1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMAddressingModes.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Type.h"
33 #include "llvm/CodeGen/CallingConvLower.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineInstrBuilder.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/CodeGen/PseudoSourceValue.h"
40 #include "llvm/CodeGen/SelectionDAG.h"
41 #include "llvm/MC/MCSectionMachO.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/ADT/VectorExtras.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
52 STATISTIC(NumTailCalls, "Number of tail calls");
54 // This option should go away when tail calls fully work.
56 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
57 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions."),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
70 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
71 CCValAssign::LocInfo &LocInfo,
72 ISD::ArgFlagsTy &ArgFlags,
74 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
75 CCValAssign::LocInfo &LocInfo,
76 ISD::ArgFlagsTy &ArgFlags,
78 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
79 CCValAssign::LocInfo &LocInfo,
80 ISD::ArgFlagsTy &ArgFlags,
82 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
83 CCValAssign::LocInfo &LocInfo,
84 ISD::ArgFlagsTy &ArgFlags,
87 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
88 EVT PromotedBitwiseVT) {
89 if (VT != PromotedLdStVT) {
90 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
91 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
92 PromotedLdStVT.getSimpleVT());
94 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
95 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
96 PromotedLdStVT.getSimpleVT());
99 EVT ElemTy = VT.getVectorElementType();
100 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
101 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
102 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
103 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
104 if (ElemTy != MVT::i32) {
105 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
106 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
108 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
110 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
112 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
113 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
116 if (VT.isInteger()) {
117 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
119 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
122 // Promote all bit-wise operations.
123 if (VT.isInteger() && VT != PromotedBitwiseVT) {
124 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
125 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
126 PromotedBitwiseVT.getSimpleVT());
127 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
128 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
129 PromotedBitwiseVT.getSimpleVT());
130 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
131 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
132 PromotedBitwiseVT.getSimpleVT());
135 // Neon does not support vector divide/remainder operations.
136 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
137 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
138 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
139 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
141 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
144 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
145 addRegisterClass(VT, ARM::DPRRegisterClass);
146 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
149 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
150 addRegisterClass(VT, ARM::QPRRegisterClass);
151 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
154 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
155 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
156 return new TargetLoweringObjectFileMachO();
158 return new ARMElfTargetObjectFile();
161 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
162 : TargetLowering(TM, createTLOF(TM)) {
163 Subtarget = &TM.getSubtarget<ARMSubtarget>();
165 if (Subtarget->isTargetDarwin()) {
166 // Uses VFP for Thumb libfuncs if available.
167 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
168 // Single-precision floating-point arithmetic.
169 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
170 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
171 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
172 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
174 // Double-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
176 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
177 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
178 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
180 // Single-precision comparisons.
181 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
182 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
183 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
184 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
185 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
186 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
187 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
188 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
190 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
199 // Double-precision comparisons.
200 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
201 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
202 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
203 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
204 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
205 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
206 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
207 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
209 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
218 // Floating-point to integer conversions.
219 // i64 conversions are done via library routines even when generating VFP
220 // instructions, so use the same ones.
221 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
222 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
223 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
224 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
226 // Conversions between floating types.
227 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
228 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
230 // Integer to floating-point conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
234 // e.g., __floatunsidf vs. __floatunssidfvfp.
235 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
236 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
237 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
238 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
242 // These libcalls are not available in 32-bit.
243 setLibcallName(RTLIB::SHL_I128, 0);
244 setLibcallName(RTLIB::SRL_I128, 0);
245 setLibcallName(RTLIB::SRA_I128, 0);
247 // Libcalls should use the AAPCS base standard ABI, even if hard float
248 // is in effect, as per the ARM RTABI specification, section 4.1.2.
249 if (Subtarget->isAAPCS_ABI()) {
250 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
251 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
252 CallingConv::ARM_AAPCS);
256 if (Subtarget->isThumb1Only())
257 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
259 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
260 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
261 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
262 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
264 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
267 if (Subtarget->hasNEON()) {
268 addDRTypeForNEON(MVT::v2f32);
269 addDRTypeForNEON(MVT::v8i8);
270 addDRTypeForNEON(MVT::v4i16);
271 addDRTypeForNEON(MVT::v2i32);
272 addDRTypeForNEON(MVT::v1i64);
274 addQRTypeForNEON(MVT::v4f32);
275 addQRTypeForNEON(MVT::v2f64);
276 addQRTypeForNEON(MVT::v16i8);
277 addQRTypeForNEON(MVT::v8i16);
278 addQRTypeForNEON(MVT::v4i32);
279 addQRTypeForNEON(MVT::v2i64);
281 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
282 // neither Neon nor VFP support any arithmetic operations on it.
283 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
284 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
285 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
286 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
287 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
288 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
289 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
290 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
291 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
292 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
293 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
294 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
295 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
296 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
297 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
298 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
299 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
300 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
301 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
302 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
303 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
304 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
305 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
306 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
308 // Neon does not support some operations on v1i64 and v2i64 types.
309 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
310 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
311 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
312 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
314 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
315 setTargetDAGCombine(ISD::SHL);
316 setTargetDAGCombine(ISD::SRL);
317 setTargetDAGCombine(ISD::SRA);
318 setTargetDAGCombine(ISD::SIGN_EXTEND);
319 setTargetDAGCombine(ISD::ZERO_EXTEND);
320 setTargetDAGCombine(ISD::ANY_EXTEND);
321 setTargetDAGCombine(ISD::SELECT_CC);
324 computeRegisterProperties();
326 // ARM does not have f32 extending load.
327 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
329 // ARM does not have i1 sign extending load.
330 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
332 // ARM supports all 4 flavors of integer indexed load / store.
333 if (!Subtarget->isThumb1Only()) {
334 for (unsigned im = (unsigned)ISD::PRE_INC;
335 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
336 setIndexedLoadAction(im, MVT::i1, Legal);
337 setIndexedLoadAction(im, MVT::i8, Legal);
338 setIndexedLoadAction(im, MVT::i16, Legal);
339 setIndexedLoadAction(im, MVT::i32, Legal);
340 setIndexedStoreAction(im, MVT::i1, Legal);
341 setIndexedStoreAction(im, MVT::i8, Legal);
342 setIndexedStoreAction(im, MVT::i16, Legal);
343 setIndexedStoreAction(im, MVT::i32, Legal);
347 // i64 operation support.
348 if (Subtarget->isThumb1Only()) {
349 setOperationAction(ISD::MUL, MVT::i64, Expand);
350 setOperationAction(ISD::MULHU, MVT::i32, Expand);
351 setOperationAction(ISD::MULHS, MVT::i32, Expand);
352 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
353 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
355 setOperationAction(ISD::MUL, MVT::i64, Expand);
356 setOperationAction(ISD::MULHU, MVT::i32, Expand);
357 if (!Subtarget->hasV6Ops())
358 setOperationAction(ISD::MULHS, MVT::i32, Expand);
360 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
361 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
362 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
363 setOperationAction(ISD::SRL, MVT::i64, Custom);
364 setOperationAction(ISD::SRA, MVT::i64, Custom);
366 // ARM does not have ROTL.
367 setOperationAction(ISD::ROTL, MVT::i32, Expand);
368 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
369 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
370 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
371 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
373 // Only ARMv6 has BSWAP.
374 if (!Subtarget->hasV6Ops())
375 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
377 // These are expanded into libcalls.
378 if (!Subtarget->hasDivide()) {
379 // v7M has a hardware divider
380 setOperationAction(ISD::SDIV, MVT::i32, Expand);
381 setOperationAction(ISD::UDIV, MVT::i32, Expand);
383 setOperationAction(ISD::SREM, MVT::i32, Expand);
384 setOperationAction(ISD::UREM, MVT::i32, Expand);
385 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
386 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
388 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
389 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
390 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
391 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
392 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
394 setOperationAction(ISD::TRAP, MVT::Other, Legal);
396 // Use the default implementation.
397 setOperationAction(ISD::VASTART, MVT::Other, Custom);
398 setOperationAction(ISD::VAARG, MVT::Other, Expand);
399 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
400 setOperationAction(ISD::VAEND, MVT::Other, Expand);
401 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
402 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
403 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
404 // FIXME: Shouldn't need this, since no register is used, but the legalizer
405 // doesn't yet know how to not do that for SjLj.
406 setExceptionSelectorRegister(ARM::R0);
407 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
408 // Handle atomics directly for ARMv[67] (except for Thumb1), otherwise
409 // use the default expansion.
410 bool canHandleAtomics =
411 (Subtarget->hasV7Ops() ||
412 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only()));
413 if (canHandleAtomics) {
414 // membarrier needs custom lowering; the rest are legal and handled
416 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
418 // Set them all for expansion, which will force libcalls.
419 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
420 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
421 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
422 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
423 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
424 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
425 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
426 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
427 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
428 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
429 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
430 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
431 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
432 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
445 // 64-bit versions are always libcalls (for now)
446 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
447 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
449 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
450 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
451 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
452 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
453 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
455 // If the subtarget does not have extract instructions, sign_extend_inreg
456 // needs to be expanded. Extract is available in ARM mode on v6 and up,
457 // and on most Thumb2 implementations.
458 if (!Subtarget->hasV6Ops()
459 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
460 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
461 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
463 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
465 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
466 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
467 // iff target supports vfp2.
468 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
470 // We want to custom lower some of our intrinsics.
471 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
473 setOperationAction(ISD::SETCC, MVT::i32, Expand);
474 setOperationAction(ISD::SETCC, MVT::f32, Expand);
475 setOperationAction(ISD::SETCC, MVT::f64, Expand);
476 setOperationAction(ISD::SELECT, MVT::i32, Expand);
477 setOperationAction(ISD::SELECT, MVT::f32, Expand);
478 setOperationAction(ISD::SELECT, MVT::f64, Expand);
479 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
480 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
481 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
483 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
484 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
485 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
486 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
487 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
489 // We don't support sin/cos/fmod/copysign/pow
490 setOperationAction(ISD::FSIN, MVT::f64, Expand);
491 setOperationAction(ISD::FSIN, MVT::f32, Expand);
492 setOperationAction(ISD::FCOS, MVT::f32, Expand);
493 setOperationAction(ISD::FCOS, MVT::f64, Expand);
494 setOperationAction(ISD::FREM, MVT::f64, Expand);
495 setOperationAction(ISD::FREM, MVT::f32, Expand);
496 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
497 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
498 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
500 setOperationAction(ISD::FPOW, MVT::f64, Expand);
501 setOperationAction(ISD::FPOW, MVT::f32, Expand);
503 // Various VFP goodness
504 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
505 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
506 if (Subtarget->hasVFP2()) {
507 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
508 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
509 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
510 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
512 // Special handling for half-precision FP.
513 if (!Subtarget->hasFP16()) {
514 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
515 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
519 // We have target-specific dag combine patterns for the following nodes:
520 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
521 setTargetDAGCombine(ISD::ADD);
522 setTargetDAGCombine(ISD::SUB);
523 setTargetDAGCombine(ISD::MUL);
525 setStackPointerRegisterToSaveRestore(ARM::SP);
527 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
528 setSchedulingPreference(Sched::RegPressure);
530 setSchedulingPreference(Sched::Hybrid);
532 // FIXME: If-converter should use instruction latency to determine
533 // profitability rather than relying on fixed limits.
534 if (Subtarget->getCPUString() == "generic") {
535 // Generic (and overly aggressive) if-conversion limits.
536 setIfCvtBlockSizeLimit(10);
537 setIfCvtDupBlockSizeLimit(2);
538 } else if (Subtarget->hasV7Ops()) {
539 setIfCvtBlockSizeLimit(3);
540 setIfCvtDupBlockSizeLimit(1);
541 } else if (Subtarget->hasV6Ops()) {
542 setIfCvtBlockSizeLimit(2);
543 setIfCvtDupBlockSizeLimit(1);
545 setIfCvtBlockSizeLimit(3);
546 setIfCvtDupBlockSizeLimit(2);
549 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
550 // Do not enable CodePlacementOpt for now: it currently runs after the
551 // ARMConstantIslandPass and messes up branch relaxation and placement
552 // of constant islands.
553 // benefitFromCodePlacementOpt = true;
556 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
559 case ARMISD::Wrapper: return "ARMISD::Wrapper";
560 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
561 case ARMISD::CALL: return "ARMISD::CALL";
562 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
563 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
564 case ARMISD::tCALL: return "ARMISD::tCALL";
565 case ARMISD::BRCOND: return "ARMISD::BRCOND";
566 case ARMISD::BR_JT: return "ARMISD::BR_JT";
567 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
568 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
569 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
570 case ARMISD::CMP: return "ARMISD::CMP";
571 case ARMISD::CMPZ: return "ARMISD::CMPZ";
572 case ARMISD::CMPFP: return "ARMISD::CMPFP";
573 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
574 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
575 case ARMISD::CMOV: return "ARMISD::CMOV";
576 case ARMISD::CNEG: return "ARMISD::CNEG";
578 case ARMISD::RBIT: return "ARMISD::RBIT";
580 case ARMISD::FTOSI: return "ARMISD::FTOSI";
581 case ARMISD::FTOUI: return "ARMISD::FTOUI";
582 case ARMISD::SITOF: return "ARMISD::SITOF";
583 case ARMISD::UITOF: return "ARMISD::UITOF";
585 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
586 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
587 case ARMISD::RRX: return "ARMISD::RRX";
589 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
590 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
592 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
593 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
595 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
597 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
599 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
601 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
602 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
604 case ARMISD::VCEQ: return "ARMISD::VCEQ";
605 case ARMISD::VCGE: return "ARMISD::VCGE";
606 case ARMISD::VCGEU: return "ARMISD::VCGEU";
607 case ARMISD::VCGT: return "ARMISD::VCGT";
608 case ARMISD::VCGTU: return "ARMISD::VCGTU";
609 case ARMISD::VTST: return "ARMISD::VTST";
611 case ARMISD::VSHL: return "ARMISD::VSHL";
612 case ARMISD::VSHRs: return "ARMISD::VSHRs";
613 case ARMISD::VSHRu: return "ARMISD::VSHRu";
614 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
615 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
616 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
617 case ARMISD::VSHRN: return "ARMISD::VSHRN";
618 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
619 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
620 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
621 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
622 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
623 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
624 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
625 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
626 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
627 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
628 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
629 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
630 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
631 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
632 case ARMISD::VDUP: return "ARMISD::VDUP";
633 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
634 case ARMISD::VEXT: return "ARMISD::VEXT";
635 case ARMISD::VREV64: return "ARMISD::VREV64";
636 case ARMISD::VREV32: return "ARMISD::VREV32";
637 case ARMISD::VREV16: return "ARMISD::VREV16";
638 case ARMISD::VZIP: return "ARMISD::VZIP";
639 case ARMISD::VUZP: return "ARMISD::VUZP";
640 case ARMISD::VTRN: return "ARMISD::VTRN";
641 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
642 case ARMISD::FMAX: return "ARMISD::FMAX";
643 case ARMISD::FMIN: return "ARMISD::FMIN";
647 /// getRegClassFor - Return the register class that should be used for the
648 /// specified value type.
649 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
650 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
651 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
652 // load / store 4 to 8 consecutive D registers.
653 if (Subtarget->hasNEON()) {
654 if (VT == MVT::v4i64)
655 return ARM::QQPRRegisterClass;
656 else if (VT == MVT::v8i64)
657 return ARM::QQQQPRRegisterClass;
659 return TargetLowering::getRegClassFor(VT);
662 /// getFunctionAlignment - Return the Log2 alignment of this function.
663 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
664 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
667 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
668 unsigned NumVals = N->getNumValues();
670 return Sched::RegPressure;
672 for (unsigned i = 0; i != NumVals; ++i) {
673 EVT VT = N->getValueType(i);
674 if (VT.isFloatingPoint() || VT.isVector())
675 return Sched::Latency;
678 if (!N->isMachineOpcode())
679 return Sched::RegPressure;
681 // Load are scheduled for latency even if there instruction itinerary
683 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
684 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
686 return Sched::Latency;
688 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
689 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
690 return Sched::Latency;
691 return Sched::RegPressure;
694 //===----------------------------------------------------------------------===//
696 //===----------------------------------------------------------------------===//
698 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
699 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
701 default: llvm_unreachable("Unknown condition code!");
702 case ISD::SETNE: return ARMCC::NE;
703 case ISD::SETEQ: return ARMCC::EQ;
704 case ISD::SETGT: return ARMCC::GT;
705 case ISD::SETGE: return ARMCC::GE;
706 case ISD::SETLT: return ARMCC::LT;
707 case ISD::SETLE: return ARMCC::LE;
708 case ISD::SETUGT: return ARMCC::HI;
709 case ISD::SETUGE: return ARMCC::HS;
710 case ISD::SETULT: return ARMCC::LO;
711 case ISD::SETULE: return ARMCC::LS;
715 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
716 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
717 ARMCC::CondCodes &CondCode2) {
718 CondCode2 = ARMCC::AL;
720 default: llvm_unreachable("Unknown FP condition!");
722 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
724 case ISD::SETOGT: CondCode = ARMCC::GT; break;
726 case ISD::SETOGE: CondCode = ARMCC::GE; break;
727 case ISD::SETOLT: CondCode = ARMCC::MI; break;
728 case ISD::SETOLE: CondCode = ARMCC::LS; break;
729 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
730 case ISD::SETO: CondCode = ARMCC::VC; break;
731 case ISD::SETUO: CondCode = ARMCC::VS; break;
732 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
733 case ISD::SETUGT: CondCode = ARMCC::HI; break;
734 case ISD::SETUGE: CondCode = ARMCC::PL; break;
736 case ISD::SETULT: CondCode = ARMCC::LT; break;
738 case ISD::SETULE: CondCode = ARMCC::LE; break;
740 case ISD::SETUNE: CondCode = ARMCC::NE; break;
744 //===----------------------------------------------------------------------===//
745 // Calling Convention Implementation
746 //===----------------------------------------------------------------------===//
748 #include "ARMGenCallingConv.inc"
750 // APCS f64 is in register pairs, possibly split to stack
751 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
752 CCValAssign::LocInfo &LocInfo,
753 CCState &State, bool CanFail) {
754 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
756 // Try to get the first register.
757 if (unsigned Reg = State.AllocateReg(RegList, 4))
758 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
760 // For the 2nd half of a v2f64, do not fail.
764 // Put the whole thing on the stack.
765 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
766 State.AllocateStack(8, 4),
771 // Try to get the second register.
772 if (unsigned Reg = State.AllocateReg(RegList, 4))
773 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
775 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
776 State.AllocateStack(4, 4),
781 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
782 CCValAssign::LocInfo &LocInfo,
783 ISD::ArgFlagsTy &ArgFlags,
785 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
787 if (LocVT == MVT::v2f64 &&
788 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
790 return true; // we handled it
793 // AAPCS f64 is in aligned register pairs
794 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
795 CCValAssign::LocInfo &LocInfo,
796 CCState &State, bool CanFail) {
797 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
798 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
800 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
802 // For the 2nd half of a v2f64, do not just fail.
806 // Put the whole thing on the stack.
807 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
808 State.AllocateStack(8, 8),
814 for (i = 0; i < 2; ++i)
815 if (HiRegList[i] == Reg)
818 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
819 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
824 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
825 CCValAssign::LocInfo &LocInfo,
826 ISD::ArgFlagsTy &ArgFlags,
828 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
830 if (LocVT == MVT::v2f64 &&
831 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
833 return true; // we handled it
836 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
837 CCValAssign::LocInfo &LocInfo, CCState &State) {
838 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
839 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
841 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
843 return false; // we didn't handle it
846 for (i = 0; i < 2; ++i)
847 if (HiRegList[i] == Reg)
850 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
851 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
856 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
857 CCValAssign::LocInfo &LocInfo,
858 ISD::ArgFlagsTy &ArgFlags,
860 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
862 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
864 return true; // we handled it
867 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
868 CCValAssign::LocInfo &LocInfo,
869 ISD::ArgFlagsTy &ArgFlags,
871 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
875 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
876 /// given CallingConvention value.
877 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
879 bool isVarArg) const {
882 llvm_unreachable("Unsupported calling convention");
884 case CallingConv::Fast:
885 // Use target triple & subtarget features to do actual dispatch.
886 if (Subtarget->isAAPCS_ABI()) {
887 if (Subtarget->hasVFP2() &&
888 FloatABIType == FloatABI::Hard && !isVarArg)
889 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
891 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
893 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
894 case CallingConv::ARM_AAPCS_VFP:
895 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
896 case CallingConv::ARM_AAPCS:
897 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
898 case CallingConv::ARM_APCS:
899 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
903 /// LowerCallResult - Lower the result values of a call into the
904 /// appropriate copies out of appropriate physical registers.
906 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
907 CallingConv::ID CallConv, bool isVarArg,
908 const SmallVectorImpl<ISD::InputArg> &Ins,
909 DebugLoc dl, SelectionDAG &DAG,
910 SmallVectorImpl<SDValue> &InVals) const {
912 // Assign locations to each value returned by this call.
913 SmallVector<CCValAssign, 16> RVLocs;
914 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
915 RVLocs, *DAG.getContext());
916 CCInfo.AnalyzeCallResult(Ins,
917 CCAssignFnForNode(CallConv, /* Return*/ true,
920 // Copy all of the result registers out of their specified physreg.
921 for (unsigned i = 0; i != RVLocs.size(); ++i) {
922 CCValAssign VA = RVLocs[i];
925 if (VA.needsCustom()) {
926 // Handle f64 or half of a v2f64.
927 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
929 Chain = Lo.getValue(1);
930 InFlag = Lo.getValue(2);
931 VA = RVLocs[++i]; // skip ahead to next loc
932 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
934 Chain = Hi.getValue(1);
935 InFlag = Hi.getValue(2);
936 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
938 if (VA.getLocVT() == MVT::v2f64) {
939 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
940 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
941 DAG.getConstant(0, MVT::i32));
943 VA = RVLocs[++i]; // skip ahead to next loc
944 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
945 Chain = Lo.getValue(1);
946 InFlag = Lo.getValue(2);
947 VA = RVLocs[++i]; // skip ahead to next loc
948 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
949 Chain = Hi.getValue(1);
950 InFlag = Hi.getValue(2);
951 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
952 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
953 DAG.getConstant(1, MVT::i32));
956 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
958 Chain = Val.getValue(1);
959 InFlag = Val.getValue(2);
962 switch (VA.getLocInfo()) {
963 default: llvm_unreachable("Unknown loc info!");
964 case CCValAssign::Full: break;
965 case CCValAssign::BCvt:
966 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
970 InVals.push_back(Val);
976 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
977 /// by "Src" to address "Dst" of size "Size". Alignment information is
978 /// specified by the specific parameter attribute. The copy will be passed as
979 /// a byval function parameter.
980 /// Sometimes what we are copying is the end of a larger object, the part that
981 /// does not fit in registers.
983 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
984 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
986 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
987 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
988 /*isVolatile=*/false, /*AlwaysInline=*/false,
992 /// LowerMemOpCallTo - Store the argument to the stack.
994 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
995 SDValue StackPtr, SDValue Arg,
996 DebugLoc dl, SelectionDAG &DAG,
997 const CCValAssign &VA,
998 ISD::ArgFlagsTy Flags) const {
999 unsigned LocMemOffset = VA.getLocMemOffset();
1000 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1001 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1002 if (Flags.isByVal()) {
1003 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1005 return DAG.getStore(Chain, dl, Arg, PtrOff,
1006 PseudoSourceValue::getStack(), LocMemOffset,
1010 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1011 SDValue Chain, SDValue &Arg,
1012 RegsToPassVector &RegsToPass,
1013 CCValAssign &VA, CCValAssign &NextVA,
1015 SmallVector<SDValue, 8> &MemOpChains,
1016 ISD::ArgFlagsTy Flags) const {
1018 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1019 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1020 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1022 if (NextVA.isRegLoc())
1023 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1025 assert(NextVA.isMemLoc());
1026 if (StackPtr.getNode() == 0)
1027 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1029 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1035 /// LowerCall - Lowering a call into a callseq_start <-
1036 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1039 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1040 CallingConv::ID CallConv, bool isVarArg,
1042 const SmallVectorImpl<ISD::OutputArg> &Outs,
1043 const SmallVectorImpl<ISD::InputArg> &Ins,
1044 DebugLoc dl, SelectionDAG &DAG,
1045 SmallVectorImpl<SDValue> &InVals) const {
1046 MachineFunction &MF = DAG.getMachineFunction();
1047 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1048 bool IsSibCall = false;
1049 // Temporarily disable tail calls so things don't break.
1050 if (!EnableARMTailCalls)
1053 // Check if it's really possible to do a tail call.
1054 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1055 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1057 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1058 // detected sibcalls.
1065 // Analyze operands of the call, assigning locations to each operand.
1066 SmallVector<CCValAssign, 16> ArgLocs;
1067 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1069 CCInfo.AnalyzeCallOperands(Outs,
1070 CCAssignFnForNode(CallConv, /* Return*/ false,
1073 // Get a count of how many bytes are to be pushed on the stack.
1074 unsigned NumBytes = CCInfo.getNextStackOffset();
1076 // For tail calls, memory operands are available in our caller's stack.
1080 // Adjust the stack pointer for the new arguments...
1081 // These operations are automatically eliminated by the prolog/epilog pass
1083 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1085 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1087 RegsToPassVector RegsToPass;
1088 SmallVector<SDValue, 8> MemOpChains;
1090 // Walk the register/memloc assignments, inserting copies/loads. In the case
1091 // of tail call optimization, arguments are handled later.
1092 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1094 ++i, ++realArgIdx) {
1095 CCValAssign &VA = ArgLocs[i];
1096 SDValue Arg = Outs[realArgIdx].Val;
1097 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1099 // Promote the value if needed.
1100 switch (VA.getLocInfo()) {
1101 default: llvm_unreachable("Unknown loc info!");
1102 case CCValAssign::Full: break;
1103 case CCValAssign::SExt:
1104 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1106 case CCValAssign::ZExt:
1107 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1109 case CCValAssign::AExt:
1110 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1112 case CCValAssign::BCvt:
1113 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1117 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1118 if (VA.needsCustom()) {
1119 if (VA.getLocVT() == MVT::v2f64) {
1120 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1121 DAG.getConstant(0, MVT::i32));
1122 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1123 DAG.getConstant(1, MVT::i32));
1125 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1126 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1128 VA = ArgLocs[++i]; // skip ahead to next loc
1129 if (VA.isRegLoc()) {
1130 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1131 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1133 assert(VA.isMemLoc());
1135 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1136 dl, DAG, VA, Flags));
1139 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1140 StackPtr, MemOpChains, Flags);
1142 } else if (VA.isRegLoc()) {
1143 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1144 } else if (!IsSibCall) {
1145 assert(VA.isMemLoc());
1147 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1148 dl, DAG, VA, Flags));
1152 if (!MemOpChains.empty())
1153 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1154 &MemOpChains[0], MemOpChains.size());
1156 // Build a sequence of copy-to-reg nodes chained together with token chain
1157 // and flag operands which copy the outgoing args into the appropriate regs.
1159 // Tail call byval lowering might overwrite argument registers so in case of
1160 // tail call optimization the copies to registers are lowered later.
1162 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1163 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1164 RegsToPass[i].second, InFlag);
1165 InFlag = Chain.getValue(1);
1168 // For tail calls lower the arguments to the 'real' stack slot.
1170 // Force all the incoming stack arguments to be loaded from the stack
1171 // before any new outgoing arguments are stored to the stack, because the
1172 // outgoing stack slots may alias the incoming argument stack slots, and
1173 // the alias isn't otherwise explicit. This is slightly more conservative
1174 // than necessary, because it means that each store effectively depends
1175 // on every argument instead of just those arguments it would clobber.
1177 // Do not flag preceeding copytoreg stuff together with the following stuff.
1179 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1180 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1181 RegsToPass[i].second, InFlag);
1182 InFlag = Chain.getValue(1);
1187 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1188 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1189 // node so that legalize doesn't hack it.
1190 bool isDirect = false;
1191 bool isARMFunc = false;
1192 bool isLocalARMFunc = false;
1193 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1195 if (EnableARMLongCalls) {
1196 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1197 && "long-calls with non-static relocation model!");
1198 // Handle a global address or an external symbol. If it's not one of
1199 // those, the target's already in a register, so we don't need to do
1201 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1202 const GlobalValue *GV = G->getGlobal();
1203 // Create a constant pool entry for the callee address
1204 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1205 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1208 // Get the address of the callee into a register
1209 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1210 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1211 Callee = DAG.getLoad(getPointerTy(), dl,
1212 DAG.getEntryNode(), CPAddr,
1213 PseudoSourceValue::getConstantPool(), 0,
1215 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1216 const char *Sym = S->getSymbol();
1218 // Create a constant pool entry for the callee address
1219 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1220 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1221 Sym, ARMPCLabelIndex, 0);
1222 // Get the address of the callee into a register
1223 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1224 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1225 Callee = DAG.getLoad(getPointerTy(), dl,
1226 DAG.getEntryNode(), CPAddr,
1227 PseudoSourceValue::getConstantPool(), 0,
1230 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1231 const GlobalValue *GV = G->getGlobal();
1233 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1234 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1235 getTargetMachine().getRelocationModel() != Reloc::Static;
1236 isARMFunc = !Subtarget->isThumb() || isStub;
1237 // ARM call to a local ARM function is predicable.
1238 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1239 // tBX takes a register source operand.
1240 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1241 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1242 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1245 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1246 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1247 Callee = DAG.getLoad(getPointerTy(), dl,
1248 DAG.getEntryNode(), CPAddr,
1249 PseudoSourceValue::getConstantPool(), 0,
1251 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1252 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1253 getPointerTy(), Callee, PICLabel);
1255 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1256 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1258 bool isStub = Subtarget->isTargetDarwin() &&
1259 getTargetMachine().getRelocationModel() != Reloc::Static;
1260 isARMFunc = !Subtarget->isThumb() || isStub;
1261 // tBX takes a register source operand.
1262 const char *Sym = S->getSymbol();
1263 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1264 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1265 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1266 Sym, ARMPCLabelIndex, 4);
1267 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1268 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1269 Callee = DAG.getLoad(getPointerTy(), dl,
1270 DAG.getEntryNode(), CPAddr,
1271 PseudoSourceValue::getConstantPool(), 0,
1273 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1274 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1275 getPointerTy(), Callee, PICLabel);
1277 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1280 // FIXME: handle tail calls differently.
1282 if (Subtarget->isThumb()) {
1283 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1284 CallOpc = ARMISD::CALL_NOLINK;
1286 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1288 CallOpc = (isDirect || Subtarget->hasV5TOps())
1289 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1290 : ARMISD::CALL_NOLINK;
1292 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1293 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1294 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1295 InFlag = Chain.getValue(1);
1298 std::vector<SDValue> Ops;
1299 Ops.push_back(Chain);
1300 Ops.push_back(Callee);
1302 // Add argument registers to the end of the list so that they are known live
1304 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1305 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1306 RegsToPass[i].second.getValueType()));
1308 if (InFlag.getNode())
1309 Ops.push_back(InFlag);
1311 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1313 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1315 // Returns a chain and a flag for retval copy to use.
1316 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1317 InFlag = Chain.getValue(1);
1319 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1320 DAG.getIntPtrConstant(0, true), InFlag);
1322 InFlag = Chain.getValue(1);
1324 // Handle result values, copying them out of physregs into vregs that we
1326 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1330 /// MatchingStackOffset - Return true if the given stack call argument is
1331 /// already available in the same position (relatively) of the caller's
1332 /// incoming argument stack.
1334 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1335 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1336 const ARMInstrInfo *TII) {
1337 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1339 if (Arg.getOpcode() == ISD::CopyFromReg) {
1340 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1341 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1343 MachineInstr *Def = MRI->getVRegDef(VR);
1346 if (!Flags.isByVal()) {
1347 if (!TII->isLoadFromStackSlot(Def, FI))
1350 // unsigned Opcode = Def->getOpcode();
1351 // if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
1352 // Def->getOperand(1).isFI()) {
1353 // FI = Def->getOperand(1).getIndex();
1354 // Bytes = Flags.getByValSize();
1358 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1359 if (Flags.isByVal())
1360 // ByVal argument is passed in as a pointer but it's now being
1361 // dereferenced. e.g.
1362 // define @foo(%struct.X* %A) {
1363 // tail call @bar(%struct.X* byval %A)
1366 SDValue Ptr = Ld->getBasePtr();
1367 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1370 FI = FINode->getIndex();
1374 assert(FI != INT_MAX);
1375 if (!MFI->isFixedObjectIndex(FI))
1377 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1380 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1381 /// for tail call optimization. Targets which want to do tail call
1382 /// optimization should implement this function.
1384 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1385 CallingConv::ID CalleeCC,
1387 bool isCalleeStructRet,
1388 bool isCallerStructRet,
1389 const SmallVectorImpl<ISD::OutputArg> &Outs,
1390 const SmallVectorImpl<ISD::InputArg> &Ins,
1391 SelectionDAG& DAG) const {
1392 const Function *CallerF = DAG.getMachineFunction().getFunction();
1393 CallingConv::ID CallerCC = CallerF->getCallingConv();
1394 bool CCMatch = CallerCC == CalleeCC;
1396 // Look for obvious safe cases to perform tail call optimization that do not
1397 // require ABI changes. This is what gcc calls sibcall.
1399 // Do not sibcall optimize vararg calls unless the call site is not passing
1401 if (isVarArg && !Outs.empty())
1404 // Also avoid sibcall optimization if either caller or callee uses struct
1405 // return semantics.
1406 if (isCalleeStructRet || isCallerStructRet)
1409 // FIXME: Completely disable sibcal for Thumb1 since Thumb1RegisterInfo::
1410 // emitEpilogue is not ready for them.
1411 if (Subtarget->isThumb1Only())
1414 if (isa<ExternalSymbolSDNode>(Callee))
1417 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1418 if (Subtarget->isThumb1Only())
1421 // On Thumb, for the moment, we can only do this to functions defined in this
1422 // compilation, or to indirect calls. A Thumb B to an ARM function is not
1423 // easily fixed up in the linker, unlike BL.
1424 if (Subtarget->isThumb()) {
1425 const GlobalValue *GV = G->getGlobal();
1426 if (GV->isDeclaration() || GV->isWeakForLinker())
1432 // If the calling conventions do not match, then we'd better make sure the
1433 // results are returned in the same way as what the caller expects.
1435 SmallVector<CCValAssign, 16> RVLocs1;
1436 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1437 RVLocs1, *DAG.getContext());
1438 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1440 SmallVector<CCValAssign, 16> RVLocs2;
1441 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1442 RVLocs2, *DAG.getContext());
1443 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1445 if (RVLocs1.size() != RVLocs2.size())
1447 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1448 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1450 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1452 if (RVLocs1[i].isRegLoc()) {
1453 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1456 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1462 // If the callee takes no arguments then go on to check the results of the
1464 if (!Outs.empty()) {
1465 // Check if stack adjustment is needed. For now, do not do this if any
1466 // argument is passed on the stack.
1467 SmallVector<CCValAssign, 16> ArgLocs;
1468 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1469 ArgLocs, *DAG.getContext());
1470 CCInfo.AnalyzeCallOperands(Outs,
1471 CCAssignFnForNode(CalleeCC, false, isVarArg));
1472 if (CCInfo.getNextStackOffset()) {
1473 MachineFunction &MF = DAG.getMachineFunction();
1475 // Check if the arguments are already laid out in the right way as
1476 // the caller's fixed stack objects.
1477 MachineFrameInfo *MFI = MF.getFrameInfo();
1478 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1479 const ARMInstrInfo *TII =
1480 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1481 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1483 ++i, ++realArgIdx) {
1484 CCValAssign &VA = ArgLocs[i];
1485 EVT RegVT = VA.getLocVT();
1486 SDValue Arg = Outs[realArgIdx].Val;
1487 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1488 if (VA.getLocInfo() == CCValAssign::Indirect)
1490 if (VA.needsCustom()) {
1491 // f64 and vector types are split into multiple registers or
1492 // register/stack-slot combinations. The types will not match
1493 // the registers; give up on memory f64 refs until we figure
1494 // out what to do about this.
1497 if (!ArgLocs[++i].isRegLoc())
1499 if (RegVT == MVT::v2f64) {
1500 if (!ArgLocs[++i].isRegLoc())
1502 if (!ArgLocs[++i].isRegLoc())
1505 } else if (!VA.isRegLoc()) {
1506 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1518 ARMTargetLowering::LowerReturn(SDValue Chain,
1519 CallingConv::ID CallConv, bool isVarArg,
1520 const SmallVectorImpl<ISD::OutputArg> &Outs,
1521 DebugLoc dl, SelectionDAG &DAG) const {
1523 // CCValAssign - represent the assignment of the return value to a location.
1524 SmallVector<CCValAssign, 16> RVLocs;
1526 // CCState - Info about the registers and stack slots.
1527 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1530 // Analyze outgoing return values.
1531 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1534 // If this is the first return lowered for this function, add
1535 // the regs to the liveout set for the function.
1536 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1537 for (unsigned i = 0; i != RVLocs.size(); ++i)
1538 if (RVLocs[i].isRegLoc())
1539 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1544 // Copy the result values into the output registers.
1545 for (unsigned i = 0, realRVLocIdx = 0;
1547 ++i, ++realRVLocIdx) {
1548 CCValAssign &VA = RVLocs[i];
1549 assert(VA.isRegLoc() && "Can only return in registers!");
1551 SDValue Arg = Outs[realRVLocIdx].Val;
1553 switch (VA.getLocInfo()) {
1554 default: llvm_unreachable("Unknown loc info!");
1555 case CCValAssign::Full: break;
1556 case CCValAssign::BCvt:
1557 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1561 if (VA.needsCustom()) {
1562 if (VA.getLocVT() == MVT::v2f64) {
1563 // Extract the first half and return it in two registers.
1564 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1565 DAG.getConstant(0, MVT::i32));
1566 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1567 DAG.getVTList(MVT::i32, MVT::i32), Half);
1569 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1570 Flag = Chain.getValue(1);
1571 VA = RVLocs[++i]; // skip ahead to next loc
1572 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1573 HalfGPRs.getValue(1), Flag);
1574 Flag = Chain.getValue(1);
1575 VA = RVLocs[++i]; // skip ahead to next loc
1577 // Extract the 2nd half and fall through to handle it as an f64 value.
1578 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1579 DAG.getConstant(1, MVT::i32));
1581 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1583 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1584 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1585 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1586 Flag = Chain.getValue(1);
1587 VA = RVLocs[++i]; // skip ahead to next loc
1588 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1591 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1593 // Guarantee that all emitted copies are
1594 // stuck together, avoiding something bad.
1595 Flag = Chain.getValue(1);
1600 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1602 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1607 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1608 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1609 // one of the above mentioned nodes. It has to be wrapped because otherwise
1610 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1611 // be used to form addressing mode. These wrapped nodes will be selected
1613 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1614 EVT PtrVT = Op.getValueType();
1615 // FIXME there is no actual debug info here
1616 DebugLoc dl = Op.getDebugLoc();
1617 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1619 if (CP->isMachineConstantPoolEntry())
1620 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1621 CP->getAlignment());
1623 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1624 CP->getAlignment());
1625 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1628 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1629 SelectionDAG &DAG) const {
1630 MachineFunction &MF = DAG.getMachineFunction();
1631 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1632 unsigned ARMPCLabelIndex = 0;
1633 DebugLoc DL = Op.getDebugLoc();
1634 EVT PtrVT = getPointerTy();
1635 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1636 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1638 if (RelocM == Reloc::Static) {
1639 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1641 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1642 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1643 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1644 ARMCP::CPBlockAddress,
1646 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1648 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1649 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1650 PseudoSourceValue::getConstantPool(), 0,
1652 if (RelocM == Reloc::Static)
1654 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1655 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1658 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1660 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1661 SelectionDAG &DAG) const {
1662 DebugLoc dl = GA->getDebugLoc();
1663 EVT PtrVT = getPointerTy();
1664 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1665 MachineFunction &MF = DAG.getMachineFunction();
1666 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1667 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1668 ARMConstantPoolValue *CPV =
1669 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1670 ARMCP::CPValue, PCAdj, "tlsgd", true);
1671 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1672 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1673 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1674 PseudoSourceValue::getConstantPool(), 0,
1676 SDValue Chain = Argument.getValue(1);
1678 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1679 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1681 // call __tls_get_addr.
1684 Entry.Node = Argument;
1685 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1686 Args.push_back(Entry);
1687 // FIXME: is there useful debug info available here?
1688 std::pair<SDValue, SDValue> CallResult =
1689 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1690 false, false, false, false,
1691 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1692 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1693 return CallResult.first;
1696 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1697 // "local exec" model.
1699 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1700 SelectionDAG &DAG) const {
1701 const GlobalValue *GV = GA->getGlobal();
1702 DebugLoc dl = GA->getDebugLoc();
1704 SDValue Chain = DAG.getEntryNode();
1705 EVT PtrVT = getPointerTy();
1706 // Get the Thread Pointer
1707 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1709 if (GV->isDeclaration()) {
1710 MachineFunction &MF = DAG.getMachineFunction();
1711 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1712 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1713 // Initial exec model.
1714 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1715 ARMConstantPoolValue *CPV =
1716 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1717 ARMCP::CPValue, PCAdj, "gottpoff", true);
1718 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1719 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1720 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1721 PseudoSourceValue::getConstantPool(), 0,
1723 Chain = Offset.getValue(1);
1725 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1726 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1728 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1729 PseudoSourceValue::getConstantPool(), 0,
1733 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1734 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1735 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1736 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1737 PseudoSourceValue::getConstantPool(), 0,
1741 // The address of the thread local variable is the add of the thread
1742 // pointer with the offset of the variable.
1743 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1747 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1748 // TODO: implement the "local dynamic" model
1749 assert(Subtarget->isTargetELF() &&
1750 "TLS not implemented for non-ELF targets");
1751 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1752 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1753 // otherwise use the "Local Exec" TLS Model
1754 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1755 return LowerToTLSGeneralDynamicModel(GA, DAG);
1757 return LowerToTLSExecModels(GA, DAG);
1760 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1761 SelectionDAG &DAG) const {
1762 EVT PtrVT = getPointerTy();
1763 DebugLoc dl = Op.getDebugLoc();
1764 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1765 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1766 if (RelocM == Reloc::PIC_) {
1767 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1768 ARMConstantPoolValue *CPV =
1769 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1770 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1771 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1772 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1774 PseudoSourceValue::getConstantPool(), 0,
1776 SDValue Chain = Result.getValue(1);
1777 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1778 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1780 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1781 PseudoSourceValue::getGOT(), 0,
1785 // If we have T2 ops, we can materialize the address directly via movt/movw
1786 // pair. This is always cheaper.
1787 if (Subtarget->useMovt()) {
1788 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1789 DAG.getTargetGlobalAddress(GV, PtrVT));
1791 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1792 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1793 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1794 PseudoSourceValue::getConstantPool(), 0,
1800 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1801 SelectionDAG &DAG) const {
1802 MachineFunction &MF = DAG.getMachineFunction();
1803 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1804 unsigned ARMPCLabelIndex = 0;
1805 EVT PtrVT = getPointerTy();
1806 DebugLoc dl = Op.getDebugLoc();
1807 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1808 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1810 if (RelocM == Reloc::Static)
1811 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1813 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1814 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1815 ARMConstantPoolValue *CPV =
1816 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1817 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1819 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1821 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1822 PseudoSourceValue::getConstantPool(), 0,
1824 SDValue Chain = Result.getValue(1);
1826 if (RelocM == Reloc::PIC_) {
1827 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1828 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1831 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1832 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1833 PseudoSourceValue::getGOT(), 0,
1839 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1840 SelectionDAG &DAG) const {
1841 assert(Subtarget->isTargetELF() &&
1842 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1843 MachineFunction &MF = DAG.getMachineFunction();
1844 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1845 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1846 EVT PtrVT = getPointerTy();
1847 DebugLoc dl = Op.getDebugLoc();
1848 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1849 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1850 "_GLOBAL_OFFSET_TABLE_",
1851 ARMPCLabelIndex, PCAdj);
1852 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1853 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1854 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1855 PseudoSourceValue::getConstantPool(), 0,
1857 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1858 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1862 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1863 DebugLoc dl = Op.getDebugLoc();
1864 SDValue Val = DAG.getConstant(0, MVT::i32);
1865 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1866 Op.getOperand(1), Val);
1870 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1871 DebugLoc dl = Op.getDebugLoc();
1872 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1873 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1877 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1878 const ARMSubtarget *Subtarget) const {
1879 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1880 DebugLoc dl = Op.getDebugLoc();
1882 default: return SDValue(); // Don't custom lower most intrinsics.
1883 case Intrinsic::arm_thread_pointer: {
1884 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1885 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1887 case Intrinsic::eh_sjlj_lsda: {
1888 MachineFunction &MF = DAG.getMachineFunction();
1889 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1890 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1891 EVT PtrVT = getPointerTy();
1892 DebugLoc dl = Op.getDebugLoc();
1893 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1895 unsigned PCAdj = (RelocM != Reloc::PIC_)
1896 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1897 ARMConstantPoolValue *CPV =
1898 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1899 ARMCP::CPLSDA, PCAdj);
1900 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1901 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1903 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1904 PseudoSourceValue::getConstantPool(), 0,
1906 SDValue Chain = Result.getValue(1);
1908 if (RelocM == Reloc::PIC_) {
1909 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1910 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1917 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1918 const ARMSubtarget *Subtarget) {
1919 DebugLoc dl = Op.getDebugLoc();
1920 SDValue Op5 = Op.getOperand(5);
1921 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1922 // v6 and v7 can both handle barriers directly, but need handled a bit
1923 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1925 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1926 if (Subtarget->hasV7Ops())
1927 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1928 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1929 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1930 DAG.getConstant(0, MVT::i32));
1931 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1935 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1936 MachineFunction &MF = DAG.getMachineFunction();
1937 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1939 // vastart just stores the address of the VarArgsFrameIndex slot into the
1940 // memory location argument.
1941 DebugLoc dl = Op.getDebugLoc();
1942 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1943 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1944 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1945 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1950 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1951 SelectionDAG &DAG) const {
1952 SDNode *Node = Op.getNode();
1953 DebugLoc dl = Node->getDebugLoc();
1954 EVT VT = Node->getValueType(0);
1955 SDValue Chain = Op.getOperand(0);
1956 SDValue Size = Op.getOperand(1);
1957 SDValue Align = Op.getOperand(2);
1959 // Chain the dynamic stack allocation so that it doesn't modify the stack
1960 // pointer when other instructions are using the stack.
1961 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1963 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1964 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1965 if (AlignVal > StackAlign)
1966 // Do this now since selection pass cannot introduce new target
1967 // independent node.
1968 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1970 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1971 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1972 // do even more horrible hack later.
1973 MachineFunction &MF = DAG.getMachineFunction();
1974 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1975 if (AFI->isThumb1OnlyFunction()) {
1977 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1979 uint32_t Val = C->getZExtValue();
1980 if (Val <= 508 && ((Val & 3) == 0))
1984 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1987 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1988 SDValue Ops1[] = { Chain, Size, Align };
1989 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1990 Chain = Res.getValue(1);
1991 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1992 DAG.getIntPtrConstant(0, true), SDValue());
1993 SDValue Ops2[] = { Res, Chain };
1994 return DAG.getMergeValues(Ops2, 2, dl);
1998 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1999 SDValue &Root, SelectionDAG &DAG,
2000 DebugLoc dl) const {
2001 MachineFunction &MF = DAG.getMachineFunction();
2002 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2004 TargetRegisterClass *RC;
2005 if (AFI->isThumb1OnlyFunction())
2006 RC = ARM::tGPRRegisterClass;
2008 RC = ARM::GPRRegisterClass;
2010 // Transform the arguments stored in physical registers into virtual ones.
2011 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2012 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2015 if (NextVA.isMemLoc()) {
2016 MachineFrameInfo *MFI = MF.getFrameInfo();
2017 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
2019 // Create load node to retrieve arguments from the stack.
2020 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2021 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2022 PseudoSourceValue::getFixedStack(FI), 0,
2025 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2026 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2029 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2033 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2034 CallingConv::ID CallConv, bool isVarArg,
2035 const SmallVectorImpl<ISD::InputArg>
2037 DebugLoc dl, SelectionDAG &DAG,
2038 SmallVectorImpl<SDValue> &InVals)
2041 MachineFunction &MF = DAG.getMachineFunction();
2042 MachineFrameInfo *MFI = MF.getFrameInfo();
2044 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2046 // Assign locations to all of the incoming arguments.
2047 SmallVector<CCValAssign, 16> ArgLocs;
2048 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2050 CCInfo.AnalyzeFormalArguments(Ins,
2051 CCAssignFnForNode(CallConv, /* Return*/ false,
2054 SmallVector<SDValue, 16> ArgValues;
2056 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2057 CCValAssign &VA = ArgLocs[i];
2059 // Arguments stored in registers.
2060 if (VA.isRegLoc()) {
2061 EVT RegVT = VA.getLocVT();
2064 if (VA.needsCustom()) {
2065 // f64 and vector types are split up into multiple registers or
2066 // combinations of registers and stack slots.
2067 if (VA.getLocVT() == MVT::v2f64) {
2068 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2070 VA = ArgLocs[++i]; // skip ahead to next loc
2072 if (VA.isMemLoc()) {
2073 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
2075 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2076 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2077 PseudoSourceValue::getFixedStack(FI), 0,
2080 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2083 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2084 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2085 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2086 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2087 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2089 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2092 TargetRegisterClass *RC;
2094 if (RegVT == MVT::f32)
2095 RC = ARM::SPRRegisterClass;
2096 else if (RegVT == MVT::f64)
2097 RC = ARM::DPRRegisterClass;
2098 else if (RegVT == MVT::v2f64)
2099 RC = ARM::QPRRegisterClass;
2100 else if (RegVT == MVT::i32)
2101 RC = (AFI->isThumb1OnlyFunction() ?
2102 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2104 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2106 // Transform the arguments in physical registers into virtual ones.
2107 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2108 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2111 // If this is an 8 or 16-bit value, it is really passed promoted
2112 // to 32 bits. Insert an assert[sz]ext to capture this, then
2113 // truncate to the right size.
2114 switch (VA.getLocInfo()) {
2115 default: llvm_unreachable("Unknown loc info!");
2116 case CCValAssign::Full: break;
2117 case CCValAssign::BCvt:
2118 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2120 case CCValAssign::SExt:
2121 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2122 DAG.getValueType(VA.getValVT()));
2123 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2125 case CCValAssign::ZExt:
2126 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2127 DAG.getValueType(VA.getValVT()));
2128 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2132 InVals.push_back(ArgValue);
2134 } else { // VA.isRegLoc()
2137 assert(VA.isMemLoc());
2138 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2140 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2141 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2144 // Create load nodes to retrieve arguments from the stack.
2145 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2146 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2147 PseudoSourceValue::getFixedStack(FI), 0,
2154 static const unsigned GPRArgRegs[] = {
2155 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2158 unsigned NumGPRs = CCInfo.getFirstUnallocated
2159 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2161 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2162 unsigned VARegSize = (4 - NumGPRs) * 4;
2163 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2164 unsigned ArgOffset = CCInfo.getNextStackOffset();
2165 if (VARegSaveSize) {
2166 // If this function is vararg, store any remaining integer argument regs
2167 // to their spots on the stack so that they may be loaded by deferencing
2168 // the result of va_next.
2169 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2170 AFI->setVarArgsFrameIndex(
2171 MFI->CreateFixedObject(VARegSaveSize,
2172 ArgOffset + VARegSaveSize - VARegSize,
2174 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2177 SmallVector<SDValue, 4> MemOps;
2178 for (; NumGPRs < 4; ++NumGPRs) {
2179 TargetRegisterClass *RC;
2180 if (AFI->isThumb1OnlyFunction())
2181 RC = ARM::tGPRRegisterClass;
2183 RC = ARM::GPRRegisterClass;
2185 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2186 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2188 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2189 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2190 0, false, false, 0);
2191 MemOps.push_back(Store);
2192 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2193 DAG.getConstant(4, getPointerTy()));
2195 if (!MemOps.empty())
2196 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2197 &MemOps[0], MemOps.size());
2199 // This will point to the next argument passed via stack.
2200 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
2207 /// isFloatingPointZero - Return true if this is +0.0.
2208 static bool isFloatingPointZero(SDValue Op) {
2209 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2210 return CFP->getValueAPF().isPosZero();
2211 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2212 // Maybe this has already been legalized into the constant pool?
2213 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2214 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2215 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2216 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2217 return CFP->getValueAPF().isPosZero();
2223 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2224 /// the given operands.
2226 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2227 SDValue &ARMCC, SelectionDAG &DAG,
2228 DebugLoc dl) const {
2229 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2230 unsigned C = RHSC->getZExtValue();
2231 if (!isLegalICmpImmediate(C)) {
2232 // Constant does not fit, try adjusting it by one?
2237 if (isLegalICmpImmediate(C-1)) {
2238 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2239 RHS = DAG.getConstant(C-1, MVT::i32);
2244 if (C > 0 && isLegalICmpImmediate(C-1)) {
2245 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2246 RHS = DAG.getConstant(C-1, MVT::i32);
2251 if (isLegalICmpImmediate(C+1)) {
2252 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2253 RHS = DAG.getConstant(C+1, MVT::i32);
2258 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2259 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2260 RHS = DAG.getConstant(C+1, MVT::i32);
2267 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2268 ARMISD::NodeType CompareType;
2271 CompareType = ARMISD::CMP;
2276 CompareType = ARMISD::CMPZ;
2279 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2280 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2283 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2284 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2287 if (!isFloatingPointZero(RHS))
2288 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2290 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2291 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2294 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2295 EVT VT = Op.getValueType();
2296 SDValue LHS = Op.getOperand(0);
2297 SDValue RHS = Op.getOperand(1);
2298 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2299 SDValue TrueVal = Op.getOperand(2);
2300 SDValue FalseVal = Op.getOperand(3);
2301 DebugLoc dl = Op.getDebugLoc();
2303 if (LHS.getValueType() == MVT::i32) {
2305 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2306 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2307 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2310 ARMCC::CondCodes CondCode, CondCode2;
2311 FPCCToARMCC(CC, CondCode, CondCode2);
2313 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2314 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2315 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2316 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2318 if (CondCode2 != ARMCC::AL) {
2319 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2320 // FIXME: Needs another CMP because flag can have but one use.
2321 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2322 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2323 Result, TrueVal, ARMCC2, CCR, Cmp2);
2328 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2329 SDValue Chain = Op.getOperand(0);
2330 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2331 SDValue LHS = Op.getOperand(2);
2332 SDValue RHS = Op.getOperand(3);
2333 SDValue Dest = Op.getOperand(4);
2334 DebugLoc dl = Op.getDebugLoc();
2336 if (LHS.getValueType() == MVT::i32) {
2338 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2339 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2340 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2341 Chain, Dest, ARMCC, CCR,Cmp);
2344 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2345 ARMCC::CondCodes CondCode, CondCode2;
2346 FPCCToARMCC(CC, CondCode, CondCode2);
2348 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2349 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2350 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2351 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2352 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2353 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2354 if (CondCode2 != ARMCC::AL) {
2355 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2356 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2357 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2362 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2363 SDValue Chain = Op.getOperand(0);
2364 SDValue Table = Op.getOperand(1);
2365 SDValue Index = Op.getOperand(2);
2366 DebugLoc dl = Op.getDebugLoc();
2368 EVT PTy = getPointerTy();
2369 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2370 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2371 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2372 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2373 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2374 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2375 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2376 if (Subtarget->isThumb2()) {
2377 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2378 // which does another jump to the destination. This also makes it easier
2379 // to translate it to TBB / TBH later.
2380 // FIXME: This might not work if the function is extremely large.
2381 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2382 Addr, Op.getOperand(2), JTI, UId);
2384 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2385 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2386 PseudoSourceValue::getJumpTable(), 0,
2388 Chain = Addr.getValue(1);
2389 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2390 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2392 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2393 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2394 Chain = Addr.getValue(1);
2395 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2399 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2400 DebugLoc dl = Op.getDebugLoc();
2403 switch (Op.getOpcode()) {
2405 assert(0 && "Invalid opcode!");
2406 case ISD::FP_TO_SINT:
2407 Opc = ARMISD::FTOSI;
2409 case ISD::FP_TO_UINT:
2410 Opc = ARMISD::FTOUI;
2413 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2414 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2417 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2418 EVT VT = Op.getValueType();
2419 DebugLoc dl = Op.getDebugLoc();
2422 switch (Op.getOpcode()) {
2424 assert(0 && "Invalid opcode!");
2425 case ISD::SINT_TO_FP:
2426 Opc = ARMISD::SITOF;
2428 case ISD::UINT_TO_FP:
2429 Opc = ARMISD::UITOF;
2433 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2434 return DAG.getNode(Opc, dl, VT, Op);
2437 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2438 // Implement fcopysign with a fabs and a conditional fneg.
2439 SDValue Tmp0 = Op.getOperand(0);
2440 SDValue Tmp1 = Op.getOperand(1);
2441 DebugLoc dl = Op.getDebugLoc();
2442 EVT VT = Op.getValueType();
2443 EVT SrcVT = Tmp1.getValueType();
2444 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2445 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2446 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2447 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2448 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2451 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2452 MachineFunction &MF = DAG.getMachineFunction();
2453 MachineFrameInfo *MFI = MF.getFrameInfo();
2454 MFI->setReturnAddressIsTaken(true);
2456 EVT VT = Op.getValueType();
2457 DebugLoc dl = Op.getDebugLoc();
2458 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2460 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2461 SDValue Offset = DAG.getConstant(4, MVT::i32);
2462 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2463 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2464 NULL, 0, false, false, 0);
2467 // Return LR, which contains the return address. Mark it an implicit live-in.
2468 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2469 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2472 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2473 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2474 MFI->setFrameAddressIsTaken(true);
2476 EVT VT = Op.getValueType();
2477 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2478 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2479 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2480 ? ARM::R7 : ARM::R11;
2481 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2483 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2488 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2489 /// expand a bit convert where either the source or destination type is i64 to
2490 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2491 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2492 /// vectors), since the legalizer won't know what to do with that.
2493 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2494 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2495 DebugLoc dl = N->getDebugLoc();
2496 SDValue Op = N->getOperand(0);
2498 // This function is only supposed to be called for i64 types, either as the
2499 // source or destination of the bit convert.
2500 EVT SrcVT = Op.getValueType();
2501 EVT DstVT = N->getValueType(0);
2502 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2503 "ExpandBIT_CONVERT called for non-i64 type");
2505 // Turn i64->f64 into VMOVDRR.
2506 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2507 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2508 DAG.getConstant(0, MVT::i32));
2509 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2510 DAG.getConstant(1, MVT::i32));
2511 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2512 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2515 // Turn f64->i64 into VMOVRRD.
2516 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2517 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2518 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2519 // Merge the pieces into a single i64 value.
2520 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2526 /// getZeroVector - Returns a vector of specified type with all zero elements.
2528 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2529 assert(VT.isVector() && "Expected a vector type");
2531 // Zero vectors are used to represent vector negation and in those cases
2532 // will be implemented with the NEON VNEG instruction. However, VNEG does
2533 // not support i64 elements, so sometimes the zero vectors will need to be
2534 // explicitly constructed. For those cases, and potentially other uses in
2535 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2536 // to their dest type. This ensures they get CSE'd.
2538 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2539 SmallVector<SDValue, 8> Ops;
2542 if (VT.getSizeInBits() == 64) {
2543 Ops.assign(8, Cst); TVT = MVT::v8i8;
2545 Ops.assign(16, Cst); TVT = MVT::v16i8;
2547 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2549 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2552 /// getOnesVector - Returns a vector of specified type with all bits set.
2554 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2555 assert(VT.isVector() && "Expected a vector type");
2557 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2558 // dest type. This ensures they get CSE'd.
2560 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2561 SmallVector<SDValue, 8> Ops;
2564 if (VT.getSizeInBits() == 64) {
2565 Ops.assign(8, Cst); TVT = MVT::v8i8;
2567 Ops.assign(16, Cst); TVT = MVT::v16i8;
2569 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2571 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2574 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2575 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2576 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2577 SelectionDAG &DAG) const {
2578 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2579 EVT VT = Op.getValueType();
2580 unsigned VTBits = VT.getSizeInBits();
2581 DebugLoc dl = Op.getDebugLoc();
2582 SDValue ShOpLo = Op.getOperand(0);
2583 SDValue ShOpHi = Op.getOperand(1);
2584 SDValue ShAmt = Op.getOperand(2);
2586 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2588 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2590 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2591 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2592 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2593 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2594 DAG.getConstant(VTBits, MVT::i32));
2595 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2596 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2597 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2599 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2600 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2602 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2603 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2606 SDValue Ops[2] = { Lo, Hi };
2607 return DAG.getMergeValues(Ops, 2, dl);
2610 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2611 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2612 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2613 SelectionDAG &DAG) const {
2614 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2615 EVT VT = Op.getValueType();
2616 unsigned VTBits = VT.getSizeInBits();
2617 DebugLoc dl = Op.getDebugLoc();
2618 SDValue ShOpLo = Op.getOperand(0);
2619 SDValue ShOpHi = Op.getOperand(1);
2620 SDValue ShAmt = Op.getOperand(2);
2623 assert(Op.getOpcode() == ISD::SHL_PARTS);
2624 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2625 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2626 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2627 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2628 DAG.getConstant(VTBits, MVT::i32));
2629 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2630 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2632 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2633 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2634 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2636 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2637 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2640 SDValue Ops[2] = { Lo, Hi };
2641 return DAG.getMergeValues(Ops, 2, dl);
2644 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2645 const ARMSubtarget *ST) {
2646 EVT VT = N->getValueType(0);
2647 DebugLoc dl = N->getDebugLoc();
2649 if (!ST->hasV6T2Ops())
2652 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2653 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2656 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2657 const ARMSubtarget *ST) {
2658 EVT VT = N->getValueType(0);
2659 DebugLoc dl = N->getDebugLoc();
2661 // Lower vector shifts on NEON to use VSHL.
2662 if (VT.isVector()) {
2663 assert(ST->hasNEON() && "unexpected vector shift");
2665 // Left shifts translate directly to the vshiftu intrinsic.
2666 if (N->getOpcode() == ISD::SHL)
2667 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2668 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2669 N->getOperand(0), N->getOperand(1));
2671 assert((N->getOpcode() == ISD::SRA ||
2672 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2674 // NEON uses the same intrinsics for both left and right shifts. For
2675 // right shifts, the shift amounts are negative, so negate the vector of
2677 EVT ShiftVT = N->getOperand(1).getValueType();
2678 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2679 getZeroVector(ShiftVT, DAG, dl),
2681 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2682 Intrinsic::arm_neon_vshifts :
2683 Intrinsic::arm_neon_vshiftu);
2684 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2685 DAG.getConstant(vshiftInt, MVT::i32),
2686 N->getOperand(0), NegatedCount);
2689 // We can get here for a node like i32 = ISD::SHL i32, i64
2693 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2694 "Unknown shift to lower!");
2696 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2697 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2698 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2701 // If we are in thumb mode, we don't have RRX.
2702 if (ST->isThumb1Only()) return SDValue();
2704 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2705 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2706 DAG.getConstant(0, MVT::i32));
2707 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2708 DAG.getConstant(1, MVT::i32));
2710 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2711 // captures the result into a carry flag.
2712 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2713 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2715 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2716 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2718 // Merge the pieces into a single i64 value.
2719 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2722 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2723 SDValue TmpOp0, TmpOp1;
2724 bool Invert = false;
2728 SDValue Op0 = Op.getOperand(0);
2729 SDValue Op1 = Op.getOperand(1);
2730 SDValue CC = Op.getOperand(2);
2731 EVT VT = Op.getValueType();
2732 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2733 DebugLoc dl = Op.getDebugLoc();
2735 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2736 switch (SetCCOpcode) {
2737 default: llvm_unreachable("Illegal FP comparison"); break;
2739 case ISD::SETNE: Invert = true; // Fallthrough
2741 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2743 case ISD::SETLT: Swap = true; // Fallthrough
2745 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2747 case ISD::SETLE: Swap = true; // Fallthrough
2749 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2750 case ISD::SETUGE: Swap = true; // Fallthrough
2751 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2752 case ISD::SETUGT: Swap = true; // Fallthrough
2753 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2754 case ISD::SETUEQ: Invert = true; // Fallthrough
2756 // Expand this to (OLT | OGT).
2760 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2761 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2763 case ISD::SETUO: Invert = true; // Fallthrough
2765 // Expand this to (OLT | OGE).
2769 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2770 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2774 // Integer comparisons.
2775 switch (SetCCOpcode) {
2776 default: llvm_unreachable("Illegal integer comparison"); break;
2777 case ISD::SETNE: Invert = true;
2778 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2779 case ISD::SETLT: Swap = true;
2780 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2781 case ISD::SETLE: Swap = true;
2782 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2783 case ISD::SETULT: Swap = true;
2784 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2785 case ISD::SETULE: Swap = true;
2786 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2789 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2790 if (Opc == ARMISD::VCEQ) {
2793 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2795 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2798 // Ignore bitconvert.
2799 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2800 AndOp = AndOp.getOperand(0);
2802 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2804 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2805 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2812 std::swap(Op0, Op1);
2814 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2817 Result = DAG.getNOT(dl, Result, VT);
2822 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2823 /// valid vector constant for a NEON instruction with a "modified immediate"
2824 /// operand (e.g., VMOV). If so, return either the constant being
2825 /// splatted or the encoded value, depending on the DoEncode parameter. The
2826 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2827 /// bits7-0=Immediate.
2828 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2829 unsigned SplatBitSize, SelectionDAG &DAG,
2830 bool isVMOV, bool DoEncode) {
2831 unsigned Op, Cmode, Imm;
2834 // SplatBitSize is set to the smallest size that splats the vector, so a
2835 // zero vector will always have SplatBitSize == 8. However, NEON modified
2836 // immediate instructions others than VMOV do not support the 8-bit encoding
2837 // of a zero vector, and the default encoding of zero is supposed to be the
2843 switch (SplatBitSize) {
2845 // Any 1-byte value is OK. Op=0, Cmode=1110.
2846 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2853 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2855 if ((SplatBits & ~0xff) == 0) {
2856 // Value = 0x00nn: Op=x, Cmode=100x.
2861 if ((SplatBits & ~0xff00) == 0) {
2862 // Value = 0xnn00: Op=x, Cmode=101x.
2864 Imm = SplatBits >> 8;
2870 // NEON's 32-bit VMOV supports splat values where:
2871 // * only one byte is nonzero, or
2872 // * the least significant byte is 0xff and the second byte is nonzero, or
2873 // * the least significant 2 bytes are 0xff and the third is nonzero.
2875 if ((SplatBits & ~0xff) == 0) {
2876 // Value = 0x000000nn: Op=x, Cmode=000x.
2881 if ((SplatBits & ~0xff00) == 0) {
2882 // Value = 0x0000nn00: Op=x, Cmode=001x.
2884 Imm = SplatBits >> 8;
2887 if ((SplatBits & ~0xff0000) == 0) {
2888 // Value = 0x00nn0000: Op=x, Cmode=010x.
2890 Imm = SplatBits >> 16;
2893 if ((SplatBits & ~0xff000000) == 0) {
2894 // Value = 0xnn000000: Op=x, Cmode=011x.
2896 Imm = SplatBits >> 24;
2900 if ((SplatBits & ~0xffff) == 0 &&
2901 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2902 // Value = 0x0000nnff: Op=x, Cmode=1100.
2904 Imm = SplatBits >> 8;
2909 if ((SplatBits & ~0xffffff) == 0 &&
2910 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2911 // Value = 0x00nnffff: Op=x, Cmode=1101.
2913 Imm = SplatBits >> 16;
2914 SplatBits |= 0xffff;
2918 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2919 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2920 // VMOV.I32. A (very) minor optimization would be to replicate the value
2921 // and fall through here to test for a valid 64-bit splat. But, then the
2922 // caller would also need to check and handle the change in size.
2926 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2929 uint64_t BitMask = 0xff;
2931 unsigned ImmMask = 1;
2933 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2934 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2937 } else if ((SplatBits & BitMask) != 0) {
2943 // Op=1, Cmode=1110.
2952 llvm_unreachable("unexpected size for isNEONModifiedImm");
2957 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2958 return DAG.getTargetConstant(SplatBits, VT);
2962 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2963 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2964 /// size, return the encoded value for that immediate. The ByteSize field
2965 /// indicates the number of bytes of each element [1248].
2966 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2967 SelectionDAG &DAG) {
2968 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2969 APInt SplatBits, SplatUndef;
2970 unsigned SplatBitSize;
2972 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2973 HasAnyUndefs, ByteSize * 8))
2976 if (SplatBitSize > ByteSize * 8)
2979 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2980 SplatBitSize, DAG, isVMOV, true);
2983 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2984 bool &ReverseVEXT, unsigned &Imm) {
2985 unsigned NumElts = VT.getVectorNumElements();
2986 ReverseVEXT = false;
2989 // If this is a VEXT shuffle, the immediate value is the index of the first
2990 // element. The other shuffle indices must be the successive elements after
2992 unsigned ExpectedElt = Imm;
2993 for (unsigned i = 1; i < NumElts; ++i) {
2994 // Increment the expected index. If it wraps around, it may still be
2995 // a VEXT but the source vectors must be swapped.
2997 if (ExpectedElt == NumElts * 2) {
3002 if (ExpectedElt != static_cast<unsigned>(M[i]))
3006 // Adjust the index value if the source operands will be swapped.
3013 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3014 /// instruction with the specified blocksize. (The order of the elements
3015 /// within each block of the vector is reversed.)
3016 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3017 unsigned BlockSize) {
3018 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3019 "Only possible block sizes for VREV are: 16, 32, 64");
3021 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3025 unsigned NumElts = VT.getVectorNumElements();
3026 unsigned BlockElts = M[0] + 1;
3028 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3031 for (unsigned i = 0; i < NumElts; ++i) {
3032 if ((unsigned) M[i] !=
3033 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3040 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3041 unsigned &WhichResult) {
3042 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3046 unsigned NumElts = VT.getVectorNumElements();
3047 WhichResult = (M[0] == 0 ? 0 : 1);
3048 for (unsigned i = 0; i < NumElts; i += 2) {
3049 if ((unsigned) M[i] != i + WhichResult ||
3050 (unsigned) M[i+1] != i + NumElts + WhichResult)
3056 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3057 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3058 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3059 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3060 unsigned &WhichResult) {
3061 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3065 unsigned NumElts = VT.getVectorNumElements();
3066 WhichResult = (M[0] == 0 ? 0 : 1);
3067 for (unsigned i = 0; i < NumElts; i += 2) {
3068 if ((unsigned) M[i] != i + WhichResult ||
3069 (unsigned) M[i+1] != i + WhichResult)
3075 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3076 unsigned &WhichResult) {
3077 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3081 unsigned NumElts = VT.getVectorNumElements();
3082 WhichResult = (M[0] == 0 ? 0 : 1);
3083 for (unsigned i = 0; i != NumElts; ++i) {
3084 if ((unsigned) M[i] != 2 * i + WhichResult)
3088 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3089 if (VT.is64BitVector() && EltSz == 32)
3095 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3096 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3097 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3098 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3099 unsigned &WhichResult) {
3100 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3104 unsigned Half = VT.getVectorNumElements() / 2;
3105 WhichResult = (M[0] == 0 ? 0 : 1);
3106 for (unsigned j = 0; j != 2; ++j) {
3107 unsigned Idx = WhichResult;
3108 for (unsigned i = 0; i != Half; ++i) {
3109 if ((unsigned) M[i + j * Half] != Idx)
3115 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3116 if (VT.is64BitVector() && EltSz == 32)
3122 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3123 unsigned &WhichResult) {
3124 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3128 unsigned NumElts = VT.getVectorNumElements();
3129 WhichResult = (M[0] == 0 ? 0 : 1);
3130 unsigned Idx = WhichResult * NumElts / 2;
3131 for (unsigned i = 0; i != NumElts; i += 2) {
3132 if ((unsigned) M[i] != Idx ||
3133 (unsigned) M[i+1] != Idx + NumElts)
3138 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3139 if (VT.is64BitVector() && EltSz == 32)
3145 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3146 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3147 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3148 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3149 unsigned &WhichResult) {
3150 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3154 unsigned NumElts = VT.getVectorNumElements();
3155 WhichResult = (M[0] == 0 ? 0 : 1);
3156 unsigned Idx = WhichResult * NumElts / 2;
3157 for (unsigned i = 0; i != NumElts; i += 2) {
3158 if ((unsigned) M[i] != Idx ||
3159 (unsigned) M[i+1] != Idx)
3164 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3165 if (VT.is64BitVector() && EltSz == 32)
3172 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3173 // Canonicalize all-zeros and all-ones vectors.
3174 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3175 if (ConstVal->isNullValue())
3176 return getZeroVector(VT, DAG, dl);
3177 if (ConstVal->isAllOnesValue())
3178 return getOnesVector(VT, DAG, dl);
3181 if (VT.is64BitVector()) {
3182 switch (Val.getValueType().getSizeInBits()) {
3183 case 8: CanonicalVT = MVT::v8i8; break;
3184 case 16: CanonicalVT = MVT::v4i16; break;
3185 case 32: CanonicalVT = MVT::v2i32; break;
3186 case 64: CanonicalVT = MVT::v1i64; break;
3187 default: llvm_unreachable("unexpected splat element type"); break;
3190 assert(VT.is128BitVector() && "unknown splat vector size");
3191 switch (Val.getValueType().getSizeInBits()) {
3192 case 8: CanonicalVT = MVT::v16i8; break;
3193 case 16: CanonicalVT = MVT::v8i16; break;
3194 case 32: CanonicalVT = MVT::v4i32; break;
3195 case 64: CanonicalVT = MVT::v2i64; break;
3196 default: llvm_unreachable("unexpected splat element type"); break;
3200 // Build a canonical splat for this value.
3201 SmallVector<SDValue, 8> Ops;
3202 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3203 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3205 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3208 // If this is a case we can't handle, return null and let the default
3209 // expansion code take care of it.
3210 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3211 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3212 DebugLoc dl = Op.getDebugLoc();
3213 EVT VT = Op.getValueType();
3215 APInt SplatBits, SplatUndef;
3216 unsigned SplatBitSize;
3218 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3219 if (SplatBitSize <= 64) {
3220 // Check if an immediate VMOV works.
3221 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3222 SplatUndef.getZExtValue(),
3223 SplatBitSize, DAG, true, false);
3225 return BuildSplat(Val, VT, DAG, dl);
3229 // Scan through the operands to see if only one value is used.
3230 unsigned NumElts = VT.getVectorNumElements();
3231 bool isOnlyLowElement = true;
3232 bool usesOnlyOneValue = true;
3233 bool isConstant = true;
3235 for (unsigned i = 0; i < NumElts; ++i) {
3236 SDValue V = Op.getOperand(i);
3237 if (V.getOpcode() == ISD::UNDEF)
3240 isOnlyLowElement = false;
3241 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3244 if (!Value.getNode())
3246 else if (V != Value)
3247 usesOnlyOneValue = false;
3250 if (!Value.getNode())
3251 return DAG.getUNDEF(VT);
3253 if (isOnlyLowElement)
3254 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3256 // If all elements are constants, fall back to the default expansion, which
3257 // will generate a load from the constant pool.
3261 // Use VDUP for non-constant splats.
3262 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3263 if (usesOnlyOneValue && EltSize <= 32)
3264 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3266 // Vectors with 32- or 64-bit elements can be built by directly assigning
3267 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3268 // will be legalized.
3269 if (EltSize >= 32) {
3270 // Do the expansion with floating-point types, since that is what the VFP
3271 // registers are defined to use, and since i64 is not legal.
3272 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3273 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3274 SmallVector<SDValue, 8> Ops;
3275 for (unsigned i = 0; i < NumElts; ++i)
3276 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3277 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3278 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3284 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3285 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3286 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3287 /// are assumed to be legal.
3289 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3291 if (VT.getVectorNumElements() == 4 &&
3292 (VT.is128BitVector() || VT.is64BitVector())) {
3293 unsigned PFIndexes[4];
3294 for (unsigned i = 0; i != 4; ++i) {
3298 PFIndexes[i] = M[i];
3301 // Compute the index in the perfect shuffle table.
3302 unsigned PFTableIndex =
3303 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3304 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3305 unsigned Cost = (PFEntry >> 30);
3312 unsigned Imm, WhichResult;
3314 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3315 return (EltSize >= 32 ||
3316 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3317 isVREVMask(M, VT, 64) ||
3318 isVREVMask(M, VT, 32) ||
3319 isVREVMask(M, VT, 16) ||
3320 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3321 isVTRNMask(M, VT, WhichResult) ||
3322 isVUZPMask(M, VT, WhichResult) ||
3323 isVZIPMask(M, VT, WhichResult) ||
3324 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3325 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3326 isVZIP_v_undef_Mask(M, VT, WhichResult));
3329 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3330 /// the specified operations to build the shuffle.
3331 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3332 SDValue RHS, SelectionDAG &DAG,
3334 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3335 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3336 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3339 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3348 OP_VUZPL, // VUZP, left result
3349 OP_VUZPR, // VUZP, right result
3350 OP_VZIPL, // VZIP, left result
3351 OP_VZIPR, // VZIP, right result
3352 OP_VTRNL, // VTRN, left result
3353 OP_VTRNR // VTRN, right result
3356 if (OpNum == OP_COPY) {
3357 if (LHSID == (1*9+2)*9+3) return LHS;
3358 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3362 SDValue OpLHS, OpRHS;
3363 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3364 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3365 EVT VT = OpLHS.getValueType();
3368 default: llvm_unreachable("Unknown shuffle opcode!");
3370 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3375 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3376 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3380 return DAG.getNode(ARMISD::VEXT, dl, VT,
3382 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3385 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3386 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3389 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3390 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3393 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3394 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3398 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3399 SDValue V1 = Op.getOperand(0);
3400 SDValue V2 = Op.getOperand(1);
3401 DebugLoc dl = Op.getDebugLoc();
3402 EVT VT = Op.getValueType();
3403 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3404 SmallVector<int, 8> ShuffleMask;
3406 // Convert shuffles that are directly supported on NEON to target-specific
3407 // DAG nodes, instead of keeping them as shuffles and matching them again
3408 // during code selection. This is more efficient and avoids the possibility
3409 // of inconsistencies between legalization and selection.
3410 // FIXME: floating-point vectors should be canonicalized to integer vectors
3411 // of the same time so that they get CSEd properly.
3412 SVN->getMask(ShuffleMask);
3414 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3415 if (EltSize <= 32) {
3416 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3417 int Lane = SVN->getSplatIndex();
3418 // If this is undef splat, generate it via "just" vdup, if possible.
3419 if (Lane == -1) Lane = 0;
3421 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3422 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3424 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3425 DAG.getConstant(Lane, MVT::i32));
3430 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3433 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3434 DAG.getConstant(Imm, MVT::i32));
3437 if (isVREVMask(ShuffleMask, VT, 64))
3438 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3439 if (isVREVMask(ShuffleMask, VT, 32))
3440 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3441 if (isVREVMask(ShuffleMask, VT, 16))
3442 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3444 // Check for Neon shuffles that modify both input vectors in place.
3445 // If both results are used, i.e., if there are two shuffles with the same
3446 // source operands and with masks corresponding to both results of one of
3447 // these operations, DAG memoization will ensure that a single node is
3448 // used for both shuffles.
3449 unsigned WhichResult;
3450 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3451 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3452 V1, V2).getValue(WhichResult);
3453 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3454 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3455 V1, V2).getValue(WhichResult);
3456 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3457 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3458 V1, V2).getValue(WhichResult);
3460 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3461 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3462 V1, V1).getValue(WhichResult);
3463 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3464 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3465 V1, V1).getValue(WhichResult);
3466 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3467 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3468 V1, V1).getValue(WhichResult);
3471 // If the shuffle is not directly supported and it has 4 elements, use
3472 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3473 unsigned NumElts = VT.getVectorNumElements();
3475 unsigned PFIndexes[4];
3476 for (unsigned i = 0; i != 4; ++i) {
3477 if (ShuffleMask[i] < 0)
3480 PFIndexes[i] = ShuffleMask[i];
3483 // Compute the index in the perfect shuffle table.
3484 unsigned PFTableIndex =
3485 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3486 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3487 unsigned Cost = (PFEntry >> 30);
3490 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3493 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3494 if (EltSize >= 32) {
3495 // Do the expansion with floating-point types, since that is what the VFP
3496 // registers are defined to use, and since i64 is not legal.
3497 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3498 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3499 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3500 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3501 SmallVector<SDValue, 8> Ops;
3502 for (unsigned i = 0; i < NumElts; ++i) {
3503 if (ShuffleMask[i] < 0)
3504 Ops.push_back(DAG.getUNDEF(EltVT));
3506 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3507 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3508 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3511 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3512 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3518 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3519 EVT VT = Op.getValueType();
3520 DebugLoc dl = Op.getDebugLoc();
3521 SDValue Vec = Op.getOperand(0);
3522 SDValue Lane = Op.getOperand(1);
3523 assert(VT == MVT::i32 &&
3524 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3525 "unexpected type for custom-lowering vector extract");
3526 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3529 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3530 // The only time a CONCAT_VECTORS operation can have legal types is when
3531 // two 64-bit vectors are concatenated to a 128-bit vector.
3532 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3533 "unexpected CONCAT_VECTORS");
3534 DebugLoc dl = Op.getDebugLoc();
3535 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3536 SDValue Op0 = Op.getOperand(0);
3537 SDValue Op1 = Op.getOperand(1);
3538 if (Op0.getOpcode() != ISD::UNDEF)
3539 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3540 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3541 DAG.getIntPtrConstant(0));
3542 if (Op1.getOpcode() != ISD::UNDEF)
3543 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3544 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3545 DAG.getIntPtrConstant(1));
3546 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3549 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3550 switch (Op.getOpcode()) {
3551 default: llvm_unreachable("Don't know how to custom lower this!");
3552 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3553 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3554 case ISD::GlobalAddress:
3555 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3556 LowerGlobalAddressELF(Op, DAG);
3557 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3558 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3559 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3560 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3561 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3562 case ISD::VASTART: return LowerVASTART(Op, DAG);
3563 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3564 case ISD::SINT_TO_FP:
3565 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3566 case ISD::FP_TO_SINT:
3567 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3568 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3569 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3570 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3571 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3572 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3573 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3574 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3576 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3579 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3580 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3581 case ISD::SRL_PARTS:
3582 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3583 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3584 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3585 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3586 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3587 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3588 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3593 /// ReplaceNodeResults - Replace the results of node with an illegal result
3594 /// type with new values built out of custom code.
3595 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3596 SmallVectorImpl<SDValue>&Results,
3597 SelectionDAG &DAG) const {
3599 switch (N->getOpcode()) {
3601 llvm_unreachable("Don't know how to custom expand this!");
3603 case ISD::BIT_CONVERT:
3604 Res = ExpandBIT_CONVERT(N, DAG);
3608 Res = LowerShift(N, DAG, Subtarget);
3612 Results.push_back(Res);
3615 //===----------------------------------------------------------------------===//
3616 // ARM Scheduler Hooks
3617 //===----------------------------------------------------------------------===//
3620 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3621 MachineBasicBlock *BB,
3622 unsigned Size) const {
3623 unsigned dest = MI->getOperand(0).getReg();
3624 unsigned ptr = MI->getOperand(1).getReg();
3625 unsigned oldval = MI->getOperand(2).getReg();
3626 unsigned newval = MI->getOperand(3).getReg();
3627 unsigned scratch = BB->getParent()->getRegInfo()
3628 .createVirtualRegister(ARM::GPRRegisterClass);
3629 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3630 DebugLoc dl = MI->getDebugLoc();
3631 bool isThumb2 = Subtarget->isThumb2();
3633 unsigned ldrOpc, strOpc;
3635 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3637 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3638 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3641 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3642 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3645 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3646 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3650 MachineFunction *MF = BB->getParent();
3651 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3652 MachineFunction::iterator It = BB;
3653 ++It; // insert the new blocks after the current block
3655 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3656 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3657 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3658 MF->insert(It, loop1MBB);
3659 MF->insert(It, loop2MBB);
3660 MF->insert(It, exitMBB);
3661 exitMBB->transferSuccessors(BB);
3665 // fallthrough --> loop1MBB
3666 BB->addSuccessor(loop1MBB);
3669 // ldrex dest, [ptr]
3673 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3674 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3675 .addReg(dest).addReg(oldval));
3676 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3677 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3678 BB->addSuccessor(loop2MBB);
3679 BB->addSuccessor(exitMBB);
3682 // strex scratch, newval, [ptr]
3686 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3688 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3689 .addReg(scratch).addImm(0));
3690 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3691 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3692 BB->addSuccessor(loop1MBB);
3693 BB->addSuccessor(exitMBB);
3699 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3705 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3706 unsigned Size, unsigned BinOpcode) const {
3707 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3708 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3710 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3711 MachineFunction *MF = BB->getParent();
3712 MachineFunction::iterator It = BB;
3715 unsigned dest = MI->getOperand(0).getReg();
3716 unsigned ptr = MI->getOperand(1).getReg();
3717 unsigned incr = MI->getOperand(2).getReg();
3718 DebugLoc dl = MI->getDebugLoc();
3720 bool isThumb2 = Subtarget->isThumb2();
3721 unsigned ldrOpc, strOpc;
3723 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3725 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3726 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3729 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3730 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3733 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3734 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3738 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3739 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3740 MF->insert(It, loopMBB);
3741 MF->insert(It, exitMBB);
3742 exitMBB->transferSuccessors(BB);
3744 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3745 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3746 unsigned scratch2 = (!BinOpcode) ? incr :
3747 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3751 // fallthrough --> loopMBB
3752 BB->addSuccessor(loopMBB);
3756 // <binop> scratch2, dest, incr
3757 // strex scratch, scratch2, ptr
3760 // fallthrough --> exitMBB
3762 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3764 // operand order needs to go the other way for NAND
3765 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3766 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3767 addReg(incr).addReg(dest)).addReg(0);
3769 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3770 addReg(dest).addReg(incr)).addReg(0);
3773 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3775 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3776 .addReg(scratch).addImm(0));
3777 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3778 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3780 BB->addSuccessor(loopMBB);
3781 BB->addSuccessor(exitMBB);
3787 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3793 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3794 MachineBasicBlock *BB) const {
3795 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3796 DebugLoc dl = MI->getDebugLoc();
3797 bool isThumb2 = Subtarget->isThumb2();
3798 switch (MI->getOpcode()) {
3801 llvm_unreachable("Unexpected instr type to insert");
3803 case ARM::ATOMIC_LOAD_ADD_I8:
3804 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3805 case ARM::ATOMIC_LOAD_ADD_I16:
3806 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3807 case ARM::ATOMIC_LOAD_ADD_I32:
3808 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3810 case ARM::ATOMIC_LOAD_AND_I8:
3811 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3812 case ARM::ATOMIC_LOAD_AND_I16:
3813 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3814 case ARM::ATOMIC_LOAD_AND_I32:
3815 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3817 case ARM::ATOMIC_LOAD_OR_I8:
3818 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3819 case ARM::ATOMIC_LOAD_OR_I16:
3820 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3821 case ARM::ATOMIC_LOAD_OR_I32:
3822 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3824 case ARM::ATOMIC_LOAD_XOR_I8:
3825 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3826 case ARM::ATOMIC_LOAD_XOR_I16:
3827 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3828 case ARM::ATOMIC_LOAD_XOR_I32:
3829 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3831 case ARM::ATOMIC_LOAD_NAND_I8:
3832 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3833 case ARM::ATOMIC_LOAD_NAND_I16:
3834 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3835 case ARM::ATOMIC_LOAD_NAND_I32:
3836 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3838 case ARM::ATOMIC_LOAD_SUB_I8:
3839 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3840 case ARM::ATOMIC_LOAD_SUB_I16:
3841 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3842 case ARM::ATOMIC_LOAD_SUB_I32:
3843 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3845 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3846 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3847 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3849 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3850 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3851 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3853 case ARM::tMOVCCr_pseudo: {
3854 // To "insert" a SELECT_CC instruction, we actually have to insert the
3855 // diamond control-flow pattern. The incoming instruction knows the
3856 // destination vreg to set, the condition code register to branch on, the
3857 // true/false values to select between, and a branch opcode to use.
3858 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3859 MachineFunction::iterator It = BB;
3865 // cmpTY ccX, r1, r2
3867 // fallthrough --> copy0MBB
3868 MachineBasicBlock *thisMBB = BB;
3869 MachineFunction *F = BB->getParent();
3870 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3871 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3872 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3873 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3874 F->insert(It, copy0MBB);
3875 F->insert(It, sinkMBB);
3876 // Update machine-CFG edges by first adding all successors of the current
3877 // block to the new block which will contain the Phi node for the select.
3878 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3879 E = BB->succ_end(); I != E; ++I)
3880 sinkMBB->addSuccessor(*I);
3881 // Next, remove all successors of the current block, and add the true
3882 // and fallthrough blocks as its successors.
3883 while (!BB->succ_empty())
3884 BB->removeSuccessor(BB->succ_begin());
3885 BB->addSuccessor(copy0MBB);
3886 BB->addSuccessor(sinkMBB);
3889 // %FalseValue = ...
3890 // # fallthrough to sinkMBB
3893 // Update machine-CFG edges
3894 BB->addSuccessor(sinkMBB);
3897 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3900 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3901 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3902 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3904 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3911 case ARM::t2SUBrSPi_:
3912 case ARM::t2SUBrSPi12_:
3913 case ARM::t2SUBrSPs_: {
3914 MachineFunction *MF = BB->getParent();
3915 unsigned DstReg = MI->getOperand(0).getReg();
3916 unsigned SrcReg = MI->getOperand(1).getReg();
3917 bool DstIsDead = MI->getOperand(0).isDead();
3918 bool SrcIsKill = MI->getOperand(1).isKill();
3920 if (SrcReg != ARM::SP) {
3921 // Copy the source to SP from virtual register.
3922 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3923 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3924 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3925 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3926 .addReg(SrcReg, getKillRegState(SrcIsKill));
3930 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3931 switch (MI->getOpcode()) {
3933 llvm_unreachable("Unexpected pseudo instruction!");
3939 OpOpc = ARM::tADDspr;
3942 OpOpc = ARM::tSUBspi;
3944 case ARM::t2SUBrSPi_:
3945 OpOpc = ARM::t2SUBrSPi;
3946 NeedPred = true; NeedCC = true;
3948 case ARM::t2SUBrSPi12_:
3949 OpOpc = ARM::t2SUBrSPi12;
3952 case ARM::t2SUBrSPs_:
3953 OpOpc = ARM::t2SUBrSPs;
3954 NeedPred = true; NeedCC = true; NeedOp3 = true;
3957 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3958 if (OpOpc == ARM::tAND)
3959 AddDefaultT1CC(MIB);
3960 MIB.addReg(ARM::SP);
3961 MIB.addOperand(MI->getOperand(2));
3963 MIB.addOperand(MI->getOperand(3));
3965 AddDefaultPred(MIB);
3969 // Copy the result from SP to virtual register.
3970 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3971 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3972 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3973 BuildMI(BB, dl, TII->get(CopyOpc))
3974 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3976 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3982 //===----------------------------------------------------------------------===//
3983 // ARM Optimization Hooks
3984 //===----------------------------------------------------------------------===//
3987 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3988 TargetLowering::DAGCombinerInfo &DCI) {
3989 SelectionDAG &DAG = DCI.DAG;
3990 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3991 EVT VT = N->getValueType(0);
3992 unsigned Opc = N->getOpcode();
3993 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3994 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3995 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3996 ISD::CondCode CC = ISD::SETCC_INVALID;
3999 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4001 SDValue CCOp = Slct.getOperand(0);
4002 if (CCOp.getOpcode() == ISD::SETCC)
4003 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4006 bool DoXform = false;
4008 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4011 if (LHS.getOpcode() == ISD::Constant &&
4012 cast<ConstantSDNode>(LHS)->isNullValue()) {
4014 } else if (CC != ISD::SETCC_INVALID &&
4015 RHS.getOpcode() == ISD::Constant &&
4016 cast<ConstantSDNode>(RHS)->isNullValue()) {
4017 std::swap(LHS, RHS);
4018 SDValue Op0 = Slct.getOperand(0);
4019 EVT OpVT = isSlctCC ? Op0.getValueType() :
4020 Op0.getOperand(0).getValueType();
4021 bool isInt = OpVT.isInteger();
4022 CC = ISD::getSetCCInverse(CC, isInt);
4024 if (!TLI.isCondCodeLegal(CC, OpVT))
4025 return SDValue(); // Inverse operator isn't legal.
4032 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4034 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4035 Slct.getOperand(0), Slct.getOperand(1), CC);
4036 SDValue CCOp = Slct.getOperand(0);
4038 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4039 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4040 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4041 CCOp, OtherOp, Result);
4046 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4047 static SDValue PerformADDCombine(SDNode *N,
4048 TargetLowering::DAGCombinerInfo &DCI) {
4049 // added by evan in r37685 with no testcase.
4050 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4052 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4053 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4054 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4055 if (Result.getNode()) return Result;
4057 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4058 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4059 if (Result.getNode()) return Result;
4065 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4066 static SDValue PerformSUBCombine(SDNode *N,
4067 TargetLowering::DAGCombinerInfo &DCI) {
4068 // added by evan in r37685 with no testcase.
4069 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4071 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4072 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4073 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4074 if (Result.getNode()) return Result;
4080 static SDValue PerformMULCombine(SDNode *N,
4081 TargetLowering::DAGCombinerInfo &DCI,
4082 const ARMSubtarget *Subtarget) {
4083 SelectionDAG &DAG = DCI.DAG;
4085 if (Subtarget->isThumb1Only())
4088 if (DAG.getMachineFunction().
4089 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4092 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4095 EVT VT = N->getValueType(0);
4099 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4103 uint64_t MulAmt = C->getZExtValue();
4104 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4105 ShiftAmt = ShiftAmt & (32 - 1);
4106 SDValue V = N->getOperand(0);
4107 DebugLoc DL = N->getDebugLoc();
4110 MulAmt >>= ShiftAmt;
4111 if (isPowerOf2_32(MulAmt - 1)) {
4112 // (mul x, 2^N + 1) => (add (shl x, N), x)
4113 Res = DAG.getNode(ISD::ADD, DL, VT,
4114 V, DAG.getNode(ISD::SHL, DL, VT,
4115 V, DAG.getConstant(Log2_32(MulAmt-1),
4117 } else if (isPowerOf2_32(MulAmt + 1)) {
4118 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4119 Res = DAG.getNode(ISD::SUB, DL, VT,
4120 DAG.getNode(ISD::SHL, DL, VT,
4121 V, DAG.getConstant(Log2_32(MulAmt+1),
4128 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4129 DAG.getConstant(ShiftAmt, MVT::i32));
4131 // Do not add new nodes to DAG combiner worklist.
4132 DCI.CombineTo(N, Res, false);
4136 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4137 /// ARMISD::VMOVRRD.
4138 static SDValue PerformVMOVRRDCombine(SDNode *N,
4139 TargetLowering::DAGCombinerInfo &DCI) {
4140 // fmrrd(fmdrr x, y) -> x,y
4141 SDValue InDouble = N->getOperand(0);
4142 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4143 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4147 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4148 /// operand of a vector shift operation, where all the elements of the
4149 /// build_vector must have the same constant integer value.
4150 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4151 // Ignore bit_converts.
4152 while (Op.getOpcode() == ISD::BIT_CONVERT)
4153 Op = Op.getOperand(0);
4154 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4155 APInt SplatBits, SplatUndef;
4156 unsigned SplatBitSize;
4158 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4159 HasAnyUndefs, ElementBits) ||
4160 SplatBitSize > ElementBits)
4162 Cnt = SplatBits.getSExtValue();
4166 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4167 /// operand of a vector shift left operation. That value must be in the range:
4168 /// 0 <= Value < ElementBits for a left shift; or
4169 /// 0 <= Value <= ElementBits for a long left shift.
4170 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4171 assert(VT.isVector() && "vector shift count is not a vector type");
4172 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4173 if (! getVShiftImm(Op, ElementBits, Cnt))
4175 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4178 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4179 /// operand of a vector shift right operation. For a shift opcode, the value
4180 /// is positive, but for an intrinsic the value count must be negative. The
4181 /// absolute value must be in the range:
4182 /// 1 <= |Value| <= ElementBits for a right shift; or
4183 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4184 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4186 assert(VT.isVector() && "vector shift count is not a vector type");
4187 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4188 if (! getVShiftImm(Op, ElementBits, Cnt))
4192 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4195 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4196 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4197 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4200 // Don't do anything for most intrinsics.
4203 // Vector shifts: check for immediate versions and lower them.
4204 // Note: This is done during DAG combining instead of DAG legalizing because
4205 // the build_vectors for 64-bit vector element shift counts are generally
4206 // not legal, and it is hard to see their values after they get legalized to
4207 // loads from a constant pool.
4208 case Intrinsic::arm_neon_vshifts:
4209 case Intrinsic::arm_neon_vshiftu:
4210 case Intrinsic::arm_neon_vshiftls:
4211 case Intrinsic::arm_neon_vshiftlu:
4212 case Intrinsic::arm_neon_vshiftn:
4213 case Intrinsic::arm_neon_vrshifts:
4214 case Intrinsic::arm_neon_vrshiftu:
4215 case Intrinsic::arm_neon_vrshiftn:
4216 case Intrinsic::arm_neon_vqshifts:
4217 case Intrinsic::arm_neon_vqshiftu:
4218 case Intrinsic::arm_neon_vqshiftsu:
4219 case Intrinsic::arm_neon_vqshiftns:
4220 case Intrinsic::arm_neon_vqshiftnu:
4221 case Intrinsic::arm_neon_vqshiftnsu:
4222 case Intrinsic::arm_neon_vqrshiftns:
4223 case Intrinsic::arm_neon_vqrshiftnu:
4224 case Intrinsic::arm_neon_vqrshiftnsu: {
4225 EVT VT = N->getOperand(1).getValueType();
4227 unsigned VShiftOpc = 0;
4230 case Intrinsic::arm_neon_vshifts:
4231 case Intrinsic::arm_neon_vshiftu:
4232 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4233 VShiftOpc = ARMISD::VSHL;
4236 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4237 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4238 ARMISD::VSHRs : ARMISD::VSHRu);
4243 case Intrinsic::arm_neon_vshiftls:
4244 case Intrinsic::arm_neon_vshiftlu:
4245 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4247 llvm_unreachable("invalid shift count for vshll intrinsic");
4249 case Intrinsic::arm_neon_vrshifts:
4250 case Intrinsic::arm_neon_vrshiftu:
4251 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4255 case Intrinsic::arm_neon_vqshifts:
4256 case Intrinsic::arm_neon_vqshiftu:
4257 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4261 case Intrinsic::arm_neon_vqshiftsu:
4262 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4264 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4266 case Intrinsic::arm_neon_vshiftn:
4267 case Intrinsic::arm_neon_vrshiftn:
4268 case Intrinsic::arm_neon_vqshiftns:
4269 case Intrinsic::arm_neon_vqshiftnu:
4270 case Intrinsic::arm_neon_vqshiftnsu:
4271 case Intrinsic::arm_neon_vqrshiftns:
4272 case Intrinsic::arm_neon_vqrshiftnu:
4273 case Intrinsic::arm_neon_vqrshiftnsu:
4274 // Narrowing shifts require an immediate right shift.
4275 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4277 llvm_unreachable("invalid shift count for narrowing vector shift "
4281 llvm_unreachable("unhandled vector shift");
4285 case Intrinsic::arm_neon_vshifts:
4286 case Intrinsic::arm_neon_vshiftu:
4287 // Opcode already set above.
4289 case Intrinsic::arm_neon_vshiftls:
4290 case Intrinsic::arm_neon_vshiftlu:
4291 if (Cnt == VT.getVectorElementType().getSizeInBits())
4292 VShiftOpc = ARMISD::VSHLLi;
4294 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4295 ARMISD::VSHLLs : ARMISD::VSHLLu);
4297 case Intrinsic::arm_neon_vshiftn:
4298 VShiftOpc = ARMISD::VSHRN; break;
4299 case Intrinsic::arm_neon_vrshifts:
4300 VShiftOpc = ARMISD::VRSHRs; break;
4301 case Intrinsic::arm_neon_vrshiftu:
4302 VShiftOpc = ARMISD::VRSHRu; break;
4303 case Intrinsic::arm_neon_vrshiftn:
4304 VShiftOpc = ARMISD::VRSHRN; break;
4305 case Intrinsic::arm_neon_vqshifts:
4306 VShiftOpc = ARMISD::VQSHLs; break;
4307 case Intrinsic::arm_neon_vqshiftu:
4308 VShiftOpc = ARMISD::VQSHLu; break;
4309 case Intrinsic::arm_neon_vqshiftsu:
4310 VShiftOpc = ARMISD::VQSHLsu; break;
4311 case Intrinsic::arm_neon_vqshiftns:
4312 VShiftOpc = ARMISD::VQSHRNs; break;
4313 case Intrinsic::arm_neon_vqshiftnu:
4314 VShiftOpc = ARMISD::VQSHRNu; break;
4315 case Intrinsic::arm_neon_vqshiftnsu:
4316 VShiftOpc = ARMISD::VQSHRNsu; break;
4317 case Intrinsic::arm_neon_vqrshiftns:
4318 VShiftOpc = ARMISD::VQRSHRNs; break;
4319 case Intrinsic::arm_neon_vqrshiftnu:
4320 VShiftOpc = ARMISD::VQRSHRNu; break;
4321 case Intrinsic::arm_neon_vqrshiftnsu:
4322 VShiftOpc = ARMISD::VQRSHRNsu; break;
4325 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4326 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4329 case Intrinsic::arm_neon_vshiftins: {
4330 EVT VT = N->getOperand(1).getValueType();
4332 unsigned VShiftOpc = 0;
4334 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4335 VShiftOpc = ARMISD::VSLI;
4336 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4337 VShiftOpc = ARMISD::VSRI;
4339 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4342 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4343 N->getOperand(1), N->getOperand(2),
4344 DAG.getConstant(Cnt, MVT::i32));
4347 case Intrinsic::arm_neon_vqrshifts:
4348 case Intrinsic::arm_neon_vqrshiftu:
4349 // No immediate versions of these to check for.
4356 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4357 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4358 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4359 /// vector element shift counts are generally not legal, and it is hard to see
4360 /// their values after they get legalized to loads from a constant pool.
4361 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4362 const ARMSubtarget *ST) {
4363 EVT VT = N->getValueType(0);
4365 // Nothing to be done for scalar shifts.
4366 if (! VT.isVector())
4369 assert(ST->hasNEON() && "unexpected vector shift");
4372 switch (N->getOpcode()) {
4373 default: llvm_unreachable("unexpected shift opcode");
4376 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4377 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4378 DAG.getConstant(Cnt, MVT::i32));
4383 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4384 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4385 ARMISD::VSHRs : ARMISD::VSHRu);
4386 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4387 DAG.getConstant(Cnt, MVT::i32));
4393 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4394 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4395 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4396 const ARMSubtarget *ST) {
4397 SDValue N0 = N->getOperand(0);
4399 // Check for sign- and zero-extensions of vector extract operations of 8-
4400 // and 16-bit vector elements. NEON supports these directly. They are
4401 // handled during DAG combining because type legalization will promote them
4402 // to 32-bit types and it is messy to recognize the operations after that.
4403 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4404 SDValue Vec = N0.getOperand(0);
4405 SDValue Lane = N0.getOperand(1);
4406 EVT VT = N->getValueType(0);
4407 EVT EltVT = N0.getValueType();
4408 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4410 if (VT == MVT::i32 &&
4411 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4412 TLI.isTypeLegal(Vec.getValueType())) {
4415 switch (N->getOpcode()) {
4416 default: llvm_unreachable("unexpected opcode");
4417 case ISD::SIGN_EXTEND:
4418 Opc = ARMISD::VGETLANEs;
4420 case ISD::ZERO_EXTEND:
4421 case ISD::ANY_EXTEND:
4422 Opc = ARMISD::VGETLANEu;
4425 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4432 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4433 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4434 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4435 const ARMSubtarget *ST) {
4436 // If the target supports NEON, try to use vmax/vmin instructions for f32
4437 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4438 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4439 // a NaN; only do the transformation when it matches that behavior.
4441 // For now only do this when using NEON for FP operations; if using VFP, it
4442 // is not obvious that the benefit outweighs the cost of switching to the
4444 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4445 N->getValueType(0) != MVT::f32)
4448 SDValue CondLHS = N->getOperand(0);
4449 SDValue CondRHS = N->getOperand(1);
4450 SDValue LHS = N->getOperand(2);
4451 SDValue RHS = N->getOperand(3);
4452 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4454 unsigned Opcode = 0;
4456 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4457 IsReversed = false; // x CC y ? x : y
4458 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4459 IsReversed = true ; // x CC y ? y : x
4473 // If LHS is NaN, an ordered comparison will be false and the result will
4474 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4475 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4476 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4477 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4479 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4480 // will return -0, so vmin can only be used for unsafe math or if one of
4481 // the operands is known to be nonzero.
4482 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4484 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4486 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4495 // If LHS is NaN, an ordered comparison will be false and the result will
4496 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4497 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4498 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4499 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4501 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4502 // will return +0, so vmax can only be used for unsafe math or if one of
4503 // the operands is known to be nonzero.
4504 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4506 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4508 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4514 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4517 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4518 DAGCombinerInfo &DCI) const {
4519 switch (N->getOpcode()) {
4521 case ISD::ADD: return PerformADDCombine(N, DCI);
4522 case ISD::SUB: return PerformSUBCombine(N, DCI);
4523 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4524 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4525 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4528 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4529 case ISD::SIGN_EXTEND:
4530 case ISD::ZERO_EXTEND:
4531 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4532 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4537 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4538 if (!Subtarget->hasV6Ops())
4539 // Pre-v6 does not support unaligned mem access.
4542 // v6+ may or may not support unaligned mem access depending on the system
4544 // FIXME: This is pretty conservative. Should we provide cmdline option to
4545 // control the behaviour?
4546 if (!Subtarget->isTargetDarwin())
4550 switch (VT.getSimpleVT().SimpleTy) {
4557 // FIXME: VLD1 etc with standard alignment is legal.
4561 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4566 switch (VT.getSimpleVT().SimpleTy) {
4567 default: return false;
4582 if ((V & (Scale - 1)) != 0)
4585 return V == (V & ((1LL << 5) - 1));
4588 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4589 const ARMSubtarget *Subtarget) {
4596 switch (VT.getSimpleVT().SimpleTy) {
4597 default: return false;
4602 // + imm12 or - imm8
4604 return V == (V & ((1LL << 8) - 1));
4605 return V == (V & ((1LL << 12) - 1));
4608 // Same as ARM mode. FIXME: NEON?
4609 if (!Subtarget->hasVFP2())
4614 return V == (V & ((1LL << 8) - 1));
4618 /// isLegalAddressImmediate - Return true if the integer value can be used
4619 /// as the offset of the target addressing mode for load / store of the
4621 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4622 const ARMSubtarget *Subtarget) {
4629 if (Subtarget->isThumb1Only())
4630 return isLegalT1AddressImmediate(V, VT);
4631 else if (Subtarget->isThumb2())
4632 return isLegalT2AddressImmediate(V, VT, Subtarget);
4637 switch (VT.getSimpleVT().SimpleTy) {
4638 default: return false;
4643 return V == (V & ((1LL << 12) - 1));
4646 return V == (V & ((1LL << 8) - 1));
4649 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4654 return V == (V & ((1LL << 8) - 1));
4658 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4660 int Scale = AM.Scale;
4664 switch (VT.getSimpleVT().SimpleTy) {
4665 default: return false;
4674 return Scale == 2 || Scale == 4 || Scale == 8;
4677 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4681 // Note, we allow "void" uses (basically, uses that aren't loads or
4682 // stores), because arm allows folding a scale into many arithmetic
4683 // operations. This should be made more precise and revisited later.
4685 // Allow r << imm, but the imm has to be a multiple of two.
4686 if (Scale & 1) return false;
4687 return isPowerOf2_32(Scale);
4691 /// isLegalAddressingMode - Return true if the addressing mode represented
4692 /// by AM is legal for this target, for a load/store of the specified type.
4693 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4694 const Type *Ty) const {
4695 EVT VT = getValueType(Ty, true);
4696 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4699 // Can never fold addr of global into load/store.
4704 case 0: // no scale reg, must be "r+i" or "r", or "i".
4707 if (Subtarget->isThumb1Only())
4711 // ARM doesn't support any R+R*scale+imm addr modes.
4718 if (Subtarget->isThumb2())
4719 return isLegalT2ScaledAddressingMode(AM, VT);
4721 int Scale = AM.Scale;
4722 switch (VT.getSimpleVT().SimpleTy) {
4723 default: return false;
4727 if (Scale < 0) Scale = -Scale;
4731 return isPowerOf2_32(Scale & ~1);
4735 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4740 // Note, we allow "void" uses (basically, uses that aren't loads or
4741 // stores), because arm allows folding a scale into many arithmetic
4742 // operations. This should be made more precise and revisited later.
4744 // Allow r << imm, but the imm has to be a multiple of two.
4745 if (Scale & 1) return false;
4746 return isPowerOf2_32(Scale);
4753 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4754 /// icmp immediate, that is the target has icmp instructions which can compare
4755 /// a register against the immediate without having to materialize the
4756 /// immediate into a register.
4757 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4758 if (!Subtarget->isThumb())
4759 return ARM_AM::getSOImmVal(Imm) != -1;
4760 if (Subtarget->isThumb2())
4761 return ARM_AM::getT2SOImmVal(Imm) != -1;
4762 return Imm >= 0 && Imm <= 255;
4765 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4766 bool isSEXTLoad, SDValue &Base,
4767 SDValue &Offset, bool &isInc,
4768 SelectionDAG &DAG) {
4769 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4772 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4774 Base = Ptr->getOperand(0);
4775 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4776 int RHSC = (int)RHS->getZExtValue();
4777 if (RHSC < 0 && RHSC > -256) {
4778 assert(Ptr->getOpcode() == ISD::ADD);
4780 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4784 isInc = (Ptr->getOpcode() == ISD::ADD);
4785 Offset = Ptr->getOperand(1);
4787 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4789 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4790 int RHSC = (int)RHS->getZExtValue();
4791 if (RHSC < 0 && RHSC > -0x1000) {
4792 assert(Ptr->getOpcode() == ISD::ADD);
4794 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4795 Base = Ptr->getOperand(0);
4800 if (Ptr->getOpcode() == ISD::ADD) {
4802 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4803 if (ShOpcVal != ARM_AM::no_shift) {
4804 Base = Ptr->getOperand(1);
4805 Offset = Ptr->getOperand(0);
4807 Base = Ptr->getOperand(0);
4808 Offset = Ptr->getOperand(1);
4813 isInc = (Ptr->getOpcode() == ISD::ADD);
4814 Base = Ptr->getOperand(0);
4815 Offset = Ptr->getOperand(1);
4819 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4823 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4824 bool isSEXTLoad, SDValue &Base,
4825 SDValue &Offset, bool &isInc,
4826 SelectionDAG &DAG) {
4827 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4830 Base = Ptr->getOperand(0);
4831 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4832 int RHSC = (int)RHS->getZExtValue();
4833 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4834 assert(Ptr->getOpcode() == ISD::ADD);
4836 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4838 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4839 isInc = Ptr->getOpcode() == ISD::ADD;
4840 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4848 /// getPreIndexedAddressParts - returns true by value, base pointer and
4849 /// offset pointer and addressing mode by reference if the node's address
4850 /// can be legally represented as pre-indexed load / store address.
4852 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4854 ISD::MemIndexedMode &AM,
4855 SelectionDAG &DAG) const {
4856 if (Subtarget->isThumb1Only())
4861 bool isSEXTLoad = false;
4862 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4863 Ptr = LD->getBasePtr();
4864 VT = LD->getMemoryVT();
4865 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4866 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4867 Ptr = ST->getBasePtr();
4868 VT = ST->getMemoryVT();
4873 bool isLegal = false;
4874 if (Subtarget->isThumb2())
4875 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4876 Offset, isInc, DAG);
4878 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4879 Offset, isInc, DAG);
4883 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4887 /// getPostIndexedAddressParts - returns true by value, base pointer and
4888 /// offset pointer and addressing mode by reference if this node can be
4889 /// combined with a load / store to form a post-indexed load / store.
4890 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4893 ISD::MemIndexedMode &AM,
4894 SelectionDAG &DAG) const {
4895 if (Subtarget->isThumb1Only())
4900 bool isSEXTLoad = false;
4901 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4902 VT = LD->getMemoryVT();
4903 Ptr = LD->getBasePtr();
4904 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4905 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4906 VT = ST->getMemoryVT();
4907 Ptr = ST->getBasePtr();
4912 bool isLegal = false;
4913 if (Subtarget->isThumb2())
4914 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4917 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4923 // Swap base ptr and offset to catch more post-index load / store when
4924 // it's legal. In Thumb2 mode, offset must be an immediate.
4925 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4926 !Subtarget->isThumb2())
4927 std::swap(Base, Offset);
4929 // Post-indexed load / store update the base pointer.
4934 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4938 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4942 const SelectionDAG &DAG,
4943 unsigned Depth) const {
4944 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4945 switch (Op.getOpcode()) {
4947 case ARMISD::CMOV: {
4948 // Bits are known zero/one if known on the LHS and RHS.
4949 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4950 if (KnownZero == 0 && KnownOne == 0) return;
4952 APInt KnownZeroRHS, KnownOneRHS;
4953 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4954 KnownZeroRHS, KnownOneRHS, Depth+1);
4955 KnownZero &= KnownZeroRHS;
4956 KnownOne &= KnownOneRHS;
4962 //===----------------------------------------------------------------------===//
4963 // ARM Inline Assembly Support
4964 //===----------------------------------------------------------------------===//
4966 /// getConstraintType - Given a constraint letter, return the type of
4967 /// constraint it is for this target.
4968 ARMTargetLowering::ConstraintType
4969 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4970 if (Constraint.size() == 1) {
4971 switch (Constraint[0]) {
4973 case 'l': return C_RegisterClass;
4974 case 'w': return C_RegisterClass;
4977 return TargetLowering::getConstraintType(Constraint);
4980 std::pair<unsigned, const TargetRegisterClass*>
4981 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4983 if (Constraint.size() == 1) {
4984 // GCC ARM Constraint Letters
4985 switch (Constraint[0]) {
4987 if (Subtarget->isThumb())
4988 return std::make_pair(0U, ARM::tGPRRegisterClass);
4990 return std::make_pair(0U, ARM::GPRRegisterClass);
4992 return std::make_pair(0U, ARM::GPRRegisterClass);
4995 return std::make_pair(0U, ARM::SPRRegisterClass);
4996 if (VT.getSizeInBits() == 64)
4997 return std::make_pair(0U, ARM::DPRRegisterClass);
4998 if (VT.getSizeInBits() == 128)
4999 return std::make_pair(0U, ARM::QPRRegisterClass);
5003 if (StringRef("{cc}").equals_lower(Constraint))
5004 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5006 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5009 std::vector<unsigned> ARMTargetLowering::
5010 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5012 if (Constraint.size() != 1)
5013 return std::vector<unsigned>();
5015 switch (Constraint[0]) { // GCC ARM Constraint Letters
5018 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5019 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5022 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5023 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5024 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5025 ARM::R12, ARM::LR, 0);
5028 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5029 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5030 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5031 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5032 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5033 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5034 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5035 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5036 if (VT.getSizeInBits() == 64)
5037 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5038 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5039 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5040 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5041 if (VT.getSizeInBits() == 128)
5042 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5043 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5047 return std::vector<unsigned>();
5050 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5051 /// vector. If it is invalid, don't add anything to Ops.
5052 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5055 std::vector<SDValue>&Ops,
5056 SelectionDAG &DAG) const {
5057 SDValue Result(0, 0);
5059 switch (Constraint) {
5061 case 'I': case 'J': case 'K': case 'L':
5062 case 'M': case 'N': case 'O':
5063 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5067 int64_t CVal64 = C->getSExtValue();
5068 int CVal = (int) CVal64;
5069 // None of these constraints allow values larger than 32 bits. Check
5070 // that the value fits in an int.
5074 switch (Constraint) {
5076 if (Subtarget->isThumb1Only()) {
5077 // This must be a constant between 0 and 255, for ADD
5079 if (CVal >= 0 && CVal <= 255)
5081 } else if (Subtarget->isThumb2()) {
5082 // A constant that can be used as an immediate value in a
5083 // data-processing instruction.
5084 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5087 // A constant that can be used as an immediate value in a
5088 // data-processing instruction.
5089 if (ARM_AM::getSOImmVal(CVal) != -1)
5095 if (Subtarget->isThumb()) { // FIXME thumb2
5096 // This must be a constant between -255 and -1, for negated ADD
5097 // immediates. This can be used in GCC with an "n" modifier that
5098 // prints the negated value, for use with SUB instructions. It is
5099 // not useful otherwise but is implemented for compatibility.
5100 if (CVal >= -255 && CVal <= -1)
5103 // This must be a constant between -4095 and 4095. It is not clear
5104 // what this constraint is intended for. Implemented for
5105 // compatibility with GCC.
5106 if (CVal >= -4095 && CVal <= 4095)
5112 if (Subtarget->isThumb1Only()) {
5113 // A 32-bit value where only one byte has a nonzero value. Exclude
5114 // zero to match GCC. This constraint is used by GCC internally for
5115 // constants that can be loaded with a move/shift combination.
5116 // It is not useful otherwise but is implemented for compatibility.
5117 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5119 } else if (Subtarget->isThumb2()) {
5120 // A constant whose bitwise inverse can be used as an immediate
5121 // value in a data-processing instruction. This can be used in GCC
5122 // with a "B" modifier that prints the inverted value, for use with
5123 // BIC and MVN instructions. It is not useful otherwise but is
5124 // implemented for compatibility.
5125 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5128 // A constant whose bitwise inverse can be used as an immediate
5129 // value in a data-processing instruction. This can be used in GCC
5130 // with a "B" modifier that prints the inverted value, for use with
5131 // BIC and MVN instructions. It is not useful otherwise but is
5132 // implemented for compatibility.
5133 if (ARM_AM::getSOImmVal(~CVal) != -1)
5139 if (Subtarget->isThumb1Only()) {
5140 // This must be a constant between -7 and 7,
5141 // for 3-operand ADD/SUB immediate instructions.
5142 if (CVal >= -7 && CVal < 7)
5144 } else if (Subtarget->isThumb2()) {
5145 // A constant whose negation can be used as an immediate value in a
5146 // data-processing instruction. This can be used in GCC with an "n"
5147 // modifier that prints the negated value, for use with SUB
5148 // instructions. It is not useful otherwise but is implemented for
5150 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5153 // A constant whose negation can be used as an immediate value in a
5154 // data-processing instruction. This can be used in GCC with an "n"
5155 // modifier that prints the negated value, for use with SUB
5156 // instructions. It is not useful otherwise but is implemented for
5158 if (ARM_AM::getSOImmVal(-CVal) != -1)
5164 if (Subtarget->isThumb()) { // FIXME thumb2
5165 // This must be a multiple of 4 between 0 and 1020, for
5166 // ADD sp + immediate.
5167 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5170 // A power of two or a constant between 0 and 32. This is used in
5171 // GCC for the shift amount on shifted register operands, but it is
5172 // useful in general for any shift amounts.
5173 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5179 if (Subtarget->isThumb()) { // FIXME thumb2
5180 // This must be a constant between 0 and 31, for shift amounts.
5181 if (CVal >= 0 && CVal <= 31)
5187 if (Subtarget->isThumb()) { // FIXME thumb2
5188 // This must be a multiple of 4 between -508 and 508, for
5189 // ADD/SUB sp = sp + immediate.
5190 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5195 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5199 if (Result.getNode()) {
5200 Ops.push_back(Result);
5203 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
5208 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5209 // The ARM target isn't yet aware of offsets.
5213 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5214 APInt Imm = FPImm.bitcastToAPInt();
5215 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5216 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5217 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5219 // We can handle 4 bits of mantissa.
5220 // mantissa = (16+UInt(e:f:g:h))/16.
5221 if (Mantissa & 0x7ffff)
5224 if ((Mantissa & 0xf) != Mantissa)
5227 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5228 if (Exp < -3 || Exp > 4)
5230 Exp = ((Exp+3) & 0x7) ^ 4;
5232 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5235 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5236 APInt Imm = FPImm.bitcastToAPInt();
5237 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5238 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5239 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5241 // We can handle 4 bits of mantissa.
5242 // mantissa = (16+UInt(e:f:g:h))/16.
5243 if (Mantissa & 0xffffffffffffLL)
5246 if ((Mantissa & 0xf) != Mantissa)
5249 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5250 if (Exp < -3 || Exp > 4)
5252 Exp = ((Exp+3) & 0x7) ^ 4;
5254 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5257 /// isFPImmLegal - Returns true if the target can instruction select the
5258 /// specified FP immediate natively. If false, the legalizer will
5259 /// materialize the FP immediate as a load from a constant pool.
5260 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5261 if (!Subtarget->hasVFP3())
5264 return ARM::getVFPf32Imm(Imm) != -1;
5266 return ARM::getVFPf64Imm(Imm) != -1;