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/Intrinsics.h"
33 #include "llvm/Type.h"
34 #include "llvm/CodeGen/CallingConvLower.h"
35 #include "llvm/CodeGen/MachineBasicBlock.h"
36 #include "llvm/CodeGen/MachineFrameInfo.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineInstrBuilder.h"
39 #include "llvm/CodeGen/MachineRegisterInfo.h"
40 #include "llvm/CodeGen/PseudoSourceValue.h"
41 #include "llvm/CodeGen/SelectionDAG.h"
42 #include "llvm/MC/MCSectionMachO.h"
43 #include "llvm/Target/TargetOptions.h"
44 #include "llvm/ADT/VectorExtras.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
55 // This option should go away when tail calls fully work.
57 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
58 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
61 // This option should go away when Machine LICM is smart enough to hoist a
64 EnableARMVDUPsplat("arm-vdup-splat", cl::Hidden,
65 cl::desc("Generate VDUP for integer constant splats (TEMPORARY OPTION)."),
69 EnableARMLongCalls("arm-long-calls", cl::Hidden,
70 cl::desc("Generate calls via indirect call instructions"),
74 ARMInterworking("arm-interworking", cl::Hidden,
75 cl::desc("Enable / disable ARM interworking (for debugging only)"),
79 EnableARMCodePlacement("arm-code-placement", cl::Hidden,
80 cl::desc("Enable code placement pass for ARM"),
83 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
84 EVT PromotedBitwiseVT) {
85 if (VT != PromotedLdStVT) {
86 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
87 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
88 PromotedLdStVT.getSimpleVT());
90 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
91 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
92 PromotedLdStVT.getSimpleVT());
95 EVT ElemTy = VT.getVectorElementType();
96 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
97 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
98 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
99 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
100 if (ElemTy != MVT::i32) {
101 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
102 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
103 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
104 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
106 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
107 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
108 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
109 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
110 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
111 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
112 if (VT.isInteger()) {
113 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
114 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
115 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
116 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
117 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
119 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
121 // Promote all bit-wise operations.
122 if (VT.isInteger() && VT != PromotedBitwiseVT) {
123 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
124 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
125 PromotedBitwiseVT.getSimpleVT());
126 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
127 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
128 PromotedBitwiseVT.getSimpleVT());
129 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
130 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
131 PromotedBitwiseVT.getSimpleVT());
134 // Neon does not support vector divide/remainder operations.
135 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
136 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
137 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
138 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
139 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
143 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
144 addRegisterClass(VT, ARM::DPRRegisterClass);
145 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
148 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
149 addRegisterClass(VT, ARM::QPRRegisterClass);
150 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
153 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
154 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
155 return new TargetLoweringObjectFileMachO();
157 return new ARMElfTargetObjectFile();
160 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
161 : TargetLowering(TM, createTLOF(TM)) {
162 Subtarget = &TM.getSubtarget<ARMSubtarget>();
163 RegInfo = TM.getRegisterInfo();
164 Itins = TM.getInstrItineraryData();
166 if (Subtarget->isTargetDarwin()) {
167 // Uses VFP for Thumb libfuncs if available.
168 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
169 // Single-precision floating-point arithmetic.
170 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
171 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
172 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
173 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
175 // Double-precision floating-point arithmetic.
176 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
177 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
178 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
179 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
181 // Single-precision comparisons.
182 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
183 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
184 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
185 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
186 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
187 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
188 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
189 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
191 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
200 // Double-precision comparisons.
201 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
202 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
203 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
204 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
205 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
206 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
207 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
208 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
210 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
219 // Floating-point to integer conversions.
220 // i64 conversions are done via library routines even when generating VFP
221 // instructions, so use the same ones.
222 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
223 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
224 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
225 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
227 // Conversions between floating types.
228 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
229 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
231 // Integer to floating-point conversions.
232 // i64 conversions are done via library routines even when generating VFP
233 // instructions, so use the same ones.
234 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
235 // e.g., __floatunsidf vs. __floatunssidfvfp.
236 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
237 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
238 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
239 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
243 // These libcalls are not available in 32-bit.
244 setLibcallName(RTLIB::SHL_I128, 0);
245 setLibcallName(RTLIB::SRL_I128, 0);
246 setLibcallName(RTLIB::SRA_I128, 0);
248 // Libcalls should use the AAPCS base standard ABI, even if hard float
249 // is in effect, as per the ARM RTABI specification, section 4.1.2.
250 if (Subtarget->isAAPCS_ABI()) {
251 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
252 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
253 CallingConv::ARM_AAPCS);
257 if (Subtarget->isThumb1Only())
258 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
260 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
261 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
262 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
263 if (!Subtarget->isFPOnlySP())
264 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
266 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
269 if (Subtarget->hasNEON()) {
270 addDRTypeForNEON(MVT::v2f32);
271 addDRTypeForNEON(MVT::v8i8);
272 addDRTypeForNEON(MVT::v4i16);
273 addDRTypeForNEON(MVT::v2i32);
274 addDRTypeForNEON(MVT::v1i64);
276 addQRTypeForNEON(MVT::v4f32);
277 addQRTypeForNEON(MVT::v2f64);
278 addQRTypeForNEON(MVT::v16i8);
279 addQRTypeForNEON(MVT::v8i16);
280 addQRTypeForNEON(MVT::v4i32);
281 addQRTypeForNEON(MVT::v2i64);
283 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
284 // neither Neon nor VFP support any arithmetic operations on it.
285 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
286 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
287 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
288 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
289 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
290 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
291 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
292 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
293 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
294 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
295 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
296 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
297 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
298 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
299 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
300 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
301 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
302 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
303 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
304 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
305 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
306 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
307 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
308 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
310 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
312 // Neon does not support some operations on v1i64 and v2i64 types.
313 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
314 // Custom handling for some quad-vector types to detect VMULL.
315 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
316 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
317 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
318 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
319 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
321 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
322 setTargetDAGCombine(ISD::SHL);
323 setTargetDAGCombine(ISD::SRL);
324 setTargetDAGCombine(ISD::SRA);
325 setTargetDAGCombine(ISD::SIGN_EXTEND);
326 setTargetDAGCombine(ISD::ZERO_EXTEND);
327 setTargetDAGCombine(ISD::ANY_EXTEND);
328 setTargetDAGCombine(ISD::SELECT_CC);
331 computeRegisterProperties();
333 // ARM does not have f32 extending load.
334 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
336 // ARM does not have i1 sign extending load.
337 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
339 // ARM supports all 4 flavors of integer indexed load / store.
340 if (!Subtarget->isThumb1Only()) {
341 for (unsigned im = (unsigned)ISD::PRE_INC;
342 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
343 setIndexedLoadAction(im, MVT::i1, Legal);
344 setIndexedLoadAction(im, MVT::i8, Legal);
345 setIndexedLoadAction(im, MVT::i16, Legal);
346 setIndexedLoadAction(im, MVT::i32, Legal);
347 setIndexedStoreAction(im, MVT::i1, Legal);
348 setIndexedStoreAction(im, MVT::i8, Legal);
349 setIndexedStoreAction(im, MVT::i16, Legal);
350 setIndexedStoreAction(im, MVT::i32, Legal);
354 // i64 operation support.
355 if (Subtarget->isThumb1Only()) {
356 setOperationAction(ISD::MUL, MVT::i64, Expand);
357 setOperationAction(ISD::MULHU, MVT::i32, Expand);
358 setOperationAction(ISD::MULHS, MVT::i32, Expand);
359 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
360 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
362 setOperationAction(ISD::MUL, MVT::i64, Expand);
363 setOperationAction(ISD::MULHU, MVT::i32, Expand);
364 if (!Subtarget->hasV6Ops())
365 setOperationAction(ISD::MULHS, MVT::i32, Expand);
367 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
368 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
369 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
370 setOperationAction(ISD::SRL, MVT::i64, Custom);
371 setOperationAction(ISD::SRA, MVT::i64, Custom);
373 // ARM does not have ROTL.
374 setOperationAction(ISD::ROTL, MVT::i32, Expand);
375 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
376 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
377 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
378 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
380 // Only ARMv6 has BSWAP.
381 if (!Subtarget->hasV6Ops())
382 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
384 // These are expanded into libcalls.
385 if (!Subtarget->hasDivide()) {
386 // v7M has a hardware divider
387 setOperationAction(ISD::SDIV, MVT::i32, Expand);
388 setOperationAction(ISD::UDIV, MVT::i32, Expand);
390 setOperationAction(ISD::SREM, MVT::i32, Expand);
391 setOperationAction(ISD::UREM, MVT::i32, Expand);
392 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
393 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
395 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
396 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
397 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
398 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
399 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
401 setOperationAction(ISD::TRAP, MVT::Other, Legal);
403 // Use the default implementation.
404 setOperationAction(ISD::VASTART, MVT::Other, Custom);
405 setOperationAction(ISD::VAARG, MVT::Other, Expand);
406 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
407 setOperationAction(ISD::VAEND, MVT::Other, Expand);
408 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
409 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
410 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
411 // FIXME: Shouldn't need this, since no register is used, but the legalizer
412 // doesn't yet know how to not do that for SjLj.
413 setExceptionSelectorRegister(ARM::R0);
414 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
415 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
416 // the default expansion.
417 if (Subtarget->hasDataBarrier() ||
418 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
419 // membarrier needs custom lowering; the rest are legal and handled
421 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
423 // Set them all for expansion, which will force libcalls.
424 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
425 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
426 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
427 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
428 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
429 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
430 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
431 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
432 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
449 // Since the libcalls include locking, fold in the fences
450 setShouldFoldAtomicFences(true);
452 // 64-bit versions are always libcalls (for now)
453 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
454 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
455 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
456 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
457 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
458 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
459 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
460 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
462 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
463 if (!Subtarget->hasV6Ops()) {
464 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
465 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
467 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
469 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
470 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
471 // iff target supports vfp2.
472 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
473 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
476 // We want to custom lower some of our intrinsics.
477 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
478 if (Subtarget->isTargetDarwin()) {
479 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
480 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
483 setOperationAction(ISD::SETCC, MVT::i32, Expand);
484 setOperationAction(ISD::SETCC, MVT::f32, Expand);
485 setOperationAction(ISD::SETCC, MVT::f64, Expand);
486 setOperationAction(ISD::SELECT, MVT::i32, Custom);
487 setOperationAction(ISD::SELECT, MVT::f32, Custom);
488 setOperationAction(ISD::SELECT, MVT::f64, Custom);
489 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
490 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
491 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
493 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
494 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
495 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
496 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
497 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
499 // We don't support sin/cos/fmod/copysign/pow
500 setOperationAction(ISD::FSIN, MVT::f64, Expand);
501 setOperationAction(ISD::FSIN, MVT::f32, Expand);
502 setOperationAction(ISD::FCOS, MVT::f32, Expand);
503 setOperationAction(ISD::FCOS, MVT::f64, Expand);
504 setOperationAction(ISD::FREM, MVT::f64, Expand);
505 setOperationAction(ISD::FREM, MVT::f32, Expand);
506 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
507 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
508 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
510 setOperationAction(ISD::FPOW, MVT::f64, Expand);
511 setOperationAction(ISD::FPOW, MVT::f32, Expand);
513 // Various VFP goodness
514 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
515 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
516 if (Subtarget->hasVFP2()) {
517 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
518 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
519 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
520 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
522 // Special handling for half-precision FP.
523 if (!Subtarget->hasFP16()) {
524 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
525 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
529 // We have target-specific dag combine patterns for the following nodes:
530 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
531 setTargetDAGCombine(ISD::ADD);
532 setTargetDAGCombine(ISD::SUB);
533 setTargetDAGCombine(ISD::MUL);
535 if (Subtarget->hasV6T2Ops())
536 setTargetDAGCombine(ISD::OR);
538 setStackPointerRegisterToSaveRestore(ARM::SP);
540 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
541 setSchedulingPreference(Sched::RegPressure);
543 setSchedulingPreference(Sched::Hybrid);
545 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
547 // On ARM arguments smaller than 4 bytes are extended, so all arguments
548 // are at least 4 bytes aligned.
549 setMinStackArgumentAlignment(4);
551 if (EnableARMCodePlacement)
552 benefitFromCodePlacementOpt = true;
555 std::pair<const TargetRegisterClass*, uint8_t>
556 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
557 const TargetRegisterClass *RRC = 0;
559 switch (VT.getSimpleVT().SimpleTy) {
561 return TargetLowering::findRepresentativeClass(VT);
562 // Use DPR as representative register class for all floating point
563 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
564 // the cost is 1 for both f32 and f64.
565 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
566 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
567 RRC = ARM::DPRRegisterClass;
569 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
570 case MVT::v4f32: case MVT::v2f64:
571 RRC = ARM::DPRRegisterClass;
575 RRC = ARM::DPRRegisterClass;
579 RRC = ARM::DPRRegisterClass;
583 return std::make_pair(RRC, Cost);
586 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
589 case ARMISD::Wrapper: return "ARMISD::Wrapper";
590 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
591 case ARMISD::CALL: return "ARMISD::CALL";
592 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
593 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
594 case ARMISD::tCALL: return "ARMISD::tCALL";
595 case ARMISD::BRCOND: return "ARMISD::BRCOND";
596 case ARMISD::BR_JT: return "ARMISD::BR_JT";
597 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
598 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
599 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
600 case ARMISD::AND: return "ARMISD::AND";
601 case ARMISD::CMP: return "ARMISD::CMP";
602 case ARMISD::CMPZ: return "ARMISD::CMPZ";
603 case ARMISD::CMPFP: return "ARMISD::CMPFP";
604 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
605 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
606 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
607 case ARMISD::CMOV: return "ARMISD::CMOV";
608 case ARMISD::CNEG: return "ARMISD::CNEG";
610 case ARMISD::RBIT: return "ARMISD::RBIT";
612 case ARMISD::FTOSI: return "ARMISD::FTOSI";
613 case ARMISD::FTOUI: return "ARMISD::FTOUI";
614 case ARMISD::SITOF: return "ARMISD::SITOF";
615 case ARMISD::UITOF: return "ARMISD::UITOF";
617 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
618 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
619 case ARMISD::RRX: return "ARMISD::RRX";
621 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
622 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
624 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
625 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
627 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
629 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
631 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
633 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
634 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
636 case ARMISD::VCEQ: return "ARMISD::VCEQ";
637 case ARMISD::VCGE: return "ARMISD::VCGE";
638 case ARMISD::VCGEU: return "ARMISD::VCGEU";
639 case ARMISD::VCGT: return "ARMISD::VCGT";
640 case ARMISD::VCGTU: return "ARMISD::VCGTU";
641 case ARMISD::VTST: return "ARMISD::VTST";
643 case ARMISD::VSHL: return "ARMISD::VSHL";
644 case ARMISD::VSHRs: return "ARMISD::VSHRs";
645 case ARMISD::VSHRu: return "ARMISD::VSHRu";
646 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
647 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
648 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
649 case ARMISD::VSHRN: return "ARMISD::VSHRN";
650 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
651 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
652 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
653 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
654 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
655 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
656 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
657 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
658 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
659 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
660 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
661 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
662 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
663 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
664 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
665 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
666 case ARMISD::VDUP: return "ARMISD::VDUP";
667 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
668 case ARMISD::VEXT: return "ARMISD::VEXT";
669 case ARMISD::VREV64: return "ARMISD::VREV64";
670 case ARMISD::VREV32: return "ARMISD::VREV32";
671 case ARMISD::VREV16: return "ARMISD::VREV16";
672 case ARMISD::VZIP: return "ARMISD::VZIP";
673 case ARMISD::VUZP: return "ARMISD::VUZP";
674 case ARMISD::VTRN: return "ARMISD::VTRN";
675 case ARMISD::VMULLs: return "ARMISD::VMULLs";
676 case ARMISD::VMULLu: return "ARMISD::VMULLu";
677 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
678 case ARMISD::FMAX: return "ARMISD::FMAX";
679 case ARMISD::FMIN: return "ARMISD::FMIN";
680 case ARMISD::BFI: return "ARMISD::BFI";
684 /// getRegClassFor - Return the register class that should be used for the
685 /// specified value type.
686 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
687 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
688 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
689 // load / store 4 to 8 consecutive D registers.
690 if (Subtarget->hasNEON()) {
691 if (VT == MVT::v4i64)
692 return ARM::QQPRRegisterClass;
693 else if (VT == MVT::v8i64)
694 return ARM::QQQQPRRegisterClass;
696 return TargetLowering::getRegClassFor(VT);
699 // Create a fast isel object.
701 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
702 return ARM::createFastISel(funcInfo);
705 /// getFunctionAlignment - Return the Log2 alignment of this function.
706 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
707 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
710 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
711 /// be used for loads / stores from the global.
712 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
713 return (Subtarget->isThumb1Only() ? 127 : 4095);
716 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
717 unsigned NumVals = N->getNumValues();
719 return Sched::RegPressure;
721 for (unsigned i = 0; i != NumVals; ++i) {
722 EVT VT = N->getValueType(i);
723 if (VT.isFloatingPoint() || VT.isVector())
724 return Sched::Latency;
727 if (!N->isMachineOpcode())
728 return Sched::RegPressure;
730 // Load are scheduled for latency even if there instruction itinerary
732 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
733 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
735 return Sched::Latency;
737 if (!Itins->isEmpty() && Itins->getStageLatency(TID.getSchedClass()) > 2)
738 return Sched::Latency;
739 return Sched::RegPressure;
743 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
744 MachineFunction &MF) const {
745 switch (RC->getID()) {
748 case ARM::tGPRRegClassID:
749 return RegInfo->hasFP(MF) ? 4 : 5;
750 case ARM::GPRRegClassID: {
751 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
752 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
754 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
755 case ARM::DPRRegClassID:
760 //===----------------------------------------------------------------------===//
762 //===----------------------------------------------------------------------===//
764 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
765 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
767 default: llvm_unreachable("Unknown condition code!");
768 case ISD::SETNE: return ARMCC::NE;
769 case ISD::SETEQ: return ARMCC::EQ;
770 case ISD::SETGT: return ARMCC::GT;
771 case ISD::SETGE: return ARMCC::GE;
772 case ISD::SETLT: return ARMCC::LT;
773 case ISD::SETLE: return ARMCC::LE;
774 case ISD::SETUGT: return ARMCC::HI;
775 case ISD::SETUGE: return ARMCC::HS;
776 case ISD::SETULT: return ARMCC::LO;
777 case ISD::SETULE: return ARMCC::LS;
781 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
782 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
783 ARMCC::CondCodes &CondCode2) {
784 CondCode2 = ARMCC::AL;
786 default: llvm_unreachable("Unknown FP condition!");
788 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
790 case ISD::SETOGT: CondCode = ARMCC::GT; break;
792 case ISD::SETOGE: CondCode = ARMCC::GE; break;
793 case ISD::SETOLT: CondCode = ARMCC::MI; break;
794 case ISD::SETOLE: CondCode = ARMCC::LS; break;
795 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
796 case ISD::SETO: CondCode = ARMCC::VC; break;
797 case ISD::SETUO: CondCode = ARMCC::VS; break;
798 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
799 case ISD::SETUGT: CondCode = ARMCC::HI; break;
800 case ISD::SETUGE: CondCode = ARMCC::PL; break;
802 case ISD::SETULT: CondCode = ARMCC::LT; break;
804 case ISD::SETULE: CondCode = ARMCC::LE; break;
806 case ISD::SETUNE: CondCode = ARMCC::NE; break;
810 //===----------------------------------------------------------------------===//
811 // Calling Convention Implementation
812 //===----------------------------------------------------------------------===//
814 #include "ARMGenCallingConv.inc"
816 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
817 /// given CallingConvention value.
818 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
820 bool isVarArg) const {
823 llvm_unreachable("Unsupported calling convention");
825 case CallingConv::Fast:
826 // Use target triple & subtarget features to do actual dispatch.
827 if (Subtarget->isAAPCS_ABI()) {
828 if (Subtarget->hasVFP2() &&
829 FloatABIType == FloatABI::Hard && !isVarArg)
830 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
832 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
834 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
835 case CallingConv::ARM_AAPCS_VFP:
836 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
837 case CallingConv::ARM_AAPCS:
838 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
839 case CallingConv::ARM_APCS:
840 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
844 /// LowerCallResult - Lower the result values of a call into the
845 /// appropriate copies out of appropriate physical registers.
847 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
848 CallingConv::ID CallConv, bool isVarArg,
849 const SmallVectorImpl<ISD::InputArg> &Ins,
850 DebugLoc dl, SelectionDAG &DAG,
851 SmallVectorImpl<SDValue> &InVals) const {
853 // Assign locations to each value returned by this call.
854 SmallVector<CCValAssign, 16> RVLocs;
855 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
856 RVLocs, *DAG.getContext());
857 CCInfo.AnalyzeCallResult(Ins,
858 CCAssignFnForNode(CallConv, /* Return*/ true,
861 // Copy all of the result registers out of their specified physreg.
862 for (unsigned i = 0; i != RVLocs.size(); ++i) {
863 CCValAssign VA = RVLocs[i];
866 if (VA.needsCustom()) {
867 // Handle f64 or half of a v2f64.
868 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
870 Chain = Lo.getValue(1);
871 InFlag = Lo.getValue(2);
872 VA = RVLocs[++i]; // skip ahead to next loc
873 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
875 Chain = Hi.getValue(1);
876 InFlag = Hi.getValue(2);
877 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
879 if (VA.getLocVT() == MVT::v2f64) {
880 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
881 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
882 DAG.getConstant(0, MVT::i32));
884 VA = RVLocs[++i]; // skip ahead to next loc
885 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
886 Chain = Lo.getValue(1);
887 InFlag = Lo.getValue(2);
888 VA = RVLocs[++i]; // skip ahead to next loc
889 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
890 Chain = Hi.getValue(1);
891 InFlag = Hi.getValue(2);
892 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
893 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
894 DAG.getConstant(1, MVT::i32));
897 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
899 Chain = Val.getValue(1);
900 InFlag = Val.getValue(2);
903 switch (VA.getLocInfo()) {
904 default: llvm_unreachable("Unknown loc info!");
905 case CCValAssign::Full: break;
906 case CCValAssign::BCvt:
907 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
911 InVals.push_back(Val);
917 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
918 /// by "Src" to address "Dst" of size "Size". Alignment information is
919 /// specified by the specific parameter attribute. The copy will be passed as
920 /// a byval function parameter.
921 /// Sometimes what we are copying is the end of a larger object, the part that
922 /// does not fit in registers.
924 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
925 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
927 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
928 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
929 /*isVolatile=*/false, /*AlwaysInline=*/false,
933 /// LowerMemOpCallTo - Store the argument to the stack.
935 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
936 SDValue StackPtr, SDValue Arg,
937 DebugLoc dl, SelectionDAG &DAG,
938 const CCValAssign &VA,
939 ISD::ArgFlagsTy Flags) const {
940 unsigned LocMemOffset = VA.getLocMemOffset();
941 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
942 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
943 if (Flags.isByVal()) {
944 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
946 return DAG.getStore(Chain, dl, Arg, PtrOff,
947 PseudoSourceValue::getStack(), LocMemOffset,
951 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
952 SDValue Chain, SDValue &Arg,
953 RegsToPassVector &RegsToPass,
954 CCValAssign &VA, CCValAssign &NextVA,
956 SmallVector<SDValue, 8> &MemOpChains,
957 ISD::ArgFlagsTy Flags) const {
959 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
960 DAG.getVTList(MVT::i32, MVT::i32), Arg);
961 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
963 if (NextVA.isRegLoc())
964 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
966 assert(NextVA.isMemLoc());
967 if (StackPtr.getNode() == 0)
968 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
970 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
976 /// LowerCall - Lowering a call into a callseq_start <-
977 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
980 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
981 CallingConv::ID CallConv, bool isVarArg,
983 const SmallVectorImpl<ISD::OutputArg> &Outs,
984 const SmallVectorImpl<SDValue> &OutVals,
985 const SmallVectorImpl<ISD::InputArg> &Ins,
986 DebugLoc dl, SelectionDAG &DAG,
987 SmallVectorImpl<SDValue> &InVals) const {
988 MachineFunction &MF = DAG.getMachineFunction();
989 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
990 bool IsSibCall = false;
991 // Temporarily disable tail calls so things don't break.
992 if (!EnableARMTailCalls)
995 // Check if it's really possible to do a tail call.
996 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
997 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
998 Outs, OutVals, Ins, DAG);
999 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1000 // detected sibcalls.
1007 // Analyze operands of the call, assigning locations to each operand.
1008 SmallVector<CCValAssign, 16> ArgLocs;
1009 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1011 CCInfo.AnalyzeCallOperands(Outs,
1012 CCAssignFnForNode(CallConv, /* Return*/ false,
1015 // Get a count of how many bytes are to be pushed on the stack.
1016 unsigned NumBytes = CCInfo.getNextStackOffset();
1018 // For tail calls, memory operands are available in our caller's stack.
1022 // Adjust the stack pointer for the new arguments...
1023 // These operations are automatically eliminated by the prolog/epilog pass
1025 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1027 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1029 RegsToPassVector RegsToPass;
1030 SmallVector<SDValue, 8> MemOpChains;
1032 // Walk the register/memloc assignments, inserting copies/loads. In the case
1033 // of tail call optimization, arguments are handled later.
1034 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1036 ++i, ++realArgIdx) {
1037 CCValAssign &VA = ArgLocs[i];
1038 SDValue Arg = OutVals[realArgIdx];
1039 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1041 // Promote the value if needed.
1042 switch (VA.getLocInfo()) {
1043 default: llvm_unreachable("Unknown loc info!");
1044 case CCValAssign::Full: break;
1045 case CCValAssign::SExt:
1046 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1048 case CCValAssign::ZExt:
1049 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1051 case CCValAssign::AExt:
1052 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1054 case CCValAssign::BCvt:
1055 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1059 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1060 if (VA.needsCustom()) {
1061 if (VA.getLocVT() == MVT::v2f64) {
1062 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1063 DAG.getConstant(0, MVT::i32));
1064 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1065 DAG.getConstant(1, MVT::i32));
1067 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1068 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1070 VA = ArgLocs[++i]; // skip ahead to next loc
1071 if (VA.isRegLoc()) {
1072 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1073 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1075 assert(VA.isMemLoc());
1077 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1078 dl, DAG, VA, Flags));
1081 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1082 StackPtr, MemOpChains, Flags);
1084 } else if (VA.isRegLoc()) {
1085 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1086 } else if (!IsSibCall) {
1087 assert(VA.isMemLoc());
1089 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1090 dl, DAG, VA, Flags));
1094 if (!MemOpChains.empty())
1095 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1096 &MemOpChains[0], MemOpChains.size());
1098 // Build a sequence of copy-to-reg nodes chained together with token chain
1099 // and flag operands which copy the outgoing args into the appropriate regs.
1101 // Tail call byval lowering might overwrite argument registers so in case of
1102 // tail call optimization the copies to registers are lowered later.
1104 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1105 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1106 RegsToPass[i].second, InFlag);
1107 InFlag = Chain.getValue(1);
1110 // For tail calls lower the arguments to the 'real' stack slot.
1112 // Force all the incoming stack arguments to be loaded from the stack
1113 // before any new outgoing arguments are stored to the stack, because the
1114 // outgoing stack slots may alias the incoming argument stack slots, and
1115 // the alias isn't otherwise explicit. This is slightly more conservative
1116 // than necessary, because it means that each store effectively depends
1117 // on every argument instead of just those arguments it would clobber.
1119 // Do not flag preceeding copytoreg stuff together with the following stuff.
1121 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1122 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1123 RegsToPass[i].second, InFlag);
1124 InFlag = Chain.getValue(1);
1129 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1130 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1131 // node so that legalize doesn't hack it.
1132 bool isDirect = false;
1133 bool isARMFunc = false;
1134 bool isLocalARMFunc = false;
1135 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1137 if (EnableARMLongCalls) {
1138 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1139 && "long-calls with non-static relocation model!");
1140 // Handle a global address or an external symbol. If it's not one of
1141 // those, the target's already in a register, so we don't need to do
1143 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1144 const GlobalValue *GV = G->getGlobal();
1145 // Create a constant pool entry for the callee address
1146 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1147 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1150 // Get the address of the callee into a register
1151 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1152 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1153 Callee = DAG.getLoad(getPointerTy(), dl,
1154 DAG.getEntryNode(), CPAddr,
1155 PseudoSourceValue::getConstantPool(), 0,
1157 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1158 const char *Sym = S->getSymbol();
1160 // Create a constant pool entry for the callee address
1161 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1162 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1163 Sym, ARMPCLabelIndex, 0);
1164 // Get the address of the callee into a register
1165 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1166 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1167 Callee = DAG.getLoad(getPointerTy(), dl,
1168 DAG.getEntryNode(), CPAddr,
1169 PseudoSourceValue::getConstantPool(), 0,
1172 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1173 const GlobalValue *GV = G->getGlobal();
1175 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1176 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1177 getTargetMachine().getRelocationModel() != Reloc::Static;
1178 isARMFunc = !Subtarget->isThumb() || isStub;
1179 // ARM call to a local ARM function is predicable.
1180 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1181 // tBX takes a register source operand.
1182 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1183 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1184 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1187 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1188 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1189 Callee = DAG.getLoad(getPointerTy(), dl,
1190 DAG.getEntryNode(), CPAddr,
1191 PseudoSourceValue::getConstantPool(), 0,
1193 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1194 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1195 getPointerTy(), Callee, PICLabel);
1197 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1198 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1200 bool isStub = Subtarget->isTargetDarwin() &&
1201 getTargetMachine().getRelocationModel() != Reloc::Static;
1202 isARMFunc = !Subtarget->isThumb() || isStub;
1203 // tBX takes a register source operand.
1204 const char *Sym = S->getSymbol();
1205 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1206 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1207 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1208 Sym, ARMPCLabelIndex, 4);
1209 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1210 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1211 Callee = DAG.getLoad(getPointerTy(), dl,
1212 DAG.getEntryNode(), CPAddr,
1213 PseudoSourceValue::getConstantPool(), 0,
1215 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1216 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1217 getPointerTy(), Callee, PICLabel);
1219 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1222 // FIXME: handle tail calls differently.
1224 if (Subtarget->isThumb()) {
1225 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1226 CallOpc = ARMISD::CALL_NOLINK;
1228 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1230 CallOpc = (isDirect || Subtarget->hasV5TOps())
1231 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1232 : ARMISD::CALL_NOLINK;
1235 std::vector<SDValue> Ops;
1236 Ops.push_back(Chain);
1237 Ops.push_back(Callee);
1239 // Add argument registers to the end of the list so that they are known live
1241 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1242 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1243 RegsToPass[i].second.getValueType()));
1245 if (InFlag.getNode())
1246 Ops.push_back(InFlag);
1248 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1250 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1252 // Returns a chain and a flag for retval copy to use.
1253 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1254 InFlag = Chain.getValue(1);
1256 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1257 DAG.getIntPtrConstant(0, true), InFlag);
1259 InFlag = Chain.getValue(1);
1261 // Handle result values, copying them out of physregs into vregs that we
1263 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1267 /// MatchingStackOffset - Return true if the given stack call argument is
1268 /// already available in the same position (relatively) of the caller's
1269 /// incoming argument stack.
1271 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1272 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1273 const ARMInstrInfo *TII) {
1274 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1276 if (Arg.getOpcode() == ISD::CopyFromReg) {
1277 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1278 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1280 MachineInstr *Def = MRI->getVRegDef(VR);
1283 if (!Flags.isByVal()) {
1284 if (!TII->isLoadFromStackSlot(Def, FI))
1289 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1290 if (Flags.isByVal())
1291 // ByVal argument is passed in as a pointer but it's now being
1292 // dereferenced. e.g.
1293 // define @foo(%struct.X* %A) {
1294 // tail call @bar(%struct.X* byval %A)
1297 SDValue Ptr = Ld->getBasePtr();
1298 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1301 FI = FINode->getIndex();
1305 assert(FI != INT_MAX);
1306 if (!MFI->isFixedObjectIndex(FI))
1308 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1311 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1312 /// for tail call optimization. Targets which want to do tail call
1313 /// optimization should implement this function.
1315 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1316 CallingConv::ID CalleeCC,
1318 bool isCalleeStructRet,
1319 bool isCallerStructRet,
1320 const SmallVectorImpl<ISD::OutputArg> &Outs,
1321 const SmallVectorImpl<SDValue> &OutVals,
1322 const SmallVectorImpl<ISD::InputArg> &Ins,
1323 SelectionDAG& DAG) const {
1324 const Function *CallerF = DAG.getMachineFunction().getFunction();
1325 CallingConv::ID CallerCC = CallerF->getCallingConv();
1326 bool CCMatch = CallerCC == CalleeCC;
1328 // Look for obvious safe cases to perform tail call optimization that do not
1329 // require ABI changes. This is what gcc calls sibcall.
1331 // Do not sibcall optimize vararg calls unless the call site is not passing
1333 if (isVarArg && !Outs.empty())
1336 // Also avoid sibcall optimization if either caller or callee uses struct
1337 // return semantics.
1338 if (isCalleeStructRet || isCallerStructRet)
1341 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1342 // emitEpilogue is not ready for them.
1343 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1344 // LR. This means if we need to reload LR, it takes an extra instructions,
1345 // which outweighs the value of the tail call; but here we don't know yet
1346 // whether LR is going to be used. Probably the right approach is to
1347 // generate the tail call here and turn it back into CALL/RET in
1348 // emitEpilogue if LR is used.
1349 if (Subtarget->isThumb1Only())
1352 // For the moment, we can only do this to functions defined in this
1353 // compilation, or to indirect calls. A Thumb B to an ARM function,
1354 // or vice versa, is not easily fixed up in the linker unlike BL.
1355 // (We could do this by loading the address of the callee into a register;
1356 // that is an extra instruction over the direct call and burns a register
1357 // as well, so is not likely to be a win.)
1359 // It might be safe to remove this restriction on non-Darwin.
1361 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1362 // but we need to make sure there are enough registers; the only valid
1363 // registers are the 4 used for parameters. We don't currently do this
1365 if (isa<ExternalSymbolSDNode>(Callee))
1368 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1369 const GlobalValue *GV = G->getGlobal();
1370 if (GV->isDeclaration() || GV->isWeakForLinker())
1374 // If the calling conventions do not match, then we'd better make sure the
1375 // results are returned in the same way as what the caller expects.
1377 SmallVector<CCValAssign, 16> RVLocs1;
1378 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1379 RVLocs1, *DAG.getContext());
1380 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1382 SmallVector<CCValAssign, 16> RVLocs2;
1383 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1384 RVLocs2, *DAG.getContext());
1385 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1387 if (RVLocs1.size() != RVLocs2.size())
1389 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1390 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1392 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1394 if (RVLocs1[i].isRegLoc()) {
1395 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1398 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1404 // If the callee takes no arguments then go on to check the results of the
1406 if (!Outs.empty()) {
1407 // Check if stack adjustment is needed. For now, do not do this if any
1408 // argument is passed on the stack.
1409 SmallVector<CCValAssign, 16> ArgLocs;
1410 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1411 ArgLocs, *DAG.getContext());
1412 CCInfo.AnalyzeCallOperands(Outs,
1413 CCAssignFnForNode(CalleeCC, false, isVarArg));
1414 if (CCInfo.getNextStackOffset()) {
1415 MachineFunction &MF = DAG.getMachineFunction();
1417 // Check if the arguments are already laid out in the right way as
1418 // the caller's fixed stack objects.
1419 MachineFrameInfo *MFI = MF.getFrameInfo();
1420 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1421 const ARMInstrInfo *TII =
1422 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1423 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1425 ++i, ++realArgIdx) {
1426 CCValAssign &VA = ArgLocs[i];
1427 EVT RegVT = VA.getLocVT();
1428 SDValue Arg = OutVals[realArgIdx];
1429 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1430 if (VA.getLocInfo() == CCValAssign::Indirect)
1432 if (VA.needsCustom()) {
1433 // f64 and vector types are split into multiple registers or
1434 // register/stack-slot combinations. The types will not match
1435 // the registers; give up on memory f64 refs until we figure
1436 // out what to do about this.
1439 if (!ArgLocs[++i].isRegLoc())
1441 if (RegVT == MVT::v2f64) {
1442 if (!ArgLocs[++i].isRegLoc())
1444 if (!ArgLocs[++i].isRegLoc())
1447 } else if (!VA.isRegLoc()) {
1448 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1460 ARMTargetLowering::LowerReturn(SDValue Chain,
1461 CallingConv::ID CallConv, bool isVarArg,
1462 const SmallVectorImpl<ISD::OutputArg> &Outs,
1463 const SmallVectorImpl<SDValue> &OutVals,
1464 DebugLoc dl, SelectionDAG &DAG) const {
1466 // CCValAssign - represent the assignment of the return value to a location.
1467 SmallVector<CCValAssign, 16> RVLocs;
1469 // CCState - Info about the registers and stack slots.
1470 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1473 // Analyze outgoing return values.
1474 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1477 // If this is the first return lowered for this function, add
1478 // the regs to the liveout set for the function.
1479 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1480 for (unsigned i = 0; i != RVLocs.size(); ++i)
1481 if (RVLocs[i].isRegLoc())
1482 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1487 // Copy the result values into the output registers.
1488 for (unsigned i = 0, realRVLocIdx = 0;
1490 ++i, ++realRVLocIdx) {
1491 CCValAssign &VA = RVLocs[i];
1492 assert(VA.isRegLoc() && "Can only return in registers!");
1494 SDValue Arg = OutVals[realRVLocIdx];
1496 switch (VA.getLocInfo()) {
1497 default: llvm_unreachable("Unknown loc info!");
1498 case CCValAssign::Full: break;
1499 case CCValAssign::BCvt:
1500 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1504 if (VA.needsCustom()) {
1505 if (VA.getLocVT() == MVT::v2f64) {
1506 // Extract the first half and return it in two registers.
1507 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1508 DAG.getConstant(0, MVT::i32));
1509 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1510 DAG.getVTList(MVT::i32, MVT::i32), Half);
1512 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1513 Flag = Chain.getValue(1);
1514 VA = RVLocs[++i]; // skip ahead to next loc
1515 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1516 HalfGPRs.getValue(1), Flag);
1517 Flag = Chain.getValue(1);
1518 VA = RVLocs[++i]; // skip ahead to next loc
1520 // Extract the 2nd half and fall through to handle it as an f64 value.
1521 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1522 DAG.getConstant(1, MVT::i32));
1524 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1526 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1527 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1528 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1529 Flag = Chain.getValue(1);
1530 VA = RVLocs[++i]; // skip ahead to next loc
1531 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1534 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1536 // Guarantee that all emitted copies are
1537 // stuck together, avoiding something bad.
1538 Flag = Chain.getValue(1);
1543 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1545 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1550 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1551 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1552 // one of the above mentioned nodes. It has to be wrapped because otherwise
1553 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1554 // be used to form addressing mode. These wrapped nodes will be selected
1556 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1557 EVT PtrVT = Op.getValueType();
1558 // FIXME there is no actual debug info here
1559 DebugLoc dl = Op.getDebugLoc();
1560 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1562 if (CP->isMachineConstantPoolEntry())
1563 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1564 CP->getAlignment());
1566 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1567 CP->getAlignment());
1568 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1571 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1572 return MachineJumpTableInfo::EK_Inline;
1575 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1576 SelectionDAG &DAG) const {
1577 MachineFunction &MF = DAG.getMachineFunction();
1578 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1579 unsigned ARMPCLabelIndex = 0;
1580 DebugLoc DL = Op.getDebugLoc();
1581 EVT PtrVT = getPointerTy();
1582 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1583 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1585 if (RelocM == Reloc::Static) {
1586 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1588 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1589 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1590 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1591 ARMCP::CPBlockAddress,
1593 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1595 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1596 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1597 PseudoSourceValue::getConstantPool(), 0,
1599 if (RelocM == Reloc::Static)
1601 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1602 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1605 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1607 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1608 SelectionDAG &DAG) const {
1609 DebugLoc dl = GA->getDebugLoc();
1610 EVT PtrVT = getPointerTy();
1611 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1612 MachineFunction &MF = DAG.getMachineFunction();
1613 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1614 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1615 ARMConstantPoolValue *CPV =
1616 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1617 ARMCP::CPValue, PCAdj, "tlsgd", true);
1618 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1619 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1620 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1621 PseudoSourceValue::getConstantPool(), 0,
1623 SDValue Chain = Argument.getValue(1);
1625 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1626 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1628 // call __tls_get_addr.
1631 Entry.Node = Argument;
1632 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1633 Args.push_back(Entry);
1634 // FIXME: is there useful debug info available here?
1635 std::pair<SDValue, SDValue> CallResult =
1636 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1637 false, false, false, false,
1638 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1639 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1640 return CallResult.first;
1643 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1644 // "local exec" model.
1646 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1647 SelectionDAG &DAG) const {
1648 const GlobalValue *GV = GA->getGlobal();
1649 DebugLoc dl = GA->getDebugLoc();
1651 SDValue Chain = DAG.getEntryNode();
1652 EVT PtrVT = getPointerTy();
1653 // Get the Thread Pointer
1654 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1656 if (GV->isDeclaration()) {
1657 MachineFunction &MF = DAG.getMachineFunction();
1658 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1659 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1660 // Initial exec model.
1661 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1662 ARMConstantPoolValue *CPV =
1663 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1664 ARMCP::CPValue, PCAdj, "gottpoff", true);
1665 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1666 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1667 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1668 PseudoSourceValue::getConstantPool(), 0,
1670 Chain = Offset.getValue(1);
1672 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1673 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1675 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1676 PseudoSourceValue::getConstantPool(), 0,
1680 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1681 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1682 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1683 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1684 PseudoSourceValue::getConstantPool(), 0,
1688 // The address of the thread local variable is the add of the thread
1689 // pointer with the offset of the variable.
1690 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1694 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1695 // TODO: implement the "local dynamic" model
1696 assert(Subtarget->isTargetELF() &&
1697 "TLS not implemented for non-ELF targets");
1698 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1699 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1700 // otherwise use the "Local Exec" TLS Model
1701 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1702 return LowerToTLSGeneralDynamicModel(GA, DAG);
1704 return LowerToTLSExecModels(GA, DAG);
1707 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1708 SelectionDAG &DAG) const {
1709 EVT PtrVT = getPointerTy();
1710 DebugLoc dl = Op.getDebugLoc();
1711 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1712 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1713 if (RelocM == Reloc::PIC_) {
1714 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1715 ARMConstantPoolValue *CPV =
1716 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1717 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1718 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1719 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1721 PseudoSourceValue::getConstantPool(), 0,
1723 SDValue Chain = Result.getValue(1);
1724 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1725 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1727 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1728 PseudoSourceValue::getGOT(), 0,
1732 // If we have T2 ops, we can materialize the address directly via movt/movw
1733 // pair. This is always cheaper.
1734 if (Subtarget->useMovt()) {
1735 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1736 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1738 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1739 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1740 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1741 PseudoSourceValue::getConstantPool(), 0,
1747 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1748 SelectionDAG &DAG) const {
1749 MachineFunction &MF = DAG.getMachineFunction();
1750 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1751 unsigned ARMPCLabelIndex = 0;
1752 EVT PtrVT = getPointerTy();
1753 DebugLoc dl = Op.getDebugLoc();
1754 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1755 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1757 if (RelocM == Reloc::Static)
1758 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1760 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1761 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1762 ARMConstantPoolValue *CPV =
1763 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1764 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1766 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1768 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1769 PseudoSourceValue::getConstantPool(), 0,
1771 SDValue Chain = Result.getValue(1);
1773 if (RelocM == Reloc::PIC_) {
1774 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1775 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1778 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1779 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1780 PseudoSourceValue::getGOT(), 0,
1786 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1787 SelectionDAG &DAG) const {
1788 assert(Subtarget->isTargetELF() &&
1789 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1790 MachineFunction &MF = DAG.getMachineFunction();
1791 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1792 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1793 EVT PtrVT = getPointerTy();
1794 DebugLoc dl = Op.getDebugLoc();
1795 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1796 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1797 "_GLOBAL_OFFSET_TABLE_",
1798 ARMPCLabelIndex, PCAdj);
1799 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1800 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1801 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1802 PseudoSourceValue::getConstantPool(), 0,
1804 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1805 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1809 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1810 DebugLoc dl = Op.getDebugLoc();
1811 SDValue Val = DAG.getConstant(0, MVT::i32);
1812 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1813 Op.getOperand(1), Val);
1817 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1818 DebugLoc dl = Op.getDebugLoc();
1819 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1820 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1824 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1825 const ARMSubtarget *Subtarget) const {
1826 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1827 DebugLoc dl = Op.getDebugLoc();
1829 default: return SDValue(); // Don't custom lower most intrinsics.
1830 case Intrinsic::arm_thread_pointer: {
1831 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1832 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1834 case Intrinsic::eh_sjlj_lsda: {
1835 MachineFunction &MF = DAG.getMachineFunction();
1836 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1837 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1838 EVT PtrVT = getPointerTy();
1839 DebugLoc dl = Op.getDebugLoc();
1840 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1842 unsigned PCAdj = (RelocM != Reloc::PIC_)
1843 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1844 ARMConstantPoolValue *CPV =
1845 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1846 ARMCP::CPLSDA, PCAdj);
1847 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1848 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1850 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1851 PseudoSourceValue::getConstantPool(), 0,
1854 if (RelocM == Reloc::PIC_) {
1855 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1856 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1863 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1864 const ARMSubtarget *Subtarget) {
1865 DebugLoc dl = Op.getDebugLoc();
1866 SDValue Op5 = Op.getOperand(5);
1867 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1868 // Some subtargets which have dmb and dsb instructions can handle barriers
1869 // directly. Some ARMv6 cpus can support them with the help of mcr
1870 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1872 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1873 if (Subtarget->hasDataBarrier())
1874 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1876 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
1877 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1878 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1879 DAG.getConstant(0, MVT::i32));
1883 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1884 MachineFunction &MF = DAG.getMachineFunction();
1885 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1887 // vastart just stores the address of the VarArgsFrameIndex slot into the
1888 // memory location argument.
1889 DebugLoc dl = Op.getDebugLoc();
1890 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1891 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1892 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1893 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1898 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1899 SDValue &Root, SelectionDAG &DAG,
1900 DebugLoc dl) const {
1901 MachineFunction &MF = DAG.getMachineFunction();
1902 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1904 TargetRegisterClass *RC;
1905 if (AFI->isThumb1OnlyFunction())
1906 RC = ARM::tGPRRegisterClass;
1908 RC = ARM::GPRRegisterClass;
1910 // Transform the arguments stored in physical registers into virtual ones.
1911 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1912 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1915 if (NextVA.isMemLoc()) {
1916 MachineFrameInfo *MFI = MF.getFrameInfo();
1917 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
1919 // Create load node to retrieve arguments from the stack.
1920 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1921 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1922 PseudoSourceValue::getFixedStack(FI), 0,
1925 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1926 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1929 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1933 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1934 CallingConv::ID CallConv, bool isVarArg,
1935 const SmallVectorImpl<ISD::InputArg>
1937 DebugLoc dl, SelectionDAG &DAG,
1938 SmallVectorImpl<SDValue> &InVals)
1941 MachineFunction &MF = DAG.getMachineFunction();
1942 MachineFrameInfo *MFI = MF.getFrameInfo();
1944 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1946 // Assign locations to all of the incoming arguments.
1947 SmallVector<CCValAssign, 16> ArgLocs;
1948 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1950 CCInfo.AnalyzeFormalArguments(Ins,
1951 CCAssignFnForNode(CallConv, /* Return*/ false,
1954 SmallVector<SDValue, 16> ArgValues;
1956 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1957 CCValAssign &VA = ArgLocs[i];
1959 // Arguments stored in registers.
1960 if (VA.isRegLoc()) {
1961 EVT RegVT = VA.getLocVT();
1964 if (VA.needsCustom()) {
1965 // f64 and vector types are split up into multiple registers or
1966 // combinations of registers and stack slots.
1967 if (VA.getLocVT() == MVT::v2f64) {
1968 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1970 VA = ArgLocs[++i]; // skip ahead to next loc
1972 if (VA.isMemLoc()) {
1973 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
1974 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1975 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1976 PseudoSourceValue::getFixedStack(FI), 0,
1979 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1982 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1983 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1984 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1985 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1986 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1988 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1991 TargetRegisterClass *RC;
1993 if (RegVT == MVT::f32)
1994 RC = ARM::SPRRegisterClass;
1995 else if (RegVT == MVT::f64)
1996 RC = ARM::DPRRegisterClass;
1997 else if (RegVT == MVT::v2f64)
1998 RC = ARM::QPRRegisterClass;
1999 else if (RegVT == MVT::i32)
2000 RC = (AFI->isThumb1OnlyFunction() ?
2001 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2003 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2005 // Transform the arguments in physical registers into virtual ones.
2006 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2007 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2010 // If this is an 8 or 16-bit value, it is really passed promoted
2011 // to 32 bits. Insert an assert[sz]ext to capture this, then
2012 // truncate to the right size.
2013 switch (VA.getLocInfo()) {
2014 default: llvm_unreachable("Unknown loc info!");
2015 case CCValAssign::Full: break;
2016 case CCValAssign::BCvt:
2017 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2019 case CCValAssign::SExt:
2020 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2021 DAG.getValueType(VA.getValVT()));
2022 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2024 case CCValAssign::ZExt:
2025 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2026 DAG.getValueType(VA.getValVT()));
2027 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2031 InVals.push_back(ArgValue);
2033 } else { // VA.isRegLoc()
2036 assert(VA.isMemLoc());
2037 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2039 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2040 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2042 // Create load nodes to retrieve arguments from the stack.
2043 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2044 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2045 PseudoSourceValue::getFixedStack(FI), 0,
2052 static const unsigned GPRArgRegs[] = {
2053 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2056 unsigned NumGPRs = CCInfo.getFirstUnallocated
2057 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2059 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2060 unsigned VARegSize = (4 - NumGPRs) * 4;
2061 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2062 unsigned ArgOffset = CCInfo.getNextStackOffset();
2063 if (VARegSaveSize) {
2064 // If this function is vararg, store any remaining integer argument regs
2065 // to their spots on the stack so that they may be loaded by deferencing
2066 // the result of va_next.
2067 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2068 AFI->setVarArgsFrameIndex(
2069 MFI->CreateFixedObject(VARegSaveSize,
2070 ArgOffset + VARegSaveSize - VARegSize,
2072 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2075 SmallVector<SDValue, 4> MemOps;
2076 for (; NumGPRs < 4; ++NumGPRs) {
2077 TargetRegisterClass *RC;
2078 if (AFI->isThumb1OnlyFunction())
2079 RC = ARM::tGPRRegisterClass;
2081 RC = ARM::GPRRegisterClass;
2083 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2084 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2086 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2087 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2088 0, false, false, 0);
2089 MemOps.push_back(Store);
2090 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2091 DAG.getConstant(4, getPointerTy()));
2093 if (!MemOps.empty())
2094 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2095 &MemOps[0], MemOps.size());
2097 // This will point to the next argument passed via stack.
2098 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2104 /// isFloatingPointZero - Return true if this is +0.0.
2105 static bool isFloatingPointZero(SDValue Op) {
2106 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2107 return CFP->getValueAPF().isPosZero();
2108 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2109 // Maybe this has already been legalized into the constant pool?
2110 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2111 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2112 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2113 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2114 return CFP->getValueAPF().isPosZero();
2120 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2121 /// the given operands.
2123 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2124 SDValue &ARMcc, SelectionDAG &DAG,
2125 DebugLoc dl) const {
2126 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2127 unsigned C = RHSC->getZExtValue();
2128 if (!isLegalICmpImmediate(C)) {
2129 // Constant does not fit, try adjusting it by one?
2134 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2135 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2136 RHS = DAG.getConstant(C-1, MVT::i32);
2141 if (C != 0 && isLegalICmpImmediate(C-1)) {
2142 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2143 RHS = DAG.getConstant(C-1, MVT::i32);
2148 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2149 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2150 RHS = DAG.getConstant(C+1, MVT::i32);
2155 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2156 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2157 RHS = DAG.getConstant(C+1, MVT::i32);
2164 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2165 ARMISD::NodeType CompareType;
2168 CompareType = ARMISD::CMP;
2173 CompareType = ARMISD::CMPZ;
2176 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2177 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2180 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2182 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2183 DebugLoc dl) const {
2185 if (!isFloatingPointZero(RHS))
2186 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2188 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2189 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2192 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2193 SDValue Cond = Op.getOperand(0);
2194 SDValue SelectTrue = Op.getOperand(1);
2195 SDValue SelectFalse = Op.getOperand(2);
2196 DebugLoc dl = Op.getDebugLoc();
2200 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2201 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2203 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2204 const ConstantSDNode *CMOVTrue =
2205 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2206 const ConstantSDNode *CMOVFalse =
2207 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2209 if (CMOVTrue && CMOVFalse) {
2210 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2211 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2215 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2217 False = SelectFalse;
2218 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2223 if (True.getNode() && False.getNode()) {
2224 EVT VT = Cond.getValueType();
2225 SDValue ARMcc = Cond.getOperand(2);
2226 SDValue CCR = Cond.getOperand(3);
2227 SDValue Cmp = Cond.getOperand(4);
2228 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2233 return DAG.getSelectCC(dl, Cond,
2234 DAG.getConstant(0, Cond.getValueType()),
2235 SelectTrue, SelectFalse, ISD::SETNE);
2238 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2239 EVT VT = Op.getValueType();
2240 SDValue LHS = Op.getOperand(0);
2241 SDValue RHS = Op.getOperand(1);
2242 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2243 SDValue TrueVal = Op.getOperand(2);
2244 SDValue FalseVal = Op.getOperand(3);
2245 DebugLoc dl = Op.getDebugLoc();
2247 if (LHS.getValueType() == MVT::i32) {
2249 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2250 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2251 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2254 ARMCC::CondCodes CondCode, CondCode2;
2255 FPCCToARMCC(CC, CondCode, CondCode2);
2257 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2258 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2259 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2260 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2262 if (CondCode2 != ARMCC::AL) {
2263 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2264 // FIXME: Needs another CMP because flag can have but one use.
2265 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2266 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2267 Result, TrueVal, ARMcc2, CCR, Cmp2);
2272 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2273 /// to morph to an integer compare sequence.
2274 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2275 const ARMSubtarget *Subtarget) {
2276 SDNode *N = Op.getNode();
2277 if (!N->hasOneUse())
2278 // Otherwise it requires moving the value from fp to integer registers.
2280 if (!N->getNumValues())
2282 EVT VT = Op.getValueType();
2283 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2284 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2285 // vmrs are very slow, e.g. cortex-a8.
2288 if (isFloatingPointZero(Op)) {
2292 return ISD::isNormalLoad(N);
2295 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2296 if (isFloatingPointZero(Op))
2297 return DAG.getConstant(0, MVT::i32);
2299 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2300 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2301 Ld->getChain(), Ld->getBasePtr(),
2302 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2303 Ld->isVolatile(), Ld->isNonTemporal(),
2304 Ld->getAlignment());
2306 llvm_unreachable("Unknown VFP cmp argument!");
2309 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2310 SDValue &RetVal1, SDValue &RetVal2) {
2311 if (isFloatingPointZero(Op)) {
2312 RetVal1 = DAG.getConstant(0, MVT::i32);
2313 RetVal2 = DAG.getConstant(0, MVT::i32);
2317 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2318 SDValue Ptr = Ld->getBasePtr();
2319 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2320 Ld->getChain(), Ptr,
2321 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2322 Ld->isVolatile(), Ld->isNonTemporal(),
2323 Ld->getAlignment());
2325 EVT PtrType = Ptr.getValueType();
2326 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2327 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2328 PtrType, Ptr, DAG.getConstant(4, PtrType));
2329 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2330 Ld->getChain(), NewPtr,
2331 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2332 Ld->isVolatile(), Ld->isNonTemporal(),
2337 llvm_unreachable("Unknown VFP cmp argument!");
2340 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2341 /// f32 and even f64 comparisons to integer ones.
2343 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2344 SDValue Chain = Op.getOperand(0);
2345 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2346 SDValue LHS = Op.getOperand(2);
2347 SDValue RHS = Op.getOperand(3);
2348 SDValue Dest = Op.getOperand(4);
2349 DebugLoc dl = Op.getDebugLoc();
2351 bool SeenZero = false;
2352 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2353 canChangeToInt(RHS, SeenZero, Subtarget) &&
2354 // If one of the operand is zero, it's safe to ignore the NaN case since
2355 // we only care about equality comparisons.
2356 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2357 // If unsafe fp math optimization is enabled and there are no othter uses of
2358 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2359 // to an integer comparison.
2360 if (CC == ISD::SETOEQ)
2362 else if (CC == ISD::SETUNE)
2366 if (LHS.getValueType() == MVT::f32) {
2367 LHS = bitcastf32Toi32(LHS, DAG);
2368 RHS = bitcastf32Toi32(RHS, DAG);
2369 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2370 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2371 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2372 Chain, Dest, ARMcc, CCR, Cmp);
2377 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2378 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2379 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2380 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2381 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2382 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2383 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2389 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2390 SDValue Chain = Op.getOperand(0);
2391 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2392 SDValue LHS = Op.getOperand(2);
2393 SDValue RHS = Op.getOperand(3);
2394 SDValue Dest = Op.getOperand(4);
2395 DebugLoc dl = Op.getDebugLoc();
2397 if (LHS.getValueType() == MVT::i32) {
2399 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2400 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2401 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2402 Chain, Dest, ARMcc, CCR, Cmp);
2405 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2408 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2409 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2410 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2411 if (Result.getNode())
2415 ARMCC::CondCodes CondCode, CondCode2;
2416 FPCCToARMCC(CC, CondCode, CondCode2);
2418 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2419 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2420 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2421 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2422 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2423 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2424 if (CondCode2 != ARMCC::AL) {
2425 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2426 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2427 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2432 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2433 SDValue Chain = Op.getOperand(0);
2434 SDValue Table = Op.getOperand(1);
2435 SDValue Index = Op.getOperand(2);
2436 DebugLoc dl = Op.getDebugLoc();
2438 EVT PTy = getPointerTy();
2439 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2440 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2441 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2442 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2443 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2444 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2445 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2446 if (Subtarget->isThumb2()) {
2447 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2448 // which does another jump to the destination. This also makes it easier
2449 // to translate it to TBB / TBH later.
2450 // FIXME: This might not work if the function is extremely large.
2451 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2452 Addr, Op.getOperand(2), JTI, UId);
2454 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2455 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2456 PseudoSourceValue::getJumpTable(), 0,
2458 Chain = Addr.getValue(1);
2459 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2460 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2462 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2463 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2464 Chain = Addr.getValue(1);
2465 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2469 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2470 DebugLoc dl = Op.getDebugLoc();
2473 switch (Op.getOpcode()) {
2475 assert(0 && "Invalid opcode!");
2476 case ISD::FP_TO_SINT:
2477 Opc = ARMISD::FTOSI;
2479 case ISD::FP_TO_UINT:
2480 Opc = ARMISD::FTOUI;
2483 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2484 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2487 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2488 EVT VT = Op.getValueType();
2489 DebugLoc dl = Op.getDebugLoc();
2492 switch (Op.getOpcode()) {
2494 assert(0 && "Invalid opcode!");
2495 case ISD::SINT_TO_FP:
2496 Opc = ARMISD::SITOF;
2498 case ISD::UINT_TO_FP:
2499 Opc = ARMISD::UITOF;
2503 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2504 return DAG.getNode(Opc, dl, VT, Op);
2507 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2508 // Implement fcopysign with a fabs and a conditional fneg.
2509 SDValue Tmp0 = Op.getOperand(0);
2510 SDValue Tmp1 = Op.getOperand(1);
2511 DebugLoc dl = Op.getDebugLoc();
2512 EVT VT = Op.getValueType();
2513 EVT SrcVT = Tmp1.getValueType();
2514 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2515 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2516 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2517 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2518 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2519 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2522 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2523 MachineFunction &MF = DAG.getMachineFunction();
2524 MachineFrameInfo *MFI = MF.getFrameInfo();
2525 MFI->setReturnAddressIsTaken(true);
2527 EVT VT = Op.getValueType();
2528 DebugLoc dl = Op.getDebugLoc();
2529 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2531 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2532 SDValue Offset = DAG.getConstant(4, MVT::i32);
2533 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2534 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2535 NULL, 0, false, false, 0);
2538 // Return LR, which contains the return address. Mark it an implicit live-in.
2539 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2540 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2543 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2544 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2545 MFI->setFrameAddressIsTaken(true);
2547 EVT VT = Op.getValueType();
2548 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2549 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2550 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2551 ? ARM::R7 : ARM::R11;
2552 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2554 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2559 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2560 /// expand a bit convert where either the source or destination type is i64 to
2561 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2562 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2563 /// vectors), since the legalizer won't know what to do with that.
2564 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2565 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2566 DebugLoc dl = N->getDebugLoc();
2567 SDValue Op = N->getOperand(0);
2569 // This function is only supposed to be called for i64 types, either as the
2570 // source or destination of the bit convert.
2571 EVT SrcVT = Op.getValueType();
2572 EVT DstVT = N->getValueType(0);
2573 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2574 "ExpandBIT_CONVERT called for non-i64 type");
2576 // Turn i64->f64 into VMOVDRR.
2577 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2578 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2579 DAG.getConstant(0, MVT::i32));
2580 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2581 DAG.getConstant(1, MVT::i32));
2582 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2583 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2586 // Turn f64->i64 into VMOVRRD.
2587 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2588 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2589 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2590 // Merge the pieces into a single i64 value.
2591 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2597 /// getZeroVector - Returns a vector of specified type with all zero elements.
2598 /// Zero vectors are used to represent vector negation and in those cases
2599 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2600 /// not support i64 elements, so sometimes the zero vectors will need to be
2601 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2603 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2604 assert(VT.isVector() && "Expected a vector type");
2605 // The canonical modified immediate encoding of a zero vector is....0!
2606 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2607 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2608 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2609 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2612 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2613 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2614 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2615 SelectionDAG &DAG) const {
2616 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2617 EVT VT = Op.getValueType();
2618 unsigned VTBits = VT.getSizeInBits();
2619 DebugLoc dl = Op.getDebugLoc();
2620 SDValue ShOpLo = Op.getOperand(0);
2621 SDValue ShOpHi = Op.getOperand(1);
2622 SDValue ShAmt = Op.getOperand(2);
2624 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2626 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2628 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2629 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2630 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2631 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2632 DAG.getConstant(VTBits, MVT::i32));
2633 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2634 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2635 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2637 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2638 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2640 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2641 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2644 SDValue Ops[2] = { Lo, Hi };
2645 return DAG.getMergeValues(Ops, 2, dl);
2648 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2649 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2650 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2651 SelectionDAG &DAG) const {
2652 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2653 EVT VT = Op.getValueType();
2654 unsigned VTBits = VT.getSizeInBits();
2655 DebugLoc dl = Op.getDebugLoc();
2656 SDValue ShOpLo = Op.getOperand(0);
2657 SDValue ShOpHi = Op.getOperand(1);
2658 SDValue ShAmt = Op.getOperand(2);
2661 assert(Op.getOpcode() == ISD::SHL_PARTS);
2662 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2663 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2664 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2665 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2666 DAG.getConstant(VTBits, MVT::i32));
2667 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2668 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2670 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2671 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2672 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2674 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2675 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2678 SDValue Ops[2] = { Lo, Hi };
2679 return DAG.getMergeValues(Ops, 2, dl);
2682 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2683 SelectionDAG &DAG) const {
2684 // The rounding mode is in bits 23:22 of the FPSCR.
2685 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2686 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2687 // so that the shift + and get folded into a bitfield extract.
2688 DebugLoc dl = Op.getDebugLoc();
2689 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2690 DAG.getConstant(Intrinsic::arm_get_fpscr,
2692 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2693 DAG.getConstant(1U << 22, MVT::i32));
2694 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2695 DAG.getConstant(22, MVT::i32));
2696 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2697 DAG.getConstant(3, MVT::i32));
2700 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2701 const ARMSubtarget *ST) {
2702 EVT VT = N->getValueType(0);
2703 DebugLoc dl = N->getDebugLoc();
2705 if (!ST->hasV6T2Ops())
2708 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2709 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2712 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2713 const ARMSubtarget *ST) {
2714 EVT VT = N->getValueType(0);
2715 DebugLoc dl = N->getDebugLoc();
2717 // Lower vector shifts on NEON to use VSHL.
2718 if (VT.isVector()) {
2719 assert(ST->hasNEON() && "unexpected vector shift");
2721 // Left shifts translate directly to the vshiftu intrinsic.
2722 if (N->getOpcode() == ISD::SHL)
2723 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2724 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2725 N->getOperand(0), N->getOperand(1));
2727 assert((N->getOpcode() == ISD::SRA ||
2728 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2730 // NEON uses the same intrinsics for both left and right shifts. For
2731 // right shifts, the shift amounts are negative, so negate the vector of
2733 EVT ShiftVT = N->getOperand(1).getValueType();
2734 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2735 getZeroVector(ShiftVT, DAG, dl),
2737 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2738 Intrinsic::arm_neon_vshifts :
2739 Intrinsic::arm_neon_vshiftu);
2740 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2741 DAG.getConstant(vshiftInt, MVT::i32),
2742 N->getOperand(0), NegatedCount);
2745 // We can get here for a node like i32 = ISD::SHL i32, i64
2749 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2750 "Unknown shift to lower!");
2752 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2753 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2754 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2757 // If we are in thumb mode, we don't have RRX.
2758 if (ST->isThumb1Only()) return SDValue();
2760 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2761 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2762 DAG.getConstant(0, MVT::i32));
2763 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2764 DAG.getConstant(1, MVT::i32));
2766 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2767 // captures the result into a carry flag.
2768 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2769 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2771 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2772 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2774 // Merge the pieces into a single i64 value.
2775 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2778 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2779 SDValue TmpOp0, TmpOp1;
2780 bool Invert = false;
2784 SDValue Op0 = Op.getOperand(0);
2785 SDValue Op1 = Op.getOperand(1);
2786 SDValue CC = Op.getOperand(2);
2787 EVT VT = Op.getValueType();
2788 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2789 DebugLoc dl = Op.getDebugLoc();
2791 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2792 switch (SetCCOpcode) {
2793 default: llvm_unreachable("Illegal FP comparison"); break;
2795 case ISD::SETNE: Invert = true; // Fallthrough
2797 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2799 case ISD::SETLT: Swap = true; // Fallthrough
2801 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2803 case ISD::SETLE: Swap = true; // Fallthrough
2805 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2806 case ISD::SETUGE: Swap = true; // Fallthrough
2807 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2808 case ISD::SETUGT: Swap = true; // Fallthrough
2809 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2810 case ISD::SETUEQ: Invert = true; // Fallthrough
2812 // Expand this to (OLT | OGT).
2816 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2817 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2819 case ISD::SETUO: Invert = true; // Fallthrough
2821 // Expand this to (OLT | OGE).
2825 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2826 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2830 // Integer comparisons.
2831 switch (SetCCOpcode) {
2832 default: llvm_unreachable("Illegal integer comparison"); break;
2833 case ISD::SETNE: Invert = true;
2834 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2835 case ISD::SETLT: Swap = true;
2836 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2837 case ISD::SETLE: Swap = true;
2838 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2839 case ISD::SETULT: Swap = true;
2840 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2841 case ISD::SETULE: Swap = true;
2842 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2845 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2846 if (Opc == ARMISD::VCEQ) {
2849 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2851 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2854 // Ignore bitconvert.
2855 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2856 AndOp = AndOp.getOperand(0);
2858 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2860 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2861 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2868 std::swap(Op0, Op1);
2870 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2873 Result = DAG.getNOT(dl, Result, VT);
2878 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2879 /// valid vector constant for a NEON instruction with a "modified immediate"
2880 /// operand (e.g., VMOV). If so, return the encoded value.
2881 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2882 unsigned SplatBitSize, SelectionDAG &DAG,
2883 EVT &VT, bool is128Bits, bool isVMOV) {
2884 unsigned OpCmode, Imm;
2886 // SplatBitSize is set to the smallest size that splats the vector, so a
2887 // zero vector will always have SplatBitSize == 8. However, NEON modified
2888 // immediate instructions others than VMOV do not support the 8-bit encoding
2889 // of a zero vector, and the default encoding of zero is supposed to be the
2894 switch (SplatBitSize) {
2898 // Any 1-byte value is OK. Op=0, Cmode=1110.
2899 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2902 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2906 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2907 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2908 if ((SplatBits & ~0xff) == 0) {
2909 // Value = 0x00nn: Op=x, Cmode=100x.
2914 if ((SplatBits & ~0xff00) == 0) {
2915 // Value = 0xnn00: Op=x, Cmode=101x.
2917 Imm = SplatBits >> 8;
2923 // NEON's 32-bit VMOV supports splat values where:
2924 // * only one byte is nonzero, or
2925 // * the least significant byte is 0xff and the second byte is nonzero, or
2926 // * the least significant 2 bytes are 0xff and the third is nonzero.
2927 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
2928 if ((SplatBits & ~0xff) == 0) {
2929 // Value = 0x000000nn: Op=x, Cmode=000x.
2934 if ((SplatBits & ~0xff00) == 0) {
2935 // Value = 0x0000nn00: Op=x, Cmode=001x.
2937 Imm = SplatBits >> 8;
2940 if ((SplatBits & ~0xff0000) == 0) {
2941 // Value = 0x00nn0000: Op=x, Cmode=010x.
2943 Imm = SplatBits >> 16;
2946 if ((SplatBits & ~0xff000000) == 0) {
2947 // Value = 0xnn000000: Op=x, Cmode=011x.
2949 Imm = SplatBits >> 24;
2953 if ((SplatBits & ~0xffff) == 0 &&
2954 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2955 // Value = 0x0000nnff: Op=x, Cmode=1100.
2957 Imm = SplatBits >> 8;
2962 if ((SplatBits & ~0xffffff) == 0 &&
2963 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2964 // Value = 0x00nnffff: Op=x, Cmode=1101.
2966 Imm = SplatBits >> 16;
2967 SplatBits |= 0xffff;
2971 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2972 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2973 // VMOV.I32. A (very) minor optimization would be to replicate the value
2974 // and fall through here to test for a valid 64-bit splat. But, then the
2975 // caller would also need to check and handle the change in size.
2981 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2982 uint64_t BitMask = 0xff;
2984 unsigned ImmMask = 1;
2986 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2987 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2990 } else if ((SplatBits & BitMask) != 0) {
2996 // Op=1, Cmode=1110.
2999 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3004 llvm_unreachable("unexpected size for isNEONModifiedImm");
3008 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3009 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3012 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3013 bool &ReverseVEXT, unsigned &Imm) {
3014 unsigned NumElts = VT.getVectorNumElements();
3015 ReverseVEXT = false;
3017 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3023 // If this is a VEXT shuffle, the immediate value is the index of the first
3024 // element. The other shuffle indices must be the successive elements after
3026 unsigned ExpectedElt = Imm;
3027 for (unsigned i = 1; i < NumElts; ++i) {
3028 // Increment the expected index. If it wraps around, it may still be
3029 // a VEXT but the source vectors must be swapped.
3031 if (ExpectedElt == NumElts * 2) {
3036 if (M[i] < 0) continue; // ignore UNDEF indices
3037 if (ExpectedElt != static_cast<unsigned>(M[i]))
3041 // Adjust the index value if the source operands will be swapped.
3048 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3049 /// instruction with the specified blocksize. (The order of the elements
3050 /// within each block of the vector is reversed.)
3051 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3052 unsigned BlockSize) {
3053 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3054 "Only possible block sizes for VREV are: 16, 32, 64");
3056 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3060 unsigned NumElts = VT.getVectorNumElements();
3061 unsigned BlockElts = M[0] + 1;
3062 // If the first shuffle index is UNDEF, be optimistic.
3064 BlockElts = BlockSize / EltSz;
3066 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3069 for (unsigned i = 0; i < NumElts; ++i) {
3070 if (M[i] < 0) continue; // ignore UNDEF indices
3071 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3078 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3079 unsigned &WhichResult) {
3080 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3084 unsigned NumElts = VT.getVectorNumElements();
3085 WhichResult = (M[0] == 0 ? 0 : 1);
3086 for (unsigned i = 0; i < NumElts; i += 2) {
3087 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3088 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3094 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3095 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3096 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3097 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3098 unsigned &WhichResult) {
3099 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3103 unsigned NumElts = VT.getVectorNumElements();
3104 WhichResult = (M[0] == 0 ? 0 : 1);
3105 for (unsigned i = 0; i < NumElts; i += 2) {
3106 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3107 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3113 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3114 unsigned &WhichResult) {
3115 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3119 unsigned NumElts = VT.getVectorNumElements();
3120 WhichResult = (M[0] == 0 ? 0 : 1);
3121 for (unsigned i = 0; i != NumElts; ++i) {
3122 if (M[i] < 0) continue; // ignore UNDEF indices
3123 if ((unsigned) M[i] != 2 * i + WhichResult)
3127 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3128 if (VT.is64BitVector() && EltSz == 32)
3134 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3135 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3136 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3137 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3138 unsigned &WhichResult) {
3139 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3143 unsigned Half = VT.getVectorNumElements() / 2;
3144 WhichResult = (M[0] == 0 ? 0 : 1);
3145 for (unsigned j = 0; j != 2; ++j) {
3146 unsigned Idx = WhichResult;
3147 for (unsigned i = 0; i != Half; ++i) {
3148 int MIdx = M[i + j * Half];
3149 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3155 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3156 if (VT.is64BitVector() && EltSz == 32)
3162 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3163 unsigned &WhichResult) {
3164 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3168 unsigned NumElts = VT.getVectorNumElements();
3169 WhichResult = (M[0] == 0 ? 0 : 1);
3170 unsigned Idx = WhichResult * NumElts / 2;
3171 for (unsigned i = 0; i != NumElts; i += 2) {
3172 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3173 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3178 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3179 if (VT.is64BitVector() && EltSz == 32)
3185 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3186 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3187 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3188 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3189 unsigned &WhichResult) {
3190 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3194 unsigned NumElts = VT.getVectorNumElements();
3195 WhichResult = (M[0] == 0 ? 0 : 1);
3196 unsigned Idx = WhichResult * NumElts / 2;
3197 for (unsigned i = 0; i != NumElts; i += 2) {
3198 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3199 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3204 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3205 if (VT.is64BitVector() && EltSz == 32)
3211 // If N is an integer constant that can be moved into a register in one
3212 // instruction, return an SDValue of such a constant (will become a MOV
3213 // instruction). Otherwise return null.
3214 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3215 const ARMSubtarget *ST, DebugLoc dl) {
3217 if (!isa<ConstantSDNode>(N))
3219 Val = cast<ConstantSDNode>(N)->getZExtValue();
3221 if (ST->isThumb1Only()) {
3222 if (Val <= 255 || ~Val <= 255)
3223 return DAG.getConstant(Val, MVT::i32);
3225 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3226 return DAG.getConstant(Val, MVT::i32);
3231 // If this is a case we can't handle, return null and let the default
3232 // expansion code take care of it.
3233 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3234 const ARMSubtarget *ST) {
3235 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3236 DebugLoc dl = Op.getDebugLoc();
3237 EVT VT = Op.getValueType();
3239 APInt SplatBits, SplatUndef;
3240 unsigned SplatBitSize;
3242 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3243 if (SplatBitSize <= 64) {
3244 // Check if an immediate VMOV works.
3246 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3247 SplatUndef.getZExtValue(), SplatBitSize,
3248 DAG, VmovVT, VT.is128BitVector(), true);
3249 if (Val.getNode()) {
3250 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3251 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3254 // Try an immediate VMVN.
3255 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3256 ((1LL << SplatBitSize) - 1));
3257 Val = isNEONModifiedImm(NegatedImm,
3258 SplatUndef.getZExtValue(), SplatBitSize,
3259 DAG, VmovVT, VT.is128BitVector(), false);
3260 if (Val.getNode()) {
3261 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3262 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3267 // Scan through the operands to see if only one value is used.
3268 unsigned NumElts = VT.getVectorNumElements();
3269 bool isOnlyLowElement = true;
3270 bool usesOnlyOneValue = true;
3271 bool isConstant = true;
3273 for (unsigned i = 0; i < NumElts; ++i) {
3274 SDValue V = Op.getOperand(i);
3275 if (V.getOpcode() == ISD::UNDEF)
3278 isOnlyLowElement = false;
3279 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3282 if (!Value.getNode())
3284 else if (V != Value)
3285 usesOnlyOneValue = false;
3288 if (!Value.getNode())
3289 return DAG.getUNDEF(VT);
3291 if (isOnlyLowElement)
3292 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3294 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3296 if (EnableARMVDUPsplat) {
3297 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3298 // i32 and try again.
3299 if (usesOnlyOneValue && EltSize <= 32) {
3301 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3302 if (VT.getVectorElementType().isFloatingPoint()) {
3303 SmallVector<SDValue, 8> Ops;
3304 for (unsigned i = 0; i < NumElts; ++i)
3305 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3307 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3309 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3310 LowerBUILD_VECTOR(Val, DAG, ST));
3312 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3314 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3318 // If all elements are constants and the case above didn't get hit, fall back
3319 // to the default expansion, which will generate a load from the constant
3324 if (!EnableARMVDUPsplat) {
3325 // Use VDUP for non-constant splats.
3326 if (usesOnlyOneValue && EltSize <= 32)
3327 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3330 // Vectors with 32- or 64-bit elements can be built by directly assigning
3331 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3332 // will be legalized.
3333 if (EltSize >= 32) {
3334 // Do the expansion with floating-point types, since that is what the VFP
3335 // registers are defined to use, and since i64 is not legal.
3336 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3337 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3338 SmallVector<SDValue, 8> Ops;
3339 for (unsigned i = 0; i < NumElts; ++i)
3340 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3341 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3342 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3348 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3349 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3350 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3351 /// are assumed to be legal.
3353 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3355 if (VT.getVectorNumElements() == 4 &&
3356 (VT.is128BitVector() || VT.is64BitVector())) {
3357 unsigned PFIndexes[4];
3358 for (unsigned i = 0; i != 4; ++i) {
3362 PFIndexes[i] = M[i];
3365 // Compute the index in the perfect shuffle table.
3366 unsigned PFTableIndex =
3367 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3368 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3369 unsigned Cost = (PFEntry >> 30);
3376 unsigned Imm, WhichResult;
3378 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3379 return (EltSize >= 32 ||
3380 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3381 isVREVMask(M, VT, 64) ||
3382 isVREVMask(M, VT, 32) ||
3383 isVREVMask(M, VT, 16) ||
3384 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3385 isVTRNMask(M, VT, WhichResult) ||
3386 isVUZPMask(M, VT, WhichResult) ||
3387 isVZIPMask(M, VT, WhichResult) ||
3388 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3389 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3390 isVZIP_v_undef_Mask(M, VT, WhichResult));
3393 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3394 /// the specified operations to build the shuffle.
3395 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3396 SDValue RHS, SelectionDAG &DAG,
3398 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3399 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3400 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3403 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3412 OP_VUZPL, // VUZP, left result
3413 OP_VUZPR, // VUZP, right result
3414 OP_VZIPL, // VZIP, left result
3415 OP_VZIPR, // VZIP, right result
3416 OP_VTRNL, // VTRN, left result
3417 OP_VTRNR // VTRN, right result
3420 if (OpNum == OP_COPY) {
3421 if (LHSID == (1*9+2)*9+3) return LHS;
3422 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3426 SDValue OpLHS, OpRHS;
3427 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3428 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3429 EVT VT = OpLHS.getValueType();
3432 default: llvm_unreachable("Unknown shuffle opcode!");
3434 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3439 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3440 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3444 return DAG.getNode(ARMISD::VEXT, dl, VT,
3446 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3449 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3450 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3453 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3454 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3457 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3458 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3462 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3463 SDValue V1 = Op.getOperand(0);
3464 SDValue V2 = Op.getOperand(1);
3465 DebugLoc dl = Op.getDebugLoc();
3466 EVT VT = Op.getValueType();
3467 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3468 SmallVector<int, 8> ShuffleMask;
3470 // Convert shuffles that are directly supported on NEON to target-specific
3471 // DAG nodes, instead of keeping them as shuffles and matching them again
3472 // during code selection. This is more efficient and avoids the possibility
3473 // of inconsistencies between legalization and selection.
3474 // FIXME: floating-point vectors should be canonicalized to integer vectors
3475 // of the same time so that they get CSEd properly.
3476 SVN->getMask(ShuffleMask);
3478 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3479 if (EltSize <= 32) {
3480 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3481 int Lane = SVN->getSplatIndex();
3482 // If this is undef splat, generate it via "just" vdup, if possible.
3483 if (Lane == -1) Lane = 0;
3485 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3486 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3488 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3489 DAG.getConstant(Lane, MVT::i32));
3494 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3497 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3498 DAG.getConstant(Imm, MVT::i32));
3501 if (isVREVMask(ShuffleMask, VT, 64))
3502 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3503 if (isVREVMask(ShuffleMask, VT, 32))
3504 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3505 if (isVREVMask(ShuffleMask, VT, 16))
3506 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3508 // Check for Neon shuffles that modify both input vectors in place.
3509 // If both results are used, i.e., if there are two shuffles with the same
3510 // source operands and with masks corresponding to both results of one of
3511 // these operations, DAG memoization will ensure that a single node is
3512 // used for both shuffles.
3513 unsigned WhichResult;
3514 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3515 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3516 V1, V2).getValue(WhichResult);
3517 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3518 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3519 V1, V2).getValue(WhichResult);
3520 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3521 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3522 V1, V2).getValue(WhichResult);
3524 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3525 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3526 V1, V1).getValue(WhichResult);
3527 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3528 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3529 V1, V1).getValue(WhichResult);
3530 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3531 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3532 V1, V1).getValue(WhichResult);
3535 // If the shuffle is not directly supported and it has 4 elements, use
3536 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3537 unsigned NumElts = VT.getVectorNumElements();
3539 unsigned PFIndexes[4];
3540 for (unsigned i = 0; i != 4; ++i) {
3541 if (ShuffleMask[i] < 0)
3544 PFIndexes[i] = ShuffleMask[i];
3547 // Compute the index in the perfect shuffle table.
3548 unsigned PFTableIndex =
3549 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3550 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3551 unsigned Cost = (PFEntry >> 30);
3554 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3557 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3558 if (EltSize >= 32) {
3559 // Do the expansion with floating-point types, since that is what the VFP
3560 // registers are defined to use, and since i64 is not legal.
3561 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3562 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3563 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3564 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3565 SmallVector<SDValue, 8> Ops;
3566 for (unsigned i = 0; i < NumElts; ++i) {
3567 if (ShuffleMask[i] < 0)
3568 Ops.push_back(DAG.getUNDEF(EltVT));
3570 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3571 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3572 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3575 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3576 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3582 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3583 EVT VT = Op.getValueType();
3584 DebugLoc dl = Op.getDebugLoc();
3585 SDValue Vec = Op.getOperand(0);
3586 SDValue Lane = Op.getOperand(1);
3587 assert(VT == MVT::i32 &&
3588 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3589 "unexpected type for custom-lowering vector extract");
3590 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3593 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3594 // The only time a CONCAT_VECTORS operation can have legal types is when
3595 // two 64-bit vectors are concatenated to a 128-bit vector.
3596 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3597 "unexpected CONCAT_VECTORS");
3598 DebugLoc dl = Op.getDebugLoc();
3599 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3600 SDValue Op0 = Op.getOperand(0);
3601 SDValue Op1 = Op.getOperand(1);
3602 if (Op0.getOpcode() != ISD::UNDEF)
3603 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3604 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3605 DAG.getIntPtrConstant(0));
3606 if (Op1.getOpcode() != ISD::UNDEF)
3607 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3608 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3609 DAG.getIntPtrConstant(1));
3610 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3613 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3614 /// an extending load, return the unextended value.
3615 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3616 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3617 return N->getOperand(0);
3618 LoadSDNode *LD = cast<LoadSDNode>(N);
3619 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3620 LD->getBasePtr(), LD->getSrcValue(),
3621 LD->getSrcValueOffset(), LD->isVolatile(),
3622 LD->isNonTemporal(), LD->getAlignment());
3625 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3626 // Multiplications are only custom-lowered for 128-bit vectors so that
3627 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3628 EVT VT = Op.getValueType();
3629 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3630 SDNode *N0 = Op.getOperand(0).getNode();
3631 SDNode *N1 = Op.getOperand(1).getNode();
3632 unsigned NewOpc = 0;
3633 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3634 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3635 NewOpc = ARMISD::VMULLs;
3636 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3637 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3638 NewOpc = ARMISD::VMULLu;
3639 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3640 // Fall through to expand this. It is not legal.
3643 // Other vector multiplications are legal.
3647 // Legalize to a VMULL instruction.
3648 DebugLoc DL = Op.getDebugLoc();
3649 SDValue Op0 = SkipExtension(N0, DAG);
3650 SDValue Op1 = SkipExtension(N1, DAG);
3652 assert(Op0.getValueType().is64BitVector() &&
3653 Op1.getValueType().is64BitVector() &&
3654 "unexpected types for extended operands to VMULL");
3655 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3658 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3659 switch (Op.getOpcode()) {
3660 default: llvm_unreachable("Don't know how to custom lower this!");
3661 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3662 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3663 case ISD::GlobalAddress:
3664 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3665 LowerGlobalAddressELF(Op, DAG);
3666 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3667 case ISD::SELECT: return LowerSELECT(Op, DAG);
3668 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3669 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3670 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3671 case ISD::VASTART: return LowerVASTART(Op, DAG);
3672 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3673 case ISD::SINT_TO_FP:
3674 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3675 case ISD::FP_TO_SINT:
3676 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3677 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3678 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3679 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3680 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3681 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3682 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3683 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3685 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3688 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3689 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3690 case ISD::SRL_PARTS:
3691 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3692 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3693 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3694 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3695 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3696 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3697 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3698 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3699 case ISD::MUL: return LowerMUL(Op, DAG);
3704 /// ReplaceNodeResults - Replace the results of node with an illegal result
3705 /// type with new values built out of custom code.
3706 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3707 SmallVectorImpl<SDValue>&Results,
3708 SelectionDAG &DAG) const {
3710 switch (N->getOpcode()) {
3712 llvm_unreachable("Don't know how to custom expand this!");
3714 case ISD::BIT_CONVERT:
3715 Res = ExpandBIT_CONVERT(N, DAG);
3719 Res = LowerShift(N, DAG, Subtarget);
3723 Results.push_back(Res);
3726 //===----------------------------------------------------------------------===//
3727 // ARM Scheduler Hooks
3728 //===----------------------------------------------------------------------===//
3731 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3732 MachineBasicBlock *BB,
3733 unsigned Size) const {
3734 unsigned dest = MI->getOperand(0).getReg();
3735 unsigned ptr = MI->getOperand(1).getReg();
3736 unsigned oldval = MI->getOperand(2).getReg();
3737 unsigned newval = MI->getOperand(3).getReg();
3738 unsigned scratch = BB->getParent()->getRegInfo()
3739 .createVirtualRegister(ARM::GPRRegisterClass);
3740 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3741 DebugLoc dl = MI->getDebugLoc();
3742 bool isThumb2 = Subtarget->isThumb2();
3744 unsigned ldrOpc, strOpc;
3746 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3748 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3749 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3752 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3753 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3756 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3757 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3761 MachineFunction *MF = BB->getParent();
3762 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3763 MachineFunction::iterator It = BB;
3764 ++It; // insert the new blocks after the current block
3766 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3767 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3768 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3769 MF->insert(It, loop1MBB);
3770 MF->insert(It, loop2MBB);
3771 MF->insert(It, exitMBB);
3773 // Transfer the remainder of BB and its successor edges to exitMBB.
3774 exitMBB->splice(exitMBB->begin(), BB,
3775 llvm::next(MachineBasicBlock::iterator(MI)),
3777 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3781 // fallthrough --> loop1MBB
3782 BB->addSuccessor(loop1MBB);
3785 // ldrex dest, [ptr]
3789 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3790 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3791 .addReg(dest).addReg(oldval));
3792 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3793 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3794 BB->addSuccessor(loop2MBB);
3795 BB->addSuccessor(exitMBB);
3798 // strex scratch, newval, [ptr]
3802 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3804 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3805 .addReg(scratch).addImm(0));
3806 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3807 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3808 BB->addSuccessor(loop1MBB);
3809 BB->addSuccessor(exitMBB);
3815 MI->eraseFromParent(); // The instruction is gone now.
3821 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3822 unsigned Size, unsigned BinOpcode) const {
3823 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3824 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3826 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3827 MachineFunction *MF = BB->getParent();
3828 MachineFunction::iterator It = BB;
3831 unsigned dest = MI->getOperand(0).getReg();
3832 unsigned ptr = MI->getOperand(1).getReg();
3833 unsigned incr = MI->getOperand(2).getReg();
3834 DebugLoc dl = MI->getDebugLoc();
3836 bool isThumb2 = Subtarget->isThumb2();
3837 unsigned ldrOpc, strOpc;
3839 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3841 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3842 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3845 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3846 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3849 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3850 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3854 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3855 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3856 MF->insert(It, loopMBB);
3857 MF->insert(It, exitMBB);
3859 // Transfer the remainder of BB and its successor edges to exitMBB.
3860 exitMBB->splice(exitMBB->begin(), BB,
3861 llvm::next(MachineBasicBlock::iterator(MI)),
3863 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3865 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3866 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3867 unsigned scratch2 = (!BinOpcode) ? incr :
3868 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3872 // fallthrough --> loopMBB
3873 BB->addSuccessor(loopMBB);
3877 // <binop> scratch2, dest, incr
3878 // strex scratch, scratch2, ptr
3881 // fallthrough --> exitMBB
3883 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3885 // operand order needs to go the other way for NAND
3886 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3887 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3888 addReg(incr).addReg(dest)).addReg(0);
3890 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3891 addReg(dest).addReg(incr)).addReg(0);
3894 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3896 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3897 .addReg(scratch).addImm(0));
3898 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3899 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3901 BB->addSuccessor(loopMBB);
3902 BB->addSuccessor(exitMBB);
3908 MI->eraseFromParent(); // The instruction is gone now.
3914 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3915 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3916 E = MBB->succ_end(); I != E; ++I)
3919 llvm_unreachable("Expecting a BB with two successors!");
3923 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3924 MachineBasicBlock *BB) const {
3925 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3926 DebugLoc dl = MI->getDebugLoc();
3927 bool isThumb2 = Subtarget->isThumb2();
3928 switch (MI->getOpcode()) {
3931 llvm_unreachable("Unexpected instr type to insert");
3933 case ARM::ATOMIC_LOAD_ADD_I8:
3934 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3935 case ARM::ATOMIC_LOAD_ADD_I16:
3936 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3937 case ARM::ATOMIC_LOAD_ADD_I32:
3938 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3940 case ARM::ATOMIC_LOAD_AND_I8:
3941 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3942 case ARM::ATOMIC_LOAD_AND_I16:
3943 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3944 case ARM::ATOMIC_LOAD_AND_I32:
3945 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3947 case ARM::ATOMIC_LOAD_OR_I8:
3948 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3949 case ARM::ATOMIC_LOAD_OR_I16:
3950 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3951 case ARM::ATOMIC_LOAD_OR_I32:
3952 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3954 case ARM::ATOMIC_LOAD_XOR_I8:
3955 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3956 case ARM::ATOMIC_LOAD_XOR_I16:
3957 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3958 case ARM::ATOMIC_LOAD_XOR_I32:
3959 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3961 case ARM::ATOMIC_LOAD_NAND_I8:
3962 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3963 case ARM::ATOMIC_LOAD_NAND_I16:
3964 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3965 case ARM::ATOMIC_LOAD_NAND_I32:
3966 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3968 case ARM::ATOMIC_LOAD_SUB_I8:
3969 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3970 case ARM::ATOMIC_LOAD_SUB_I16:
3971 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3972 case ARM::ATOMIC_LOAD_SUB_I32:
3973 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3975 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3976 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3977 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3979 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3980 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3981 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3983 case ARM::tMOVCCr_pseudo: {
3984 // To "insert" a SELECT_CC instruction, we actually have to insert the
3985 // diamond control-flow pattern. The incoming instruction knows the
3986 // destination vreg to set, the condition code register to branch on, the
3987 // true/false values to select between, and a branch opcode to use.
3988 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3989 MachineFunction::iterator It = BB;
3995 // cmpTY ccX, r1, r2
3997 // fallthrough --> copy0MBB
3998 MachineBasicBlock *thisMBB = BB;
3999 MachineFunction *F = BB->getParent();
4000 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4001 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4002 F->insert(It, copy0MBB);
4003 F->insert(It, sinkMBB);
4005 // Transfer the remainder of BB and its successor edges to sinkMBB.
4006 sinkMBB->splice(sinkMBB->begin(), BB,
4007 llvm::next(MachineBasicBlock::iterator(MI)),
4009 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4011 BB->addSuccessor(copy0MBB);
4012 BB->addSuccessor(sinkMBB);
4014 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4015 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4018 // %FalseValue = ...
4019 // # fallthrough to sinkMBB
4022 // Update machine-CFG edges
4023 BB->addSuccessor(sinkMBB);
4026 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4029 BuildMI(*BB, BB->begin(), dl,
4030 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4031 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4032 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4034 MI->eraseFromParent(); // The pseudo instruction is gone now.
4039 case ARM::BCCZi64: {
4040 // Compare both parts that make up the double comparison separately for
4042 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4044 unsigned LHS1 = MI->getOperand(1).getReg();
4045 unsigned LHS2 = MI->getOperand(2).getReg();
4047 AddDefaultPred(BuildMI(BB, dl,
4048 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4049 .addReg(LHS1).addImm(0));
4050 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4051 .addReg(LHS2).addImm(0)
4052 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4054 unsigned RHS1 = MI->getOperand(3).getReg();
4055 unsigned RHS2 = MI->getOperand(4).getReg();
4056 AddDefaultPred(BuildMI(BB, dl,
4057 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4058 .addReg(LHS1).addReg(RHS1));
4059 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4060 .addReg(LHS2).addReg(RHS2)
4061 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4064 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4065 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4066 if (MI->getOperand(0).getImm() == ARMCC::NE)
4067 std::swap(destMBB, exitMBB);
4069 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4070 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4071 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4074 MI->eraseFromParent(); // The pseudo instruction is gone now.
4080 //===----------------------------------------------------------------------===//
4081 // ARM Optimization Hooks
4082 //===----------------------------------------------------------------------===//
4085 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4086 TargetLowering::DAGCombinerInfo &DCI) {
4087 SelectionDAG &DAG = DCI.DAG;
4088 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4089 EVT VT = N->getValueType(0);
4090 unsigned Opc = N->getOpcode();
4091 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4092 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4093 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4094 ISD::CondCode CC = ISD::SETCC_INVALID;
4097 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4099 SDValue CCOp = Slct.getOperand(0);
4100 if (CCOp.getOpcode() == ISD::SETCC)
4101 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4104 bool DoXform = false;
4106 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4109 if (LHS.getOpcode() == ISD::Constant &&
4110 cast<ConstantSDNode>(LHS)->isNullValue()) {
4112 } else if (CC != ISD::SETCC_INVALID &&
4113 RHS.getOpcode() == ISD::Constant &&
4114 cast<ConstantSDNode>(RHS)->isNullValue()) {
4115 std::swap(LHS, RHS);
4116 SDValue Op0 = Slct.getOperand(0);
4117 EVT OpVT = isSlctCC ? Op0.getValueType() :
4118 Op0.getOperand(0).getValueType();
4119 bool isInt = OpVT.isInteger();
4120 CC = ISD::getSetCCInverse(CC, isInt);
4122 if (!TLI.isCondCodeLegal(CC, OpVT))
4123 return SDValue(); // Inverse operator isn't legal.
4130 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4132 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4133 Slct.getOperand(0), Slct.getOperand(1), CC);
4134 SDValue CCOp = Slct.getOperand(0);
4136 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4137 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4138 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4139 CCOp, OtherOp, Result);
4144 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4145 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4146 /// called with the default operands, and if that fails, with commuted
4148 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4149 TargetLowering::DAGCombinerInfo &DCI) {
4150 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4151 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4152 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4153 if (Result.getNode()) return Result;
4158 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4160 static SDValue PerformADDCombine(SDNode *N,
4161 TargetLowering::DAGCombinerInfo &DCI) {
4162 SDValue N0 = N->getOperand(0);
4163 SDValue N1 = N->getOperand(1);
4165 // First try with the default operand order.
4166 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4167 if (Result.getNode())
4170 // If that didn't work, try again with the operands commuted.
4171 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4174 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4176 static SDValue PerformSUBCombine(SDNode *N,
4177 TargetLowering::DAGCombinerInfo &DCI) {
4178 SDValue N0 = N->getOperand(0);
4179 SDValue N1 = N->getOperand(1);
4181 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4182 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4183 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4184 if (Result.getNode()) return Result;
4190 static SDValue PerformMULCombine(SDNode *N,
4191 TargetLowering::DAGCombinerInfo &DCI,
4192 const ARMSubtarget *Subtarget) {
4193 SelectionDAG &DAG = DCI.DAG;
4195 if (Subtarget->isThumb1Only())
4198 if (DAG.getMachineFunction().
4199 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4202 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4205 EVT VT = N->getValueType(0);
4209 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4213 uint64_t MulAmt = C->getZExtValue();
4214 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4215 ShiftAmt = ShiftAmt & (32 - 1);
4216 SDValue V = N->getOperand(0);
4217 DebugLoc DL = N->getDebugLoc();
4220 MulAmt >>= ShiftAmt;
4221 if (isPowerOf2_32(MulAmt - 1)) {
4222 // (mul x, 2^N + 1) => (add (shl x, N), x)
4223 Res = DAG.getNode(ISD::ADD, DL, VT,
4224 V, DAG.getNode(ISD::SHL, DL, VT,
4225 V, DAG.getConstant(Log2_32(MulAmt-1),
4227 } else if (isPowerOf2_32(MulAmt + 1)) {
4228 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4229 Res = DAG.getNode(ISD::SUB, DL, VT,
4230 DAG.getNode(ISD::SHL, DL, VT,
4231 V, DAG.getConstant(Log2_32(MulAmt+1),
4238 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4239 DAG.getConstant(ShiftAmt, MVT::i32));
4241 // Do not add new nodes to DAG combiner worklist.
4242 DCI.CombineTo(N, Res, false);
4246 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4247 static SDValue PerformORCombine(SDNode *N,
4248 TargetLowering::DAGCombinerInfo &DCI,
4249 const ARMSubtarget *Subtarget) {
4250 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4253 // BFI is only available on V6T2+
4254 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4257 SelectionDAG &DAG = DCI.DAG;
4258 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4259 DebugLoc DL = N->getDebugLoc();
4260 // 1) or (and A, mask), val => ARMbfi A, val, mask
4261 // iff (val & mask) == val
4263 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4264 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4265 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4266 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4267 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4268 // (i.e., copy a bitfield value into another bitfield of the same width)
4269 if (N0.getOpcode() != ISD::AND)
4272 EVT VT = N->getValueType(0);
4277 // The value and the mask need to be constants so we can verify this is
4278 // actually a bitfield set. If the mask is 0xffff, we can do better
4279 // via a movt instruction, so don't use BFI in that case.
4280 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4283 unsigned Mask = C->getZExtValue();
4287 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4288 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4289 unsigned Val = C->getZExtValue();
4290 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4292 Val >>= CountTrailingZeros_32(~Mask);
4294 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4295 DAG.getConstant(Val, MVT::i32),
4296 DAG.getConstant(Mask, MVT::i32));
4298 // Do not add new nodes to DAG combiner worklist.
4299 DCI.CombineTo(N, Res, false);
4300 } else if (N1.getOpcode() == ISD::AND) {
4301 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4302 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4305 unsigned Mask2 = C->getZExtValue();
4307 if (ARM::isBitFieldInvertedMask(Mask) &&
4308 ARM::isBitFieldInvertedMask(~Mask2) &&
4309 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4310 // The pack halfword instruction works better for masks that fit it,
4311 // so use that when it's available.
4312 if (Subtarget->hasT2ExtractPack() &&
4313 (Mask == 0xffff || Mask == 0xffff0000))
4316 unsigned lsb = CountTrailingZeros_32(Mask2);
4317 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4318 DAG.getConstant(lsb, MVT::i32));
4319 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4320 DAG.getConstant(Mask, MVT::i32));
4321 // Do not add new nodes to DAG combiner worklist.
4322 DCI.CombineTo(N, Res, false);
4323 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4324 ARM::isBitFieldInvertedMask(Mask2) &&
4325 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4326 // The pack halfword instruction works better for masks that fit it,
4327 // so use that when it's available.
4328 if (Subtarget->hasT2ExtractPack() &&
4329 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4332 unsigned lsb = CountTrailingZeros_32(Mask);
4333 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4334 DAG.getConstant(lsb, MVT::i32));
4335 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4336 DAG.getConstant(Mask2, MVT::i32));
4337 // Do not add new nodes to DAG combiner worklist.
4338 DCI.CombineTo(N, Res, false);
4345 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4346 /// ARMISD::VMOVRRD.
4347 static SDValue PerformVMOVRRDCombine(SDNode *N,
4348 TargetLowering::DAGCombinerInfo &DCI) {
4349 // fmrrd(fmdrr x, y) -> x,y
4350 SDValue InDouble = N->getOperand(0);
4351 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4352 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4356 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4357 /// ARMISD::VDUPLANE.
4358 static SDValue PerformVDUPLANECombine(SDNode *N,
4359 TargetLowering::DAGCombinerInfo &DCI) {
4360 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4362 SDValue Op = N->getOperand(0);
4363 EVT VT = N->getValueType(0);
4365 // Ignore bit_converts.
4366 while (Op.getOpcode() == ISD::BIT_CONVERT)
4367 Op = Op.getOperand(0);
4368 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4371 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4372 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4373 // The canonical VMOV for a zero vector uses a 32-bit element size.
4374 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4376 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4378 if (EltSize > VT.getVectorElementType().getSizeInBits())
4381 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4382 return DCI.CombineTo(N, Res, false);
4385 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4386 /// operand of a vector shift operation, where all the elements of the
4387 /// build_vector must have the same constant integer value.
4388 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4389 // Ignore bit_converts.
4390 while (Op.getOpcode() == ISD::BIT_CONVERT)
4391 Op = Op.getOperand(0);
4392 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4393 APInt SplatBits, SplatUndef;
4394 unsigned SplatBitSize;
4396 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4397 HasAnyUndefs, ElementBits) ||
4398 SplatBitSize > ElementBits)
4400 Cnt = SplatBits.getSExtValue();
4404 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4405 /// operand of a vector shift left operation. That value must be in the range:
4406 /// 0 <= Value < ElementBits for a left shift; or
4407 /// 0 <= Value <= ElementBits for a long left shift.
4408 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4409 assert(VT.isVector() && "vector shift count is not a vector type");
4410 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4411 if (! getVShiftImm(Op, ElementBits, Cnt))
4413 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4416 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4417 /// operand of a vector shift right operation. For a shift opcode, the value
4418 /// is positive, but for an intrinsic the value count must be negative. The
4419 /// absolute value must be in the range:
4420 /// 1 <= |Value| <= ElementBits for a right shift; or
4421 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4422 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4424 assert(VT.isVector() && "vector shift count is not a vector type");
4425 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4426 if (! getVShiftImm(Op, ElementBits, Cnt))
4430 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4433 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4434 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4435 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4438 // Don't do anything for most intrinsics.
4441 // Vector shifts: check for immediate versions and lower them.
4442 // Note: This is done during DAG combining instead of DAG legalizing because
4443 // the build_vectors for 64-bit vector element shift counts are generally
4444 // not legal, and it is hard to see their values after they get legalized to
4445 // loads from a constant pool.
4446 case Intrinsic::arm_neon_vshifts:
4447 case Intrinsic::arm_neon_vshiftu:
4448 case Intrinsic::arm_neon_vshiftls:
4449 case Intrinsic::arm_neon_vshiftlu:
4450 case Intrinsic::arm_neon_vshiftn:
4451 case Intrinsic::arm_neon_vrshifts:
4452 case Intrinsic::arm_neon_vrshiftu:
4453 case Intrinsic::arm_neon_vrshiftn:
4454 case Intrinsic::arm_neon_vqshifts:
4455 case Intrinsic::arm_neon_vqshiftu:
4456 case Intrinsic::arm_neon_vqshiftsu:
4457 case Intrinsic::arm_neon_vqshiftns:
4458 case Intrinsic::arm_neon_vqshiftnu:
4459 case Intrinsic::arm_neon_vqshiftnsu:
4460 case Intrinsic::arm_neon_vqrshiftns:
4461 case Intrinsic::arm_neon_vqrshiftnu:
4462 case Intrinsic::arm_neon_vqrshiftnsu: {
4463 EVT VT = N->getOperand(1).getValueType();
4465 unsigned VShiftOpc = 0;
4468 case Intrinsic::arm_neon_vshifts:
4469 case Intrinsic::arm_neon_vshiftu:
4470 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4471 VShiftOpc = ARMISD::VSHL;
4474 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4475 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4476 ARMISD::VSHRs : ARMISD::VSHRu);
4481 case Intrinsic::arm_neon_vshiftls:
4482 case Intrinsic::arm_neon_vshiftlu:
4483 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4485 llvm_unreachable("invalid shift count for vshll intrinsic");
4487 case Intrinsic::arm_neon_vrshifts:
4488 case Intrinsic::arm_neon_vrshiftu:
4489 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4493 case Intrinsic::arm_neon_vqshifts:
4494 case Intrinsic::arm_neon_vqshiftu:
4495 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4499 case Intrinsic::arm_neon_vqshiftsu:
4500 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4502 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4504 case Intrinsic::arm_neon_vshiftn:
4505 case Intrinsic::arm_neon_vrshiftn:
4506 case Intrinsic::arm_neon_vqshiftns:
4507 case Intrinsic::arm_neon_vqshiftnu:
4508 case Intrinsic::arm_neon_vqshiftnsu:
4509 case Intrinsic::arm_neon_vqrshiftns:
4510 case Intrinsic::arm_neon_vqrshiftnu:
4511 case Intrinsic::arm_neon_vqrshiftnsu:
4512 // Narrowing shifts require an immediate right shift.
4513 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4515 llvm_unreachable("invalid shift count for narrowing vector shift "
4519 llvm_unreachable("unhandled vector shift");
4523 case Intrinsic::arm_neon_vshifts:
4524 case Intrinsic::arm_neon_vshiftu:
4525 // Opcode already set above.
4527 case Intrinsic::arm_neon_vshiftls:
4528 case Intrinsic::arm_neon_vshiftlu:
4529 if (Cnt == VT.getVectorElementType().getSizeInBits())
4530 VShiftOpc = ARMISD::VSHLLi;
4532 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4533 ARMISD::VSHLLs : ARMISD::VSHLLu);
4535 case Intrinsic::arm_neon_vshiftn:
4536 VShiftOpc = ARMISD::VSHRN; break;
4537 case Intrinsic::arm_neon_vrshifts:
4538 VShiftOpc = ARMISD::VRSHRs; break;
4539 case Intrinsic::arm_neon_vrshiftu:
4540 VShiftOpc = ARMISD::VRSHRu; break;
4541 case Intrinsic::arm_neon_vrshiftn:
4542 VShiftOpc = ARMISD::VRSHRN; break;
4543 case Intrinsic::arm_neon_vqshifts:
4544 VShiftOpc = ARMISD::VQSHLs; break;
4545 case Intrinsic::arm_neon_vqshiftu:
4546 VShiftOpc = ARMISD::VQSHLu; break;
4547 case Intrinsic::arm_neon_vqshiftsu:
4548 VShiftOpc = ARMISD::VQSHLsu; break;
4549 case Intrinsic::arm_neon_vqshiftns:
4550 VShiftOpc = ARMISD::VQSHRNs; break;
4551 case Intrinsic::arm_neon_vqshiftnu:
4552 VShiftOpc = ARMISD::VQSHRNu; break;
4553 case Intrinsic::arm_neon_vqshiftnsu:
4554 VShiftOpc = ARMISD::VQSHRNsu; break;
4555 case Intrinsic::arm_neon_vqrshiftns:
4556 VShiftOpc = ARMISD::VQRSHRNs; break;
4557 case Intrinsic::arm_neon_vqrshiftnu:
4558 VShiftOpc = ARMISD::VQRSHRNu; break;
4559 case Intrinsic::arm_neon_vqrshiftnsu:
4560 VShiftOpc = ARMISD::VQRSHRNsu; break;
4563 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4564 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4567 case Intrinsic::arm_neon_vshiftins: {
4568 EVT VT = N->getOperand(1).getValueType();
4570 unsigned VShiftOpc = 0;
4572 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4573 VShiftOpc = ARMISD::VSLI;
4574 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4575 VShiftOpc = ARMISD::VSRI;
4577 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4580 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4581 N->getOperand(1), N->getOperand(2),
4582 DAG.getConstant(Cnt, MVT::i32));
4585 case Intrinsic::arm_neon_vqrshifts:
4586 case Intrinsic::arm_neon_vqrshiftu:
4587 // No immediate versions of these to check for.
4594 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4595 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4596 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4597 /// vector element shift counts are generally not legal, and it is hard to see
4598 /// their values after they get legalized to loads from a constant pool.
4599 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4600 const ARMSubtarget *ST) {
4601 EVT VT = N->getValueType(0);
4603 // Nothing to be done for scalar shifts.
4604 if (! VT.isVector())
4607 assert(ST->hasNEON() && "unexpected vector shift");
4610 switch (N->getOpcode()) {
4611 default: llvm_unreachable("unexpected shift opcode");
4614 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4615 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4616 DAG.getConstant(Cnt, MVT::i32));
4621 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4622 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4623 ARMISD::VSHRs : ARMISD::VSHRu);
4624 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4625 DAG.getConstant(Cnt, MVT::i32));
4631 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4632 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4633 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4634 const ARMSubtarget *ST) {
4635 SDValue N0 = N->getOperand(0);
4637 // Check for sign- and zero-extensions of vector extract operations of 8-
4638 // and 16-bit vector elements. NEON supports these directly. They are
4639 // handled during DAG combining because type legalization will promote them
4640 // to 32-bit types and it is messy to recognize the operations after that.
4641 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4642 SDValue Vec = N0.getOperand(0);
4643 SDValue Lane = N0.getOperand(1);
4644 EVT VT = N->getValueType(0);
4645 EVT EltVT = N0.getValueType();
4646 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4648 if (VT == MVT::i32 &&
4649 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4650 TLI.isTypeLegal(Vec.getValueType())) {
4653 switch (N->getOpcode()) {
4654 default: llvm_unreachable("unexpected opcode");
4655 case ISD::SIGN_EXTEND:
4656 Opc = ARMISD::VGETLANEs;
4658 case ISD::ZERO_EXTEND:
4659 case ISD::ANY_EXTEND:
4660 Opc = ARMISD::VGETLANEu;
4663 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4670 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4671 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4672 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4673 const ARMSubtarget *ST) {
4674 // If the target supports NEON, try to use vmax/vmin instructions for f32
4675 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4676 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4677 // a NaN; only do the transformation when it matches that behavior.
4679 // For now only do this when using NEON for FP operations; if using VFP, it
4680 // is not obvious that the benefit outweighs the cost of switching to the
4682 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4683 N->getValueType(0) != MVT::f32)
4686 SDValue CondLHS = N->getOperand(0);
4687 SDValue CondRHS = N->getOperand(1);
4688 SDValue LHS = N->getOperand(2);
4689 SDValue RHS = N->getOperand(3);
4690 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4692 unsigned Opcode = 0;
4694 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4695 IsReversed = false; // x CC y ? x : y
4696 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4697 IsReversed = true ; // x CC y ? y : x
4711 // If LHS is NaN, an ordered comparison will be false and the result will
4712 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4713 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4714 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4715 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4717 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4718 // will return -0, so vmin can only be used for unsafe math or if one of
4719 // the operands is known to be nonzero.
4720 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4722 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4724 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4733 // If LHS is NaN, an ordered comparison will be false and the result will
4734 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4735 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4736 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4737 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4739 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4740 // will return +0, so vmax can only be used for unsafe math or if one of
4741 // the operands is known to be nonzero.
4742 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4744 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4746 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4752 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4755 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4756 DAGCombinerInfo &DCI) const {
4757 switch (N->getOpcode()) {
4759 case ISD::ADD: return PerformADDCombine(N, DCI);
4760 case ISD::SUB: return PerformSUBCombine(N, DCI);
4761 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4762 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4763 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4764 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4765 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4768 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4769 case ISD::SIGN_EXTEND:
4770 case ISD::ZERO_EXTEND:
4771 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4772 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4777 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4778 if (!Subtarget->hasV6Ops())
4779 // Pre-v6 does not support unaligned mem access.
4782 // v6+ may or may not support unaligned mem access depending on the system
4784 // FIXME: This is pretty conservative. Should we provide cmdline option to
4785 // control the behaviour?
4786 if (!Subtarget->isTargetDarwin())
4789 switch (VT.getSimpleVT().SimpleTy) {
4796 // FIXME: VLD1 etc with standard alignment is legal.
4800 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4805 switch (VT.getSimpleVT().SimpleTy) {
4806 default: return false;
4821 if ((V & (Scale - 1)) != 0)
4824 return V == (V & ((1LL << 5) - 1));
4827 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4828 const ARMSubtarget *Subtarget) {
4835 switch (VT.getSimpleVT().SimpleTy) {
4836 default: return false;
4841 // + imm12 or - imm8
4843 return V == (V & ((1LL << 8) - 1));
4844 return V == (V & ((1LL << 12) - 1));
4847 // Same as ARM mode. FIXME: NEON?
4848 if (!Subtarget->hasVFP2())
4853 return V == (V & ((1LL << 8) - 1));
4857 /// isLegalAddressImmediate - Return true if the integer value can be used
4858 /// as the offset of the target addressing mode for load / store of the
4860 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4861 const ARMSubtarget *Subtarget) {
4868 if (Subtarget->isThumb1Only())
4869 return isLegalT1AddressImmediate(V, VT);
4870 else if (Subtarget->isThumb2())
4871 return isLegalT2AddressImmediate(V, VT, Subtarget);
4876 switch (VT.getSimpleVT().SimpleTy) {
4877 default: return false;
4882 return V == (V & ((1LL << 12) - 1));
4885 return V == (V & ((1LL << 8) - 1));
4888 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4893 return V == (V & ((1LL << 8) - 1));
4897 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4899 int Scale = AM.Scale;
4903 switch (VT.getSimpleVT().SimpleTy) {
4904 default: return false;
4913 return Scale == 2 || Scale == 4 || Scale == 8;
4916 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4920 // Note, we allow "void" uses (basically, uses that aren't loads or
4921 // stores), because arm allows folding a scale into many arithmetic
4922 // operations. This should be made more precise and revisited later.
4924 // Allow r << imm, but the imm has to be a multiple of two.
4925 if (Scale & 1) return false;
4926 return isPowerOf2_32(Scale);
4930 /// isLegalAddressingMode - Return true if the addressing mode represented
4931 /// by AM is legal for this target, for a load/store of the specified type.
4932 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4933 const Type *Ty) const {
4934 EVT VT = getValueType(Ty, true);
4935 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4938 // Can never fold addr of global into load/store.
4943 case 0: // no scale reg, must be "r+i" or "r", or "i".
4946 if (Subtarget->isThumb1Only())
4950 // ARM doesn't support any R+R*scale+imm addr modes.
4957 if (Subtarget->isThumb2())
4958 return isLegalT2ScaledAddressingMode(AM, VT);
4960 int Scale = AM.Scale;
4961 switch (VT.getSimpleVT().SimpleTy) {
4962 default: return false;
4966 if (Scale < 0) Scale = -Scale;
4970 return isPowerOf2_32(Scale & ~1);
4974 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4979 // Note, we allow "void" uses (basically, uses that aren't loads or
4980 // stores), because arm allows folding a scale into many arithmetic
4981 // operations. This should be made more precise and revisited later.
4983 // Allow r << imm, but the imm has to be a multiple of two.
4984 if (Scale & 1) return false;
4985 return isPowerOf2_32(Scale);
4992 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4993 /// icmp immediate, that is the target has icmp instructions which can compare
4994 /// a register against the immediate without having to materialize the
4995 /// immediate into a register.
4996 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4997 if (!Subtarget->isThumb())
4998 return ARM_AM::getSOImmVal(Imm) != -1;
4999 if (Subtarget->isThumb2())
5000 return ARM_AM::getT2SOImmVal(Imm) != -1;
5001 return Imm >= 0 && Imm <= 255;
5004 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5005 bool isSEXTLoad, SDValue &Base,
5006 SDValue &Offset, bool &isInc,
5007 SelectionDAG &DAG) {
5008 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5011 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5013 Base = Ptr->getOperand(0);
5014 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5015 int RHSC = (int)RHS->getZExtValue();
5016 if (RHSC < 0 && RHSC > -256) {
5017 assert(Ptr->getOpcode() == ISD::ADD);
5019 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5023 isInc = (Ptr->getOpcode() == ISD::ADD);
5024 Offset = Ptr->getOperand(1);
5026 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5028 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5029 int RHSC = (int)RHS->getZExtValue();
5030 if (RHSC < 0 && RHSC > -0x1000) {
5031 assert(Ptr->getOpcode() == ISD::ADD);
5033 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5034 Base = Ptr->getOperand(0);
5039 if (Ptr->getOpcode() == ISD::ADD) {
5041 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5042 if (ShOpcVal != ARM_AM::no_shift) {
5043 Base = Ptr->getOperand(1);
5044 Offset = Ptr->getOperand(0);
5046 Base = Ptr->getOperand(0);
5047 Offset = Ptr->getOperand(1);
5052 isInc = (Ptr->getOpcode() == ISD::ADD);
5053 Base = Ptr->getOperand(0);
5054 Offset = Ptr->getOperand(1);
5058 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5062 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5063 bool isSEXTLoad, SDValue &Base,
5064 SDValue &Offset, bool &isInc,
5065 SelectionDAG &DAG) {
5066 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5069 Base = Ptr->getOperand(0);
5070 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5071 int RHSC = (int)RHS->getZExtValue();
5072 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5073 assert(Ptr->getOpcode() == ISD::ADD);
5075 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5077 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5078 isInc = Ptr->getOpcode() == ISD::ADD;
5079 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5087 /// getPreIndexedAddressParts - returns true by value, base pointer and
5088 /// offset pointer and addressing mode by reference if the node's address
5089 /// can be legally represented as pre-indexed load / store address.
5091 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5093 ISD::MemIndexedMode &AM,
5094 SelectionDAG &DAG) const {
5095 if (Subtarget->isThumb1Only())
5100 bool isSEXTLoad = false;
5101 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5102 Ptr = LD->getBasePtr();
5103 VT = LD->getMemoryVT();
5104 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5105 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5106 Ptr = ST->getBasePtr();
5107 VT = ST->getMemoryVT();
5112 bool isLegal = false;
5113 if (Subtarget->isThumb2())
5114 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5115 Offset, isInc, DAG);
5117 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5118 Offset, isInc, DAG);
5122 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5126 /// getPostIndexedAddressParts - returns true by value, base pointer and
5127 /// offset pointer and addressing mode by reference if this node can be
5128 /// combined with a load / store to form a post-indexed load / store.
5129 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5132 ISD::MemIndexedMode &AM,
5133 SelectionDAG &DAG) const {
5134 if (Subtarget->isThumb1Only())
5139 bool isSEXTLoad = false;
5140 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5141 VT = LD->getMemoryVT();
5142 Ptr = LD->getBasePtr();
5143 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5144 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5145 VT = ST->getMemoryVT();
5146 Ptr = ST->getBasePtr();
5151 bool isLegal = false;
5152 if (Subtarget->isThumb2())
5153 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5156 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5162 // Swap base ptr and offset to catch more post-index load / store when
5163 // it's legal. In Thumb2 mode, offset must be an immediate.
5164 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5165 !Subtarget->isThumb2())
5166 std::swap(Base, Offset);
5168 // Post-indexed load / store update the base pointer.
5173 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5177 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5181 const SelectionDAG &DAG,
5182 unsigned Depth) const {
5183 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5184 switch (Op.getOpcode()) {
5186 case ARMISD::CMOV: {
5187 // Bits are known zero/one if known on the LHS and RHS.
5188 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5189 if (KnownZero == 0 && KnownOne == 0) return;
5191 APInt KnownZeroRHS, KnownOneRHS;
5192 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5193 KnownZeroRHS, KnownOneRHS, Depth+1);
5194 KnownZero &= KnownZeroRHS;
5195 KnownOne &= KnownOneRHS;
5201 //===----------------------------------------------------------------------===//
5202 // ARM Inline Assembly Support
5203 //===----------------------------------------------------------------------===//
5205 /// getConstraintType - Given a constraint letter, return the type of
5206 /// constraint it is for this target.
5207 ARMTargetLowering::ConstraintType
5208 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5209 if (Constraint.size() == 1) {
5210 switch (Constraint[0]) {
5212 case 'l': return C_RegisterClass;
5213 case 'w': return C_RegisterClass;
5216 return TargetLowering::getConstraintType(Constraint);
5219 std::pair<unsigned, const TargetRegisterClass*>
5220 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5222 if (Constraint.size() == 1) {
5223 // GCC ARM Constraint Letters
5224 switch (Constraint[0]) {
5226 if (Subtarget->isThumb())
5227 return std::make_pair(0U, ARM::tGPRRegisterClass);
5229 return std::make_pair(0U, ARM::GPRRegisterClass);
5231 return std::make_pair(0U, ARM::GPRRegisterClass);
5234 return std::make_pair(0U, ARM::SPRRegisterClass);
5235 if (VT.getSizeInBits() == 64)
5236 return std::make_pair(0U, ARM::DPRRegisterClass);
5237 if (VT.getSizeInBits() == 128)
5238 return std::make_pair(0U, ARM::QPRRegisterClass);
5242 if (StringRef("{cc}").equals_lower(Constraint))
5243 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5245 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5248 std::vector<unsigned> ARMTargetLowering::
5249 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5251 if (Constraint.size() != 1)
5252 return std::vector<unsigned>();
5254 switch (Constraint[0]) { // GCC ARM Constraint Letters
5257 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5258 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5261 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5262 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5263 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5264 ARM::R12, ARM::LR, 0);
5267 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5268 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5269 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5270 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5271 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5272 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5273 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5274 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5275 if (VT.getSizeInBits() == 64)
5276 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5277 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5278 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5279 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5280 if (VT.getSizeInBits() == 128)
5281 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5282 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5286 return std::vector<unsigned>();
5289 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5290 /// vector. If it is invalid, don't add anything to Ops.
5291 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5293 std::vector<SDValue>&Ops,
5294 SelectionDAG &DAG) const {
5295 SDValue Result(0, 0);
5297 switch (Constraint) {
5299 case 'I': case 'J': case 'K': case 'L':
5300 case 'M': case 'N': case 'O':
5301 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5305 int64_t CVal64 = C->getSExtValue();
5306 int CVal = (int) CVal64;
5307 // None of these constraints allow values larger than 32 bits. Check
5308 // that the value fits in an int.
5312 switch (Constraint) {
5314 if (Subtarget->isThumb1Only()) {
5315 // This must be a constant between 0 and 255, for ADD
5317 if (CVal >= 0 && CVal <= 255)
5319 } else if (Subtarget->isThumb2()) {
5320 // A constant that can be used as an immediate value in a
5321 // data-processing instruction.
5322 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5325 // A constant that can be used as an immediate value in a
5326 // data-processing instruction.
5327 if (ARM_AM::getSOImmVal(CVal) != -1)
5333 if (Subtarget->isThumb()) { // FIXME thumb2
5334 // This must be a constant between -255 and -1, for negated ADD
5335 // immediates. This can be used in GCC with an "n" modifier that
5336 // prints the negated value, for use with SUB instructions. It is
5337 // not useful otherwise but is implemented for compatibility.
5338 if (CVal >= -255 && CVal <= -1)
5341 // This must be a constant between -4095 and 4095. It is not clear
5342 // what this constraint is intended for. Implemented for
5343 // compatibility with GCC.
5344 if (CVal >= -4095 && CVal <= 4095)
5350 if (Subtarget->isThumb1Only()) {
5351 // A 32-bit value where only one byte has a nonzero value. Exclude
5352 // zero to match GCC. This constraint is used by GCC internally for
5353 // constants that can be loaded with a move/shift combination.
5354 // It is not useful otherwise but is implemented for compatibility.
5355 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5357 } else if (Subtarget->isThumb2()) {
5358 // A constant whose bitwise inverse can be used as an immediate
5359 // value in a data-processing instruction. This can be used in GCC
5360 // with a "B" modifier that prints the inverted value, for use with
5361 // BIC and MVN instructions. It is not useful otherwise but is
5362 // implemented for compatibility.
5363 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5366 // A constant whose bitwise inverse can be used as an immediate
5367 // value in a data-processing instruction. This can be used in GCC
5368 // with a "B" modifier that prints the inverted value, for use with
5369 // BIC and MVN instructions. It is not useful otherwise but is
5370 // implemented for compatibility.
5371 if (ARM_AM::getSOImmVal(~CVal) != -1)
5377 if (Subtarget->isThumb1Only()) {
5378 // This must be a constant between -7 and 7,
5379 // for 3-operand ADD/SUB immediate instructions.
5380 if (CVal >= -7 && CVal < 7)
5382 } else if (Subtarget->isThumb2()) {
5383 // A constant whose negation can be used as an immediate value in a
5384 // data-processing instruction. This can be used in GCC with an "n"
5385 // modifier that prints the negated value, for use with SUB
5386 // instructions. It is not useful otherwise but is implemented for
5388 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5391 // A constant whose negation can be used as an immediate value in a
5392 // data-processing instruction. This can be used in GCC with an "n"
5393 // modifier that prints the negated value, for use with SUB
5394 // instructions. It is not useful otherwise but is implemented for
5396 if (ARM_AM::getSOImmVal(-CVal) != -1)
5402 if (Subtarget->isThumb()) { // FIXME thumb2
5403 // This must be a multiple of 4 between 0 and 1020, for
5404 // ADD sp + immediate.
5405 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5408 // A power of two or a constant between 0 and 32. This is used in
5409 // GCC for the shift amount on shifted register operands, but it is
5410 // useful in general for any shift amounts.
5411 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5417 if (Subtarget->isThumb()) { // FIXME thumb2
5418 // This must be a constant between 0 and 31, for shift amounts.
5419 if (CVal >= 0 && CVal <= 31)
5425 if (Subtarget->isThumb()) { // FIXME thumb2
5426 // This must be a multiple of 4 between -508 and 508, for
5427 // ADD/SUB sp = sp + immediate.
5428 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5433 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5437 if (Result.getNode()) {
5438 Ops.push_back(Result);
5441 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5445 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5446 // The ARM target isn't yet aware of offsets.
5450 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5451 APInt Imm = FPImm.bitcastToAPInt();
5452 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5453 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5454 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5456 // We can handle 4 bits of mantissa.
5457 // mantissa = (16+UInt(e:f:g:h))/16.
5458 if (Mantissa & 0x7ffff)
5461 if ((Mantissa & 0xf) != Mantissa)
5464 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5465 if (Exp < -3 || Exp > 4)
5467 Exp = ((Exp+3) & 0x7) ^ 4;
5469 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5472 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5473 APInt Imm = FPImm.bitcastToAPInt();
5474 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5475 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5476 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5478 // We can handle 4 bits of mantissa.
5479 // mantissa = (16+UInt(e:f:g:h))/16.
5480 if (Mantissa & 0xffffffffffffLL)
5483 if ((Mantissa & 0xf) != Mantissa)
5486 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5487 if (Exp < -3 || Exp > 4)
5489 Exp = ((Exp+3) & 0x7) ^ 4;
5491 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5494 bool ARM::isBitFieldInvertedMask(unsigned v) {
5495 if (v == 0xffffffff)
5497 // there can be 1's on either or both "outsides", all the "inside"
5499 unsigned int lsb = 0, msb = 31;
5500 while (v & (1 << msb)) --msb;
5501 while (v & (1 << lsb)) ++lsb;
5502 for (unsigned int i = lsb; i <= msb; ++i) {
5509 /// isFPImmLegal - Returns true if the target can instruction select the
5510 /// specified FP immediate natively. If false, the legalizer will
5511 /// materialize the FP immediate as a load from a constant pool.
5512 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5513 if (!Subtarget->hasVFP3())
5516 return ARM::getVFPf32Imm(Imm) != -1;
5518 return ARM::getVFPf64Imm(Imm) != -1;