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 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
416 // the default expansion.
417 if (Subtarget->hasDataBarrier() ||
418 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
419 // membarrier needs custom lowering; the rest are legal and handled
421 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
423 // Set them all for expansion, which will force libcalls.
424 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
425 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
426 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
427 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
428 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
429 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
430 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
431 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
432 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
449 // Since the libcalls include locking, fold in the fences
450 setShouldFoldAtomicFences(true);
452 // 64-bit versions are always libcalls (for now)
453 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
454 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
455 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
456 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
457 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
458 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
459 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
460 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
462 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
463 if (!Subtarget->hasV6Ops()) {
464 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
465 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
467 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
469 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
470 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
471 // iff target supports vfp2.
472 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
473 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
476 // We want to custom lower some of our intrinsics.
477 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
478 if (Subtarget->isTargetDarwin()) {
479 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
480 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
483 setOperationAction(ISD::SETCC, MVT::i32, Expand);
484 setOperationAction(ISD::SETCC, MVT::f32, Expand);
485 setOperationAction(ISD::SETCC, MVT::f64, Expand);
486 setOperationAction(ISD::SELECT, MVT::i32, Custom);
487 setOperationAction(ISD::SELECT, MVT::f32, Custom);
488 setOperationAction(ISD::SELECT, MVT::f64, Custom);
489 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
490 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
491 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
493 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
494 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
495 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
496 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
497 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
499 // We don't support sin/cos/fmod/copysign/pow
500 setOperationAction(ISD::FSIN, MVT::f64, Expand);
501 setOperationAction(ISD::FSIN, MVT::f32, Expand);
502 setOperationAction(ISD::FCOS, MVT::f32, Expand);
503 setOperationAction(ISD::FCOS, MVT::f64, Expand);
504 setOperationAction(ISD::FREM, MVT::f64, Expand);
505 setOperationAction(ISD::FREM, MVT::f32, Expand);
506 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
507 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
508 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
510 setOperationAction(ISD::FPOW, MVT::f64, Expand);
511 setOperationAction(ISD::FPOW, MVT::f32, Expand);
513 // Various VFP goodness
514 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
515 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
516 if (Subtarget->hasVFP2()) {
517 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
518 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
519 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
520 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
522 // Special handling for half-precision FP.
523 if (!Subtarget->hasFP16()) {
524 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
525 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
529 // We have target-specific dag combine patterns for the following nodes:
530 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
531 setTargetDAGCombine(ISD::ADD);
532 setTargetDAGCombine(ISD::SUB);
533 setTargetDAGCombine(ISD::MUL);
535 if (Subtarget->hasV6T2Ops())
536 setTargetDAGCombine(ISD::OR);
538 setStackPointerRegisterToSaveRestore(ARM::SP);
540 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
541 setSchedulingPreference(Sched::RegPressure);
543 setSchedulingPreference(Sched::Hybrid);
545 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
547 // On ARM arguments smaller than 4 bytes are extended, so all arguments
548 // are at least 4 bytes aligned.
549 setMinStackArgumentAlignment(4);
551 if (EnableARMCodePlacement)
552 benefitFromCodePlacementOpt = true;
555 std::pair<const TargetRegisterClass*, uint8_t>
556 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
557 const TargetRegisterClass *RRC = 0;
559 switch (VT.getSimpleVT().SimpleTy) {
561 return TargetLowering::findRepresentativeClass(VT);
562 // Use DPR as representative register class for all floating point
563 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
564 // the cost is 1 for both f32 and f64.
565 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
566 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
567 RRC = ARM::DPRRegisterClass;
569 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
570 case MVT::v4f32: case MVT::v2f64:
571 RRC = ARM::DPRRegisterClass;
575 RRC = ARM::DPRRegisterClass;
579 RRC = ARM::DPRRegisterClass;
583 return std::make_pair(RRC, Cost);
586 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
589 case ARMISD::Wrapper: return "ARMISD::Wrapper";
590 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
591 case ARMISD::CALL: return "ARMISD::CALL";
592 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
593 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
594 case ARMISD::tCALL: return "ARMISD::tCALL";
595 case ARMISD::BRCOND: return "ARMISD::BRCOND";
596 case ARMISD::BR_JT: return "ARMISD::BR_JT";
597 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
598 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
599 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
600 case ARMISD::CMP: return "ARMISD::CMP";
601 case ARMISD::CMPZ: return "ARMISD::CMPZ";
602 case ARMISD::CMPFP: return "ARMISD::CMPFP";
603 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
604 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
605 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
606 case ARMISD::CMOV: return "ARMISD::CMOV";
607 case ARMISD::CNEG: return "ARMISD::CNEG";
609 case ARMISD::RBIT: return "ARMISD::RBIT";
611 case ARMISD::FTOSI: return "ARMISD::FTOSI";
612 case ARMISD::FTOUI: return "ARMISD::FTOUI";
613 case ARMISD::SITOF: return "ARMISD::SITOF";
614 case ARMISD::UITOF: return "ARMISD::UITOF";
616 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
617 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
618 case ARMISD::RRX: return "ARMISD::RRX";
620 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
621 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
623 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
624 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
626 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
628 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
630 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
632 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
633 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
635 case ARMISD::VCEQ: return "ARMISD::VCEQ";
636 case ARMISD::VCGE: return "ARMISD::VCGE";
637 case ARMISD::VCGEU: return "ARMISD::VCGEU";
638 case ARMISD::VCGT: return "ARMISD::VCGT";
639 case ARMISD::VCGTU: return "ARMISD::VCGTU";
640 case ARMISD::VTST: return "ARMISD::VTST";
642 case ARMISD::VSHL: return "ARMISD::VSHL";
643 case ARMISD::VSHRs: return "ARMISD::VSHRs";
644 case ARMISD::VSHRu: return "ARMISD::VSHRu";
645 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
646 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
647 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
648 case ARMISD::VSHRN: return "ARMISD::VSHRN";
649 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
650 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
651 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
652 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
653 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
654 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
655 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
656 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
657 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
658 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
659 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
660 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
661 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
662 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
663 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
664 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
665 case ARMISD::VDUP: return "ARMISD::VDUP";
666 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
667 case ARMISD::VEXT: return "ARMISD::VEXT";
668 case ARMISD::VREV64: return "ARMISD::VREV64";
669 case ARMISD::VREV32: return "ARMISD::VREV32";
670 case ARMISD::VREV16: return "ARMISD::VREV16";
671 case ARMISD::VZIP: return "ARMISD::VZIP";
672 case ARMISD::VUZP: return "ARMISD::VUZP";
673 case ARMISD::VTRN: return "ARMISD::VTRN";
674 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
675 case ARMISD::FMAX: return "ARMISD::FMAX";
676 case ARMISD::FMIN: return "ARMISD::FMIN";
677 case ARMISD::BFI: return "ARMISD::BFI";
681 /// getRegClassFor - Return the register class that should be used for the
682 /// specified value type.
683 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
684 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
685 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
686 // load / store 4 to 8 consecutive D registers.
687 if (Subtarget->hasNEON()) {
688 if (VT == MVT::v4i64)
689 return ARM::QQPRRegisterClass;
690 else if (VT == MVT::v8i64)
691 return ARM::QQQQPRRegisterClass;
693 return TargetLowering::getRegClassFor(VT);
696 // Create a fast isel object.
698 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
699 return ARM::createFastISel(funcInfo);
702 /// getFunctionAlignment - Return the Log2 alignment of this function.
703 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
704 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
707 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
708 /// be used for loads / stores from the global.
709 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
710 return (Subtarget->isThumb1Only() ? 127 : 4095);
713 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
714 unsigned NumVals = N->getNumValues();
716 return Sched::RegPressure;
718 for (unsigned i = 0; i != NumVals; ++i) {
719 EVT VT = N->getValueType(i);
720 if (VT.isFloatingPoint() || VT.isVector())
721 return Sched::Latency;
724 if (!N->isMachineOpcode())
725 return Sched::RegPressure;
727 // Load are scheduled for latency even if there instruction itinerary
729 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
730 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
732 return Sched::Latency;
734 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
735 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
736 return Sched::Latency;
737 return Sched::RegPressure;
741 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
742 MachineFunction &MF) const {
743 switch (RC->getID()) {
746 case ARM::tGPRRegClassID:
747 return RegInfo->hasFP(MF) ? 4 : 5;
748 case ARM::GPRRegClassID: {
749 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
750 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
752 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
753 case ARM::DPRRegClassID:
758 //===----------------------------------------------------------------------===//
760 //===----------------------------------------------------------------------===//
762 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
763 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
765 default: llvm_unreachable("Unknown condition code!");
766 case ISD::SETNE: return ARMCC::NE;
767 case ISD::SETEQ: return ARMCC::EQ;
768 case ISD::SETGT: return ARMCC::GT;
769 case ISD::SETGE: return ARMCC::GE;
770 case ISD::SETLT: return ARMCC::LT;
771 case ISD::SETLE: return ARMCC::LE;
772 case ISD::SETUGT: return ARMCC::HI;
773 case ISD::SETUGE: return ARMCC::HS;
774 case ISD::SETULT: return ARMCC::LO;
775 case ISD::SETULE: return ARMCC::LS;
779 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
780 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
781 ARMCC::CondCodes &CondCode2) {
782 CondCode2 = ARMCC::AL;
784 default: llvm_unreachable("Unknown FP condition!");
786 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
788 case ISD::SETOGT: CondCode = ARMCC::GT; break;
790 case ISD::SETOGE: CondCode = ARMCC::GE; break;
791 case ISD::SETOLT: CondCode = ARMCC::MI; break;
792 case ISD::SETOLE: CondCode = ARMCC::LS; break;
793 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
794 case ISD::SETO: CondCode = ARMCC::VC; break;
795 case ISD::SETUO: CondCode = ARMCC::VS; break;
796 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
797 case ISD::SETUGT: CondCode = ARMCC::HI; break;
798 case ISD::SETUGE: CondCode = ARMCC::PL; break;
800 case ISD::SETULT: CondCode = ARMCC::LT; break;
802 case ISD::SETULE: CondCode = ARMCC::LE; break;
804 case ISD::SETUNE: CondCode = ARMCC::NE; break;
808 //===----------------------------------------------------------------------===//
809 // Calling Convention Implementation
810 //===----------------------------------------------------------------------===//
812 #include "ARMGenCallingConv.inc"
814 // APCS f64 is in register pairs, possibly split to stack
815 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
816 CCValAssign::LocInfo &LocInfo,
817 CCState &State, bool CanFail) {
818 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
820 // Try to get the first register.
821 if (unsigned Reg = State.AllocateReg(RegList, 4))
822 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
824 // For the 2nd half of a v2f64, do not fail.
828 // Put the whole thing on the stack.
829 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
830 State.AllocateStack(8, 4),
835 // Try to get the second register.
836 if (unsigned Reg = State.AllocateReg(RegList, 4))
837 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
839 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
840 State.AllocateStack(4, 4),
845 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
846 CCValAssign::LocInfo &LocInfo,
847 ISD::ArgFlagsTy &ArgFlags,
849 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
851 if (LocVT == MVT::v2f64 &&
852 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
854 return true; // we handled it
857 // AAPCS f64 is in aligned register pairs
858 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
859 CCValAssign::LocInfo &LocInfo,
860 CCState &State, bool CanFail) {
861 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
862 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
863 static const unsigned ShadowRegList[] = { ARM::R0, ARM::R1 };
865 unsigned Reg = State.AllocateReg(HiRegList, ShadowRegList, 2);
867 // For the 2nd half of a v2f64, do not just fail.
871 // Put the whole thing on the stack.
872 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
873 State.AllocateStack(8, 8),
879 for (i = 0; i < 2; ++i)
880 if (HiRegList[i] == Reg)
883 unsigned T = State.AllocateReg(LoRegList[i]);
885 assert(T == LoRegList[i] && "Could not allocate register");
887 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
888 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
893 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
894 CCValAssign::LocInfo &LocInfo,
895 ISD::ArgFlagsTy &ArgFlags,
897 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
899 if (LocVT == MVT::v2f64 &&
900 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
902 return true; // we handled it
905 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
906 CCValAssign::LocInfo &LocInfo, CCState &State) {
907 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
908 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
910 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
912 return false; // we didn't handle it
915 for (i = 0; i < 2; ++i)
916 if (HiRegList[i] == Reg)
919 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
920 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
925 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
926 CCValAssign::LocInfo &LocInfo,
927 ISD::ArgFlagsTy &ArgFlags,
929 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
931 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
933 return true; // we handled it
936 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
937 CCValAssign::LocInfo &LocInfo,
938 ISD::ArgFlagsTy &ArgFlags,
940 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
944 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
945 /// given CallingConvention value.
946 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
948 bool isVarArg) const {
951 llvm_unreachable("Unsupported calling convention");
953 case CallingConv::Fast:
954 // Use target triple & subtarget features to do actual dispatch.
955 if (Subtarget->isAAPCS_ABI()) {
956 if (Subtarget->hasVFP2() &&
957 FloatABIType == FloatABI::Hard && !isVarArg)
958 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
960 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
962 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
963 case CallingConv::ARM_AAPCS_VFP:
964 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
965 case CallingConv::ARM_AAPCS:
966 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
967 case CallingConv::ARM_APCS:
968 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
972 /// LowerCallResult - Lower the result values of a call into the
973 /// appropriate copies out of appropriate physical registers.
975 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
976 CallingConv::ID CallConv, bool isVarArg,
977 const SmallVectorImpl<ISD::InputArg> &Ins,
978 DebugLoc dl, SelectionDAG &DAG,
979 SmallVectorImpl<SDValue> &InVals) const {
981 // Assign locations to each value returned by this call.
982 SmallVector<CCValAssign, 16> RVLocs;
983 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
984 RVLocs, *DAG.getContext());
985 CCInfo.AnalyzeCallResult(Ins,
986 CCAssignFnForNode(CallConv, /* Return*/ true,
989 // Copy all of the result registers out of their specified physreg.
990 for (unsigned i = 0; i != RVLocs.size(); ++i) {
991 CCValAssign VA = RVLocs[i];
994 if (VA.needsCustom()) {
995 // Handle f64 or half of a v2f64.
996 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
998 Chain = Lo.getValue(1);
999 InFlag = Lo.getValue(2);
1000 VA = RVLocs[++i]; // skip ahead to next loc
1001 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1003 Chain = Hi.getValue(1);
1004 InFlag = Hi.getValue(2);
1005 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1007 if (VA.getLocVT() == MVT::v2f64) {
1008 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1009 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1010 DAG.getConstant(0, MVT::i32));
1012 VA = RVLocs[++i]; // skip ahead to next loc
1013 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1014 Chain = Lo.getValue(1);
1015 InFlag = Lo.getValue(2);
1016 VA = RVLocs[++i]; // skip ahead to next loc
1017 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1018 Chain = Hi.getValue(1);
1019 InFlag = Hi.getValue(2);
1020 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1021 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1022 DAG.getConstant(1, MVT::i32));
1025 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1027 Chain = Val.getValue(1);
1028 InFlag = Val.getValue(2);
1031 switch (VA.getLocInfo()) {
1032 default: llvm_unreachable("Unknown loc info!");
1033 case CCValAssign::Full: break;
1034 case CCValAssign::BCvt:
1035 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1039 InVals.push_back(Val);
1045 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1046 /// by "Src" to address "Dst" of size "Size". Alignment information is
1047 /// specified by the specific parameter attribute. The copy will be passed as
1048 /// a byval function parameter.
1049 /// Sometimes what we are copying is the end of a larger object, the part that
1050 /// does not fit in registers.
1052 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1053 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1055 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1056 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1057 /*isVolatile=*/false, /*AlwaysInline=*/false,
1061 /// LowerMemOpCallTo - Store the argument to the stack.
1063 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1064 SDValue StackPtr, SDValue Arg,
1065 DebugLoc dl, SelectionDAG &DAG,
1066 const CCValAssign &VA,
1067 ISD::ArgFlagsTy Flags) const {
1068 unsigned LocMemOffset = VA.getLocMemOffset();
1069 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1070 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1071 if (Flags.isByVal()) {
1072 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1074 return DAG.getStore(Chain, dl, Arg, PtrOff,
1075 PseudoSourceValue::getStack(), LocMemOffset,
1079 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1080 SDValue Chain, SDValue &Arg,
1081 RegsToPassVector &RegsToPass,
1082 CCValAssign &VA, CCValAssign &NextVA,
1084 SmallVector<SDValue, 8> &MemOpChains,
1085 ISD::ArgFlagsTy Flags) const {
1087 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1088 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1089 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1091 if (NextVA.isRegLoc())
1092 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1094 assert(NextVA.isMemLoc());
1095 if (StackPtr.getNode() == 0)
1096 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1098 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1104 /// LowerCall - Lowering a call into a callseq_start <-
1105 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1108 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1109 CallingConv::ID CallConv, bool isVarArg,
1111 const SmallVectorImpl<ISD::OutputArg> &Outs,
1112 const SmallVectorImpl<SDValue> &OutVals,
1113 const SmallVectorImpl<ISD::InputArg> &Ins,
1114 DebugLoc dl, SelectionDAG &DAG,
1115 SmallVectorImpl<SDValue> &InVals) const {
1116 MachineFunction &MF = DAG.getMachineFunction();
1117 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1118 bool IsSibCall = false;
1120 // Check if it's really possible to do a tail call.
1121 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1122 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1123 Outs, OutVals, Ins, DAG);
1124 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1125 // detected sibcalls.
1132 // Analyze operands of the call, assigning locations to each operand.
1133 SmallVector<CCValAssign, 16> ArgLocs;
1134 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1136 CCInfo.AnalyzeCallOperands(Outs,
1137 CCAssignFnForNode(CallConv, /* Return*/ false,
1140 // Get a count of how many bytes are to be pushed on the stack.
1141 unsigned NumBytes = CCInfo.getNextStackOffset();
1143 // For tail calls, memory operands are available in our caller's stack.
1147 // Adjust the stack pointer for the new arguments...
1148 // These operations are automatically eliminated by the prolog/epilog pass
1150 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1152 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1154 RegsToPassVector RegsToPass;
1155 SmallVector<SDValue, 8> MemOpChains;
1157 // Walk the register/memloc assignments, inserting copies/loads. In the case
1158 // of tail call optimization, arguments are handled later.
1159 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1161 ++i, ++realArgIdx) {
1162 CCValAssign &VA = ArgLocs[i];
1163 SDValue Arg = OutVals[realArgIdx];
1164 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1166 // Promote the value if needed.
1167 switch (VA.getLocInfo()) {
1168 default: llvm_unreachable("Unknown loc info!");
1169 case CCValAssign::Full: break;
1170 case CCValAssign::SExt:
1171 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1173 case CCValAssign::ZExt:
1174 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1176 case CCValAssign::AExt:
1177 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1179 case CCValAssign::BCvt:
1180 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1184 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1185 if (VA.needsCustom()) {
1186 if (VA.getLocVT() == MVT::v2f64) {
1187 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1188 DAG.getConstant(0, MVT::i32));
1189 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1190 DAG.getConstant(1, MVT::i32));
1192 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1193 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1195 VA = ArgLocs[++i]; // skip ahead to next loc
1196 if (VA.isRegLoc()) {
1197 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1198 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1200 assert(VA.isMemLoc());
1202 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1203 dl, DAG, VA, Flags));
1206 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1207 StackPtr, MemOpChains, Flags);
1209 } else if (VA.isRegLoc()) {
1210 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1211 } else if (!IsSibCall) {
1212 assert(VA.isMemLoc());
1214 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1215 dl, DAG, VA, Flags));
1219 if (!MemOpChains.empty())
1220 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1221 &MemOpChains[0], MemOpChains.size());
1223 // Build a sequence of copy-to-reg nodes chained together with token chain
1224 // and flag operands which copy the outgoing args into the appropriate regs.
1226 // Tail call byval lowering might overwrite argument registers so in case of
1227 // tail call optimization the copies to registers are lowered later.
1229 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1230 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1231 RegsToPass[i].second, InFlag);
1232 InFlag = Chain.getValue(1);
1235 // For tail calls lower the arguments to the 'real' stack slot.
1237 // Force all the incoming stack arguments to be loaded from the stack
1238 // before any new outgoing arguments are stored to the stack, because the
1239 // outgoing stack slots may alias the incoming argument stack slots, and
1240 // the alias isn't otherwise explicit. This is slightly more conservative
1241 // than necessary, because it means that each store effectively depends
1242 // on every argument instead of just those arguments it would clobber.
1244 // Do not flag preceeding copytoreg stuff together with the following stuff.
1246 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1247 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1248 RegsToPass[i].second, InFlag);
1249 InFlag = Chain.getValue(1);
1254 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1255 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1256 // node so that legalize doesn't hack it.
1257 bool isDirect = false;
1258 bool isARMFunc = false;
1259 bool isLocalARMFunc = false;
1260 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1262 if (EnableARMLongCalls) {
1263 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1264 && "long-calls with non-static relocation model!");
1265 // Handle a global address or an external symbol. If it's not one of
1266 // those, the target's already in a register, so we don't need to do
1268 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1269 const GlobalValue *GV = G->getGlobal();
1270 // Create a constant pool entry for the callee address
1271 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1272 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1275 // Get the address of the callee into a register
1276 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1277 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1278 Callee = DAG.getLoad(getPointerTy(), dl,
1279 DAG.getEntryNode(), CPAddr,
1280 PseudoSourceValue::getConstantPool(), 0,
1282 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1283 const char *Sym = S->getSymbol();
1285 // Create a constant pool entry for the callee address
1286 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1287 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1288 Sym, ARMPCLabelIndex, 0);
1289 // Get the address of the callee into a register
1290 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1291 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1292 Callee = DAG.getLoad(getPointerTy(), dl,
1293 DAG.getEntryNode(), CPAddr,
1294 PseudoSourceValue::getConstantPool(), 0,
1297 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1298 const GlobalValue *GV = G->getGlobal();
1300 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1301 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1302 getTargetMachine().getRelocationModel() != Reloc::Static;
1303 isARMFunc = !Subtarget->isThumb() || isStub;
1304 // ARM call to a local ARM function is predicable.
1305 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1306 // tBX takes a register source operand.
1307 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1308 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1309 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1312 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1313 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1314 Callee = DAG.getLoad(getPointerTy(), dl,
1315 DAG.getEntryNode(), CPAddr,
1316 PseudoSourceValue::getConstantPool(), 0,
1318 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1319 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1320 getPointerTy(), Callee, PICLabel);
1322 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1323 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1325 bool isStub = Subtarget->isTargetDarwin() &&
1326 getTargetMachine().getRelocationModel() != Reloc::Static;
1327 isARMFunc = !Subtarget->isThumb() || isStub;
1328 // tBX takes a register source operand.
1329 const char *Sym = S->getSymbol();
1330 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1331 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1332 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1333 Sym, ARMPCLabelIndex, 4);
1334 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1335 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1336 Callee = DAG.getLoad(getPointerTy(), dl,
1337 DAG.getEntryNode(), CPAddr,
1338 PseudoSourceValue::getConstantPool(), 0,
1340 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1341 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1342 getPointerTy(), Callee, PICLabel);
1344 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1347 // FIXME: handle tail calls differently.
1349 if (Subtarget->isThumb()) {
1350 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1351 CallOpc = ARMISD::CALL_NOLINK;
1353 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1355 CallOpc = (isDirect || Subtarget->hasV5TOps())
1356 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1357 : ARMISD::CALL_NOLINK;
1360 std::vector<SDValue> Ops;
1361 Ops.push_back(Chain);
1362 Ops.push_back(Callee);
1364 // Add argument registers to the end of the list so that they are known live
1366 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1367 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1368 RegsToPass[i].second.getValueType()));
1370 if (InFlag.getNode())
1371 Ops.push_back(InFlag);
1373 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1375 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1377 // Returns a chain and a flag for retval copy to use.
1378 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1379 InFlag = Chain.getValue(1);
1381 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1382 DAG.getIntPtrConstant(0, true), InFlag);
1384 InFlag = Chain.getValue(1);
1386 // Handle result values, copying them out of physregs into vregs that we
1388 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1392 /// MatchingStackOffset - Return true if the given stack call argument is
1393 /// already available in the same position (relatively) of the caller's
1394 /// incoming argument stack.
1396 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1397 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1398 const ARMInstrInfo *TII) {
1399 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1401 if (Arg.getOpcode() == ISD::CopyFromReg) {
1402 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1403 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1405 MachineInstr *Def = MRI->getVRegDef(VR);
1408 if (!Flags.isByVal()) {
1409 if (!TII->isLoadFromStackSlot(Def, FI))
1414 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1415 if (Flags.isByVal())
1416 // ByVal argument is passed in as a pointer but it's now being
1417 // dereferenced. e.g.
1418 // define @foo(%struct.X* %A) {
1419 // tail call @bar(%struct.X* byval %A)
1422 SDValue Ptr = Ld->getBasePtr();
1423 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1426 FI = FINode->getIndex();
1430 assert(FI != INT_MAX);
1431 if (!MFI->isFixedObjectIndex(FI))
1433 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1436 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1437 /// for tail call optimization. Targets which want to do tail call
1438 /// optimization should implement this function.
1440 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1441 CallingConv::ID CalleeCC,
1443 bool isCalleeStructRet,
1444 bool isCallerStructRet,
1445 const SmallVectorImpl<ISD::OutputArg> &Outs,
1446 const SmallVectorImpl<SDValue> &OutVals,
1447 const SmallVectorImpl<ISD::InputArg> &Ins,
1448 SelectionDAG& DAG) const {
1449 const Function *CallerF = DAG.getMachineFunction().getFunction();
1450 CallingConv::ID CallerCC = CallerF->getCallingConv();
1451 bool CCMatch = CallerCC == CalleeCC;
1453 // Look for obvious safe cases to perform tail call optimization that do not
1454 // require ABI changes. This is what gcc calls sibcall.
1456 // Do not sibcall optimize vararg calls unless the call site is not passing
1458 if (isVarArg && !Outs.empty())
1461 // Also avoid sibcall optimization if either caller or callee uses struct
1462 // return semantics.
1463 if (isCalleeStructRet || isCallerStructRet)
1466 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1467 // emitEpilogue is not ready for them.
1468 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1469 // LR. This means if we need to reload LR, it takes an extra instructions,
1470 // which outweighs the value of the tail call; but here we don't know yet
1471 // whether LR is going to be used. Probably the right approach is to
1472 // generate the tail call here and turn it back into CALL/RET in
1473 // emitEpilogue if LR is used.
1474 if (Subtarget->isThumb1Only())
1477 // For the moment, we can only do this to functions defined in this
1478 // compilation, or to indirect calls. A Thumb B to an ARM function,
1479 // or vice versa, is not easily fixed up in the linker unlike BL.
1480 // (We could do this by loading the address of the callee into a register;
1481 // that is an extra instruction over the direct call and burns a register
1482 // as well, so is not likely to be a win.)
1484 // It might be safe to remove this restriction on non-Darwin.
1486 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1487 // but we need to make sure there are enough registers; the only valid
1488 // registers are the 4 used for parameters. We don't currently do this
1490 if (isa<ExternalSymbolSDNode>(Callee))
1493 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1494 const GlobalValue *GV = G->getGlobal();
1495 if (GV->isDeclaration() || GV->isWeakForLinker())
1499 // If the calling conventions do not match, then we'd better make sure the
1500 // results are returned in the same way as what the caller expects.
1502 SmallVector<CCValAssign, 16> RVLocs1;
1503 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1504 RVLocs1, *DAG.getContext());
1505 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1507 SmallVector<CCValAssign, 16> RVLocs2;
1508 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1509 RVLocs2, *DAG.getContext());
1510 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1512 if (RVLocs1.size() != RVLocs2.size())
1514 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1515 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1517 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1519 if (RVLocs1[i].isRegLoc()) {
1520 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1523 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1529 // If the callee takes no arguments then go on to check the results of the
1531 if (!Outs.empty()) {
1532 // Check if stack adjustment is needed. For now, do not do this if any
1533 // argument is passed on the stack.
1534 SmallVector<CCValAssign, 16> ArgLocs;
1535 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1536 ArgLocs, *DAG.getContext());
1537 CCInfo.AnalyzeCallOperands(Outs,
1538 CCAssignFnForNode(CalleeCC, false, isVarArg));
1539 if (CCInfo.getNextStackOffset()) {
1540 MachineFunction &MF = DAG.getMachineFunction();
1542 // Check if the arguments are already laid out in the right way as
1543 // the caller's fixed stack objects.
1544 MachineFrameInfo *MFI = MF.getFrameInfo();
1545 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1546 const ARMInstrInfo *TII =
1547 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1548 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1550 ++i, ++realArgIdx) {
1551 CCValAssign &VA = ArgLocs[i];
1552 EVT RegVT = VA.getLocVT();
1553 SDValue Arg = OutVals[realArgIdx];
1554 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1555 if (VA.getLocInfo() == CCValAssign::Indirect)
1557 if (VA.needsCustom()) {
1558 // f64 and vector types are split into multiple registers or
1559 // register/stack-slot combinations. The types will not match
1560 // the registers; give up on memory f64 refs until we figure
1561 // out what to do about this.
1564 if (!ArgLocs[++i].isRegLoc())
1566 if (RegVT == MVT::v2f64) {
1567 if (!ArgLocs[++i].isRegLoc())
1569 if (!ArgLocs[++i].isRegLoc())
1572 } else if (!VA.isRegLoc()) {
1573 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1585 ARMTargetLowering::LowerReturn(SDValue Chain,
1586 CallingConv::ID CallConv, bool isVarArg,
1587 const SmallVectorImpl<ISD::OutputArg> &Outs,
1588 const SmallVectorImpl<SDValue> &OutVals,
1589 DebugLoc dl, SelectionDAG &DAG) const {
1591 // CCValAssign - represent the assignment of the return value to a location.
1592 SmallVector<CCValAssign, 16> RVLocs;
1594 // CCState - Info about the registers and stack slots.
1595 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1598 // Analyze outgoing return values.
1599 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1602 // If this is the first return lowered for this function, add
1603 // the regs to the liveout set for the function.
1604 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1605 for (unsigned i = 0; i != RVLocs.size(); ++i)
1606 if (RVLocs[i].isRegLoc())
1607 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1612 // Copy the result values into the output registers.
1613 for (unsigned i = 0, realRVLocIdx = 0;
1615 ++i, ++realRVLocIdx) {
1616 CCValAssign &VA = RVLocs[i];
1617 assert(VA.isRegLoc() && "Can only return in registers!");
1619 SDValue Arg = OutVals[realRVLocIdx];
1621 switch (VA.getLocInfo()) {
1622 default: llvm_unreachable("Unknown loc info!");
1623 case CCValAssign::Full: break;
1624 case CCValAssign::BCvt:
1625 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1629 if (VA.needsCustom()) {
1630 if (VA.getLocVT() == MVT::v2f64) {
1631 // Extract the first half and return it in two registers.
1632 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1633 DAG.getConstant(0, MVT::i32));
1634 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1635 DAG.getVTList(MVT::i32, MVT::i32), Half);
1637 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1638 Flag = Chain.getValue(1);
1639 VA = RVLocs[++i]; // skip ahead to next loc
1640 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1641 HalfGPRs.getValue(1), Flag);
1642 Flag = Chain.getValue(1);
1643 VA = RVLocs[++i]; // skip ahead to next loc
1645 // Extract the 2nd half and fall through to handle it as an f64 value.
1646 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1647 DAG.getConstant(1, MVT::i32));
1649 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1651 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1652 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1653 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1654 Flag = Chain.getValue(1);
1655 VA = RVLocs[++i]; // skip ahead to next loc
1656 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1659 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1661 // Guarantee that all emitted copies are
1662 // stuck together, avoiding something bad.
1663 Flag = Chain.getValue(1);
1668 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1670 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1675 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1676 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1677 // one of the above mentioned nodes. It has to be wrapped because otherwise
1678 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1679 // be used to form addressing mode. These wrapped nodes will be selected
1681 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1682 EVT PtrVT = Op.getValueType();
1683 // FIXME there is no actual debug info here
1684 DebugLoc dl = Op.getDebugLoc();
1685 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1687 if (CP->isMachineConstantPoolEntry())
1688 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1689 CP->getAlignment());
1691 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1692 CP->getAlignment());
1693 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1696 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1697 return MachineJumpTableInfo::EK_Inline;
1700 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1701 SelectionDAG &DAG) const {
1702 MachineFunction &MF = DAG.getMachineFunction();
1703 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1704 unsigned ARMPCLabelIndex = 0;
1705 DebugLoc DL = Op.getDebugLoc();
1706 EVT PtrVT = getPointerTy();
1707 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1708 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1710 if (RelocM == Reloc::Static) {
1711 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1713 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1714 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1715 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1716 ARMCP::CPBlockAddress,
1718 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1720 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1721 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1722 PseudoSourceValue::getConstantPool(), 0,
1724 if (RelocM == Reloc::Static)
1726 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1727 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1730 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1732 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1733 SelectionDAG &DAG) const {
1734 DebugLoc dl = GA->getDebugLoc();
1735 EVT PtrVT = getPointerTy();
1736 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1737 MachineFunction &MF = DAG.getMachineFunction();
1738 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1739 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1740 ARMConstantPoolValue *CPV =
1741 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1742 ARMCP::CPValue, PCAdj, "tlsgd", true);
1743 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1744 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1745 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1746 PseudoSourceValue::getConstantPool(), 0,
1748 SDValue Chain = Argument.getValue(1);
1750 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1751 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1753 // call __tls_get_addr.
1756 Entry.Node = Argument;
1757 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1758 Args.push_back(Entry);
1759 // FIXME: is there useful debug info available here?
1760 std::pair<SDValue, SDValue> CallResult =
1761 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1762 false, false, false, false,
1763 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1764 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1765 return CallResult.first;
1768 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1769 // "local exec" model.
1771 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1772 SelectionDAG &DAG) const {
1773 const GlobalValue *GV = GA->getGlobal();
1774 DebugLoc dl = GA->getDebugLoc();
1776 SDValue Chain = DAG.getEntryNode();
1777 EVT PtrVT = getPointerTy();
1778 // Get the Thread Pointer
1779 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1781 if (GV->isDeclaration()) {
1782 MachineFunction &MF = DAG.getMachineFunction();
1783 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1784 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1785 // Initial exec model.
1786 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1787 ARMConstantPoolValue *CPV =
1788 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1789 ARMCP::CPValue, PCAdj, "gottpoff", true);
1790 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1791 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1792 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1793 PseudoSourceValue::getConstantPool(), 0,
1795 Chain = Offset.getValue(1);
1797 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1798 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1800 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1801 PseudoSourceValue::getConstantPool(), 0,
1805 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1806 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1807 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1808 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1809 PseudoSourceValue::getConstantPool(), 0,
1813 // The address of the thread local variable is the add of the thread
1814 // pointer with the offset of the variable.
1815 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1819 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1820 // TODO: implement the "local dynamic" model
1821 assert(Subtarget->isTargetELF() &&
1822 "TLS not implemented for non-ELF targets");
1823 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1824 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1825 // otherwise use the "Local Exec" TLS Model
1826 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1827 return LowerToTLSGeneralDynamicModel(GA, DAG);
1829 return LowerToTLSExecModels(GA, DAG);
1832 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1833 SelectionDAG &DAG) const {
1834 EVT PtrVT = getPointerTy();
1835 DebugLoc dl = Op.getDebugLoc();
1836 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1837 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1838 if (RelocM == Reloc::PIC_) {
1839 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1840 ARMConstantPoolValue *CPV =
1841 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1842 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1843 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1844 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1846 PseudoSourceValue::getConstantPool(), 0,
1848 SDValue Chain = Result.getValue(1);
1849 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1850 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1852 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1853 PseudoSourceValue::getGOT(), 0,
1857 // If we have T2 ops, we can materialize the address directly via movt/movw
1858 // pair. This is always cheaper.
1859 if (Subtarget->useMovt()) {
1860 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1861 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1863 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1864 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1865 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1866 PseudoSourceValue::getConstantPool(), 0,
1872 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1873 SelectionDAG &DAG) const {
1874 MachineFunction &MF = DAG.getMachineFunction();
1875 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1876 unsigned ARMPCLabelIndex = 0;
1877 EVT PtrVT = getPointerTy();
1878 DebugLoc dl = Op.getDebugLoc();
1879 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1880 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1882 if (RelocM == Reloc::Static)
1883 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1885 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1886 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1887 ARMConstantPoolValue *CPV =
1888 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1889 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1891 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1893 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1894 PseudoSourceValue::getConstantPool(), 0,
1896 SDValue Chain = Result.getValue(1);
1898 if (RelocM == Reloc::PIC_) {
1899 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1900 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1903 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1904 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1905 PseudoSourceValue::getGOT(), 0,
1911 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1912 SelectionDAG &DAG) const {
1913 assert(Subtarget->isTargetELF() &&
1914 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1915 MachineFunction &MF = DAG.getMachineFunction();
1916 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1917 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1918 EVT PtrVT = getPointerTy();
1919 DebugLoc dl = Op.getDebugLoc();
1920 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1921 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1922 "_GLOBAL_OFFSET_TABLE_",
1923 ARMPCLabelIndex, PCAdj);
1924 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1925 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1926 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1927 PseudoSourceValue::getConstantPool(), 0,
1929 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1930 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1934 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1935 DebugLoc dl = Op.getDebugLoc();
1936 SDValue Val = DAG.getConstant(0, MVT::i32);
1937 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1938 Op.getOperand(1), Val);
1942 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1943 DebugLoc dl = Op.getDebugLoc();
1944 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1945 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1949 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1950 const ARMSubtarget *Subtarget) const {
1951 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1952 DebugLoc dl = Op.getDebugLoc();
1954 default: return SDValue(); // Don't custom lower most intrinsics.
1955 case Intrinsic::arm_thread_pointer: {
1956 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1957 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1959 case Intrinsic::eh_sjlj_lsda: {
1960 MachineFunction &MF = DAG.getMachineFunction();
1961 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1962 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1963 EVT PtrVT = getPointerTy();
1964 DebugLoc dl = Op.getDebugLoc();
1965 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1967 unsigned PCAdj = (RelocM != Reloc::PIC_)
1968 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1969 ARMConstantPoolValue *CPV =
1970 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1971 ARMCP::CPLSDA, PCAdj);
1972 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1973 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1975 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1976 PseudoSourceValue::getConstantPool(), 0,
1979 if (RelocM == Reloc::PIC_) {
1980 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1981 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1988 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1989 const ARMSubtarget *Subtarget) {
1990 DebugLoc dl = Op.getDebugLoc();
1991 SDValue Op5 = Op.getOperand(5);
1992 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1993 // Some subtargets which have dmb and dsb instructions can handle barriers
1994 // directly. Some ARMv6 cpus can support them with the help of mcr
1995 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1997 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1998 if (Subtarget->hasDataBarrier())
1999 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
2001 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
2002 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2003 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
2004 DAG.getConstant(0, MVT::i32));
2008 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2009 MachineFunction &MF = DAG.getMachineFunction();
2010 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2012 // vastart just stores the address of the VarArgsFrameIndex slot into the
2013 // memory location argument.
2014 DebugLoc dl = Op.getDebugLoc();
2015 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2016 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2017 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2018 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
2023 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2024 SDValue &Root, SelectionDAG &DAG,
2025 DebugLoc dl) const {
2026 MachineFunction &MF = DAG.getMachineFunction();
2027 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2029 TargetRegisterClass *RC;
2030 if (AFI->isThumb1OnlyFunction())
2031 RC = ARM::tGPRRegisterClass;
2033 RC = ARM::GPRRegisterClass;
2035 // Transform the arguments stored in physical registers into virtual ones.
2036 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2037 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2040 if (NextVA.isMemLoc()) {
2041 MachineFrameInfo *MFI = MF.getFrameInfo();
2042 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2044 // Create load node to retrieve arguments from the stack.
2045 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2046 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2047 PseudoSourceValue::getFixedStack(FI), 0,
2050 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2051 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2054 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2058 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2059 CallingConv::ID CallConv, bool isVarArg,
2060 const SmallVectorImpl<ISD::InputArg>
2062 DebugLoc dl, SelectionDAG &DAG,
2063 SmallVectorImpl<SDValue> &InVals)
2066 MachineFunction &MF = DAG.getMachineFunction();
2067 MachineFrameInfo *MFI = MF.getFrameInfo();
2069 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2071 // Assign locations to all of the incoming arguments.
2072 SmallVector<CCValAssign, 16> ArgLocs;
2073 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2075 CCInfo.AnalyzeFormalArguments(Ins,
2076 CCAssignFnForNode(CallConv, /* Return*/ false,
2079 SmallVector<SDValue, 16> ArgValues;
2081 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2082 CCValAssign &VA = ArgLocs[i];
2084 // Arguments stored in registers.
2085 if (VA.isRegLoc()) {
2086 EVT RegVT = VA.getLocVT();
2089 if (VA.needsCustom()) {
2090 // f64 and vector types are split up into multiple registers or
2091 // combinations of registers and stack slots.
2092 if (VA.getLocVT() == MVT::v2f64) {
2093 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2095 VA = ArgLocs[++i]; // skip ahead to next loc
2097 if (VA.isMemLoc()) {
2098 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2099 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2100 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2101 PseudoSourceValue::getFixedStack(FI), 0,
2104 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2107 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2108 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2109 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2110 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2111 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2113 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2116 TargetRegisterClass *RC;
2118 if (RegVT == MVT::f32)
2119 RC = ARM::SPRRegisterClass;
2120 else if (RegVT == MVT::f64)
2121 RC = ARM::DPRRegisterClass;
2122 else if (RegVT == MVT::v2f64)
2123 RC = ARM::QPRRegisterClass;
2124 else if (RegVT == MVT::i32)
2125 RC = (AFI->isThumb1OnlyFunction() ?
2126 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2128 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2130 // Transform the arguments in physical registers into virtual ones.
2131 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2132 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2135 // If this is an 8 or 16-bit value, it is really passed promoted
2136 // to 32 bits. Insert an assert[sz]ext to capture this, then
2137 // truncate to the right size.
2138 switch (VA.getLocInfo()) {
2139 default: llvm_unreachable("Unknown loc info!");
2140 case CCValAssign::Full: break;
2141 case CCValAssign::BCvt:
2142 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2144 case CCValAssign::SExt:
2145 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2146 DAG.getValueType(VA.getValVT()));
2147 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2149 case CCValAssign::ZExt:
2150 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2151 DAG.getValueType(VA.getValVT()));
2152 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2156 InVals.push_back(ArgValue);
2158 } else { // VA.isRegLoc()
2161 assert(VA.isMemLoc());
2162 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2164 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2165 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2167 // Create load nodes to retrieve arguments from the stack.
2168 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2169 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2170 PseudoSourceValue::getFixedStack(FI), 0,
2177 static const unsigned GPRArgRegs[] = {
2178 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2181 unsigned NumGPRs = CCInfo.getFirstUnallocated
2182 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2184 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2185 unsigned VARegSize = (4 - NumGPRs) * 4;
2186 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2187 unsigned ArgOffset = CCInfo.getNextStackOffset();
2188 if (VARegSaveSize) {
2189 // If this function is vararg, store any remaining integer argument regs
2190 // to their spots on the stack so that they may be loaded by deferencing
2191 // the result of va_next.
2192 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2193 AFI->setVarArgsFrameIndex(
2194 MFI->CreateFixedObject(VARegSaveSize,
2195 ArgOffset + VARegSaveSize - VARegSize,
2197 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2200 SmallVector<SDValue, 4> MemOps;
2201 for (; NumGPRs < 4; ++NumGPRs) {
2202 TargetRegisterClass *RC;
2203 if (AFI->isThumb1OnlyFunction())
2204 RC = ARM::tGPRRegisterClass;
2206 RC = ARM::GPRRegisterClass;
2208 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2209 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2211 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2212 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2213 0, false, false, 0);
2214 MemOps.push_back(Store);
2215 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2216 DAG.getConstant(4, getPointerTy()));
2218 if (!MemOps.empty())
2219 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2220 &MemOps[0], MemOps.size());
2222 // This will point to the next argument passed via stack.
2223 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2229 /// isFloatingPointZero - Return true if this is +0.0.
2230 static bool isFloatingPointZero(SDValue Op) {
2231 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2232 return CFP->getValueAPF().isPosZero();
2233 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2234 // Maybe this has already been legalized into the constant pool?
2235 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2236 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2237 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2238 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2239 return CFP->getValueAPF().isPosZero();
2245 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2246 /// the given operands.
2248 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2249 SDValue &ARMcc, SelectionDAG &DAG,
2250 DebugLoc dl) const {
2251 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2252 unsigned C = RHSC->getZExtValue();
2253 if (!isLegalICmpImmediate(C)) {
2254 // Constant does not fit, try adjusting it by one?
2259 if (isLegalICmpImmediate(C-1)) {
2260 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2261 RHS = DAG.getConstant(C-1, MVT::i32);
2266 if (C > 0 && isLegalICmpImmediate(C-1)) {
2267 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2268 RHS = DAG.getConstant(C-1, MVT::i32);
2273 if (isLegalICmpImmediate(C+1)) {
2274 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2275 RHS = DAG.getConstant(C+1, MVT::i32);
2280 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2281 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2282 RHS = DAG.getConstant(C+1, MVT::i32);
2289 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2290 ARMISD::NodeType CompareType;
2293 CompareType = ARMISD::CMP;
2298 CompareType = ARMISD::CMPZ;
2301 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2302 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2305 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2307 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2308 DebugLoc dl) const {
2310 if (!isFloatingPointZero(RHS))
2311 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2313 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2314 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2317 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2318 SDValue Cond = Op.getOperand(0);
2319 SDValue SelectTrue = Op.getOperand(1);
2320 SDValue SelectFalse = Op.getOperand(2);
2321 DebugLoc dl = Op.getDebugLoc();
2325 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2326 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2328 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2329 const ConstantSDNode *CMOVTrue =
2330 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2331 const ConstantSDNode *CMOVFalse =
2332 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2334 if (CMOVTrue && CMOVFalse) {
2335 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2336 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2340 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2342 False = SelectFalse;
2343 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2348 if (True.getNode() && False.getNode()) {
2349 EVT VT = Cond.getValueType();
2350 SDValue ARMcc = Cond.getOperand(2);
2351 SDValue CCR = Cond.getOperand(3);
2352 SDValue Cmp = Cond.getOperand(4);
2353 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2358 return DAG.getSelectCC(dl, Cond,
2359 DAG.getConstant(0, Cond.getValueType()),
2360 SelectTrue, SelectFalse, ISD::SETNE);
2363 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2364 EVT VT = Op.getValueType();
2365 SDValue LHS = Op.getOperand(0);
2366 SDValue RHS = Op.getOperand(1);
2367 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2368 SDValue TrueVal = Op.getOperand(2);
2369 SDValue FalseVal = Op.getOperand(3);
2370 DebugLoc dl = Op.getDebugLoc();
2372 if (LHS.getValueType() == MVT::i32) {
2374 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2375 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2376 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2379 ARMCC::CondCodes CondCode, CondCode2;
2380 FPCCToARMCC(CC, CondCode, CondCode2);
2382 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2383 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2384 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2385 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2387 if (CondCode2 != ARMCC::AL) {
2388 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2389 // FIXME: Needs another CMP because flag can have but one use.
2390 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2391 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2392 Result, TrueVal, ARMcc2, CCR, Cmp2);
2397 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2398 /// to morph to an integer compare sequence.
2399 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2400 const ARMSubtarget *Subtarget) {
2401 SDNode *N = Op.getNode();
2402 if (!N->hasOneUse())
2403 // Otherwise it requires moving the value from fp to integer registers.
2405 if (!N->getNumValues())
2407 EVT VT = Op.getValueType();
2408 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2409 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2410 // vmrs are very slow, e.g. cortex-a8.
2413 if (isFloatingPointZero(Op)) {
2417 return ISD::isNormalLoad(N);
2420 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2421 if (isFloatingPointZero(Op))
2422 return DAG.getConstant(0, MVT::i32);
2424 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2425 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2426 Ld->getChain(), Ld->getBasePtr(),
2427 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2428 Ld->isVolatile(), Ld->isNonTemporal(),
2429 Ld->getAlignment());
2431 llvm_unreachable("Unknown VFP cmp argument!");
2434 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2435 SDValue &RetVal1, SDValue &RetVal2) {
2436 if (isFloatingPointZero(Op)) {
2437 RetVal1 = DAG.getConstant(0, MVT::i32);
2438 RetVal2 = DAG.getConstant(0, MVT::i32);
2442 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2443 SDValue Ptr = Ld->getBasePtr();
2444 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2445 Ld->getChain(), Ptr,
2446 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2447 Ld->isVolatile(), Ld->isNonTemporal(),
2448 Ld->getAlignment());
2450 EVT PtrType = Ptr.getValueType();
2451 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2452 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2453 PtrType, Ptr, DAG.getConstant(4, PtrType));
2454 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2455 Ld->getChain(), NewPtr,
2456 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2457 Ld->isVolatile(), Ld->isNonTemporal(),
2462 llvm_unreachable("Unknown VFP cmp argument!");
2465 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2466 /// f32 and even f64 comparisons to integer ones.
2468 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2469 SDValue Chain = Op.getOperand(0);
2470 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2471 SDValue LHS = Op.getOperand(2);
2472 SDValue RHS = Op.getOperand(3);
2473 SDValue Dest = Op.getOperand(4);
2474 DebugLoc dl = Op.getDebugLoc();
2476 bool SeenZero = false;
2477 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2478 canChangeToInt(RHS, SeenZero, Subtarget) &&
2479 // If one of the operand is zero, it's safe to ignore the NaN case since
2480 // we only care about equality comparisons.
2481 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2482 // If unsafe fp math optimization is enabled and there are no othter uses of
2483 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2484 // to an integer comparison.
2485 if (CC == ISD::SETOEQ)
2487 else if (CC == ISD::SETUNE)
2491 if (LHS.getValueType() == MVT::f32) {
2492 LHS = bitcastf32Toi32(LHS, DAG);
2493 RHS = bitcastf32Toi32(RHS, DAG);
2494 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2495 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2496 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2497 Chain, Dest, ARMcc, CCR, Cmp);
2502 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2503 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2504 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2505 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2506 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2507 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2508 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2514 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2515 SDValue Chain = Op.getOperand(0);
2516 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2517 SDValue LHS = Op.getOperand(2);
2518 SDValue RHS = Op.getOperand(3);
2519 SDValue Dest = Op.getOperand(4);
2520 DebugLoc dl = Op.getDebugLoc();
2522 if (LHS.getValueType() == MVT::i32) {
2524 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2525 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2526 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2527 Chain, Dest, ARMcc, CCR, Cmp);
2530 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2533 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2534 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2535 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2536 if (Result.getNode())
2540 ARMCC::CondCodes CondCode, CondCode2;
2541 FPCCToARMCC(CC, CondCode, CondCode2);
2543 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2544 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2545 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2546 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2547 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2548 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2549 if (CondCode2 != ARMCC::AL) {
2550 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2551 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2552 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2557 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2558 SDValue Chain = Op.getOperand(0);
2559 SDValue Table = Op.getOperand(1);
2560 SDValue Index = Op.getOperand(2);
2561 DebugLoc dl = Op.getDebugLoc();
2563 EVT PTy = getPointerTy();
2564 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2565 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2566 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2567 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2568 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2569 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2570 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2571 if (Subtarget->isThumb2()) {
2572 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2573 // which does another jump to the destination. This also makes it easier
2574 // to translate it to TBB / TBH later.
2575 // FIXME: This might not work if the function is extremely large.
2576 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2577 Addr, Op.getOperand(2), JTI, UId);
2579 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2580 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2581 PseudoSourceValue::getJumpTable(), 0,
2583 Chain = Addr.getValue(1);
2584 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2585 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2587 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2588 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2589 Chain = Addr.getValue(1);
2590 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2594 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2595 DebugLoc dl = Op.getDebugLoc();
2598 switch (Op.getOpcode()) {
2600 assert(0 && "Invalid opcode!");
2601 case ISD::FP_TO_SINT:
2602 Opc = ARMISD::FTOSI;
2604 case ISD::FP_TO_UINT:
2605 Opc = ARMISD::FTOUI;
2608 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2609 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2612 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2613 EVT VT = Op.getValueType();
2614 DebugLoc dl = Op.getDebugLoc();
2617 switch (Op.getOpcode()) {
2619 assert(0 && "Invalid opcode!");
2620 case ISD::SINT_TO_FP:
2621 Opc = ARMISD::SITOF;
2623 case ISD::UINT_TO_FP:
2624 Opc = ARMISD::UITOF;
2628 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2629 return DAG.getNode(Opc, dl, VT, Op);
2632 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2633 // Implement fcopysign with a fabs and a conditional fneg.
2634 SDValue Tmp0 = Op.getOperand(0);
2635 SDValue Tmp1 = Op.getOperand(1);
2636 DebugLoc dl = Op.getDebugLoc();
2637 EVT VT = Op.getValueType();
2638 EVT SrcVT = Tmp1.getValueType();
2639 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2640 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2641 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2642 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2643 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2644 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2647 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2648 MachineFunction &MF = DAG.getMachineFunction();
2649 MachineFrameInfo *MFI = MF.getFrameInfo();
2650 MFI->setReturnAddressIsTaken(true);
2652 EVT VT = Op.getValueType();
2653 DebugLoc dl = Op.getDebugLoc();
2654 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2656 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2657 SDValue Offset = DAG.getConstant(4, MVT::i32);
2658 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2659 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2660 NULL, 0, false, false, 0);
2663 // Return LR, which contains the return address. Mark it an implicit live-in.
2664 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2665 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2668 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2669 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2670 MFI->setFrameAddressIsTaken(true);
2672 EVT VT = Op.getValueType();
2673 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2674 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2675 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2676 ? ARM::R7 : ARM::R11;
2677 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2679 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2684 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2685 /// expand a bit convert where either the source or destination type is i64 to
2686 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2687 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2688 /// vectors), since the legalizer won't know what to do with that.
2689 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2690 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2691 DebugLoc dl = N->getDebugLoc();
2692 SDValue Op = N->getOperand(0);
2694 // This function is only supposed to be called for i64 types, either as the
2695 // source or destination of the bit convert.
2696 EVT SrcVT = Op.getValueType();
2697 EVT DstVT = N->getValueType(0);
2698 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2699 "ExpandBIT_CONVERT called for non-i64 type");
2701 // Turn i64->f64 into VMOVDRR.
2702 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2703 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2704 DAG.getConstant(0, MVT::i32));
2705 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2706 DAG.getConstant(1, MVT::i32));
2707 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2708 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2711 // Turn f64->i64 into VMOVRRD.
2712 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2713 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2714 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2715 // Merge the pieces into a single i64 value.
2716 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2722 /// getZeroVector - Returns a vector of specified type with all zero elements.
2723 /// Zero vectors are used to represent vector negation and in those cases
2724 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2725 /// not support i64 elements, so sometimes the zero vectors will need to be
2726 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2728 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2729 assert(VT.isVector() && "Expected a vector type");
2730 // The canonical modified immediate encoding of a zero vector is....0!
2731 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2732 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2733 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2734 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2737 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2738 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2739 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2740 SelectionDAG &DAG) const {
2741 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2742 EVT VT = Op.getValueType();
2743 unsigned VTBits = VT.getSizeInBits();
2744 DebugLoc dl = Op.getDebugLoc();
2745 SDValue ShOpLo = Op.getOperand(0);
2746 SDValue ShOpHi = Op.getOperand(1);
2747 SDValue ShAmt = Op.getOperand(2);
2749 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2751 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2753 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2754 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2755 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2756 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2757 DAG.getConstant(VTBits, MVT::i32));
2758 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2759 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2760 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2762 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2763 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2765 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2766 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2769 SDValue Ops[2] = { Lo, Hi };
2770 return DAG.getMergeValues(Ops, 2, dl);
2773 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2774 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2775 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2776 SelectionDAG &DAG) const {
2777 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2778 EVT VT = Op.getValueType();
2779 unsigned VTBits = VT.getSizeInBits();
2780 DebugLoc dl = Op.getDebugLoc();
2781 SDValue ShOpLo = Op.getOperand(0);
2782 SDValue ShOpHi = Op.getOperand(1);
2783 SDValue ShAmt = Op.getOperand(2);
2786 assert(Op.getOpcode() == ISD::SHL_PARTS);
2787 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2788 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2789 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2790 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2791 DAG.getConstant(VTBits, MVT::i32));
2792 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2793 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2795 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2796 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2797 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2799 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2800 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2803 SDValue Ops[2] = { Lo, Hi };
2804 return DAG.getMergeValues(Ops, 2, dl);
2807 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2808 SelectionDAG &DAG) const {
2809 // The rounding mode is in bits 23:22 of the FPSCR.
2810 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2811 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2812 // so that the shift + and get folded into a bitfield extract.
2813 DebugLoc dl = Op.getDebugLoc();
2814 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2815 DAG.getConstant(Intrinsic::arm_get_fpscr,
2817 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2818 DAG.getConstant(1U << 22, MVT::i32));
2819 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2820 DAG.getConstant(22, MVT::i32));
2821 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2822 DAG.getConstant(3, MVT::i32));
2825 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2826 const ARMSubtarget *ST) {
2827 EVT VT = N->getValueType(0);
2828 DebugLoc dl = N->getDebugLoc();
2830 if (!ST->hasV6T2Ops())
2833 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2834 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2837 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2838 const ARMSubtarget *ST) {
2839 EVT VT = N->getValueType(0);
2840 DebugLoc dl = N->getDebugLoc();
2842 // Lower vector shifts on NEON to use VSHL.
2843 if (VT.isVector()) {
2844 assert(ST->hasNEON() && "unexpected vector shift");
2846 // Left shifts translate directly to the vshiftu intrinsic.
2847 if (N->getOpcode() == ISD::SHL)
2848 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2849 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2850 N->getOperand(0), N->getOperand(1));
2852 assert((N->getOpcode() == ISD::SRA ||
2853 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2855 // NEON uses the same intrinsics for both left and right shifts. For
2856 // right shifts, the shift amounts are negative, so negate the vector of
2858 EVT ShiftVT = N->getOperand(1).getValueType();
2859 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2860 getZeroVector(ShiftVT, DAG, dl),
2862 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2863 Intrinsic::arm_neon_vshifts :
2864 Intrinsic::arm_neon_vshiftu);
2865 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2866 DAG.getConstant(vshiftInt, MVT::i32),
2867 N->getOperand(0), NegatedCount);
2870 // We can get here for a node like i32 = ISD::SHL i32, i64
2874 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2875 "Unknown shift to lower!");
2877 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2878 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2879 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2882 // If we are in thumb mode, we don't have RRX.
2883 if (ST->isThumb1Only()) return SDValue();
2885 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2886 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2887 DAG.getConstant(0, MVT::i32));
2888 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2889 DAG.getConstant(1, MVT::i32));
2891 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2892 // captures the result into a carry flag.
2893 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2894 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2896 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2897 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2899 // Merge the pieces into a single i64 value.
2900 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2903 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2904 SDValue TmpOp0, TmpOp1;
2905 bool Invert = false;
2909 SDValue Op0 = Op.getOperand(0);
2910 SDValue Op1 = Op.getOperand(1);
2911 SDValue CC = Op.getOperand(2);
2912 EVT VT = Op.getValueType();
2913 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2914 DebugLoc dl = Op.getDebugLoc();
2916 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2917 switch (SetCCOpcode) {
2918 default: llvm_unreachable("Illegal FP comparison"); break;
2920 case ISD::SETNE: Invert = true; // Fallthrough
2922 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2924 case ISD::SETLT: Swap = true; // Fallthrough
2926 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2928 case ISD::SETLE: Swap = true; // Fallthrough
2930 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2931 case ISD::SETUGE: Swap = true; // Fallthrough
2932 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2933 case ISD::SETUGT: Swap = true; // Fallthrough
2934 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2935 case ISD::SETUEQ: Invert = true; // Fallthrough
2937 // Expand this to (OLT | OGT).
2941 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2942 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2944 case ISD::SETUO: Invert = true; // Fallthrough
2946 // Expand this to (OLT | OGE).
2950 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2951 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2955 // Integer comparisons.
2956 switch (SetCCOpcode) {
2957 default: llvm_unreachable("Illegal integer comparison"); break;
2958 case ISD::SETNE: Invert = true;
2959 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2960 case ISD::SETLT: Swap = true;
2961 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2962 case ISD::SETLE: Swap = true;
2963 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2964 case ISD::SETULT: Swap = true;
2965 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2966 case ISD::SETULE: Swap = true;
2967 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2970 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2971 if (Opc == ARMISD::VCEQ) {
2974 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2976 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2979 // Ignore bitconvert.
2980 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2981 AndOp = AndOp.getOperand(0);
2983 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2985 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2986 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2993 std::swap(Op0, Op1);
2995 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2998 Result = DAG.getNOT(dl, Result, VT);
3003 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3004 /// valid vector constant for a NEON instruction with a "modified immediate"
3005 /// operand (e.g., VMOV). If so, return the encoded value.
3006 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3007 unsigned SplatBitSize, SelectionDAG &DAG,
3008 EVT &VT, bool is128Bits, bool isVMOV) {
3009 unsigned OpCmode, Imm;
3011 // SplatBitSize is set to the smallest size that splats the vector, so a
3012 // zero vector will always have SplatBitSize == 8. However, NEON modified
3013 // immediate instructions others than VMOV do not support the 8-bit encoding
3014 // of a zero vector, and the default encoding of zero is supposed to be the
3019 switch (SplatBitSize) {
3023 // Any 1-byte value is OK. Op=0, Cmode=1110.
3024 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3027 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3031 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3032 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3033 if ((SplatBits & ~0xff) == 0) {
3034 // Value = 0x00nn: Op=x, Cmode=100x.
3039 if ((SplatBits & ~0xff00) == 0) {
3040 // Value = 0xnn00: Op=x, Cmode=101x.
3042 Imm = SplatBits >> 8;
3048 // NEON's 32-bit VMOV supports splat values where:
3049 // * only one byte is nonzero, or
3050 // * the least significant byte is 0xff and the second byte is nonzero, or
3051 // * the least significant 2 bytes are 0xff and the third is nonzero.
3052 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3053 if ((SplatBits & ~0xff) == 0) {
3054 // Value = 0x000000nn: Op=x, Cmode=000x.
3059 if ((SplatBits & ~0xff00) == 0) {
3060 // Value = 0x0000nn00: Op=x, Cmode=001x.
3062 Imm = SplatBits >> 8;
3065 if ((SplatBits & ~0xff0000) == 0) {
3066 // Value = 0x00nn0000: Op=x, Cmode=010x.
3068 Imm = SplatBits >> 16;
3071 if ((SplatBits & ~0xff000000) == 0) {
3072 // Value = 0xnn000000: Op=x, Cmode=011x.
3074 Imm = SplatBits >> 24;
3078 if ((SplatBits & ~0xffff) == 0 &&
3079 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3080 // Value = 0x0000nnff: Op=x, Cmode=1100.
3082 Imm = SplatBits >> 8;
3087 if ((SplatBits & ~0xffffff) == 0 &&
3088 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3089 // Value = 0x00nnffff: Op=x, Cmode=1101.
3091 Imm = SplatBits >> 16;
3092 SplatBits |= 0xffff;
3096 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3097 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3098 // VMOV.I32. A (very) minor optimization would be to replicate the value
3099 // and fall through here to test for a valid 64-bit splat. But, then the
3100 // caller would also need to check and handle the change in size.
3106 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3107 uint64_t BitMask = 0xff;
3109 unsigned ImmMask = 1;
3111 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3112 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3115 } else if ((SplatBits & BitMask) != 0) {
3121 // Op=1, Cmode=1110.
3124 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3129 llvm_unreachable("unexpected size for isNEONModifiedImm");
3133 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3134 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3137 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3138 bool &ReverseVEXT, unsigned &Imm) {
3139 unsigned NumElts = VT.getVectorNumElements();
3140 ReverseVEXT = false;
3143 // If this is a VEXT shuffle, the immediate value is the index of the first
3144 // element. The other shuffle indices must be the successive elements after
3146 unsigned ExpectedElt = Imm;
3147 for (unsigned i = 1; i < NumElts; ++i) {
3148 // Increment the expected index. If it wraps around, it may still be
3149 // a VEXT but the source vectors must be swapped.
3151 if (ExpectedElt == NumElts * 2) {
3156 if (ExpectedElt != static_cast<unsigned>(M[i]))
3160 // Adjust the index value if the source operands will be swapped.
3167 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3168 /// instruction with the specified blocksize. (The order of the elements
3169 /// within each block of the vector is reversed.)
3170 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3171 unsigned BlockSize) {
3172 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3173 "Only possible block sizes for VREV are: 16, 32, 64");
3175 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3179 unsigned NumElts = VT.getVectorNumElements();
3180 unsigned BlockElts = M[0] + 1;
3182 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3185 for (unsigned i = 0; i < NumElts; ++i) {
3186 if ((unsigned) M[i] !=
3187 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3194 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3195 unsigned &WhichResult) {
3196 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3200 unsigned NumElts = VT.getVectorNumElements();
3201 WhichResult = (M[0] == 0 ? 0 : 1);
3202 for (unsigned i = 0; i < NumElts; i += 2) {
3203 if ((unsigned) M[i] != i + WhichResult ||
3204 (unsigned) M[i+1] != i + NumElts + WhichResult)
3210 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3211 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3212 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3213 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3214 unsigned &WhichResult) {
3215 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3219 unsigned NumElts = VT.getVectorNumElements();
3220 WhichResult = (M[0] == 0 ? 0 : 1);
3221 for (unsigned i = 0; i < NumElts; i += 2) {
3222 if ((unsigned) M[i] != i + WhichResult ||
3223 (unsigned) M[i+1] != i + WhichResult)
3229 static bool isVUZPMask(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 for (unsigned i = 0; i != NumElts; ++i) {
3238 if ((unsigned) M[i] != 2 * i + WhichResult)
3242 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3243 if (VT.is64BitVector() && EltSz == 32)
3249 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3250 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3251 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3252 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3253 unsigned &WhichResult) {
3254 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3258 unsigned Half = VT.getVectorNumElements() / 2;
3259 WhichResult = (M[0] == 0 ? 0 : 1);
3260 for (unsigned j = 0; j != 2; ++j) {
3261 unsigned Idx = WhichResult;
3262 for (unsigned i = 0; i != Half; ++i) {
3263 if ((unsigned) M[i + j * Half] != Idx)
3269 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3270 if (VT.is64BitVector() && EltSz == 32)
3276 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3277 unsigned &WhichResult) {
3278 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3282 unsigned NumElts = VT.getVectorNumElements();
3283 WhichResult = (M[0] == 0 ? 0 : 1);
3284 unsigned Idx = WhichResult * NumElts / 2;
3285 for (unsigned i = 0; i != NumElts; i += 2) {
3286 if ((unsigned) M[i] != Idx ||
3287 (unsigned) M[i+1] != Idx + NumElts)
3292 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3293 if (VT.is64BitVector() && EltSz == 32)
3299 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3300 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3301 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3302 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3303 unsigned &WhichResult) {
3304 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3308 unsigned NumElts = VT.getVectorNumElements();
3309 WhichResult = (M[0] == 0 ? 0 : 1);
3310 unsigned Idx = WhichResult * NumElts / 2;
3311 for (unsigned i = 0; i != NumElts; i += 2) {
3312 if ((unsigned) M[i] != Idx ||
3313 (unsigned) M[i+1] != Idx)
3318 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3319 if (VT.is64BitVector() && EltSz == 32)
3325 // If N is an integer constant that can be moved into a register in one
3326 // instruction, return an SDValue of such a constant (will become a MOV
3327 // instruction). Otherwise return null.
3328 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3329 const ARMSubtarget *ST, DebugLoc dl) {
3331 if (!isa<ConstantSDNode>(N))
3333 Val = cast<ConstantSDNode>(N)->getZExtValue();
3335 if (ST->isThumb1Only()) {
3336 if (Val <= 255 || ~Val <= 255)
3337 return DAG.getConstant(Val, MVT::i32);
3339 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3340 return DAG.getConstant(Val, MVT::i32);
3345 // If this is a case we can't handle, return null and let the default
3346 // expansion code take care of it.
3347 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3348 const ARMSubtarget *ST) {
3349 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3350 DebugLoc dl = Op.getDebugLoc();
3351 EVT VT = Op.getValueType();
3353 APInt SplatBits, SplatUndef;
3354 unsigned SplatBitSize;
3356 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3357 if (SplatBitSize <= 64) {
3358 // Check if an immediate VMOV works.
3360 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3361 SplatUndef.getZExtValue(), SplatBitSize,
3362 DAG, VmovVT, VT.is128BitVector(), true);
3363 if (Val.getNode()) {
3364 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3365 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3368 // Try an immediate VMVN.
3369 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3370 ((1LL << SplatBitSize) - 1));
3371 Val = isNEONModifiedImm(NegatedImm,
3372 SplatUndef.getZExtValue(), SplatBitSize,
3373 DAG, VmovVT, VT.is128BitVector(), false);
3374 if (Val.getNode()) {
3375 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3376 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3381 // Scan through the operands to see if only one value is used.
3382 unsigned NumElts = VT.getVectorNumElements();
3383 bool isOnlyLowElement = true;
3384 bool usesOnlyOneValue = true;
3385 bool isConstant = true;
3387 for (unsigned i = 0; i < NumElts; ++i) {
3388 SDValue V = Op.getOperand(i);
3389 if (V.getOpcode() == ISD::UNDEF)
3392 isOnlyLowElement = false;
3393 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3396 if (!Value.getNode())
3398 else if (V != Value)
3399 usesOnlyOneValue = false;
3402 if (!Value.getNode())
3403 return DAG.getUNDEF(VT);
3405 if (isOnlyLowElement)
3406 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3408 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3410 if (EnableARMVDUPsplat) {
3411 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3412 // i32 and try again.
3413 if (usesOnlyOneValue && EltSize <= 32) {
3415 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3416 if (VT.getVectorElementType().isFloatingPoint()) {
3417 SmallVector<SDValue, 8> Ops;
3418 for (unsigned i = 0; i < NumElts; ++i)
3419 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3421 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3423 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3424 LowerBUILD_VECTOR(Val, DAG, ST));
3426 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3428 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3432 // If all elements are constants and the case above didn't get hit, fall back
3433 // to the default expansion, which will generate a load from the constant
3438 if (!EnableARMVDUPsplat) {
3439 // Use VDUP for non-constant splats.
3440 if (usesOnlyOneValue && EltSize <= 32)
3441 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3444 // Vectors with 32- or 64-bit elements can be built by directly assigning
3445 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3446 // will be legalized.
3447 if (EltSize >= 32) {
3448 // Do the expansion with floating-point types, since that is what the VFP
3449 // registers are defined to use, and since i64 is not legal.
3450 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3451 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3452 SmallVector<SDValue, 8> Ops;
3453 for (unsigned i = 0; i < NumElts; ++i)
3454 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3455 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3456 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3462 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3463 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3464 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3465 /// are assumed to be legal.
3467 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3469 if (VT.getVectorNumElements() == 4 &&
3470 (VT.is128BitVector() || VT.is64BitVector())) {
3471 unsigned PFIndexes[4];
3472 for (unsigned i = 0; i != 4; ++i) {
3476 PFIndexes[i] = M[i];
3479 // Compute the index in the perfect shuffle table.
3480 unsigned PFTableIndex =
3481 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3482 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3483 unsigned Cost = (PFEntry >> 30);
3490 unsigned Imm, WhichResult;
3492 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3493 return (EltSize >= 32 ||
3494 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3495 isVREVMask(M, VT, 64) ||
3496 isVREVMask(M, VT, 32) ||
3497 isVREVMask(M, VT, 16) ||
3498 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3499 isVTRNMask(M, VT, WhichResult) ||
3500 isVUZPMask(M, VT, WhichResult) ||
3501 isVZIPMask(M, VT, WhichResult) ||
3502 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3503 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3504 isVZIP_v_undef_Mask(M, VT, WhichResult));
3507 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3508 /// the specified operations to build the shuffle.
3509 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3510 SDValue RHS, SelectionDAG &DAG,
3512 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3513 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3514 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3517 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3526 OP_VUZPL, // VUZP, left result
3527 OP_VUZPR, // VUZP, right result
3528 OP_VZIPL, // VZIP, left result
3529 OP_VZIPR, // VZIP, right result
3530 OP_VTRNL, // VTRN, left result
3531 OP_VTRNR // VTRN, right result
3534 if (OpNum == OP_COPY) {
3535 if (LHSID == (1*9+2)*9+3) return LHS;
3536 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3540 SDValue OpLHS, OpRHS;
3541 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3542 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3543 EVT VT = OpLHS.getValueType();
3546 default: llvm_unreachable("Unknown shuffle opcode!");
3548 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3553 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3554 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3558 return DAG.getNode(ARMISD::VEXT, dl, VT,
3560 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3563 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3564 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3567 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3568 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3571 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3572 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3576 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3577 SDValue V1 = Op.getOperand(0);
3578 SDValue V2 = Op.getOperand(1);
3579 DebugLoc dl = Op.getDebugLoc();
3580 EVT VT = Op.getValueType();
3581 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3582 SmallVector<int, 8> ShuffleMask;
3584 // Convert shuffles that are directly supported on NEON to target-specific
3585 // DAG nodes, instead of keeping them as shuffles and matching them again
3586 // during code selection. This is more efficient and avoids the possibility
3587 // of inconsistencies between legalization and selection.
3588 // FIXME: floating-point vectors should be canonicalized to integer vectors
3589 // of the same time so that they get CSEd properly.
3590 SVN->getMask(ShuffleMask);
3592 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3593 if (EltSize <= 32) {
3594 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3595 int Lane = SVN->getSplatIndex();
3596 // If this is undef splat, generate it via "just" vdup, if possible.
3597 if (Lane == -1) Lane = 0;
3599 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3600 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3602 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3603 DAG.getConstant(Lane, MVT::i32));
3608 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3611 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3612 DAG.getConstant(Imm, MVT::i32));
3615 if (isVREVMask(ShuffleMask, VT, 64))
3616 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3617 if (isVREVMask(ShuffleMask, VT, 32))
3618 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3619 if (isVREVMask(ShuffleMask, VT, 16))
3620 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3622 // Check for Neon shuffles that modify both input vectors in place.
3623 // If both results are used, i.e., if there are two shuffles with the same
3624 // source operands and with masks corresponding to both results of one of
3625 // these operations, DAG memoization will ensure that a single node is
3626 // used for both shuffles.
3627 unsigned WhichResult;
3628 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3629 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3630 V1, V2).getValue(WhichResult);
3631 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3632 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3633 V1, V2).getValue(WhichResult);
3634 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3635 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3636 V1, V2).getValue(WhichResult);
3638 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3639 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3640 V1, V1).getValue(WhichResult);
3641 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3642 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3643 V1, V1).getValue(WhichResult);
3644 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3645 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3646 V1, V1).getValue(WhichResult);
3649 // If the shuffle is not directly supported and it has 4 elements, use
3650 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3651 unsigned NumElts = VT.getVectorNumElements();
3653 unsigned PFIndexes[4];
3654 for (unsigned i = 0; i != 4; ++i) {
3655 if (ShuffleMask[i] < 0)
3658 PFIndexes[i] = ShuffleMask[i];
3661 // Compute the index in the perfect shuffle table.
3662 unsigned PFTableIndex =
3663 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3664 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3665 unsigned Cost = (PFEntry >> 30);
3668 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3671 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3672 if (EltSize >= 32) {
3673 // Do the expansion with floating-point types, since that is what the VFP
3674 // registers are defined to use, and since i64 is not legal.
3675 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3676 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3677 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3678 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3679 SmallVector<SDValue, 8> Ops;
3680 for (unsigned i = 0; i < NumElts; ++i) {
3681 if (ShuffleMask[i] < 0)
3682 Ops.push_back(DAG.getUNDEF(EltVT));
3684 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3685 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3686 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3689 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3690 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3696 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3697 EVT VT = Op.getValueType();
3698 DebugLoc dl = Op.getDebugLoc();
3699 SDValue Vec = Op.getOperand(0);
3700 SDValue Lane = Op.getOperand(1);
3701 assert(VT == MVT::i32 &&
3702 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3703 "unexpected type for custom-lowering vector extract");
3704 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3707 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3708 // The only time a CONCAT_VECTORS operation can have legal types is when
3709 // two 64-bit vectors are concatenated to a 128-bit vector.
3710 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3711 "unexpected CONCAT_VECTORS");
3712 DebugLoc dl = Op.getDebugLoc();
3713 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3714 SDValue Op0 = Op.getOperand(0);
3715 SDValue Op1 = Op.getOperand(1);
3716 if (Op0.getOpcode() != ISD::UNDEF)
3717 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3718 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3719 DAG.getIntPtrConstant(0));
3720 if (Op1.getOpcode() != ISD::UNDEF)
3721 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3722 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3723 DAG.getIntPtrConstant(1));
3724 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3727 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3728 switch (Op.getOpcode()) {
3729 default: llvm_unreachable("Don't know how to custom lower this!");
3730 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3731 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3732 case ISD::GlobalAddress:
3733 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3734 LowerGlobalAddressELF(Op, DAG);
3735 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3736 case ISD::SELECT: return LowerSELECT(Op, DAG);
3737 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3738 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3739 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3740 case ISD::VASTART: return LowerVASTART(Op, DAG);
3741 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3742 case ISD::SINT_TO_FP:
3743 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3744 case ISD::FP_TO_SINT:
3745 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3746 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3747 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3748 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3749 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3750 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3751 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3752 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3754 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3757 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3758 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3759 case ISD::SRL_PARTS:
3760 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3761 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3762 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3763 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3764 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3765 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3766 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3767 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3772 /// ReplaceNodeResults - Replace the results of node with an illegal result
3773 /// type with new values built out of custom code.
3774 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3775 SmallVectorImpl<SDValue>&Results,
3776 SelectionDAG &DAG) const {
3778 switch (N->getOpcode()) {
3780 llvm_unreachable("Don't know how to custom expand this!");
3782 case ISD::BIT_CONVERT:
3783 Res = ExpandBIT_CONVERT(N, DAG);
3787 Res = LowerShift(N, DAG, Subtarget);
3791 Results.push_back(Res);
3794 //===----------------------------------------------------------------------===//
3795 // ARM Scheduler Hooks
3796 //===----------------------------------------------------------------------===//
3799 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3800 MachineBasicBlock *BB,
3801 unsigned Size) const {
3802 unsigned dest = MI->getOperand(0).getReg();
3803 unsigned ptr = MI->getOperand(1).getReg();
3804 unsigned oldval = MI->getOperand(2).getReg();
3805 unsigned newval = MI->getOperand(3).getReg();
3806 unsigned scratch = BB->getParent()->getRegInfo()
3807 .createVirtualRegister(ARM::GPRRegisterClass);
3808 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3809 DebugLoc dl = MI->getDebugLoc();
3810 bool isThumb2 = Subtarget->isThumb2();
3812 unsigned ldrOpc, strOpc;
3814 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3816 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3817 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3820 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3821 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3824 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3825 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3829 MachineFunction *MF = BB->getParent();
3830 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3831 MachineFunction::iterator It = BB;
3832 ++It; // insert the new blocks after the current block
3834 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3835 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3836 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3837 MF->insert(It, loop1MBB);
3838 MF->insert(It, loop2MBB);
3839 MF->insert(It, exitMBB);
3841 // Transfer the remainder of BB and its successor edges to exitMBB.
3842 exitMBB->splice(exitMBB->begin(), BB,
3843 llvm::next(MachineBasicBlock::iterator(MI)),
3845 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3849 // fallthrough --> loop1MBB
3850 BB->addSuccessor(loop1MBB);
3853 // ldrex dest, [ptr]
3857 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3858 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3859 .addReg(dest).addReg(oldval));
3860 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3861 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3862 BB->addSuccessor(loop2MBB);
3863 BB->addSuccessor(exitMBB);
3866 // strex scratch, newval, [ptr]
3870 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3872 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3873 .addReg(scratch).addImm(0));
3874 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3875 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3876 BB->addSuccessor(loop1MBB);
3877 BB->addSuccessor(exitMBB);
3883 MI->eraseFromParent(); // The instruction is gone now.
3889 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3890 unsigned Size, unsigned BinOpcode) const {
3891 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3892 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3894 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3895 MachineFunction *MF = BB->getParent();
3896 MachineFunction::iterator It = BB;
3899 unsigned dest = MI->getOperand(0).getReg();
3900 unsigned ptr = MI->getOperand(1).getReg();
3901 unsigned incr = MI->getOperand(2).getReg();
3902 DebugLoc dl = MI->getDebugLoc();
3904 bool isThumb2 = Subtarget->isThumb2();
3905 unsigned ldrOpc, strOpc;
3907 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3909 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3910 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3913 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3914 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3917 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3918 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3922 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3923 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3924 MF->insert(It, loopMBB);
3925 MF->insert(It, exitMBB);
3927 // Transfer the remainder of BB and its successor edges to exitMBB.
3928 exitMBB->splice(exitMBB->begin(), BB,
3929 llvm::next(MachineBasicBlock::iterator(MI)),
3931 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3933 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3934 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3935 unsigned scratch2 = (!BinOpcode) ? incr :
3936 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3940 // fallthrough --> loopMBB
3941 BB->addSuccessor(loopMBB);
3945 // <binop> scratch2, dest, incr
3946 // strex scratch, scratch2, ptr
3949 // fallthrough --> exitMBB
3951 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3953 // operand order needs to go the other way for NAND
3954 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3955 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3956 addReg(incr).addReg(dest)).addReg(0);
3958 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3959 addReg(dest).addReg(incr)).addReg(0);
3962 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3964 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3965 .addReg(scratch).addImm(0));
3966 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3967 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3969 BB->addSuccessor(loopMBB);
3970 BB->addSuccessor(exitMBB);
3976 MI->eraseFromParent(); // The instruction is gone now.
3982 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3983 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3984 E = MBB->succ_end(); I != E; ++I)
3987 llvm_unreachable("Expecting a BB with two successors!");
3991 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3992 MachineBasicBlock *BB) const {
3993 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3994 DebugLoc dl = MI->getDebugLoc();
3995 bool isThumb2 = Subtarget->isThumb2();
3996 switch (MI->getOpcode()) {
3999 llvm_unreachable("Unexpected instr type to insert");
4001 case ARM::ATOMIC_LOAD_ADD_I8:
4002 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4003 case ARM::ATOMIC_LOAD_ADD_I16:
4004 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4005 case ARM::ATOMIC_LOAD_ADD_I32:
4006 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
4008 case ARM::ATOMIC_LOAD_AND_I8:
4009 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4010 case ARM::ATOMIC_LOAD_AND_I16:
4011 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4012 case ARM::ATOMIC_LOAD_AND_I32:
4013 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
4015 case ARM::ATOMIC_LOAD_OR_I8:
4016 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4017 case ARM::ATOMIC_LOAD_OR_I16:
4018 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4019 case ARM::ATOMIC_LOAD_OR_I32:
4020 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
4022 case ARM::ATOMIC_LOAD_XOR_I8:
4023 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4024 case ARM::ATOMIC_LOAD_XOR_I16:
4025 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4026 case ARM::ATOMIC_LOAD_XOR_I32:
4027 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
4029 case ARM::ATOMIC_LOAD_NAND_I8:
4030 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4031 case ARM::ATOMIC_LOAD_NAND_I16:
4032 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4033 case ARM::ATOMIC_LOAD_NAND_I32:
4034 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
4036 case ARM::ATOMIC_LOAD_SUB_I8:
4037 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4038 case ARM::ATOMIC_LOAD_SUB_I16:
4039 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4040 case ARM::ATOMIC_LOAD_SUB_I32:
4041 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
4043 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
4044 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
4045 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
4047 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
4048 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
4049 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
4051 case ARM::tMOVCCr_pseudo: {
4052 // To "insert" a SELECT_CC instruction, we actually have to insert the
4053 // diamond control-flow pattern. The incoming instruction knows the
4054 // destination vreg to set, the condition code register to branch on, the
4055 // true/false values to select between, and a branch opcode to use.
4056 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4057 MachineFunction::iterator It = BB;
4063 // cmpTY ccX, r1, r2
4065 // fallthrough --> copy0MBB
4066 MachineBasicBlock *thisMBB = BB;
4067 MachineFunction *F = BB->getParent();
4068 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4069 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4070 F->insert(It, copy0MBB);
4071 F->insert(It, sinkMBB);
4073 // Transfer the remainder of BB and its successor edges to sinkMBB.
4074 sinkMBB->splice(sinkMBB->begin(), BB,
4075 llvm::next(MachineBasicBlock::iterator(MI)),
4077 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4079 BB->addSuccessor(copy0MBB);
4080 BB->addSuccessor(sinkMBB);
4082 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4083 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4086 // %FalseValue = ...
4087 // # fallthrough to sinkMBB
4090 // Update machine-CFG edges
4091 BB->addSuccessor(sinkMBB);
4094 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4097 BuildMI(*BB, BB->begin(), dl,
4098 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4099 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4100 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4102 MI->eraseFromParent(); // The pseudo instruction is gone now.
4107 case ARM::BCCZi64: {
4108 // Compare both parts that make up the double comparison separately for
4110 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4112 unsigned LHS1 = MI->getOperand(1).getReg();
4113 unsigned LHS2 = MI->getOperand(2).getReg();
4115 AddDefaultPred(BuildMI(BB, dl,
4116 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4117 .addReg(LHS1).addImm(0));
4118 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4119 .addReg(LHS2).addImm(0)
4120 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4122 unsigned RHS1 = MI->getOperand(3).getReg();
4123 unsigned RHS2 = MI->getOperand(4).getReg();
4124 AddDefaultPred(BuildMI(BB, dl,
4125 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4126 .addReg(LHS1).addReg(RHS1));
4127 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4128 .addReg(LHS2).addReg(RHS2)
4129 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4132 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4133 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4134 if (MI->getOperand(0).getImm() == ARMCC::NE)
4135 std::swap(destMBB, exitMBB);
4137 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4138 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4139 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4142 MI->eraseFromParent(); // The pseudo instruction is gone now.
4148 //===----------------------------------------------------------------------===//
4149 // ARM Optimization Hooks
4150 //===----------------------------------------------------------------------===//
4153 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4154 TargetLowering::DAGCombinerInfo &DCI) {
4155 SelectionDAG &DAG = DCI.DAG;
4156 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4157 EVT VT = N->getValueType(0);
4158 unsigned Opc = N->getOpcode();
4159 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4160 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4161 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4162 ISD::CondCode CC = ISD::SETCC_INVALID;
4165 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4167 SDValue CCOp = Slct.getOperand(0);
4168 if (CCOp.getOpcode() == ISD::SETCC)
4169 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4172 bool DoXform = false;
4174 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4177 if (LHS.getOpcode() == ISD::Constant &&
4178 cast<ConstantSDNode>(LHS)->isNullValue()) {
4180 } else if (CC != ISD::SETCC_INVALID &&
4181 RHS.getOpcode() == ISD::Constant &&
4182 cast<ConstantSDNode>(RHS)->isNullValue()) {
4183 std::swap(LHS, RHS);
4184 SDValue Op0 = Slct.getOperand(0);
4185 EVT OpVT = isSlctCC ? Op0.getValueType() :
4186 Op0.getOperand(0).getValueType();
4187 bool isInt = OpVT.isInteger();
4188 CC = ISD::getSetCCInverse(CC, isInt);
4190 if (!TLI.isCondCodeLegal(CC, OpVT))
4191 return SDValue(); // Inverse operator isn't legal.
4198 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4200 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4201 Slct.getOperand(0), Slct.getOperand(1), CC);
4202 SDValue CCOp = Slct.getOperand(0);
4204 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4205 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4206 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4207 CCOp, OtherOp, Result);
4212 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4213 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4214 /// called with the default operands, and if that fails, with commuted
4216 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4217 TargetLowering::DAGCombinerInfo &DCI) {
4218 SelectionDAG &DAG = DCI.DAG;
4220 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4221 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4222 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4223 if (Result.getNode()) return Result;
4226 // fold (add (arm_neon_vabd a, b) c) -> (arm_neon_vaba c, a, b)
4227 EVT VT = N->getValueType(0);
4228 if (N0.getOpcode() == ISD::INTRINSIC_WO_CHAIN && VT.isInteger()) {
4229 unsigned IntNo = cast<ConstantSDNode>(N0.getOperand(0))->getZExtValue();
4230 if (IntNo == Intrinsic::arm_neon_vabds)
4231 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), VT,
4232 DAG.getConstant(Intrinsic::arm_neon_vabas, MVT::i32),
4233 N1, N0.getOperand(1), N0.getOperand(2));
4234 if (IntNo == Intrinsic::arm_neon_vabdu)
4235 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), VT,
4236 DAG.getConstant(Intrinsic::arm_neon_vabau, MVT::i32),
4237 N1, N0.getOperand(1), N0.getOperand(2));
4243 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4245 static SDValue PerformADDCombine(SDNode *N,
4246 TargetLowering::DAGCombinerInfo &DCI) {
4247 SDValue N0 = N->getOperand(0);
4248 SDValue N1 = N->getOperand(1);
4250 // First try with the default operand order.
4251 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4252 if (Result.getNode())
4255 // If that didn't work, try again with the operands commuted.
4256 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4259 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4261 static SDValue PerformSUBCombine(SDNode *N,
4262 TargetLowering::DAGCombinerInfo &DCI) {
4263 SDValue N0 = N->getOperand(0);
4264 SDValue N1 = N->getOperand(1);
4266 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4267 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4268 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4269 if (Result.getNode()) return Result;
4275 static SDValue PerformMULCombine(SDNode *N,
4276 TargetLowering::DAGCombinerInfo &DCI,
4277 const ARMSubtarget *Subtarget) {
4278 SelectionDAG &DAG = DCI.DAG;
4280 if (Subtarget->isThumb1Only())
4283 if (DAG.getMachineFunction().
4284 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4287 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4290 EVT VT = N->getValueType(0);
4294 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4298 uint64_t MulAmt = C->getZExtValue();
4299 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4300 ShiftAmt = ShiftAmt & (32 - 1);
4301 SDValue V = N->getOperand(0);
4302 DebugLoc DL = N->getDebugLoc();
4305 MulAmt >>= ShiftAmt;
4306 if (isPowerOf2_32(MulAmt - 1)) {
4307 // (mul x, 2^N + 1) => (add (shl x, N), x)
4308 Res = DAG.getNode(ISD::ADD, DL, VT,
4309 V, DAG.getNode(ISD::SHL, DL, VT,
4310 V, DAG.getConstant(Log2_32(MulAmt-1),
4312 } else if (isPowerOf2_32(MulAmt + 1)) {
4313 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4314 Res = DAG.getNode(ISD::SUB, DL, VT,
4315 DAG.getNode(ISD::SHL, DL, VT,
4316 V, DAG.getConstant(Log2_32(MulAmt+1),
4323 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4324 DAG.getConstant(ShiftAmt, MVT::i32));
4326 // Do not add new nodes to DAG combiner worklist.
4327 DCI.CombineTo(N, Res, false);
4331 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4332 static SDValue PerformORCombine(SDNode *N,
4333 TargetLowering::DAGCombinerInfo &DCI,
4334 const ARMSubtarget *Subtarget) {
4335 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4338 // BFI is only available on V6T2+
4339 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4342 SelectionDAG &DAG = DCI.DAG;
4343 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4344 DebugLoc DL = N->getDebugLoc();
4345 // 1) or (and A, mask), val => ARMbfi A, val, mask
4346 // iff (val & mask) == val
4348 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4349 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4350 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4351 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4352 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4353 // (i.e., copy a bitfield value into another bitfield of the same width)
4354 if (N0.getOpcode() != ISD::AND)
4357 EVT VT = N->getValueType(0);
4362 // The value and the mask need to be constants so we can verify this is
4363 // actually a bitfield set. If the mask is 0xffff, we can do better
4364 // via a movt instruction, so don't use BFI in that case.
4365 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4368 unsigned Mask = C->getZExtValue();
4372 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4373 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4374 unsigned Val = C->getZExtValue();
4375 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4377 Val >>= CountTrailingZeros_32(~Mask);
4379 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4380 DAG.getConstant(Val, MVT::i32),
4381 DAG.getConstant(Mask, MVT::i32));
4383 // Do not add new nodes to DAG combiner worklist.
4384 DCI.CombineTo(N, Res, false);
4385 } else if (N1.getOpcode() == ISD::AND) {
4386 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4387 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4390 unsigned Mask2 = C->getZExtValue();
4392 if (ARM::isBitFieldInvertedMask(Mask) &&
4393 ARM::isBitFieldInvertedMask(~Mask2) &&
4394 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4395 // The pack halfword instruction works better for masks that fit it,
4396 // so use that when it's available.
4397 if (Subtarget->hasT2ExtractPack() &&
4398 (Mask == 0xffff || Mask == 0xffff0000))
4401 unsigned lsb = CountTrailingZeros_32(Mask2);
4402 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4403 DAG.getConstant(lsb, MVT::i32));
4404 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4405 DAG.getConstant(Mask, MVT::i32));
4406 // Do not add new nodes to DAG combiner worklist.
4407 DCI.CombineTo(N, Res, false);
4408 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4409 ARM::isBitFieldInvertedMask(Mask2) &&
4410 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4411 // The pack halfword instruction works better for masks that fit it,
4412 // so use that when it's available.
4413 if (Subtarget->hasT2ExtractPack() &&
4414 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4417 unsigned lsb = CountTrailingZeros_32(Mask);
4418 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4419 DAG.getConstant(lsb, MVT::i32));
4420 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4421 DAG.getConstant(Mask2, MVT::i32));
4422 // Do not add new nodes to DAG combiner worklist.
4423 DCI.CombineTo(N, Res, false);
4430 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4431 /// ARMISD::VMOVRRD.
4432 static SDValue PerformVMOVRRDCombine(SDNode *N,
4433 TargetLowering::DAGCombinerInfo &DCI) {
4434 // fmrrd(fmdrr x, y) -> x,y
4435 SDValue InDouble = N->getOperand(0);
4436 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4437 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4441 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4442 /// ARMISD::VDUPLANE.
4443 static SDValue PerformVDUPLANECombine(SDNode *N,
4444 TargetLowering::DAGCombinerInfo &DCI) {
4445 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4447 SDValue Op = N->getOperand(0);
4448 EVT VT = N->getValueType(0);
4450 // Ignore bit_converts.
4451 while (Op.getOpcode() == ISD::BIT_CONVERT)
4452 Op = Op.getOperand(0);
4453 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4456 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4457 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4458 // The canonical VMOV for a zero vector uses a 32-bit element size.
4459 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4461 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4463 if (EltSize > VT.getVectorElementType().getSizeInBits())
4466 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4467 return DCI.CombineTo(N, Res, false);
4470 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4471 /// operand of a vector shift operation, where all the elements of the
4472 /// build_vector must have the same constant integer value.
4473 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4474 // Ignore bit_converts.
4475 while (Op.getOpcode() == ISD::BIT_CONVERT)
4476 Op = Op.getOperand(0);
4477 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4478 APInt SplatBits, SplatUndef;
4479 unsigned SplatBitSize;
4481 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4482 HasAnyUndefs, ElementBits) ||
4483 SplatBitSize > ElementBits)
4485 Cnt = SplatBits.getSExtValue();
4489 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4490 /// operand of a vector shift left operation. That value must be in the range:
4491 /// 0 <= Value < ElementBits for a left shift; or
4492 /// 0 <= Value <= ElementBits for a long left shift.
4493 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4494 assert(VT.isVector() && "vector shift count is not a vector type");
4495 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4496 if (! getVShiftImm(Op, ElementBits, Cnt))
4498 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4501 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4502 /// operand of a vector shift right operation. For a shift opcode, the value
4503 /// is positive, but for an intrinsic the value count must be negative. The
4504 /// absolute value must be in the range:
4505 /// 1 <= |Value| <= ElementBits for a right shift; or
4506 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4507 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4509 assert(VT.isVector() && "vector shift count is not a vector type");
4510 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4511 if (! getVShiftImm(Op, ElementBits, Cnt))
4515 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4518 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4519 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4520 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4523 // Don't do anything for most intrinsics.
4526 // Vector shifts: check for immediate versions and lower them.
4527 // Note: This is done during DAG combining instead of DAG legalizing because
4528 // the build_vectors for 64-bit vector element shift counts are generally
4529 // not legal, and it is hard to see their values after they get legalized to
4530 // loads from a constant pool.
4531 case Intrinsic::arm_neon_vshifts:
4532 case Intrinsic::arm_neon_vshiftu:
4533 case Intrinsic::arm_neon_vshiftls:
4534 case Intrinsic::arm_neon_vshiftlu:
4535 case Intrinsic::arm_neon_vshiftn:
4536 case Intrinsic::arm_neon_vrshifts:
4537 case Intrinsic::arm_neon_vrshiftu:
4538 case Intrinsic::arm_neon_vrshiftn:
4539 case Intrinsic::arm_neon_vqshifts:
4540 case Intrinsic::arm_neon_vqshiftu:
4541 case Intrinsic::arm_neon_vqshiftsu:
4542 case Intrinsic::arm_neon_vqshiftns:
4543 case Intrinsic::arm_neon_vqshiftnu:
4544 case Intrinsic::arm_neon_vqshiftnsu:
4545 case Intrinsic::arm_neon_vqrshiftns:
4546 case Intrinsic::arm_neon_vqrshiftnu:
4547 case Intrinsic::arm_neon_vqrshiftnsu: {
4548 EVT VT = N->getOperand(1).getValueType();
4550 unsigned VShiftOpc = 0;
4553 case Intrinsic::arm_neon_vshifts:
4554 case Intrinsic::arm_neon_vshiftu:
4555 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4556 VShiftOpc = ARMISD::VSHL;
4559 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4560 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4561 ARMISD::VSHRs : ARMISD::VSHRu);
4566 case Intrinsic::arm_neon_vshiftls:
4567 case Intrinsic::arm_neon_vshiftlu:
4568 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4570 llvm_unreachable("invalid shift count for vshll intrinsic");
4572 case Intrinsic::arm_neon_vrshifts:
4573 case Intrinsic::arm_neon_vrshiftu:
4574 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4578 case Intrinsic::arm_neon_vqshifts:
4579 case Intrinsic::arm_neon_vqshiftu:
4580 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4584 case Intrinsic::arm_neon_vqshiftsu:
4585 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4587 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4589 case Intrinsic::arm_neon_vshiftn:
4590 case Intrinsic::arm_neon_vrshiftn:
4591 case Intrinsic::arm_neon_vqshiftns:
4592 case Intrinsic::arm_neon_vqshiftnu:
4593 case Intrinsic::arm_neon_vqshiftnsu:
4594 case Intrinsic::arm_neon_vqrshiftns:
4595 case Intrinsic::arm_neon_vqrshiftnu:
4596 case Intrinsic::arm_neon_vqrshiftnsu:
4597 // Narrowing shifts require an immediate right shift.
4598 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4600 llvm_unreachable("invalid shift count for narrowing vector shift "
4604 llvm_unreachable("unhandled vector shift");
4608 case Intrinsic::arm_neon_vshifts:
4609 case Intrinsic::arm_neon_vshiftu:
4610 // Opcode already set above.
4612 case Intrinsic::arm_neon_vshiftls:
4613 case Intrinsic::arm_neon_vshiftlu:
4614 if (Cnt == VT.getVectorElementType().getSizeInBits())
4615 VShiftOpc = ARMISD::VSHLLi;
4617 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4618 ARMISD::VSHLLs : ARMISD::VSHLLu);
4620 case Intrinsic::arm_neon_vshiftn:
4621 VShiftOpc = ARMISD::VSHRN; break;
4622 case Intrinsic::arm_neon_vrshifts:
4623 VShiftOpc = ARMISD::VRSHRs; break;
4624 case Intrinsic::arm_neon_vrshiftu:
4625 VShiftOpc = ARMISD::VRSHRu; break;
4626 case Intrinsic::arm_neon_vrshiftn:
4627 VShiftOpc = ARMISD::VRSHRN; break;
4628 case Intrinsic::arm_neon_vqshifts:
4629 VShiftOpc = ARMISD::VQSHLs; break;
4630 case Intrinsic::arm_neon_vqshiftu:
4631 VShiftOpc = ARMISD::VQSHLu; break;
4632 case Intrinsic::arm_neon_vqshiftsu:
4633 VShiftOpc = ARMISD::VQSHLsu; break;
4634 case Intrinsic::arm_neon_vqshiftns:
4635 VShiftOpc = ARMISD::VQSHRNs; break;
4636 case Intrinsic::arm_neon_vqshiftnu:
4637 VShiftOpc = ARMISD::VQSHRNu; break;
4638 case Intrinsic::arm_neon_vqshiftnsu:
4639 VShiftOpc = ARMISD::VQSHRNsu; break;
4640 case Intrinsic::arm_neon_vqrshiftns:
4641 VShiftOpc = ARMISD::VQRSHRNs; break;
4642 case Intrinsic::arm_neon_vqrshiftnu:
4643 VShiftOpc = ARMISD::VQRSHRNu; break;
4644 case Intrinsic::arm_neon_vqrshiftnsu:
4645 VShiftOpc = ARMISD::VQRSHRNsu; break;
4648 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4649 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4652 case Intrinsic::arm_neon_vshiftins: {
4653 EVT VT = N->getOperand(1).getValueType();
4655 unsigned VShiftOpc = 0;
4657 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4658 VShiftOpc = ARMISD::VSLI;
4659 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4660 VShiftOpc = ARMISD::VSRI;
4662 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4665 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4666 N->getOperand(1), N->getOperand(2),
4667 DAG.getConstant(Cnt, MVT::i32));
4670 case Intrinsic::arm_neon_vqrshifts:
4671 case Intrinsic::arm_neon_vqrshiftu:
4672 // No immediate versions of these to check for.
4679 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4680 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4681 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4682 /// vector element shift counts are generally not legal, and it is hard to see
4683 /// their values after they get legalized to loads from a constant pool.
4684 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4685 const ARMSubtarget *ST) {
4686 EVT VT = N->getValueType(0);
4688 // Nothing to be done for scalar shifts.
4689 if (! VT.isVector())
4692 assert(ST->hasNEON() && "unexpected vector shift");
4695 switch (N->getOpcode()) {
4696 default: llvm_unreachable("unexpected shift opcode");
4699 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4700 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4701 DAG.getConstant(Cnt, MVT::i32));
4706 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4707 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4708 ARMISD::VSHRs : ARMISD::VSHRu);
4709 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4710 DAG.getConstant(Cnt, MVT::i32));
4716 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4717 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4718 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4719 const ARMSubtarget *ST) {
4720 SDValue N0 = N->getOperand(0);
4722 // Check for sign- and zero-extensions of vector extract operations of 8-
4723 // and 16-bit vector elements. NEON supports these directly. They are
4724 // handled during DAG combining because type legalization will promote them
4725 // to 32-bit types and it is messy to recognize the operations after that.
4726 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4727 SDValue Vec = N0.getOperand(0);
4728 SDValue Lane = N0.getOperand(1);
4729 EVT VT = N->getValueType(0);
4730 EVT EltVT = N0.getValueType();
4731 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4733 if (VT == MVT::i32 &&
4734 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4735 TLI.isTypeLegal(Vec.getValueType())) {
4738 switch (N->getOpcode()) {
4739 default: llvm_unreachable("unexpected opcode");
4740 case ISD::SIGN_EXTEND:
4741 Opc = ARMISD::VGETLANEs;
4743 case ISD::ZERO_EXTEND:
4744 case ISD::ANY_EXTEND:
4745 Opc = ARMISD::VGETLANEu;
4748 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4755 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4756 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4757 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4758 const ARMSubtarget *ST) {
4759 // If the target supports NEON, try to use vmax/vmin instructions for f32
4760 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4761 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4762 // a NaN; only do the transformation when it matches that behavior.
4764 // For now only do this when using NEON for FP operations; if using VFP, it
4765 // is not obvious that the benefit outweighs the cost of switching to the
4767 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4768 N->getValueType(0) != MVT::f32)
4771 SDValue CondLHS = N->getOperand(0);
4772 SDValue CondRHS = N->getOperand(1);
4773 SDValue LHS = N->getOperand(2);
4774 SDValue RHS = N->getOperand(3);
4775 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4777 unsigned Opcode = 0;
4779 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4780 IsReversed = false; // x CC y ? x : y
4781 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4782 IsReversed = true ; // x CC y ? y : x
4796 // If LHS is NaN, an ordered comparison will be false and the result will
4797 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4798 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4799 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4800 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4802 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4803 // will return -0, so vmin can only be used for unsafe math or if one of
4804 // the operands is known to be nonzero.
4805 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4807 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4809 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4818 // If LHS is NaN, an ordered comparison will be false and the result will
4819 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4820 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4821 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4822 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4824 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4825 // will return +0, so vmax can only be used for unsafe math or if one of
4826 // the operands is known to be nonzero.
4827 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4829 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4831 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4837 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4840 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4841 DAGCombinerInfo &DCI) const {
4842 switch (N->getOpcode()) {
4844 case ISD::ADD: return PerformADDCombine(N, DCI);
4845 case ISD::SUB: return PerformSUBCombine(N, DCI);
4846 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4847 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4848 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4849 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4850 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4853 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4854 case ISD::SIGN_EXTEND:
4855 case ISD::ZERO_EXTEND:
4856 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4857 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4862 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4863 if (!Subtarget->hasV6Ops())
4864 // Pre-v6 does not support unaligned mem access.
4867 // v6+ may or may not support unaligned mem access depending on the system
4869 // FIXME: This is pretty conservative. Should we provide cmdline option to
4870 // control the behaviour?
4871 if (!Subtarget->isTargetDarwin())
4874 switch (VT.getSimpleVT().SimpleTy) {
4881 // FIXME: VLD1 etc with standard alignment is legal.
4885 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4890 switch (VT.getSimpleVT().SimpleTy) {
4891 default: return false;
4906 if ((V & (Scale - 1)) != 0)
4909 return V == (V & ((1LL << 5) - 1));
4912 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4913 const ARMSubtarget *Subtarget) {
4920 switch (VT.getSimpleVT().SimpleTy) {
4921 default: return false;
4926 // + imm12 or - imm8
4928 return V == (V & ((1LL << 8) - 1));
4929 return V == (V & ((1LL << 12) - 1));
4932 // Same as ARM mode. FIXME: NEON?
4933 if (!Subtarget->hasVFP2())
4938 return V == (V & ((1LL << 8) - 1));
4942 /// isLegalAddressImmediate - Return true if the integer value can be used
4943 /// as the offset of the target addressing mode for load / store of the
4945 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4946 const ARMSubtarget *Subtarget) {
4953 if (Subtarget->isThumb1Only())
4954 return isLegalT1AddressImmediate(V, VT);
4955 else if (Subtarget->isThumb2())
4956 return isLegalT2AddressImmediate(V, VT, Subtarget);
4961 switch (VT.getSimpleVT().SimpleTy) {
4962 default: return false;
4967 return V == (V & ((1LL << 12) - 1));
4970 return V == (V & ((1LL << 8) - 1));
4973 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4978 return V == (V & ((1LL << 8) - 1));
4982 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4984 int Scale = AM.Scale;
4988 switch (VT.getSimpleVT().SimpleTy) {
4989 default: return false;
4998 return Scale == 2 || Scale == 4 || Scale == 8;
5001 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5005 // Note, we allow "void" uses (basically, uses that aren't loads or
5006 // stores), because arm allows folding a scale into many arithmetic
5007 // operations. This should be made more precise and revisited later.
5009 // Allow r << imm, but the imm has to be a multiple of two.
5010 if (Scale & 1) return false;
5011 return isPowerOf2_32(Scale);
5015 /// isLegalAddressingMode - Return true if the addressing mode represented
5016 /// by AM is legal for this target, for a load/store of the specified type.
5017 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5018 const Type *Ty) const {
5019 EVT VT = getValueType(Ty, true);
5020 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
5023 // Can never fold addr of global into load/store.
5028 case 0: // no scale reg, must be "r+i" or "r", or "i".
5031 if (Subtarget->isThumb1Only())
5035 // ARM doesn't support any R+R*scale+imm addr modes.
5042 if (Subtarget->isThumb2())
5043 return isLegalT2ScaledAddressingMode(AM, VT);
5045 int Scale = AM.Scale;
5046 switch (VT.getSimpleVT().SimpleTy) {
5047 default: return false;
5051 if (Scale < 0) Scale = -Scale;
5055 return isPowerOf2_32(Scale & ~1);
5059 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5064 // Note, we allow "void" uses (basically, uses that aren't loads or
5065 // stores), because arm allows folding a scale into many arithmetic
5066 // operations. This should be made more precise and revisited later.
5068 // Allow r << imm, but the imm has to be a multiple of two.
5069 if (Scale & 1) return false;
5070 return isPowerOf2_32(Scale);
5077 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5078 /// icmp immediate, that is the target has icmp instructions which can compare
5079 /// a register against the immediate without having to materialize the
5080 /// immediate into a register.
5081 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5082 if (!Subtarget->isThumb())
5083 return ARM_AM::getSOImmVal(Imm) != -1;
5084 if (Subtarget->isThumb2())
5085 return ARM_AM::getT2SOImmVal(Imm) != -1;
5086 return Imm >= 0 && Imm <= 255;
5089 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5090 bool isSEXTLoad, SDValue &Base,
5091 SDValue &Offset, bool &isInc,
5092 SelectionDAG &DAG) {
5093 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5096 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5098 Base = Ptr->getOperand(0);
5099 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5100 int RHSC = (int)RHS->getZExtValue();
5101 if (RHSC < 0 && RHSC > -256) {
5102 assert(Ptr->getOpcode() == ISD::ADD);
5104 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5108 isInc = (Ptr->getOpcode() == ISD::ADD);
5109 Offset = Ptr->getOperand(1);
5111 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5113 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5114 int RHSC = (int)RHS->getZExtValue();
5115 if (RHSC < 0 && RHSC > -0x1000) {
5116 assert(Ptr->getOpcode() == ISD::ADD);
5118 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5119 Base = Ptr->getOperand(0);
5124 if (Ptr->getOpcode() == ISD::ADD) {
5126 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5127 if (ShOpcVal != ARM_AM::no_shift) {
5128 Base = Ptr->getOperand(1);
5129 Offset = Ptr->getOperand(0);
5131 Base = Ptr->getOperand(0);
5132 Offset = Ptr->getOperand(1);
5137 isInc = (Ptr->getOpcode() == ISD::ADD);
5138 Base = Ptr->getOperand(0);
5139 Offset = Ptr->getOperand(1);
5143 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5147 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5148 bool isSEXTLoad, SDValue &Base,
5149 SDValue &Offset, bool &isInc,
5150 SelectionDAG &DAG) {
5151 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5154 Base = Ptr->getOperand(0);
5155 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5156 int RHSC = (int)RHS->getZExtValue();
5157 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5158 assert(Ptr->getOpcode() == ISD::ADD);
5160 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5162 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5163 isInc = Ptr->getOpcode() == ISD::ADD;
5164 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5172 /// getPreIndexedAddressParts - returns true by value, base pointer and
5173 /// offset pointer and addressing mode by reference if the node's address
5174 /// can be legally represented as pre-indexed load / store address.
5176 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5178 ISD::MemIndexedMode &AM,
5179 SelectionDAG &DAG) const {
5180 if (Subtarget->isThumb1Only())
5185 bool isSEXTLoad = false;
5186 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5187 Ptr = LD->getBasePtr();
5188 VT = LD->getMemoryVT();
5189 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5190 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5191 Ptr = ST->getBasePtr();
5192 VT = ST->getMemoryVT();
5197 bool isLegal = false;
5198 if (Subtarget->isThumb2())
5199 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5200 Offset, isInc, DAG);
5202 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5203 Offset, isInc, DAG);
5207 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5211 /// getPostIndexedAddressParts - returns true by value, base pointer and
5212 /// offset pointer and addressing mode by reference if this node can be
5213 /// combined with a load / store to form a post-indexed load / store.
5214 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5217 ISD::MemIndexedMode &AM,
5218 SelectionDAG &DAG) const {
5219 if (Subtarget->isThumb1Only())
5224 bool isSEXTLoad = false;
5225 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5226 VT = LD->getMemoryVT();
5227 Ptr = LD->getBasePtr();
5228 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5229 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5230 VT = ST->getMemoryVT();
5231 Ptr = ST->getBasePtr();
5236 bool isLegal = false;
5237 if (Subtarget->isThumb2())
5238 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5241 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5247 // Swap base ptr and offset to catch more post-index load / store when
5248 // it's legal. In Thumb2 mode, offset must be an immediate.
5249 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5250 !Subtarget->isThumb2())
5251 std::swap(Base, Offset);
5253 // Post-indexed load / store update the base pointer.
5258 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5262 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5266 const SelectionDAG &DAG,
5267 unsigned Depth) const {
5268 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5269 switch (Op.getOpcode()) {
5271 case ARMISD::CMOV: {
5272 // Bits are known zero/one if known on the LHS and RHS.
5273 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5274 if (KnownZero == 0 && KnownOne == 0) return;
5276 APInt KnownZeroRHS, KnownOneRHS;
5277 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5278 KnownZeroRHS, KnownOneRHS, Depth+1);
5279 KnownZero &= KnownZeroRHS;
5280 KnownOne &= KnownOneRHS;
5286 //===----------------------------------------------------------------------===//
5287 // ARM Inline Assembly Support
5288 //===----------------------------------------------------------------------===//
5290 /// getConstraintType - Given a constraint letter, return the type of
5291 /// constraint it is for this target.
5292 ARMTargetLowering::ConstraintType
5293 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5294 if (Constraint.size() == 1) {
5295 switch (Constraint[0]) {
5297 case 'l': return C_RegisterClass;
5298 case 'w': return C_RegisterClass;
5301 return TargetLowering::getConstraintType(Constraint);
5304 std::pair<unsigned, const TargetRegisterClass*>
5305 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5307 if (Constraint.size() == 1) {
5308 // GCC ARM Constraint Letters
5309 switch (Constraint[0]) {
5311 if (Subtarget->isThumb())
5312 return std::make_pair(0U, ARM::tGPRRegisterClass);
5314 return std::make_pair(0U, ARM::GPRRegisterClass);
5316 return std::make_pair(0U, ARM::GPRRegisterClass);
5319 return std::make_pair(0U, ARM::SPRRegisterClass);
5320 if (VT.getSizeInBits() == 64)
5321 return std::make_pair(0U, ARM::DPRRegisterClass);
5322 if (VT.getSizeInBits() == 128)
5323 return std::make_pair(0U, ARM::QPRRegisterClass);
5327 if (StringRef("{cc}").equals_lower(Constraint))
5328 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5330 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5333 std::vector<unsigned> ARMTargetLowering::
5334 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5336 if (Constraint.size() != 1)
5337 return std::vector<unsigned>();
5339 switch (Constraint[0]) { // GCC ARM Constraint Letters
5342 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5343 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5346 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5347 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5348 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5349 ARM::R12, ARM::LR, 0);
5352 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5353 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5354 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5355 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5356 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5357 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5358 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5359 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5360 if (VT.getSizeInBits() == 64)
5361 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5362 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5363 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5364 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5365 if (VT.getSizeInBits() == 128)
5366 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5367 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5371 return std::vector<unsigned>();
5374 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5375 /// vector. If it is invalid, don't add anything to Ops.
5376 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5378 std::vector<SDValue>&Ops,
5379 SelectionDAG &DAG) const {
5380 SDValue Result(0, 0);
5382 switch (Constraint) {
5384 case 'I': case 'J': case 'K': case 'L':
5385 case 'M': case 'N': case 'O':
5386 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5390 int64_t CVal64 = C->getSExtValue();
5391 int CVal = (int) CVal64;
5392 // None of these constraints allow values larger than 32 bits. Check
5393 // that the value fits in an int.
5397 switch (Constraint) {
5399 if (Subtarget->isThumb1Only()) {
5400 // This must be a constant between 0 and 255, for ADD
5402 if (CVal >= 0 && CVal <= 255)
5404 } else if (Subtarget->isThumb2()) {
5405 // A constant that can be used as an immediate value in a
5406 // data-processing instruction.
5407 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5410 // A constant that can be used as an immediate value in a
5411 // data-processing instruction.
5412 if (ARM_AM::getSOImmVal(CVal) != -1)
5418 if (Subtarget->isThumb()) { // FIXME thumb2
5419 // This must be a constant between -255 and -1, for negated ADD
5420 // immediates. This can be used in GCC with an "n" modifier that
5421 // prints the negated value, for use with SUB instructions. It is
5422 // not useful otherwise but is implemented for compatibility.
5423 if (CVal >= -255 && CVal <= -1)
5426 // This must be a constant between -4095 and 4095. It is not clear
5427 // what this constraint is intended for. Implemented for
5428 // compatibility with GCC.
5429 if (CVal >= -4095 && CVal <= 4095)
5435 if (Subtarget->isThumb1Only()) {
5436 // A 32-bit value where only one byte has a nonzero value. Exclude
5437 // zero to match GCC. This constraint is used by GCC internally for
5438 // constants that can be loaded with a move/shift combination.
5439 // It is not useful otherwise but is implemented for compatibility.
5440 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5442 } else if (Subtarget->isThumb2()) {
5443 // A constant whose bitwise inverse can be used as an immediate
5444 // value in a data-processing instruction. This can be used in GCC
5445 // with a "B" modifier that prints the inverted value, for use with
5446 // BIC and MVN instructions. It is not useful otherwise but is
5447 // implemented for compatibility.
5448 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5451 // A constant whose bitwise inverse can be used as an immediate
5452 // value in a data-processing instruction. This can be used in GCC
5453 // with a "B" modifier that prints the inverted value, for use with
5454 // BIC and MVN instructions. It is not useful otherwise but is
5455 // implemented for compatibility.
5456 if (ARM_AM::getSOImmVal(~CVal) != -1)
5462 if (Subtarget->isThumb1Only()) {
5463 // This must be a constant between -7 and 7,
5464 // for 3-operand ADD/SUB immediate instructions.
5465 if (CVal >= -7 && CVal < 7)
5467 } else if (Subtarget->isThumb2()) {
5468 // A constant whose negation can be used as an immediate value in a
5469 // data-processing instruction. This can be used in GCC with an "n"
5470 // modifier that prints the negated value, for use with SUB
5471 // instructions. It is not useful otherwise but is implemented for
5473 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5476 // A constant whose negation can be used as an immediate value in a
5477 // data-processing instruction. This can be used in GCC with an "n"
5478 // modifier that prints the negated value, for use with SUB
5479 // instructions. It is not useful otherwise but is implemented for
5481 if (ARM_AM::getSOImmVal(-CVal) != -1)
5487 if (Subtarget->isThumb()) { // FIXME thumb2
5488 // This must be a multiple of 4 between 0 and 1020, for
5489 // ADD sp + immediate.
5490 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5493 // A power of two or a constant between 0 and 32. This is used in
5494 // GCC for the shift amount on shifted register operands, but it is
5495 // useful in general for any shift amounts.
5496 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5502 if (Subtarget->isThumb()) { // FIXME thumb2
5503 // This must be a constant between 0 and 31, for shift amounts.
5504 if (CVal >= 0 && CVal <= 31)
5510 if (Subtarget->isThumb()) { // FIXME thumb2
5511 // This must be a multiple of 4 between -508 and 508, for
5512 // ADD/SUB sp = sp + immediate.
5513 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5518 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5522 if (Result.getNode()) {
5523 Ops.push_back(Result);
5526 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5530 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5531 // The ARM target isn't yet aware of offsets.
5535 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5536 APInt Imm = FPImm.bitcastToAPInt();
5537 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5538 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5539 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5541 // We can handle 4 bits of mantissa.
5542 // mantissa = (16+UInt(e:f:g:h))/16.
5543 if (Mantissa & 0x7ffff)
5546 if ((Mantissa & 0xf) != Mantissa)
5549 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5550 if (Exp < -3 || Exp > 4)
5552 Exp = ((Exp+3) & 0x7) ^ 4;
5554 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5557 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5558 APInt Imm = FPImm.bitcastToAPInt();
5559 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5560 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5561 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5563 // We can handle 4 bits of mantissa.
5564 // mantissa = (16+UInt(e:f:g:h))/16.
5565 if (Mantissa & 0xffffffffffffLL)
5568 if ((Mantissa & 0xf) != Mantissa)
5571 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5572 if (Exp < -3 || Exp > 4)
5574 Exp = ((Exp+3) & 0x7) ^ 4;
5576 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5579 bool ARM::isBitFieldInvertedMask(unsigned v) {
5580 if (v == 0xffffffff)
5582 // there can be 1's on either or both "outsides", all the "inside"
5584 unsigned int lsb = 0, msb = 31;
5585 while (v & (1 << msb)) --msb;
5586 while (v & (1 << lsb)) ++lsb;
5587 for (unsigned int i = lsb; i <= msb; ++i) {
5594 /// isFPImmLegal - Returns true if the target can instruction select the
5595 /// specified FP immediate natively. If false, the legalizer will
5596 /// materialize the FP immediate as a load from a constant pool.
5597 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5598 if (!Subtarget->hasVFP3())
5601 return ARM::getVFPf32Imm(Imm) != -1;
5603 return ARM::getVFPf64Imm(Imm) != -1;