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);
329 setTargetDAGCombine(ISD::BUILD_VECTOR);
332 computeRegisterProperties();
334 // ARM does not have f32 extending load.
335 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
337 // ARM does not have i1 sign extending load.
338 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
340 // ARM supports all 4 flavors of integer indexed load / store.
341 if (!Subtarget->isThumb1Only()) {
342 for (unsigned im = (unsigned)ISD::PRE_INC;
343 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
344 setIndexedLoadAction(im, MVT::i1, Legal);
345 setIndexedLoadAction(im, MVT::i8, Legal);
346 setIndexedLoadAction(im, MVT::i16, Legal);
347 setIndexedLoadAction(im, MVT::i32, Legal);
348 setIndexedStoreAction(im, MVT::i1, Legal);
349 setIndexedStoreAction(im, MVT::i8, Legal);
350 setIndexedStoreAction(im, MVT::i16, Legal);
351 setIndexedStoreAction(im, MVT::i32, Legal);
355 // i64 operation support.
356 if (Subtarget->isThumb1Only()) {
357 setOperationAction(ISD::MUL, MVT::i64, Expand);
358 setOperationAction(ISD::MULHU, MVT::i32, Expand);
359 setOperationAction(ISD::MULHS, MVT::i32, Expand);
360 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
361 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
363 setOperationAction(ISD::MUL, MVT::i64, Expand);
364 setOperationAction(ISD::MULHU, MVT::i32, Expand);
365 if (!Subtarget->hasV6Ops())
366 setOperationAction(ISD::MULHS, MVT::i32, Expand);
368 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
369 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
370 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
371 setOperationAction(ISD::SRL, MVT::i64, Custom);
372 setOperationAction(ISD::SRA, MVT::i64, Custom);
374 // ARM does not have ROTL.
375 setOperationAction(ISD::ROTL, MVT::i32, Expand);
376 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
377 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
378 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
379 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
381 // Only ARMv6 has BSWAP.
382 if (!Subtarget->hasV6Ops())
383 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
385 // These are expanded into libcalls.
386 if (!Subtarget->hasDivide()) {
387 // v7M has a hardware divider
388 setOperationAction(ISD::SDIV, MVT::i32, Expand);
389 setOperationAction(ISD::UDIV, MVT::i32, Expand);
391 setOperationAction(ISD::SREM, MVT::i32, Expand);
392 setOperationAction(ISD::UREM, MVT::i32, Expand);
393 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
394 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
396 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
397 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
398 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
399 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
400 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
402 setOperationAction(ISD::TRAP, MVT::Other, Legal);
404 // Use the default implementation.
405 setOperationAction(ISD::VASTART, MVT::Other, Custom);
406 setOperationAction(ISD::VAARG, MVT::Other, Expand);
407 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
408 setOperationAction(ISD::VAEND, MVT::Other, Expand);
409 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
410 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
411 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
412 // FIXME: Shouldn't need this, since no register is used, but the legalizer
413 // doesn't yet know how to not do that for SjLj.
414 setExceptionSelectorRegister(ARM::R0);
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
416 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
417 // the default expansion.
418 if (Subtarget->hasDataBarrier() ||
419 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) {
420 // membarrier needs custom lowering; the rest are legal and handled
422 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
424 // Set them all for expansion, which will force libcalls.
425 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
426 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
427 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
428 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
429 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
430 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
431 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
432 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
449 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
450 // Since the libcalls include locking, fold in the fences
451 setShouldFoldAtomicFences(true);
453 // 64-bit versions are always libcalls (for now)
454 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
455 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
456 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
457 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
458 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
459 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
460 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
461 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
463 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
464 if (!Subtarget->hasV6Ops()) {
465 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
466 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
468 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
470 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
471 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
472 // iff target supports vfp2.
473 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
474 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
477 // We want to custom lower some of our intrinsics.
478 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
479 if (Subtarget->isTargetDarwin()) {
480 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
481 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
484 setOperationAction(ISD::SETCC, MVT::i32, Expand);
485 setOperationAction(ISD::SETCC, MVT::f32, Expand);
486 setOperationAction(ISD::SETCC, MVT::f64, Expand);
487 setOperationAction(ISD::SELECT, MVT::i32, Custom);
488 setOperationAction(ISD::SELECT, MVT::f32, Custom);
489 setOperationAction(ISD::SELECT, MVT::f64, Custom);
490 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
491 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
492 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
494 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
495 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
496 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
497 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
498 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
500 // We don't support sin/cos/fmod/copysign/pow
501 setOperationAction(ISD::FSIN, MVT::f64, Expand);
502 setOperationAction(ISD::FSIN, MVT::f32, Expand);
503 setOperationAction(ISD::FCOS, MVT::f32, Expand);
504 setOperationAction(ISD::FCOS, MVT::f64, Expand);
505 setOperationAction(ISD::FREM, MVT::f64, Expand);
506 setOperationAction(ISD::FREM, MVT::f32, Expand);
507 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
508 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
509 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
511 setOperationAction(ISD::FPOW, MVT::f64, Expand);
512 setOperationAction(ISD::FPOW, MVT::f32, Expand);
514 // Various VFP goodness
515 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
516 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
517 if (Subtarget->hasVFP2()) {
518 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
519 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
520 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
521 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
523 // Special handling for half-precision FP.
524 if (!Subtarget->hasFP16()) {
525 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
526 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
530 // We have target-specific dag combine patterns for the following nodes:
531 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
532 setTargetDAGCombine(ISD::ADD);
533 setTargetDAGCombine(ISD::SUB);
534 setTargetDAGCombine(ISD::MUL);
536 if (Subtarget->hasV6T2Ops())
537 setTargetDAGCombine(ISD::OR);
539 setStackPointerRegisterToSaveRestore(ARM::SP);
541 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
542 setSchedulingPreference(Sched::RegPressure);
544 setSchedulingPreference(Sched::Hybrid);
546 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
548 // On ARM arguments smaller than 4 bytes are extended, so all arguments
549 // are at least 4 bytes aligned.
550 setMinStackArgumentAlignment(4);
552 if (EnableARMCodePlacement)
553 benefitFromCodePlacementOpt = true;
556 std::pair<const TargetRegisterClass*, uint8_t>
557 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
558 const TargetRegisterClass *RRC = 0;
560 switch (VT.getSimpleVT().SimpleTy) {
562 return TargetLowering::findRepresentativeClass(VT);
563 // Use DPR as representative register class for all floating point
564 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
565 // the cost is 1 for both f32 and f64.
566 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
567 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
568 RRC = ARM::DPRRegisterClass;
570 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
571 case MVT::v4f32: case MVT::v2f64:
572 RRC = ARM::DPRRegisterClass;
576 RRC = ARM::DPRRegisterClass;
580 RRC = ARM::DPRRegisterClass;
584 return std::make_pair(RRC, Cost);
587 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
590 case ARMISD::Wrapper: return "ARMISD::Wrapper";
591 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
592 case ARMISD::CALL: return "ARMISD::CALL";
593 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
594 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
595 case ARMISD::tCALL: return "ARMISD::tCALL";
596 case ARMISD::BRCOND: return "ARMISD::BRCOND";
597 case ARMISD::BR_JT: return "ARMISD::BR_JT";
598 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
599 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
600 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
601 case ARMISD::AND: return "ARMISD::AND";
602 case ARMISD::CMP: return "ARMISD::CMP";
603 case ARMISD::CMPZ: return "ARMISD::CMPZ";
604 case ARMISD::CMPFP: return "ARMISD::CMPFP";
605 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
606 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
607 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
608 case ARMISD::CMOV: return "ARMISD::CMOV";
609 case ARMISD::CNEG: return "ARMISD::CNEG";
611 case ARMISD::RBIT: return "ARMISD::RBIT";
613 case ARMISD::FTOSI: return "ARMISD::FTOSI";
614 case ARMISD::FTOUI: return "ARMISD::FTOUI";
615 case ARMISD::SITOF: return "ARMISD::SITOF";
616 case ARMISD::UITOF: return "ARMISD::UITOF";
618 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
619 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
620 case ARMISD::RRX: return "ARMISD::RRX";
622 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
623 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
625 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
626 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
628 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
630 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
632 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
634 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
635 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
637 case ARMISD::VCEQ: return "ARMISD::VCEQ";
638 case ARMISD::VCGE: return "ARMISD::VCGE";
639 case ARMISD::VCGEU: return "ARMISD::VCGEU";
640 case ARMISD::VCGT: return "ARMISD::VCGT";
641 case ARMISD::VCGTU: return "ARMISD::VCGTU";
642 case ARMISD::VTST: return "ARMISD::VTST";
644 case ARMISD::VSHL: return "ARMISD::VSHL";
645 case ARMISD::VSHRs: return "ARMISD::VSHRs";
646 case ARMISD::VSHRu: return "ARMISD::VSHRu";
647 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
648 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
649 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
650 case ARMISD::VSHRN: return "ARMISD::VSHRN";
651 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
652 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
653 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
654 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
655 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
656 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
657 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
658 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
659 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
660 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
661 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
662 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
663 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
664 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
665 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
666 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
667 case ARMISD::VDUP: return "ARMISD::VDUP";
668 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
669 case ARMISD::VEXT: return "ARMISD::VEXT";
670 case ARMISD::VREV64: return "ARMISD::VREV64";
671 case ARMISD::VREV32: return "ARMISD::VREV32";
672 case ARMISD::VREV16: return "ARMISD::VREV16";
673 case ARMISD::VZIP: return "ARMISD::VZIP";
674 case ARMISD::VUZP: return "ARMISD::VUZP";
675 case ARMISD::VTRN: return "ARMISD::VTRN";
676 case ARMISD::VMULLs: return "ARMISD::VMULLs";
677 case ARMISD::VMULLu: return "ARMISD::VMULLu";
678 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
679 case ARMISD::FMAX: return "ARMISD::FMAX";
680 case ARMISD::FMIN: return "ARMISD::FMIN";
681 case ARMISD::BFI: return "ARMISD::BFI";
685 /// getRegClassFor - Return the register class that should be used for the
686 /// specified value type.
687 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
688 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
689 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
690 // load / store 4 to 8 consecutive D registers.
691 if (Subtarget->hasNEON()) {
692 if (VT == MVT::v4i64)
693 return ARM::QQPRRegisterClass;
694 else if (VT == MVT::v8i64)
695 return ARM::QQQQPRRegisterClass;
697 return TargetLowering::getRegClassFor(VT);
700 // Create a fast isel object.
702 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
703 return ARM::createFastISel(funcInfo);
706 /// getFunctionAlignment - Return the Log2 alignment of this function.
707 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
708 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
711 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
712 /// be used for loads / stores from the global.
713 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
714 return (Subtarget->isThumb1Only() ? 127 : 4095);
717 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
718 unsigned NumVals = N->getNumValues();
720 return Sched::RegPressure;
722 for (unsigned i = 0; i != NumVals; ++i) {
723 EVT VT = N->getValueType(i);
724 if (VT.isFloatingPoint() || VT.isVector())
725 return Sched::Latency;
728 if (!N->isMachineOpcode())
729 return Sched::RegPressure;
731 // Load are scheduled for latency even if there instruction itinerary
733 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
734 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
736 return Sched::Latency;
738 if (!Itins->isEmpty() && Itins->getStageLatency(TID.getSchedClass()) > 2)
739 return Sched::Latency;
740 return Sched::RegPressure;
744 ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
745 MachineFunction &MF) const {
746 switch (RC->getID()) {
749 case ARM::tGPRRegClassID:
750 return RegInfo->hasFP(MF) ? 4 : 5;
751 case ARM::GPRRegClassID: {
752 unsigned FP = RegInfo->hasFP(MF) ? 1 : 0;
753 return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
755 case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
756 case ARM::DPRRegClassID:
761 //===----------------------------------------------------------------------===//
763 //===----------------------------------------------------------------------===//
765 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
766 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
768 default: llvm_unreachable("Unknown condition code!");
769 case ISD::SETNE: return ARMCC::NE;
770 case ISD::SETEQ: return ARMCC::EQ;
771 case ISD::SETGT: return ARMCC::GT;
772 case ISD::SETGE: return ARMCC::GE;
773 case ISD::SETLT: return ARMCC::LT;
774 case ISD::SETLE: return ARMCC::LE;
775 case ISD::SETUGT: return ARMCC::HI;
776 case ISD::SETUGE: return ARMCC::HS;
777 case ISD::SETULT: return ARMCC::LO;
778 case ISD::SETULE: return ARMCC::LS;
782 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
783 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
784 ARMCC::CondCodes &CondCode2) {
785 CondCode2 = ARMCC::AL;
787 default: llvm_unreachable("Unknown FP condition!");
789 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
791 case ISD::SETOGT: CondCode = ARMCC::GT; break;
793 case ISD::SETOGE: CondCode = ARMCC::GE; break;
794 case ISD::SETOLT: CondCode = ARMCC::MI; break;
795 case ISD::SETOLE: CondCode = ARMCC::LS; break;
796 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
797 case ISD::SETO: CondCode = ARMCC::VC; break;
798 case ISD::SETUO: CondCode = ARMCC::VS; break;
799 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
800 case ISD::SETUGT: CondCode = ARMCC::HI; break;
801 case ISD::SETUGE: CondCode = ARMCC::PL; break;
803 case ISD::SETULT: CondCode = ARMCC::LT; break;
805 case ISD::SETULE: CondCode = ARMCC::LE; break;
807 case ISD::SETUNE: CondCode = ARMCC::NE; break;
811 //===----------------------------------------------------------------------===//
812 // Calling Convention Implementation
813 //===----------------------------------------------------------------------===//
815 #include "ARMGenCallingConv.inc"
817 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
818 /// given CallingConvention value.
819 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
821 bool isVarArg) const {
824 llvm_unreachable("Unsupported calling convention");
826 case CallingConv::Fast:
827 // Use target triple & subtarget features to do actual dispatch.
828 if (Subtarget->isAAPCS_ABI()) {
829 if (Subtarget->hasVFP2() &&
830 FloatABIType == FloatABI::Hard && !isVarArg)
831 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
833 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
835 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
836 case CallingConv::ARM_AAPCS_VFP:
837 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
838 case CallingConv::ARM_AAPCS:
839 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
840 case CallingConv::ARM_APCS:
841 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
845 /// LowerCallResult - Lower the result values of a call into the
846 /// appropriate copies out of appropriate physical registers.
848 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
849 CallingConv::ID CallConv, bool isVarArg,
850 const SmallVectorImpl<ISD::InputArg> &Ins,
851 DebugLoc dl, SelectionDAG &DAG,
852 SmallVectorImpl<SDValue> &InVals) const {
854 // Assign locations to each value returned by this call.
855 SmallVector<CCValAssign, 16> RVLocs;
856 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
857 RVLocs, *DAG.getContext());
858 CCInfo.AnalyzeCallResult(Ins,
859 CCAssignFnForNode(CallConv, /* Return*/ true,
862 // Copy all of the result registers out of their specified physreg.
863 for (unsigned i = 0; i != RVLocs.size(); ++i) {
864 CCValAssign VA = RVLocs[i];
867 if (VA.needsCustom()) {
868 // Handle f64 or half of a v2f64.
869 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
871 Chain = Lo.getValue(1);
872 InFlag = Lo.getValue(2);
873 VA = RVLocs[++i]; // skip ahead to next loc
874 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
876 Chain = Hi.getValue(1);
877 InFlag = Hi.getValue(2);
878 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
880 if (VA.getLocVT() == MVT::v2f64) {
881 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
882 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
883 DAG.getConstant(0, MVT::i32));
885 VA = RVLocs[++i]; // skip ahead to next loc
886 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
887 Chain = Lo.getValue(1);
888 InFlag = Lo.getValue(2);
889 VA = RVLocs[++i]; // skip ahead to next loc
890 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
891 Chain = Hi.getValue(1);
892 InFlag = Hi.getValue(2);
893 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
894 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
895 DAG.getConstant(1, MVT::i32));
898 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
900 Chain = Val.getValue(1);
901 InFlag = Val.getValue(2);
904 switch (VA.getLocInfo()) {
905 default: llvm_unreachable("Unknown loc info!");
906 case CCValAssign::Full: break;
907 case CCValAssign::BCvt:
908 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
912 InVals.push_back(Val);
918 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
919 /// by "Src" to address "Dst" of size "Size". Alignment information is
920 /// specified by the specific parameter attribute. The copy will be passed as
921 /// a byval function parameter.
922 /// Sometimes what we are copying is the end of a larger object, the part that
923 /// does not fit in registers.
925 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
926 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
928 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
929 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
930 /*isVolatile=*/false, /*AlwaysInline=*/false,
931 MachinePointerInfo(0), MachinePointerInfo(0));
934 /// LowerMemOpCallTo - Store the argument to the stack.
936 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
937 SDValue StackPtr, SDValue Arg,
938 DebugLoc dl, SelectionDAG &DAG,
939 const CCValAssign &VA,
940 ISD::ArgFlagsTy Flags) const {
941 unsigned LocMemOffset = VA.getLocMemOffset();
942 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
943 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
944 if (Flags.isByVal()) {
945 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
947 return DAG.getStore(Chain, dl, Arg, PtrOff,
948 PseudoSourceValue::getStack(), LocMemOffset,
952 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
953 SDValue Chain, SDValue &Arg,
954 RegsToPassVector &RegsToPass,
955 CCValAssign &VA, CCValAssign &NextVA,
957 SmallVector<SDValue, 8> &MemOpChains,
958 ISD::ArgFlagsTy Flags) const {
960 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
961 DAG.getVTList(MVT::i32, MVT::i32), Arg);
962 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
964 if (NextVA.isRegLoc())
965 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
967 assert(NextVA.isMemLoc());
968 if (StackPtr.getNode() == 0)
969 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
971 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
977 /// LowerCall - Lowering a call into a callseq_start <-
978 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
981 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
982 CallingConv::ID CallConv, bool isVarArg,
984 const SmallVectorImpl<ISD::OutputArg> &Outs,
985 const SmallVectorImpl<SDValue> &OutVals,
986 const SmallVectorImpl<ISD::InputArg> &Ins,
987 DebugLoc dl, SelectionDAG &DAG,
988 SmallVectorImpl<SDValue> &InVals) const {
989 MachineFunction &MF = DAG.getMachineFunction();
990 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
991 bool IsSibCall = false;
992 // Temporarily disable tail calls so things don't break.
993 if (!EnableARMTailCalls)
996 // Check if it's really possible to do a tail call.
997 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
998 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
999 Outs, OutVals, Ins, DAG);
1000 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1001 // detected sibcalls.
1008 // Analyze operands of the call, assigning locations to each operand.
1009 SmallVector<CCValAssign, 16> ArgLocs;
1010 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1012 CCInfo.AnalyzeCallOperands(Outs,
1013 CCAssignFnForNode(CallConv, /* Return*/ false,
1016 // Get a count of how many bytes are to be pushed on the stack.
1017 unsigned NumBytes = CCInfo.getNextStackOffset();
1019 // For tail calls, memory operands are available in our caller's stack.
1023 // Adjust the stack pointer for the new arguments...
1024 // These operations are automatically eliminated by the prolog/epilog pass
1026 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1028 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1030 RegsToPassVector RegsToPass;
1031 SmallVector<SDValue, 8> MemOpChains;
1033 // Walk the register/memloc assignments, inserting copies/loads. In the case
1034 // of tail call optimization, arguments are handled later.
1035 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1037 ++i, ++realArgIdx) {
1038 CCValAssign &VA = ArgLocs[i];
1039 SDValue Arg = OutVals[realArgIdx];
1040 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1042 // Promote the value if needed.
1043 switch (VA.getLocInfo()) {
1044 default: llvm_unreachable("Unknown loc info!");
1045 case CCValAssign::Full: break;
1046 case CCValAssign::SExt:
1047 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1049 case CCValAssign::ZExt:
1050 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1052 case CCValAssign::AExt:
1053 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1055 case CCValAssign::BCvt:
1056 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1060 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1061 if (VA.needsCustom()) {
1062 if (VA.getLocVT() == MVT::v2f64) {
1063 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1064 DAG.getConstant(0, MVT::i32));
1065 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1066 DAG.getConstant(1, MVT::i32));
1068 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1069 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1071 VA = ArgLocs[++i]; // skip ahead to next loc
1072 if (VA.isRegLoc()) {
1073 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1074 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1076 assert(VA.isMemLoc());
1078 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1079 dl, DAG, VA, Flags));
1082 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1083 StackPtr, MemOpChains, Flags);
1085 } else if (VA.isRegLoc()) {
1086 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1087 } else if (!IsSibCall) {
1088 assert(VA.isMemLoc());
1090 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1091 dl, DAG, VA, Flags));
1095 if (!MemOpChains.empty())
1096 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1097 &MemOpChains[0], MemOpChains.size());
1099 // Build a sequence of copy-to-reg nodes chained together with token chain
1100 // and flag operands which copy the outgoing args into the appropriate regs.
1102 // Tail call byval lowering might overwrite argument registers so in case of
1103 // tail call optimization the copies to registers are lowered later.
1105 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1106 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1107 RegsToPass[i].second, InFlag);
1108 InFlag = Chain.getValue(1);
1111 // For tail calls lower the arguments to the 'real' stack slot.
1113 // Force all the incoming stack arguments to be loaded from the stack
1114 // before any new outgoing arguments are stored to the stack, because the
1115 // outgoing stack slots may alias the incoming argument stack slots, and
1116 // the alias isn't otherwise explicit. This is slightly more conservative
1117 // than necessary, because it means that each store effectively depends
1118 // on every argument instead of just those arguments it would clobber.
1120 // Do not flag preceeding copytoreg stuff together with the following stuff.
1122 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1123 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1124 RegsToPass[i].second, InFlag);
1125 InFlag = Chain.getValue(1);
1130 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1131 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1132 // node so that legalize doesn't hack it.
1133 bool isDirect = false;
1134 bool isARMFunc = false;
1135 bool isLocalARMFunc = false;
1136 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1138 if (EnableARMLongCalls) {
1139 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1140 && "long-calls with non-static relocation model!");
1141 // Handle a global address or an external symbol. If it's not one of
1142 // those, the target's already in a register, so we don't need to do
1144 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1145 const GlobalValue *GV = G->getGlobal();
1146 // Create a constant pool entry for the callee address
1147 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1148 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1151 // Get the address of the callee into a register
1152 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1153 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1154 Callee = DAG.getLoad(getPointerTy(), dl,
1155 DAG.getEntryNode(), CPAddr,
1156 MachinePointerInfo::getConstantPool(),
1158 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1159 const char *Sym = S->getSymbol();
1161 // Create a constant pool entry for the callee address
1162 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1163 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1164 Sym, ARMPCLabelIndex, 0);
1165 // Get the address of the callee into a register
1166 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1167 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1168 Callee = DAG.getLoad(getPointerTy(), dl,
1169 DAG.getEntryNode(), CPAddr,
1170 MachinePointerInfo::getConstantPool(),
1173 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1174 const GlobalValue *GV = G->getGlobal();
1176 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1177 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1178 getTargetMachine().getRelocationModel() != Reloc::Static;
1179 isARMFunc = !Subtarget->isThumb() || isStub;
1180 // ARM call to a local ARM function is predicable.
1181 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1182 // tBX takes a register source operand.
1183 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1184 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1185 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1188 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1189 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1190 Callee = DAG.getLoad(getPointerTy(), dl,
1191 DAG.getEntryNode(), CPAddr,
1192 MachinePointerInfo::getConstantPool(),
1194 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1195 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1196 getPointerTy(), Callee, PICLabel);
1198 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1199 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1201 bool isStub = Subtarget->isTargetDarwin() &&
1202 getTargetMachine().getRelocationModel() != Reloc::Static;
1203 isARMFunc = !Subtarget->isThumb() || isStub;
1204 // tBX takes a register source operand.
1205 const char *Sym = S->getSymbol();
1206 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1207 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1208 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1209 Sym, ARMPCLabelIndex, 4);
1210 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1211 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1212 Callee = DAG.getLoad(getPointerTy(), dl,
1213 DAG.getEntryNode(), CPAddr,
1214 MachinePointerInfo::getConstantPool(),
1216 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1217 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1218 getPointerTy(), Callee, PICLabel);
1220 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1223 // FIXME: handle tail calls differently.
1225 if (Subtarget->isThumb()) {
1226 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1227 CallOpc = ARMISD::CALL_NOLINK;
1229 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1231 CallOpc = (isDirect || Subtarget->hasV5TOps())
1232 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1233 : ARMISD::CALL_NOLINK;
1236 std::vector<SDValue> Ops;
1237 Ops.push_back(Chain);
1238 Ops.push_back(Callee);
1240 // Add argument registers to the end of the list so that they are known live
1242 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1243 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1244 RegsToPass[i].second.getValueType()));
1246 if (InFlag.getNode())
1247 Ops.push_back(InFlag);
1249 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1251 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1253 // Returns a chain and a flag for retval copy to use.
1254 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1255 InFlag = Chain.getValue(1);
1257 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1258 DAG.getIntPtrConstant(0, true), InFlag);
1260 InFlag = Chain.getValue(1);
1262 // Handle result values, copying them out of physregs into vregs that we
1264 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1268 /// MatchingStackOffset - Return true if the given stack call argument is
1269 /// already available in the same position (relatively) of the caller's
1270 /// incoming argument stack.
1272 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1273 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1274 const ARMInstrInfo *TII) {
1275 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1277 if (Arg.getOpcode() == ISD::CopyFromReg) {
1278 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1279 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1281 MachineInstr *Def = MRI->getVRegDef(VR);
1284 if (!Flags.isByVal()) {
1285 if (!TII->isLoadFromStackSlot(Def, FI))
1290 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1291 if (Flags.isByVal())
1292 // ByVal argument is passed in as a pointer but it's now being
1293 // dereferenced. e.g.
1294 // define @foo(%struct.X* %A) {
1295 // tail call @bar(%struct.X* byval %A)
1298 SDValue Ptr = Ld->getBasePtr();
1299 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1302 FI = FINode->getIndex();
1306 assert(FI != INT_MAX);
1307 if (!MFI->isFixedObjectIndex(FI))
1309 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1312 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1313 /// for tail call optimization. Targets which want to do tail call
1314 /// optimization should implement this function.
1316 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1317 CallingConv::ID CalleeCC,
1319 bool isCalleeStructRet,
1320 bool isCallerStructRet,
1321 const SmallVectorImpl<ISD::OutputArg> &Outs,
1322 const SmallVectorImpl<SDValue> &OutVals,
1323 const SmallVectorImpl<ISD::InputArg> &Ins,
1324 SelectionDAG& DAG) const {
1325 const Function *CallerF = DAG.getMachineFunction().getFunction();
1326 CallingConv::ID CallerCC = CallerF->getCallingConv();
1327 bool CCMatch = CallerCC == CalleeCC;
1329 // Look for obvious safe cases to perform tail call optimization that do not
1330 // require ABI changes. This is what gcc calls sibcall.
1332 // Do not sibcall optimize vararg calls unless the call site is not passing
1334 if (isVarArg && !Outs.empty())
1337 // Also avoid sibcall optimization if either caller or callee uses struct
1338 // return semantics.
1339 if (isCalleeStructRet || isCallerStructRet)
1342 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1343 // emitEpilogue is not ready for them.
1344 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1345 // LR. This means if we need to reload LR, it takes an extra instructions,
1346 // which outweighs the value of the tail call; but here we don't know yet
1347 // whether LR is going to be used. Probably the right approach is to
1348 // generate the tail call here and turn it back into CALL/RET in
1349 // emitEpilogue if LR is used.
1350 if (Subtarget->isThumb1Only())
1353 // For the moment, we can only do this to functions defined in this
1354 // compilation, or to indirect calls. A Thumb B to an ARM function,
1355 // or vice versa, is not easily fixed up in the linker unlike BL.
1356 // (We could do this by loading the address of the callee into a register;
1357 // that is an extra instruction over the direct call and burns a register
1358 // as well, so is not likely to be a win.)
1360 // It might be safe to remove this restriction on non-Darwin.
1362 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1363 // but we need to make sure there are enough registers; the only valid
1364 // registers are the 4 used for parameters. We don't currently do this
1366 if (isa<ExternalSymbolSDNode>(Callee))
1369 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1370 const GlobalValue *GV = G->getGlobal();
1371 if (GV->isDeclaration() || GV->isWeakForLinker())
1375 // If the calling conventions do not match, then we'd better make sure the
1376 // results are returned in the same way as what the caller expects.
1378 SmallVector<CCValAssign, 16> RVLocs1;
1379 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1380 RVLocs1, *DAG.getContext());
1381 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1383 SmallVector<CCValAssign, 16> RVLocs2;
1384 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1385 RVLocs2, *DAG.getContext());
1386 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1388 if (RVLocs1.size() != RVLocs2.size())
1390 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1391 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1393 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1395 if (RVLocs1[i].isRegLoc()) {
1396 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1399 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1405 // If the callee takes no arguments then go on to check the results of the
1407 if (!Outs.empty()) {
1408 // Check if stack adjustment is needed. For now, do not do this if any
1409 // argument is passed on the stack.
1410 SmallVector<CCValAssign, 16> ArgLocs;
1411 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1412 ArgLocs, *DAG.getContext());
1413 CCInfo.AnalyzeCallOperands(Outs,
1414 CCAssignFnForNode(CalleeCC, false, isVarArg));
1415 if (CCInfo.getNextStackOffset()) {
1416 MachineFunction &MF = DAG.getMachineFunction();
1418 // Check if the arguments are already laid out in the right way as
1419 // the caller's fixed stack objects.
1420 MachineFrameInfo *MFI = MF.getFrameInfo();
1421 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1422 const ARMInstrInfo *TII =
1423 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1424 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1426 ++i, ++realArgIdx) {
1427 CCValAssign &VA = ArgLocs[i];
1428 EVT RegVT = VA.getLocVT();
1429 SDValue Arg = OutVals[realArgIdx];
1430 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1431 if (VA.getLocInfo() == CCValAssign::Indirect)
1433 if (VA.needsCustom()) {
1434 // f64 and vector types are split into multiple registers or
1435 // register/stack-slot combinations. The types will not match
1436 // the registers; give up on memory f64 refs until we figure
1437 // out what to do about this.
1440 if (!ArgLocs[++i].isRegLoc())
1442 if (RegVT == MVT::v2f64) {
1443 if (!ArgLocs[++i].isRegLoc())
1445 if (!ArgLocs[++i].isRegLoc())
1448 } else if (!VA.isRegLoc()) {
1449 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1461 ARMTargetLowering::LowerReturn(SDValue Chain,
1462 CallingConv::ID CallConv, bool isVarArg,
1463 const SmallVectorImpl<ISD::OutputArg> &Outs,
1464 const SmallVectorImpl<SDValue> &OutVals,
1465 DebugLoc dl, SelectionDAG &DAG) const {
1467 // CCValAssign - represent the assignment of the return value to a location.
1468 SmallVector<CCValAssign, 16> RVLocs;
1470 // CCState - Info about the registers and stack slots.
1471 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1474 // Analyze outgoing return values.
1475 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1478 // If this is the first return lowered for this function, add
1479 // the regs to the liveout set for the function.
1480 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1481 for (unsigned i = 0; i != RVLocs.size(); ++i)
1482 if (RVLocs[i].isRegLoc())
1483 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1488 // Copy the result values into the output registers.
1489 for (unsigned i = 0, realRVLocIdx = 0;
1491 ++i, ++realRVLocIdx) {
1492 CCValAssign &VA = RVLocs[i];
1493 assert(VA.isRegLoc() && "Can only return in registers!");
1495 SDValue Arg = OutVals[realRVLocIdx];
1497 switch (VA.getLocInfo()) {
1498 default: llvm_unreachable("Unknown loc info!");
1499 case CCValAssign::Full: break;
1500 case CCValAssign::BCvt:
1501 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1505 if (VA.needsCustom()) {
1506 if (VA.getLocVT() == MVT::v2f64) {
1507 // Extract the first half and return it in two registers.
1508 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1509 DAG.getConstant(0, MVT::i32));
1510 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1511 DAG.getVTList(MVT::i32, MVT::i32), Half);
1513 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1514 Flag = Chain.getValue(1);
1515 VA = RVLocs[++i]; // skip ahead to next loc
1516 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1517 HalfGPRs.getValue(1), Flag);
1518 Flag = Chain.getValue(1);
1519 VA = RVLocs[++i]; // skip ahead to next loc
1521 // Extract the 2nd half and fall through to handle it as an f64 value.
1522 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1523 DAG.getConstant(1, MVT::i32));
1525 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1527 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1528 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1529 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1530 Flag = Chain.getValue(1);
1531 VA = RVLocs[++i]; // skip ahead to next loc
1532 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1535 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1537 // Guarantee that all emitted copies are
1538 // stuck together, avoiding something bad.
1539 Flag = Chain.getValue(1);
1544 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1546 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1551 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1552 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1553 // one of the above mentioned nodes. It has to be wrapped because otherwise
1554 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1555 // be used to form addressing mode. These wrapped nodes will be selected
1557 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1558 EVT PtrVT = Op.getValueType();
1559 // FIXME there is no actual debug info here
1560 DebugLoc dl = Op.getDebugLoc();
1561 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1563 if (CP->isMachineConstantPoolEntry())
1564 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1565 CP->getAlignment());
1567 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1568 CP->getAlignment());
1569 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1572 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1573 return MachineJumpTableInfo::EK_Inline;
1576 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1577 SelectionDAG &DAG) const {
1578 MachineFunction &MF = DAG.getMachineFunction();
1579 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1580 unsigned ARMPCLabelIndex = 0;
1581 DebugLoc DL = Op.getDebugLoc();
1582 EVT PtrVT = getPointerTy();
1583 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1584 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1586 if (RelocM == Reloc::Static) {
1587 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1589 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1590 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1591 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1592 ARMCP::CPBlockAddress,
1594 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1596 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1597 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1598 MachinePointerInfo::getConstantPool(),
1600 if (RelocM == Reloc::Static)
1602 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1603 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1606 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1608 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1609 SelectionDAG &DAG) const {
1610 DebugLoc dl = GA->getDebugLoc();
1611 EVT PtrVT = getPointerTy();
1612 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1613 MachineFunction &MF = DAG.getMachineFunction();
1614 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1615 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1616 ARMConstantPoolValue *CPV =
1617 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1618 ARMCP::CPValue, PCAdj, "tlsgd", true);
1619 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1620 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1621 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1622 MachinePointerInfo::getConstantPool(),
1624 SDValue Chain = Argument.getValue(1);
1626 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1627 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1629 // call __tls_get_addr.
1632 Entry.Node = Argument;
1633 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1634 Args.push_back(Entry);
1635 // FIXME: is there useful debug info available here?
1636 std::pair<SDValue, SDValue> CallResult =
1637 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1638 false, false, false, false,
1639 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1640 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1641 return CallResult.first;
1644 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1645 // "local exec" model.
1647 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1648 SelectionDAG &DAG) const {
1649 const GlobalValue *GV = GA->getGlobal();
1650 DebugLoc dl = GA->getDebugLoc();
1652 SDValue Chain = DAG.getEntryNode();
1653 EVT PtrVT = getPointerTy();
1654 // Get the Thread Pointer
1655 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1657 if (GV->isDeclaration()) {
1658 MachineFunction &MF = DAG.getMachineFunction();
1659 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1660 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1661 // Initial exec model.
1662 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1663 ARMConstantPoolValue *CPV =
1664 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1665 ARMCP::CPValue, PCAdj, "gottpoff", true);
1666 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1667 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1668 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1669 MachinePointerInfo::getConstantPool(),
1671 Chain = Offset.getValue(1);
1673 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1674 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1676 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1677 MachinePointerInfo::getConstantPool(),
1681 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1682 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1683 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1684 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1685 MachinePointerInfo::getConstantPool(),
1689 // The address of the thread local variable is the add of the thread
1690 // pointer with the offset of the variable.
1691 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1695 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1696 // TODO: implement the "local dynamic" model
1697 assert(Subtarget->isTargetELF() &&
1698 "TLS not implemented for non-ELF targets");
1699 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1700 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1701 // otherwise use the "Local Exec" TLS Model
1702 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1703 return LowerToTLSGeneralDynamicModel(GA, DAG);
1705 return LowerToTLSExecModels(GA, DAG);
1708 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1709 SelectionDAG &DAG) const {
1710 EVT PtrVT = getPointerTy();
1711 DebugLoc dl = Op.getDebugLoc();
1712 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1713 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1714 if (RelocM == Reloc::PIC_) {
1715 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1716 ARMConstantPoolValue *CPV =
1717 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1718 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1719 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1720 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1722 MachinePointerInfo::getConstantPool(),
1724 SDValue Chain = Result.getValue(1);
1725 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1726 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1728 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1729 MachinePointerInfo::getGOT(), false, false, 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 MachinePointerInfo::getConstantPool(),
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 MachinePointerInfo::getConstantPool(),
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, MachinePointerInfo::getGOT(),
1785 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1786 SelectionDAG &DAG) const {
1787 assert(Subtarget->isTargetELF() &&
1788 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1789 MachineFunction &MF = DAG.getMachineFunction();
1790 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1791 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1792 EVT PtrVT = getPointerTy();
1793 DebugLoc dl = Op.getDebugLoc();
1794 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1795 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1796 "_GLOBAL_OFFSET_TABLE_",
1797 ARMPCLabelIndex, PCAdj);
1798 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1799 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1800 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1801 MachinePointerInfo::getConstantPool(),
1803 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1804 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1808 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1809 DebugLoc dl = Op.getDebugLoc();
1810 SDValue Val = DAG.getConstant(0, MVT::i32);
1811 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1812 Op.getOperand(1), Val);
1816 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1817 DebugLoc dl = Op.getDebugLoc();
1818 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1819 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1823 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1824 const ARMSubtarget *Subtarget) const {
1825 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1826 DebugLoc dl = Op.getDebugLoc();
1828 default: return SDValue(); // Don't custom lower most intrinsics.
1829 case Intrinsic::arm_thread_pointer: {
1830 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1831 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1833 case Intrinsic::eh_sjlj_lsda: {
1834 MachineFunction &MF = DAG.getMachineFunction();
1835 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1836 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1837 EVT PtrVT = getPointerTy();
1838 DebugLoc dl = Op.getDebugLoc();
1839 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1841 unsigned PCAdj = (RelocM != Reloc::PIC_)
1842 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1843 ARMConstantPoolValue *CPV =
1844 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1845 ARMCP::CPLSDA, PCAdj);
1846 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1847 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1849 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1850 MachinePointerInfo::getConstantPool(),
1853 if (RelocM == Reloc::PIC_) {
1854 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1855 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1862 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1863 const ARMSubtarget *Subtarget) {
1864 DebugLoc dl = Op.getDebugLoc();
1865 SDValue Op5 = Op.getOperand(5);
1866 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1867 // Some subtargets which have dmb and dsb instructions can handle barriers
1868 // directly. Some ARMv6 cpus can support them with the help of mcr
1869 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1871 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1872 if (Subtarget->hasDataBarrier())
1873 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1875 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
1876 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1877 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1878 DAG.getConstant(0, MVT::i32));
1882 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1883 MachineFunction &MF = DAG.getMachineFunction();
1884 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1886 // vastart just stores the address of the VarArgsFrameIndex slot into the
1887 // memory location argument.
1888 DebugLoc dl = Op.getDebugLoc();
1889 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1890 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1891 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1892 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1897 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1898 SDValue &Root, SelectionDAG &DAG,
1899 DebugLoc dl) const {
1900 MachineFunction &MF = DAG.getMachineFunction();
1901 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1903 TargetRegisterClass *RC;
1904 if (AFI->isThumb1OnlyFunction())
1905 RC = ARM::tGPRRegisterClass;
1907 RC = ARM::GPRRegisterClass;
1909 // Transform the arguments stored in physical registers into virtual ones.
1910 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1911 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1914 if (NextVA.isMemLoc()) {
1915 MachineFrameInfo *MFI = MF.getFrameInfo();
1916 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
1918 // Create load node to retrieve arguments from the stack.
1919 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1920 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1921 MachinePointerInfo::getFixedStack(FI),
1924 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1925 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1928 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1932 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1933 CallingConv::ID CallConv, bool isVarArg,
1934 const SmallVectorImpl<ISD::InputArg>
1936 DebugLoc dl, SelectionDAG &DAG,
1937 SmallVectorImpl<SDValue> &InVals)
1940 MachineFunction &MF = DAG.getMachineFunction();
1941 MachineFrameInfo *MFI = MF.getFrameInfo();
1943 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1945 // Assign locations to all of the incoming arguments.
1946 SmallVector<CCValAssign, 16> ArgLocs;
1947 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1949 CCInfo.AnalyzeFormalArguments(Ins,
1950 CCAssignFnForNode(CallConv, /* Return*/ false,
1953 SmallVector<SDValue, 16> ArgValues;
1955 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1956 CCValAssign &VA = ArgLocs[i];
1958 // Arguments stored in registers.
1959 if (VA.isRegLoc()) {
1960 EVT RegVT = VA.getLocVT();
1963 if (VA.needsCustom()) {
1964 // f64 and vector types are split up into multiple registers or
1965 // combinations of registers and stack slots.
1966 if (VA.getLocVT() == MVT::v2f64) {
1967 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1969 VA = ArgLocs[++i]; // skip ahead to next loc
1971 if (VA.isMemLoc()) {
1972 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
1973 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1974 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1975 MachinePointerInfo::getFixedStack(FI),
1978 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1981 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1982 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1983 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1984 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1985 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1987 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1990 TargetRegisterClass *RC;
1992 if (RegVT == MVT::f32)
1993 RC = ARM::SPRRegisterClass;
1994 else if (RegVT == MVT::f64)
1995 RC = ARM::DPRRegisterClass;
1996 else if (RegVT == MVT::v2f64)
1997 RC = ARM::QPRRegisterClass;
1998 else if (RegVT == MVT::i32)
1999 RC = (AFI->isThumb1OnlyFunction() ?
2000 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2002 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2004 // Transform the arguments in physical registers into virtual ones.
2005 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2006 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2009 // If this is an 8 or 16-bit value, it is really passed promoted
2010 // to 32 bits. Insert an assert[sz]ext to capture this, then
2011 // truncate to the right size.
2012 switch (VA.getLocInfo()) {
2013 default: llvm_unreachable("Unknown loc info!");
2014 case CCValAssign::Full: break;
2015 case CCValAssign::BCvt:
2016 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2018 case CCValAssign::SExt:
2019 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2020 DAG.getValueType(VA.getValVT()));
2021 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2023 case CCValAssign::ZExt:
2024 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2025 DAG.getValueType(VA.getValVT()));
2026 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2030 InVals.push_back(ArgValue);
2032 } else { // VA.isRegLoc()
2035 assert(VA.isMemLoc());
2036 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2038 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2039 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2041 // Create load nodes to retrieve arguments from the stack.
2042 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2043 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2044 MachinePointerInfo::getFixedStack(FI),
2051 static const unsigned GPRArgRegs[] = {
2052 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2055 unsigned NumGPRs = CCInfo.getFirstUnallocated
2056 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2058 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2059 unsigned VARegSize = (4 - NumGPRs) * 4;
2060 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2061 unsigned ArgOffset = CCInfo.getNextStackOffset();
2062 if (VARegSaveSize) {
2063 // If this function is vararg, store any remaining integer argument regs
2064 // to their spots on the stack so that they may be loaded by deferencing
2065 // the result of va_next.
2066 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2067 AFI->setVarArgsFrameIndex(
2068 MFI->CreateFixedObject(VARegSaveSize,
2069 ArgOffset + VARegSaveSize - VARegSize,
2071 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2074 SmallVector<SDValue, 4> MemOps;
2075 for (; NumGPRs < 4; ++NumGPRs) {
2076 TargetRegisterClass *RC;
2077 if (AFI->isThumb1OnlyFunction())
2078 RC = ARM::tGPRRegisterClass;
2080 RC = ARM::GPRRegisterClass;
2082 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2083 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2085 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2086 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2087 0, false, false, 0);
2088 MemOps.push_back(Store);
2089 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2090 DAG.getConstant(4, getPointerTy()));
2092 if (!MemOps.empty())
2093 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2094 &MemOps[0], MemOps.size());
2096 // This will point to the next argument passed via stack.
2097 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2103 /// isFloatingPointZero - Return true if this is +0.0.
2104 static bool isFloatingPointZero(SDValue Op) {
2105 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2106 return CFP->getValueAPF().isPosZero();
2107 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2108 // Maybe this has already been legalized into the constant pool?
2109 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2110 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2111 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2112 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2113 return CFP->getValueAPF().isPosZero();
2119 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2120 /// the given operands.
2122 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2123 SDValue &ARMcc, SelectionDAG &DAG,
2124 DebugLoc dl) const {
2125 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2126 unsigned C = RHSC->getZExtValue();
2127 if (!isLegalICmpImmediate(C)) {
2128 // Constant does not fit, try adjusting it by one?
2133 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2134 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2135 RHS = DAG.getConstant(C-1, MVT::i32);
2140 if (C != 0 && isLegalICmpImmediate(C-1)) {
2141 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2142 RHS = DAG.getConstant(C-1, MVT::i32);
2147 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2148 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2149 RHS = DAG.getConstant(C+1, MVT::i32);
2154 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2155 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2156 RHS = DAG.getConstant(C+1, MVT::i32);
2163 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2164 ARMISD::NodeType CompareType;
2167 CompareType = ARMISD::CMP;
2172 CompareType = ARMISD::CMPZ;
2175 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2176 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2179 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2181 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2182 DebugLoc dl) const {
2184 if (!isFloatingPointZero(RHS))
2185 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2187 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2188 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2191 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2192 SDValue Cond = Op.getOperand(0);
2193 SDValue SelectTrue = Op.getOperand(1);
2194 SDValue SelectFalse = Op.getOperand(2);
2195 DebugLoc dl = Op.getDebugLoc();
2199 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2200 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2202 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2203 const ConstantSDNode *CMOVTrue =
2204 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2205 const ConstantSDNode *CMOVFalse =
2206 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2208 if (CMOVTrue && CMOVFalse) {
2209 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2210 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2214 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2216 False = SelectFalse;
2217 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2222 if (True.getNode() && False.getNode()) {
2223 EVT VT = Cond.getValueType();
2224 SDValue ARMcc = Cond.getOperand(2);
2225 SDValue CCR = Cond.getOperand(3);
2226 SDValue Cmp = Cond.getOperand(4);
2227 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2232 return DAG.getSelectCC(dl, Cond,
2233 DAG.getConstant(0, Cond.getValueType()),
2234 SelectTrue, SelectFalse, ISD::SETNE);
2237 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2238 EVT VT = Op.getValueType();
2239 SDValue LHS = Op.getOperand(0);
2240 SDValue RHS = Op.getOperand(1);
2241 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2242 SDValue TrueVal = Op.getOperand(2);
2243 SDValue FalseVal = Op.getOperand(3);
2244 DebugLoc dl = Op.getDebugLoc();
2246 if (LHS.getValueType() == MVT::i32) {
2248 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2249 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2250 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2253 ARMCC::CondCodes CondCode, CondCode2;
2254 FPCCToARMCC(CC, CondCode, CondCode2);
2256 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2257 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2258 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2259 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2261 if (CondCode2 != ARMCC::AL) {
2262 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2263 // FIXME: Needs another CMP because flag can have but one use.
2264 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2265 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2266 Result, TrueVal, ARMcc2, CCR, Cmp2);
2271 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2272 /// to morph to an integer compare sequence.
2273 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2274 const ARMSubtarget *Subtarget) {
2275 SDNode *N = Op.getNode();
2276 if (!N->hasOneUse())
2277 // Otherwise it requires moving the value from fp to integer registers.
2279 if (!N->getNumValues())
2281 EVT VT = Op.getValueType();
2282 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2283 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2284 // vmrs are very slow, e.g. cortex-a8.
2287 if (isFloatingPointZero(Op)) {
2291 return ISD::isNormalLoad(N);
2294 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2295 if (isFloatingPointZero(Op))
2296 return DAG.getConstant(0, MVT::i32);
2298 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2299 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2300 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2301 Ld->isVolatile(), Ld->isNonTemporal(),
2302 Ld->getAlignment());
2304 llvm_unreachable("Unknown VFP cmp argument!");
2307 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2308 SDValue &RetVal1, SDValue &RetVal2) {
2309 if (isFloatingPointZero(Op)) {
2310 RetVal1 = DAG.getConstant(0, MVT::i32);
2311 RetVal2 = DAG.getConstant(0, MVT::i32);
2315 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2316 SDValue Ptr = Ld->getBasePtr();
2317 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2318 Ld->getChain(), Ptr,
2319 Ld->getPointerInfo(),
2320 Ld->isVolatile(), Ld->isNonTemporal(),
2321 Ld->getAlignment());
2323 EVT PtrType = Ptr.getValueType();
2324 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2325 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2326 PtrType, Ptr, DAG.getConstant(4, PtrType));
2327 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2328 Ld->getChain(), NewPtr,
2329 Ld->getPointerInfo().getWithOffset(4),
2330 Ld->isVolatile(), Ld->isNonTemporal(),
2335 llvm_unreachable("Unknown VFP cmp argument!");
2338 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2339 /// f32 and even f64 comparisons to integer ones.
2341 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2342 SDValue Chain = Op.getOperand(0);
2343 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2344 SDValue LHS = Op.getOperand(2);
2345 SDValue RHS = Op.getOperand(3);
2346 SDValue Dest = Op.getOperand(4);
2347 DebugLoc dl = Op.getDebugLoc();
2349 bool SeenZero = false;
2350 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2351 canChangeToInt(RHS, SeenZero, Subtarget) &&
2352 // If one of the operand is zero, it's safe to ignore the NaN case since
2353 // we only care about equality comparisons.
2354 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2355 // If unsafe fp math optimization is enabled and there are no othter uses of
2356 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2357 // to an integer comparison.
2358 if (CC == ISD::SETOEQ)
2360 else if (CC == ISD::SETUNE)
2364 if (LHS.getValueType() == MVT::f32) {
2365 LHS = bitcastf32Toi32(LHS, DAG);
2366 RHS = bitcastf32Toi32(RHS, DAG);
2367 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2368 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2369 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2370 Chain, Dest, ARMcc, CCR, Cmp);
2375 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2376 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2377 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2378 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2379 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2380 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2381 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2387 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2388 SDValue Chain = Op.getOperand(0);
2389 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2390 SDValue LHS = Op.getOperand(2);
2391 SDValue RHS = Op.getOperand(3);
2392 SDValue Dest = Op.getOperand(4);
2393 DebugLoc dl = Op.getDebugLoc();
2395 if (LHS.getValueType() == MVT::i32) {
2397 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2398 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2399 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2400 Chain, Dest, ARMcc, CCR, Cmp);
2403 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2406 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2407 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2408 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2409 if (Result.getNode())
2413 ARMCC::CondCodes CondCode, CondCode2;
2414 FPCCToARMCC(CC, CondCode, CondCode2);
2416 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2417 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2418 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2419 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2420 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2421 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2422 if (CondCode2 != ARMCC::AL) {
2423 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2424 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2425 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2430 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2431 SDValue Chain = Op.getOperand(0);
2432 SDValue Table = Op.getOperand(1);
2433 SDValue Index = Op.getOperand(2);
2434 DebugLoc dl = Op.getDebugLoc();
2436 EVT PTy = getPointerTy();
2437 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2438 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2439 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2440 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2441 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2442 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2443 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2444 if (Subtarget->isThumb2()) {
2445 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2446 // which does another jump to the destination. This also makes it easier
2447 // to translate it to TBB / TBH later.
2448 // FIXME: This might not work if the function is extremely large.
2449 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2450 Addr, Op.getOperand(2), JTI, UId);
2452 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2453 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2454 MachinePointerInfo::getJumpTable(),
2456 Chain = Addr.getValue(1);
2457 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2458 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2460 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2461 MachinePointerInfo::getJumpTable(), false, false, 0);
2462 Chain = Addr.getValue(1);
2463 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2467 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2468 DebugLoc dl = Op.getDebugLoc();
2471 switch (Op.getOpcode()) {
2473 assert(0 && "Invalid opcode!");
2474 case ISD::FP_TO_SINT:
2475 Opc = ARMISD::FTOSI;
2477 case ISD::FP_TO_UINT:
2478 Opc = ARMISD::FTOUI;
2481 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2482 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2485 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2486 EVT VT = Op.getValueType();
2487 DebugLoc dl = Op.getDebugLoc();
2490 switch (Op.getOpcode()) {
2492 assert(0 && "Invalid opcode!");
2493 case ISD::SINT_TO_FP:
2494 Opc = ARMISD::SITOF;
2496 case ISD::UINT_TO_FP:
2497 Opc = ARMISD::UITOF;
2501 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2502 return DAG.getNode(Opc, dl, VT, Op);
2505 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2506 // Implement fcopysign with a fabs and a conditional fneg.
2507 SDValue Tmp0 = Op.getOperand(0);
2508 SDValue Tmp1 = Op.getOperand(1);
2509 DebugLoc dl = Op.getDebugLoc();
2510 EVT VT = Op.getValueType();
2511 EVT SrcVT = Tmp1.getValueType();
2512 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2513 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2514 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2515 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2516 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2517 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2520 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2521 MachineFunction &MF = DAG.getMachineFunction();
2522 MachineFrameInfo *MFI = MF.getFrameInfo();
2523 MFI->setReturnAddressIsTaken(true);
2525 EVT VT = Op.getValueType();
2526 DebugLoc dl = Op.getDebugLoc();
2527 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2529 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2530 SDValue Offset = DAG.getConstant(4, MVT::i32);
2531 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2532 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2533 MachinePointerInfo(), false, false, 0);
2536 // Return LR, which contains the return address. Mark it an implicit live-in.
2537 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2538 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2541 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2542 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2543 MFI->setFrameAddressIsTaken(true);
2545 EVT VT = Op.getValueType();
2546 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2547 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2548 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2549 ? ARM::R7 : ARM::R11;
2550 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2552 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2553 MachinePointerInfo(),
2558 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2559 /// expand a bit convert where either the source or destination type is i64 to
2560 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2561 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2562 /// vectors), since the legalizer won't know what to do with that.
2563 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2564 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2565 DebugLoc dl = N->getDebugLoc();
2566 SDValue Op = N->getOperand(0);
2568 // This function is only supposed to be called for i64 types, either as the
2569 // source or destination of the bit convert.
2570 EVT SrcVT = Op.getValueType();
2571 EVT DstVT = N->getValueType(0);
2572 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2573 "ExpandBIT_CONVERT called for non-i64 type");
2575 // Turn i64->f64 into VMOVDRR.
2576 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2577 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2578 DAG.getConstant(0, MVT::i32));
2579 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2580 DAG.getConstant(1, MVT::i32));
2581 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2582 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2585 // Turn f64->i64 into VMOVRRD.
2586 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2587 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2588 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2589 // Merge the pieces into a single i64 value.
2590 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2596 /// getZeroVector - Returns a vector of specified type with all zero elements.
2597 /// Zero vectors are used to represent vector negation and in those cases
2598 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2599 /// not support i64 elements, so sometimes the zero vectors will need to be
2600 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2602 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2603 assert(VT.isVector() && "Expected a vector type");
2604 // The canonical modified immediate encoding of a zero vector is....0!
2605 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2606 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2607 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2608 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2611 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2612 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2613 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2614 SelectionDAG &DAG) const {
2615 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2616 EVT VT = Op.getValueType();
2617 unsigned VTBits = VT.getSizeInBits();
2618 DebugLoc dl = Op.getDebugLoc();
2619 SDValue ShOpLo = Op.getOperand(0);
2620 SDValue ShOpHi = Op.getOperand(1);
2621 SDValue ShAmt = Op.getOperand(2);
2623 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2625 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2627 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2628 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2629 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2630 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2631 DAG.getConstant(VTBits, MVT::i32));
2632 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2633 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2634 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2636 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2637 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2639 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2640 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2643 SDValue Ops[2] = { Lo, Hi };
2644 return DAG.getMergeValues(Ops, 2, dl);
2647 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2648 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2649 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2650 SelectionDAG &DAG) const {
2651 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2652 EVT VT = Op.getValueType();
2653 unsigned VTBits = VT.getSizeInBits();
2654 DebugLoc dl = Op.getDebugLoc();
2655 SDValue ShOpLo = Op.getOperand(0);
2656 SDValue ShOpHi = Op.getOperand(1);
2657 SDValue ShAmt = Op.getOperand(2);
2660 assert(Op.getOpcode() == ISD::SHL_PARTS);
2661 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2662 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2663 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2664 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2665 DAG.getConstant(VTBits, MVT::i32));
2666 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2667 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2669 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2670 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2671 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2673 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2674 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2677 SDValue Ops[2] = { Lo, Hi };
2678 return DAG.getMergeValues(Ops, 2, dl);
2681 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2682 SelectionDAG &DAG) const {
2683 // The rounding mode is in bits 23:22 of the FPSCR.
2684 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2685 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2686 // so that the shift + and get folded into a bitfield extract.
2687 DebugLoc dl = Op.getDebugLoc();
2688 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2689 DAG.getConstant(Intrinsic::arm_get_fpscr,
2691 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2692 DAG.getConstant(1U << 22, MVT::i32));
2693 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2694 DAG.getConstant(22, MVT::i32));
2695 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2696 DAG.getConstant(3, MVT::i32));
2699 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2700 const ARMSubtarget *ST) {
2701 EVT VT = N->getValueType(0);
2702 DebugLoc dl = N->getDebugLoc();
2704 if (!ST->hasV6T2Ops())
2707 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2708 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2711 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2712 const ARMSubtarget *ST) {
2713 EVT VT = N->getValueType(0);
2714 DebugLoc dl = N->getDebugLoc();
2716 // Lower vector shifts on NEON to use VSHL.
2717 if (VT.isVector()) {
2718 assert(ST->hasNEON() && "unexpected vector shift");
2720 // Left shifts translate directly to the vshiftu intrinsic.
2721 if (N->getOpcode() == ISD::SHL)
2722 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2723 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2724 N->getOperand(0), N->getOperand(1));
2726 assert((N->getOpcode() == ISD::SRA ||
2727 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2729 // NEON uses the same intrinsics for both left and right shifts. For
2730 // right shifts, the shift amounts are negative, so negate the vector of
2732 EVT ShiftVT = N->getOperand(1).getValueType();
2733 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2734 getZeroVector(ShiftVT, DAG, dl),
2736 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2737 Intrinsic::arm_neon_vshifts :
2738 Intrinsic::arm_neon_vshiftu);
2739 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2740 DAG.getConstant(vshiftInt, MVT::i32),
2741 N->getOperand(0), NegatedCount);
2744 // We can get here for a node like i32 = ISD::SHL i32, i64
2748 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2749 "Unknown shift to lower!");
2751 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2752 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2753 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2756 // If we are in thumb mode, we don't have RRX.
2757 if (ST->isThumb1Only()) return SDValue();
2759 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2760 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2761 DAG.getConstant(0, MVT::i32));
2762 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2763 DAG.getConstant(1, MVT::i32));
2765 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2766 // captures the result into a carry flag.
2767 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2768 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2770 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2771 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2773 // Merge the pieces into a single i64 value.
2774 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2777 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2778 SDValue TmpOp0, TmpOp1;
2779 bool Invert = false;
2783 SDValue Op0 = Op.getOperand(0);
2784 SDValue Op1 = Op.getOperand(1);
2785 SDValue CC = Op.getOperand(2);
2786 EVT VT = Op.getValueType();
2787 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2788 DebugLoc dl = Op.getDebugLoc();
2790 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2791 switch (SetCCOpcode) {
2792 default: llvm_unreachable("Illegal FP comparison"); break;
2794 case ISD::SETNE: Invert = true; // Fallthrough
2796 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2798 case ISD::SETLT: Swap = true; // Fallthrough
2800 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2802 case ISD::SETLE: Swap = true; // Fallthrough
2804 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2805 case ISD::SETUGE: Swap = true; // Fallthrough
2806 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2807 case ISD::SETUGT: Swap = true; // Fallthrough
2808 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2809 case ISD::SETUEQ: Invert = true; // Fallthrough
2811 // Expand this to (OLT | OGT).
2815 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2816 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2818 case ISD::SETUO: Invert = true; // Fallthrough
2820 // Expand this to (OLT | OGE).
2824 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2825 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2829 // Integer comparisons.
2830 switch (SetCCOpcode) {
2831 default: llvm_unreachable("Illegal integer comparison"); break;
2832 case ISD::SETNE: Invert = true;
2833 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2834 case ISD::SETLT: Swap = true;
2835 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2836 case ISD::SETLE: Swap = true;
2837 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2838 case ISD::SETULT: Swap = true;
2839 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2840 case ISD::SETULE: Swap = true;
2841 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2844 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2845 if (Opc == ARMISD::VCEQ) {
2848 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2850 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2853 // Ignore bitconvert.
2854 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2855 AndOp = AndOp.getOperand(0);
2857 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2859 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2860 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2867 std::swap(Op0, Op1);
2869 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2872 Result = DAG.getNOT(dl, Result, VT);
2877 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2878 /// valid vector constant for a NEON instruction with a "modified immediate"
2879 /// operand (e.g., VMOV). If so, return the encoded value.
2880 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2881 unsigned SplatBitSize, SelectionDAG &DAG,
2882 EVT &VT, bool is128Bits, bool isVMOV) {
2883 unsigned OpCmode, Imm;
2885 // SplatBitSize is set to the smallest size that splats the vector, so a
2886 // zero vector will always have SplatBitSize == 8. However, NEON modified
2887 // immediate instructions others than VMOV do not support the 8-bit encoding
2888 // of a zero vector, and the default encoding of zero is supposed to be the
2893 switch (SplatBitSize) {
2897 // Any 1-byte value is OK. Op=0, Cmode=1110.
2898 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2901 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2905 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2906 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2907 if ((SplatBits & ~0xff) == 0) {
2908 // Value = 0x00nn: Op=x, Cmode=100x.
2913 if ((SplatBits & ~0xff00) == 0) {
2914 // Value = 0xnn00: Op=x, Cmode=101x.
2916 Imm = SplatBits >> 8;
2922 // NEON's 32-bit VMOV supports splat values where:
2923 // * only one byte is nonzero, or
2924 // * the least significant byte is 0xff and the second byte is nonzero, or
2925 // * the least significant 2 bytes are 0xff and the third is nonzero.
2926 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
2927 if ((SplatBits & ~0xff) == 0) {
2928 // Value = 0x000000nn: Op=x, Cmode=000x.
2933 if ((SplatBits & ~0xff00) == 0) {
2934 // Value = 0x0000nn00: Op=x, Cmode=001x.
2936 Imm = SplatBits >> 8;
2939 if ((SplatBits & ~0xff0000) == 0) {
2940 // Value = 0x00nn0000: Op=x, Cmode=010x.
2942 Imm = SplatBits >> 16;
2945 if ((SplatBits & ~0xff000000) == 0) {
2946 // Value = 0xnn000000: Op=x, Cmode=011x.
2948 Imm = SplatBits >> 24;
2952 if ((SplatBits & ~0xffff) == 0 &&
2953 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2954 // Value = 0x0000nnff: Op=x, Cmode=1100.
2956 Imm = SplatBits >> 8;
2961 if ((SplatBits & ~0xffffff) == 0 &&
2962 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2963 // Value = 0x00nnffff: Op=x, Cmode=1101.
2965 Imm = SplatBits >> 16;
2966 SplatBits |= 0xffff;
2970 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2971 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2972 // VMOV.I32. A (very) minor optimization would be to replicate the value
2973 // and fall through here to test for a valid 64-bit splat. But, then the
2974 // caller would also need to check and handle the change in size.
2980 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2981 uint64_t BitMask = 0xff;
2983 unsigned ImmMask = 1;
2985 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2986 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2989 } else if ((SplatBits & BitMask) != 0) {
2995 // Op=1, Cmode=1110.
2998 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3003 llvm_unreachable("unexpected size for isNEONModifiedImm");
3007 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3008 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3011 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3012 bool &ReverseVEXT, unsigned &Imm) {
3013 unsigned NumElts = VT.getVectorNumElements();
3014 ReverseVEXT = false;
3016 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3022 // If this is a VEXT shuffle, the immediate value is the index of the first
3023 // element. The other shuffle indices must be the successive elements after
3025 unsigned ExpectedElt = Imm;
3026 for (unsigned i = 1; i < NumElts; ++i) {
3027 // Increment the expected index. If it wraps around, it may still be
3028 // a VEXT but the source vectors must be swapped.
3030 if (ExpectedElt == NumElts * 2) {
3035 if (M[i] < 0) continue; // ignore UNDEF indices
3036 if (ExpectedElt != static_cast<unsigned>(M[i]))
3040 // Adjust the index value if the source operands will be swapped.
3047 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3048 /// instruction with the specified blocksize. (The order of the elements
3049 /// within each block of the vector is reversed.)
3050 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3051 unsigned BlockSize) {
3052 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3053 "Only possible block sizes for VREV are: 16, 32, 64");
3055 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3059 unsigned NumElts = VT.getVectorNumElements();
3060 unsigned BlockElts = M[0] + 1;
3061 // If the first shuffle index is UNDEF, be optimistic.
3063 BlockElts = BlockSize / EltSz;
3065 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3068 for (unsigned i = 0; i < NumElts; ++i) {
3069 if (M[i] < 0) continue; // ignore UNDEF indices
3070 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3077 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3078 unsigned &WhichResult) {
3079 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3083 unsigned NumElts = VT.getVectorNumElements();
3084 WhichResult = (M[0] == 0 ? 0 : 1);
3085 for (unsigned i = 0; i < NumElts; i += 2) {
3086 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3087 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3093 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3094 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3095 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3096 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3097 unsigned &WhichResult) {
3098 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3102 unsigned NumElts = VT.getVectorNumElements();
3103 WhichResult = (M[0] == 0 ? 0 : 1);
3104 for (unsigned i = 0; i < NumElts; i += 2) {
3105 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3106 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3112 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3113 unsigned &WhichResult) {
3114 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3118 unsigned NumElts = VT.getVectorNumElements();
3119 WhichResult = (M[0] == 0 ? 0 : 1);
3120 for (unsigned i = 0; i != NumElts; ++i) {
3121 if (M[i] < 0) continue; // ignore UNDEF indices
3122 if ((unsigned) M[i] != 2 * i + WhichResult)
3126 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3127 if (VT.is64BitVector() && EltSz == 32)
3133 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3134 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3135 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3136 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3137 unsigned &WhichResult) {
3138 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3142 unsigned Half = VT.getVectorNumElements() / 2;
3143 WhichResult = (M[0] == 0 ? 0 : 1);
3144 for (unsigned j = 0; j != 2; ++j) {
3145 unsigned Idx = WhichResult;
3146 for (unsigned i = 0; i != Half; ++i) {
3147 int MIdx = M[i + j * Half];
3148 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3154 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3155 if (VT.is64BitVector() && EltSz == 32)
3161 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3162 unsigned &WhichResult) {
3163 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3167 unsigned NumElts = VT.getVectorNumElements();
3168 WhichResult = (M[0] == 0 ? 0 : 1);
3169 unsigned Idx = WhichResult * NumElts / 2;
3170 for (unsigned i = 0; i != NumElts; i += 2) {
3171 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3172 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3177 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3178 if (VT.is64BitVector() && EltSz == 32)
3184 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3185 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3186 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3187 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3188 unsigned &WhichResult) {
3189 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3193 unsigned NumElts = VT.getVectorNumElements();
3194 WhichResult = (M[0] == 0 ? 0 : 1);
3195 unsigned Idx = WhichResult * NumElts / 2;
3196 for (unsigned i = 0; i != NumElts; i += 2) {
3197 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3198 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3203 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3204 if (VT.is64BitVector() && EltSz == 32)
3210 // If N is an integer constant that can be moved into a register in one
3211 // instruction, return an SDValue of such a constant (will become a MOV
3212 // instruction). Otherwise return null.
3213 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3214 const ARMSubtarget *ST, DebugLoc dl) {
3216 if (!isa<ConstantSDNode>(N))
3218 Val = cast<ConstantSDNode>(N)->getZExtValue();
3220 if (ST->isThumb1Only()) {
3221 if (Val <= 255 || ~Val <= 255)
3222 return DAG.getConstant(Val, MVT::i32);
3224 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3225 return DAG.getConstant(Val, MVT::i32);
3230 // If this is a case we can't handle, return null and let the default
3231 // expansion code take care of it.
3232 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3233 const ARMSubtarget *ST) {
3234 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3235 DebugLoc dl = Op.getDebugLoc();
3236 EVT VT = Op.getValueType();
3238 APInt SplatBits, SplatUndef;
3239 unsigned SplatBitSize;
3241 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3242 if (SplatBitSize <= 64) {
3243 // Check if an immediate VMOV works.
3245 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3246 SplatUndef.getZExtValue(), SplatBitSize,
3247 DAG, VmovVT, VT.is128BitVector(), true);
3248 if (Val.getNode()) {
3249 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3250 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3253 // Try an immediate VMVN.
3254 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3255 ((1LL << SplatBitSize) - 1));
3256 Val = isNEONModifiedImm(NegatedImm,
3257 SplatUndef.getZExtValue(), SplatBitSize,
3258 DAG, VmovVT, VT.is128BitVector(), false);
3259 if (Val.getNode()) {
3260 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3261 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3266 // Scan through the operands to see if only one value is used.
3267 unsigned NumElts = VT.getVectorNumElements();
3268 bool isOnlyLowElement = true;
3269 bool usesOnlyOneValue = true;
3270 bool isConstant = true;
3272 for (unsigned i = 0; i < NumElts; ++i) {
3273 SDValue V = Op.getOperand(i);
3274 if (V.getOpcode() == ISD::UNDEF)
3277 isOnlyLowElement = false;
3278 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3281 if (!Value.getNode())
3283 else if (V != Value)
3284 usesOnlyOneValue = false;
3287 if (!Value.getNode())
3288 return DAG.getUNDEF(VT);
3290 if (isOnlyLowElement)
3291 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3293 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3295 if (EnableARMVDUPsplat) {
3296 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3297 // i32 and try again.
3298 if (usesOnlyOneValue && EltSize <= 32) {
3300 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3301 if (VT.getVectorElementType().isFloatingPoint()) {
3302 SmallVector<SDValue, 8> Ops;
3303 for (unsigned i = 0; i < NumElts; ++i)
3304 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3306 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3308 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3309 LowerBUILD_VECTOR(Val, DAG, ST));
3311 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3313 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3317 // If all elements are constants and the case above didn't get hit, fall back
3318 // to the default expansion, which will generate a load from the constant
3323 if (!EnableARMVDUPsplat) {
3324 // Use VDUP for non-constant splats.
3325 if (usesOnlyOneValue && EltSize <= 32)
3326 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3329 // Vectors with 32- or 64-bit elements can be built by directly assigning
3330 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3331 // will be legalized.
3332 if (EltSize >= 32) {
3333 // Do the expansion with floating-point types, since that is what the VFP
3334 // registers are defined to use, and since i64 is not legal.
3335 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3336 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3337 SmallVector<SDValue, 8> Ops;
3338 for (unsigned i = 0; i < NumElts; ++i)
3339 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3340 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3341 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3347 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3348 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3349 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3350 /// are assumed to be legal.
3352 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3354 if (VT.getVectorNumElements() == 4 &&
3355 (VT.is128BitVector() || VT.is64BitVector())) {
3356 unsigned PFIndexes[4];
3357 for (unsigned i = 0; i != 4; ++i) {
3361 PFIndexes[i] = M[i];
3364 // Compute the index in the perfect shuffle table.
3365 unsigned PFTableIndex =
3366 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3367 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3368 unsigned Cost = (PFEntry >> 30);
3375 unsigned Imm, WhichResult;
3377 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3378 return (EltSize >= 32 ||
3379 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3380 isVREVMask(M, VT, 64) ||
3381 isVREVMask(M, VT, 32) ||
3382 isVREVMask(M, VT, 16) ||
3383 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3384 isVTRNMask(M, VT, WhichResult) ||
3385 isVUZPMask(M, VT, WhichResult) ||
3386 isVZIPMask(M, VT, WhichResult) ||
3387 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3388 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3389 isVZIP_v_undef_Mask(M, VT, WhichResult));
3392 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3393 /// the specified operations to build the shuffle.
3394 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3395 SDValue RHS, SelectionDAG &DAG,
3397 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3398 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3399 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3402 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3411 OP_VUZPL, // VUZP, left result
3412 OP_VUZPR, // VUZP, right result
3413 OP_VZIPL, // VZIP, left result
3414 OP_VZIPR, // VZIP, right result
3415 OP_VTRNL, // VTRN, left result
3416 OP_VTRNR // VTRN, right result
3419 if (OpNum == OP_COPY) {
3420 if (LHSID == (1*9+2)*9+3) return LHS;
3421 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3425 SDValue OpLHS, OpRHS;
3426 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3427 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3428 EVT VT = OpLHS.getValueType();
3431 default: llvm_unreachable("Unknown shuffle opcode!");
3433 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3438 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3439 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3443 return DAG.getNode(ARMISD::VEXT, dl, VT,
3445 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3448 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3449 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3452 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3453 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3456 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3457 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3461 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3462 SDValue V1 = Op.getOperand(0);
3463 SDValue V2 = Op.getOperand(1);
3464 DebugLoc dl = Op.getDebugLoc();
3465 EVT VT = Op.getValueType();
3466 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3467 SmallVector<int, 8> ShuffleMask;
3469 // Convert shuffles that are directly supported on NEON to target-specific
3470 // DAG nodes, instead of keeping them as shuffles and matching them again
3471 // during code selection. This is more efficient and avoids the possibility
3472 // of inconsistencies between legalization and selection.
3473 // FIXME: floating-point vectors should be canonicalized to integer vectors
3474 // of the same time so that they get CSEd properly.
3475 SVN->getMask(ShuffleMask);
3477 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3478 if (EltSize <= 32) {
3479 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3480 int Lane = SVN->getSplatIndex();
3481 // If this is undef splat, generate it via "just" vdup, if possible.
3482 if (Lane == -1) Lane = 0;
3484 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3485 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3487 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3488 DAG.getConstant(Lane, MVT::i32));
3493 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3496 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3497 DAG.getConstant(Imm, MVT::i32));
3500 if (isVREVMask(ShuffleMask, VT, 64))
3501 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3502 if (isVREVMask(ShuffleMask, VT, 32))
3503 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3504 if (isVREVMask(ShuffleMask, VT, 16))
3505 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3507 // Check for Neon shuffles that modify both input vectors in place.
3508 // If both results are used, i.e., if there are two shuffles with the same
3509 // source operands and with masks corresponding to both results of one of
3510 // these operations, DAG memoization will ensure that a single node is
3511 // used for both shuffles.
3512 unsigned WhichResult;
3513 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3514 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3515 V1, V2).getValue(WhichResult);
3516 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3517 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3518 V1, V2).getValue(WhichResult);
3519 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3520 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3521 V1, V2).getValue(WhichResult);
3523 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3524 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3525 V1, V1).getValue(WhichResult);
3526 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3527 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3528 V1, V1).getValue(WhichResult);
3529 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3530 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3531 V1, V1).getValue(WhichResult);
3534 // If the shuffle is not directly supported and it has 4 elements, use
3535 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3536 unsigned NumElts = VT.getVectorNumElements();
3538 unsigned PFIndexes[4];
3539 for (unsigned i = 0; i != 4; ++i) {
3540 if (ShuffleMask[i] < 0)
3543 PFIndexes[i] = ShuffleMask[i];
3546 // Compute the index in the perfect shuffle table.
3547 unsigned PFTableIndex =
3548 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3549 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3550 unsigned Cost = (PFEntry >> 30);
3553 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3556 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3557 if (EltSize >= 32) {
3558 // Do the expansion with floating-point types, since that is what the VFP
3559 // registers are defined to use, and since i64 is not legal.
3560 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3561 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3562 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3563 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3564 SmallVector<SDValue, 8> Ops;
3565 for (unsigned i = 0; i < NumElts; ++i) {
3566 if (ShuffleMask[i] < 0)
3567 Ops.push_back(DAG.getUNDEF(EltVT));
3569 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3570 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3571 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3574 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3575 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3581 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3582 EVT VT = Op.getValueType();
3583 DebugLoc dl = Op.getDebugLoc();
3584 SDValue Vec = Op.getOperand(0);
3585 SDValue Lane = Op.getOperand(1);
3586 assert(VT == MVT::i32 &&
3587 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3588 "unexpected type for custom-lowering vector extract");
3589 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3592 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3593 // The only time a CONCAT_VECTORS operation can have legal types is when
3594 // two 64-bit vectors are concatenated to a 128-bit vector.
3595 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3596 "unexpected CONCAT_VECTORS");
3597 DebugLoc dl = Op.getDebugLoc();
3598 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3599 SDValue Op0 = Op.getOperand(0);
3600 SDValue Op1 = Op.getOperand(1);
3601 if (Op0.getOpcode() != ISD::UNDEF)
3602 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3603 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3604 DAG.getIntPtrConstant(0));
3605 if (Op1.getOpcode() != ISD::UNDEF)
3606 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3607 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3608 DAG.getIntPtrConstant(1));
3609 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3612 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3613 /// an extending load, return the unextended value.
3614 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3615 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3616 return N->getOperand(0);
3617 LoadSDNode *LD = cast<LoadSDNode>(N);
3618 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3619 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
3620 LD->isNonTemporal(), LD->getAlignment());
3623 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3624 // Multiplications are only custom-lowered for 128-bit vectors so that
3625 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3626 EVT VT = Op.getValueType();
3627 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3628 SDNode *N0 = Op.getOperand(0).getNode();
3629 SDNode *N1 = Op.getOperand(1).getNode();
3630 unsigned NewOpc = 0;
3631 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3632 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3633 NewOpc = ARMISD::VMULLs;
3634 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3635 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3636 NewOpc = ARMISD::VMULLu;
3637 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3638 // Fall through to expand this. It is not legal.
3641 // Other vector multiplications are legal.
3645 // Legalize to a VMULL instruction.
3646 DebugLoc DL = Op.getDebugLoc();
3647 SDValue Op0 = SkipExtension(N0, DAG);
3648 SDValue Op1 = SkipExtension(N1, DAG);
3650 assert(Op0.getValueType().is64BitVector() &&
3651 Op1.getValueType().is64BitVector() &&
3652 "unexpected types for extended operands to VMULL");
3653 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3656 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3657 switch (Op.getOpcode()) {
3658 default: llvm_unreachable("Don't know how to custom lower this!");
3659 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3660 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3661 case ISD::GlobalAddress:
3662 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3663 LowerGlobalAddressELF(Op, DAG);
3664 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3665 case ISD::SELECT: return LowerSELECT(Op, DAG);
3666 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3667 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3668 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3669 case ISD::VASTART: return LowerVASTART(Op, DAG);
3670 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3671 case ISD::SINT_TO_FP:
3672 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3673 case ISD::FP_TO_SINT:
3674 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3675 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3676 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3677 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3678 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3679 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3680 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3681 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3683 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3686 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3687 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3688 case ISD::SRL_PARTS:
3689 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3690 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3691 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3692 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3693 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3694 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3695 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3696 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3697 case ISD::MUL: return LowerMUL(Op, DAG);
3702 /// ReplaceNodeResults - Replace the results of node with an illegal result
3703 /// type with new values built out of custom code.
3704 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3705 SmallVectorImpl<SDValue>&Results,
3706 SelectionDAG &DAG) const {
3708 switch (N->getOpcode()) {
3710 llvm_unreachable("Don't know how to custom expand this!");
3712 case ISD::BIT_CONVERT:
3713 Res = ExpandBIT_CONVERT(N, DAG);
3717 Res = LowerShift(N, DAG, Subtarget);
3721 Results.push_back(Res);
3724 //===----------------------------------------------------------------------===//
3725 // ARM Scheduler Hooks
3726 //===----------------------------------------------------------------------===//
3729 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3730 MachineBasicBlock *BB,
3731 unsigned Size) const {
3732 unsigned dest = MI->getOperand(0).getReg();
3733 unsigned ptr = MI->getOperand(1).getReg();
3734 unsigned oldval = MI->getOperand(2).getReg();
3735 unsigned newval = MI->getOperand(3).getReg();
3736 unsigned scratch = BB->getParent()->getRegInfo()
3737 .createVirtualRegister(ARM::GPRRegisterClass);
3738 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3739 DebugLoc dl = MI->getDebugLoc();
3740 bool isThumb2 = Subtarget->isThumb2();
3742 unsigned ldrOpc, strOpc;
3744 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3746 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3747 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3750 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3751 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3754 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3755 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3759 MachineFunction *MF = BB->getParent();
3760 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3761 MachineFunction::iterator It = BB;
3762 ++It; // insert the new blocks after the current block
3764 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3765 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3766 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3767 MF->insert(It, loop1MBB);
3768 MF->insert(It, loop2MBB);
3769 MF->insert(It, exitMBB);
3771 // Transfer the remainder of BB and its successor edges to exitMBB.
3772 exitMBB->splice(exitMBB->begin(), BB,
3773 llvm::next(MachineBasicBlock::iterator(MI)),
3775 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3779 // fallthrough --> loop1MBB
3780 BB->addSuccessor(loop1MBB);
3783 // ldrex dest, [ptr]
3787 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3788 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3789 .addReg(dest).addReg(oldval));
3790 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3791 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3792 BB->addSuccessor(loop2MBB);
3793 BB->addSuccessor(exitMBB);
3796 // strex scratch, newval, [ptr]
3800 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3802 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3803 .addReg(scratch).addImm(0));
3804 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3805 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3806 BB->addSuccessor(loop1MBB);
3807 BB->addSuccessor(exitMBB);
3813 MI->eraseFromParent(); // The instruction is gone now.
3819 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3820 unsigned Size, unsigned BinOpcode) const {
3821 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3822 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3824 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3825 MachineFunction *MF = BB->getParent();
3826 MachineFunction::iterator It = BB;
3829 unsigned dest = MI->getOperand(0).getReg();
3830 unsigned ptr = MI->getOperand(1).getReg();
3831 unsigned incr = MI->getOperand(2).getReg();
3832 DebugLoc dl = MI->getDebugLoc();
3834 bool isThumb2 = Subtarget->isThumb2();
3835 unsigned ldrOpc, strOpc;
3837 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3839 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3840 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3843 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3844 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3847 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3848 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3852 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3853 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3854 MF->insert(It, loopMBB);
3855 MF->insert(It, exitMBB);
3857 // Transfer the remainder of BB and its successor edges to exitMBB.
3858 exitMBB->splice(exitMBB->begin(), BB,
3859 llvm::next(MachineBasicBlock::iterator(MI)),
3861 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3863 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3864 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3865 unsigned scratch2 = (!BinOpcode) ? incr :
3866 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3870 // fallthrough --> loopMBB
3871 BB->addSuccessor(loopMBB);
3875 // <binop> scratch2, dest, incr
3876 // strex scratch, scratch2, ptr
3879 // fallthrough --> exitMBB
3881 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3883 // operand order needs to go the other way for NAND
3884 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3885 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3886 addReg(incr).addReg(dest)).addReg(0);
3888 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3889 addReg(dest).addReg(incr)).addReg(0);
3892 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3894 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3895 .addReg(scratch).addImm(0));
3896 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3897 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3899 BB->addSuccessor(loopMBB);
3900 BB->addSuccessor(exitMBB);
3906 MI->eraseFromParent(); // The instruction is gone now.
3912 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3913 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3914 E = MBB->succ_end(); I != E; ++I)
3917 llvm_unreachable("Expecting a BB with two successors!");
3921 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3922 MachineBasicBlock *BB) const {
3923 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3924 DebugLoc dl = MI->getDebugLoc();
3925 bool isThumb2 = Subtarget->isThumb2();
3926 switch (MI->getOpcode()) {
3929 llvm_unreachable("Unexpected instr type to insert");
3931 case ARM::ATOMIC_LOAD_ADD_I8:
3932 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3933 case ARM::ATOMIC_LOAD_ADD_I16:
3934 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3935 case ARM::ATOMIC_LOAD_ADD_I32:
3936 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3938 case ARM::ATOMIC_LOAD_AND_I8:
3939 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3940 case ARM::ATOMIC_LOAD_AND_I16:
3941 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3942 case ARM::ATOMIC_LOAD_AND_I32:
3943 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3945 case ARM::ATOMIC_LOAD_OR_I8:
3946 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3947 case ARM::ATOMIC_LOAD_OR_I16:
3948 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3949 case ARM::ATOMIC_LOAD_OR_I32:
3950 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3952 case ARM::ATOMIC_LOAD_XOR_I8:
3953 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3954 case ARM::ATOMIC_LOAD_XOR_I16:
3955 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3956 case ARM::ATOMIC_LOAD_XOR_I32:
3957 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3959 case ARM::ATOMIC_LOAD_NAND_I8:
3960 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3961 case ARM::ATOMIC_LOAD_NAND_I16:
3962 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3963 case ARM::ATOMIC_LOAD_NAND_I32:
3964 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3966 case ARM::ATOMIC_LOAD_SUB_I8:
3967 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3968 case ARM::ATOMIC_LOAD_SUB_I16:
3969 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3970 case ARM::ATOMIC_LOAD_SUB_I32:
3971 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3973 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3974 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3975 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3977 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3978 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3979 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3981 case ARM::tMOVCCr_pseudo: {
3982 // To "insert" a SELECT_CC instruction, we actually have to insert the
3983 // diamond control-flow pattern. The incoming instruction knows the
3984 // destination vreg to set, the condition code register to branch on, the
3985 // true/false values to select between, and a branch opcode to use.
3986 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3987 MachineFunction::iterator It = BB;
3993 // cmpTY ccX, r1, r2
3995 // fallthrough --> copy0MBB
3996 MachineBasicBlock *thisMBB = BB;
3997 MachineFunction *F = BB->getParent();
3998 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3999 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4000 F->insert(It, copy0MBB);
4001 F->insert(It, sinkMBB);
4003 // Transfer the remainder of BB and its successor edges to sinkMBB.
4004 sinkMBB->splice(sinkMBB->begin(), BB,
4005 llvm::next(MachineBasicBlock::iterator(MI)),
4007 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4009 BB->addSuccessor(copy0MBB);
4010 BB->addSuccessor(sinkMBB);
4012 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4013 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4016 // %FalseValue = ...
4017 // # fallthrough to sinkMBB
4020 // Update machine-CFG edges
4021 BB->addSuccessor(sinkMBB);
4024 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4027 BuildMI(*BB, BB->begin(), dl,
4028 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4029 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4030 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4032 MI->eraseFromParent(); // The pseudo instruction is gone now.
4037 case ARM::BCCZi64: {
4038 // Compare both parts that make up the double comparison separately for
4040 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4042 unsigned LHS1 = MI->getOperand(1).getReg();
4043 unsigned LHS2 = MI->getOperand(2).getReg();
4045 AddDefaultPred(BuildMI(BB, dl,
4046 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4047 .addReg(LHS1).addImm(0));
4048 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4049 .addReg(LHS2).addImm(0)
4050 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4052 unsigned RHS1 = MI->getOperand(3).getReg();
4053 unsigned RHS2 = MI->getOperand(4).getReg();
4054 AddDefaultPred(BuildMI(BB, dl,
4055 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4056 .addReg(LHS1).addReg(RHS1));
4057 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4058 .addReg(LHS2).addReg(RHS2)
4059 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4062 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4063 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4064 if (MI->getOperand(0).getImm() == ARMCC::NE)
4065 std::swap(destMBB, exitMBB);
4067 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4068 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4069 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4072 MI->eraseFromParent(); // The pseudo instruction is gone now.
4078 //===----------------------------------------------------------------------===//
4079 // ARM Optimization Hooks
4080 //===----------------------------------------------------------------------===//
4083 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4084 TargetLowering::DAGCombinerInfo &DCI) {
4085 SelectionDAG &DAG = DCI.DAG;
4086 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4087 EVT VT = N->getValueType(0);
4088 unsigned Opc = N->getOpcode();
4089 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4090 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4091 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4092 ISD::CondCode CC = ISD::SETCC_INVALID;
4095 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4097 SDValue CCOp = Slct.getOperand(0);
4098 if (CCOp.getOpcode() == ISD::SETCC)
4099 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4102 bool DoXform = false;
4104 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4107 if (LHS.getOpcode() == ISD::Constant &&
4108 cast<ConstantSDNode>(LHS)->isNullValue()) {
4110 } else if (CC != ISD::SETCC_INVALID &&
4111 RHS.getOpcode() == ISD::Constant &&
4112 cast<ConstantSDNode>(RHS)->isNullValue()) {
4113 std::swap(LHS, RHS);
4114 SDValue Op0 = Slct.getOperand(0);
4115 EVT OpVT = isSlctCC ? Op0.getValueType() :
4116 Op0.getOperand(0).getValueType();
4117 bool isInt = OpVT.isInteger();
4118 CC = ISD::getSetCCInverse(CC, isInt);
4120 if (!TLI.isCondCodeLegal(CC, OpVT))
4121 return SDValue(); // Inverse operator isn't legal.
4128 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4130 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4131 Slct.getOperand(0), Slct.getOperand(1), CC);
4132 SDValue CCOp = Slct.getOperand(0);
4134 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4135 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4136 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4137 CCOp, OtherOp, Result);
4142 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4143 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4144 /// called with the default operands, and if that fails, with commuted
4146 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4147 TargetLowering::DAGCombinerInfo &DCI) {
4148 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4149 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4150 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4151 if (Result.getNode()) return Result;
4156 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4158 static SDValue PerformADDCombine(SDNode *N,
4159 TargetLowering::DAGCombinerInfo &DCI) {
4160 SDValue N0 = N->getOperand(0);
4161 SDValue N1 = N->getOperand(1);
4163 // First try with the default operand order.
4164 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4165 if (Result.getNode())
4168 // If that didn't work, try again with the operands commuted.
4169 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4172 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4174 static SDValue PerformSUBCombine(SDNode *N,
4175 TargetLowering::DAGCombinerInfo &DCI) {
4176 SDValue N0 = N->getOperand(0);
4177 SDValue N1 = N->getOperand(1);
4179 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4180 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4181 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4182 if (Result.getNode()) return Result;
4188 static SDValue PerformMULCombine(SDNode *N,
4189 TargetLowering::DAGCombinerInfo &DCI,
4190 const ARMSubtarget *Subtarget) {
4191 SelectionDAG &DAG = DCI.DAG;
4193 if (Subtarget->isThumb1Only())
4196 if (DAG.getMachineFunction().
4197 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4200 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4203 EVT VT = N->getValueType(0);
4207 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4211 uint64_t MulAmt = C->getZExtValue();
4212 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4213 ShiftAmt = ShiftAmt & (32 - 1);
4214 SDValue V = N->getOperand(0);
4215 DebugLoc DL = N->getDebugLoc();
4218 MulAmt >>= ShiftAmt;
4219 if (isPowerOf2_32(MulAmt - 1)) {
4220 // (mul x, 2^N + 1) => (add (shl x, N), x)
4221 Res = DAG.getNode(ISD::ADD, DL, VT,
4222 V, DAG.getNode(ISD::SHL, DL, VT,
4223 V, DAG.getConstant(Log2_32(MulAmt-1),
4225 } else if (isPowerOf2_32(MulAmt + 1)) {
4226 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4227 Res = DAG.getNode(ISD::SUB, DL, VT,
4228 DAG.getNode(ISD::SHL, DL, VT,
4229 V, DAG.getConstant(Log2_32(MulAmt+1),
4236 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4237 DAG.getConstant(ShiftAmt, MVT::i32));
4239 // Do not add new nodes to DAG combiner worklist.
4240 DCI.CombineTo(N, Res, false);
4244 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4245 static SDValue PerformORCombine(SDNode *N,
4246 TargetLowering::DAGCombinerInfo &DCI,
4247 const ARMSubtarget *Subtarget) {
4248 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4251 // BFI is only available on V6T2+
4252 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4255 SelectionDAG &DAG = DCI.DAG;
4256 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4257 DebugLoc DL = N->getDebugLoc();
4258 // 1) or (and A, mask), val => ARMbfi A, val, mask
4259 // iff (val & mask) == val
4261 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4262 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4263 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4264 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4265 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4266 // (i.e., copy a bitfield value into another bitfield of the same width)
4267 if (N0.getOpcode() != ISD::AND)
4270 EVT VT = N->getValueType(0);
4275 // The value and the mask need to be constants so we can verify this is
4276 // actually a bitfield set. If the mask is 0xffff, we can do better
4277 // via a movt instruction, so don't use BFI in that case.
4278 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4281 unsigned Mask = C->getZExtValue();
4285 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4286 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4287 unsigned Val = C->getZExtValue();
4288 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4290 Val >>= CountTrailingZeros_32(~Mask);
4292 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4293 DAG.getConstant(Val, MVT::i32),
4294 DAG.getConstant(Mask, MVT::i32));
4296 // Do not add new nodes to DAG combiner worklist.
4297 DCI.CombineTo(N, Res, false);
4298 } else if (N1.getOpcode() == ISD::AND) {
4299 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4300 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4303 unsigned Mask2 = C->getZExtValue();
4305 if (ARM::isBitFieldInvertedMask(Mask) &&
4306 ARM::isBitFieldInvertedMask(~Mask2) &&
4307 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4308 // The pack halfword instruction works better for masks that fit it,
4309 // so use that when it's available.
4310 if (Subtarget->hasT2ExtractPack() &&
4311 (Mask == 0xffff || Mask == 0xffff0000))
4314 unsigned lsb = CountTrailingZeros_32(Mask2);
4315 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4316 DAG.getConstant(lsb, MVT::i32));
4317 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4318 DAG.getConstant(Mask, MVT::i32));
4319 // Do not add new nodes to DAG combiner worklist.
4320 DCI.CombineTo(N, Res, false);
4321 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4322 ARM::isBitFieldInvertedMask(Mask2) &&
4323 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4324 // The pack halfword instruction works better for masks that fit it,
4325 // so use that when it's available.
4326 if (Subtarget->hasT2ExtractPack() &&
4327 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4330 unsigned lsb = CountTrailingZeros_32(Mask);
4331 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4332 DAG.getConstant(lsb, MVT::i32));
4333 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4334 DAG.getConstant(Mask2, MVT::i32));
4335 // Do not add new nodes to DAG combiner worklist.
4336 DCI.CombineTo(N, Res, false);
4343 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4344 /// ISD::BUILD_VECTOR.
4345 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4346 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4347 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4348 // into a pair of GPRs, which is fine when the value is used as a scalar,
4349 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4350 if (N->getNumOperands() == 2) {
4351 SDValue Op0 = N->getOperand(0);
4352 SDValue Op1 = N->getOperand(1);
4353 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4354 Op0 = Op0.getOperand(0);
4355 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4356 Op1 = Op1.getOperand(0);
4357 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4358 Op0.getNode() == Op1.getNode() &&
4359 Op0.getResNo() == 0 && Op1.getResNo() == 1) {
4360 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4361 N->getValueType(0), Op0.getOperand(0));
4368 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4369 /// ARMISD::VMOVRRD.
4370 static SDValue PerformVMOVRRDCombine(SDNode *N,
4371 TargetLowering::DAGCombinerInfo &DCI) {
4372 // fmrrd(fmdrr x, y) -> x,y
4373 SDValue InDouble = N->getOperand(0);
4374 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4375 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4379 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4380 /// ARMISD::VDUPLANE.
4381 static SDValue PerformVDUPLANECombine(SDNode *N,
4382 TargetLowering::DAGCombinerInfo &DCI) {
4383 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4385 SDValue Op = N->getOperand(0);
4386 EVT VT = N->getValueType(0);
4388 // Ignore bit_converts.
4389 while (Op.getOpcode() == ISD::BIT_CONVERT)
4390 Op = Op.getOperand(0);
4391 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4394 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4395 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4396 // The canonical VMOV for a zero vector uses a 32-bit element size.
4397 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4399 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4401 if (EltSize > VT.getVectorElementType().getSizeInBits())
4404 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4405 return DCI.CombineTo(N, Res, false);
4408 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4409 /// operand of a vector shift operation, where all the elements of the
4410 /// build_vector must have the same constant integer value.
4411 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4412 // Ignore bit_converts.
4413 while (Op.getOpcode() == ISD::BIT_CONVERT)
4414 Op = Op.getOperand(0);
4415 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4416 APInt SplatBits, SplatUndef;
4417 unsigned SplatBitSize;
4419 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4420 HasAnyUndefs, ElementBits) ||
4421 SplatBitSize > ElementBits)
4423 Cnt = SplatBits.getSExtValue();
4427 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4428 /// operand of a vector shift left operation. That value must be in the range:
4429 /// 0 <= Value < ElementBits for a left shift; or
4430 /// 0 <= Value <= ElementBits for a long left shift.
4431 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4432 assert(VT.isVector() && "vector shift count is not a vector type");
4433 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4434 if (! getVShiftImm(Op, ElementBits, Cnt))
4436 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4439 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4440 /// operand of a vector shift right operation. For a shift opcode, the value
4441 /// is positive, but for an intrinsic the value count must be negative. The
4442 /// absolute value must be in the range:
4443 /// 1 <= |Value| <= ElementBits for a right shift; or
4444 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4445 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4447 assert(VT.isVector() && "vector shift count is not a vector type");
4448 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4449 if (! getVShiftImm(Op, ElementBits, Cnt))
4453 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4456 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4457 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4458 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4461 // Don't do anything for most intrinsics.
4464 // Vector shifts: check for immediate versions and lower them.
4465 // Note: This is done during DAG combining instead of DAG legalizing because
4466 // the build_vectors for 64-bit vector element shift counts are generally
4467 // not legal, and it is hard to see their values after they get legalized to
4468 // loads from a constant pool.
4469 case Intrinsic::arm_neon_vshifts:
4470 case Intrinsic::arm_neon_vshiftu:
4471 case Intrinsic::arm_neon_vshiftls:
4472 case Intrinsic::arm_neon_vshiftlu:
4473 case Intrinsic::arm_neon_vshiftn:
4474 case Intrinsic::arm_neon_vrshifts:
4475 case Intrinsic::arm_neon_vrshiftu:
4476 case Intrinsic::arm_neon_vrshiftn:
4477 case Intrinsic::arm_neon_vqshifts:
4478 case Intrinsic::arm_neon_vqshiftu:
4479 case Intrinsic::arm_neon_vqshiftsu:
4480 case Intrinsic::arm_neon_vqshiftns:
4481 case Intrinsic::arm_neon_vqshiftnu:
4482 case Intrinsic::arm_neon_vqshiftnsu:
4483 case Intrinsic::arm_neon_vqrshiftns:
4484 case Intrinsic::arm_neon_vqrshiftnu:
4485 case Intrinsic::arm_neon_vqrshiftnsu: {
4486 EVT VT = N->getOperand(1).getValueType();
4488 unsigned VShiftOpc = 0;
4491 case Intrinsic::arm_neon_vshifts:
4492 case Intrinsic::arm_neon_vshiftu:
4493 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4494 VShiftOpc = ARMISD::VSHL;
4497 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4498 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4499 ARMISD::VSHRs : ARMISD::VSHRu);
4504 case Intrinsic::arm_neon_vshiftls:
4505 case Intrinsic::arm_neon_vshiftlu:
4506 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4508 llvm_unreachable("invalid shift count for vshll intrinsic");
4510 case Intrinsic::arm_neon_vrshifts:
4511 case Intrinsic::arm_neon_vrshiftu:
4512 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4516 case Intrinsic::arm_neon_vqshifts:
4517 case Intrinsic::arm_neon_vqshiftu:
4518 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4522 case Intrinsic::arm_neon_vqshiftsu:
4523 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4525 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4527 case Intrinsic::arm_neon_vshiftn:
4528 case Intrinsic::arm_neon_vrshiftn:
4529 case Intrinsic::arm_neon_vqshiftns:
4530 case Intrinsic::arm_neon_vqshiftnu:
4531 case Intrinsic::arm_neon_vqshiftnsu:
4532 case Intrinsic::arm_neon_vqrshiftns:
4533 case Intrinsic::arm_neon_vqrshiftnu:
4534 case Intrinsic::arm_neon_vqrshiftnsu:
4535 // Narrowing shifts require an immediate right shift.
4536 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4538 llvm_unreachable("invalid shift count for narrowing vector shift "
4542 llvm_unreachable("unhandled vector shift");
4546 case Intrinsic::arm_neon_vshifts:
4547 case Intrinsic::arm_neon_vshiftu:
4548 // Opcode already set above.
4550 case Intrinsic::arm_neon_vshiftls:
4551 case Intrinsic::arm_neon_vshiftlu:
4552 if (Cnt == VT.getVectorElementType().getSizeInBits())
4553 VShiftOpc = ARMISD::VSHLLi;
4555 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4556 ARMISD::VSHLLs : ARMISD::VSHLLu);
4558 case Intrinsic::arm_neon_vshiftn:
4559 VShiftOpc = ARMISD::VSHRN; break;
4560 case Intrinsic::arm_neon_vrshifts:
4561 VShiftOpc = ARMISD::VRSHRs; break;
4562 case Intrinsic::arm_neon_vrshiftu:
4563 VShiftOpc = ARMISD::VRSHRu; break;
4564 case Intrinsic::arm_neon_vrshiftn:
4565 VShiftOpc = ARMISD::VRSHRN; break;
4566 case Intrinsic::arm_neon_vqshifts:
4567 VShiftOpc = ARMISD::VQSHLs; break;
4568 case Intrinsic::arm_neon_vqshiftu:
4569 VShiftOpc = ARMISD::VQSHLu; break;
4570 case Intrinsic::arm_neon_vqshiftsu:
4571 VShiftOpc = ARMISD::VQSHLsu; break;
4572 case Intrinsic::arm_neon_vqshiftns:
4573 VShiftOpc = ARMISD::VQSHRNs; break;
4574 case Intrinsic::arm_neon_vqshiftnu:
4575 VShiftOpc = ARMISD::VQSHRNu; break;
4576 case Intrinsic::arm_neon_vqshiftnsu:
4577 VShiftOpc = ARMISD::VQSHRNsu; break;
4578 case Intrinsic::arm_neon_vqrshiftns:
4579 VShiftOpc = ARMISD::VQRSHRNs; break;
4580 case Intrinsic::arm_neon_vqrshiftnu:
4581 VShiftOpc = ARMISD::VQRSHRNu; break;
4582 case Intrinsic::arm_neon_vqrshiftnsu:
4583 VShiftOpc = ARMISD::VQRSHRNsu; break;
4586 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4587 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4590 case Intrinsic::arm_neon_vshiftins: {
4591 EVT VT = N->getOperand(1).getValueType();
4593 unsigned VShiftOpc = 0;
4595 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4596 VShiftOpc = ARMISD::VSLI;
4597 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4598 VShiftOpc = ARMISD::VSRI;
4600 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4603 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4604 N->getOperand(1), N->getOperand(2),
4605 DAG.getConstant(Cnt, MVT::i32));
4608 case Intrinsic::arm_neon_vqrshifts:
4609 case Intrinsic::arm_neon_vqrshiftu:
4610 // No immediate versions of these to check for.
4617 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4618 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4619 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4620 /// vector element shift counts are generally not legal, and it is hard to see
4621 /// their values after they get legalized to loads from a constant pool.
4622 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4623 const ARMSubtarget *ST) {
4624 EVT VT = N->getValueType(0);
4626 // Nothing to be done for scalar shifts.
4627 if (! VT.isVector())
4630 assert(ST->hasNEON() && "unexpected vector shift");
4633 switch (N->getOpcode()) {
4634 default: llvm_unreachable("unexpected shift opcode");
4637 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4638 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4639 DAG.getConstant(Cnt, MVT::i32));
4644 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4645 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4646 ARMISD::VSHRs : ARMISD::VSHRu);
4647 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4648 DAG.getConstant(Cnt, MVT::i32));
4654 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4655 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4656 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4657 const ARMSubtarget *ST) {
4658 SDValue N0 = N->getOperand(0);
4660 // Check for sign- and zero-extensions of vector extract operations of 8-
4661 // and 16-bit vector elements. NEON supports these directly. They are
4662 // handled during DAG combining because type legalization will promote them
4663 // to 32-bit types and it is messy to recognize the operations after that.
4664 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4665 SDValue Vec = N0.getOperand(0);
4666 SDValue Lane = N0.getOperand(1);
4667 EVT VT = N->getValueType(0);
4668 EVT EltVT = N0.getValueType();
4669 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4671 if (VT == MVT::i32 &&
4672 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4673 TLI.isTypeLegal(Vec.getValueType())) {
4676 switch (N->getOpcode()) {
4677 default: llvm_unreachable("unexpected opcode");
4678 case ISD::SIGN_EXTEND:
4679 Opc = ARMISD::VGETLANEs;
4681 case ISD::ZERO_EXTEND:
4682 case ISD::ANY_EXTEND:
4683 Opc = ARMISD::VGETLANEu;
4686 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4693 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4694 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4695 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4696 const ARMSubtarget *ST) {
4697 // If the target supports NEON, try to use vmax/vmin instructions for f32
4698 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4699 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4700 // a NaN; only do the transformation when it matches that behavior.
4702 // For now only do this when using NEON for FP operations; if using VFP, it
4703 // is not obvious that the benefit outweighs the cost of switching to the
4705 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4706 N->getValueType(0) != MVT::f32)
4709 SDValue CondLHS = N->getOperand(0);
4710 SDValue CondRHS = N->getOperand(1);
4711 SDValue LHS = N->getOperand(2);
4712 SDValue RHS = N->getOperand(3);
4713 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4715 unsigned Opcode = 0;
4717 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4718 IsReversed = false; // x CC y ? x : y
4719 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4720 IsReversed = true ; // x CC y ? y : x
4734 // If LHS is NaN, an ordered comparison will be false and the result will
4735 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4736 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4737 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4738 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4740 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4741 // will return -0, so vmin can only be used for unsafe math or if one of
4742 // the operands is known to be nonzero.
4743 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4745 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4747 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4756 // If LHS is NaN, an ordered comparison will be false and the result will
4757 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4758 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4759 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4760 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4762 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4763 // will return +0, so vmax can only be used for unsafe math or if one of
4764 // the operands is known to be nonzero.
4765 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4767 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4769 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4775 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4778 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4779 DAGCombinerInfo &DCI) const {
4780 switch (N->getOpcode()) {
4782 case ISD::ADD: return PerformADDCombine(N, DCI);
4783 case ISD::SUB: return PerformSUBCombine(N, DCI);
4784 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4785 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4786 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
4787 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4788 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4789 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4792 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4793 case ISD::SIGN_EXTEND:
4794 case ISD::ZERO_EXTEND:
4795 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4796 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4801 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4802 if (!Subtarget->hasV6Ops())
4803 // Pre-v6 does not support unaligned mem access.
4806 // v6+ may or may not support unaligned mem access depending on the system
4808 // FIXME: This is pretty conservative. Should we provide cmdline option to
4809 // control the behaviour?
4810 if (!Subtarget->isTargetDarwin())
4813 switch (VT.getSimpleVT().SimpleTy) {
4820 // FIXME: VLD1 etc with standard alignment is legal.
4824 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4829 switch (VT.getSimpleVT().SimpleTy) {
4830 default: return false;
4845 if ((V & (Scale - 1)) != 0)
4848 return V == (V & ((1LL << 5) - 1));
4851 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4852 const ARMSubtarget *Subtarget) {
4859 switch (VT.getSimpleVT().SimpleTy) {
4860 default: return false;
4865 // + imm12 or - imm8
4867 return V == (V & ((1LL << 8) - 1));
4868 return V == (V & ((1LL << 12) - 1));
4871 // Same as ARM mode. FIXME: NEON?
4872 if (!Subtarget->hasVFP2())
4877 return V == (V & ((1LL << 8) - 1));
4881 /// isLegalAddressImmediate - Return true if the integer value can be used
4882 /// as the offset of the target addressing mode for load / store of the
4884 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4885 const ARMSubtarget *Subtarget) {
4892 if (Subtarget->isThumb1Only())
4893 return isLegalT1AddressImmediate(V, VT);
4894 else if (Subtarget->isThumb2())
4895 return isLegalT2AddressImmediate(V, VT, Subtarget);
4900 switch (VT.getSimpleVT().SimpleTy) {
4901 default: return false;
4906 return V == (V & ((1LL << 12) - 1));
4909 return V == (V & ((1LL << 8) - 1));
4912 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4917 return V == (V & ((1LL << 8) - 1));
4921 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4923 int Scale = AM.Scale;
4927 switch (VT.getSimpleVT().SimpleTy) {
4928 default: return false;
4937 return Scale == 2 || Scale == 4 || Scale == 8;
4940 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4944 // Note, we allow "void" uses (basically, uses that aren't loads or
4945 // stores), because arm allows folding a scale into many arithmetic
4946 // operations. This should be made more precise and revisited later.
4948 // Allow r << imm, but the imm has to be a multiple of two.
4949 if (Scale & 1) return false;
4950 return isPowerOf2_32(Scale);
4954 /// isLegalAddressingMode - Return true if the addressing mode represented
4955 /// by AM is legal for this target, for a load/store of the specified type.
4956 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4957 const Type *Ty) const {
4958 EVT VT = getValueType(Ty, true);
4959 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4962 // Can never fold addr of global into load/store.
4967 case 0: // no scale reg, must be "r+i" or "r", or "i".
4970 if (Subtarget->isThumb1Only())
4974 // ARM doesn't support any R+R*scale+imm addr modes.
4981 if (Subtarget->isThumb2())
4982 return isLegalT2ScaledAddressingMode(AM, VT);
4984 int Scale = AM.Scale;
4985 switch (VT.getSimpleVT().SimpleTy) {
4986 default: return false;
4990 if (Scale < 0) Scale = -Scale;
4994 return isPowerOf2_32(Scale & ~1);
4998 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5003 // Note, we allow "void" uses (basically, uses that aren't loads or
5004 // stores), because arm allows folding a scale into many arithmetic
5005 // operations. This should be made more precise and revisited later.
5007 // Allow r << imm, but the imm has to be a multiple of two.
5008 if (Scale & 1) return false;
5009 return isPowerOf2_32(Scale);
5016 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5017 /// icmp immediate, that is the target has icmp instructions which can compare
5018 /// a register against the immediate without having to materialize the
5019 /// immediate into a register.
5020 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5021 if (!Subtarget->isThumb())
5022 return ARM_AM::getSOImmVal(Imm) != -1;
5023 if (Subtarget->isThumb2())
5024 return ARM_AM::getT2SOImmVal(Imm) != -1;
5025 return Imm >= 0 && Imm <= 255;
5028 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5029 bool isSEXTLoad, SDValue &Base,
5030 SDValue &Offset, bool &isInc,
5031 SelectionDAG &DAG) {
5032 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5035 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5037 Base = Ptr->getOperand(0);
5038 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5039 int RHSC = (int)RHS->getZExtValue();
5040 if (RHSC < 0 && RHSC > -256) {
5041 assert(Ptr->getOpcode() == ISD::ADD);
5043 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5047 isInc = (Ptr->getOpcode() == ISD::ADD);
5048 Offset = Ptr->getOperand(1);
5050 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5052 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5053 int RHSC = (int)RHS->getZExtValue();
5054 if (RHSC < 0 && RHSC > -0x1000) {
5055 assert(Ptr->getOpcode() == ISD::ADD);
5057 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5058 Base = Ptr->getOperand(0);
5063 if (Ptr->getOpcode() == ISD::ADD) {
5065 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5066 if (ShOpcVal != ARM_AM::no_shift) {
5067 Base = Ptr->getOperand(1);
5068 Offset = Ptr->getOperand(0);
5070 Base = Ptr->getOperand(0);
5071 Offset = Ptr->getOperand(1);
5076 isInc = (Ptr->getOpcode() == ISD::ADD);
5077 Base = Ptr->getOperand(0);
5078 Offset = Ptr->getOperand(1);
5082 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5086 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5087 bool isSEXTLoad, SDValue &Base,
5088 SDValue &Offset, bool &isInc,
5089 SelectionDAG &DAG) {
5090 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5093 Base = Ptr->getOperand(0);
5094 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5095 int RHSC = (int)RHS->getZExtValue();
5096 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5097 assert(Ptr->getOpcode() == ISD::ADD);
5099 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5101 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5102 isInc = Ptr->getOpcode() == ISD::ADD;
5103 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5111 /// getPreIndexedAddressParts - returns true by value, base pointer and
5112 /// offset pointer and addressing mode by reference if the node's address
5113 /// can be legally represented as pre-indexed load / store address.
5115 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5117 ISD::MemIndexedMode &AM,
5118 SelectionDAG &DAG) const {
5119 if (Subtarget->isThumb1Only())
5124 bool isSEXTLoad = false;
5125 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5126 Ptr = LD->getBasePtr();
5127 VT = LD->getMemoryVT();
5128 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5129 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5130 Ptr = ST->getBasePtr();
5131 VT = ST->getMemoryVT();
5136 bool isLegal = false;
5137 if (Subtarget->isThumb2())
5138 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5139 Offset, isInc, DAG);
5141 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5142 Offset, isInc, DAG);
5146 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5150 /// getPostIndexedAddressParts - returns true by value, base pointer and
5151 /// offset pointer and addressing mode by reference if this node can be
5152 /// combined with a load / store to form a post-indexed load / store.
5153 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5156 ISD::MemIndexedMode &AM,
5157 SelectionDAG &DAG) const {
5158 if (Subtarget->isThumb1Only())
5163 bool isSEXTLoad = false;
5164 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5165 VT = LD->getMemoryVT();
5166 Ptr = LD->getBasePtr();
5167 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5168 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5169 VT = ST->getMemoryVT();
5170 Ptr = ST->getBasePtr();
5175 bool isLegal = false;
5176 if (Subtarget->isThumb2())
5177 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5180 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5186 // Swap base ptr and offset to catch more post-index load / store when
5187 // it's legal. In Thumb2 mode, offset must be an immediate.
5188 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5189 !Subtarget->isThumb2())
5190 std::swap(Base, Offset);
5192 // Post-indexed load / store update the base pointer.
5197 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5201 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5205 const SelectionDAG &DAG,
5206 unsigned Depth) const {
5207 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5208 switch (Op.getOpcode()) {
5210 case ARMISD::CMOV: {
5211 // Bits are known zero/one if known on the LHS and RHS.
5212 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5213 if (KnownZero == 0 && KnownOne == 0) return;
5215 APInt KnownZeroRHS, KnownOneRHS;
5216 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5217 KnownZeroRHS, KnownOneRHS, Depth+1);
5218 KnownZero &= KnownZeroRHS;
5219 KnownOne &= KnownOneRHS;
5225 //===----------------------------------------------------------------------===//
5226 // ARM Inline Assembly Support
5227 //===----------------------------------------------------------------------===//
5229 /// getConstraintType - Given a constraint letter, return the type of
5230 /// constraint it is for this target.
5231 ARMTargetLowering::ConstraintType
5232 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5233 if (Constraint.size() == 1) {
5234 switch (Constraint[0]) {
5236 case 'l': return C_RegisterClass;
5237 case 'w': return C_RegisterClass;
5240 return TargetLowering::getConstraintType(Constraint);
5243 std::pair<unsigned, const TargetRegisterClass*>
5244 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5246 if (Constraint.size() == 1) {
5247 // GCC ARM Constraint Letters
5248 switch (Constraint[0]) {
5250 if (Subtarget->isThumb())
5251 return std::make_pair(0U, ARM::tGPRRegisterClass);
5253 return std::make_pair(0U, ARM::GPRRegisterClass);
5255 return std::make_pair(0U, ARM::GPRRegisterClass);
5258 return std::make_pair(0U, ARM::SPRRegisterClass);
5259 if (VT.getSizeInBits() == 64)
5260 return std::make_pair(0U, ARM::DPRRegisterClass);
5261 if (VT.getSizeInBits() == 128)
5262 return std::make_pair(0U, ARM::QPRRegisterClass);
5266 if (StringRef("{cc}").equals_lower(Constraint))
5267 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5269 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5272 std::vector<unsigned> ARMTargetLowering::
5273 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5275 if (Constraint.size() != 1)
5276 return std::vector<unsigned>();
5278 switch (Constraint[0]) { // GCC ARM Constraint Letters
5281 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5282 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5285 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5286 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5287 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5288 ARM::R12, ARM::LR, 0);
5291 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5292 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5293 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5294 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5295 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5296 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5297 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5298 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5299 if (VT.getSizeInBits() == 64)
5300 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5301 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5302 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5303 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5304 if (VT.getSizeInBits() == 128)
5305 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5306 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5310 return std::vector<unsigned>();
5313 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5314 /// vector. If it is invalid, don't add anything to Ops.
5315 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5317 std::vector<SDValue>&Ops,
5318 SelectionDAG &DAG) const {
5319 SDValue Result(0, 0);
5321 switch (Constraint) {
5323 case 'I': case 'J': case 'K': case 'L':
5324 case 'M': case 'N': case 'O':
5325 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5329 int64_t CVal64 = C->getSExtValue();
5330 int CVal = (int) CVal64;
5331 // None of these constraints allow values larger than 32 bits. Check
5332 // that the value fits in an int.
5336 switch (Constraint) {
5338 if (Subtarget->isThumb1Only()) {
5339 // This must be a constant between 0 and 255, for ADD
5341 if (CVal >= 0 && CVal <= 255)
5343 } else if (Subtarget->isThumb2()) {
5344 // A constant that can be used as an immediate value in a
5345 // data-processing instruction.
5346 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5349 // A constant that can be used as an immediate value in a
5350 // data-processing instruction.
5351 if (ARM_AM::getSOImmVal(CVal) != -1)
5357 if (Subtarget->isThumb()) { // FIXME thumb2
5358 // This must be a constant between -255 and -1, for negated ADD
5359 // immediates. This can be used in GCC with an "n" modifier that
5360 // prints the negated value, for use with SUB instructions. It is
5361 // not useful otherwise but is implemented for compatibility.
5362 if (CVal >= -255 && CVal <= -1)
5365 // This must be a constant between -4095 and 4095. It is not clear
5366 // what this constraint is intended for. Implemented for
5367 // compatibility with GCC.
5368 if (CVal >= -4095 && CVal <= 4095)
5374 if (Subtarget->isThumb1Only()) {
5375 // A 32-bit value where only one byte has a nonzero value. Exclude
5376 // zero to match GCC. This constraint is used by GCC internally for
5377 // constants that can be loaded with a move/shift combination.
5378 // It is not useful otherwise but is implemented for compatibility.
5379 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5381 } else if (Subtarget->isThumb2()) {
5382 // A constant whose bitwise inverse can be used as an immediate
5383 // value in a data-processing instruction. This can be used in GCC
5384 // with a "B" modifier that prints the inverted value, for use with
5385 // BIC and MVN instructions. It is not useful otherwise but is
5386 // implemented for compatibility.
5387 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5390 // A constant whose bitwise inverse can be used as an immediate
5391 // value in a data-processing instruction. This can be used in GCC
5392 // with a "B" modifier that prints the inverted value, for use with
5393 // BIC and MVN instructions. It is not useful otherwise but is
5394 // implemented for compatibility.
5395 if (ARM_AM::getSOImmVal(~CVal) != -1)
5401 if (Subtarget->isThumb1Only()) {
5402 // This must be a constant between -7 and 7,
5403 // for 3-operand ADD/SUB immediate instructions.
5404 if (CVal >= -7 && CVal < 7)
5406 } else if (Subtarget->isThumb2()) {
5407 // A constant whose negation can be used as an immediate value in a
5408 // data-processing instruction. This can be used in GCC with an "n"
5409 // modifier that prints the negated value, for use with SUB
5410 // instructions. It is not useful otherwise but is implemented for
5412 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5415 // A constant whose negation can be used as an immediate value in a
5416 // data-processing instruction. This can be used in GCC with an "n"
5417 // modifier that prints the negated value, for use with SUB
5418 // instructions. It is not useful otherwise but is implemented for
5420 if (ARM_AM::getSOImmVal(-CVal) != -1)
5426 if (Subtarget->isThumb()) { // FIXME thumb2
5427 // This must be a multiple of 4 between 0 and 1020, for
5428 // ADD sp + immediate.
5429 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5432 // A power of two or a constant between 0 and 32. This is used in
5433 // GCC for the shift amount on shifted register operands, but it is
5434 // useful in general for any shift amounts.
5435 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5441 if (Subtarget->isThumb()) { // FIXME thumb2
5442 // This must be a constant between 0 and 31, for shift amounts.
5443 if (CVal >= 0 && CVal <= 31)
5449 if (Subtarget->isThumb()) { // FIXME thumb2
5450 // This must be a multiple of 4 between -508 and 508, for
5451 // ADD/SUB sp = sp + immediate.
5452 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5457 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5461 if (Result.getNode()) {
5462 Ops.push_back(Result);
5465 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5469 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5470 // The ARM target isn't yet aware of offsets.
5474 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5475 APInt Imm = FPImm.bitcastToAPInt();
5476 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5477 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5478 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5480 // We can handle 4 bits of mantissa.
5481 // mantissa = (16+UInt(e:f:g:h))/16.
5482 if (Mantissa & 0x7ffff)
5485 if ((Mantissa & 0xf) != Mantissa)
5488 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5489 if (Exp < -3 || Exp > 4)
5491 Exp = ((Exp+3) & 0x7) ^ 4;
5493 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5496 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5497 APInt Imm = FPImm.bitcastToAPInt();
5498 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5499 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5500 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5502 // We can handle 4 bits of mantissa.
5503 // mantissa = (16+UInt(e:f:g:h))/16.
5504 if (Mantissa & 0xffffffffffffLL)
5507 if ((Mantissa & 0xf) != Mantissa)
5510 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5511 if (Exp < -3 || Exp > 4)
5513 Exp = ((Exp+3) & 0x7) ^ 4;
5515 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5518 bool ARM::isBitFieldInvertedMask(unsigned v) {
5519 if (v == 0xffffffff)
5521 // there can be 1's on either or both "outsides", all the "inside"
5523 unsigned int lsb = 0, msb = 31;
5524 while (v & (1 << msb)) --msb;
5525 while (v & (1 << lsb)) ++lsb;
5526 for (unsigned int i = lsb; i <= msb; ++i) {
5533 /// isFPImmLegal - Returns true if the target can instruction select the
5534 /// specified FP immediate natively. If false, the legalizer will
5535 /// materialize the FP immediate as a load from a constant pool.
5536 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5537 if (!Subtarget->hasVFP3())
5540 return ARM::getVFPf32Imm(Imm) != -1;
5542 return ARM::getVFPf64Imm(Imm) != -1;