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,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getConstantPool(), 0,
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 PseudoSourceValue::getGOT(), 0,
1733 // If we have T2 ops, we can materialize the address directly via movt/movw
1734 // pair. This is always cheaper.
1735 if (Subtarget->useMovt()) {
1736 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1737 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1739 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1740 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1741 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1742 PseudoSourceValue::getConstantPool(), 0,
1748 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1749 SelectionDAG &DAG) const {
1750 MachineFunction &MF = DAG.getMachineFunction();
1751 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1752 unsigned ARMPCLabelIndex = 0;
1753 EVT PtrVT = getPointerTy();
1754 DebugLoc dl = Op.getDebugLoc();
1755 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1756 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1758 if (RelocM == Reloc::Static)
1759 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1761 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1762 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1763 ARMConstantPoolValue *CPV =
1764 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1765 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1767 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1769 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1770 PseudoSourceValue::getConstantPool(), 0,
1772 SDValue Chain = Result.getValue(1);
1774 if (RelocM == Reloc::PIC_) {
1775 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1776 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1779 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1780 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1781 PseudoSourceValue::getGOT(), 0,
1787 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1788 SelectionDAG &DAG) const {
1789 assert(Subtarget->isTargetELF() &&
1790 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1791 MachineFunction &MF = DAG.getMachineFunction();
1792 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1793 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1794 EVT PtrVT = getPointerTy();
1795 DebugLoc dl = Op.getDebugLoc();
1796 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1797 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1798 "_GLOBAL_OFFSET_TABLE_",
1799 ARMPCLabelIndex, PCAdj);
1800 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1801 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1802 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1803 PseudoSourceValue::getConstantPool(), 0,
1805 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1806 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1810 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1811 DebugLoc dl = Op.getDebugLoc();
1812 SDValue Val = DAG.getConstant(0, MVT::i32);
1813 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1814 Op.getOperand(1), Val);
1818 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1819 DebugLoc dl = Op.getDebugLoc();
1820 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1821 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1825 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1826 const ARMSubtarget *Subtarget) const {
1827 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1828 DebugLoc dl = Op.getDebugLoc();
1830 default: return SDValue(); // Don't custom lower most intrinsics.
1831 case Intrinsic::arm_thread_pointer: {
1832 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1833 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1835 case Intrinsic::eh_sjlj_lsda: {
1836 MachineFunction &MF = DAG.getMachineFunction();
1837 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1838 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1839 EVT PtrVT = getPointerTy();
1840 DebugLoc dl = Op.getDebugLoc();
1841 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1843 unsigned PCAdj = (RelocM != Reloc::PIC_)
1844 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1845 ARMConstantPoolValue *CPV =
1846 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1847 ARMCP::CPLSDA, PCAdj);
1848 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1849 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1851 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1852 PseudoSourceValue::getConstantPool(), 0,
1855 if (RelocM == Reloc::PIC_) {
1856 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1857 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1864 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1865 const ARMSubtarget *Subtarget) {
1866 DebugLoc dl = Op.getDebugLoc();
1867 SDValue Op5 = Op.getOperand(5);
1868 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1869 // Some subtargets which have dmb and dsb instructions can handle barriers
1870 // directly. Some ARMv6 cpus can support them with the help of mcr
1871 // instruction. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1873 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1874 if (Subtarget->hasDataBarrier())
1875 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1877 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb1Only() &&
1878 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1879 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1880 DAG.getConstant(0, MVT::i32));
1884 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1885 MachineFunction &MF = DAG.getMachineFunction();
1886 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1888 // vastart just stores the address of the VarArgsFrameIndex slot into the
1889 // memory location argument.
1890 DebugLoc dl = Op.getDebugLoc();
1891 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1892 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1893 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1894 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1899 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1900 SDValue &Root, SelectionDAG &DAG,
1901 DebugLoc dl) const {
1902 MachineFunction &MF = DAG.getMachineFunction();
1903 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1905 TargetRegisterClass *RC;
1906 if (AFI->isThumb1OnlyFunction())
1907 RC = ARM::tGPRRegisterClass;
1909 RC = ARM::GPRRegisterClass;
1911 // Transform the arguments stored in physical registers into virtual ones.
1912 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1913 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1916 if (NextVA.isMemLoc()) {
1917 MachineFrameInfo *MFI = MF.getFrameInfo();
1918 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
1920 // Create load node to retrieve arguments from the stack.
1921 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1922 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1923 PseudoSourceValue::getFixedStack(FI), 0,
1926 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1927 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1930 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1934 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1935 CallingConv::ID CallConv, bool isVarArg,
1936 const SmallVectorImpl<ISD::InputArg>
1938 DebugLoc dl, SelectionDAG &DAG,
1939 SmallVectorImpl<SDValue> &InVals)
1942 MachineFunction &MF = DAG.getMachineFunction();
1943 MachineFrameInfo *MFI = MF.getFrameInfo();
1945 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1947 // Assign locations to all of the incoming arguments.
1948 SmallVector<CCValAssign, 16> ArgLocs;
1949 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1951 CCInfo.AnalyzeFormalArguments(Ins,
1952 CCAssignFnForNode(CallConv, /* Return*/ false,
1955 SmallVector<SDValue, 16> ArgValues;
1957 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1958 CCValAssign &VA = ArgLocs[i];
1960 // Arguments stored in registers.
1961 if (VA.isRegLoc()) {
1962 EVT RegVT = VA.getLocVT();
1965 if (VA.needsCustom()) {
1966 // f64 and vector types are split up into multiple registers or
1967 // combinations of registers and stack slots.
1968 if (VA.getLocVT() == MVT::v2f64) {
1969 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1971 VA = ArgLocs[++i]; // skip ahead to next loc
1973 if (VA.isMemLoc()) {
1974 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
1975 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1976 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1977 PseudoSourceValue::getFixedStack(FI), 0,
1980 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1983 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1984 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1985 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1986 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1987 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1989 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1992 TargetRegisterClass *RC;
1994 if (RegVT == MVT::f32)
1995 RC = ARM::SPRRegisterClass;
1996 else if (RegVT == MVT::f64)
1997 RC = ARM::DPRRegisterClass;
1998 else if (RegVT == MVT::v2f64)
1999 RC = ARM::QPRRegisterClass;
2000 else if (RegVT == MVT::i32)
2001 RC = (AFI->isThumb1OnlyFunction() ?
2002 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2004 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2006 // Transform the arguments in physical registers into virtual ones.
2007 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2008 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2011 // If this is an 8 or 16-bit value, it is really passed promoted
2012 // to 32 bits. Insert an assert[sz]ext to capture this, then
2013 // truncate to the right size.
2014 switch (VA.getLocInfo()) {
2015 default: llvm_unreachable("Unknown loc info!");
2016 case CCValAssign::Full: break;
2017 case CCValAssign::BCvt:
2018 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2020 case CCValAssign::SExt:
2021 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2022 DAG.getValueType(VA.getValVT()));
2023 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2025 case CCValAssign::ZExt:
2026 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2027 DAG.getValueType(VA.getValVT()));
2028 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2032 InVals.push_back(ArgValue);
2034 } else { // VA.isRegLoc()
2037 assert(VA.isMemLoc());
2038 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2040 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2041 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2043 // Create load nodes to retrieve arguments from the stack.
2044 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2045 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2046 PseudoSourceValue::getFixedStack(FI), 0,
2053 static const unsigned GPRArgRegs[] = {
2054 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2057 unsigned NumGPRs = CCInfo.getFirstUnallocated
2058 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2060 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2061 unsigned VARegSize = (4 - NumGPRs) * 4;
2062 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2063 unsigned ArgOffset = CCInfo.getNextStackOffset();
2064 if (VARegSaveSize) {
2065 // If this function is vararg, store any remaining integer argument regs
2066 // to their spots on the stack so that they may be loaded by deferencing
2067 // the result of va_next.
2068 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2069 AFI->setVarArgsFrameIndex(
2070 MFI->CreateFixedObject(VARegSaveSize,
2071 ArgOffset + VARegSaveSize - VARegSize,
2073 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2076 SmallVector<SDValue, 4> MemOps;
2077 for (; NumGPRs < 4; ++NumGPRs) {
2078 TargetRegisterClass *RC;
2079 if (AFI->isThumb1OnlyFunction())
2080 RC = ARM::tGPRRegisterClass;
2082 RC = ARM::GPRRegisterClass;
2084 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2085 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2087 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2088 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2089 0, false, false, 0);
2090 MemOps.push_back(Store);
2091 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2092 DAG.getConstant(4, getPointerTy()));
2094 if (!MemOps.empty())
2095 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2096 &MemOps[0], MemOps.size());
2098 // This will point to the next argument passed via stack.
2099 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2105 /// isFloatingPointZero - Return true if this is +0.0.
2106 static bool isFloatingPointZero(SDValue Op) {
2107 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2108 return CFP->getValueAPF().isPosZero();
2109 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2110 // Maybe this has already been legalized into the constant pool?
2111 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2112 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2113 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2114 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2115 return CFP->getValueAPF().isPosZero();
2121 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2122 /// the given operands.
2124 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2125 SDValue &ARMcc, SelectionDAG &DAG,
2126 DebugLoc dl) const {
2127 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2128 unsigned C = RHSC->getZExtValue();
2129 if (!isLegalICmpImmediate(C)) {
2130 // Constant does not fit, try adjusting it by one?
2135 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2136 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2137 RHS = DAG.getConstant(C-1, MVT::i32);
2142 if (C != 0 && isLegalICmpImmediate(C-1)) {
2143 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2144 RHS = DAG.getConstant(C-1, MVT::i32);
2149 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2150 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2151 RHS = DAG.getConstant(C+1, MVT::i32);
2156 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2157 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2158 RHS = DAG.getConstant(C+1, MVT::i32);
2165 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2166 ARMISD::NodeType CompareType;
2169 CompareType = ARMISD::CMP;
2174 CompareType = ARMISD::CMPZ;
2177 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2178 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2181 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2183 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2184 DebugLoc dl) const {
2186 if (!isFloatingPointZero(RHS))
2187 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2189 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2190 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2193 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2194 SDValue Cond = Op.getOperand(0);
2195 SDValue SelectTrue = Op.getOperand(1);
2196 SDValue SelectFalse = Op.getOperand(2);
2197 DebugLoc dl = Op.getDebugLoc();
2201 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2202 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2204 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2205 const ConstantSDNode *CMOVTrue =
2206 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2207 const ConstantSDNode *CMOVFalse =
2208 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2210 if (CMOVTrue && CMOVFalse) {
2211 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2212 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2216 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2218 False = SelectFalse;
2219 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2224 if (True.getNode() && False.getNode()) {
2225 EVT VT = Cond.getValueType();
2226 SDValue ARMcc = Cond.getOperand(2);
2227 SDValue CCR = Cond.getOperand(3);
2228 SDValue Cmp = Cond.getOperand(4);
2229 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2234 return DAG.getSelectCC(dl, Cond,
2235 DAG.getConstant(0, Cond.getValueType()),
2236 SelectTrue, SelectFalse, ISD::SETNE);
2239 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2240 EVT VT = Op.getValueType();
2241 SDValue LHS = Op.getOperand(0);
2242 SDValue RHS = Op.getOperand(1);
2243 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2244 SDValue TrueVal = Op.getOperand(2);
2245 SDValue FalseVal = Op.getOperand(3);
2246 DebugLoc dl = Op.getDebugLoc();
2248 if (LHS.getValueType() == MVT::i32) {
2250 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2251 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2252 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2255 ARMCC::CondCodes CondCode, CondCode2;
2256 FPCCToARMCC(CC, CondCode, CondCode2);
2258 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2259 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2260 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2261 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2263 if (CondCode2 != ARMCC::AL) {
2264 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2265 // FIXME: Needs another CMP because flag can have but one use.
2266 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2267 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2268 Result, TrueVal, ARMcc2, CCR, Cmp2);
2273 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2274 /// to morph to an integer compare sequence.
2275 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2276 const ARMSubtarget *Subtarget) {
2277 SDNode *N = Op.getNode();
2278 if (!N->hasOneUse())
2279 // Otherwise it requires moving the value from fp to integer registers.
2281 if (!N->getNumValues())
2283 EVT VT = Op.getValueType();
2284 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2285 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2286 // vmrs are very slow, e.g. cortex-a8.
2289 if (isFloatingPointZero(Op)) {
2293 return ISD::isNormalLoad(N);
2296 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2297 if (isFloatingPointZero(Op))
2298 return DAG.getConstant(0, MVT::i32);
2300 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2301 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2302 Ld->getChain(), Ld->getBasePtr(),
2303 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2304 Ld->isVolatile(), Ld->isNonTemporal(),
2305 Ld->getAlignment());
2307 llvm_unreachable("Unknown VFP cmp argument!");
2310 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2311 SDValue &RetVal1, SDValue &RetVal2) {
2312 if (isFloatingPointZero(Op)) {
2313 RetVal1 = DAG.getConstant(0, MVT::i32);
2314 RetVal2 = DAG.getConstant(0, MVT::i32);
2318 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2319 SDValue Ptr = Ld->getBasePtr();
2320 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2321 Ld->getChain(), Ptr,
2322 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2323 Ld->isVolatile(), Ld->isNonTemporal(),
2324 Ld->getAlignment());
2326 EVT PtrType = Ptr.getValueType();
2327 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2328 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2329 PtrType, Ptr, DAG.getConstant(4, PtrType));
2330 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2331 Ld->getChain(), NewPtr,
2332 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2333 Ld->isVolatile(), Ld->isNonTemporal(),
2338 llvm_unreachable("Unknown VFP cmp argument!");
2341 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2342 /// f32 and even f64 comparisons to integer ones.
2344 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2345 SDValue Chain = Op.getOperand(0);
2346 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2347 SDValue LHS = Op.getOperand(2);
2348 SDValue RHS = Op.getOperand(3);
2349 SDValue Dest = Op.getOperand(4);
2350 DebugLoc dl = Op.getDebugLoc();
2352 bool SeenZero = false;
2353 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2354 canChangeToInt(RHS, SeenZero, Subtarget) &&
2355 // If one of the operand is zero, it's safe to ignore the NaN case since
2356 // we only care about equality comparisons.
2357 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2358 // If unsafe fp math optimization is enabled and there are no othter uses of
2359 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2360 // to an integer comparison.
2361 if (CC == ISD::SETOEQ)
2363 else if (CC == ISD::SETUNE)
2367 if (LHS.getValueType() == MVT::f32) {
2368 LHS = bitcastf32Toi32(LHS, DAG);
2369 RHS = bitcastf32Toi32(RHS, DAG);
2370 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2371 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2372 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2373 Chain, Dest, ARMcc, CCR, Cmp);
2378 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2379 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2380 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2381 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2382 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2383 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2384 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2390 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2391 SDValue Chain = Op.getOperand(0);
2392 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2393 SDValue LHS = Op.getOperand(2);
2394 SDValue RHS = Op.getOperand(3);
2395 SDValue Dest = Op.getOperand(4);
2396 DebugLoc dl = Op.getDebugLoc();
2398 if (LHS.getValueType() == MVT::i32) {
2400 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2401 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2402 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2403 Chain, Dest, ARMcc, CCR, Cmp);
2406 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2409 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2410 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2411 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2412 if (Result.getNode())
2416 ARMCC::CondCodes CondCode, CondCode2;
2417 FPCCToARMCC(CC, CondCode, CondCode2);
2419 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2420 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2421 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2422 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2423 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2424 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2425 if (CondCode2 != ARMCC::AL) {
2426 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2427 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2428 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2433 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2434 SDValue Chain = Op.getOperand(0);
2435 SDValue Table = Op.getOperand(1);
2436 SDValue Index = Op.getOperand(2);
2437 DebugLoc dl = Op.getDebugLoc();
2439 EVT PTy = getPointerTy();
2440 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2441 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2442 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2443 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2444 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2445 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2446 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2447 if (Subtarget->isThumb2()) {
2448 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2449 // which does another jump to the destination. This also makes it easier
2450 // to translate it to TBB / TBH later.
2451 // FIXME: This might not work if the function is extremely large.
2452 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2453 Addr, Op.getOperand(2), JTI, UId);
2455 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2456 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2457 PseudoSourceValue::getJumpTable(), 0,
2459 Chain = Addr.getValue(1);
2460 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2461 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2463 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2464 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2465 Chain = Addr.getValue(1);
2466 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2470 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2471 DebugLoc dl = Op.getDebugLoc();
2474 switch (Op.getOpcode()) {
2476 assert(0 && "Invalid opcode!");
2477 case ISD::FP_TO_SINT:
2478 Opc = ARMISD::FTOSI;
2480 case ISD::FP_TO_UINT:
2481 Opc = ARMISD::FTOUI;
2484 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2485 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2488 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2489 EVT VT = Op.getValueType();
2490 DebugLoc dl = Op.getDebugLoc();
2493 switch (Op.getOpcode()) {
2495 assert(0 && "Invalid opcode!");
2496 case ISD::SINT_TO_FP:
2497 Opc = ARMISD::SITOF;
2499 case ISD::UINT_TO_FP:
2500 Opc = ARMISD::UITOF;
2504 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2505 return DAG.getNode(Opc, dl, VT, Op);
2508 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2509 // Implement fcopysign with a fabs and a conditional fneg.
2510 SDValue Tmp0 = Op.getOperand(0);
2511 SDValue Tmp1 = Op.getOperand(1);
2512 DebugLoc dl = Op.getDebugLoc();
2513 EVT VT = Op.getValueType();
2514 EVT SrcVT = Tmp1.getValueType();
2515 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2516 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2517 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2518 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2519 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2520 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2523 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2524 MachineFunction &MF = DAG.getMachineFunction();
2525 MachineFrameInfo *MFI = MF.getFrameInfo();
2526 MFI->setReturnAddressIsTaken(true);
2528 EVT VT = Op.getValueType();
2529 DebugLoc dl = Op.getDebugLoc();
2530 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2532 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2533 SDValue Offset = DAG.getConstant(4, MVT::i32);
2534 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2535 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2536 NULL, 0, false, false, 0);
2539 // Return LR, which contains the return address. Mark it an implicit live-in.
2540 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
2541 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2544 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2545 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2546 MFI->setFrameAddressIsTaken(true);
2548 EVT VT = Op.getValueType();
2549 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2550 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2551 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2552 ? ARM::R7 : ARM::R11;
2553 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2555 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2560 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2561 /// expand a bit convert where either the source or destination type is i64 to
2562 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2563 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2564 /// vectors), since the legalizer won't know what to do with that.
2565 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2566 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2567 DebugLoc dl = N->getDebugLoc();
2568 SDValue Op = N->getOperand(0);
2570 // This function is only supposed to be called for i64 types, either as the
2571 // source or destination of the bit convert.
2572 EVT SrcVT = Op.getValueType();
2573 EVT DstVT = N->getValueType(0);
2574 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2575 "ExpandBIT_CONVERT called for non-i64 type");
2577 // Turn i64->f64 into VMOVDRR.
2578 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2579 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2580 DAG.getConstant(0, MVT::i32));
2581 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2582 DAG.getConstant(1, MVT::i32));
2583 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2584 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2587 // Turn f64->i64 into VMOVRRD.
2588 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2589 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2590 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2591 // Merge the pieces into a single i64 value.
2592 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2598 /// getZeroVector - Returns a vector of specified type with all zero elements.
2599 /// Zero vectors are used to represent vector negation and in those cases
2600 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2601 /// not support i64 elements, so sometimes the zero vectors will need to be
2602 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2604 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2605 assert(VT.isVector() && "Expected a vector type");
2606 // The canonical modified immediate encoding of a zero vector is....0!
2607 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2608 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2609 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2610 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2613 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2614 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2615 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2616 SelectionDAG &DAG) const {
2617 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2618 EVT VT = Op.getValueType();
2619 unsigned VTBits = VT.getSizeInBits();
2620 DebugLoc dl = Op.getDebugLoc();
2621 SDValue ShOpLo = Op.getOperand(0);
2622 SDValue ShOpHi = Op.getOperand(1);
2623 SDValue ShAmt = Op.getOperand(2);
2625 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2627 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2629 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2630 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2631 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2632 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2633 DAG.getConstant(VTBits, MVT::i32));
2634 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2635 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2636 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2638 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2639 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2641 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2642 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2645 SDValue Ops[2] = { Lo, Hi };
2646 return DAG.getMergeValues(Ops, 2, dl);
2649 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2650 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2651 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2652 SelectionDAG &DAG) const {
2653 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2654 EVT VT = Op.getValueType();
2655 unsigned VTBits = VT.getSizeInBits();
2656 DebugLoc dl = Op.getDebugLoc();
2657 SDValue ShOpLo = Op.getOperand(0);
2658 SDValue ShOpHi = Op.getOperand(1);
2659 SDValue ShAmt = Op.getOperand(2);
2662 assert(Op.getOpcode() == ISD::SHL_PARTS);
2663 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2664 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2665 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2666 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2667 DAG.getConstant(VTBits, MVT::i32));
2668 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2669 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2671 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2672 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2673 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2675 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2676 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2679 SDValue Ops[2] = { Lo, Hi };
2680 return DAG.getMergeValues(Ops, 2, dl);
2683 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
2684 SelectionDAG &DAG) const {
2685 // The rounding mode is in bits 23:22 of the FPSCR.
2686 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
2687 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
2688 // so that the shift + and get folded into a bitfield extract.
2689 DebugLoc dl = Op.getDebugLoc();
2690 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
2691 DAG.getConstant(Intrinsic::arm_get_fpscr,
2693 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
2694 DAG.getConstant(1U << 22, MVT::i32));
2695 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
2696 DAG.getConstant(22, MVT::i32));
2697 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
2698 DAG.getConstant(3, MVT::i32));
2701 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2702 const ARMSubtarget *ST) {
2703 EVT VT = N->getValueType(0);
2704 DebugLoc dl = N->getDebugLoc();
2706 if (!ST->hasV6T2Ops())
2709 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2710 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2713 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2714 const ARMSubtarget *ST) {
2715 EVT VT = N->getValueType(0);
2716 DebugLoc dl = N->getDebugLoc();
2718 // Lower vector shifts on NEON to use VSHL.
2719 if (VT.isVector()) {
2720 assert(ST->hasNEON() && "unexpected vector shift");
2722 // Left shifts translate directly to the vshiftu intrinsic.
2723 if (N->getOpcode() == ISD::SHL)
2724 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2725 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2726 N->getOperand(0), N->getOperand(1));
2728 assert((N->getOpcode() == ISD::SRA ||
2729 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2731 // NEON uses the same intrinsics for both left and right shifts. For
2732 // right shifts, the shift amounts are negative, so negate the vector of
2734 EVT ShiftVT = N->getOperand(1).getValueType();
2735 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2736 getZeroVector(ShiftVT, DAG, dl),
2738 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2739 Intrinsic::arm_neon_vshifts :
2740 Intrinsic::arm_neon_vshiftu);
2741 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2742 DAG.getConstant(vshiftInt, MVT::i32),
2743 N->getOperand(0), NegatedCount);
2746 // We can get here for a node like i32 = ISD::SHL i32, i64
2750 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2751 "Unknown shift to lower!");
2753 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2754 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2755 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2758 // If we are in thumb mode, we don't have RRX.
2759 if (ST->isThumb1Only()) return SDValue();
2761 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2762 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2763 DAG.getConstant(0, MVT::i32));
2764 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2765 DAG.getConstant(1, MVT::i32));
2767 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2768 // captures the result into a carry flag.
2769 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2770 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2772 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2773 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2775 // Merge the pieces into a single i64 value.
2776 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2779 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2780 SDValue TmpOp0, TmpOp1;
2781 bool Invert = false;
2785 SDValue Op0 = Op.getOperand(0);
2786 SDValue Op1 = Op.getOperand(1);
2787 SDValue CC = Op.getOperand(2);
2788 EVT VT = Op.getValueType();
2789 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2790 DebugLoc dl = Op.getDebugLoc();
2792 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2793 switch (SetCCOpcode) {
2794 default: llvm_unreachable("Illegal FP comparison"); break;
2796 case ISD::SETNE: Invert = true; // Fallthrough
2798 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2800 case ISD::SETLT: Swap = true; // Fallthrough
2802 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2804 case ISD::SETLE: Swap = true; // Fallthrough
2806 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2807 case ISD::SETUGE: Swap = true; // Fallthrough
2808 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2809 case ISD::SETUGT: Swap = true; // Fallthrough
2810 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2811 case ISD::SETUEQ: Invert = true; // Fallthrough
2813 // Expand this to (OLT | OGT).
2817 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2818 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2820 case ISD::SETUO: Invert = true; // Fallthrough
2822 // Expand this to (OLT | OGE).
2826 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2827 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2831 // Integer comparisons.
2832 switch (SetCCOpcode) {
2833 default: llvm_unreachable("Illegal integer comparison"); break;
2834 case ISD::SETNE: Invert = true;
2835 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2836 case ISD::SETLT: Swap = true;
2837 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2838 case ISD::SETLE: Swap = true;
2839 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2840 case ISD::SETULT: Swap = true;
2841 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2842 case ISD::SETULE: Swap = true;
2843 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2846 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2847 if (Opc == ARMISD::VCEQ) {
2850 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2852 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2855 // Ignore bitconvert.
2856 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2857 AndOp = AndOp.getOperand(0);
2859 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2861 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2862 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2869 std::swap(Op0, Op1);
2871 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2874 Result = DAG.getNOT(dl, Result, VT);
2879 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2880 /// valid vector constant for a NEON instruction with a "modified immediate"
2881 /// operand (e.g., VMOV). If so, return the encoded value.
2882 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2883 unsigned SplatBitSize, SelectionDAG &DAG,
2884 EVT &VT, bool is128Bits, bool isVMOV) {
2885 unsigned OpCmode, Imm;
2887 // SplatBitSize is set to the smallest size that splats the vector, so a
2888 // zero vector will always have SplatBitSize == 8. However, NEON modified
2889 // immediate instructions others than VMOV do not support the 8-bit encoding
2890 // of a zero vector, and the default encoding of zero is supposed to be the
2895 switch (SplatBitSize) {
2899 // Any 1-byte value is OK. Op=0, Cmode=1110.
2900 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2903 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2907 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2908 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2909 if ((SplatBits & ~0xff) == 0) {
2910 // Value = 0x00nn: Op=x, Cmode=100x.
2915 if ((SplatBits & ~0xff00) == 0) {
2916 // Value = 0xnn00: Op=x, Cmode=101x.
2918 Imm = SplatBits >> 8;
2924 // NEON's 32-bit VMOV supports splat values where:
2925 // * only one byte is nonzero, or
2926 // * the least significant byte is 0xff and the second byte is nonzero, or
2927 // * the least significant 2 bytes are 0xff and the third is nonzero.
2928 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
2929 if ((SplatBits & ~0xff) == 0) {
2930 // Value = 0x000000nn: Op=x, Cmode=000x.
2935 if ((SplatBits & ~0xff00) == 0) {
2936 // Value = 0x0000nn00: Op=x, Cmode=001x.
2938 Imm = SplatBits >> 8;
2941 if ((SplatBits & ~0xff0000) == 0) {
2942 // Value = 0x00nn0000: Op=x, Cmode=010x.
2944 Imm = SplatBits >> 16;
2947 if ((SplatBits & ~0xff000000) == 0) {
2948 // Value = 0xnn000000: Op=x, Cmode=011x.
2950 Imm = SplatBits >> 24;
2954 if ((SplatBits & ~0xffff) == 0 &&
2955 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2956 // Value = 0x0000nnff: Op=x, Cmode=1100.
2958 Imm = SplatBits >> 8;
2963 if ((SplatBits & ~0xffffff) == 0 &&
2964 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2965 // Value = 0x00nnffff: Op=x, Cmode=1101.
2967 Imm = SplatBits >> 16;
2968 SplatBits |= 0xffff;
2972 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2973 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2974 // VMOV.I32. A (very) minor optimization would be to replicate the value
2975 // and fall through here to test for a valid 64-bit splat. But, then the
2976 // caller would also need to check and handle the change in size.
2982 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2983 uint64_t BitMask = 0xff;
2985 unsigned ImmMask = 1;
2987 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2988 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2991 } else if ((SplatBits & BitMask) != 0) {
2997 // Op=1, Cmode=1110.
3000 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3005 llvm_unreachable("unexpected size for isNEONModifiedImm");
3009 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3010 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3013 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3014 bool &ReverseVEXT, unsigned &Imm) {
3015 unsigned NumElts = VT.getVectorNumElements();
3016 ReverseVEXT = false;
3018 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3024 // If this is a VEXT shuffle, the immediate value is the index of the first
3025 // element. The other shuffle indices must be the successive elements after
3027 unsigned ExpectedElt = Imm;
3028 for (unsigned i = 1; i < NumElts; ++i) {
3029 // Increment the expected index. If it wraps around, it may still be
3030 // a VEXT but the source vectors must be swapped.
3032 if (ExpectedElt == NumElts * 2) {
3037 if (M[i] < 0) continue; // ignore UNDEF indices
3038 if (ExpectedElt != static_cast<unsigned>(M[i]))
3042 // Adjust the index value if the source operands will be swapped.
3049 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3050 /// instruction with the specified blocksize. (The order of the elements
3051 /// within each block of the vector is reversed.)
3052 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3053 unsigned BlockSize) {
3054 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3055 "Only possible block sizes for VREV are: 16, 32, 64");
3057 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3061 unsigned NumElts = VT.getVectorNumElements();
3062 unsigned BlockElts = M[0] + 1;
3063 // If the first shuffle index is UNDEF, be optimistic.
3065 BlockElts = BlockSize / EltSz;
3067 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3070 for (unsigned i = 0; i < NumElts; ++i) {
3071 if (M[i] < 0) continue; // ignore UNDEF indices
3072 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3079 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3080 unsigned &WhichResult) {
3081 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3085 unsigned NumElts = VT.getVectorNumElements();
3086 WhichResult = (M[0] == 0 ? 0 : 1);
3087 for (unsigned i = 0; i < NumElts; i += 2) {
3088 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3089 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3095 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3096 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3097 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3098 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3099 unsigned &WhichResult) {
3100 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3104 unsigned NumElts = VT.getVectorNumElements();
3105 WhichResult = (M[0] == 0 ? 0 : 1);
3106 for (unsigned i = 0; i < NumElts; i += 2) {
3107 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3108 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3114 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3115 unsigned &WhichResult) {
3116 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3120 unsigned NumElts = VT.getVectorNumElements();
3121 WhichResult = (M[0] == 0 ? 0 : 1);
3122 for (unsigned i = 0; i != NumElts; ++i) {
3123 if (M[i] < 0) continue; // ignore UNDEF indices
3124 if ((unsigned) M[i] != 2 * i + WhichResult)
3128 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3129 if (VT.is64BitVector() && EltSz == 32)
3135 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3136 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3137 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3138 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3139 unsigned &WhichResult) {
3140 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3144 unsigned Half = VT.getVectorNumElements() / 2;
3145 WhichResult = (M[0] == 0 ? 0 : 1);
3146 for (unsigned j = 0; j != 2; ++j) {
3147 unsigned Idx = WhichResult;
3148 for (unsigned i = 0; i != Half; ++i) {
3149 int MIdx = M[i + j * Half];
3150 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3156 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3157 if (VT.is64BitVector() && EltSz == 32)
3163 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3164 unsigned &WhichResult) {
3165 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3169 unsigned NumElts = VT.getVectorNumElements();
3170 WhichResult = (M[0] == 0 ? 0 : 1);
3171 unsigned Idx = WhichResult * NumElts / 2;
3172 for (unsigned i = 0; i != NumElts; i += 2) {
3173 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3174 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3179 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3180 if (VT.is64BitVector() && EltSz == 32)
3186 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3187 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3188 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3189 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3190 unsigned &WhichResult) {
3191 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3195 unsigned NumElts = VT.getVectorNumElements();
3196 WhichResult = (M[0] == 0 ? 0 : 1);
3197 unsigned Idx = WhichResult * NumElts / 2;
3198 for (unsigned i = 0; i != NumElts; i += 2) {
3199 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3200 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3205 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3206 if (VT.is64BitVector() && EltSz == 32)
3212 // If N is an integer constant that can be moved into a register in one
3213 // instruction, return an SDValue of such a constant (will become a MOV
3214 // instruction). Otherwise return null.
3215 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3216 const ARMSubtarget *ST, DebugLoc dl) {
3218 if (!isa<ConstantSDNode>(N))
3220 Val = cast<ConstantSDNode>(N)->getZExtValue();
3222 if (ST->isThumb1Only()) {
3223 if (Val <= 255 || ~Val <= 255)
3224 return DAG.getConstant(Val, MVT::i32);
3226 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3227 return DAG.getConstant(Val, MVT::i32);
3232 // If this is a case we can't handle, return null and let the default
3233 // expansion code take care of it.
3234 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3235 const ARMSubtarget *ST) {
3236 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3237 DebugLoc dl = Op.getDebugLoc();
3238 EVT VT = Op.getValueType();
3240 APInt SplatBits, SplatUndef;
3241 unsigned SplatBitSize;
3243 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3244 if (SplatBitSize <= 64) {
3245 // Check if an immediate VMOV works.
3247 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3248 SplatUndef.getZExtValue(), SplatBitSize,
3249 DAG, VmovVT, VT.is128BitVector(), true);
3250 if (Val.getNode()) {
3251 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3252 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3255 // Try an immediate VMVN.
3256 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3257 ((1LL << SplatBitSize) - 1));
3258 Val = isNEONModifiedImm(NegatedImm,
3259 SplatUndef.getZExtValue(), SplatBitSize,
3260 DAG, VmovVT, VT.is128BitVector(), false);
3261 if (Val.getNode()) {
3262 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3263 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3268 // Scan through the operands to see if only one value is used.
3269 unsigned NumElts = VT.getVectorNumElements();
3270 bool isOnlyLowElement = true;
3271 bool usesOnlyOneValue = true;
3272 bool isConstant = true;
3274 for (unsigned i = 0; i < NumElts; ++i) {
3275 SDValue V = Op.getOperand(i);
3276 if (V.getOpcode() == ISD::UNDEF)
3279 isOnlyLowElement = false;
3280 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3283 if (!Value.getNode())
3285 else if (V != Value)
3286 usesOnlyOneValue = false;
3289 if (!Value.getNode())
3290 return DAG.getUNDEF(VT);
3292 if (isOnlyLowElement)
3293 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3295 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3297 if (EnableARMVDUPsplat) {
3298 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3299 // i32 and try again.
3300 if (usesOnlyOneValue && EltSize <= 32) {
3302 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3303 if (VT.getVectorElementType().isFloatingPoint()) {
3304 SmallVector<SDValue, 8> Ops;
3305 for (unsigned i = 0; i < NumElts; ++i)
3306 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
3308 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
3310 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3311 LowerBUILD_VECTOR(Val, DAG, ST));
3313 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3315 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3319 // If all elements are constants and the case above didn't get hit, fall back
3320 // to the default expansion, which will generate a load from the constant
3325 if (!EnableARMVDUPsplat) {
3326 // Use VDUP for non-constant splats.
3327 if (usesOnlyOneValue && EltSize <= 32)
3328 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3331 // Vectors with 32- or 64-bit elements can be built by directly assigning
3332 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3333 // will be legalized.
3334 if (EltSize >= 32) {
3335 // Do the expansion with floating-point types, since that is what the VFP
3336 // registers are defined to use, and since i64 is not legal.
3337 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3338 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3339 SmallVector<SDValue, 8> Ops;
3340 for (unsigned i = 0; i < NumElts; ++i)
3341 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3342 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3343 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3349 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3350 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3351 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3352 /// are assumed to be legal.
3354 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3356 if (VT.getVectorNumElements() == 4 &&
3357 (VT.is128BitVector() || VT.is64BitVector())) {
3358 unsigned PFIndexes[4];
3359 for (unsigned i = 0; i != 4; ++i) {
3363 PFIndexes[i] = M[i];
3366 // Compute the index in the perfect shuffle table.
3367 unsigned PFTableIndex =
3368 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3369 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3370 unsigned Cost = (PFEntry >> 30);
3377 unsigned Imm, WhichResult;
3379 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3380 return (EltSize >= 32 ||
3381 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3382 isVREVMask(M, VT, 64) ||
3383 isVREVMask(M, VT, 32) ||
3384 isVREVMask(M, VT, 16) ||
3385 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3386 isVTRNMask(M, VT, WhichResult) ||
3387 isVUZPMask(M, VT, WhichResult) ||
3388 isVZIPMask(M, VT, WhichResult) ||
3389 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3390 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3391 isVZIP_v_undef_Mask(M, VT, WhichResult));
3394 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3395 /// the specified operations to build the shuffle.
3396 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3397 SDValue RHS, SelectionDAG &DAG,
3399 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3400 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3401 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3404 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3413 OP_VUZPL, // VUZP, left result
3414 OP_VUZPR, // VUZP, right result
3415 OP_VZIPL, // VZIP, left result
3416 OP_VZIPR, // VZIP, right result
3417 OP_VTRNL, // VTRN, left result
3418 OP_VTRNR // VTRN, right result
3421 if (OpNum == OP_COPY) {
3422 if (LHSID == (1*9+2)*9+3) return LHS;
3423 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3427 SDValue OpLHS, OpRHS;
3428 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3429 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3430 EVT VT = OpLHS.getValueType();
3433 default: llvm_unreachable("Unknown shuffle opcode!");
3435 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3440 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3441 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3445 return DAG.getNode(ARMISD::VEXT, dl, VT,
3447 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3450 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3451 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3454 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3455 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3458 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3459 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3463 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3464 SDValue V1 = Op.getOperand(0);
3465 SDValue V2 = Op.getOperand(1);
3466 DebugLoc dl = Op.getDebugLoc();
3467 EVT VT = Op.getValueType();
3468 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3469 SmallVector<int, 8> ShuffleMask;
3471 // Convert shuffles that are directly supported on NEON to target-specific
3472 // DAG nodes, instead of keeping them as shuffles and matching them again
3473 // during code selection. This is more efficient and avoids the possibility
3474 // of inconsistencies between legalization and selection.
3475 // FIXME: floating-point vectors should be canonicalized to integer vectors
3476 // of the same time so that they get CSEd properly.
3477 SVN->getMask(ShuffleMask);
3479 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3480 if (EltSize <= 32) {
3481 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3482 int Lane = SVN->getSplatIndex();
3483 // If this is undef splat, generate it via "just" vdup, if possible.
3484 if (Lane == -1) Lane = 0;
3486 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3487 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3489 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3490 DAG.getConstant(Lane, MVT::i32));
3495 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3498 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3499 DAG.getConstant(Imm, MVT::i32));
3502 if (isVREVMask(ShuffleMask, VT, 64))
3503 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3504 if (isVREVMask(ShuffleMask, VT, 32))
3505 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3506 if (isVREVMask(ShuffleMask, VT, 16))
3507 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3509 // Check for Neon shuffles that modify both input vectors in place.
3510 // If both results are used, i.e., if there are two shuffles with the same
3511 // source operands and with masks corresponding to both results of one of
3512 // these operations, DAG memoization will ensure that a single node is
3513 // used for both shuffles.
3514 unsigned WhichResult;
3515 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3516 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3517 V1, V2).getValue(WhichResult);
3518 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3519 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3520 V1, V2).getValue(WhichResult);
3521 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3522 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3523 V1, V2).getValue(WhichResult);
3525 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3526 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3527 V1, V1).getValue(WhichResult);
3528 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3529 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3530 V1, V1).getValue(WhichResult);
3531 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3532 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3533 V1, V1).getValue(WhichResult);
3536 // If the shuffle is not directly supported and it has 4 elements, use
3537 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3538 unsigned NumElts = VT.getVectorNumElements();
3540 unsigned PFIndexes[4];
3541 for (unsigned i = 0; i != 4; ++i) {
3542 if (ShuffleMask[i] < 0)
3545 PFIndexes[i] = ShuffleMask[i];
3548 // Compute the index in the perfect shuffle table.
3549 unsigned PFTableIndex =
3550 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3551 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3552 unsigned Cost = (PFEntry >> 30);
3555 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3558 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3559 if (EltSize >= 32) {
3560 // Do the expansion with floating-point types, since that is what the VFP
3561 // registers are defined to use, and since i64 is not legal.
3562 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3563 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3564 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3565 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3566 SmallVector<SDValue, 8> Ops;
3567 for (unsigned i = 0; i < NumElts; ++i) {
3568 if (ShuffleMask[i] < 0)
3569 Ops.push_back(DAG.getUNDEF(EltVT));
3571 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3572 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3573 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3576 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3577 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3583 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3584 EVT VT = Op.getValueType();
3585 DebugLoc dl = Op.getDebugLoc();
3586 SDValue Vec = Op.getOperand(0);
3587 SDValue Lane = Op.getOperand(1);
3588 assert(VT == MVT::i32 &&
3589 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3590 "unexpected type for custom-lowering vector extract");
3591 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3594 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3595 // The only time a CONCAT_VECTORS operation can have legal types is when
3596 // two 64-bit vectors are concatenated to a 128-bit vector.
3597 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3598 "unexpected CONCAT_VECTORS");
3599 DebugLoc dl = Op.getDebugLoc();
3600 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3601 SDValue Op0 = Op.getOperand(0);
3602 SDValue Op1 = Op.getOperand(1);
3603 if (Op0.getOpcode() != ISD::UNDEF)
3604 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3605 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3606 DAG.getIntPtrConstant(0));
3607 if (Op1.getOpcode() != ISD::UNDEF)
3608 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3609 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3610 DAG.getIntPtrConstant(1));
3611 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3614 /// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
3615 /// an extending load, return the unextended value.
3616 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
3617 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
3618 return N->getOperand(0);
3619 LoadSDNode *LD = cast<LoadSDNode>(N);
3620 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
3621 LD->getBasePtr(), LD->getSrcValue(),
3622 LD->getSrcValueOffset(), LD->isVolatile(),
3623 LD->isNonTemporal(), LD->getAlignment());
3626 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
3627 // Multiplications are only custom-lowered for 128-bit vectors so that
3628 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
3629 EVT VT = Op.getValueType();
3630 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
3631 SDNode *N0 = Op.getOperand(0).getNode();
3632 SDNode *N1 = Op.getOperand(1).getNode();
3633 unsigned NewOpc = 0;
3634 if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
3635 (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
3636 NewOpc = ARMISD::VMULLs;
3637 } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
3638 (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
3639 NewOpc = ARMISD::VMULLu;
3640 } else if (VT.getSimpleVT().SimpleTy == MVT::v2i64) {
3641 // Fall through to expand this. It is not legal.
3644 // Other vector multiplications are legal.
3648 // Legalize to a VMULL instruction.
3649 DebugLoc DL = Op.getDebugLoc();
3650 SDValue Op0 = SkipExtension(N0, DAG);
3651 SDValue Op1 = SkipExtension(N1, DAG);
3653 assert(Op0.getValueType().is64BitVector() &&
3654 Op1.getValueType().is64BitVector() &&
3655 "unexpected types for extended operands to VMULL");
3656 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
3659 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3660 switch (Op.getOpcode()) {
3661 default: llvm_unreachable("Don't know how to custom lower this!");
3662 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3663 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3664 case ISD::GlobalAddress:
3665 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3666 LowerGlobalAddressELF(Op, DAG);
3667 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3668 case ISD::SELECT: return LowerSELECT(Op, DAG);
3669 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3670 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3671 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3672 case ISD::VASTART: return LowerVASTART(Op, DAG);
3673 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3674 case ISD::SINT_TO_FP:
3675 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3676 case ISD::FP_TO_SINT:
3677 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3678 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3679 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3680 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3681 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3682 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3683 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3684 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3686 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3689 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3690 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3691 case ISD::SRL_PARTS:
3692 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3693 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3694 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3695 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
3696 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3697 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3698 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3699 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3700 case ISD::MUL: return LowerMUL(Op, DAG);
3705 /// ReplaceNodeResults - Replace the results of node with an illegal result
3706 /// type with new values built out of custom code.
3707 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3708 SmallVectorImpl<SDValue>&Results,
3709 SelectionDAG &DAG) const {
3711 switch (N->getOpcode()) {
3713 llvm_unreachable("Don't know how to custom expand this!");
3715 case ISD::BIT_CONVERT:
3716 Res = ExpandBIT_CONVERT(N, DAG);
3720 Res = LowerShift(N, DAG, Subtarget);
3724 Results.push_back(Res);
3727 //===----------------------------------------------------------------------===//
3728 // ARM Scheduler Hooks
3729 //===----------------------------------------------------------------------===//
3732 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3733 MachineBasicBlock *BB,
3734 unsigned Size) const {
3735 unsigned dest = MI->getOperand(0).getReg();
3736 unsigned ptr = MI->getOperand(1).getReg();
3737 unsigned oldval = MI->getOperand(2).getReg();
3738 unsigned newval = MI->getOperand(3).getReg();
3739 unsigned scratch = BB->getParent()->getRegInfo()
3740 .createVirtualRegister(ARM::GPRRegisterClass);
3741 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3742 DebugLoc dl = MI->getDebugLoc();
3743 bool isThumb2 = Subtarget->isThumb2();
3745 unsigned ldrOpc, strOpc;
3747 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3749 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3750 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3753 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3754 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3757 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3758 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3762 MachineFunction *MF = BB->getParent();
3763 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3764 MachineFunction::iterator It = BB;
3765 ++It; // insert the new blocks after the current block
3767 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3768 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3769 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3770 MF->insert(It, loop1MBB);
3771 MF->insert(It, loop2MBB);
3772 MF->insert(It, exitMBB);
3774 // Transfer the remainder of BB and its successor edges to exitMBB.
3775 exitMBB->splice(exitMBB->begin(), BB,
3776 llvm::next(MachineBasicBlock::iterator(MI)),
3778 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3782 // fallthrough --> loop1MBB
3783 BB->addSuccessor(loop1MBB);
3786 // ldrex dest, [ptr]
3790 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3791 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3792 .addReg(dest).addReg(oldval));
3793 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3794 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3795 BB->addSuccessor(loop2MBB);
3796 BB->addSuccessor(exitMBB);
3799 // strex scratch, newval, [ptr]
3803 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3805 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3806 .addReg(scratch).addImm(0));
3807 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3808 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3809 BB->addSuccessor(loop1MBB);
3810 BB->addSuccessor(exitMBB);
3816 MI->eraseFromParent(); // The instruction is gone now.
3822 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3823 unsigned Size, unsigned BinOpcode) const {
3824 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3825 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3827 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3828 MachineFunction *MF = BB->getParent();
3829 MachineFunction::iterator It = BB;
3832 unsigned dest = MI->getOperand(0).getReg();
3833 unsigned ptr = MI->getOperand(1).getReg();
3834 unsigned incr = MI->getOperand(2).getReg();
3835 DebugLoc dl = MI->getDebugLoc();
3837 bool isThumb2 = Subtarget->isThumb2();
3838 unsigned ldrOpc, strOpc;
3840 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3842 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3843 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3846 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3847 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3850 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3851 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3855 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3856 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3857 MF->insert(It, loopMBB);
3858 MF->insert(It, exitMBB);
3860 // Transfer the remainder of BB and its successor edges to exitMBB.
3861 exitMBB->splice(exitMBB->begin(), BB,
3862 llvm::next(MachineBasicBlock::iterator(MI)),
3864 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3866 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3867 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3868 unsigned scratch2 = (!BinOpcode) ? incr :
3869 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3873 // fallthrough --> loopMBB
3874 BB->addSuccessor(loopMBB);
3878 // <binop> scratch2, dest, incr
3879 // strex scratch, scratch2, ptr
3882 // fallthrough --> exitMBB
3884 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3886 // operand order needs to go the other way for NAND
3887 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3888 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3889 addReg(incr).addReg(dest)).addReg(0);
3891 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3892 addReg(dest).addReg(incr)).addReg(0);
3895 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3897 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3898 .addReg(scratch).addImm(0));
3899 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3900 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3902 BB->addSuccessor(loopMBB);
3903 BB->addSuccessor(exitMBB);
3909 MI->eraseFromParent(); // The instruction is gone now.
3915 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3916 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3917 E = MBB->succ_end(); I != E; ++I)
3920 llvm_unreachable("Expecting a BB with two successors!");
3924 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3925 MachineBasicBlock *BB) const {
3926 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3927 DebugLoc dl = MI->getDebugLoc();
3928 bool isThumb2 = Subtarget->isThumb2();
3929 switch (MI->getOpcode()) {
3932 llvm_unreachable("Unexpected instr type to insert");
3934 case ARM::ATOMIC_LOAD_ADD_I8:
3935 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3936 case ARM::ATOMIC_LOAD_ADD_I16:
3937 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3938 case ARM::ATOMIC_LOAD_ADD_I32:
3939 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3941 case ARM::ATOMIC_LOAD_AND_I8:
3942 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3943 case ARM::ATOMIC_LOAD_AND_I16:
3944 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3945 case ARM::ATOMIC_LOAD_AND_I32:
3946 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3948 case ARM::ATOMIC_LOAD_OR_I8:
3949 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3950 case ARM::ATOMIC_LOAD_OR_I16:
3951 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3952 case ARM::ATOMIC_LOAD_OR_I32:
3953 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3955 case ARM::ATOMIC_LOAD_XOR_I8:
3956 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3957 case ARM::ATOMIC_LOAD_XOR_I16:
3958 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3959 case ARM::ATOMIC_LOAD_XOR_I32:
3960 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3962 case ARM::ATOMIC_LOAD_NAND_I8:
3963 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3964 case ARM::ATOMIC_LOAD_NAND_I16:
3965 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3966 case ARM::ATOMIC_LOAD_NAND_I32:
3967 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3969 case ARM::ATOMIC_LOAD_SUB_I8:
3970 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3971 case ARM::ATOMIC_LOAD_SUB_I16:
3972 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3973 case ARM::ATOMIC_LOAD_SUB_I32:
3974 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3976 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3977 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3978 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3980 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3981 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3982 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3984 case ARM::tMOVCCr_pseudo: {
3985 // To "insert" a SELECT_CC instruction, we actually have to insert the
3986 // diamond control-flow pattern. The incoming instruction knows the
3987 // destination vreg to set, the condition code register to branch on, the
3988 // true/false values to select between, and a branch opcode to use.
3989 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3990 MachineFunction::iterator It = BB;
3996 // cmpTY ccX, r1, r2
3998 // fallthrough --> copy0MBB
3999 MachineBasicBlock *thisMBB = BB;
4000 MachineFunction *F = BB->getParent();
4001 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4002 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4003 F->insert(It, copy0MBB);
4004 F->insert(It, sinkMBB);
4006 // Transfer the remainder of BB and its successor edges to sinkMBB.
4007 sinkMBB->splice(sinkMBB->begin(), BB,
4008 llvm::next(MachineBasicBlock::iterator(MI)),
4010 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4012 BB->addSuccessor(copy0MBB);
4013 BB->addSuccessor(sinkMBB);
4015 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
4016 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
4019 // %FalseValue = ...
4020 // # fallthrough to sinkMBB
4023 // Update machine-CFG edges
4024 BB->addSuccessor(sinkMBB);
4027 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4030 BuildMI(*BB, BB->begin(), dl,
4031 TII->get(ARM::PHI), MI->getOperand(0).getReg())
4032 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4033 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4035 MI->eraseFromParent(); // The pseudo instruction is gone now.
4040 case ARM::BCCZi64: {
4041 // Compare both parts that make up the double comparison separately for
4043 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
4045 unsigned LHS1 = MI->getOperand(1).getReg();
4046 unsigned LHS2 = MI->getOperand(2).getReg();
4048 AddDefaultPred(BuildMI(BB, dl,
4049 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4050 .addReg(LHS1).addImm(0));
4051 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
4052 .addReg(LHS2).addImm(0)
4053 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4055 unsigned RHS1 = MI->getOperand(3).getReg();
4056 unsigned RHS2 = MI->getOperand(4).getReg();
4057 AddDefaultPred(BuildMI(BB, dl,
4058 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4059 .addReg(LHS1).addReg(RHS1));
4060 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
4061 .addReg(LHS2).addReg(RHS2)
4062 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
4065 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
4066 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
4067 if (MI->getOperand(0).getImm() == ARMCC::NE)
4068 std::swap(destMBB, exitMBB);
4070 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
4071 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
4072 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4075 MI->eraseFromParent(); // The pseudo instruction is gone now.
4081 //===----------------------------------------------------------------------===//
4082 // ARM Optimization Hooks
4083 //===----------------------------------------------------------------------===//
4086 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4087 TargetLowering::DAGCombinerInfo &DCI) {
4088 SelectionDAG &DAG = DCI.DAG;
4089 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4090 EVT VT = N->getValueType(0);
4091 unsigned Opc = N->getOpcode();
4092 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4093 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4094 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4095 ISD::CondCode CC = ISD::SETCC_INVALID;
4098 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4100 SDValue CCOp = Slct.getOperand(0);
4101 if (CCOp.getOpcode() == ISD::SETCC)
4102 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4105 bool DoXform = false;
4107 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4110 if (LHS.getOpcode() == ISD::Constant &&
4111 cast<ConstantSDNode>(LHS)->isNullValue()) {
4113 } else if (CC != ISD::SETCC_INVALID &&
4114 RHS.getOpcode() == ISD::Constant &&
4115 cast<ConstantSDNode>(RHS)->isNullValue()) {
4116 std::swap(LHS, RHS);
4117 SDValue Op0 = Slct.getOperand(0);
4118 EVT OpVT = isSlctCC ? Op0.getValueType() :
4119 Op0.getOperand(0).getValueType();
4120 bool isInt = OpVT.isInteger();
4121 CC = ISD::getSetCCInverse(CC, isInt);
4123 if (!TLI.isCondCodeLegal(CC, OpVT))
4124 return SDValue(); // Inverse operator isn't legal.
4131 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4133 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4134 Slct.getOperand(0), Slct.getOperand(1), CC);
4135 SDValue CCOp = Slct.getOperand(0);
4137 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4138 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4139 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4140 CCOp, OtherOp, Result);
4145 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
4146 /// operands N0 and N1. This is a helper for PerformADDCombine that is
4147 /// called with the default operands, and if that fails, with commuted
4149 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
4150 TargetLowering::DAGCombinerInfo &DCI) {
4151 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4152 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4153 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4154 if (Result.getNode()) return Result;
4159 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4161 static SDValue PerformADDCombine(SDNode *N,
4162 TargetLowering::DAGCombinerInfo &DCI) {
4163 SDValue N0 = N->getOperand(0);
4164 SDValue N1 = N->getOperand(1);
4166 // First try with the default operand order.
4167 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI);
4168 if (Result.getNode())
4171 // If that didn't work, try again with the operands commuted.
4172 return PerformADDCombineWithOperands(N, N1, N0, DCI);
4175 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4177 static SDValue PerformSUBCombine(SDNode *N,
4178 TargetLowering::DAGCombinerInfo &DCI) {
4179 SDValue N0 = N->getOperand(0);
4180 SDValue N1 = N->getOperand(1);
4182 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4183 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4184 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4185 if (Result.getNode()) return Result;
4191 static SDValue PerformMULCombine(SDNode *N,
4192 TargetLowering::DAGCombinerInfo &DCI,
4193 const ARMSubtarget *Subtarget) {
4194 SelectionDAG &DAG = DCI.DAG;
4196 if (Subtarget->isThumb1Only())
4199 if (DAG.getMachineFunction().
4200 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4203 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4206 EVT VT = N->getValueType(0);
4210 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4214 uint64_t MulAmt = C->getZExtValue();
4215 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4216 ShiftAmt = ShiftAmt & (32 - 1);
4217 SDValue V = N->getOperand(0);
4218 DebugLoc DL = N->getDebugLoc();
4221 MulAmt >>= ShiftAmt;
4222 if (isPowerOf2_32(MulAmt - 1)) {
4223 // (mul x, 2^N + 1) => (add (shl x, N), x)
4224 Res = DAG.getNode(ISD::ADD, DL, VT,
4225 V, DAG.getNode(ISD::SHL, DL, VT,
4226 V, DAG.getConstant(Log2_32(MulAmt-1),
4228 } else if (isPowerOf2_32(MulAmt + 1)) {
4229 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4230 Res = DAG.getNode(ISD::SUB, DL, VT,
4231 DAG.getNode(ISD::SHL, DL, VT,
4232 V, DAG.getConstant(Log2_32(MulAmt+1),
4239 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4240 DAG.getConstant(ShiftAmt, MVT::i32));
4242 // Do not add new nodes to DAG combiner worklist.
4243 DCI.CombineTo(N, Res, false);
4247 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4248 static SDValue PerformORCombine(SDNode *N,
4249 TargetLowering::DAGCombinerInfo &DCI,
4250 const ARMSubtarget *Subtarget) {
4251 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4254 // BFI is only available on V6T2+
4255 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4258 SelectionDAG &DAG = DCI.DAG;
4259 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4260 DebugLoc DL = N->getDebugLoc();
4261 // 1) or (and A, mask), val => ARMbfi A, val, mask
4262 // iff (val & mask) == val
4264 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4265 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4266 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4267 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4268 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4269 // (i.e., copy a bitfield value into another bitfield of the same width)
4270 if (N0.getOpcode() != ISD::AND)
4273 EVT VT = N->getValueType(0);
4278 // The value and the mask need to be constants so we can verify this is
4279 // actually a bitfield set. If the mask is 0xffff, we can do better
4280 // via a movt instruction, so don't use BFI in that case.
4281 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4284 unsigned Mask = C->getZExtValue();
4288 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4289 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4290 unsigned Val = C->getZExtValue();
4291 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4293 Val >>= CountTrailingZeros_32(~Mask);
4295 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4296 DAG.getConstant(Val, MVT::i32),
4297 DAG.getConstant(Mask, MVT::i32));
4299 // Do not add new nodes to DAG combiner worklist.
4300 DCI.CombineTo(N, Res, false);
4301 } else if (N1.getOpcode() == ISD::AND) {
4302 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4303 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4306 unsigned Mask2 = C->getZExtValue();
4308 if (ARM::isBitFieldInvertedMask(Mask) &&
4309 ARM::isBitFieldInvertedMask(~Mask2) &&
4310 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4311 // The pack halfword instruction works better for masks that fit it,
4312 // so use that when it's available.
4313 if (Subtarget->hasT2ExtractPack() &&
4314 (Mask == 0xffff || Mask == 0xffff0000))
4317 unsigned lsb = CountTrailingZeros_32(Mask2);
4318 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4319 DAG.getConstant(lsb, MVT::i32));
4320 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4321 DAG.getConstant(Mask, MVT::i32));
4322 // Do not add new nodes to DAG combiner worklist.
4323 DCI.CombineTo(N, Res, false);
4324 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4325 ARM::isBitFieldInvertedMask(Mask2) &&
4326 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4327 // The pack halfword instruction works better for masks that fit it,
4328 // so use that when it's available.
4329 if (Subtarget->hasT2ExtractPack() &&
4330 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4333 unsigned lsb = CountTrailingZeros_32(Mask);
4334 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4335 DAG.getConstant(lsb, MVT::i32));
4336 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4337 DAG.getConstant(Mask2, MVT::i32));
4338 // Do not add new nodes to DAG combiner worklist.
4339 DCI.CombineTo(N, Res, false);
4346 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
4347 /// ISD::BUILD_VECTOR.
4348 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
4349 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
4350 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
4351 // into a pair of GPRs, which is fine when the value is used as a scalar,
4352 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
4353 if (N->getNumOperands() == 2) {
4354 SDValue Op0 = N->getOperand(0);
4355 SDValue Op1 = N->getOperand(1);
4356 if (Op0.getOpcode() == ISD::BIT_CONVERT)
4357 Op0 = Op0.getOperand(0);
4358 if (Op1.getOpcode() == ISD::BIT_CONVERT)
4359 Op1 = Op1.getOperand(0);
4360 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
4361 Op0.getNode() == Op1.getNode() &&
4362 Op0.getResNo() == 0 && Op1.getResNo() == 1) {
4363 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4364 N->getValueType(0), Op0.getOperand(0));
4371 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4372 /// ARMISD::VMOVRRD.
4373 static SDValue PerformVMOVRRDCombine(SDNode *N,
4374 TargetLowering::DAGCombinerInfo &DCI) {
4375 // fmrrd(fmdrr x, y) -> x,y
4376 SDValue InDouble = N->getOperand(0);
4377 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4378 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4382 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4383 /// ARMISD::VDUPLANE.
4384 static SDValue PerformVDUPLANECombine(SDNode *N,
4385 TargetLowering::DAGCombinerInfo &DCI) {
4386 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4388 SDValue Op = N->getOperand(0);
4389 EVT VT = N->getValueType(0);
4391 // Ignore bit_converts.
4392 while (Op.getOpcode() == ISD::BIT_CONVERT)
4393 Op = Op.getOperand(0);
4394 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4397 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4398 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4399 // The canonical VMOV for a zero vector uses a 32-bit element size.
4400 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4402 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4404 if (EltSize > VT.getVectorElementType().getSizeInBits())
4407 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4408 return DCI.CombineTo(N, Res, false);
4411 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4412 /// operand of a vector shift operation, where all the elements of the
4413 /// build_vector must have the same constant integer value.
4414 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4415 // Ignore bit_converts.
4416 while (Op.getOpcode() == ISD::BIT_CONVERT)
4417 Op = Op.getOperand(0);
4418 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4419 APInt SplatBits, SplatUndef;
4420 unsigned SplatBitSize;
4422 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4423 HasAnyUndefs, ElementBits) ||
4424 SplatBitSize > ElementBits)
4426 Cnt = SplatBits.getSExtValue();
4430 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4431 /// operand of a vector shift left operation. That value must be in the range:
4432 /// 0 <= Value < ElementBits for a left shift; or
4433 /// 0 <= Value <= ElementBits for a long left shift.
4434 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4435 assert(VT.isVector() && "vector shift count is not a vector type");
4436 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4437 if (! getVShiftImm(Op, ElementBits, Cnt))
4439 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4442 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4443 /// operand of a vector shift right operation. For a shift opcode, the value
4444 /// is positive, but for an intrinsic the value count must be negative. The
4445 /// absolute value must be in the range:
4446 /// 1 <= |Value| <= ElementBits for a right shift; or
4447 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4448 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4450 assert(VT.isVector() && "vector shift count is not a vector type");
4451 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4452 if (! getVShiftImm(Op, ElementBits, Cnt))
4456 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4459 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4460 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4461 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4464 // Don't do anything for most intrinsics.
4467 // Vector shifts: check for immediate versions and lower them.
4468 // Note: This is done during DAG combining instead of DAG legalizing because
4469 // the build_vectors for 64-bit vector element shift counts are generally
4470 // not legal, and it is hard to see their values after they get legalized to
4471 // loads from a constant pool.
4472 case Intrinsic::arm_neon_vshifts:
4473 case Intrinsic::arm_neon_vshiftu:
4474 case Intrinsic::arm_neon_vshiftls:
4475 case Intrinsic::arm_neon_vshiftlu:
4476 case Intrinsic::arm_neon_vshiftn:
4477 case Intrinsic::arm_neon_vrshifts:
4478 case Intrinsic::arm_neon_vrshiftu:
4479 case Intrinsic::arm_neon_vrshiftn:
4480 case Intrinsic::arm_neon_vqshifts:
4481 case Intrinsic::arm_neon_vqshiftu:
4482 case Intrinsic::arm_neon_vqshiftsu:
4483 case Intrinsic::arm_neon_vqshiftns:
4484 case Intrinsic::arm_neon_vqshiftnu:
4485 case Intrinsic::arm_neon_vqshiftnsu:
4486 case Intrinsic::arm_neon_vqrshiftns:
4487 case Intrinsic::arm_neon_vqrshiftnu:
4488 case Intrinsic::arm_neon_vqrshiftnsu: {
4489 EVT VT = N->getOperand(1).getValueType();
4491 unsigned VShiftOpc = 0;
4494 case Intrinsic::arm_neon_vshifts:
4495 case Intrinsic::arm_neon_vshiftu:
4496 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4497 VShiftOpc = ARMISD::VSHL;
4500 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4501 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4502 ARMISD::VSHRs : ARMISD::VSHRu);
4507 case Intrinsic::arm_neon_vshiftls:
4508 case Intrinsic::arm_neon_vshiftlu:
4509 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4511 llvm_unreachable("invalid shift count for vshll intrinsic");
4513 case Intrinsic::arm_neon_vrshifts:
4514 case Intrinsic::arm_neon_vrshiftu:
4515 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4519 case Intrinsic::arm_neon_vqshifts:
4520 case Intrinsic::arm_neon_vqshiftu:
4521 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4525 case Intrinsic::arm_neon_vqshiftsu:
4526 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4528 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4530 case Intrinsic::arm_neon_vshiftn:
4531 case Intrinsic::arm_neon_vrshiftn:
4532 case Intrinsic::arm_neon_vqshiftns:
4533 case Intrinsic::arm_neon_vqshiftnu:
4534 case Intrinsic::arm_neon_vqshiftnsu:
4535 case Intrinsic::arm_neon_vqrshiftns:
4536 case Intrinsic::arm_neon_vqrshiftnu:
4537 case Intrinsic::arm_neon_vqrshiftnsu:
4538 // Narrowing shifts require an immediate right shift.
4539 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4541 llvm_unreachable("invalid shift count for narrowing vector shift "
4545 llvm_unreachable("unhandled vector shift");
4549 case Intrinsic::arm_neon_vshifts:
4550 case Intrinsic::arm_neon_vshiftu:
4551 // Opcode already set above.
4553 case Intrinsic::arm_neon_vshiftls:
4554 case Intrinsic::arm_neon_vshiftlu:
4555 if (Cnt == VT.getVectorElementType().getSizeInBits())
4556 VShiftOpc = ARMISD::VSHLLi;
4558 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4559 ARMISD::VSHLLs : ARMISD::VSHLLu);
4561 case Intrinsic::arm_neon_vshiftn:
4562 VShiftOpc = ARMISD::VSHRN; break;
4563 case Intrinsic::arm_neon_vrshifts:
4564 VShiftOpc = ARMISD::VRSHRs; break;
4565 case Intrinsic::arm_neon_vrshiftu:
4566 VShiftOpc = ARMISD::VRSHRu; break;
4567 case Intrinsic::arm_neon_vrshiftn:
4568 VShiftOpc = ARMISD::VRSHRN; break;
4569 case Intrinsic::arm_neon_vqshifts:
4570 VShiftOpc = ARMISD::VQSHLs; break;
4571 case Intrinsic::arm_neon_vqshiftu:
4572 VShiftOpc = ARMISD::VQSHLu; break;
4573 case Intrinsic::arm_neon_vqshiftsu:
4574 VShiftOpc = ARMISD::VQSHLsu; break;
4575 case Intrinsic::arm_neon_vqshiftns:
4576 VShiftOpc = ARMISD::VQSHRNs; break;
4577 case Intrinsic::arm_neon_vqshiftnu:
4578 VShiftOpc = ARMISD::VQSHRNu; break;
4579 case Intrinsic::arm_neon_vqshiftnsu:
4580 VShiftOpc = ARMISD::VQSHRNsu; break;
4581 case Intrinsic::arm_neon_vqrshiftns:
4582 VShiftOpc = ARMISD::VQRSHRNs; break;
4583 case Intrinsic::arm_neon_vqrshiftnu:
4584 VShiftOpc = ARMISD::VQRSHRNu; break;
4585 case Intrinsic::arm_neon_vqrshiftnsu:
4586 VShiftOpc = ARMISD::VQRSHRNsu; break;
4589 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4590 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4593 case Intrinsic::arm_neon_vshiftins: {
4594 EVT VT = N->getOperand(1).getValueType();
4596 unsigned VShiftOpc = 0;
4598 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4599 VShiftOpc = ARMISD::VSLI;
4600 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4601 VShiftOpc = ARMISD::VSRI;
4603 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4606 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4607 N->getOperand(1), N->getOperand(2),
4608 DAG.getConstant(Cnt, MVT::i32));
4611 case Intrinsic::arm_neon_vqrshifts:
4612 case Intrinsic::arm_neon_vqrshiftu:
4613 // No immediate versions of these to check for.
4620 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4621 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4622 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4623 /// vector element shift counts are generally not legal, and it is hard to see
4624 /// their values after they get legalized to loads from a constant pool.
4625 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4626 const ARMSubtarget *ST) {
4627 EVT VT = N->getValueType(0);
4629 // Nothing to be done for scalar shifts.
4630 if (! VT.isVector())
4633 assert(ST->hasNEON() && "unexpected vector shift");
4636 switch (N->getOpcode()) {
4637 default: llvm_unreachable("unexpected shift opcode");
4640 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4641 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4642 DAG.getConstant(Cnt, MVT::i32));
4647 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4648 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4649 ARMISD::VSHRs : ARMISD::VSHRu);
4650 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4651 DAG.getConstant(Cnt, MVT::i32));
4657 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4658 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4659 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4660 const ARMSubtarget *ST) {
4661 SDValue N0 = N->getOperand(0);
4663 // Check for sign- and zero-extensions of vector extract operations of 8-
4664 // and 16-bit vector elements. NEON supports these directly. They are
4665 // handled during DAG combining because type legalization will promote them
4666 // to 32-bit types and it is messy to recognize the operations after that.
4667 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4668 SDValue Vec = N0.getOperand(0);
4669 SDValue Lane = N0.getOperand(1);
4670 EVT VT = N->getValueType(0);
4671 EVT EltVT = N0.getValueType();
4672 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4674 if (VT == MVT::i32 &&
4675 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4676 TLI.isTypeLegal(Vec.getValueType())) {
4679 switch (N->getOpcode()) {
4680 default: llvm_unreachable("unexpected opcode");
4681 case ISD::SIGN_EXTEND:
4682 Opc = ARMISD::VGETLANEs;
4684 case ISD::ZERO_EXTEND:
4685 case ISD::ANY_EXTEND:
4686 Opc = ARMISD::VGETLANEu;
4689 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4696 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4697 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4698 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4699 const ARMSubtarget *ST) {
4700 // If the target supports NEON, try to use vmax/vmin instructions for f32
4701 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4702 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4703 // a NaN; only do the transformation when it matches that behavior.
4705 // For now only do this when using NEON for FP operations; if using VFP, it
4706 // is not obvious that the benefit outweighs the cost of switching to the
4708 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4709 N->getValueType(0) != MVT::f32)
4712 SDValue CondLHS = N->getOperand(0);
4713 SDValue CondRHS = N->getOperand(1);
4714 SDValue LHS = N->getOperand(2);
4715 SDValue RHS = N->getOperand(3);
4716 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4718 unsigned Opcode = 0;
4720 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4721 IsReversed = false; // x CC y ? x : y
4722 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4723 IsReversed = true ; // x CC y ? y : x
4737 // If LHS is NaN, an ordered comparison will be false and the result will
4738 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4739 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4740 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4741 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4743 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4744 // will return -0, so vmin can only be used for unsafe math or if one of
4745 // the operands is known to be nonzero.
4746 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4748 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4750 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4759 // If LHS is NaN, an ordered comparison will be false and the result will
4760 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4761 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4762 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4763 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4765 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4766 // will return +0, so vmax can only be used for unsafe math or if one of
4767 // the operands is known to be nonzero.
4768 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4770 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4772 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4778 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4781 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4782 DAGCombinerInfo &DCI) const {
4783 switch (N->getOpcode()) {
4785 case ISD::ADD: return PerformADDCombine(N, DCI);
4786 case ISD::SUB: return PerformSUBCombine(N, DCI);
4787 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4788 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4789 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
4790 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4791 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4792 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4795 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4796 case ISD::SIGN_EXTEND:
4797 case ISD::ZERO_EXTEND:
4798 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4799 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4804 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4805 if (!Subtarget->hasV6Ops())
4806 // Pre-v6 does not support unaligned mem access.
4809 // v6+ may or may not support unaligned mem access depending on the system
4811 // FIXME: This is pretty conservative. Should we provide cmdline option to
4812 // control the behaviour?
4813 if (!Subtarget->isTargetDarwin())
4816 switch (VT.getSimpleVT().SimpleTy) {
4823 // FIXME: VLD1 etc with standard alignment is legal.
4827 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4832 switch (VT.getSimpleVT().SimpleTy) {
4833 default: return false;
4848 if ((V & (Scale - 1)) != 0)
4851 return V == (V & ((1LL << 5) - 1));
4854 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4855 const ARMSubtarget *Subtarget) {
4862 switch (VT.getSimpleVT().SimpleTy) {
4863 default: return false;
4868 // + imm12 or - imm8
4870 return V == (V & ((1LL << 8) - 1));
4871 return V == (V & ((1LL << 12) - 1));
4874 // Same as ARM mode. FIXME: NEON?
4875 if (!Subtarget->hasVFP2())
4880 return V == (V & ((1LL << 8) - 1));
4884 /// isLegalAddressImmediate - Return true if the integer value can be used
4885 /// as the offset of the target addressing mode for load / store of the
4887 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4888 const ARMSubtarget *Subtarget) {
4895 if (Subtarget->isThumb1Only())
4896 return isLegalT1AddressImmediate(V, VT);
4897 else if (Subtarget->isThumb2())
4898 return isLegalT2AddressImmediate(V, VT, Subtarget);
4903 switch (VT.getSimpleVT().SimpleTy) {
4904 default: return false;
4909 return V == (V & ((1LL << 12) - 1));
4912 return V == (V & ((1LL << 8) - 1));
4915 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4920 return V == (V & ((1LL << 8) - 1));
4924 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4926 int Scale = AM.Scale;
4930 switch (VT.getSimpleVT().SimpleTy) {
4931 default: return false;
4940 return Scale == 2 || Scale == 4 || Scale == 8;
4943 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4947 // Note, we allow "void" uses (basically, uses that aren't loads or
4948 // stores), because arm allows folding a scale into many arithmetic
4949 // operations. This should be made more precise and revisited later.
4951 // Allow r << imm, but the imm has to be a multiple of two.
4952 if (Scale & 1) return false;
4953 return isPowerOf2_32(Scale);
4957 /// isLegalAddressingMode - Return true if the addressing mode represented
4958 /// by AM is legal for this target, for a load/store of the specified type.
4959 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4960 const Type *Ty) const {
4961 EVT VT = getValueType(Ty, true);
4962 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4965 // Can never fold addr of global into load/store.
4970 case 0: // no scale reg, must be "r+i" or "r", or "i".
4973 if (Subtarget->isThumb1Only())
4977 // ARM doesn't support any R+R*scale+imm addr modes.
4984 if (Subtarget->isThumb2())
4985 return isLegalT2ScaledAddressingMode(AM, VT);
4987 int Scale = AM.Scale;
4988 switch (VT.getSimpleVT().SimpleTy) {
4989 default: return false;
4993 if (Scale < 0) Scale = -Scale;
4997 return isPowerOf2_32(Scale & ~1);
5001 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
5006 // Note, we allow "void" uses (basically, uses that aren't loads or
5007 // stores), because arm allows folding a scale into many arithmetic
5008 // operations. This should be made more precise and revisited later.
5010 // Allow r << imm, but the imm has to be a multiple of two.
5011 if (Scale & 1) return false;
5012 return isPowerOf2_32(Scale);
5019 /// isLegalICmpImmediate - Return true if the specified immediate is legal
5020 /// icmp immediate, that is the target has icmp instructions which can compare
5021 /// a register against the immediate without having to materialize the
5022 /// immediate into a register.
5023 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
5024 if (!Subtarget->isThumb())
5025 return ARM_AM::getSOImmVal(Imm) != -1;
5026 if (Subtarget->isThumb2())
5027 return ARM_AM::getT2SOImmVal(Imm) != -1;
5028 return Imm >= 0 && Imm <= 255;
5031 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
5032 bool isSEXTLoad, SDValue &Base,
5033 SDValue &Offset, bool &isInc,
5034 SelectionDAG &DAG) {
5035 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5038 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
5040 Base = Ptr->getOperand(0);
5041 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5042 int RHSC = (int)RHS->getZExtValue();
5043 if (RHSC < 0 && RHSC > -256) {
5044 assert(Ptr->getOpcode() == ISD::ADD);
5046 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5050 isInc = (Ptr->getOpcode() == ISD::ADD);
5051 Offset = Ptr->getOperand(1);
5053 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5055 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5056 int RHSC = (int)RHS->getZExtValue();
5057 if (RHSC < 0 && RHSC > -0x1000) {
5058 assert(Ptr->getOpcode() == ISD::ADD);
5060 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5061 Base = Ptr->getOperand(0);
5066 if (Ptr->getOpcode() == ISD::ADD) {
5068 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5069 if (ShOpcVal != ARM_AM::no_shift) {
5070 Base = Ptr->getOperand(1);
5071 Offset = Ptr->getOperand(0);
5073 Base = Ptr->getOperand(0);
5074 Offset = Ptr->getOperand(1);
5079 isInc = (Ptr->getOpcode() == ISD::ADD);
5080 Base = Ptr->getOperand(0);
5081 Offset = Ptr->getOperand(1);
5085 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5089 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5090 bool isSEXTLoad, SDValue &Base,
5091 SDValue &Offset, bool &isInc,
5092 SelectionDAG &DAG) {
5093 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5096 Base = Ptr->getOperand(0);
5097 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5098 int RHSC = (int)RHS->getZExtValue();
5099 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5100 assert(Ptr->getOpcode() == ISD::ADD);
5102 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5104 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5105 isInc = Ptr->getOpcode() == ISD::ADD;
5106 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5114 /// getPreIndexedAddressParts - returns true by value, base pointer and
5115 /// offset pointer and addressing mode by reference if the node's address
5116 /// can be legally represented as pre-indexed load / store address.
5118 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5120 ISD::MemIndexedMode &AM,
5121 SelectionDAG &DAG) const {
5122 if (Subtarget->isThumb1Only())
5127 bool isSEXTLoad = false;
5128 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5129 Ptr = LD->getBasePtr();
5130 VT = LD->getMemoryVT();
5131 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5132 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5133 Ptr = ST->getBasePtr();
5134 VT = ST->getMemoryVT();
5139 bool isLegal = false;
5140 if (Subtarget->isThumb2())
5141 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5142 Offset, isInc, DAG);
5144 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5145 Offset, isInc, DAG);
5149 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5153 /// getPostIndexedAddressParts - returns true by value, base pointer and
5154 /// offset pointer and addressing mode by reference if this node can be
5155 /// combined with a load / store to form a post-indexed load / store.
5156 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5159 ISD::MemIndexedMode &AM,
5160 SelectionDAG &DAG) const {
5161 if (Subtarget->isThumb1Only())
5166 bool isSEXTLoad = false;
5167 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5168 VT = LD->getMemoryVT();
5169 Ptr = LD->getBasePtr();
5170 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5171 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5172 VT = ST->getMemoryVT();
5173 Ptr = ST->getBasePtr();
5178 bool isLegal = false;
5179 if (Subtarget->isThumb2())
5180 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5183 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5189 // Swap base ptr and offset to catch more post-index load / store when
5190 // it's legal. In Thumb2 mode, offset must be an immediate.
5191 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5192 !Subtarget->isThumb2())
5193 std::swap(Base, Offset);
5195 // Post-indexed load / store update the base pointer.
5200 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5204 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5208 const SelectionDAG &DAG,
5209 unsigned Depth) const {
5210 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5211 switch (Op.getOpcode()) {
5213 case ARMISD::CMOV: {
5214 // Bits are known zero/one if known on the LHS and RHS.
5215 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5216 if (KnownZero == 0 && KnownOne == 0) return;
5218 APInt KnownZeroRHS, KnownOneRHS;
5219 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5220 KnownZeroRHS, KnownOneRHS, Depth+1);
5221 KnownZero &= KnownZeroRHS;
5222 KnownOne &= KnownOneRHS;
5228 //===----------------------------------------------------------------------===//
5229 // ARM Inline Assembly Support
5230 //===----------------------------------------------------------------------===//
5232 /// getConstraintType - Given a constraint letter, return the type of
5233 /// constraint it is for this target.
5234 ARMTargetLowering::ConstraintType
5235 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5236 if (Constraint.size() == 1) {
5237 switch (Constraint[0]) {
5239 case 'l': return C_RegisterClass;
5240 case 'w': return C_RegisterClass;
5243 return TargetLowering::getConstraintType(Constraint);
5246 std::pair<unsigned, const TargetRegisterClass*>
5247 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5249 if (Constraint.size() == 1) {
5250 // GCC ARM Constraint Letters
5251 switch (Constraint[0]) {
5253 if (Subtarget->isThumb())
5254 return std::make_pair(0U, ARM::tGPRRegisterClass);
5256 return std::make_pair(0U, ARM::GPRRegisterClass);
5258 return std::make_pair(0U, ARM::GPRRegisterClass);
5261 return std::make_pair(0U, ARM::SPRRegisterClass);
5262 if (VT.getSizeInBits() == 64)
5263 return std::make_pair(0U, ARM::DPRRegisterClass);
5264 if (VT.getSizeInBits() == 128)
5265 return std::make_pair(0U, ARM::QPRRegisterClass);
5269 if (StringRef("{cc}").equals_lower(Constraint))
5270 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5272 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5275 std::vector<unsigned> ARMTargetLowering::
5276 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5278 if (Constraint.size() != 1)
5279 return std::vector<unsigned>();
5281 switch (Constraint[0]) { // GCC ARM Constraint Letters
5284 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5285 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5288 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5289 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5290 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5291 ARM::R12, ARM::LR, 0);
5294 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5295 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5296 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5297 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5298 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5299 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5300 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5301 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5302 if (VT.getSizeInBits() == 64)
5303 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5304 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5305 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5306 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5307 if (VT.getSizeInBits() == 128)
5308 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5309 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5313 return std::vector<unsigned>();
5316 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5317 /// vector. If it is invalid, don't add anything to Ops.
5318 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5320 std::vector<SDValue>&Ops,
5321 SelectionDAG &DAG) const {
5322 SDValue Result(0, 0);
5324 switch (Constraint) {
5326 case 'I': case 'J': case 'K': case 'L':
5327 case 'M': case 'N': case 'O':
5328 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5332 int64_t CVal64 = C->getSExtValue();
5333 int CVal = (int) CVal64;
5334 // None of these constraints allow values larger than 32 bits. Check
5335 // that the value fits in an int.
5339 switch (Constraint) {
5341 if (Subtarget->isThumb1Only()) {
5342 // This must be a constant between 0 and 255, for ADD
5344 if (CVal >= 0 && CVal <= 255)
5346 } else if (Subtarget->isThumb2()) {
5347 // A constant that can be used as an immediate value in a
5348 // data-processing instruction.
5349 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5352 // A constant that can be used as an immediate value in a
5353 // data-processing instruction.
5354 if (ARM_AM::getSOImmVal(CVal) != -1)
5360 if (Subtarget->isThumb()) { // FIXME thumb2
5361 // This must be a constant between -255 and -1, for negated ADD
5362 // immediates. This can be used in GCC with an "n" modifier that
5363 // prints the negated value, for use with SUB instructions. It is
5364 // not useful otherwise but is implemented for compatibility.
5365 if (CVal >= -255 && CVal <= -1)
5368 // This must be a constant between -4095 and 4095. It is not clear
5369 // what this constraint is intended for. Implemented for
5370 // compatibility with GCC.
5371 if (CVal >= -4095 && CVal <= 4095)
5377 if (Subtarget->isThumb1Only()) {
5378 // A 32-bit value where only one byte has a nonzero value. Exclude
5379 // zero to match GCC. This constraint is used by GCC internally for
5380 // constants that can be loaded with a move/shift combination.
5381 // It is not useful otherwise but is implemented for compatibility.
5382 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5384 } else if (Subtarget->isThumb2()) {
5385 // A constant whose bitwise inverse can be used as an immediate
5386 // value in a data-processing instruction. This can be used in GCC
5387 // with a "B" modifier that prints the inverted value, for use with
5388 // BIC and MVN instructions. It is not useful otherwise but is
5389 // implemented for compatibility.
5390 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5393 // A constant whose bitwise inverse can be used as an immediate
5394 // value in a data-processing instruction. This can be used in GCC
5395 // with a "B" modifier that prints the inverted value, for use with
5396 // BIC and MVN instructions. It is not useful otherwise but is
5397 // implemented for compatibility.
5398 if (ARM_AM::getSOImmVal(~CVal) != -1)
5404 if (Subtarget->isThumb1Only()) {
5405 // This must be a constant between -7 and 7,
5406 // for 3-operand ADD/SUB immediate instructions.
5407 if (CVal >= -7 && CVal < 7)
5409 } else if (Subtarget->isThumb2()) {
5410 // A constant whose negation can be used as an immediate value in a
5411 // data-processing instruction. This can be used in GCC with an "n"
5412 // modifier that prints the negated value, for use with SUB
5413 // instructions. It is not useful otherwise but is implemented for
5415 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5418 // A constant whose negation can be used as an immediate value in a
5419 // data-processing instruction. This can be used in GCC with an "n"
5420 // modifier that prints the negated value, for use with SUB
5421 // instructions. It is not useful otherwise but is implemented for
5423 if (ARM_AM::getSOImmVal(-CVal) != -1)
5429 if (Subtarget->isThumb()) { // FIXME thumb2
5430 // This must be a multiple of 4 between 0 and 1020, for
5431 // ADD sp + immediate.
5432 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5435 // A power of two or a constant between 0 and 32. This is used in
5436 // GCC for the shift amount on shifted register operands, but it is
5437 // useful in general for any shift amounts.
5438 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5444 if (Subtarget->isThumb()) { // FIXME thumb2
5445 // This must be a constant between 0 and 31, for shift amounts.
5446 if (CVal >= 0 && CVal <= 31)
5452 if (Subtarget->isThumb()) { // FIXME thumb2
5453 // This must be a multiple of 4 between -508 and 508, for
5454 // ADD/SUB sp = sp + immediate.
5455 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5460 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5464 if (Result.getNode()) {
5465 Ops.push_back(Result);
5468 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5472 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5473 // The ARM target isn't yet aware of offsets.
5477 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5478 APInt Imm = FPImm.bitcastToAPInt();
5479 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5480 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5481 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5483 // We can handle 4 bits of mantissa.
5484 // mantissa = (16+UInt(e:f:g:h))/16.
5485 if (Mantissa & 0x7ffff)
5488 if ((Mantissa & 0xf) != Mantissa)
5491 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5492 if (Exp < -3 || Exp > 4)
5494 Exp = ((Exp+3) & 0x7) ^ 4;
5496 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5499 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5500 APInt Imm = FPImm.bitcastToAPInt();
5501 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5502 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5503 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5505 // We can handle 4 bits of mantissa.
5506 // mantissa = (16+UInt(e:f:g:h))/16.
5507 if (Mantissa & 0xffffffffffffLL)
5510 if ((Mantissa & 0xf) != Mantissa)
5513 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5514 if (Exp < -3 || Exp > 4)
5516 Exp = ((Exp+3) & 0x7) ^ 4;
5518 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5521 bool ARM::isBitFieldInvertedMask(unsigned v) {
5522 if (v == 0xffffffff)
5524 // there can be 1's on either or both "outsides", all the "inside"
5526 unsigned int lsb = 0, msb = 31;
5527 while (v & (1 << msb)) --msb;
5528 while (v & (1 << lsb)) ++lsb;
5529 for (unsigned int i = lsb; i <= msb; ++i) {
5536 /// isFPImmLegal - Returns true if the target can instruction select the
5537 /// specified FP immediate natively. If false, the legalizer will
5538 /// materialize the FP immediate as a load from a constant pool.
5539 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5540 if (!Subtarget->hasVFP3())
5543 return ARM::getVFPf32Imm(Imm) != -1;
5545 return ARM::getVFPf64Imm(Imm) != -1;