor(sinkMBB);
4034 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4035 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4038 // %FalseValue = ...
4039 // # fallthrough to sinkMBB
4042 // Update machine-CFG edges
4043 BB->addSuccessor(sinkMBB);
4046 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4049 BuildMI(*BB, BB->begin(), dl,
4050 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4051 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4052 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4054 MI->eraseFromParent(); // The pseudo instruction is gone now.
4059 case ARM::BCCZi64: {
4060 // Compare both parts that make up the double comparison separately for
4062 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4064 unsigned LHS1 = MI->getOperand(1).getReg();
4065 unsigned LHS2 = MI->getOperand(2).getReg();
4067 AddDefaultPred(BuildMI(BB, dl,
4068 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4069 .addReg(LHS1).addImm(0));
4070 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4071 .addReg(LHS2).addImm(0)
4072 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4074 unsigned RHS1 = MI->getOperand(3).getReg();
4075 unsigned RHS2 = MI->getOperand(4).getReg();
4076 AddDefaultPred(BuildMI(BB, dl,
4077 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4078 .addReg(LHS1).addReg(RHS1));
4079 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4080 .addReg(LHS2).addReg(RHS2)
4081 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4084 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4085 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4086 if (MI->getOperand(0).getImm() == ARMCC::NE)
4087 std::swap(destMBB, exitMBB);
4089 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4090 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4091 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4094 MI->eraseFromParent(); // The pseudo instruction is gone now.
4101 case ARM::t2SUBrSPi_:
4102 case ARM::t2SUBrSPi12_:
4103 case ARM::t2SUBrSPs_: {
4104 MachineFunction *MF = BB->getParent();
4105 unsigned DstReg = MI->getOperand(0).getReg();
4106 unsigned SrcReg = MI->getOperand(1).getReg();
4107 bool DstIsDead = MI->getOperand(0).isDead();
4108 bool SrcIsKill = MI->getOperand(1).isKill();
4110 if (SrcReg != ARM::SP) {
4111 // Copy the source to SP from virtual register.
4112 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
4113 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4114 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
4115 BuildMI(*BB, MI, dl, TII->get(CopyOpc), ARM::SP)
4116 .addReg(SrcReg, getKillRegState(SrcIsKill));
4120 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
4121 switch (MI->getOpcode()) {
4123 llvm_unreachable("Unexpected pseudo instruction!");
4129 OpOpc = ARM::tADDspr;
4132 OpOpc = ARM::tSUBspi;
4134 case ARM::t2SUBrSPi_:
4135 OpOpc = ARM::t2SUBrSPi;
4136 NeedPred = true; NeedCC = true;
4138 case ARM::t2SUBrSPi12_:
4139 OpOpc = ARM::t2SUBrSPi12;
4142 case ARM::t2SUBrSPs_:
4143 OpOpc = ARM::t2SUBrSPs;
4144 NeedPred = true; NeedCC = true; NeedOp3 = true;
4147 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(OpOpc), ARM::SP);
4148 if (OpOpc == ARM::tAND)
4149 AddDefaultT1CC(MIB);
4150 MIB.addReg(ARM::SP);
4151 MIB.addOperand(MI->getOperand(2));
4153 MIB.addOperand(MI->getOperand(3));
4155 AddDefaultPred(MIB);
4159 // Copy the result from SP to virtual register.
4160 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
4161 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4162 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
4163 BuildMI(*BB, MI, dl, TII->get(CopyOpc))
4164 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
4166 MI->eraseFromParent(); // The pseudo instruction is gone now.
4172 //===----------------------------------------------------------------------===//
4173 // ARM Optimization Hooks
4174 //===----------------------------------------------------------------------===//
4177 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4178 TargetLowering::DAGCombinerInfo &DCI) {
4179 SelectionDAG &DAG = DCI.DAG;
4180 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4181 EVT VT = N->getValueType(0);
4182 unsigned Opc = N->getOpcode();
4183 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4184 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4185 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4186 ISD::CondCode CC = ISD::SETCC_INVALID;
4189 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4191 SDValue CCOp = Slct.getOperand(0);
4192 if (CCOp.getOpcode() == ISD::SETCC)
4193 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4196 bool DoXform = false;
4198 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4201 if (LHS.getOpcode() == ISD::Constant &&
4202 cast<ConstantSDNode>(LHS)->isNullValue()) {
4204 } else if (CC != ISD::SETCC_INVALID &&
4205 RHS.getOpcode() == ISD::Constant &&
4206 cast<ConstantSDNode>(RHS)->isNullValue()) {
4207 std::swap(LHS, RHS);
4208 SDValue Op0 = Slct.getOperand(0);
4209 EVT OpVT = isSlctCC ? Op0.getValueType() :
4210 Op0.getOperand(0).getValueType();
4211 bool isInt = OpVT.isInteger();
4212 CC = ISD::getSetCCInverse(CC, isInt);
4214 if (!TLI.isCondCodeLegal(CC, OpVT))
4215 return SDValue(); // Inverse operator isn't legal.
4222 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4224 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4225 Slct.getOperand(0), Slct.getOperand(1), CC);
4226 SDValue CCOp = Slct.getOperand(0);
4228 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4229 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4230 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4231 CCOp, OtherOp, Result);
4236 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4237 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4238 /// called with the default operands, and if that fails, with commuted
4240 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4241 TargetLowering::DAGCombinerInfo &DCI) {
4242 SelectionDAG &DAG = DCI.DAG;
4244 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4245 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4246 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4247 if (Result.getNode()) return Result;
4250 // fold (add (arm_neon_vabd a, b) c) -> (arm_neon_vaba c, a, b)
4251 EVT VT = N->getValueType(0);
4252 if (N0.getOpcode() == ISD::INTRINSIC_WO_CHAIN && VT.isInteger()) {
4253 unsigned IntNo = cast<ConstantSDNode>(N0.getOperand(0))->getZExtValue();
4254 if (IntNo == Intrinsic::arm_neon_vabds)
4255 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), VT,
4256 DAG.getConstant(Intrinsic::arm_neon_vabas, MVT::i32),
4257 N1, N0.getOperand(1), N0.getOperand(2));
4258 if (IntNo == Intrinsic::arm_neon_vabdu)
4259 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), VT,
4260 DAG.getConstant(Intrinsic::arm_neon_vabau, MVT::i32),
4261 N1, N0.getOperand(1), N0.getOperand(2));
4267 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4269 static SDValue PerformADDCombine(SDNode *N,
4270 TargetLowering::DAGCombinerInfo &DCI) {
4271 SDValue N0 = N->getOperand(0);
4272 SDValue N1 = N->getOperand(1);
4274 // First try with the default operand order.
4275 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4276 if (Result.getNode())
4279 // If that didn't work, try again with the operands commuted.
4280 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4283 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4285 static SDValue PerformSUBCombine(SDNode *N,
4286 TargetLowering::DAGCombinerInfo &DCI) {
4287 SDValue N0 = N->getOperand(0);
4288 SDValue N1 = N->getOperand(1);
4290 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4291 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4292 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4293 if (Result.getNode()) return Result;
4299 static SDValue PerformMULCombine(SDNode *N,
4300 TargetLowering::DAGCombinerInfo &DCI,
4301 const ARMSubtarget *Subtarget) {
4302 SelectionDAG &DAG = DCI.DAG;
4304 if (Subtarget->isThumb1Only())
4307 if (DAG.getMachineFunction().
4308 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4311 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4314 EVT VT = N->getValueType(0);
4318 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4322 uint64_t MulAmt = C->getZExtValue();
4323 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4324 ShiftAmt = ShiftAmt & (32 - 1);
4325 SDValue V = N->getOperand(0);
4326 DebugLoc DL = N->getDebugLoc();
4329 MulAmt >>= ShiftAmt;
4330 if (isPowerOf2_32(MulAmt - 1)) {
4331 // (mul x, 2^N + 1) => (add (shl x, N), x)
4332 Res = DAG.getNode(ISD::ADD, DL, VT,
4333 V, DAG.getNode(ISD::SHL, DL, VT,
4334 V, DAG.getConstant(Log2_32(MulAmt-1),
4336 } else if (isPowerOf2_32(MulAmt + 1)) {
4337 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4338 Res = DAG.getNode(ISD::SUB, DL, VT,
4339 DAG.getNode(ISD::SHL, DL, VT,
4340 V, DAG.getConstant(Log2_32(MulAmt+1),
4347 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4348 DAG.getConstant(ShiftAmt, MVT::i32));
4350 // Do not add new nodes to DAG combiner worklist.
4351 DCI.CombineTo(N, Res, false);
4355 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4356 static SDValue PerformORCombine(SDNode *N,
4357 TargetLowering::DAGCombinerInfo &DCI,
4358 const ARMSubtarget *Subtarget) {
4359 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4362 // BFI is only available on V6T2+
4363 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4366 SelectionDAG &DAG = DCI.DAG;
4367 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4368 DebugLoc DL = N->getDebugLoc();
4369 // 1) or (and A, mask), val => ARMbfi A, val, mask
4370 // iff (val & mask) == val
4372 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4373 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4374 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4375 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4376 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4377 // (i.e., copy a bitfield value into another bitfield of the same width)
4378 if (N0.getOpcode() != ISD::AND)
4381 EVT VT = N->getValueType(0);
4386 // The value and the mask need to be constants so we can verify this is
4387 // actually a bitfield set. If the mask is 0xffff, we can do better
4388 // via a movt instruction, so don't use BFI in that case.
4389 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4392 unsigned Mask = C->getZExtValue();
4396 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4397 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4398 unsigned Val = C->getZExtValue();
4399 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4401 Val >>= CountTrailingZeros_32(~Mask);
4403 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4404 DAG.getConstant(Val, MVT::i32),
4405 DAG.getConstant(Mask, MVT::i32));
4407 // Do not add new nodes to DAG combiner worklist.
4408 DCI.CombineTo(N, Res, false);
4409 } else if (N1.getOpcode() == ISD::AND) {
4410 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4411 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4414 unsigned Mask2 = C->getZExtValue();
4416 if (ARM::isBitFieldInvertedMask(Mask) &&
4417 ARM::isBitFieldInvertedMask(~Mask2) &&
4418 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4419 // The pack halfword instruction works better for masks that fit it,
4420 // so use that when it's available.
4421 if (Subtarget->hasT2ExtractPack() &&
4422 (Mask == 0xffff || Mask == 0xffff0000))
4425 unsigned lsb = CountTrailingZeros_32(Mask2);
4426 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4427 DAG.getConstant(lsb, MVT::i32));
4428 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4429 DAG.getConstant(Mask, MVT::i32));
4430 // Do not add new nodes to DAG combiner worklist.
4431 DCI.CombineTo(N, Res, false);
4432 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4433 ARM::isBitFieldInvertedMask(Mask2) &&
4434 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4435 // The pack halfword instruction works better for masks that fit it,
4436 // so use that when it's available.
4437 if (Subtarget->hasT2ExtractPack() &&
4438 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4441 unsigned lsb = CountTrailingZeros_32(Mask);
4442 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4443 DAG.getConstant(lsb, MVT::i32));
4444 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4445 DAG.getConstant(Mask2, MVT::i32));
4446 // Do not add new nodes to DAG combiner worklist.
4447 DCI.CombineTo(N, Res, false);
4454 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4455 /// ARMISD::VMOVRRD.
4456 static SDValue PerformVMOVRRDCombine(SDNode *N,
4457 TargetLowering::DAGCombinerInfo &DCI) {
4458 // fmrrd(fmdrr x, y) -> x,y
4459 SDValue InDouble = N->getOperand(0);
4460 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4461 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4465 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4466 /// ARMISD::VDUPLANE.
4467 static SDValue PerformVDUPLANECombine(SDNode *N,
4468 TargetLowering::DAGCombinerInfo &DCI) {
4469 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4471 SDValue Op = N->getOperand(0);
4472 EVT VT = N->getValueType(0);
4474 // Ignore bit_converts.
4475 while (Op.getOpcode() == ISD::BIT_CONVERT)
4476 Op = Op.getOperand(0);
4477 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4480 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4481 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4482 // The canonical VMOV for a zero vector uses a 32-bit element size.
4483 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4485 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4487 if (EltSize > VT.getVectorElementType().getSizeInBits())
4490 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4491 return DCI.CombineTo(N, Res, false);
4494 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4495 /// operand of a vector shift operation, where all the elements of the
4496 /// build_vector must have the same constant integer value.
4497 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4498 // Ignore bit_converts.
4499 while (Op.getOpcode() == ISD::BIT_CONVERT)
4500 Op = Op.getOperand(0);
4501 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4502 APInt SplatBits, SplatUndef;
4503 unsigned SplatBitSize;
4505 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4506 HasAnyUndefs, ElementBits) ||
4507 SplatBitSize > ElementBits)
4509 Cnt = SplatBits.getSExtValue();
4513 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4514 /// operand of a vector shift left operation. That value must be in the range:
4515 /// 0 <= Value < ElementBits for a left shift; or
4516 /// 0 <= Value <= ElementBits for a long left shift.
4517 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4518 assert(VT.isVector() && "vector shift count is not a vector type");
4519 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4520 if (! getVShiftImm(Op, ElementBits, Cnt))
4522 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4525 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4526 /// operand of a vector shift right operation. For a shift opcode, the value
4527 /// is positive, but for an intrinsic the value count must be negative. The
4528 /// absolute value must be in the range:
4529 /// 1 <= |Value| <= ElementBits for a right shift; or
4530 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4531 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4533 assert(VT.isVector() && "vector shift count is not a vector type");
4534 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4535 if (! getVShiftImm(Op, ElementBits, Cnt))
4539 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4542 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4543 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4544 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4547 // Don't do anything for most intrinsics.
4550 // Vector shifts: check for immediate versions and lower them.
4551 // Note: This is done during DAG combining instead of DAG legalizing because
4552 // the build_vectors for 64-bit vector element shift counts are generally
4553 // not legal, and it is hard to see their values after they get legalized to
4554 // loads from a constant pool.
4555 case Intrinsic::arm_neon_vshifts:
4556 case Intrinsic::arm_neon_vshiftu:
4557 case Intrinsic::arm_neon_vshiftls:
4558 case Intrinsic::arm_neon_vshiftlu:
4559 case Intrinsic::arm_neon_vshiftn:
4560 case Intrinsic::arm_neon_vrshifts:
4561 case Intrinsic::arm_neon_vrshiftu:
4562 case Intrinsic::arm_neon_vrshiftn:
4563 case Intrinsic::arm_neon_vqshifts:
4564 case Intrinsic::arm_neon_vqshiftu:
4565 case Intrinsic::arm_neon_vqshiftsu:
4566 case Intrinsic::arm_neon_vqshiftns:
4567 case Intrinsic::arm_neon_vqshiftnu:
4568 case Intrinsic::arm_neon_vqshiftnsu:
4569 case Intrinsic::arm_neon_vqrshiftns:
4570 case Intrinsic::arm_neon_vqrshiftnu:
4571 case Intrinsic::arm_neon_vqrshiftnsu: {
4572 EVT VT = N->getOperand(1).getValueType();
4574 unsigned VShiftOpc = 0;
4577 case Intrinsic::arm_neon_vshifts:
4578 case Intrinsic::arm_neon_vshiftu:
4579 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4580 VShiftOpc = ARMISD::VSHL;
4583 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4584 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4585 ARMISD::VSHRs : ARMISD::VSHRu);
4590 case Intrinsic::arm_neon_vshiftls:
4591 case Intrinsic::arm_neon_vshiftlu:
4592 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4594 llvm_unreachable("invalid shift count for vshll intrinsic");
4596 case Intrinsic::arm_neon_vrshifts:
4597 case Intrinsic::arm_neon_vrshiftu:
4598 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4602 case Intrinsic::arm_neon_vqshifts:
4603 case Intrinsic::arm_neon_vqshiftu:
4604 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4608 case Intrinsic::arm_neon_vqshiftsu:
4609 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4611 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4613 case Intrinsic::arm_neon_vshiftn:
4614 case Intrinsic::arm_neon_vrshiftn:
4615 case Intrinsic::arm_neon_vqshiftns:
4616 case Intrinsic::arm_neon_vqshiftnu:
4617 case Intrinsic::arm_neon_vqshiftnsu:
4618 case Intrinsic::arm_neon_vqrshiftns:
4619 case Intrinsic::arm_neon_vqrshiftnu:
4620 case Intrinsic::arm_neon_vqrshiftnsu:
4621 // Narrowing shifts require an immediate right shift.
4622 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4624 llvm_unreachable("invalid shift count for narrowing vector shift "
4628 llvm_unreachable("unhandled vector shift");
4632 case Intrinsic::arm_neon_vshifts:
4633 case Intrinsic::arm_neon_vshiftu:
4634 // Opcode already set above.
4636 case Intrinsic::arm_neon_vshiftls:
4637 case Intrinsic::arm_neon_vshiftlu:
4638 if (Cnt == VT.getVectorElementType().getSizeInBits())
4639 VShiftOpc = ARMISD::VSHLLi;
4641 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4642 ARMISD::VSHLLs : ARMISD::VSHLLu);
4644 case Intrinsic::arm_neon_vshiftn:
4645 VShiftOpc = ARMISD::VSHRN; break;
4646 case Intrinsic::arm_neon_vrshifts:
4647 VShiftOpc = ARMISD::VRSHRs; break;
4648 case Intrinsic::arm_neon_vrshiftu:
4649 VShiftOpc = ARMISD::VRSHRu; break;
4650 case Intrinsic::arm_neon_vrshiftn:
4651 VShiftOpc = ARMISD::VRSHRN; break;
4652 case Intrinsic::arm_neon_vqshifts:
4653 VShiftOpc = ARMISD::VQSHLs; break;
4654 case Intrinsic::arm_neon_vqshiftu:
4655 VShiftOpc = ARMISD::VQSHLu; break;
4656 case Intrinsic::arm_neon_vqshiftsu:
4657 VShiftOpc = ARMISD::VQSHLsu; break;
4658 case Intrinsic::arm_neon_vqshiftns:
4659 VShiftOpc = ARMISD::VQSHRNs; break;
4660 case Intrinsic::arm_neon_vqshiftnu:
4661 VShiftOpc = ARMISD::VQSHRNu; break;
4662 case Intrinsic::arm_neon_vqshiftnsu:
4663 VShiftOpc = ARMISD::VQSHRNsu; break;
4664 case Intrinsic::arm_neon_vqrshiftns:
4665 VShiftOpc = ARMISD::VQRSHRNs; break;
4666 case Intrinsic::arm_neon_vqrshiftnu:
4667 VShiftOpc = ARMISD::VQRSHRNu; break;
4668 case Intrinsic::arm_neon_vqrshiftnsu:
4669 VShiftOpc = ARMISD::VQRSHRNsu; break;
4672 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4673 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4676 case Intrinsic::arm_neon_vshiftins: {
4677 EVT VT = N->getOperand(1).getValueType();
4679 unsigned VShiftOpc = 0;
4681 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4682 VShiftOpc = ARMISD::VSLI;
4683 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4684 VShiftOpc = ARMISD::VSRI;
4686 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4689 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4690 N->getOperand(1), N->getOperand(2),
4691 DAG.getConstant(Cnt, MVT::i32));
4694 case Intrinsic::arm_neon_vqrshifts:
4695 case Intrinsic::arm_neon_vqrshiftu:
4696 // No immediate versions of these to check for.
4703 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4704 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4705 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4706 /// vector element shift counts are generally not legal, and it is hard to see
4707 /// their values after they get legalized to loads from a constant pool.
4708 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4709 const ARMSubtarget *ST) {
4710 EVT VT = N->getValueType(0);
4712 // Nothing to be done for scalar shifts.
4713 if (! VT.isVector())
4716 assert(ST->hasNEON() && "unexpected vector shift");
4719 switch (N->getOpcode()) {
4720 default: llvm_unreachable("unexpected shift opcode");
4723 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4724 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4725 DAG.getConstant(Cnt, MVT::i32));
4730 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4731 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4732 ARMISD::VSHRs : ARMISD::VSHRu);
4733 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4734 DAG.getConstant(Cnt, MVT::i32));
4740 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4741 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4742 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4743 const ARMSubtarget *ST) {
4744 SDValue N0 = N->getOperand(0);
4746 // Check for sign- and zero-extensions of vector extract operations of 8-
4747 // and 16-bit vector elements. NEON supports these directly. They are
4748 // handled during DAG combining because type legalization will promote them
4749 // to 32-bit types and it is messy to recognize the operations after that.
4750 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4751 SDValue Vec = N0.getOperand(0);
4752 SDValue Lane = N0.getOperand(1);
4753 EVT VT = N->getValueType(0);
4754 EVT EltVT = N0.getValueType();
4755 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4757 if (VT == MVT::i32 &&
4758 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4759 TLI.isTypeLegal(Vec.getValueType())) {
4762 switch (N->getOpcode()) {
4763 default: llvm_unreachable("unexpected opcode");
4764 case ISD::SIGN_EXTEND:
4765 Opc = ARMISD::VGETLANEs;
4767 case ISD::ZERO_EXTEND:
4768 case ISD::ANY_EXTEND:
4769 Opc = ARMISD::VGETLANEu;
4772 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4779 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4780 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4781 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4782 const ARMSubtarget *ST) {
4783 // If the target supports NEON, try to use vmax/vmin instructions for f32
4784 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4785 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4786 // a NaN; only do the transformation when it matches that behavior.
4788 // For now only do this when using NEON for FP operations; if using VFP, it
4789 // is not obvious that the benefit outweighs the cost of switching to the
4791 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4792 N->getValueType(0) != MVT::f32)
4795 SDValue CondLHS = N->getOperand(0);
4796 SDValue CondRHS = N->getOperand(1);
4797 SDValue LHS = N->getOperand(2);
4798 SDValue RHS = N->getOperand(3);
4799 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4801 unsigned Opcode = 0;
4803 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4804 IsReversed = false; // x CC y ? x : y
4805 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4806 IsReversed = true ; // x CC y ? y : x
4820 // If LHS is NaN, an ordered comparison will be false and the result will
4821 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4822 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4823 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4824 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4826 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4827 // will return -0, so vmin can only be used for unsafe math or if one of
4828 // the operands is known to be nonzero.
4829 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4831 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4833 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4842 // If LHS is NaN, an ordered comparison will be false and the result will
4843 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4844 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4845 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4846 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4848 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4849 // will return +0, so vmax can only be used for unsafe math or if one of
4850 // the operands is known to be nonzero.
4851 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4853 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4855 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4861 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4864 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4865 DAGCombinerInfo &DCI) const {
4866 switch (N->getOpcode()) {
4868 case ISD::ADD: return PerformADDCombine(N, DCI);
4869 case ISD::SUB: return PerformSUBCombine(N, DCI);
4870 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4871 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4872 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4873 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4874 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4877 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4878 case ISD::SIGN_EXTEND:
4879 case ISD::ZERO_EXTEND:
4880 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4881 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4886 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4887 if (!Subtarget->hasV6Ops())
4888 // Pre-v6 does not support unaligned mem access.
4891 // v6+ may or may not support unaligned mem access depending on the system
4893 // FIXME: This is pretty conservative. Should we provide cmdline option to
4894 // control the behaviour?
4895 if (!Subtarget->isTargetDarwin())
4898 switch (VT.getSimpleVT().SimpleTy) {
4905 // FIXME: VLD1 etc with standard alignment is legal.
4909 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4914 switch (VT.getSimpleVT().SimpleTy) {
4915 default: return false;
4930 if ((V & (Scale - 1)) != 0)
4933 return V == (V & ((1LL << 5) - 1));
4936 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4937 const ARMSubtarget *Subtarget) {
4944 switch (VT.getSimpleVT().SimpleTy) {
4945 default: return false;
4950 // + imm12 or - imm8
4952 return V == (V & ((1LL << 8) - 1));
4953 return V == (V & ((1LL << 12) - 1));
4956 // Same as ARM mode. FIXME: NEON?
4957 if (!Subtarget->hasVFP2())
4962 return V == (V & ((1LL << 8) - 1));
4966 /// isLegalAddressImmediate - Return true if the integer value can be used
4967 /// as the offset of the target addressing mode for load / store of the
4969 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4970 const ARMSubtarget *Subtarget) {
4977 if (Subtarget->isThumb1Only())
4978 return isLegalT1AddressImmediate(V, VT);
4979 else if (Subtarget->isThumb2())
4980 return isLegalT2AddressImmediate(V, VT, Subtarget);
4985 switch (VT.getSimpleVT().SimpleTy) {
4986 default: return false;
4991 return V == (V & ((1LL << 12) - 1));
4994 return V == (V & ((1LL << 8) - 1));
4997 if (!Subtarget->hasVFP2()) // FIXME: NEON?
5002 return V == (V & ((1LL << 8) - 1));
5006 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
5008 int Scale = AM.Scale;
5012 switch (VT.getSimpleVT().SimpleTy) {
5013 default: return false;
5022 return Scale == 2 || Scale == 4 || Scale == 8;
5025 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5029 // Note, we allow "void" uses (basically, uses that aren't loads or
5030 // stores), because arm allows folding a scale into many arithmetic
5031 // operations. This should be made more precise and revisited later.
5033 // Allow r << imm, but the imm has to be a multiple of two.
5034 if (Scale & 1) return false;
5035 return isPowerOf2_32(Scale);
5039 /// isLegalAddressingMode - Return true if the addressing mode represented
5040 /// by AM is legal for this target, for a load/store of the specified type.
5041 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5042 const Type *Ty) const {
5043 EVT VT = getValueType(Ty, true);
5044 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5047 // Can never fold addr of global into load/store.
5052 case 0: // no scale reg, must be "r+i" or "r", or "i".
5055 if (Subtarget->isThumb1Only())
5059 // ARM doesn't support any R+R*scale+imm addr modes.
5066 if (Subtarget->isThumb2())
5067 return isLegalT2ScaledAddressingMode(AM, VT);
5069 int Scale = AM.Scale;
5070 switch (VT.getSimpleVT().SimpleTy) {
5071 default: return false;
5075 if (Scale < 0) Scale = -Scale;
5079 return isPowerOf2_32(Scale & ~1);
5083 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5088 // Note, we allow "void" uses (basically, uses that aren't loads or
5089 // stores), because arm allows folding a scale into many arithmetic
5090 // operations. This should be made more precise and revisited later.
5092 // Allow r << imm, but the imm has to be a multiple of two.
5093 if (Scale & 1) return false;
5094 return isPowerOf2_32(Scale);
5101 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5102 /// icmp immediate, that is the target has icmp instructions which can compare
5103 /// a register against the immediate without having to materialize the
5104 /// immediate into a register.
5105 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5106 if (!Subtarget->isThumb())
5107 return ARM_AM::getSOImmVal(Imm) != -1;
5108 if (Subtarget->isThumb2())
5109 return ARM_AM::getT2SOImmVal(Imm) != -1;
5110 return Imm >= 0 && Imm <= 255;
5113 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5114 bool isSEXTLoad, SDValue &Base,
5115 SDValue &Offset, bool &isInc,
5116 SelectionDAG &DAG) {
5117 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5120 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5122 Base = Ptr->getOperand(0);
5123 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5124 int RHSC = (int)RHS->getZExtValue();
5125 if (RHSC < 0 && RHSC > -256) {
5126 assert(Ptr->getOpcode() == ISD::ADD);
5128 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5132 isInc = (Ptr->getOpcode() == ISD::ADD);
5133 Offset = Ptr->getOperand(1);
5135 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5137 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5138 int RHSC = (int)RHS->getZExtValue();
5139 if (RHSC < 0 && RHSC > -0x1000) {
5140 assert(Ptr->getOpcode() == ISD::ADD);
5142 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5143 Base = Ptr->getOperand(0);
5148 if (Ptr->getOpcode() == ISD::ADD) {
5150 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5151 if (ShOpcVal != ARM_AM::no_shift) {
5152 Base = Ptr->getOperand(1);
5153 Offset = Ptr->getOperand(0);
5155 Base = Ptr->getOperand(0);
5156 Offset = Ptr->getOperand(1);
5161 isInc = (Ptr->getOpcode() == ISD::ADD);
5162 Base = Ptr->getOperand(0);
5163 Offset = Ptr->getOperand(1);
5167 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5171 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5172 bool isSEXTLoad, SDValue &Base,
5173 SDValue &Offset, bool &isInc,
5174 SelectionDAG &DAG) {
5175 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5178 Base = Ptr->getOperand(0);
5179 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5180 int RHSC = (int)RHS->getZExtValue();
5181 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5182 assert(Ptr->getOpcode() == ISD::ADD);
5184 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5186 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5187 isInc = Ptr->getOpcode() == ISD::ADD;
5188 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5196 /// getPreIndexedAddressParts - returns true by value, base pointer and
5197 /// offset pointer and addressing mode by reference if the node's address
5198 /// can be legally represented as pre-indexed load / store address.
5200 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5202 ISD::MemIndexedMode &AM,
5203 SelectionDAG &DAG) const {
5204 if (Subtarget->isThumb1Only())
5209 bool isSEXTLoad = false;
5210 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5211 Ptr = LD->getBasePtr();
5212 VT = LD->getMemoryVT();
5213 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5214 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5215 Ptr = ST->getBasePtr();
5216 VT = ST->getMemoryVT();
5221 bool isLegal = false;
5222 if (Subtarget->isThumb2())
5223 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5224 Offset, isInc, DAG);
5226 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5227 Offset, isInc, DAG);
5231 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5235 /// getPostIndexedAddressParts - returns true by value, base pointer and
5236 /// offset pointer and addressing mode by reference if this node can be
5237 /// combined with a load / store to form a post-indexed load / store.
5238 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5241 ISD::MemIndexedMode &AM,
5242 SelectionDAG &DAG) const {
5243 if (Subtarget->isThumb1Only())
5248 bool isSEXTLoad = false;
5249 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5250 VT = LD->getMemoryVT();
5251 Ptr = LD->getBasePtr();
5252 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5253 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5254 VT = ST->getMemoryVT();
5255 Ptr = ST->getBasePtr();
5260 bool isLegal = false;
5261 if (Subtarget->isThumb2())
5262 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5265 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5271 // Swap base ptr and offset to catch more post-index load / store when
5272 // it's legal. In Thumb2 mode, offset must be an immediate.
5273 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5274 !Subtarget->isThumb2())
5275 std::swap(Base, Offset);
5277 // Post-indexed load / store update the base pointer.
5282 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5286 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5290 const SelectionDAG &DAG,
5291 unsigned Depth) const {
5292 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5293 switch (Op.getOpcode()) {
5295 case ARMISD::CMOV: {
5296 // Bits are known zero/one if known on the LHS and RHS.
5297 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5298 if (KnownZero == 0 && KnownOne == 0) return;
5300 APInt KnownZeroRHS, KnownOneRHS;
5301 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5302 KnownZeroRHS, KnownOneRHS, Depth+1);
5303 KnownZero &= KnownZeroRHS;
5304 KnownOne &= KnownOneRHS;
5310 //===----------------------------------------------------------------------===//
5311 // ARM Inline Assembly Support
5312 //===----------------------------------------------------------------------===//
5314 /// getConstraintType - Given a constraint letter, return the type of
5315 /// constraint it is for this target.
5316 ARMTargetLowering::ConstraintType
5317 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5318 if (Constraint.size() == 1) {
5319 switch (Constraint[0]) {
5321 case 'l': return C_RegisterClass;
5322 case 'w': return C_RegisterClass;
5325 return TargetLowering::getConstraintType(Constraint);
5328 std::pair<unsigned, const TargetRegisterClass*>
5329 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5331 if (Constraint.size() == 1) {
5332 // GCC ARM Constraint Letters
5333 switch (Constraint[0]) {
5335 if (Subtarget->isThumb())
5336 return std::make_pair(0U, ARM::tGPRRegisterClass);
5338 return std::make_pair(0U, ARM::GPRRegisterClass);
5340 return std::make_pair(0U, ARM::GPRRegisterClass);
5343 return std::make_pair(0U, ARM::SPRRegisterClass);
5344 if (VT.getSizeInBits() == 64)
5345 return std::make_pair(0U, ARM::DPRRegisterClass);
5346 if (VT.getSizeInBits() == 128)
5347 return std::make_pair(0U, ARM::QPRRegisterClass);
5351 if (StringRef("{cc}").equals_lower(Constraint))
5352 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5354 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5357 std::vector<unsigned> ARMTargetLowering::
5358 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5360 if (Constraint.size() != 1)
5361 return std::vector<unsigned>();
5363 switch (Constraint[0]) { // GCC ARM Constraint Letters
5366 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5367 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5370 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5371 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5372 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5373 ARM::R12, ARM::LR, 0);
5376 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5377 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5378 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5379 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5380 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5381 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5382 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5383 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5384 if (VT.getSizeInBits() == 64)
5385 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5386 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5387 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5388 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5389 if (VT.getSizeInBits() == 128)
5390 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5391 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5395 return std::vector<unsigned>();
5398 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5399 /// vector. If it is invalid, don't add anything to Ops.
5400 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5402 std::vector<SDValue>&Ops,
5403 SelectionDAG &DAG) const {
5404 SDValue Result(0, 0);
5406 switch (Constraint) {
5408 case 'I': case 'J': case 'K': case 'L':
5409 case 'M': case 'N': case 'O':
5410 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5414 int64_t CVal64 = C->getSExtValue();
5415 int CVal = (int) CVal64;
5416 // None of these constraints allow values larger than 32 bits. Check
5417 // that the value fits in an int.
5421 switch (Constraint) {
5423 if (Subtarget->isThumb1Only()) {
5424 // This must be a constant between 0 and 255, for ADD
5426 if (CVal >= 0 && CVal <= 255)
5428 } else if (Subtarget->isThumb2()) {
5429 // A constant that can be used as an immediate value in a
5430 // data-processing instruction.
5431 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5434 // A constant that can be used as an immediate value in a
5435 // data-processing instruction.
5436 if (ARM_AM::getSOImmVal(CVal) != -1)
5442 if (Subtarget->isThumb()) { // FIXME thumb2
5443 // This must be a constant between -255 and -1, for negated ADD
5444 // immediates. This can be used in GCC with an "n" modifier that
5445 // prints the negated value, for use with SUB instructions. It is
5446 // not useful otherwise but is implemented for compatibility.
5447 if (CVal >= -255 && CVal <= -1)
5450 // This must be a constant between -4095 and 4095. It is not clear
5451 // what this constraint is intended for. Implemented for
5452 // compatibility with GCC.
5453 if (CVal >= -4095 && CVal <= 4095)
5459 if (Subtarget->isThumb1Only()) {
5460 // A 32-bit value where only one byte has a nonzero value. Exclude
5461 // zero to match GCC. This constraint is used by GCC internally for
5462 // constants that can be loaded with a move/shift combination.
5463 // It is not useful otherwise but is implemented for compatibility.
5464 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5466 } else if (Subtarget->isThumb2()) {
5467 // A constant whose bitwise inverse can be used as an immediate
5468 // value in a data-processing instruction. This can be used in GCC
5469 // with a "B" modifier that prints the inverted value, for use with
5470 // BIC and MVN instructions. It is not useful otherwise but is
5471 // implemented for compatibility.
5472 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5475 // A constant whose bitwise inverse can be used as an immediate
5476 // value in a data-processing instruction. This can be used in GCC
5477 // with a "B" modifier that prints the inverted value, for use with
5478 // BIC and MVN instructions. It is not useful otherwise but is
5479 // implemented for compatibility.
5480 if (ARM_AM::getSOImmVal(~CVal) != -1)
5486 if (Subtarget->isThumb1Only()) {
5487 // This must be a constant between -7 and 7,
5488 // for 3-operand ADD/SUB immediate instructions.
5489 if (CVal >= -7 && CVal < 7)
5491 } else if (Subtarget->isThumb2()) {
5492 // A constant whose negation can be used as an immediate value in a
5493 // data-processing instruction. This can be used in GCC with an "n"
5494 // modifier that prints the negated value, for use with SUB
5495 // instructions. It is not useful otherwise but is implemented for
5497 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5500 // A constant whose negation can be used as an immediate value in a
5501 // data-processing instruction. This can be used in GCC with an "n"
5502 // modifier that prints the negated value, for use with SUB
5503 // instructions. It is not useful otherwise but is implemented for
5505 if (ARM_AM::getSOImmVal(-CVal) != -1)
5511 if (Subtarget->isThumb()) { // FIXME thumb2
5512 // This must be a multiple of 4 between 0 and 1020, for
5513 // ADD sp + immediate.
5514 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5517 // A power of two or a constant between 0 and 32. This is used in
5518 // GCC for the shift amount on shifted register operands, but it is
5519 // useful in general for any shift amounts.
5520 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5526 if (Subtarget->isThumb()) { // FIXME thumb2
5527 // This must be a constant between 0 and 31, for shift amounts.
5528 if (CVal >= 0 && CVal <= 31)
5534 if (Subtarget->isThumb()) { // FIXME thumb2
5535 // This must be a multiple of 4 between -508 and 508, for
5536 // ADD/SUB sp = sp + immediate.
5537 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5542 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5546 if (Result.getNode()) {
5547 Ops.push_back(Result);
5550 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5554 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5555 // The ARM target isn't yet aware of offsets.
5559 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5560 APInt Imm = FPImm.bitcastToAPInt();
5561 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5562 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5563 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5565 // We can handle 4 bits of mantissa.
5566 // mantissa = (16+UInt(e:f:g:h))/16.
5567 if (Mantissa & 0x7ffff)
5570 if ((Mantissa & 0xf) != Mantissa)
5573 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5574 if (Exp < -3 || Exp > 4)
5576 Exp = ((Exp+3) & 0x7) ^ 4;
5578 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5581 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5582 APInt Imm = FPImm.bitcastToAPInt();
5583 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5584 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5585 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5587 // We can handle 4 bits of mantissa.
5588 // mantissa = (16+UInt(e:f:g:h))/16.
5589 if (Mantissa & 0xffffffffffffLL)
5592 if ((Mantissa & 0xf) != Mantissa)
5595 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5596 if (Exp < -3 || Exp > 4)
5598 Exp = ((Exp+3) & 0x7) ^ 4;
5600 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5603 bool ARM::isBitFieldInvertedMask(unsigned v) {
5604 if (v == 0xffffffff)
5606 // there can be 1's on either or both "outsides", all the "inside"
5608 unsigned int lsb = 0, msb = 31;
5609 while (v & (1 << msb)) --msb;
5610 while (v & (1 << lsb)) ++lsb;
5611 for (unsigned int i = lsb; i <= msb; ++i) {
5618 /// isFPImmLegal - Returns true if the target can instruction select the
5619 /// specified FP immediate natively. If false, the legalizer will
5620 /// materialize the FP immediate as a load from a constant pool.
5621 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5622 if (!Subtarget->hasVFP3())
5625 return ARM::getVFPf32Imm(Imm) != -1;
5627 return ARM::getVFPf64Imm(Imm) != -1;