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 "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMISelLowering.h"
21 #include "ARMMachineFunctionInfo.h"
22 #include "ARMPerfectShuffle.h"
23 #include "ARMRegisterInfo.h"
24 #include "ARMSubtarget.h"
25 #include "ARMTargetMachine.h"
26 #include "ARMTargetObjectFile.h"
27 #include "llvm/CallingConv.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Instruction.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/Intrinsics.h"
34 #include "llvm/Type.h"
35 #include "llvm/CodeGen/CallingConvLower.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/PseudoSourceValue.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
54 STATISTIC(NumTailCalls, "Number of tail calls");
56 // This option should go away when tail calls fully work.
58 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
59 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
62 // This option should go away when Machine LICM is smart enough to hoist a
65 EnableARMVDUPsplat("arm-vdup-splat", cl::Hidden,
66 cl::desc("Generate VDUP for integer constant splats (TEMPORARY OPTION)."),
70 EnableARMLongCalls("arm-long-calls", cl::Hidden,
71 cl::desc("Generate calls via indirect call instructions"),
75 ARMInterworking("arm-interworking", cl::Hidden,
76 cl::desc("Enable / disable ARM interworking (for debugging only)"),
80 EnableARMCodePlacement("arm-code-placement", cl::Hidden,
81 cl::desc("Enable code placement pass for ARM"),
84 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
85 EVT PromotedBitwiseVT) {
86 if (VT != PromotedLdStVT) {
87 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
88 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
89 PromotedLdStVT.getSimpleVT());
91 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
92 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
93 PromotedLdStVT.getSimpleVT());
96 EVT ElemTy = VT.getVectorElementType();
97 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
98 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
99 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
100 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
101 if (ElemTy != MVT::i32) {
102 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
103 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
104 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
105 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
108 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
109 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
110 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
111 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
112 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
113 if (VT.isInteger()) {
114 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
115 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
116 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
117 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
118 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
120 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
122 // Promote all bit-wise operations.
123 if (VT.isInteger() && VT != PromotedBitwiseVT) {
124 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
125 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
126 PromotedBitwiseVT.getSimpleVT());
127 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
128 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
129 PromotedBitwiseVT.getSimpleVT());
130 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
131 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
132 PromotedBitwiseVT.getSimpleVT());
135 // Neon does not support vector divide/remainder operations.
136 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
137 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
138 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
139 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
141 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
144 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
145 addRegisterClass(VT, ARM::DPRRegisterClass);
146 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
149 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
150 addRegisterClass(VT, ARM::QPRRegisterClass);
151 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
154 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
155 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
156 return new TargetLoweringObjectFileMachO();
158 return new ARMElfTargetObjectFile();
161 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
162 : TargetLowering(TM, createTLOF(TM)) {
163 Subtarget = &TM.getSubtarget<ARMSubtarget>();
164 RegInfo = TM.getRegisterInfo();
165 Itins = TM.getInstrItineraryData();
167 if (Subtarget->isTargetDarwin()) {
168 // Uses VFP for Thumb libfuncs if available.
169 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
170 // Single-precision floating-point arithmetic.
171 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
172 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
173 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
174 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
176 // Double-precision floating-point arithmetic.
177 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
178 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
179 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
180 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
182 // Single-precision comparisons.
183 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
184 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
185 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
186 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
187 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
188 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
189 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
190 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
192 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
201 // Double-precision comparisons.
202 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
203 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
204 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
205 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
206 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
207 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
208 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
209 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
211 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
220 // Floating-point to integer conversions.
221 // i64 conversions are done via library routines even when generating VFP
222 // instructions, so use the same ones.
223 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
224 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
225 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
226 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
228 // Conversions between floating types.
229 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
230 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
232 // Integer to floating-point conversions.
233 // i64 conversions are done via library routines even when generating VFP
234 // instructions, so use the same ones.
235 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
236 // e.g., __floatunsidf vs. __floatunssidfvfp.
237 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
238 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
239 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
240 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
244 // These libcalls are not available in 32-bit.
245 setLibcallName(RTLIB::SHL_I128, 0);
246 setLibcallName(RTLIB::SRL_I128, 0);
247 setLibcallName(RTLIB::SRA_I128, 0);
249 // Libcalls should use the AAPCS base standard ABI, even if hard float
250 // is in effect, as per the ARM RTABI specification, section 4.1.2.
251 if (Subtarget->isAAPCS_ABI()) {
252 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
253 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
254 CallingConv::ARM_AAPCS);
258 if (Subtarget->isThumb1Only())
259 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
261 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
262 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
263 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
264 if (!Subtarget->isFPOnlySP())
265 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
267 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
270 if (Subtarget->hasNEON()) {
271 addDRTypeForNEON(MVT::v2f32);
272 addDRTypeForNEON(MVT::v8i8);
273 addDRTypeForNEON(MVT::v4i16);
274 addDRTypeForNEON(MVT::v2i32);
275 addDRTypeForNEON(MVT::v1i64);
277 addQRTypeForNEON(MVT::v4f32);
278 addQRTypeForNEON(MVT::v2f64);
279 addQRTypeForNEON(MVT::v16i8);
280 addQRTypeForNEON(MVT::v8i16);
281 addQRTypeForNEON(MVT::v4i32);
282 addQRTypeForNEON(MVT::v2i64);
284 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
285 // neither Neon nor VFP support any arithmetic operations on it.
286 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
287 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
288 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
289 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
290 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
291 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
292 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
293 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
294 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
295 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
296 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
297 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
298 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
299 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
300 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
301 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
302 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
303 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
304 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
305 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
306 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
307 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
308 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
309 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
311 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
313 // Neon does not support some operations on v1i64 and v2i64 types.
314 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
315 // Custom handling for some quad-vector types to detect VMULL.
316 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
317 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
318 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
319 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
320 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
322 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
323 setTargetDAGCombine(ISD::SHL);
324 setTargetDAGCombine(ISD::SRL);
325 setTargetDAGCombine(ISD::SRA);
326 setTargetDAGCombine(ISD::SIGN_EXTEND);
327 setTargetDAGCombine(ISD::ZERO_EXTEND);
328 setTargetDAGCombine(ISD::ANY_EXTEND);
329 setTargetDAGCombine(ISD::SELECT_CC);
330 setTargetDAGCombine(ISD::BUILD_VECTOR);
333 computeRegisterProperties();
335 // ARM does not have f32 extending load.
336 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
338 // ARM does not have i1 sign extending load.
339 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
341 // ARM supports all 4 flavors of integer indexed load / store.
342 if (!Subtarget->isThumb1Only()) {
343 for (unsigned im = (unsigned)ISD::PRE_INC;
344 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
345 setIndexedLoadAction(im, MVT::i1, Legal);
346 setIndexedLoadAction(im, MVT::i8, Legal);
347 setIndexedLoadAction(im, MVT::i16, Legal);
348 setIndexedLoadAction(im, MVT::i32, Legal);
349 setIndexedStoreAction(im, MVT::i1, Legal);
350 setIndexedStoreAction(im, MVT::i8, Legal);
351 setIndexedStoreAction(im, MVT::i16, Legal);
352 setIndexedStoreAction(im, MVT::i32, Legal);
356 // i64 operation support.
357 if (Subtarget->isThumb1Only()) {
358 setOperationAction(ISD::MUL, MVT::i64, Expand);
359 setOperationAction(ISD::MULHU, MVT::i32, Expand);
360 setOperationAction(ISD::MULHS, MVT::i32, Expand);
361 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
362 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
364 setOperationAction(ISD::MUL, MVT::i64, Expand);
365 setOperationAction(ISD::MULHU, MVT::i32, Expand);
366 if (!Subtarget->hasV6Ops())
367 setOperationAction(ISD::MULHS, MVT::i32, Expand);
369 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
370 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
371 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
372 setOperationAction(ISD::SRL, MVT::i64, Custom);
373 setOperationAction(ISD::SRA, MVT::i64, Custom);
375 // ARM does not have ROTL.
376 setOperationAction(ISD::ROTL, MVT::i32, Expand);
377 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
378 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
379 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
380 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
382 // Only ARMv6 has BSWAP.
383 if (!Subtarget->hasV6Ops())
384 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
386 // These are expanded into libcalls.
387 if (!Subtarget->hasDivide()) {
388 // v7M has a hardware divider
389 setOperationAction(ISD::SDIV, MVT::i32, Expand);
390 setOperationAction(ISD::UDIV, MVT::i32, Expand);
392 setOperationAction(ISD::SREM, MVT::i32, Expand);
393 setOperationAction(ISD::UREM, MVT::i32, Expand);
394 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
395 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
397 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
398 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
399 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
400 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
401 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
403 setOperationAction(ISD::TRAP, MVT::Other, Legal);
405 // Use the default implementation.
406 setOperationAction(ISD::VASTART, MVT::Other, Custom);
407 setOperationAction(ISD::VAARG, MVT::Other, Expand);
408 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
409 setOperationAction(ISD::VAEND, MVT::Other, Expand);
410 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
411 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
412 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
413 // FIXME: Shouldn't need this, since no register is used, but the legalizer
414 // doesn't yet know how to not do that for SjLj.
415 setExceptionSelectorRegister(ARM::R0);
416 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
417 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
418 // the default expansion.
419 if (Subtarget->hasDataBarrier() ||
420 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
421 // membarrier needs custom lowering; the rest are legal and handled
423 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
425 // Set them all for expansion, which will force libcalls.
426 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
427 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
428 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
429 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
430 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
431 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
432 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
449 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
450 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
451 // Since the libcalls include locking, fold in the fences
452 setShouldFoldAtomicFences(true);
454 // 64-bit versions are always libcalls (for now)
455 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
456 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
457 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
458 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
459 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
460 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
461 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
462 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
464 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
465 if (!Subtarget->hasV6Ops()) {
466 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
467 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
469 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
471 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
472 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
473 // iff target supports vfp2.
474 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
475 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
478 // We want to custom lower some of our intrinsics.
479 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
480 if (Subtarget->isTargetDarwin()) {
481 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
482 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
485 setOperationAction(ISD::SETCC, MVT::i32, Expand);
486 setOperationAction(ISD::SETCC, MVT::f32, Expand);
487 setOperationAction(ISD::SETCC, MVT::f64, Expand);
488 setOperationAction(ISD::SELECT, MVT::i32, Custom);
489 setOperationAction(ISD::SELECT, MVT::f32, Custom);
490 setOperationAction(ISD::SELECT, MVT::f64, Custom);
491 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
492 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
493 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
495 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
496 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
497 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
498 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
499 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
501 // We don't support sin/cos/fmod/copysign/pow
502 setOperationAction(ISD::FSIN, MVT::f64, Expand);
503 setOperationAction(ISD::FSIN, MVT::f32, Expand);
504 setOperationAction(ISD::FCOS, MVT::f32, Expand);
505 setOperationAction(ISD::FCOS, MVT::f64, Expand);
506 setOperationAction(ISD::FREM, MVT::f64, Expand);
507 setOperationAction(ISD::FREM, MVT::f32, Expand);
508 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
509 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
510 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
512 setOperationAction(ISD::FPOW, MVT::f64, Expand);
513 setOperationAction(ISD::FPOW, MVT::f32, Expand);
515 // Various VFP goodness
516 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
517 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
518 if (Subtarget->hasVFP2()) {
519 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
520 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
521 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
522 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
524 // Special handling for half-precision FP.
525 if (!Subtarget->hasFP16()) {
526 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
527 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
531 // We have target-specific dag combine patterns for the following nodes:
532 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
533 setTargetDAGCombine(ISD::ADD);
534 setTargetDAGCombine(ISD::SUB);
535 setTargetDAGCombine(ISD::MUL);
537 if (Subtarget->hasV6T2Ops())
538 setTargetDAGCombine(ISD::OR);
540 setStackPointerRegisterToSaveRestore(ARM::SP);
542 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
543 setSchedulingPreference(Sched::RegPressure);
545 setSchedulingPreference(Sched::Hybrid);
547 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
549 // On ARM arguments smaller than 4 bytes are extended, so all arguments
550 // are at least 4 bytes aligned.
551 setMinStackArgumentAlignment(4);
553 if (EnableARMCodePlacement)
554 benefitFromCodePlacementOpt = true;
557 std::pair<const TargetRegisterClass*, uint8_t>
558 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
559 const TargetRegisterClass *RRC = 0;
561 switch (VT.getSimpleVT().SimpleTy) {
563 return TargetLowering::findRepresentativeClass(VT);
564 // Use DPR as representative register class for all floating point
565 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
566 // the cost is 1 for both f32 and f64.
567 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
568 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
569 RRC = ARM::DPRRegisterClass;
571 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
572 case MVT::v4f32: case MVT::v2f64:
573 RRC = ARM::DPRRegisterClass;
577 RRC = ARM::DPRRegisterClass;
581 RRC = ARM::DPRRegisterClass;
585 return std::make_pair(RRC, Cost);
588 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
591 case ARMISD::Wrapper: return "ARMISD::Wrapper";
592 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
593 case ARMISD::CALL: return "ARMISD::CALL";
594 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
595 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
596 case ARMISD::tCALL: return "ARMISD::tCALL";
597 case ARMISD::BRCOND: return "ARMISD::BRCOND";
598 case ARMISD::BR_JT: return "ARMISD::BR_JT";
599 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
600 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
601 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
602 case ARMISD::AND: return "ARMISD::AND";
603 case ARMISD::CMP: return "ARMISD::CMP";
604 case ARMISD::CMPZ: return "ARMISD::CMPZ";
605 case ARMISD::CMPFP: return "ARMISD::CMPFP";
606 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
607 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
608 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
609 case ARMISD::CMOV: return "ARMISD::CMOV";
610 case ARMISD::CNEG: return "ARMISD::CNEG";
612 case ARMISD::RBIT: return "ARMISD::RBIT";
614 case ARMISD::FTOSI: return "ARMISD::FTOSI";
615 case ARMISD::FTOUI: return "ARMISD::FTOUI";
616 case ARMISD::SITOF: return "ARMISD::SITOF";
617 case ARMISD::UITOF: return "ARMISD::UITOF";
619 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
620 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
621 case ARMISD::RRX: return "ARMISD::RRX";
623 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
624 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
626 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
627 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
629 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
631 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
633 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
635 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
636 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
638 case ARMISD::VCEQ: return "ARMISD::VCEQ";
639 case ARMISD::VCGE: return "ARMISD::VCGE";
640 case ARMISD::VCGEU: return "ARMISD::VCGEU";
641 case ARMISD::VCGT: return "ARMISD::VCGT";
642 case ARMISD::VCGTU: return "ARMISD::VCGTU";
643 case ARMISD::VTST: return "ARMISD::VTST";
645 case ARMISD::VSHL: return "ARMISD::VSHL";
646 case ARMISD::VSHRs: return "ARMISD::VSHRs";
647 case ARMISD::VSHRu: return "ARMISD::VSHRu";
648 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
649 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
650 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
651 case ARMISD::VSHRN: return "ARMISD::VSHRN";
652 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
653 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
654 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
655 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
656 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
657 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
658 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
659 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
660 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
661 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
662 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
663 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
664 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
665 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
666 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
667 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
668 case ARMISD::VDUP: return "ARMISD::VDUP";
669 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
670 case ARMISD::VEXT: return "ARMISD::VEXT";
671 case ARMISD::VREV64: return "ARMISD::VREV64";
672 case ARMISD::VREV32: return "ARMISD::VREV32";
673 case ARMISD::VREV16: return "ARMISD::VREV16";
674 case ARMISD::VZIP: return "ARMISD::VZIP";
675 case ARMISD::VUZP: return "ARMISD::VUZP";
676 case ARMISD::VTRN: return "ARMISD::VTRN";
677 case ARMISD::VMULLs: return "ARMISD::VMULLs";
678 case ARMISD::VMULLu: return "ARMISD::VMULLu";
679 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
680 case ARMISD::FMAX: return "ARMISD::FMAX";
681 case ARMISD::FMIN: return "ARMISD::FMIN";
682 case ARMISD::BFI: return "ARMISD::BFI";
686 /// getRegClassFor - Return the register class that should be used for the
687 /// specified value type.
688 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
689 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
690 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
691 // load / store 4 to 8 consecutive D registers.
692 if (Subtarget->hasNEON()) {
693 if (VT == MVT::v4i64)
694 return ARM::QQPRRegisterClass;
695 else if (VT == MVT::v8i64)
696 return ARM::QQQQPRRegisterClass;
698 return TargetLowering::getRegClassFor(VT);
701 // Create a fast isel object.
703 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
704 return ARM::createFastISel(funcInfo);
707 /// getFunctionAlignment - Return the Log2 alignment of this function.
708 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
709 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
712 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
713 /// be used for loads / stores from the global.
714 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
715 return (Subtarget->isThumb1Only() ? 127 : 4095);
718 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
719 unsigned NumVals = N->getNumValues();
721 return Sched::RegPressure;
723 for (unsigned i = 0; i != NumVals; ++i) {
724 EVT VT = N->getValueType(i);
725 if (VT.isFloatingPoint() || VT.isVector())
726 return Sched::Latency;
729 if (!N->isMachineOpcode())
730 return Sched::RegPressure;
732 // Load are scheduled for latency even if there instruction itinerary
734 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
735 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
737 return Sched::Latency;
739 if (!Itins->isEmpty() && Itins->getStageLatency(TID.getSchedClass()) > 2)
740 return Sched::Latency;
741 return Sched::RegPressure;
745 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
746 MachineFunction &MF) const {
747 switch (RC->getID()) {
750 case ARM::tGPRRegClassID:
751 return RegInfo->hasFP(MF) ? 4 : 5;
752 case ARM::GPRRegClassID: {
753 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
754 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
756 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
757 case ARM::DPRRegClassID:
762 //===----------------------------------------------------------------------===//
764 //===----------------------------------------------------------------------===//
766 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
767 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
769 default: llvm_unreachable("Unknown condition code!");
770 case ISD::SETNE: return ARMCC::NE;
771 case ISD::SETEQ: return ARMCC::EQ;
772 case ISD::SETGT: return ARMCC::GT;
773 case ISD::SETGE: return ARMCC::GE;
774 case ISD::SETLT: return ARMCC::LT;
775 case ISD::SETLE: return ARMCC::LE;
776 case ISD::SETUGT: return ARMCC::HI;
777 case ISD::SETUGE: return ARMCC::HS;
778 case ISD::SETULT: return ARMCC::LO;
779 case ISD::SETULE: return ARMCC::LS;
783 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
784 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
785 ARMCC::CondCodes &CondCode2) {
786 CondCode2 = ARMCC::AL;
788 default: llvm_unreachable("Unknown FP condition!");
790 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
792 case ISD::SETOGT: CondCode = ARMCC::GT; break;
794 case ISD::SETOGE: CondCode = ARMCC::GE; break;
795 case ISD::SETOLT: CondCode = ARMCC::MI; break;
796 case ISD::SETOLE: CondCode = ARMCC::LS; break;
797 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
798 case ISD::SETO: CondCode = ARMCC::VC; break;
799 case ISD::SETUO: CondCode = ARMCC::VS; break;
800 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
801 case ISD::SETUGT: CondCode = ARMCC::HI; break;
802 case ISD::SETUGE: CondCode = ARMCC::PL; break;
804 case ISD::SETULT: CondCode = ARMCC::LT; break;
806 case ISD::SETULE: CondCode = ARMCC::LE; break;
808 case ISD::SETUNE: CondCode = ARMCC::NE; break;
812 //===----------------------------------------------------------------------===//
813 // Calling Convention Implementation
814 //===----------------------------------------------------------------------===//
816 #include "ARMGenCallingConv.inc"
818 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
819 /// given CallingConvention value.
820 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
822 bool isVarArg) const {
825 llvm_unreachable("Unsupported calling convention");
827 case CallingConv::Fast:
828 // Use target triple & subtarget features to do actual dispatch.
829 if (Subtarget->isAAPCS_ABI()) {
830 if (Subtarget->hasVFP2() &&
831 FloatABIType == FloatABI::Hard && !isVarArg)
832 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
834 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
836 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
837 case CallingConv::ARM_AAPCS_VFP:
838 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
839 case CallingConv::ARM_AAPCS:
840 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
841 case CallingConv::ARM_APCS:
842 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
846 /// LowerCallResult - Lower the result values of a call into the
847 /// appropriate copies out of appropriate physical registers.
849 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
850 CallingConv::ID CallConv, bool isVarArg,
851 const SmallVectorImpl<ISD::InputArg> &Ins,
852 DebugLoc dl, SelectionDAG &DAG,
853 SmallVectorImpl<SDValue> &InVals) const {
855 // Assign locations to each value returned by this call.
856 SmallVector<CCValAssign, 16> RVLocs;
857 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
858 RVLocs, *DAG.getContext());
859 CCInfo.AnalyzeCallResult(Ins,
860 CCAssignFnForNode(CallConv, /* Return*/ true,
863 // Copy all of the result registers out of their specified physreg.
864 for (unsigned i = 0; i != RVLocs.size(); ++i) {
865 CCValAssign VA = RVLocs[i];
868 if (VA.needsCustom()) {
869 // Handle f64 or half of a v2f64.
870 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
872 Chain = Lo.getValue(1);
873 InFlag = Lo.getValue(2);
874 VA = RVLocs[++i]; // skip ahead to next loc
875 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
877 Chain = Hi.getValue(1);
878 InFlag = Hi.getValue(2);
879 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
881 if (VA.getLocVT() == MVT::v2f64) {
882 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
883 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
884 DAG.getConstant(0, MVT::i32));
886 VA = RVLocs[++i]; // skip ahead to next loc
887 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
888 Chain = Lo.getValue(1);
889 InFlag = Lo.getValue(2);
890 VA = RVLocs[++i]; // skip ahead to next loc
891 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
892 Chain = Hi.getValue(1);
893 InFlag = Hi.getValue(2);
894 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
895 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
896 DAG.getConstant(1, MVT::i32));
899 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
901 Chain = Val.getValue(1);
902 InFlag = Val.getValue(2);
905 switch (VA.getLocInfo()) {
906 default: llvm_unreachable("Unknown loc info!");
907 case CCValAssign::Full: break;
908 case CCValAssign::BCvt:
909 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
913 InVals.push_back(Val);
919 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
920 /// by "Src" to address "Dst" of size "Size". Alignment information is
921 /// specified by the specific parameter attribute. The copy will be passed as
922 /// a byval function parameter.
923 /// Sometimes what we are copying is the end of a larger object, the part that
924 /// does not fit in registers.
926 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
927 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
929 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
930 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
931 /*isVolatile=*/false, /*AlwaysInline=*/false,
932 MachinePointerInfo(0), MachinePointerInfo(0));
935 /// LowerMemOpCallTo - Store the argument to the stack.
937 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
938 SDValue StackPtr, SDValue Arg,
939 DebugLoc dl, SelectionDAG &DAG,
940 const CCValAssign &VA,
941 ISD::ArgFlagsTy Flags) const {
942 unsigned LocMemOffset = VA.getLocMemOffset();
943 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
944 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
946 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
948 return DAG.getStore(Chain, dl, Arg, PtrOff,
949 MachinePointerInfo::getStack(LocMemOffset),
953 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
954 SDValue Chain, SDValue &Arg,
955 RegsToPassVector &RegsToPass,
956 CCValAssign &VA, CCValAssign &NextVA,
958 SmallVector<SDValue, 8> &MemOpChains,
959 ISD::ArgFlagsTy Flags) const {
961 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
962 DAG.getVTList(MVT::i32, MVT::i32), Arg);
963 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
965 if (NextVA.isRegLoc())
966 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
968 assert(NextVA.isMemLoc());
969 if (StackPtr.getNode() == 0)
970 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
972 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
978 /// LowerCall - Lowering a call into a callseq_start <-
979 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
982 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
983 CallingConv::ID CallConv, bool isVarArg,
985 const SmallVectorImpl<ISD::OutputArg> &Outs,
986 const SmallVectorImpl<SDValue> &OutVals,
987 const SmallVectorImpl<ISD::InputArg> &Ins,
988 DebugLoc dl, SelectionDAG &DAG,
989 SmallVectorImpl<SDValue> &InVals) const {
990 MachineFunction &MF = DAG.getMachineFunction();
991 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
992 bool IsSibCall = false;
993 // Temporarily disable tail calls so things don't break.
994 if (!EnableARMTailCalls)
997 // Check if it's really possible to do a tail call.
998 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
999 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1000 Outs, OutVals, Ins, DAG);
1001 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1002 // detected sibcalls.
1009 // Analyze operands of the call, assigning locations to each operand.
1010 SmallVector<CCValAssign, 16> ArgLocs;
1011 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1013 CCInfo.AnalyzeCallOperands(Outs,
1014 CCAssignFnForNode(CallConv, /* Return*/ false,
1017 // Get a count of how many bytes are to be pushed on the stack.
1018 unsigned NumBytes = CCInfo.getNextStackOffset();
1020 // For tail calls, memory operands are available in our caller's stack.
1024 // Adjust the stack pointer for the new arguments...
1025 // These operations are automatically eliminated by the prolog/epilog pass
1027 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1029 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1031 RegsToPassVector RegsToPass;
1032 SmallVector<SDValue, 8> MemOpChains;
1034 // Walk the register/memloc assignments, inserting copies/loads. In the case
1035 // of tail call optimization, arguments are handled later.
1036 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1038 ++i, ++realArgIdx) {
1039 CCValAssign &VA = ArgLocs[i];
1040 SDValue Arg = OutVals[realArgIdx];
1041 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1043 // Promote the value if needed.
1044 switch (VA.getLocInfo()) {
1045 default: llvm_unreachable("Unknown loc info!");
1046 case CCValAssign::Full: break;
1047 case CCValAssign::SExt:
1048 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1050 case CCValAssign::ZExt:
1051 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1053 case CCValAssign::AExt:
1054 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1056 case CCValAssign::BCvt:
1057 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1061 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1062 if (VA.needsCustom()) {
1063 if (VA.getLocVT() == MVT::v2f64) {
1064 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1065 DAG.getConstant(0, MVT::i32));
1066 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1067 DAG.getConstant(1, MVT::i32));
1069 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1070 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1072 VA = ArgLocs[++i]; // skip ahead to next loc
1073 if (VA.isRegLoc()) {
1074 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1075 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1077 assert(VA.isMemLoc());
1079 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1080 dl, DAG, VA, Flags));
1083 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1084 StackPtr, MemOpChains, Flags);
1086 } else if (VA.isRegLoc()) {
1087 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1088 } else if (!IsSibCall) {
1089 assert(VA.isMemLoc());
1091 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1092 dl, DAG, VA, Flags));
1096 if (!MemOpChains.empty())
1097 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1098 &MemOpChains[0], MemOpChains.size());
1100 // Build a sequence of copy-to-reg nodes chained together with token chain
1101 // and flag operands which copy the outgoing args into the appropriate regs.
1103 // Tail call byval lowering might overwrite argument registers so in case of
1104 // tail call optimization the copies to registers are lowered later.
1106 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1107 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1108 RegsToPass[i].second, InFlag);
1109 InFlag = Chain.getValue(1);
1112 // For tail calls lower the arguments to the 'real' stack slot.
1114 // Force all the incoming stack arguments to be loaded from the stack
1115 // before any new outgoing arguments are stored to the stack, because the
1116 // outgoing stack slots may alias the incoming argument stack slots, and
1117 // the alias isn't otherwise explicit. This is slightly more conservative
1118 // than necessary, because it means that each store effectively depends
1119 // on every argument instead of just those arguments it would clobber.
1121 // Do not flag preceeding copytoreg stuff together with the following stuff.
1123 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1124 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1125 RegsToPass[i].second, InFlag);
1126 InFlag = Chain.getValue(1);
1131 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1132 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1133 // node so that legalize doesn't hack it.
1134 bool isDirect = false;
1135 bool isARMFunc = false;
1136 bool isLocalARMFunc = false;
1137 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1139 if (EnableARMLongCalls) {
1140 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1141 && "long-calls with non-static relocation model!");
1142 // Handle a global address or an external symbol. If it's not one of
1143 // those, the target's already in a register, so we don't need to do
1145 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1146 const GlobalValue *GV = G->getGlobal();
1147 // Create a constant pool entry for the callee address
1148 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1149 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1152 // Get the address of the callee into a register
1153 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1154 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1155 Callee = DAG.getLoad(getPointerTy(), dl,
1156 DAG.getEntryNode(), CPAddr,
1157 MachinePointerInfo::getConstantPool(),
1159 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1160 const char *Sym = S->getSymbol();
1162 // Create a constant pool entry for the callee address
1163 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1164 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1165 Sym, ARMPCLabelIndex, 0);
1166 // Get the address of the callee into a register
1167 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1168 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1169 Callee = DAG.getLoad(getPointerTy(), dl,
1170 DAG.getEntryNode(), CPAddr,
1171 MachinePointerInfo::getConstantPool(),
1174 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1175 const GlobalValue *GV = G->getGlobal();
1177 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1178 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1179 getTargetMachine().getRelocationModel() != Reloc::Static;
1180 isARMFunc = !Subtarget->isThumb() || isStub;
1181 // ARM call to a local ARM function is predicable.
1182 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1183 // tBX takes a register source operand.
1184 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1185 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1186 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1189 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1190 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1191 Callee = DAG.getLoad(getPointerTy(), dl,
1192 DAG.getEntryNode(), CPAddr,
1193 MachinePointerInfo::getConstantPool(),
1195 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1196 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1197 getPointerTy(), Callee, PICLabel);
1199 // On ELF targets for PIC code, direct calls should go through the PLT
1200 unsigned OpFlags = 0;
1201 if (Subtarget->isTargetELF() &&
1202 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1203 OpFlags = ARMII::MO_PLT;
1204 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1206 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1208 bool isStub = Subtarget->isTargetDarwin() &&
1209 getTargetMachine().getRelocationModel() != Reloc::Static;
1210 isARMFunc = !Subtarget->isThumb() || isStub;
1211 // tBX takes a register source operand.
1212 const char *Sym = S->getSymbol();
1213 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1214 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1215 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1216 Sym, ARMPCLabelIndex, 4);
1217 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1218 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1219 Callee = DAG.getLoad(getPointerTy(), dl,
1220 DAG.getEntryNode(), CPAddr,
1221 MachinePointerInfo::getConstantPool(),
1223 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1224 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1225 getPointerTy(), Callee, PICLabel);
1227 unsigned OpFlags = 0;
1228 // On ELF targets for PIC code, direct calls should go through the PLT
1229 if (Subtarget->isTargetELF() &&
1230 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1231 OpFlags = ARMII::MO_PLT;
1232 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1236 // FIXME: handle tail calls differently.
1238 if (Subtarget->isThumb()) {
1239 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1240 CallOpc = ARMISD::CALL_NOLINK;
1242 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1244 CallOpc = (isDirect || Subtarget->hasV5TOps())
1245 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1246 : ARMISD::CALL_NOLINK;
1249 std::vector<SDValue> Ops;
1250 Ops.push_back(Chain);
1251 Ops.push_back(Callee);
1253 // Add argument registers to the end of the list so that they are known live
1255 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1256 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1257 RegsToPass[i].second.getValueType()));
1259 if (InFlag.getNode())
1260 Ops.push_back(InFlag);
1262 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1264 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1266 // Returns a chain and a flag for retval copy to use.
1267 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1268 InFlag = Chain.getValue(1);
1270 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1271 DAG.getIntPtrConstant(0, true), InFlag);
1273 InFlag = Chain.getValue(1);
1275 // Handle result values, copying them out of physregs into vregs that we
1277 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1281 /// MatchingStackOffset - Return true if the given stack call argument is
1282 /// already available in the same position (relatively) of the caller's
1283 /// incoming argument stack.
1285 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1286 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1287 const ARMInstrInfo *TII) {
1288 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1290 if (Arg.getOpcode() == ISD::CopyFromReg) {
1291 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1292 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1294 MachineInstr *Def = MRI->getVRegDef(VR);
1297 if (!Flags.isByVal()) {
1298 if (!TII->isLoadFromStackSlot(Def, FI))
1303 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1304 if (Flags.isByVal())
1305 // ByVal argument is passed in as a pointer but it's now being
1306 // dereferenced. e.g.
1307 // define @foo(%struct.X* %A) {
1308 // tail call @bar(%struct.X* byval %A)
1311 SDValue Ptr = Ld->getBasePtr();
1312 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1315 FI = FINode->getIndex();
1319 assert(FI != INT_MAX);
1320 if (!MFI->isFixedObjectIndex(FI))
1322 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1325 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1326 /// for tail call optimization. Targets which want to do tail call
1327 /// optimization should implement this function.
1329 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1330 CallingConv::ID CalleeCC,
1332 bool isCalleeStructRet,
1333 bool isCallerStructRet,
1334 const SmallVectorImpl<ISD::OutputArg> &Outs,
1335 const SmallVectorImpl<SDValue> &OutVals,
1336 const SmallVectorImpl<ISD::InputArg> &Ins,
1337 SelectionDAG& DAG) const {
1338 const Function *CallerF = DAG.getMachineFunction().getFunction();
1339 CallingConv::ID CallerCC = CallerF->getCallingConv();
1340 bool CCMatch = CallerCC == CalleeCC;
1342 // Look for obvious safe cases to perform tail call optimization that do not
1343 // require ABI changes. This is what gcc calls sibcall.
1345 // Do not sibcall optimize vararg calls unless the call site is not passing
1347 if (isVarArg && !Outs.empty())
1350 // Also avoid sibcall optimization if either caller or callee uses struct
1351 // return semantics.
1352 if (isCalleeStructRet || isCallerStructRet)
1355 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1356 // emitEpilogue is not ready for them.
1357 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1358 // LR. This means if we need to reload LR, it takes an extra instructions,
1359 // which outweighs the value of the tail call; but here we don't know yet
1360 // whether LR is going to be used. Probably the right approach is to
1361 // generate the tail call here and turn it back into CALL/RET in
1362 // emitEpilogue if LR is used.
1363 if (Subtarget->isThumb1Only())
1366 // For the moment, we can only do this to functions defined in this
1367 // compilation, or to indirect calls. A Thumb B to an ARM function,
1368 // or vice versa, is not easily fixed up in the linker unlike BL.
1369 // (We could do this by loading the address of the callee into a register;
1370 // that is an extra instruction over the direct call and burns a register
1371 // as well, so is not likely to be a win.)
1373 // It might be safe to remove this restriction on non-Darwin.
1375 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1376 // but we need to make sure there are enough registers; the only valid
1377 // registers are the 4 used for parameters. We don't currently do this
1379 if (isa<ExternalSymbolSDNode>(Callee))
1382 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1383 const GlobalValue *GV = G->getGlobal();
1384 if (GV->isDeclaration() || GV->isWeakForLinker())
1388 // If the calling conventions do not match, then we'd better make sure the
1389 // results are returned in the same way as what the caller expects.
1391 SmallVector<CCValAssign, 16> RVLocs1;
1392 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1393 RVLocs1, *DAG.getContext());
1394 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1396 SmallVector<CCValAssign, 16> RVLocs2;
1397 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1398 RVLocs2, *DAG.getContext());
1399 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1401 if (RVLocs1.size() != RVLocs2.size())
1403 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1404 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1406 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1408 if (RVLocs1[i].isRegLoc()) {
1409 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1412 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1418 // If the callee takes no arguments then go on to check the results of the
1420 if (!Outs.empty()) {
1421 // Check if stack adjustment is needed. For now, do not do this if any
1422 // argument is passed on the stack.
1423 SmallVector<CCValAssign, 16> ArgLocs;
1424 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1425 ArgLocs, *DAG.getContext());
1426 CCInfo.AnalyzeCallOperands(Outs,
1427 CCAssignFnForNode(CalleeCC, false, isVarArg));
1428 if (CCInfo.getNextStackOffset()) {
1429 MachineFunction &MF = DAG.getMachineFunction();
1431 // Check if the arguments are already laid out in the right way as
1432 // the caller's fixed stack objects.
1433 MachineFrameInfo *MFI = MF.getFrameInfo();
1434 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1435 const ARMInstrInfo *TII =
1436 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1437 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1439 ++i, ++realArgIdx) {
1440 CCValAssign &VA = ArgLocs[i];
1441 EVT RegVT = VA.getLocVT();
1442 SDValue Arg = OutVals[realArgIdx];
1443 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1444 if (VA.getLocInfo() == CCValAssign::Indirect)
1446 if (VA.needsCustom()) {
1447 // f64 and vector types are split into multiple registers or
1448 // register/stack-slot combinations. The types will not match
1449 // the registers; give up on memory f64 refs until we figure
1450 // out what to do about this.
1453 if (!ArgLocs[++i].isRegLoc())
1455 if (RegVT == MVT::v2f64) {
1456 if (!ArgLocs[++i].isRegLoc())
1458 if (!ArgLocs[++i].isRegLoc())
1461 } else if (!VA.isRegLoc()) {
1462 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1474 ARMTargetLowering::LowerReturn(SDValue Chain,
1475 CallingConv::ID CallConv, bool isVarArg,
1476 const SmallVectorImpl<ISD::OutputArg> &Outs,
1477 const SmallVectorImpl<SDValue> &OutVals,
1478 DebugLoc dl, SelectionDAG &DAG) const {
1480 // CCValAssign - represent the assignment of the return value to a location.
1481 SmallVector<CCValAssign, 16> RVLocs;
1483 // CCState - Info about the registers and stack slots.
1484 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1487 // Analyze outgoing return values.
1488 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1491 // If this is the first return lowered for this function, add
1492 // the regs to the liveout set for the function.
1493 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1494 for (unsigned i = 0; i != RVLocs.size(); ++i)
1495 if (RVLocs[i].isRegLoc())
1496 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1501 // Copy the result values into the output registers.
1502 for (unsigned i = 0, realRVLocIdx = 0;
1504 ++i, ++realRVLocIdx) {
1505 CCValAssign &VA = RVLocs[i];
1506 assert(VA.isRegLoc() && "Can only return in registers!");
1508 SDValue Arg = OutVals[realRVLocIdx];
1510 switch (VA.getLocInfo()) {
1511 default: llvm_unreachable("Unknown loc info!");
1512 case CCValAssign::Full: break;
1513 case CCValAssign::BCvt:
1514 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1518 if (VA.needsCustom()) {
1519 if (VA.getLocVT() == MVT::v2f64) {
1520 // Extract the first half and return it in two registers.
1521 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1522 DAG.getConstant(0, MVT::i32));
1523 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1524 DAG.getVTList(MVT::i32, MVT::i32), Half);
1526 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1527 Flag = Chain.getValue(1);
1528 VA = RVLocs[++i]; // skip ahead to next loc
1529 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1530 HalfGPRs.getValue(1), Flag);
1531 Flag = Chain.getValue(1);
1532 VA = RVLocs[++i]; // skip ahead to next loc
1534 // Extract the 2nd half and fall through to handle it as an f64 value.
1535 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1536 DAG.getConstant(1, MVT::i32));
1538 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1540 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1541 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1542 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1543 Flag = Chain.getValue(1);
1544 VA = RVLocs[++i]; // skip ahead to next loc
1545 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1548 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1550 // Guarantee that all emitted copies are
1551 // stuck together, avoiding something bad.
1552 Flag = Chain.getValue(1);
1557 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1559 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1564 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1565 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1566 // one of the above mentioned nodes. It has to be wrapped because otherwise
1567 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1568 // be used to form addressing mode. These wrapped nodes will be selected
1570 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1571 EVT PtrVT = Op.getValueType();
1572 // FIXME there is no actual debug info here
1573 DebugLoc dl = Op.getDebugLoc();
1574 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1576 if (CP->isMachineConstantPoolEntry())
1577 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1578 CP->getAlignment());
1580 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1581 CP->getAlignment());
1582 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1585 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1586 return MachineJumpTableInfo::EK_Inline;
1589 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1590 SelectionDAG &DAG) const {
1591 MachineFunction &MF = DAG.getMachineFunction();
1592 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1593 unsigned ARMPCLabelIndex = 0;
1594 DebugLoc DL = Op.getDebugLoc();
1595 EVT PtrVT = getPointerTy();
1596 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1597 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1599 if (RelocM == Reloc::Static) {
1600 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1602 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1603 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1604 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1605 ARMCP::CPBlockAddress,
1607 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1609 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1610 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1611 MachinePointerInfo::getConstantPool(),
1613 if (RelocM == Reloc::Static)
1615 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1616 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1619 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1621 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1622 SelectionDAG &DAG) const {
1623 DebugLoc dl = GA->getDebugLoc();
1624 EVT PtrVT = getPointerTy();
1625 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1626 MachineFunction &MF = DAG.getMachineFunction();
1627 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1628 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1629 ARMConstantPoolValue *CPV =
1630 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1631 ARMCP::CPValue, PCAdj, "tlsgd", true);
1632 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1633 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1634 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1635 MachinePointerInfo::getConstantPool(),
1637 SDValue Chain = Argument.getValue(1);
1639 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1640 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1642 // call __tls_get_addr.
1645 Entry.Node = Argument;
1646 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1647 Args.push_back(Entry);
1648 // FIXME: is there useful debug info available here?
1649 std::pair<SDValue, SDValue> CallResult =
1650 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1651 false, false, false, false,
1652 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1653 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1654 return CallResult.first;
1657 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1658 // "local exec" model.
1660 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1661 SelectionDAG &DAG) const {
1662 const GlobalValue *GV = GA->getGlobal();
1663 DebugLoc dl = GA->getDebugLoc();
1665 SDValue Chain = DAG.getEntryNode();
1666 EVT PtrVT = getPointerTy();
1667 // Get the Thread Pointer
1668 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1670 if (GV->isDeclaration()) {
1671 MachineFunction &MF = DAG.getMachineFunction();
1672 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1673 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1674 // Initial exec model.
1675 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1676 ARMConstantPoolValue *CPV =
1677 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1678 ARMCP::CPValue, PCAdj, "gottpoff", true);
1679 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1680 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1681 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1682 MachinePointerInfo::getConstantPool(),
1684 Chain = Offset.getValue(1);
1686 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1687 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1689 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1690 MachinePointerInfo::getConstantPool(),
1694 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1695 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1696 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1697 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1698 MachinePointerInfo::getConstantPool(),
1702 // The address of the thread local variable is the add of the thread
1703 // pointer with the offset of the variable.
1704 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1708 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1709 // TODO: implement the "local dynamic" model
1710 assert(Subtarget->isTargetELF() &&
1711 "TLS not implemented for non-ELF targets");
1712 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1713 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1714 // otherwise use the "Local Exec" TLS Model
1715 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1716 return LowerToTLSGeneralDynamicModel(GA, DAG);
1718 return LowerToTLSExecModels(GA, DAG);
1721 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1722 SelectionDAG &DAG) const {
1723 EVT PtrVT = getPointerTy();
1724 DebugLoc dl = Op.getDebugLoc();
1725 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1726 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1727 if (RelocM == Reloc::PIC_) {
1728 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1729 ARMConstantPoolValue *CPV =
1730 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1731 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1732 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1733 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1735 MachinePointerInfo::getConstantPool(),
1737 SDValue Chain = Result.getValue(1);
1738 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1739 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1741 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1742 MachinePointerInfo::getGOT(), false, false, 0);
1745 // If we have T2 ops, we can materialize the address directly via movt/movw
1746 // pair. This is always cheaper.
1747 if (Subtarget->useMovt()) {
1748 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1749 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1751 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1752 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1753 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1754 MachinePointerInfo::getConstantPool(),
1760 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1761 SelectionDAG &DAG) const {
1762 MachineFunction &MF = DAG.getMachineFunction();
1763 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1764 unsigned ARMPCLabelIndex = 0;
1765 EVT PtrVT = getPointerTy();
1766 DebugLoc dl = Op.getDebugLoc();
1767 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1768 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1770 if (RelocM == Reloc::Static)
1771 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1773 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1774 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1775 ARMConstantPoolValue *CPV =
1776 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1777 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1779 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1781 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1782 MachinePointerInfo::getConstantPool(),
1784 SDValue Chain = Result.getValue(1);
1786 if (RelocM == Reloc::PIC_) {
1787 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1788 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1791 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1792 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
1798 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1799 SelectionDAG &DAG) const {
1800 assert(Subtarget->isTargetELF() &&
1801 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1802 MachineFunction &MF = DAG.getMachineFunction();
1803 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1804 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1805 EVT PtrVT = getPointerTy();
1806 DebugLoc dl = Op.getDebugLoc();
1807 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1808 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1809 "_GLOBAL_OFFSET_TABLE_",
1810 ARMPCLabelIndex, PCAdj);
1811 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1812 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1813 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1814 MachinePointerInfo::getConstantPool(),
1816 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1817 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1821 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1822 DebugLoc dl = Op.getDebugLoc();
1823 SDValue Val = DAG.getConstant(0, MVT::i32);
1824 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1825 Op.getOperand(1), Val);
1829 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1830 DebugLoc dl = Op.getDebugLoc();
1831 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1832 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1836 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1837 const ARMSubtarget *Subtarget) const {
1838 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1839 DebugLoc dl = Op.getDebugLoc();
1841 default: return SDValue(); // Don't custom lower most intrinsics.
1842 case Intrinsic::arm_thread_pointer: {
1843 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1844 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1846 case Intrinsic::eh_sjlj_lsda: {
1847 MachineFunction &MF = DAG.getMachineFunction();
1848 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1849 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1850 EVT PtrVT = getPointerTy();
1851 DebugLoc dl = Op.getDebugLoc();
1852 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1854 unsigned PCAdj = (RelocM != Reloc::PIC_)
1855 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1856 ARMConstantPoolValue *CPV =
1857 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1858 ARMCP::CPLSDA, PCAdj);
1859 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1860 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1862 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1863 MachinePointerInfo::getConstantPool(),
1866 if (RelocM == Reloc::PIC_) {
1867 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1868 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1875 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1876 const ARMSubtarget *Subtarget) {
1877 DebugLoc dl = Op.getDebugLoc();
1878 SDValue Op5 = Op.getOperand(5);
1879 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1880 // Some subtargets which have dmb and dsb instructions can handle barriers
1881 // directly. Some ARMv6 cpus can support them with the help of mcr
1882 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1884 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1885 if (Subtarget->hasDataBarrier())
1886 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1888 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
1889 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1890 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1891 DAG.getConstant(0, MVT::i32));
1895 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1896 MachineFunction &MF = DAG.getMachineFunction();
1897 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1899 // vastart just stores the address of the VarArgsFrameIndex slot into the
1900 // memory location argument.
1901 DebugLoc dl = Op.getDebugLoc();
1902 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1903 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1904 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1905 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1906 MachinePointerInfo(SV), false, false, 0);
1910 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1911 SDValue &Root, SelectionDAG &DAG,
1912 DebugLoc dl) const {
1913 MachineFunction &MF = DAG.getMachineFunction();
1914 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1916 TargetRegisterClass *RC;
1917 if (AFI->isThumb1OnlyFunction())
1918 RC = ARM::tGPRRegisterClass;
1920 RC = ARM::GPRRegisterClass;
1922 // Transform the arguments stored in physical registers into virtual ones.
1923 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1924 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1927 if (NextVA.isMemLoc()) {
1928 MachineFrameInfo *MFI = MF.getFrameInfo();
1929 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
1931 // Create load node to retrieve arguments from the stack.
1932 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1933 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1934 MachinePointerInfo::getFixedStack(FI),
1937 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1938 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1941 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1945 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1946 CallingConv::ID CallConv, bool isVarArg,
1947 const SmallVectorImpl<ISD::InputArg>
1949 DebugLoc dl, SelectionDAG &DAG,
1950 SmallVectorImpl<SDValue> &InVals)
1953 MachineFunction &MF = DAG.getMachineFunction();
1954 MachineFrameInfo *MFI = MF.getFrameInfo();
1956 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1958 // Assign locations to all of the incoming arguments.
1959 SmallVector<CCValAssign, 16> ArgLocs;
1960 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1962 CCInfo.AnalyzeFormalArguments(Ins,
1963 CCAssignFnForNode(CallConv, /* Return*/ false,
1966 SmallVector<SDValue, 16> ArgValues;
1968 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1969 CCValAssign &VA = ArgLocs[i];
1971 // Arguments stored in registers.
1972 if (VA.isRegLoc()) {
1973 EVT RegVT = VA.getLocVT();
1976 if (VA.needsCustom()) {
1977 // f64 and vector types are split up into multiple registers or
1978 // combinations of registers and stack slots.
1979 if (VA.getLocVT() == MVT::v2f64) {
1980 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1982 VA = ArgLocs[++i]; // skip ahead to next loc
1984 if (VA.isMemLoc()) {
1985 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
1986 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1987 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1988 MachinePointerInfo::getFixedStack(FI),
1991 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1994 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1995 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1996 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1997 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1998 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2000 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2003 TargetRegisterClass *RC;
2005 if (RegVT == MVT::f32)
2006 RC = ARM::SPRRegisterClass;
2007 else if (RegVT == MVT::f64)
2008 RC = ARM::DPRRegisterClass;
2009 else if (RegVT == MVT::v2f64)
2010 RC = ARM::QPRRegisterClass;
2011 else if (RegVT == MVT::i32)
2012 RC = (AFI->isThumb1OnlyFunction() ?
2013 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2015 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2017 // Transform the arguments in physical registers into virtual ones.
2018 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2019 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2022 // If this is an 8 or 16-bit value, it is really passed promoted
2023 // to 32 bits. Insert an assert[sz]ext to capture this, then
2024 // truncate to the right size.
2025 switch (VA.getLocInfo()) {
2026 default: llvm_unreachable("Unknown loc info!");
2027 case CCValAssign::Full: break;
2028 case CCValAssign::BCvt:
2029 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2031 case CCValAssign::SExt:
2032 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2033 DAG.getValueType(VA.getValVT()));
2034 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2036 case CCValAssign::ZExt:
2037 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2038 DAG.getValueType(VA.getValVT()));
2039 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2043 InVals.push_back(ArgValue);
2045 } else { // VA.isRegLoc()
2048 assert(VA.isMemLoc());
2049 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2051 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2052 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2054 // Create load nodes to retrieve arguments from the stack.
2055 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2056 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2057 MachinePointerInfo::getFixedStack(FI),
2064 static const unsigned GPRArgRegs[] = {
2065 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2068 unsigned NumGPRs = CCInfo.getFirstUnallocated
2069 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2071 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2072 unsigned VARegSize = (4 - NumGPRs) * 4;
2073 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2074 unsigned ArgOffset = CCInfo.getNextStackOffset();
2075 if (VARegSaveSize) {
2076 // If this function is vararg, store any remaining integer argument regs
2077 // to their spots on the stack so that they may be loaded by deferencing
2078 // the result of va_next.
2079 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2080 AFI->setVarArgsFrameIndex(
2081 MFI->CreateFixedObject(VARegSaveSize,
2082 ArgOffset + VARegSaveSize - VARegSize,
2084 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2087 SmallVector<SDValue, 4> MemOps;
2088 for (; NumGPRs < 4; ++NumGPRs) {
2089 TargetRegisterClass *RC;
2090 if (AFI->isThumb1OnlyFunction())
2091 RC = ARM::tGPRRegisterClass;
2093 RC = ARM::GPRRegisterClass;
2095 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2096 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2098 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2099 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2101 MemOps.push_back(Store);
2102 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2103 DAG.getConstant(4, getPointerTy()));
2105 if (!MemOps.empty())
2106 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2107 &MemOps[0], MemOps.size());
2109 // This will point to the next argument passed via stack.
2110 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2116 /// isFloatingPointZero - Return true if this is +0.0.
2117 static bool isFloatingPointZero(SDValue Op) {
2118 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2119 return CFP->getValueAPF().isPosZero();
2120 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2121 // Maybe this has already been legalized into the constant pool?
2122 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2123 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2124 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2125 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2126 return CFP->getValueAPF().isPosZero();
2132 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2133 /// the given operands.
2135 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2136 SDValue &ARMcc, SelectionDAG &DAG,
2137 DebugLoc dl) const {
2138 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2139 unsigned C = RHSC->getZExtValue();
2140 if (!isLegalICmpImmediate(C)) {
2141 // Constant does not fit, try adjusting it by one?
2146 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2147 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2148 RHS = DAG.getConstant(C-1, MVT::i32);
2153 if (C != 0 && isLegalICmpImmediate(C-1)) {
2154 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2155 RHS = DAG.getConstant(C-1, MVT::i32);
2160 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2161 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2162 RHS = DAG.getConstant(C+1, MVT::i32);
2167 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2168 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2169 RHS = DAG.getConstant(C+1, MVT::i32);
2176 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2177 ARMISD::NodeType CompareType;
2180 CompareType = ARMISD::CMP;
2185 CompareType = ARMISD::CMPZ;
2188 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2189 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2192 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2194 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2195 DebugLoc dl) const {
2197 if (!isFloatingPointZero(RHS))
2198 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2200 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2201 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2204 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2205 SDValue Cond = Op.getOperand(0);
2206 SDValue SelectTrue = Op.getOperand(1);
2207 SDValue SelectFalse = Op.getOperand(2);
2208 DebugLoc dl = Op.getDebugLoc();
2212 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2213 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2215 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2216 const ConstantSDNode *CMOVTrue =
2217 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2218 const ConstantSDNode *CMOVFalse =
2219 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2221 if (CMOVTrue && CMOVFalse) {
2222 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2223 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2227 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2229 False = SelectFalse;
2230 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2235 if (True.getNode() && False.getNode()) {
2236 EVT VT = Cond.getValueType();
2237 SDValue ARMcc = Cond.getOperand(2);
2238 SDValue CCR = Cond.getOperand(3);
2239 SDValue Cmp = Cond.getOperand(4);
2240 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2245 return DAG.getSelectCC(dl, Cond,
2246 DAG.getConstant(0, Cond.getValueType()),
2247 SelectTrue, SelectFalse, ISD::SETNE);
2250 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2251 EVT VT = Op.getValueType();
2252 SDValue LHS = Op.getOperand(0);
2253 SDValue RHS = Op.getOperand(1);
2254 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2255 SDValue TrueVal = Op.getOperand(2);
2256 SDValue FalseVal = Op.getOperand(3);
2257 DebugLoc dl = Op.getDebugLoc();
2259 if (LHS.getValueType() == MVT::i32) {
2261 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2262 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2263 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2266 ARMCC::CondCodes CondCode, CondCode2;
2267 FPCCToARMCC(CC, CondCode, CondCode2);
2269 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2270 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2271 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2272 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2274 if (CondCode2 != ARMCC::AL) {
2275 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2276 // FIXME: Needs another CMP because flag can have but one use.
2277 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2278 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2279 Result, TrueVal, ARMcc2, CCR, Cmp2);
2284 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2285 /// to morph to an integer compare sequence.
2286 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2287 const ARMSubtarget *Subtarget) {
2288 SDNode *N = Op.getNode();
2289 if (!N->hasOneUse())
2290 // Otherwise it requires moving the value from fp to integer registers.
2292 if (!N->getNumValues())
2294 EVT VT = Op.getValueType();
2295 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2296 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2297 // vmrs are very slow, e.g. cortex-a8.
2300 if (isFloatingPointZero(Op)) {
2304 return ISD::isNormalLoad(N);
2307 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2308 if (isFloatingPointZero(Op))
2309 return DAG.getConstant(0, MVT::i32);
2311 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2312 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2313 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2314 Ld->isVolatile(), Ld->isNonTemporal(),
2315 Ld->getAlignment());
2317 llvm_unreachable("Unknown VFP cmp argument!");
2320 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2321 SDValue &RetVal1, SDValue &RetVal2) {
2322 if (isFloatingPointZero(Op)) {
2323 RetVal1 = DAG.getConstant(0, MVT::i32);
2324 RetVal2 = DAG.getConstant(0, MVT::i32);
2328 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2329 SDValue Ptr = Ld->getBasePtr();
2330 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2331 Ld->getChain(), Ptr,
2332 Ld->getPointerInfo(),
2333 Ld->isVolatile(), Ld->isNonTemporal(),
2334 Ld->getAlignment());
2336 EVT PtrType = Ptr.getValueType();
2337 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2338 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2339 PtrType, Ptr, DAG.getConstant(4, PtrType));
2340 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2341 Ld->getChain(), NewPtr,
2342 Ld->getPointerInfo().getWithOffset(4),
2343 Ld->isVolatile(), Ld->isNonTemporal(),
2348 llvm_unreachable("Unknown VFP cmp argument!");
2351 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2352 /// f32 and even f64 comparisons to integer ones.
2354 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2355 SDValue Chain = Op.getOperand(0);
2356 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2357 SDValue LHS = Op.getOperand(2);
2358 SDValue RHS = Op.getOperand(3);
2359 SDValue Dest = Op.getOperand(4);
2360 DebugLoc dl = Op.getDebugLoc();
2362 bool SeenZero = false;
2363 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2364 canChangeToInt(RHS, SeenZero, Subtarget) &&
2365 // If one of the operand is zero, it's safe to ignore the NaN case since
2366 // we only care about equality comparisons.
2367 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2368 // If unsafe fp math optimization is enabled and there are no othter uses of
2369 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2370 // to an integer comparison.
2371 if (CC == ISD::SETOEQ)
2373 else if (CC == ISD::SETUNE)
2377 if (LHS.getValueType() == MVT::f32) {
2378 LHS = bitcastf32Toi32(LHS, DAG);
2379 RHS = bitcastf32Toi32(RHS, DAG);
2380 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2381 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2382 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2383 Chain, Dest, ARMcc, CCR, Cmp);
2388 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2389 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2390 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2391 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2392 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2393 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2394 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2400 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2401 SDValue Chain = Op.getOperand(0);
2402 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2403 SDValue LHS = Op.getOperand(2);
2404 SDValue RHS = Op.getOperand(3);
2405 SDValue Dest = Op.getOperand(4);
2406 DebugLoc dl = Op.getDebugLoc();
2408 if (LHS.getValueType() == MVT::i32) {
2410 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2411 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2412 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2413 Chain, Dest, ARMcc, CCR, Cmp);
2416 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2419 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2420 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2421 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2422 if (Result.getNode())
2426 ARMCC::CondCodes CondCode, CondCode2;
2427 FPCCToARMCC(CC, CondCode, CondCode2);
2429 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2430 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2431 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2432 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2433 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2434 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2435 if (CondCode2 != ARMCC::AL) {
2436 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2437 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2438 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2443 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2444 SDValue Chain = Op.getOperand(0);
2445 SDValue Table = Op.getOperand(1);
2446 SDValue Index = Op.getOperand(2);
2447 DebugLoc dl = Op.getDebugLoc();
2449 EVT PTy = getPointerTy();
2450 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2451 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2452 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2453 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2454 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2455 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2456 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2457 if (Subtarget->isThumb2()) {
2458 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2459 // which does another jump to the destination. This also makes it easier
2460 // to translate it to TBB / TBH later.
2461 // FIXME: This might not work if the function is extremely large.
2462 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2463 Addr, Op.getOperand(2), JTI, UId);
2465 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2466 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2467 MachinePointerInfo::getJumpTable(),
2469 Chain = Addr.getValue(1);
2470 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2471 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2473 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2474 MachinePointerInfo::getJumpTable(), false, false, 0);
2475 Chain = Addr.getValue(1);
2476 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2480 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2481 DebugLoc dl = Op.getDebugLoc();
2484 switch (Op.getOpcode()) {
2486 assert(0 && "Invalid opcode!");
2487 case ISD::FP_TO_SINT:
2488 Opc = ARMISD::FTOSI;
2490 case ISD::FP_TO_UINT:
2491 Opc = ARMISD::FTOUI;
2494 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2495 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2498 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2499 EVT VT = Op.getValueType();
2500 DebugLoc dl = Op.getDebugLoc();
2503 switch (Op.getOpcode()) {
2505 assert(0 && "Invalid opcode!");
2506 case ISD::SINT_TO_FP:
2507 Opc = ARMISD::SITOF;
2509 case ISD::UINT_TO_FP:
2510 Opc = ARMISD::UITOF;
2514 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2515 return DAG.getNode(Opc, dl, VT, Op);
2518 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2519 // Implement fcopysign with a fabs and a conditional fneg.
2520 SDValue Tmp0 = Op.getOperand(0);
2521 SDValue Tmp1 = Op.getOperand(1);
2522 DebugLoc dl = Op.getDebugLoc();
2523 EVT VT = Op.getValueType();
2524 EVT SrcVT = Tmp1.getValueType();
2525 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2526 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2527 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2528 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2529 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2530 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2533 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2534 MachineFunction &MF = DAG.getMachineFunction();
2535 MachineFrameInfo *MFI = MF.getFrameInfo();
2536 MFI->setReturnAddressIsTaken(true);
2538 EVT VT = Op.getValueType();
2539 DebugLoc dl = Op.getDebugLoc();
2540 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2542 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2543 SDValue Offset = DAG.getConstant(4, MVT::i32);
2544 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2545 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2546 MachinePointerInfo(), false, false, 0);
2549 // Return LR, which contains the return address. Mark it an implicit live-in.
2550 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2551 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2554 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2555 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2556 MFI->setFrameAddressIsTaken(true);
2558 EVT VT = Op.getValueType();
2559 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2560 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2561 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2562 ? ARM::R7 : ARM::R11;
2563 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2565 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2566 MachinePointerInfo(),
2571 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2572 /// expand a bit convert where either the source or destination type is i64 to
2573 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2574 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2575 /// vectors), since the legalizer won't know what to do with that.
2576 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2577 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2578 DebugLoc dl = N->getDebugLoc();
2579 SDValue Op = N->getOperand(0);
2581 // This function is only supposed to be called for i64 types, either as the
2582 // source or destination of the bit convert.
2583 EVT SrcVT = Op.getValueType();
2584 EVT DstVT = N->getValueType(0);
2585 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2586 "ExpandBIT_CONVERT called for non-i64 type");
2588 // Turn i64->f64 into VMOVDRR.
2589 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2590 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2591 DAG.getConstant(0, MVT::i32));
2592 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2593 DAG.getConstant(1, MVT::i32));
2594 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2595 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2598 // Turn f64->i64 into VMOVRRD.
2599 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2600 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2601 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2602 // Merge the pieces into a single i64 value.
2603 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2609 /// getZeroVector - Returns a vector of specified type with all zero elements.
2610 /// Zero vectors are used to represent vector negation and in those cases
2611 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2612 /// not support i64 elements, so sometimes the zero vectors will need to be
2613 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2615 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2616 assert(VT.isVector() && "Expected a vector type");
2617 // The canonical modified immediate encoding of a zero vector is....0!
2618 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2619 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2620 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2621 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2624 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2625 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2626 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2627 SelectionDAG &DAG) const {
2628 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2629 EVT VT = Op.getValueType();
2630 unsigned VTBits = VT.getSizeInBits();
2631 DebugLoc dl = Op.getDebugLoc();
2632 SDValue ShOpLo = Op.getOperand(0);
2633 SDValue ShOpHi = Op.getOperand(1);
2634 SDValue ShAmt = Op.getOperand(2);
2636 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2638 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2640 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2641 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2642 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2643 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2644 DAG.getConstant(VTBits, MVT::i32));
2645 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2646 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2647 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2649 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2650 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2652 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2653 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2656 SDValue Ops[2] = { Lo, Hi };
2657 return DAG.getMergeValues(Ops, 2, dl);
2660 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2661 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2662 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2663 SelectionDAG &DAG) const {
2664 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2665 EVT VT = Op.getValueType();
2666 unsigned VTBits = VT.getSizeInBits();
2667 DebugLoc dl = Op.getDebugLoc();
2668 SDValue ShOpLo = Op.getOperand(0);
2669 SDValue ShOpHi = Op.getOperand(1);
2670 SDValue ShAmt = Op.getOperand(2);
2673 assert(Op.getOpcode() == ISD::SHL_PARTS);
2674 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2675 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2676 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2677 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2678 DAG.getConstant(VTBits, MVT::i32));
2679 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2680 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2682 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2683 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2684 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2686 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2687 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2690 SDValue Ops[2] = { Lo, Hi };
2691 return DAG.getMergeValues(Ops, 2, dl);
2694 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2695 SelectionDAG &DAG) const {
2696 // The rounding mode is in bits 23:22 of the FPSCR.
2697 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2698 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2699 // so that the shift + and get folded into a bitfield extract.
2700 DebugLoc dl = Op.getDebugLoc();
2701 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2702 DAG.getConstant(Intrinsic::arm_get_fpscr,
2704 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2705 DAG.getConstant(1U << 22, MVT::i32));
2706 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2707 DAG.getConstant(22, MVT::i32));
2708 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2709 DAG.getConstant(3, MVT::i32));
2712 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2713 const ARMSubtarget *ST) {
2714 EVT VT = N->getValueType(0);
2715 DebugLoc dl = N->getDebugLoc();
2717 if (!ST->hasV6T2Ops())
2720 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2721 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2724 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2725 const ARMSubtarget *ST) {
2726 EVT VT = N->getValueType(0);
2727 DebugLoc dl = N->getDebugLoc();
2729 // Lower vector shifts on NEON to use VSHL.
2730 if (VT.isVector()) {
2731 assert(ST->hasNEON() && "unexpected vector shift");
2733 // Left shifts translate directly to the vshiftu intrinsic.
2734 if (N->getOpcode() == ISD::SHL)
2735 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2736 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2737 N->getOperand(0), N->getOperand(1));
2739 assert((N->getOpcode() == ISD::SRA ||
2740 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2742 // NEON uses the same intrinsics for both left and right shifts. For
2743 // right shifts, the shift amounts are negative, so negate the vector of
2745 EVT ShiftVT = N->getOperand(1).getValueType();
2746 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2747 getZeroVector(ShiftVT, DAG, dl),
2749 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2750 Intrinsic::arm_neon_vshifts :
2751 Intrinsic::arm_neon_vshiftu);
2752 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2753 DAG.getConstant(vshiftInt, MVT::i32),
2754 N->getOperand(0), NegatedCount);
2757 // We can get here for a node like i32 = ISD::SHL i32, i64
2761 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2762 "Unknown shift to lower!");
2764 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2765 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2766 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2769 // If we are in thumb mode, we don't have RRX.
2770 if (ST->isThumb1Only()) return SDValue();
2772 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2773 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2774 DAG.getConstant(0, MVT::i32));
2775 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2776 DAG.getConstant(1, MVT::i32));
2778 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2779 // captures the result into a carry flag.
2780 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2781 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2783 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2784 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2786 // Merge the pieces into a single i64 value.
2787 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2790 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2791 SDValue TmpOp0, TmpOp1;
2792 bool Invert = false;
2796 SDValue Op0 = Op.getOperand(0);
2797 SDValue Op1 = Op.getOperand(1);
2798 SDValue CC = Op.getOperand(2);
2799 EVT VT = Op.getValueType();
2800 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2801 DebugLoc dl = Op.getDebugLoc();
2803 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2804 switch (SetCCOpcode) {
2805 default: llvm_unreachable("Illegal FP comparison"); break;
2807 case ISD::SETNE: Invert = true; // Fallthrough
2809 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2811 case ISD::SETLT: Swap = true; // Fallthrough
2813 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2815 case ISD::SETLE: Swap = true; // Fallthrough
2817 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2818 case ISD::SETUGE: Swap = true; // Fallthrough
2819 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2820 case ISD::SETUGT: Swap = true; // Fallthrough
2821 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2822 case ISD::SETUEQ: Invert = true; // Fallthrough
2824 // Expand this to (OLT | OGT).
2828 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2829 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2831 case ISD::SETUO: Invert = true; // Fallthrough
2833 // Expand this to (OLT | OGE).
2837 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2838 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2842 // Integer comparisons.
2843 switch (SetCCOpcode) {
2844 default: llvm_unreachable("Illegal integer comparison"); break;
2845 case ISD::SETNE: Invert = true;
2846 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2847 case ISD::SETLT: Swap = true;
2848 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2849 case ISD::SETLE: Swap = true;
2850 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2851 case ISD::SETULT: Swap = true;
2852 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2853 case ISD::SETULE: Swap = true;
2854 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2857 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2858 if (Opc == ARMISD::VCEQ) {
2861 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2863 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2866 // Ignore bitconvert.
2867 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2868 AndOp = AndOp.getOperand(0);
2870 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2872 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2873 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2880 std::swap(Op0, Op1);
2882 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2885 Result = DAG.getNOT(dl, Result, VT);
2890 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2891 /// valid vector constant for a NEON instruction with a "modified immediate"
2892 /// operand (e.g., VMOV). If so, return the encoded value.
2893 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2894 unsigned SplatBitSize, SelectionDAG &DAG,
2895 EVT &VT, bool is128Bits, bool isVMOV) {
2896 unsigned OpCmode, Imm;
2898 // SplatBitSize is set to the smallest size that splats the vector, so a
2899 // zero vector will always have SplatBitSize == 8. However, NEON modified
2900 // immediate instructions others than VMOV do not support the 8-bit encoding
2901 // of a zero vector, and the default encoding of zero is supposed to be the
2906 switch (SplatBitSize) {
2910 // Any 1-byte value is OK. Op=0, Cmode=1110.
2911 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2914 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2918 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2919 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2920 if ((SplatBits & ~0xff) == 0) {
2921 // Value = 0x00nn: Op=x, Cmode=100x.
2926 if ((SplatBits & ~0xff00) == 0) {
2927 // Value = 0xnn00: Op=x, Cmode=101x.
2929 Imm = SplatBits >> 8;
2935 // NEON's 32-bit VMOV supports splat values where:
2936 // * only one byte is nonzero, or
2937 // * the least significant byte is 0xff and the second byte is nonzero, or
2938 // * the least significant 2 bytes are 0xff and the third is nonzero.
2939 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
2940 if ((SplatBits & ~0xff) == 0) {
2941 // Value = 0x000000nn: Op=x, Cmode=000x.
2946 if ((SplatBits & ~0xff00) == 0) {
2947 // Value = 0x0000nn00: Op=x, Cmode=001x.
2949 Imm = SplatBits >> 8;
2952 if ((SplatBits & ~0xff0000) == 0) {
2953 // Value = 0x00nn0000: Op=x, Cmode=010x.
2955 Imm = SplatBits >> 16;
2958 if ((SplatBits & ~0xff000000) == 0) {
2959 // Value = 0xnn000000: Op=x, Cmode=011x.
2961 Imm = SplatBits >> 24;
2965 if ((SplatBits & ~0xffff) == 0 &&
2966 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2967 // Value = 0x0000nnff: Op=x, Cmode=1100.
2969 Imm = SplatBits >> 8;
2974 if ((SplatBits & ~0xffffff) == 0 &&
2975 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2976 // Value = 0x00nnffff: Op=x, Cmode=1101.
2978 Imm = SplatBits >> 16;
2979 SplatBits |= 0xffff;
2983 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2984 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2985 // VMOV.I32. A (very) minor optimization would be to replicate the value
2986 // and fall through here to test for a valid 64-bit splat. But, then the
2987 // caller would also need to check and handle the change in size.
2993 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2994 uint64_t BitMask = 0xff;
2996 unsigned ImmMask = 1;
2998 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2999 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3002 } else if ((SplatBits & BitMask) != 0) {
3008 // Op=1, Cmode=1110.
3011 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3016 llvm_unreachable("unexpected size for isNEONModifiedImm");
3020 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3021 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3024 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3025 bool &ReverseVEXT, unsigned &Imm) {
3026 unsigned NumElts = VT.getVectorNumElements();
3027 ReverseVEXT = false;
3029 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3035 // If this is a VEXT shuffle, the immediate value is the index of the first
3036 // element. The other shuffle indices must be the successive elements after
3038 unsigned ExpectedElt = Imm;
3039 for (unsigned i = 1; i < NumElts; ++i) {
3040 // Increment the expected index. If it wraps around, it may still be
3041 // a VEXT but the source vectors must be swapped.
3043 if (ExpectedElt == NumElts * 2) {
3048 if (M[i] < 0) continue; // ignore UNDEF indices
3049 if (ExpectedElt != static_cast<unsigned>(M[i]))
3053 // Adjust the index value if the source operands will be swapped.
3060 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3061 /// instruction with the specified blocksize. (The order of the elements
3062 /// within each block of the vector is reversed.)
3063 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3064 unsigned BlockSize) {
3065 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3066 "Only possible block sizes for VREV are: 16, 32, 64");
3068 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3072 unsigned NumElts = VT.getVectorNumElements();
3073 unsigned BlockElts = M[0] + 1;
3074 // If the first shuffle index is UNDEF, be optimistic.
3076 BlockElts = BlockSize / EltSz;
3078 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3081 for (unsigned i = 0; i < NumElts; ++i) {
3082 if (M[i] < 0) continue; // ignore UNDEF indices
3083 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3090 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3091 unsigned &WhichResult) {
3092 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3096 unsigned NumElts = VT.getVectorNumElements();
3097 WhichResult = (M[0] == 0 ? 0 : 1);
3098 for (unsigned i = 0; i < NumElts; i += 2) {
3099 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3100 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3106 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3107 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3108 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3109 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3110 unsigned &WhichResult) {
3111 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3115 unsigned NumElts = VT.getVectorNumElements();
3116 WhichResult = (M[0] == 0 ? 0 : 1);
3117 for (unsigned i = 0; i < NumElts; i += 2) {
3118 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3119 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3125 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3126 unsigned &WhichResult) {
3127 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3131 unsigned NumElts = VT.getVectorNumElements();
3132 WhichResult = (M[0] == 0 ? 0 : 1);
3133 for (unsigned i = 0; i != NumElts; ++i) {
3134 if (M[i] < 0) continue; // ignore UNDEF indices
3135 if ((unsigned) M[i] != 2 * i + WhichResult)
3139 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3140 if (VT.is64BitVector() && EltSz == 32)
3146 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3147 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3148 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3149 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3150 unsigned &WhichResult) {
3151 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3155 unsigned Half = VT.getVectorNumElements() / 2;
3156 WhichResult = (M[0] == 0 ? 0 : 1);
3157 for (unsigned j = 0; j != 2; ++j) {
3158 unsigned Idx = WhichResult;
3159 for (unsigned i = 0; i != Half; ++i) {
3160 int MIdx = M[i + j * Half];
3161 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3167 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3168 if (VT.is64BitVector() && EltSz == 32)
3174 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3175 unsigned &WhichResult) {
3176 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3180 unsigned NumElts = VT.getVectorNumElements();
3181 WhichResult = (M[0] == 0 ? 0 : 1);
3182 unsigned Idx = WhichResult * NumElts / 2;
3183 for (unsigned i = 0; i != NumElts; i += 2) {
3184 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3185 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3190 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3191 if (VT.is64BitVector() && EltSz == 32)
3197 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3198 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3199 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3200 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3201 unsigned &WhichResult) {
3202 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3206 unsigned NumElts = VT.getVectorNumElements();
3207 WhichResult = (M[0] == 0 ? 0 : 1);
3208 unsigned Idx = WhichResult * NumElts / 2;
3209 for (unsigned i = 0; i != NumElts; i += 2) {
3210 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3211 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3216 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3217 if (VT.is64BitVector() && EltSz == 32)
3223 // If N is an integer constant that can be moved into a register in one
3224 // instruction, return an SDValue of such a constant (will become a MOV
3225 // instruction). Otherwise return null.
3226 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3227 const ARMSubtarget *ST, DebugLoc dl) {
3229 if (!isa<ConstantSDNode>(N))
3231 Val = cast<ConstantSDNode>(N)->getZExtValue();
3233 if (ST->isThumb1Only()) {
3234 if (Val <= 255 || ~Val <= 255)
3235 return DAG.getConstant(Val, MVT::i32);
3237 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3238 return DAG.getConstant(Val, MVT::i32);
3243 // If this is a case we can't handle, return null and let the default
3244 // expansion code take care of it.
3245 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3246 const ARMSubtarget *ST) {
3247 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3248 DebugLoc dl = Op.getDebugLoc();
3249 EVT VT = Op.getValueType();
3251 APInt SplatBits, SplatUndef;
3252 unsigned SplatBitSize;
3254 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3255 if (SplatBitSize <= 64) {
3256 // Check if an immediate VMOV works.
3258 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3259 SplatUndef.getZExtValue(), SplatBitSize,
3260 DAG, VmovVT, VT.is128BitVector(), true);
3261 if (Val.getNode()) {
3262 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3263 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3266 // Try an immediate VMVN.
3267 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3268 ((1LL << SplatBitSize) - 1));
3269 Val = isNEONModifiedImm(NegatedImm,
3270 SplatUndef.getZExtValue(), SplatBitSize,
3271 DAG, VmovVT, VT.is128BitVector(), false);
3272 if (Val.getNode()) {
3273 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3274 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3279 // Scan through the operands to see if only one value is used.
3280 unsigned NumElts = VT.getVectorNumElements();
3281 bool isOnlyLowElement = true;
3282 bool usesOnlyOneValue = true;
3283 bool isConstant = true;
3285 for (unsigned i = 0; i < NumElts; ++i) {
3286 SDValue V = Op.getOperand(i);
3287 if (V.getOpcode() == ISD::UNDEF)
3290 isOnlyLowElement = false;
3291 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3294 if (!Value.getNode())
3296 else if (V != Value)
3297 usesOnlyOneValue = false;
3300 if (!Value.getNode())
3301 return DAG.getUNDEF(VT);
3303 if (isOnlyLowElement)
3304 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3306 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3308 if (EnableARMVDUPsplat) {
3309 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3310 // i32 and try again.
3311 if (usesOnlyOneValue && EltSize <= 32) {
3313 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3314 if (VT.getVectorElementType().isFloatingPoint()) {
3315 SmallVector<SDValue, 8> Ops;
3316 for (unsigned i = 0; i < NumElts; ++i)
3317 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3319 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3321 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3322 LowerBUILD_VECTOR(Val, DAG, ST));
3324 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3326 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3330 // If all elements are constants and the case above didn't get hit, fall back
3331 // to the default expansion, which will generate a load from the constant
3336 if (!EnableARMVDUPsplat) {
3337 // Use VDUP for non-constant splats.
3338 if (usesOnlyOneValue && EltSize <= 32)
3339 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3342 // Vectors with 32- or 64-bit elements can be built by directly assigning
3343 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3344 // will be legalized.
3345 if (EltSize >= 32) {
3346 // Do the expansion with floating-point types, since that is what the VFP
3347 // registers are defined to use, and since i64 is not legal.
3348 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3349 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3350 SmallVector<SDValue, 8> Ops;
3351 for (unsigned i = 0; i < NumElts; ++i)
3352 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3353 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3354 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3360 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3361 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3362 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3363 /// are assumed to be legal.
3365 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3367 if (VT.getVectorNumElements() == 4 &&
3368 (VT.is128BitVector() || VT.is64BitVector())) {
3369 unsigned PFIndexes[4];
3370 for (unsigned i = 0; i != 4; ++i) {
3374 PFIndexes[i] = M[i];
3377 // Compute the index in the perfect shuffle table.
3378 unsigned PFTableIndex =
3379 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3380 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3381 unsigned Cost = (PFEntry >> 30);
3388 unsigned Imm, WhichResult;
3390 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3391 return (EltSize >= 32 ||
3392 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3393 isVREVMask(M, VT, 64) ||
3394 isVREVMask(M, VT, 32) ||
3395 isVREVMask(M, VT, 16) ||
3396 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3397 isVTRNMask(M, VT, WhichResult) ||
3398 isVUZPMask(M, VT, WhichResult) ||
3399 isVZIPMask(M, VT, WhichResult) ||
3400 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3401 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3402 isVZIP_v_undef_Mask(M, VT, WhichResult));
3405 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3406 /// the specified operations to build the shuffle.
3407 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3408 SDValue RHS, SelectionDAG &DAG,
3410 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3411 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3412 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3415 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3424 OP_VUZPL, // VUZP, left result
3425 OP_VUZPR, // VUZP, right result
3426 OP_VZIPL, // VZIP, left result
3427 OP_VZIPR, // VZIP, right result
3428 OP_VTRNL, // VTRN, left result
3429 OP_VTRNR // VTRN, right result
3432 if (OpNum == OP_COPY) {
3433 if (LHSID == (1*9+2)*9+3) return LHS;
3434 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3438 SDValue OpLHS, OpRHS;
3439 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3440 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3441 EVT VT = OpLHS.getValueType();
3444 default: llvm_unreachable("Unknown shuffle opcode!");
3446 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3451 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3452 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3456 return DAG.getNode(ARMISD::VEXT, dl, VT,
3458 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3461 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3462 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3465 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3466 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3469 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3470 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3474 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3475 SDValue V1 = Op.getOperand(0);
3476 SDValue V2 = Op.getOperand(1);
3477 DebugLoc dl = Op.getDebugLoc();
3478 EVT VT = Op.getValueType();
3479 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3480 SmallVector<int, 8> ShuffleMask;
3482 // Convert shuffles that are directly supported on NEON to target-specific
3483 // DAG nodes, instead of keeping them as shuffles and matching them again
3484 // during code selection. This is more efficient and avoids the possibility
3485 // of inconsistencies between legalization and selection.
3486 // FIXME: floating-point vectors should be canonicalized to integer vectors
3487 // of the same time so that they get CSEd properly.
3488 SVN->getMask(ShuffleMask);
3490 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3491 if (EltSize <= 32) {
3492 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3493 int Lane = SVN->getSplatIndex();
3494 // If this is undef splat, generate it via "just" vdup, if possible.
3495 if (Lane == -1) Lane = 0;
3497 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3498 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3500 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3501 DAG.getConstant(Lane, MVT::i32));
3506 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3509 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3510 DAG.getConstant(Imm, MVT::i32));
3513 if (isVREVMask(ShuffleMask, VT, 64))
3514 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3515 if (isVREVMask(ShuffleMask, VT, 32))
3516 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3517 if (isVREVMask(ShuffleMask, VT, 16))
3518 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3520 // Check for Neon shuffles that modify both input vectors in place.
3521 // If both results are used, i.e., if there are two shuffles with the same
3522 // source operands and with masks corresponding to both results of one of
3523 // these operations, DAG memoization will ensure that a single node is
3524 // used for both shuffles.
3525 unsigned WhichResult;
3526 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3527 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3528 V1, V2).getValue(WhichResult);
3529 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3530 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3531 V1, V2).getValue(WhichResult);
3532 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3533 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3534 V1, V2).getValue(WhichResult);
3536 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3537 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3538 V1, V1).getValue(WhichResult);
3539 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3540 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3541 V1, V1).getValue(WhichResult);
3542 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3543 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3544 V1, V1).getValue(WhichResult);
3547 // If the shuffle is not directly supported and it has 4 elements, use
3548 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3549 unsigned NumElts = VT.getVectorNumElements();
3551 unsigned PFIndexes[4];
3552 for (unsigned i = 0; i != 4; ++i) {
3553 if (ShuffleMask[i] < 0)
3556 PFIndexes[i] = ShuffleMask[i];
3559 // Compute the index in the perfect shuffle table.
3560 unsigned PFTableIndex =
3561 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3562 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3563 unsigned Cost = (PFEntry >> 30);
3566 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3569 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3570 if (EltSize >= 32) {
3571 // Do the expansion with floating-point types, since that is what the VFP
3572 // registers are defined to use, and since i64 is not legal.
3573 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3574 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3575 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3576 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3577 SmallVector<SDValue, 8> Ops;
3578 for (unsigned i = 0; i < NumElts; ++i) {
3579 if (ShuffleMask[i] < 0)
3580 Ops.push_back(DAG.getUNDEF(EltVT));
3582 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3583 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3584 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3587 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3588 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3594 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3595 EVT VT = Op.getValueType();
3596 DebugLoc dl = Op.getDebugLoc();
3597 SDValue Vec = Op.getOperand(0);
3598 SDValue Lane = Op.getOperand(1);
3599 assert(VT == MVT::i32 &&
3600 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3601 "unexpected type for custom-lowering vector extract");
3602 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3605 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3606 // The only time a CONCAT_VECTORS operation can have legal types is when
3607 // two 64-bit vectors are concatenated to a 128-bit vector.
3608 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3609 "unexpected CONCAT_VECTORS");
3610 DebugLoc dl = Op.getDebugLoc();
3611 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3612 SDValue Op0 = Op.getOperand(0);
3613 SDValue Op1 = Op.getOperand(1);
3614 if (Op0.getOpcode() != ISD::UNDEF)
3615 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3616 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3617 DAG.getIntPtrConstant(0));
3618 if (Op1.getOpcode() != ISD::UNDEF)
3619 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3620 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3621 DAG.getIntPtrConstant(1));
3622 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3625 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3626 /// an extending load, return the unextended value.
3627 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3628 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3629 return N->getOperand(0);
3630 LoadSDNode *LD = cast<LoadSDNode>(N);
3631 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3632 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3633 LD->isNonTemporal(), LD->getAlignment());
3636 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3637 // Multiplications are only custom-lowered for 128-bit vectors so that
3638 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3639 EVT VT = Op.getValueType();
3640 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3641 SDNode *N0 = Op.getOperand(0).getNode();
3642 SDNode *N1 = Op.getOperand(1).getNode();
3643 unsigned NewOpc = 0;
3644 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3645 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3646 NewOpc = ARMISD::VMULLs;
3647 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3648 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3649 NewOpc = ARMISD::VMULLu;
3650 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3651 // Fall through to expand this. It is not legal.
3654 // Other vector multiplications are legal.
3658 // Legalize to a VMULL instruction.
3659 DebugLoc DL = Op.getDebugLoc();
3660 SDValue Op0 = SkipExtension(N0, DAG);
3661 SDValue Op1 = SkipExtension(N1, DAG);
3663 assert(Op0.getValueType().is64BitVector() &&
3664 Op1.getValueType().is64BitVector() &&
3665 "unexpected types for extended operands to VMULL");
3666 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3669 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3670 switch (Op.getOpcode()) {
3671 default: llvm_unreachable("Don't know how to custom lower this!");
3672 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3673 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3674 case ISD::GlobalAddress:
3675 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3676 LowerGlobalAddressELF(Op, DAG);
3677 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3678 case ISD::SELECT: return LowerSELECT(Op, DAG);
3679 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3680 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3681 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3682 case ISD::VASTART: return LowerVASTART(Op, DAG);
3683 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3684 case ISD::SINT_TO_FP:
3685 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3686 case ISD::FP_TO_SINT:
3687 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3688 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3689 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3690 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3691 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3692 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3693 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3694 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3696 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3699 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3700 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3701 case ISD::SRL_PARTS:
3702 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3703 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3704 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3705 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3706 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3707 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3708 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3709 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3710 case ISD::MUL: return LowerMUL(Op, DAG);
3715 /// ReplaceNodeResults - Replace the results of node with an illegal result
3716 /// type with new values built out of custom code.
3717 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3718 SmallVectorImpl<SDValue>&Results,
3719 SelectionDAG &DAG) const {
3721 switch (N->getOpcode()) {
3723 llvm_unreachable("Don't know how to custom expand this!");
3725 case ISD::BIT_CONVERT:
3726 Res = ExpandBIT_CONVERT(N, DAG);
3730 Res = LowerShift(N, DAG, Subtarget);
3734 Results.push_back(Res);
3737 //===----------------------------------------------------------------------===//
3738 // ARM Scheduler Hooks
3739 //===----------------------------------------------------------------------===//
3742 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3743 MachineBasicBlock *BB,
3744 unsigned Size) const {
3745 unsigned dest = MI->getOperand(0).getReg();
3746 unsigned ptr = MI->getOperand(1).getReg();
3747 unsigned oldval = MI->getOperand(2).getReg();
3748 unsigned newval = MI->getOperand(3).getReg();
3749 unsigned scratch = BB->getParent()->getRegInfo()
3750 .createVirtualRegister(ARM::GPRRegisterClass);
3751 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3752 DebugLoc dl = MI->getDebugLoc();
3753 bool isThumb2 = Subtarget->isThumb2();
3755 unsigned ldrOpc, strOpc;
3757 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3759 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3760 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3763 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3764 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3767 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3768 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3772 MachineFunction *MF = BB->getParent();
3773 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3774 MachineFunction::iterator It = BB;
3775 ++It; // insert the new blocks after the current block
3777 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3778 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3779 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3780 MF->insert(It, loop1MBB);
3781 MF->insert(It, loop2MBB);
3782 MF->insert(It, exitMBB);
3784 // Transfer the remainder of BB and its successor edges to exitMBB.
3785 exitMBB->splice(exitMBB->begin(), BB,
3786 llvm::next(MachineBasicBlock::iterator(MI)),
3788 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3792 // fallthrough --> loop1MBB
3793 BB->addSuccessor(loop1MBB);
3796 // ldrex dest, [ptr]
3800 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3801 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3802 .addReg(dest).addReg(oldval));
3803 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3804 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3805 BB->addSuccessor(loop2MBB);
3806 BB->addSuccessor(exitMBB);
3809 // strex scratch, newval, [ptr]
3813 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3815 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3816 .addReg(scratch).addImm(0));
3817 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3818 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3819 BB->addSuccessor(loop1MBB);
3820 BB->addSuccessor(exitMBB);
3826 MI->eraseFromParent(); // The instruction is gone now.
3832 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3833 unsigned Size, unsigned BinOpcode) const {
3834 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3835 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3837 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3838 MachineFunction *MF = BB->getParent();
3839 MachineFunction::iterator It = BB;
3842 unsigned dest = MI->getOperand(0).getReg();
3843 unsigned ptr = MI->getOperand(1).getReg();
3844 unsigned incr = MI->getOperand(2).getReg();
3845 DebugLoc dl = MI->getDebugLoc();
3847 bool isThumb2 = Subtarget->isThumb2();
3848 unsigned ldrOpc, strOpc;
3850 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3852 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3853 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3856 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3857 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3860 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3861 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3865 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3866 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3867 MF->insert(It, loopMBB);
3868 MF->insert(It, exitMBB);
3870 // Transfer the remainder of BB and its successor edges to exitMBB.
3871 exitMBB->splice(exitMBB->begin(), BB,
3872 llvm::next(MachineBasicBlock::iterator(MI)),
3874 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3876 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3877 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3878 unsigned scratch2 = (!BinOpcode) ? incr :
3879 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3883 // fallthrough --> loopMBB
3884 BB->addSuccessor(loopMBB);
3888 // <binop> scratch2, dest, incr
3889 // strex scratch, scratch2, ptr
3892 // fallthrough --> exitMBB
3894 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3896 // operand order needs to go the other way for NAND
3897 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3898 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3899 addReg(incr).addReg(dest)).addReg(0);
3901 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3902 addReg(dest).addReg(incr)).addReg(0);
3905 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3907 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3908 .addReg(scratch).addImm(0));
3909 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3910 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3912 BB->addSuccessor(loopMBB);
3913 BB->addSuccessor(exitMBB);
3919 MI->eraseFromParent(); // The instruction is gone now.
3925 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3926 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3927 E = MBB->succ_end(); I != E; ++I)
3930 llvm_unreachable("Expecting a BB with two successors!");
3934 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3935 MachineBasicBlock *BB) const {
3936 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3937 DebugLoc dl = MI->getDebugLoc();
3938 bool isThumb2 = Subtarget->isThumb2();
3939 switch (MI->getOpcode()) {
3942 llvm_unreachable("Unexpected instr type to insert");
3944 case ARM::ATOMIC_LOAD_ADD_I8:
3945 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3946 case ARM::ATOMIC_LOAD_ADD_I16:
3947 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3948 case ARM::ATOMIC_LOAD_ADD_I32:
3949 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3951 case ARM::ATOMIC_LOAD_AND_I8:
3952 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3953 case ARM::ATOMIC_LOAD_AND_I16:
3954 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3955 case ARM::ATOMIC_LOAD_AND_I32:
3956 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3958 case ARM::ATOMIC_LOAD_OR_I8:
3959 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3960 case ARM::ATOMIC_LOAD_OR_I16:
3961 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3962 case ARM::ATOMIC_LOAD_OR_I32:
3963 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3965 case ARM::ATOMIC_LOAD_XOR_I8:
3966 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3967 case ARM::ATOMIC_LOAD_XOR_I16:
3968 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3969 case ARM::ATOMIC_LOAD_XOR_I32:
3970 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3972 case ARM::ATOMIC_LOAD_NAND_I8:
3973 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3974 case ARM::ATOMIC_LOAD_NAND_I16:
3975 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3976 case ARM::ATOMIC_LOAD_NAND_I32:
3977 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3979 case ARM::ATOMIC_LOAD_SUB_I8:
3980 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3981 case ARM::ATOMIC_LOAD_SUB_I16:
3982 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3983 case ARM::ATOMIC_LOAD_SUB_I32:
3984 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3986 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3987 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3988 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3990 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3991 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3992 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3994 case ARM::tMOVCCr_pseudo: {
3995 // To "insert" a SELECT_CC instruction, we actually have to insert the
3996 // diamond control-flow pattern. The incoming instruction knows the
3997 // destination vreg to set, the condition code register to branch on, the
3998 // true/false values to select between, and a branch opcode to use.
3999 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4000 MachineFunction::iterator It = BB;
4006 // cmpTY ccX, r1, r2
4008 // fallthrough --> copy0MBB
4009 MachineBasicBlock *thisMBB = BB;
4010 MachineFunction *F = BB->getParent();
4011 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4012 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4013 F->insert(It, copy0MBB);
4014 F->insert(It, sinkMBB);
4016 // Transfer the remainder of BB and its successor edges to sinkMBB.
4017 sinkMBB->splice(sinkMBB->begin(), BB,
4018 llvm::next(MachineBasicBlock::iterator(MI)),
4020 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4022 BB->addSuccessor(copy0MBB);
4023 BB->addSuccessor(sinkMBB);
4025 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4026 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4029 // %FalseValue = ...
4030 // # fallthrough to sinkMBB
4033 // Update machine-CFG edges
4034 BB->addSuccessor(sinkMBB);
4037 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4040 BuildMI(*BB, BB->begin(), dl,
4041 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4042 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4043 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4045 MI->eraseFromParent(); // The pseudo instruction is gone now.
4050 case ARM::BCCZi64: {
4051 // Compare both parts that make up the double comparison separately for
4053 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4055 unsigned LHS1 = MI->getOperand(1).getReg();
4056 unsigned LHS2 = MI->getOperand(2).getReg();
4058 AddDefaultPred(BuildMI(BB, dl,
4059 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4060 .addReg(LHS1).addImm(0));
4061 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4062 .addReg(LHS2).addImm(0)
4063 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4065 unsigned RHS1 = MI->getOperand(3).getReg();
4066 unsigned RHS2 = MI->getOperand(4).getReg();
4067 AddDefaultPred(BuildMI(BB, dl,
4068 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4069 .addReg(LHS1).addReg(RHS1));
4070 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4071 .addReg(LHS2).addReg(RHS2)
4072 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4075 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4076 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4077 if (MI->getOperand(0).getImm() == ARMCC::NE)
4078 std::swap(destMBB, exitMBB);
4080 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4081 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4082 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4085 MI->eraseFromParent(); // The pseudo instruction is gone now.
4091 //===----------------------------------------------------------------------===//
4092 // ARM Optimization Hooks
4093 //===----------------------------------------------------------------------===//
4096 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4097 TargetLowering::DAGCombinerInfo &DCI) {
4098 SelectionDAG &DAG = DCI.DAG;
4099 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4100 EVT VT = N->getValueType(0);
4101 unsigned Opc = N->getOpcode();
4102 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4103 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4104 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4105 ISD::CondCode CC = ISD::SETCC_INVALID;
4108 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4110 SDValue CCOp = Slct.getOperand(0);
4111 if (CCOp.getOpcode() == ISD::SETCC)
4112 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4115 bool DoXform = false;
4117 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4120 if (LHS.getOpcode() == ISD::Constant &&
4121 cast<ConstantSDNode>(LHS)->isNullValue()) {
4123 } else if (CC != ISD::SETCC_INVALID &&
4124 RHS.getOpcode() == ISD::Constant &&
4125 cast<ConstantSDNode>(RHS)->isNullValue()) {
4126 std::swap(LHS, RHS);
4127 SDValue Op0 = Slct.getOperand(0);
4128 EVT OpVT = isSlctCC ? Op0.getValueType() :
4129 Op0.getOperand(0).getValueType();
4130 bool isInt = OpVT.isInteger();
4131 CC = ISD::getSetCCInverse(CC, isInt);
4133 if (!TLI.isCondCodeLegal(CC, OpVT))
4134 return SDValue(); // Inverse operator isn't legal.
4141 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4143 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4144 Slct.getOperand(0), Slct.getOperand(1), CC);
4145 SDValue CCOp = Slct.getOperand(0);
4147 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4148 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4149 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4150 CCOp, OtherOp, Result);
4155 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4156 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4157 /// called with the default operands, and if that fails, with commuted
4159 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4160 TargetLowering::DAGCombinerInfo &DCI) {
4161 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4162 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4163 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4164 if (Result.getNode()) return Result;
4169 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4171 static SDValue PerformADDCombine(SDNode *N,
4172 TargetLowering::DAGCombinerInfo &DCI) {
4173 SDValue N0 = N->getOperand(0);
4174 SDValue N1 = N->getOperand(1);
4176 // First try with the default operand order.
4177 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4178 if (Result.getNode())
4181 // If that didn't work, try again with the operands commuted.
4182 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4185 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4187 static SDValue PerformSUBCombine(SDNode *N,
4188 TargetLowering::DAGCombinerInfo &DCI) {
4189 SDValue N0 = N->getOperand(0);
4190 SDValue N1 = N->getOperand(1);
4192 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4193 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4194 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4195 if (Result.getNode()) return Result;
4201 static SDValue PerformMULCombine(SDNode *N,
4202 TargetLowering::DAGCombinerInfo &DCI,
4203 const ARMSubtarget *Subtarget) {
4204 SelectionDAG &DAG = DCI.DAG;
4206 if (Subtarget->isThumb1Only())
4209 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4212 EVT VT = N->getValueType(0);
4216 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4220 uint64_t MulAmt = C->getZExtValue();
4221 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4222 ShiftAmt = ShiftAmt & (32 - 1);
4223 SDValue V = N->getOperand(0);
4224 DebugLoc DL = N->getDebugLoc();
4227 MulAmt >>= ShiftAmt;
4228 if (isPowerOf2_32(MulAmt - 1)) {
4229 // (mul x, 2^N + 1) => (add (shl x, N), x)
4230 Res = DAG.getNode(ISD::ADD, DL, VT,
4231 V, DAG.getNode(ISD::SHL, DL, VT,
4232 V, DAG.getConstant(Log2_32(MulAmt-1),
4234 } else if (isPowerOf2_32(MulAmt + 1)) {
4235 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4236 Res = DAG.getNode(ISD::SUB, DL, VT,
4237 DAG.getNode(ISD::SHL, DL, VT,
4238 V, DAG.getConstant(Log2_32(MulAmt+1),
4245 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4246 DAG.getConstant(ShiftAmt, MVT::i32));
4248 // Do not add new nodes to DAG combiner worklist.
4249 DCI.CombineTo(N, Res, false);
4253 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4254 static SDValue PerformORCombine(SDNode *N,
4255 TargetLowering::DAGCombinerInfo &DCI,
4256 const ARMSubtarget *Subtarget) {
4257 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4260 // BFI is only available on V6T2+
4261 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4264 SelectionDAG &DAG = DCI.DAG;
4265 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4266 DebugLoc DL = N->getDebugLoc();
4267 // 1) or (and A, mask), val => ARMbfi A, val, mask
4268 // iff (val & mask) == val
4270 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4271 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4272 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4273 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4274 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4275 // (i.e., copy a bitfield value into another bitfield of the same width)
4276 if (N0.getOpcode() != ISD::AND)
4279 EVT VT = N->getValueType(0);
4284 // The value and the mask need to be constants so we can verify this is
4285 // actually a bitfield set. If the mask is 0xffff, we can do better
4286 // via a movt instruction, so don't use BFI in that case.
4287 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4290 unsigned Mask = C->getZExtValue();
4294 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4295 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4296 unsigned Val = C->getZExtValue();
4297 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4299 Val >>= CountTrailingZeros_32(~Mask);
4301 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4302 DAG.getConstant(Val, MVT::i32),
4303 DAG.getConstant(Mask, MVT::i32));
4305 // Do not add new nodes to DAG combiner worklist.
4306 DCI.CombineTo(N, Res, false);
4307 } else if (N1.getOpcode() == ISD::AND) {
4308 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4309 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4312 unsigned Mask2 = C->getZExtValue();
4314 if (ARM::isBitFieldInvertedMask(Mask) &&
4315 ARM::isBitFieldInvertedMask(~Mask2) &&
4316 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4317 // The pack halfword instruction works better for masks that fit it,
4318 // so use that when it's available.
4319 if (Subtarget->hasT2ExtractPack() &&
4320 (Mask == 0xffff || Mask == 0xffff0000))
4323 unsigned lsb = CountTrailingZeros_32(Mask2);
4324 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4325 DAG.getConstant(lsb, MVT::i32));
4326 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4327 DAG.getConstant(Mask, MVT::i32));
4328 // Do not add new nodes to DAG combiner worklist.
4329 DCI.CombineTo(N, Res, false);
4330 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4331 ARM::isBitFieldInvertedMask(Mask2) &&
4332 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4333 // The pack halfword instruction works better for masks that fit it,
4334 // so use that when it's available.
4335 if (Subtarget->hasT2ExtractPack() &&
4336 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4339 unsigned lsb = CountTrailingZeros_32(Mask);
4340 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4341 DAG.getConstant(lsb, MVT::i32));
4342 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4343 DAG.getConstant(Mask2, MVT::i32));
4344 // Do not add new nodes to DAG combiner worklist.
4345 DCI.CombineTo(N, Res, false);
4352 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4353 /// ARMISD::VMOVRRD.
4354 static SDValue PerformVMOVRRDCombine(SDNode *N,
4355 TargetLowering::DAGCombinerInfo &DCI) {
4356 // vmovrrd(vmovdrr x, y) -> x,y
4357 SDValue InDouble = N->getOperand(0);
4358 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4359 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4363 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
4364 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
4365 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
4366 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
4367 SDValue Op0 = N->getOperand(0);
4368 SDValue Op1 = N->getOperand(1);
4369 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4370 Op0 = Op0.getOperand(0);
4371 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4372 Op1 = Op1.getOperand(0);
4373 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4374 Op0.getNode() == Op1.getNode() &&
4375 Op0.getResNo() == 0 && Op1.getResNo() == 1)
4376 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4377 N->getValueType(0), Op0.getOperand(0));
4381 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4382 /// ISD::BUILD_VECTOR.
4383 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4384 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4385 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4386 // into a pair of GPRs, which is fine when the value is used as a scalar,
4387 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4388 if (N->getNumOperands() == 2)
4389 return PerformVMOVDRRCombine(N, DAG);
4394 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4395 /// ARMISD::VDUPLANE.
4396 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
4397 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4399 SDValue Op = N->getOperand(0);
4400 EVT VT = N->getValueType(0);
4402 // Ignore bit_converts.
4403 while (Op.getOpcode() == ISD::BIT_CONVERT)
4404 Op = Op.getOperand(0);
4405 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4408 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4409 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4410 // The canonical VMOV for a zero vector uses a 32-bit element size.
4411 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4413 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4415 if (EltSize > VT.getVectorElementType().getSizeInBits())
4418 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4421 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4422 /// operand of a vector shift operation, where all the elements of the
4423 /// build_vector must have the same constant integer value.
4424 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4425 // Ignore bit_converts.
4426 while (Op.getOpcode() == ISD::BIT_CONVERT)
4427 Op = Op.getOperand(0);
4428 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4429 APInt SplatBits, SplatUndef;
4430 unsigned SplatBitSize;
4432 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4433 HasAnyUndefs, ElementBits) ||
4434 SplatBitSize > ElementBits)
4436 Cnt = SplatBits.getSExtValue();
4440 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4441 /// operand of a vector shift left operation. That value must be in the range:
4442 /// 0 <= Value < ElementBits for a left shift; or
4443 /// 0 <= Value <= ElementBits for a long left shift.
4444 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4445 assert(VT.isVector() && "vector shift count is not a vector type");
4446 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4447 if (! getVShiftImm(Op, ElementBits, Cnt))
4449 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4452 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4453 /// operand of a vector shift right operation. For a shift opcode, the value
4454 /// is positive, but for an intrinsic the value count must be negative. The
4455 /// absolute value must be in the range:
4456 /// 1 <= |Value| <= ElementBits for a right shift; or
4457 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4458 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4460 assert(VT.isVector() && "vector shift count is not a vector type");
4461 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4462 if (! getVShiftImm(Op, ElementBits, Cnt))
4466 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4469 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4470 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4471 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4474 // Don't do anything for most intrinsics.
4477 // Vector shifts: check for immediate versions and lower them.
4478 // Note: This is done during DAG combining instead of DAG legalizing because
4479 // the build_vectors for 64-bit vector element shift counts are generally
4480 // not legal, and it is hard to see their values after they get legalized to
4481 // loads from a constant pool.
4482 case Intrinsic::arm_neon_vshifts:
4483 case Intrinsic::arm_neon_vshiftu:
4484 case Intrinsic::arm_neon_vshiftls:
4485 case Intrinsic::arm_neon_vshiftlu:
4486 case Intrinsic::arm_neon_vshiftn:
4487 case Intrinsic::arm_neon_vrshifts:
4488 case Intrinsic::arm_neon_vrshiftu:
4489 case Intrinsic::arm_neon_vrshiftn:
4490 case Intrinsic::arm_neon_vqshifts:
4491 case Intrinsic::arm_neon_vqshiftu:
4492 case Intrinsic::arm_neon_vqshiftsu:
4493 case Intrinsic::arm_neon_vqshiftns:
4494 case Intrinsic::arm_neon_vqshiftnu:
4495 case Intrinsic::arm_neon_vqshiftnsu:
4496 case Intrinsic::arm_neon_vqrshiftns:
4497 case Intrinsic::arm_neon_vqrshiftnu:
4498 case Intrinsic::arm_neon_vqrshiftnsu: {
4499 EVT VT = N->getOperand(1).getValueType();
4501 unsigned VShiftOpc = 0;
4504 case Intrinsic::arm_neon_vshifts:
4505 case Intrinsic::arm_neon_vshiftu:
4506 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4507 VShiftOpc = ARMISD::VSHL;
4510 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4511 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4512 ARMISD::VSHRs : ARMISD::VSHRu);
4517 case Intrinsic::arm_neon_vshiftls:
4518 case Intrinsic::arm_neon_vshiftlu:
4519 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4521 llvm_unreachable("invalid shift count for vshll intrinsic");
4523 case Intrinsic::arm_neon_vrshifts:
4524 case Intrinsic::arm_neon_vrshiftu:
4525 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4529 case Intrinsic::arm_neon_vqshifts:
4530 case Intrinsic::arm_neon_vqshiftu:
4531 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4535 case Intrinsic::arm_neon_vqshiftsu:
4536 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4538 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4540 case Intrinsic::arm_neon_vshiftn:
4541 case Intrinsic::arm_neon_vrshiftn:
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 // Narrowing shifts require an immediate right shift.
4549 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4551 llvm_unreachable("invalid shift count for narrowing vector shift "
4555 llvm_unreachable("unhandled vector shift");
4559 case Intrinsic::arm_neon_vshifts:
4560 case Intrinsic::arm_neon_vshiftu:
4561 // Opcode already set above.
4563 case Intrinsic::arm_neon_vshiftls:
4564 case Intrinsic::arm_neon_vshiftlu:
4565 if (Cnt == VT.getVectorElementType().getSizeInBits())
4566 VShiftOpc = ARMISD::VSHLLi;
4568 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4569 ARMISD::VSHLLs : ARMISD::VSHLLu);
4571 case Intrinsic::arm_neon_vshiftn:
4572 VShiftOpc = ARMISD::VSHRN; break;
4573 case Intrinsic::arm_neon_vrshifts:
4574 VShiftOpc = ARMISD::VRSHRs; break;
4575 case Intrinsic::arm_neon_vrshiftu:
4576 VShiftOpc = ARMISD::VRSHRu; break;
4577 case Intrinsic::arm_neon_vrshiftn:
4578 VShiftOpc = ARMISD::VRSHRN; break;
4579 case Intrinsic::arm_neon_vqshifts:
4580 VShiftOpc = ARMISD::VQSHLs; break;
4581 case Intrinsic::arm_neon_vqshiftu:
4582 VShiftOpc = ARMISD::VQSHLu; break;
4583 case Intrinsic::arm_neon_vqshiftsu:
4584 VShiftOpc = ARMISD::VQSHLsu; break;
4585 case Intrinsic::arm_neon_vqshiftns:
4586 VShiftOpc = ARMISD::VQSHRNs; break;
4587 case Intrinsic::arm_neon_vqshiftnu:
4588 VShiftOpc = ARMISD::VQSHRNu; break;
4589 case Intrinsic::arm_neon_vqshiftnsu:
4590 VShiftOpc = ARMISD::VQSHRNsu; break;
4591 case Intrinsic::arm_neon_vqrshiftns:
4592 VShiftOpc = ARMISD::VQRSHRNs; break;
4593 case Intrinsic::arm_neon_vqrshiftnu:
4594 VShiftOpc = ARMISD::VQRSHRNu; break;
4595 case Intrinsic::arm_neon_vqrshiftnsu:
4596 VShiftOpc = ARMISD::VQRSHRNsu; break;
4599 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4600 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4603 case Intrinsic::arm_neon_vshiftins: {
4604 EVT VT = N->getOperand(1).getValueType();
4606 unsigned VShiftOpc = 0;
4608 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4609 VShiftOpc = ARMISD::VSLI;
4610 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4611 VShiftOpc = ARMISD::VSRI;
4613 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4616 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4617 N->getOperand(1), N->getOperand(2),
4618 DAG.getConstant(Cnt, MVT::i32));
4621 case Intrinsic::arm_neon_vqrshifts:
4622 case Intrinsic::arm_neon_vqrshiftu:
4623 // No immediate versions of these to check for.
4630 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4631 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4632 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4633 /// vector element shift counts are generally not legal, and it is hard to see
4634 /// their values after they get legalized to loads from a constant pool.
4635 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4636 const ARMSubtarget *ST) {
4637 EVT VT = N->getValueType(0);
4639 // Nothing to be done for scalar shifts.
4640 if (! VT.isVector())
4643 assert(ST->hasNEON() && "unexpected vector shift");
4646 switch (N->getOpcode()) {
4647 default: llvm_unreachable("unexpected shift opcode");
4650 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4651 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4652 DAG.getConstant(Cnt, MVT::i32));
4657 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4658 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4659 ARMISD::VSHRs : ARMISD::VSHRu);
4660 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4661 DAG.getConstant(Cnt, MVT::i32));
4667 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4668 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4669 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4670 const ARMSubtarget *ST) {
4671 SDValue N0 = N->getOperand(0);
4673 // Check for sign- and zero-extensions of vector extract operations of 8-
4674 // and 16-bit vector elements. NEON supports these directly. They are
4675 // handled during DAG combining because type legalization will promote them
4676 // to 32-bit types and it is messy to recognize the operations after that.
4677 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4678 SDValue Vec = N0.getOperand(0);
4679 SDValue Lane = N0.getOperand(1);
4680 EVT VT = N->getValueType(0);
4681 EVT EltVT = N0.getValueType();
4682 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4684 if (VT == MVT::i32 &&
4685 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4686 TLI.isTypeLegal(Vec.getValueType())) {
4689 switch (N->getOpcode()) {
4690 default: llvm_unreachable("unexpected opcode");
4691 case ISD::SIGN_EXTEND:
4692 Opc = ARMISD::VGETLANEs;
4694 case ISD::ZERO_EXTEND:
4695 case ISD::ANY_EXTEND:
4696 Opc = ARMISD::VGETLANEu;
4699 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4706 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4707 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4708 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4709 const ARMSubtarget *ST) {
4710 // If the target supports NEON, try to use vmax/vmin instructions for f32
4711 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4712 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4713 // a NaN; only do the transformation when it matches that behavior.
4715 // For now only do this when using NEON for FP operations; if using VFP, it
4716 // is not obvious that the benefit outweighs the cost of switching to the
4718 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4719 N->getValueType(0) != MVT::f32)
4722 SDValue CondLHS = N->getOperand(0);
4723 SDValue CondRHS = N->getOperand(1);
4724 SDValue LHS = N->getOperand(2);
4725 SDValue RHS = N->getOperand(3);
4726 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4728 unsigned Opcode = 0;
4730 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4731 IsReversed = false; // x CC y ? x : y
4732 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4733 IsReversed = true ; // x CC y ? y : x
4747 // If LHS is NaN, an ordered comparison will be false and the result will
4748 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4749 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4750 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4751 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4753 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4754 // will return -0, so vmin can only be used for unsafe math or if one of
4755 // the operands is known to be nonzero.
4756 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4758 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4760 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4769 // If LHS is NaN, an ordered comparison will be false and the result will
4770 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4771 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4772 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4773 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4775 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4776 // will return +0, so vmax can only be used for unsafe math or if one of
4777 // the operands is known to be nonzero.
4778 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4780 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4782 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4788 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4791 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4792 DAGCombinerInfo &DCI) const {
4793 switch (N->getOpcode()) {
4795 case ISD::ADD: return PerformADDCombine(N, DCI);
4796 case ISD::SUB: return PerformSUBCombine(N, DCI);
4797 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4798 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4799 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4800 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
4801 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
4802 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
4803 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4806 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4807 case ISD::SIGN_EXTEND:
4808 case ISD::ZERO_EXTEND:
4809 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4810 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4815 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4816 if (!Subtarget->hasV6Ops())
4817 // Pre-v6 does not support unaligned mem access.
4820 // v6+ may or may not support unaligned mem access depending on the system
4822 // FIXME: This is pretty conservative. Should we provide cmdline option to
4823 // control the behaviour?
4824 if (!Subtarget->isTargetDarwin())
4827 switch (VT.getSimpleVT().SimpleTy) {
4834 // FIXME: VLD1 etc with standard alignment is legal.
4838 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4843 switch (VT.getSimpleVT().SimpleTy) {
4844 default: return false;
4859 if ((V & (Scale - 1)) != 0)
4862 return V == (V & ((1LL << 5) - 1));
4865 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4866 const ARMSubtarget *Subtarget) {
4873 switch (VT.getSimpleVT().SimpleTy) {
4874 default: return false;
4879 // + imm12 or - imm8
4881 return V == (V & ((1LL << 8) - 1));
4882 return V == (V & ((1LL << 12) - 1));
4885 // Same as ARM mode. FIXME: NEON?
4886 if (!Subtarget->hasVFP2())
4891 return V == (V & ((1LL << 8) - 1));
4895 /// isLegalAddressImmediate - Return true if the integer value can be used
4896 /// as the offset of the target addressing mode for load / store of the
4898 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4899 const ARMSubtarget *Subtarget) {
4906 if (Subtarget->isThumb1Only())
4907 return isLegalT1AddressImmediate(V, VT);
4908 else if (Subtarget->isThumb2())
4909 return isLegalT2AddressImmediate(V, VT, Subtarget);
4914 switch (VT.getSimpleVT().SimpleTy) {
4915 default: return false;
4920 return V == (V & ((1LL << 12) - 1));
4923 return V == (V & ((1LL << 8) - 1));
4926 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4931 return V == (V & ((1LL << 8) - 1));
4935 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4937 int Scale = AM.Scale;
4941 switch (VT.getSimpleVT().SimpleTy) {
4942 default: return false;
4951 return Scale == 2 || Scale == 4 || Scale == 8;
4954 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4958 // Note, we allow "void" uses (basically, uses that aren't loads or
4959 // stores), because arm allows folding a scale into many arithmetic
4960 // operations. This should be made more precise and revisited later.
4962 // Allow r << imm, but the imm has to be a multiple of two.
4963 if (Scale & 1) return false;
4964 return isPowerOf2_32(Scale);
4968 /// isLegalAddressingMode - Return true if the addressing mode represented
4969 /// by AM is legal for this target, for a load/store of the specified type.
4970 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4971 const Type *Ty) const {
4972 EVT VT = getValueType(Ty, true);
4973 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4976 // Can never fold addr of global into load/store.
4981 case 0: // no scale reg, must be "r+i" or "r", or "i".
4984 if (Subtarget->isThumb1Only())
4988 // ARM doesn't support any R+R*scale+imm addr modes.
4995 if (Subtarget->isThumb2())
4996 return isLegalT2ScaledAddressingMode(AM, VT);
4998 int Scale = AM.Scale;
4999 switch (VT.getSimpleVT().SimpleTy) {
5000 default: return false;
5004 if (Scale < 0) Scale = -Scale;
5008 return isPowerOf2_32(Scale & ~1);
5012 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5017 // Note, we allow "void" uses (basically, uses that aren't loads or
5018 // stores), because arm allows folding a scale into many arithmetic
5019 // operations. This should be made more precise and revisited later.
5021 // Allow r << imm, but the imm has to be a multiple of two.
5022 if (Scale & 1) return false;
5023 return isPowerOf2_32(Scale);
5030 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5031 /// icmp immediate, that is the target has icmp instructions which can compare
5032 /// a register against the immediate without having to materialize the
5033 /// immediate into a register.
5034 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5035 if (!Subtarget->isThumb())
5036 return ARM_AM::getSOImmVal(Imm) != -1;
5037 if (Subtarget->isThumb2())
5038 return ARM_AM::getT2SOImmVal(Imm) != -1;
5039 return Imm >= 0 && Imm <= 255;
5042 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5043 bool isSEXTLoad, SDValue &Base,
5044 SDValue &Offset, bool &isInc,
5045 SelectionDAG &DAG) {
5046 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5049 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5051 Base = Ptr->getOperand(0);
5052 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5053 int RHSC = (int)RHS->getZExtValue();
5054 if (RHSC < 0 && RHSC > -256) {
5055 assert(Ptr->getOpcode() == ISD::ADD);
5057 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5061 isInc = (Ptr->getOpcode() == ISD::ADD);
5062 Offset = Ptr->getOperand(1);
5064 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5066 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5067 int RHSC = (int)RHS->getZExtValue();
5068 if (RHSC < 0 && RHSC > -0x1000) {
5069 assert(Ptr->getOpcode() == ISD::ADD);
5071 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5072 Base = Ptr->getOperand(0);
5077 if (Ptr->getOpcode() == ISD::ADD) {
5079 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5080 if (ShOpcVal != ARM_AM::no_shift) {
5081 Base = Ptr->getOperand(1);
5082 Offset = Ptr->getOperand(0);
5084 Base = Ptr->getOperand(0);
5085 Offset = Ptr->getOperand(1);
5090 isInc = (Ptr->getOpcode() == ISD::ADD);
5091 Base = Ptr->getOperand(0);
5092 Offset = Ptr->getOperand(1);
5096 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5100 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5101 bool isSEXTLoad, SDValue &Base,
5102 SDValue &Offset, bool &isInc,
5103 SelectionDAG &DAG) {
5104 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5107 Base = Ptr->getOperand(0);
5108 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5109 int RHSC = (int)RHS->getZExtValue();
5110 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5111 assert(Ptr->getOpcode() == ISD::ADD);
5113 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5115 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5116 isInc = Ptr->getOpcode() == ISD::ADD;
5117 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5125 /// getPreIndexedAddressParts - returns true by value, base pointer and
5126 /// offset pointer and addressing mode by reference if the node's address
5127 /// can be legally represented as pre-indexed load / store address.
5129 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5131 ISD::MemIndexedMode &AM,
5132 SelectionDAG &DAG) const {
5133 if (Subtarget->isThumb1Only())
5138 bool isSEXTLoad = false;
5139 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5140 Ptr = LD->getBasePtr();
5141 VT = LD->getMemoryVT();
5142 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5143 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5144 Ptr = ST->getBasePtr();
5145 VT = ST->getMemoryVT();
5150 bool isLegal = false;
5151 if (Subtarget->isThumb2())
5152 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5153 Offset, isInc, DAG);
5155 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5156 Offset, isInc, DAG);
5160 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5164 /// getPostIndexedAddressParts - returns true by value, base pointer and
5165 /// offset pointer and addressing mode by reference if this node can be
5166 /// combined with a load / store to form a post-indexed load / store.
5167 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5170 ISD::MemIndexedMode &AM,
5171 SelectionDAG &DAG) const {
5172 if (Subtarget->isThumb1Only())
5177 bool isSEXTLoad = false;
5178 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5179 VT = LD->getMemoryVT();
5180 Ptr = LD->getBasePtr();
5181 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5182 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5183 VT = ST->getMemoryVT();
5184 Ptr = ST->getBasePtr();
5189 bool isLegal = false;
5190 if (Subtarget->isThumb2())
5191 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5194 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5200 // Swap base ptr and offset to catch more post-index load / store when
5201 // it's legal. In Thumb2 mode, offset must be an immediate.
5202 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5203 !Subtarget->isThumb2())
5204 std::swap(Base, Offset);
5206 // Post-indexed load / store update the base pointer.
5211 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5215 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5219 const SelectionDAG &DAG,
5220 unsigned Depth) const {
5221 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5222 switch (Op.getOpcode()) {
5224 case ARMISD::CMOV: {
5225 // Bits are known zero/one if known on the LHS and RHS.
5226 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5227 if (KnownZero == 0 && KnownOne == 0) return;
5229 APInt KnownZeroRHS, KnownOneRHS;
5230 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5231 KnownZeroRHS, KnownOneRHS, Depth+1);
5232 KnownZero &= KnownZeroRHS;
5233 KnownOne &= KnownOneRHS;
5239 //===----------------------------------------------------------------------===//
5240 // ARM Inline Assembly Support
5241 //===----------------------------------------------------------------------===//
5243 /// getConstraintType - Given a constraint letter, return the type of
5244 /// constraint it is for this target.
5245 ARMTargetLowering::ConstraintType
5246 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5247 if (Constraint.size() == 1) {
5248 switch (Constraint[0]) {
5250 case 'l': return C_RegisterClass;
5251 case 'w': return C_RegisterClass;
5254 return TargetLowering::getConstraintType(Constraint);
5257 std::pair<unsigned, const TargetRegisterClass*>
5258 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5260 if (Constraint.size() == 1) {
5261 // GCC ARM Constraint Letters
5262 switch (Constraint[0]) {
5264 if (Subtarget->isThumb())
5265 return std::make_pair(0U, ARM::tGPRRegisterClass);
5267 return std::make_pair(0U, ARM::GPRRegisterClass);
5269 return std::make_pair(0U, ARM::GPRRegisterClass);
5272 return std::make_pair(0U, ARM::SPRRegisterClass);
5273 if (VT.getSizeInBits() == 64)
5274 return std::make_pair(0U, ARM::DPRRegisterClass);
5275 if (VT.getSizeInBits() == 128)
5276 return std::make_pair(0U, ARM::QPRRegisterClass);
5280 if (StringRef("{cc}").equals_lower(Constraint))
5281 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5283 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5286 std::vector<unsigned> ARMTargetLowering::
5287 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5289 if (Constraint.size() != 1)
5290 return std::vector<unsigned>();
5292 switch (Constraint[0]) { // GCC ARM Constraint Letters
5295 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5296 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5299 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5300 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5301 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5302 ARM::R12, ARM::LR, 0);
5305 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5306 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5307 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5308 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5309 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5310 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5311 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5312 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5313 if (VT.getSizeInBits() == 64)
5314 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5315 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5316 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5317 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5318 if (VT.getSizeInBits() == 128)
5319 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5320 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5324 return std::vector<unsigned>();
5327 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5328 /// vector. If it is invalid, don't add anything to Ops.
5329 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5331 std::vector<SDValue>&Ops,
5332 SelectionDAG &DAG) const {
5333 SDValue Result(0, 0);
5335 switch (Constraint) {
5337 case 'I': case 'J': case 'K': case 'L':
5338 case 'M': case 'N': case 'O':
5339 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5343 int64_t CVal64 = C->getSExtValue();
5344 int CVal = (int) CVal64;
5345 // None of these constraints allow values larger than 32 bits. Check
5346 // that the value fits in an int.
5350 switch (Constraint) {
5352 if (Subtarget->isThumb1Only()) {
5353 // This must be a constant between 0 and 255, for ADD
5355 if (CVal >= 0 && CVal <= 255)
5357 } else if (Subtarget->isThumb2()) {
5358 // A constant that can be used as an immediate value in a
5359 // data-processing instruction.
5360 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5363 // A constant that can be used as an immediate value in a
5364 // data-processing instruction.
5365 if (ARM_AM::getSOImmVal(CVal) != -1)
5371 if (Subtarget->isThumb()) { // FIXME thumb2
5372 // This must be a constant between -255 and -1, for negated ADD
5373 // immediates. This can be used in GCC with an "n" modifier that
5374 // prints the negated value, for use with SUB instructions. It is
5375 // not useful otherwise but is implemented for compatibility.
5376 if (CVal >= -255 && CVal <= -1)
5379 // This must be a constant between -4095 and 4095. It is not clear
5380 // what this constraint is intended for. Implemented for
5381 // compatibility with GCC.
5382 if (CVal >= -4095 && CVal <= 4095)
5388 if (Subtarget->isThumb1Only()) {
5389 // A 32-bit value where only one byte has a nonzero value. Exclude
5390 // zero to match GCC. This constraint is used by GCC internally for
5391 // constants that can be loaded with a move/shift combination.
5392 // It is not useful otherwise but is implemented for compatibility.
5393 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5395 } else if (Subtarget->isThumb2()) {
5396 // A constant whose bitwise inverse can be used as an immediate
5397 // value in a data-processing instruction. This can be used in GCC
5398 // with a "B" modifier that prints the inverted value, for use with
5399 // BIC and MVN instructions. It is not useful otherwise but is
5400 // implemented for compatibility.
5401 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5404 // A constant whose bitwise inverse can be used as an immediate
5405 // value in a data-processing instruction. This can be used in GCC
5406 // with a "B" modifier that prints the inverted value, for use with
5407 // BIC and MVN instructions. It is not useful otherwise but is
5408 // implemented for compatibility.
5409 if (ARM_AM::getSOImmVal(~CVal) != -1)
5415 if (Subtarget->isThumb1Only()) {
5416 // This must be a constant between -7 and 7,
5417 // for 3-operand ADD/SUB immediate instructions.
5418 if (CVal >= -7 && CVal < 7)
5420 } else if (Subtarget->isThumb2()) {
5421 // A constant whose negation can be used as an immediate value in a
5422 // data-processing instruction. This can be used in GCC with an "n"
5423 // modifier that prints the negated value, for use with SUB
5424 // instructions. It is not useful otherwise but is implemented for
5426 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5429 // A constant whose negation can be used as an immediate value in a
5430 // data-processing instruction. This can be used in GCC with an "n"
5431 // modifier that prints the negated value, for use with SUB
5432 // instructions. It is not useful otherwise but is implemented for
5434 if (ARM_AM::getSOImmVal(-CVal) != -1)
5440 if (Subtarget->isThumb()) { // FIXME thumb2
5441 // This must be a multiple of 4 between 0 and 1020, for
5442 // ADD sp + immediate.
5443 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5446 // A power of two or a constant between 0 and 32. This is used in
5447 // GCC for the shift amount on shifted register operands, but it is
5448 // useful in general for any shift amounts.
5449 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5455 if (Subtarget->isThumb()) { // FIXME thumb2
5456 // This must be a constant between 0 and 31, for shift amounts.
5457 if (CVal >= 0 && CVal <= 31)
5463 if (Subtarget->isThumb()) { // FIXME thumb2
5464 // This must be a multiple of 4 between -508 and 508, for
5465 // ADD/SUB sp = sp + immediate.
5466 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5471 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5475 if (Result.getNode()) {
5476 Ops.push_back(Result);
5479 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5483 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5484 // The ARM target isn't yet aware of offsets.
5488 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5489 APInt Imm = FPImm.bitcastToAPInt();
5490 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5491 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5492 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5494 // We can handle 4 bits of mantissa.
5495 // mantissa = (16+UInt(e:f:g:h))/16.
5496 if (Mantissa & 0x7ffff)
5499 if ((Mantissa & 0xf) != Mantissa)
5502 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5503 if (Exp < -3 || Exp > 4)
5505 Exp = ((Exp+3) & 0x7) ^ 4;
5507 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5510 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5511 APInt Imm = FPImm.bitcastToAPInt();
5512 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5513 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5514 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5516 // We can handle 4 bits of mantissa.
5517 // mantissa = (16+UInt(e:f:g:h))/16.
5518 if (Mantissa & 0xffffffffffffLL)
5521 if ((Mantissa & 0xf) != Mantissa)
5524 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5525 if (Exp < -3 || Exp > 4)
5527 Exp = ((Exp+3) & 0x7) ^ 4;
5529 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5532 bool ARM::isBitFieldInvertedMask(unsigned v) {
5533 if (v == 0xffffffff)
5535 // there can be 1's on either or both "outsides", all the "inside"
5537 unsigned int lsb = 0, msb = 31;
5538 while (v & (1 << msb)) --msb;
5539 while (v & (1 << lsb)) ++lsb;
5540 for (unsigned int i = lsb; i <= msb; ++i) {
5547 /// isFPImmLegal - Returns true if the target can instruction select the
5548 /// specified FP immediate natively. If false, the legalizer will
5549 /// materialize the FP immediate as a load from a constant pool.
5550 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5551 if (!Subtarget->hasVFP3())
5554 return ARM::getVFPf32Imm(Imm) != -1;
5556 return ARM::getVFPf64Imm(Imm) != -1;
5560 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5561 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5562 /// specified in the intrinsic calls.
5563 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5565 unsigned Intrinsic) const {
5566 switch (Intrinsic) {
5567 case Intrinsic::arm_neon_vld1:
5568 case Intrinsic::arm_neon_vld2:
5569 case Intrinsic::arm_neon_vld3:
5570 case Intrinsic::arm_neon_vld4:
5571 case Intrinsic::arm_neon_vld2lane:
5572 case Intrinsic::arm_neon_vld3lane:
5573 case Intrinsic::arm_neon_vld4lane: {
5574 Info.opc = ISD::INTRINSIC_W_CHAIN;
5575 // Conservatively set memVT to the entire set of vectors loaded.
5576 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
5577 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5578 Info.ptrVal = I.getArgOperand(0);
5580 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5581 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5582 Info.vol = false; // volatile loads with NEON intrinsics not supported
5583 Info.readMem = true;
5584 Info.writeMem = false;
5587 case Intrinsic::arm_neon_vst1:
5588 case Intrinsic::arm_neon_vst2:
5589 case Intrinsic::arm_neon_vst3:
5590 case Intrinsic::arm_neon_vst4:
5591 case Intrinsic::arm_neon_vst2lane:
5592 case Intrinsic::arm_neon_vst3lane:
5593 case Intrinsic::arm_neon_vst4lane: {
5594 Info.opc = ISD::INTRINSIC_VOID;
5595 // Conservatively set memVT to the entire set of vectors stored.
5596 unsigned NumElts = 0;
5597 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5598 const Type *ArgTy = I.getArgOperand(ArgI)->getType();
5599 if (!ArgTy->isVectorTy())
5601 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
5603 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5604 Info.ptrVal = I.getArgOperand(0);
5606 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5607 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5608 Info.vol = false; // volatile stores with NEON intrinsics not supported
5609 Info.readMem = false;
5610 Info.writeMem = true;