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);
444 // Since the libcalls include locking, fold in the fences
445 setShouldFoldAtomicFences(true);
447 // 64-bit versions are always libcalls (for now)
448 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
449 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
450 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
451 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
452 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
453 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
454 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
455 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
457 // If the subtarget does not have extract instructions, sign_extend_inreg
458 // needs to be expanded. Extract is available in ARM mode on v6 and up,
459 // and on most Thumb2 implementations.
460 if (!Subtarget->hasV6Ops()
461 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
462 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
463 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
465 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
467 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
468 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
469 // iff target supports vfp2.
470 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
472 // We want to custom lower some of our intrinsics.
473 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
475 setOperationAction(ISD::SETCC, MVT::i32, Expand);
476 setOperationAction(ISD::SETCC, MVT::f32, Expand);
477 setOperationAction(ISD::SETCC, MVT::f64, Expand);
478 setOperationAction(ISD::SELECT, MVT::i32, Expand);
479 setOperationAction(ISD::SELECT, MVT::f32, Expand);
480 setOperationAction(ISD::SELECT, MVT::f64, Expand);
481 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
482 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
483 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
485 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
486 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
487 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
488 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
489 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
491 // We don't support sin/cos/fmod/copysign/pow
492 setOperationAction(ISD::FSIN, MVT::f64, Expand);
493 setOperationAction(ISD::FSIN, MVT::f32, Expand);
494 setOperationAction(ISD::FCOS, MVT::f32, Expand);
495 setOperationAction(ISD::FCOS, MVT::f64, Expand);
496 setOperationAction(ISD::FREM, MVT::f64, Expand);
497 setOperationAction(ISD::FREM, MVT::f32, Expand);
498 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
499 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
500 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
502 setOperationAction(ISD::FPOW, MVT::f64, Expand);
503 setOperationAction(ISD::FPOW, MVT::f32, Expand);
505 // Various VFP goodness
506 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
507 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
508 if (Subtarget->hasVFP2()) {
509 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
510 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
511 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
512 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
514 // Special handling for half-precision FP.
515 if (!Subtarget->hasFP16()) {
516 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
517 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
521 // We have target-specific dag combine patterns for the following nodes:
522 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
523 setTargetDAGCombine(ISD::ADD);
524 setTargetDAGCombine(ISD::SUB);
525 setTargetDAGCombine(ISD::MUL);
527 setStackPointerRegisterToSaveRestore(ARM::SP);
529 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
530 setSchedulingPreference(Sched::RegPressure);
532 setSchedulingPreference(Sched::Hybrid);
534 // FIXME: If-converter should use instruction latency to determine
535 // profitability rather than relying on fixed limits.
536 if (Subtarget->getCPUString() == "generic") {
537 // Generic (and overly aggressive) if-conversion limits.
538 setIfCvtBlockSizeLimit(10);
539 setIfCvtDupBlockSizeLimit(2);
540 } else if (Subtarget->hasV7Ops()) {
541 setIfCvtBlockSizeLimit(3);
542 setIfCvtDupBlockSizeLimit(1);
543 } else if (Subtarget->hasV6Ops()) {
544 setIfCvtBlockSizeLimit(2);
545 setIfCvtDupBlockSizeLimit(1);
547 setIfCvtBlockSizeLimit(3);
548 setIfCvtDupBlockSizeLimit(2);
551 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
552 // Do not enable CodePlacementOpt for now: it currently runs after the
553 // ARMConstantIslandPass and messes up branch relaxation and placement
554 // of constant islands.
555 // benefitFromCodePlacementOpt = true;
558 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
561 case ARMISD::Wrapper: return "ARMISD::Wrapper";
562 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
563 case ARMISD::CALL: return "ARMISD::CALL";
564 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
565 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
566 case ARMISD::tCALL: return "ARMISD::tCALL";
567 case ARMISD::BRCOND: return "ARMISD::BRCOND";
568 case ARMISD::BR_JT: return "ARMISD::BR_JT";
569 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
570 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
571 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
572 case ARMISD::CMP: return "ARMISD::CMP";
573 case ARMISD::CMPZ: return "ARMISD::CMPZ";
574 case ARMISD::CMPFP: return "ARMISD::CMPFP";
575 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
576 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
577 case ARMISD::CMOV: return "ARMISD::CMOV";
578 case ARMISD::CNEG: return "ARMISD::CNEG";
580 case ARMISD::RBIT: return "ARMISD::RBIT";
582 case ARMISD::FTOSI: return "ARMISD::FTOSI";
583 case ARMISD::FTOUI: return "ARMISD::FTOUI";
584 case ARMISD::SITOF: return "ARMISD::SITOF";
585 case ARMISD::UITOF: return "ARMISD::UITOF";
587 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
588 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
589 case ARMISD::RRX: return "ARMISD::RRX";
591 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
592 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
594 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
595 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
597 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
599 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
601 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
603 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
604 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
606 case ARMISD::VCEQ: return "ARMISD::VCEQ";
607 case ARMISD::VCGE: return "ARMISD::VCGE";
608 case ARMISD::VCGEU: return "ARMISD::VCGEU";
609 case ARMISD::VCGT: return "ARMISD::VCGT";
610 case ARMISD::VCGTU: return "ARMISD::VCGTU";
611 case ARMISD::VTST: return "ARMISD::VTST";
613 case ARMISD::VSHL: return "ARMISD::VSHL";
614 case ARMISD::VSHRs: return "ARMISD::VSHRs";
615 case ARMISD::VSHRu: return "ARMISD::VSHRu";
616 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
617 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
618 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
619 case ARMISD::VSHRN: return "ARMISD::VSHRN";
620 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
621 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
622 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
623 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
624 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
625 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
626 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
627 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
628 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
629 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
630 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
631 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
632 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
633 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
634 case ARMISD::VDUP: return "ARMISD::VDUP";
635 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
636 case ARMISD::VEXT: return "ARMISD::VEXT";
637 case ARMISD::VREV64: return "ARMISD::VREV64";
638 case ARMISD::VREV32: return "ARMISD::VREV32";
639 case ARMISD::VREV16: return "ARMISD::VREV16";
640 case ARMISD::VZIP: return "ARMISD::VZIP";
641 case ARMISD::VUZP: return "ARMISD::VUZP";
642 case ARMISD::VTRN: return "ARMISD::VTRN";
643 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
644 case ARMISD::FMAX: return "ARMISD::FMAX";
645 case ARMISD::FMIN: return "ARMISD::FMIN";
649 /// getRegClassFor - Return the register class that should be used for the
650 /// specified value type.
651 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
652 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
653 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
654 // load / store 4 to 8 consecutive D registers.
655 if (Subtarget->hasNEON()) {
656 if (VT == MVT::v4i64)
657 return ARM::QQPRRegisterClass;
658 else if (VT == MVT::v8i64)
659 return ARM::QQQQPRRegisterClass;
661 return TargetLowering::getRegClassFor(VT);
664 /// getFunctionAlignment - Return the Log2 alignment of this function.
665 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
666 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
669 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
670 unsigned NumVals = N->getNumValues();
672 return Sched::RegPressure;
674 for (unsigned i = 0; i != NumVals; ++i) {
675 EVT VT = N->getValueType(i);
676 if (VT.isFloatingPoint() || VT.isVector())
677 return Sched::Latency;
680 if (!N->isMachineOpcode())
681 return Sched::RegPressure;
683 // Load are scheduled for latency even if there instruction itinerary
685 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
686 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
688 return Sched::Latency;
690 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
691 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
692 return Sched::Latency;
693 return Sched::RegPressure;
696 //===----------------------------------------------------------------------===//
698 //===----------------------------------------------------------------------===//
700 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
701 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
703 default: llvm_unreachable("Unknown condition code!");
704 case ISD::SETNE: return ARMCC::NE;
705 case ISD::SETEQ: return ARMCC::EQ;
706 case ISD::SETGT: return ARMCC::GT;
707 case ISD::SETGE: return ARMCC::GE;
708 case ISD::SETLT: return ARMCC::LT;
709 case ISD::SETLE: return ARMCC::LE;
710 case ISD::SETUGT: return ARMCC::HI;
711 case ISD::SETUGE: return ARMCC::HS;
712 case ISD::SETULT: return ARMCC::LO;
713 case ISD::SETULE: return ARMCC::LS;
717 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
718 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
719 ARMCC::CondCodes &CondCode2) {
720 CondCode2 = ARMCC::AL;
722 default: llvm_unreachable("Unknown FP condition!");
724 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
726 case ISD::SETOGT: CondCode = ARMCC::GT; break;
728 case ISD::SETOGE: CondCode = ARMCC::GE; break;
729 case ISD::SETOLT: CondCode = ARMCC::MI; break;
730 case ISD::SETOLE: CondCode = ARMCC::LS; break;
731 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
732 case ISD::SETO: CondCode = ARMCC::VC; break;
733 case ISD::SETUO: CondCode = ARMCC::VS; break;
734 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
735 case ISD::SETUGT: CondCode = ARMCC::HI; break;
736 case ISD::SETUGE: CondCode = ARMCC::PL; break;
738 case ISD::SETULT: CondCode = ARMCC::LT; break;
740 case ISD::SETULE: CondCode = ARMCC::LE; break;
742 case ISD::SETUNE: CondCode = ARMCC::NE; break;
746 //===----------------------------------------------------------------------===//
747 // Calling Convention Implementation
748 //===----------------------------------------------------------------------===//
750 #include "ARMGenCallingConv.inc"
752 // APCS f64 is in register pairs, possibly split to stack
753 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
754 CCValAssign::LocInfo &LocInfo,
755 CCState &State, bool CanFail) {
756 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
758 // Try to get the first register.
759 if (unsigned Reg = State.AllocateReg(RegList, 4))
760 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
762 // For the 2nd half of a v2f64, do not fail.
766 // Put the whole thing on the stack.
767 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
768 State.AllocateStack(8, 4),
773 // Try to get the second register.
774 if (unsigned Reg = State.AllocateReg(RegList, 4))
775 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
777 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
778 State.AllocateStack(4, 4),
783 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
784 CCValAssign::LocInfo &LocInfo,
785 ISD::ArgFlagsTy &ArgFlags,
787 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
789 if (LocVT == MVT::v2f64 &&
790 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
792 return true; // we handled it
795 // AAPCS f64 is in aligned register pairs
796 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
797 CCValAssign::LocInfo &LocInfo,
798 CCState &State, bool CanFail) {
799 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
800 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
802 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
804 // For the 2nd half of a v2f64, do not just fail.
808 // Put the whole thing on the stack.
809 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
810 State.AllocateStack(8, 8),
816 for (i = 0; i < 2; ++i)
817 if (HiRegList[i] == Reg)
820 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
821 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
826 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
827 CCValAssign::LocInfo &LocInfo,
828 ISD::ArgFlagsTy &ArgFlags,
830 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
832 if (LocVT == MVT::v2f64 &&
833 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
835 return true; // we handled it
838 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
839 CCValAssign::LocInfo &LocInfo, CCState &State) {
840 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
841 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
843 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
845 return false; // we didn't handle it
848 for (i = 0; i < 2; ++i)
849 if (HiRegList[i] == Reg)
852 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
853 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
858 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
859 CCValAssign::LocInfo &LocInfo,
860 ISD::ArgFlagsTy &ArgFlags,
862 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
864 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
866 return true; // we handled it
869 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
870 CCValAssign::LocInfo &LocInfo,
871 ISD::ArgFlagsTy &ArgFlags,
873 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
877 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
878 /// given CallingConvention value.
879 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
881 bool isVarArg) const {
884 llvm_unreachable("Unsupported calling convention");
886 case CallingConv::Fast:
887 // Use target triple & subtarget features to do actual dispatch.
888 if (Subtarget->isAAPCS_ABI()) {
889 if (Subtarget->hasVFP2() &&
890 FloatABIType == FloatABI::Hard && !isVarArg)
891 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
893 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
895 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
896 case CallingConv::ARM_AAPCS_VFP:
897 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
898 case CallingConv::ARM_AAPCS:
899 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
900 case CallingConv::ARM_APCS:
901 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
905 /// LowerCallResult - Lower the result values of a call into the
906 /// appropriate copies out of appropriate physical registers.
908 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
909 CallingConv::ID CallConv, bool isVarArg,
910 const SmallVectorImpl<ISD::InputArg> &Ins,
911 DebugLoc dl, SelectionDAG &DAG,
912 SmallVectorImpl<SDValue> &InVals) const {
914 // Assign locations to each value returned by this call.
915 SmallVector<CCValAssign, 16> RVLocs;
916 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
917 RVLocs, *DAG.getContext());
918 CCInfo.AnalyzeCallResult(Ins,
919 CCAssignFnForNode(CallConv, /* Return*/ true,
922 // Copy all of the result registers out of their specified physreg.
923 for (unsigned i = 0; i != RVLocs.size(); ++i) {
924 CCValAssign VA = RVLocs[i];
927 if (VA.needsCustom()) {
928 // Handle f64 or half of a v2f64.
929 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
931 Chain = Lo.getValue(1);
932 InFlag = Lo.getValue(2);
933 VA = RVLocs[++i]; // skip ahead to next loc
934 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
936 Chain = Hi.getValue(1);
937 InFlag = Hi.getValue(2);
938 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
940 if (VA.getLocVT() == MVT::v2f64) {
941 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
942 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
943 DAG.getConstant(0, MVT::i32));
945 VA = RVLocs[++i]; // skip ahead to next loc
946 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
947 Chain = Lo.getValue(1);
948 InFlag = Lo.getValue(2);
949 VA = RVLocs[++i]; // skip ahead to next loc
950 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
951 Chain = Hi.getValue(1);
952 InFlag = Hi.getValue(2);
953 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
954 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
955 DAG.getConstant(1, MVT::i32));
958 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
960 Chain = Val.getValue(1);
961 InFlag = Val.getValue(2);
964 switch (VA.getLocInfo()) {
965 default: llvm_unreachable("Unknown loc info!");
966 case CCValAssign::Full: break;
967 case CCValAssign::BCvt:
968 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
972 InVals.push_back(Val);
978 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
979 /// by "Src" to address "Dst" of size "Size". Alignment information is
980 /// specified by the specific parameter attribute. The copy will be passed as
981 /// a byval function parameter.
982 /// Sometimes what we are copying is the end of a larger object, the part that
983 /// does not fit in registers.
985 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
986 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
988 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
989 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
990 /*isVolatile=*/false, /*AlwaysInline=*/false,
994 /// LowerMemOpCallTo - Store the argument to the stack.
996 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
997 SDValue StackPtr, SDValue Arg,
998 DebugLoc dl, SelectionDAG &DAG,
999 const CCValAssign &VA,
1000 ISD::ArgFlagsTy Flags) const {
1001 unsigned LocMemOffset = VA.getLocMemOffset();
1002 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1003 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1004 if (Flags.isByVal()) {
1005 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1007 return DAG.getStore(Chain, dl, Arg, PtrOff,
1008 PseudoSourceValue::getStack(), LocMemOffset,
1012 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1013 SDValue Chain, SDValue &Arg,
1014 RegsToPassVector &RegsToPass,
1015 CCValAssign &VA, CCValAssign &NextVA,
1017 SmallVector<SDValue, 8> &MemOpChains,
1018 ISD::ArgFlagsTy Flags) const {
1020 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1021 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1022 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1024 if (NextVA.isRegLoc())
1025 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1027 assert(NextVA.isMemLoc());
1028 if (StackPtr.getNode() == 0)
1029 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1031 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1037 /// LowerCall - Lowering a call into a callseq_start <-
1038 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1041 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1042 CallingConv::ID CallConv, bool isVarArg,
1044 const SmallVectorImpl<ISD::OutputArg> &Outs,
1045 const SmallVectorImpl<ISD::InputArg> &Ins,
1046 DebugLoc dl, SelectionDAG &DAG,
1047 SmallVectorImpl<SDValue> &InVals) const {
1048 MachineFunction &MF = DAG.getMachineFunction();
1049 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1050 bool IsSibCall = false;
1051 // Temporarily disable tail calls so things don't break.
1052 if (!EnableARMTailCalls)
1055 // Check if it's really possible to do a tail call.
1056 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1057 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1059 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1060 // detected sibcalls.
1067 // Analyze operands of the call, assigning locations to each operand.
1068 SmallVector<CCValAssign, 16> ArgLocs;
1069 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1071 CCInfo.AnalyzeCallOperands(Outs,
1072 CCAssignFnForNode(CallConv, /* Return*/ false,
1075 // Get a count of how many bytes are to be pushed on the stack.
1076 unsigned NumBytes = CCInfo.getNextStackOffset();
1078 // For tail calls, memory operands are available in our caller's stack.
1082 // Adjust the stack pointer for the new arguments...
1083 // These operations are automatically eliminated by the prolog/epilog pass
1085 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1087 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1089 RegsToPassVector RegsToPass;
1090 SmallVector<SDValue, 8> MemOpChains;
1092 // Walk the register/memloc assignments, inserting copies/loads. In the case
1093 // of tail call optimization, arguments are handled later.
1094 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1096 ++i, ++realArgIdx) {
1097 CCValAssign &VA = ArgLocs[i];
1098 SDValue Arg = Outs[realArgIdx].Val;
1099 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1101 // Promote the value if needed.
1102 switch (VA.getLocInfo()) {
1103 default: llvm_unreachable("Unknown loc info!");
1104 case CCValAssign::Full: break;
1105 case CCValAssign::SExt:
1106 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1108 case CCValAssign::ZExt:
1109 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1111 case CCValAssign::AExt:
1112 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1114 case CCValAssign::BCvt:
1115 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1119 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1120 if (VA.needsCustom()) {
1121 if (VA.getLocVT() == MVT::v2f64) {
1122 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1123 DAG.getConstant(0, MVT::i32));
1124 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1125 DAG.getConstant(1, MVT::i32));
1127 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1128 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1130 VA = ArgLocs[++i]; // skip ahead to next loc
1131 if (VA.isRegLoc()) {
1132 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1133 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1135 assert(VA.isMemLoc());
1137 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1138 dl, DAG, VA, Flags));
1141 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1142 StackPtr, MemOpChains, Flags);
1144 } else if (VA.isRegLoc()) {
1145 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1146 } else if (!IsSibCall) {
1147 assert(VA.isMemLoc());
1149 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1150 dl, DAG, VA, Flags));
1154 if (!MemOpChains.empty())
1155 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1156 &MemOpChains[0], MemOpChains.size());
1158 // Build a sequence of copy-to-reg nodes chained together with token chain
1159 // and flag operands which copy the outgoing args into the appropriate regs.
1161 // Tail call byval lowering might overwrite argument registers so in case of
1162 // tail call optimization the copies to registers are lowered later.
1164 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1165 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1166 RegsToPass[i].second, InFlag);
1167 InFlag = Chain.getValue(1);
1170 // For tail calls lower the arguments to the 'real' stack slot.
1172 // Force all the incoming stack arguments to be loaded from the stack
1173 // before any new outgoing arguments are stored to the stack, because the
1174 // outgoing stack slots may alias the incoming argument stack slots, and
1175 // the alias isn't otherwise explicit. This is slightly more conservative
1176 // than necessary, because it means that each store effectively depends
1177 // on every argument instead of just those arguments it would clobber.
1179 // Do not flag preceeding copytoreg stuff together with the following stuff.
1181 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1182 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1183 RegsToPass[i].second, InFlag);
1184 InFlag = Chain.getValue(1);
1189 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1190 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1191 // node so that legalize doesn't hack it.
1192 bool isDirect = false;
1193 bool isARMFunc = false;
1194 bool isLocalARMFunc = false;
1195 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1197 if (EnableARMLongCalls) {
1198 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1199 && "long-calls with non-static relocation model!");
1200 // Handle a global address or an external symbol. If it's not one of
1201 // those, the target's already in a register, so we don't need to do
1203 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1204 const GlobalValue *GV = G->getGlobal();
1205 // Create a constant pool entry for the callee address
1206 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1207 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1210 // Get the address of the callee into a register
1211 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1212 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1213 Callee = DAG.getLoad(getPointerTy(), dl,
1214 DAG.getEntryNode(), CPAddr,
1215 PseudoSourceValue::getConstantPool(), 0,
1217 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1218 const char *Sym = S->getSymbol();
1220 // Create a constant pool entry for the callee address
1221 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1222 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1223 Sym, ARMPCLabelIndex, 0);
1224 // Get the address of the callee into a register
1225 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1226 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1227 Callee = DAG.getLoad(getPointerTy(), dl,
1228 DAG.getEntryNode(), CPAddr,
1229 PseudoSourceValue::getConstantPool(), 0,
1232 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1233 const GlobalValue *GV = G->getGlobal();
1235 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1236 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1237 getTargetMachine().getRelocationModel() != Reloc::Static;
1238 isARMFunc = !Subtarget->isThumb() || isStub;
1239 // ARM call to a local ARM function is predicable.
1240 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1241 // tBX takes a register source operand.
1242 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1243 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1244 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1247 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1248 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1249 Callee = DAG.getLoad(getPointerTy(), dl,
1250 DAG.getEntryNode(), CPAddr,
1251 PseudoSourceValue::getConstantPool(), 0,
1253 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1254 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1255 getPointerTy(), Callee, PICLabel);
1257 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1258 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1260 bool isStub = Subtarget->isTargetDarwin() &&
1261 getTargetMachine().getRelocationModel() != Reloc::Static;
1262 isARMFunc = !Subtarget->isThumb() || isStub;
1263 // tBX takes a register source operand.
1264 const char *Sym = S->getSymbol();
1265 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1266 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1267 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1268 Sym, ARMPCLabelIndex, 4);
1269 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1270 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1271 Callee = DAG.getLoad(getPointerTy(), dl,
1272 DAG.getEntryNode(), CPAddr,
1273 PseudoSourceValue::getConstantPool(), 0,
1275 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1276 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1277 getPointerTy(), Callee, PICLabel);
1279 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1282 // FIXME: handle tail calls differently.
1284 if (Subtarget->isThumb()) {
1285 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1286 CallOpc = ARMISD::CALL_NOLINK;
1288 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1290 CallOpc = (isDirect || Subtarget->hasV5TOps())
1291 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1292 : ARMISD::CALL_NOLINK;
1294 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1295 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1296 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1297 InFlag = Chain.getValue(1);
1300 std::vector<SDValue> Ops;
1301 Ops.push_back(Chain);
1302 Ops.push_back(Callee);
1304 // Add argument registers to the end of the list so that they are known live
1306 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1307 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1308 RegsToPass[i].second.getValueType()));
1310 if (InFlag.getNode())
1311 Ops.push_back(InFlag);
1313 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1315 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1317 // Returns a chain and a flag for retval copy to use.
1318 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1319 InFlag = Chain.getValue(1);
1321 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1322 DAG.getIntPtrConstant(0, true), InFlag);
1324 InFlag = Chain.getValue(1);
1326 // Handle result values, copying them out of physregs into vregs that we
1328 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1332 /// MatchingStackOffset - Return true if the given stack call argument is
1333 /// already available in the same position (relatively) of the caller's
1334 /// incoming argument stack.
1336 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1337 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1338 const ARMInstrInfo *TII) {
1339 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1341 if (Arg.getOpcode() == ISD::CopyFromReg) {
1342 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1343 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1345 MachineInstr *Def = MRI->getVRegDef(VR);
1348 if (!Flags.isByVal()) {
1349 if (!TII->isLoadFromStackSlot(Def, FI))
1352 // unsigned Opcode = Def->getOpcode();
1353 // if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
1354 // Def->getOperand(1).isFI()) {
1355 // FI = Def->getOperand(1).getIndex();
1356 // Bytes = Flags.getByValSize();
1360 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1361 if (Flags.isByVal())
1362 // ByVal argument is passed in as a pointer but it's now being
1363 // dereferenced. e.g.
1364 // define @foo(%struct.X* %A) {
1365 // tail call @bar(%struct.X* byval %A)
1368 SDValue Ptr = Ld->getBasePtr();
1369 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1372 FI = FINode->getIndex();
1376 assert(FI != INT_MAX);
1377 if (!MFI->isFixedObjectIndex(FI))
1379 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1382 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1383 /// for tail call optimization. Targets which want to do tail call
1384 /// optimization should implement this function.
1386 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1387 CallingConv::ID CalleeCC,
1389 bool isCalleeStructRet,
1390 bool isCallerStructRet,
1391 const SmallVectorImpl<ISD::OutputArg> &Outs,
1392 const SmallVectorImpl<ISD::InputArg> &Ins,
1393 SelectionDAG& DAG) const {
1394 const Function *CallerF = DAG.getMachineFunction().getFunction();
1395 CallingConv::ID CallerCC = CallerF->getCallingConv();
1396 bool CCMatch = CallerCC == CalleeCC;
1398 // Look for obvious safe cases to perform tail call optimization that do not
1399 // require ABI changes. This is what gcc calls sibcall.
1401 // Do not sibcall optimize vararg calls unless the call site is not passing
1403 if (isVarArg && !Outs.empty())
1406 // Also avoid sibcall optimization if either caller or callee uses struct
1407 // return semantics.
1408 if (isCalleeStructRet || isCallerStructRet)
1411 // FIXME: Completely disable sibcal for Thumb1 since Thumb1RegisterInfo::
1412 // emitEpilogue is not ready for them.
1413 if (Subtarget->isThumb1Only())
1416 if (isa<ExternalSymbolSDNode>(Callee))
1419 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1420 if (Subtarget->isThumb1Only())
1423 // On Thumb, for the moment, we can only do this to functions defined in this
1424 // compilation, or to indirect calls. A Thumb B to an ARM function is not
1425 // easily fixed up in the linker, unlike BL.
1426 if (Subtarget->isThumb()) {
1427 const GlobalValue *GV = G->getGlobal();
1428 if (GV->isDeclaration() || GV->isWeakForLinker())
1434 // If the calling conventions do not match, then we'd better make sure the
1435 // results are returned in the same way as what the caller expects.
1437 SmallVector<CCValAssign, 16> RVLocs1;
1438 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1439 RVLocs1, *DAG.getContext());
1440 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1442 SmallVector<CCValAssign, 16> RVLocs2;
1443 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1444 RVLocs2, *DAG.getContext());
1445 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1447 if (RVLocs1.size() != RVLocs2.size())
1449 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1450 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1452 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1454 if (RVLocs1[i].isRegLoc()) {
1455 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1458 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1464 // If the callee takes no arguments then go on to check the results of the
1466 if (!Outs.empty()) {
1467 // Check if stack adjustment is needed. For now, do not do this if any
1468 // argument is passed on the stack.
1469 SmallVector<CCValAssign, 16> ArgLocs;
1470 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1471 ArgLocs, *DAG.getContext());
1472 CCInfo.AnalyzeCallOperands(Outs,
1473 CCAssignFnForNode(CalleeCC, false, isVarArg));
1474 if (CCInfo.getNextStackOffset()) {
1475 MachineFunction &MF = DAG.getMachineFunction();
1477 // Check if the arguments are already laid out in the right way as
1478 // the caller's fixed stack objects.
1479 MachineFrameInfo *MFI = MF.getFrameInfo();
1480 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1481 const ARMInstrInfo *TII =
1482 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1483 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1485 ++i, ++realArgIdx) {
1486 CCValAssign &VA = ArgLocs[i];
1487 EVT RegVT = VA.getLocVT();
1488 SDValue Arg = Outs[realArgIdx].Val;
1489 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1490 if (VA.getLocInfo() == CCValAssign::Indirect)
1492 if (VA.needsCustom()) {
1493 // f64 and vector types are split into multiple registers or
1494 // register/stack-slot combinations. The types will not match
1495 // the registers; give up on memory f64 refs until we figure
1496 // out what to do about this.
1499 if (!ArgLocs[++i].isRegLoc())
1501 if (RegVT == MVT::v2f64) {
1502 if (!ArgLocs[++i].isRegLoc())
1504 if (!ArgLocs[++i].isRegLoc())
1507 } else if (!VA.isRegLoc()) {
1508 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1520 ARMTargetLowering::LowerReturn(SDValue Chain,
1521 CallingConv::ID CallConv, bool isVarArg,
1522 const SmallVectorImpl<ISD::OutputArg> &Outs,
1523 DebugLoc dl, SelectionDAG &DAG) const {
1525 // CCValAssign - represent the assignment of the return value to a location.
1526 SmallVector<CCValAssign, 16> RVLocs;
1528 // CCState - Info about the registers and stack slots.
1529 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1532 // Analyze outgoing return values.
1533 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1536 // If this is the first return lowered for this function, add
1537 // the regs to the liveout set for the function.
1538 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1539 for (unsigned i = 0; i != RVLocs.size(); ++i)
1540 if (RVLocs[i].isRegLoc())
1541 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1546 // Copy the result values into the output registers.
1547 for (unsigned i = 0, realRVLocIdx = 0;
1549 ++i, ++realRVLocIdx) {
1550 CCValAssign &VA = RVLocs[i];
1551 assert(VA.isRegLoc() && "Can only return in registers!");
1553 SDValue Arg = Outs[realRVLocIdx].Val;
1555 switch (VA.getLocInfo()) {
1556 default: llvm_unreachable("Unknown loc info!");
1557 case CCValAssign::Full: break;
1558 case CCValAssign::BCvt:
1559 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1563 if (VA.needsCustom()) {
1564 if (VA.getLocVT() == MVT::v2f64) {
1565 // Extract the first half and return it in two registers.
1566 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1567 DAG.getConstant(0, MVT::i32));
1568 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1569 DAG.getVTList(MVT::i32, MVT::i32), Half);
1571 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1572 Flag = Chain.getValue(1);
1573 VA = RVLocs[++i]; // skip ahead to next loc
1574 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1575 HalfGPRs.getValue(1), Flag);
1576 Flag = Chain.getValue(1);
1577 VA = RVLocs[++i]; // skip ahead to next loc
1579 // Extract the 2nd half and fall through to handle it as an f64 value.
1580 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1581 DAG.getConstant(1, MVT::i32));
1583 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1585 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1586 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1587 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1588 Flag = Chain.getValue(1);
1589 VA = RVLocs[++i]; // skip ahead to next loc
1590 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1593 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1595 // Guarantee that all emitted copies are
1596 // stuck together, avoiding something bad.
1597 Flag = Chain.getValue(1);
1602 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1604 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1609 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1610 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1611 // one of the above mentioned nodes. It has to be wrapped because otherwise
1612 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1613 // be used to form addressing mode. These wrapped nodes will be selected
1615 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1616 EVT PtrVT = Op.getValueType();
1617 // FIXME there is no actual debug info here
1618 DebugLoc dl = Op.getDebugLoc();
1619 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1621 if (CP->isMachineConstantPoolEntry())
1622 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1623 CP->getAlignment());
1625 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1626 CP->getAlignment());
1627 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1630 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1631 SelectionDAG &DAG) const {
1632 MachineFunction &MF = DAG.getMachineFunction();
1633 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1634 unsigned ARMPCLabelIndex = 0;
1635 DebugLoc DL = Op.getDebugLoc();
1636 EVT PtrVT = getPointerTy();
1637 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1638 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1640 if (RelocM == Reloc::Static) {
1641 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1643 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1644 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1645 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1646 ARMCP::CPBlockAddress,
1648 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1650 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1651 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1652 PseudoSourceValue::getConstantPool(), 0,
1654 if (RelocM == Reloc::Static)
1656 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1657 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1660 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1662 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1663 SelectionDAG &DAG) const {
1664 DebugLoc dl = GA->getDebugLoc();
1665 EVT PtrVT = getPointerTy();
1666 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1667 MachineFunction &MF = DAG.getMachineFunction();
1668 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1669 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1670 ARMConstantPoolValue *CPV =
1671 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1672 ARMCP::CPValue, PCAdj, "tlsgd", true);
1673 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1674 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1675 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1676 PseudoSourceValue::getConstantPool(), 0,
1678 SDValue Chain = Argument.getValue(1);
1680 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1681 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1683 // call __tls_get_addr.
1686 Entry.Node = Argument;
1687 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1688 Args.push_back(Entry);
1689 // FIXME: is there useful debug info available here?
1690 std::pair<SDValue, SDValue> CallResult =
1691 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1692 false, false, false, false,
1693 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1694 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1695 return CallResult.first;
1698 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1699 // "local exec" model.
1701 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1702 SelectionDAG &DAG) const {
1703 const GlobalValue *GV = GA->getGlobal();
1704 DebugLoc dl = GA->getDebugLoc();
1706 SDValue Chain = DAG.getEntryNode();
1707 EVT PtrVT = getPointerTy();
1708 // Get the Thread Pointer
1709 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1711 if (GV->isDeclaration()) {
1712 MachineFunction &MF = DAG.getMachineFunction();
1713 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1714 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1715 // Initial exec model.
1716 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1717 ARMConstantPoolValue *CPV =
1718 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1719 ARMCP::CPValue, PCAdj, "gottpoff", true);
1720 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1721 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1722 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1723 PseudoSourceValue::getConstantPool(), 0,
1725 Chain = Offset.getValue(1);
1727 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1728 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1730 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1731 PseudoSourceValue::getConstantPool(), 0,
1735 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1736 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1737 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1738 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1739 PseudoSourceValue::getConstantPool(), 0,
1743 // The address of the thread local variable is the add of the thread
1744 // pointer with the offset of the variable.
1745 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1749 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1750 // TODO: implement the "local dynamic" model
1751 assert(Subtarget->isTargetELF() &&
1752 "TLS not implemented for non-ELF targets");
1753 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1754 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1755 // otherwise use the "Local Exec" TLS Model
1756 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1757 return LowerToTLSGeneralDynamicModel(GA, DAG);
1759 return LowerToTLSExecModels(GA, DAG);
1762 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1763 SelectionDAG &DAG) const {
1764 EVT PtrVT = getPointerTy();
1765 DebugLoc dl = Op.getDebugLoc();
1766 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1767 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1768 if (RelocM == Reloc::PIC_) {
1769 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1770 ARMConstantPoolValue *CPV =
1771 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1772 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1773 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1774 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1776 PseudoSourceValue::getConstantPool(), 0,
1778 SDValue Chain = Result.getValue(1);
1779 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1780 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1782 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1783 PseudoSourceValue::getGOT(), 0,
1787 // If we have T2 ops, we can materialize the address directly via movt/movw
1788 // pair. This is always cheaper.
1789 if (Subtarget->useMovt()) {
1790 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1791 DAG.getTargetGlobalAddress(GV, PtrVT));
1793 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1794 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1795 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1796 PseudoSourceValue::getConstantPool(), 0,
1802 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1803 SelectionDAG &DAG) const {
1804 MachineFunction &MF = DAG.getMachineFunction();
1805 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1806 unsigned ARMPCLabelIndex = 0;
1807 EVT PtrVT = getPointerTy();
1808 DebugLoc dl = Op.getDebugLoc();
1809 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1810 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1812 if (RelocM == Reloc::Static)
1813 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1815 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1816 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1817 ARMConstantPoolValue *CPV =
1818 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1819 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1821 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1823 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1824 PseudoSourceValue::getConstantPool(), 0,
1826 SDValue Chain = Result.getValue(1);
1828 if (RelocM == Reloc::PIC_) {
1829 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1830 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1833 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1834 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1835 PseudoSourceValue::getGOT(), 0,
1841 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1842 SelectionDAG &DAG) const {
1843 assert(Subtarget->isTargetELF() &&
1844 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1845 MachineFunction &MF = DAG.getMachineFunction();
1846 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1847 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1848 EVT PtrVT = getPointerTy();
1849 DebugLoc dl = Op.getDebugLoc();
1850 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1851 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1852 "_GLOBAL_OFFSET_TABLE_",
1853 ARMPCLabelIndex, PCAdj);
1854 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1855 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1856 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1857 PseudoSourceValue::getConstantPool(), 0,
1859 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1860 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1864 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1865 DebugLoc dl = Op.getDebugLoc();
1866 SDValue Val = DAG.getConstant(0, MVT::i32);
1867 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1868 Op.getOperand(1), Val);
1872 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1873 DebugLoc dl = Op.getDebugLoc();
1874 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1875 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1879 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1880 const ARMSubtarget *Subtarget) const {
1881 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1882 DebugLoc dl = Op.getDebugLoc();
1884 default: return SDValue(); // Don't custom lower most intrinsics.
1885 case Intrinsic::arm_thread_pointer: {
1886 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1887 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1889 case Intrinsic::eh_sjlj_lsda: {
1890 MachineFunction &MF = DAG.getMachineFunction();
1891 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1892 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1893 EVT PtrVT = getPointerTy();
1894 DebugLoc dl = Op.getDebugLoc();
1895 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1897 unsigned PCAdj = (RelocM != Reloc::PIC_)
1898 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1899 ARMConstantPoolValue *CPV =
1900 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1901 ARMCP::CPLSDA, PCAdj);
1902 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1903 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1905 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1906 PseudoSourceValue::getConstantPool(), 0,
1908 SDValue Chain = Result.getValue(1);
1910 if (RelocM == Reloc::PIC_) {
1911 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1912 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1919 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1920 const ARMSubtarget *Subtarget) {
1921 DebugLoc dl = Op.getDebugLoc();
1922 SDValue Op5 = Op.getOperand(5);
1923 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1924 // v6 and v7 can both handle barriers directly, but need handled a bit
1925 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1927 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1928 if (Subtarget->hasV7Ops())
1929 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1930 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1931 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1932 DAG.getConstant(0, MVT::i32));
1933 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1937 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1938 MachineFunction &MF = DAG.getMachineFunction();
1939 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1941 // vastart just stores the address of the VarArgsFrameIndex slot into the
1942 // memory location argument.
1943 DebugLoc dl = Op.getDebugLoc();
1944 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1945 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1946 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1947 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1952 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1953 SelectionDAG &DAG) const {
1954 SDNode *Node = Op.getNode();
1955 DebugLoc dl = Node->getDebugLoc();
1956 EVT VT = Node->getValueType(0);
1957 SDValue Chain = Op.getOperand(0);
1958 SDValue Size = Op.getOperand(1);
1959 SDValue Align = Op.getOperand(2);
1961 // Chain the dynamic stack allocation so that it doesn't modify the stack
1962 // pointer when other instructions are using the stack.
1963 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1965 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1966 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1967 if (AlignVal > StackAlign)
1968 // Do this now since selection pass cannot introduce new target
1969 // independent node.
1970 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1972 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1973 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1974 // do even more horrible hack later.
1975 MachineFunction &MF = DAG.getMachineFunction();
1976 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1977 if (AFI->isThumb1OnlyFunction()) {
1979 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1981 uint32_t Val = C->getZExtValue();
1982 if (Val <= 508 && ((Val & 3) == 0))
1986 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1989 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1990 SDValue Ops1[] = { Chain, Size, Align };
1991 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1992 Chain = Res.getValue(1);
1993 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1994 DAG.getIntPtrConstant(0, true), SDValue());
1995 SDValue Ops2[] = { Res, Chain };
1996 return DAG.getMergeValues(Ops2, 2, dl);
2000 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2001 SDValue &Root, SelectionDAG &DAG,
2002 DebugLoc dl) const {
2003 MachineFunction &MF = DAG.getMachineFunction();
2004 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2006 TargetRegisterClass *RC;
2007 if (AFI->isThumb1OnlyFunction())
2008 RC = ARM::tGPRRegisterClass;
2010 RC = ARM::GPRRegisterClass;
2012 // Transform the arguments stored in physical registers into virtual ones.
2013 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2014 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2017 if (NextVA.isMemLoc()) {
2018 MachineFrameInfo *MFI = MF.getFrameInfo();
2019 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
2021 // Create load node to retrieve arguments from the stack.
2022 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2023 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2024 PseudoSourceValue::getFixedStack(FI), 0,
2027 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2028 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2031 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2035 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2036 CallingConv::ID CallConv, bool isVarArg,
2037 const SmallVectorImpl<ISD::InputArg>
2039 DebugLoc dl, SelectionDAG &DAG,
2040 SmallVectorImpl<SDValue> &InVals)
2043 MachineFunction &MF = DAG.getMachineFunction();
2044 MachineFrameInfo *MFI = MF.getFrameInfo();
2046 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2048 // Assign locations to all of the incoming arguments.
2049 SmallVector<CCValAssign, 16> ArgLocs;
2050 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2052 CCInfo.AnalyzeFormalArguments(Ins,
2053 CCAssignFnForNode(CallConv, /* Return*/ false,
2056 SmallVector<SDValue, 16> ArgValues;
2058 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2059 CCValAssign &VA = ArgLocs[i];
2061 // Arguments stored in registers.
2062 if (VA.isRegLoc()) {
2063 EVT RegVT = VA.getLocVT();
2066 if (VA.needsCustom()) {
2067 // f64 and vector types are split up into multiple registers or
2068 // combinations of registers and stack slots.
2069 if (VA.getLocVT() == MVT::v2f64) {
2070 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2072 VA = ArgLocs[++i]; // skip ahead to next loc
2074 if (VA.isMemLoc()) {
2075 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
2077 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2078 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2079 PseudoSourceValue::getFixedStack(FI), 0,
2082 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2085 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2086 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2087 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2088 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2089 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2091 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2094 TargetRegisterClass *RC;
2096 if (RegVT == MVT::f32)
2097 RC = ARM::SPRRegisterClass;
2098 else if (RegVT == MVT::f64)
2099 RC = ARM::DPRRegisterClass;
2100 else if (RegVT == MVT::v2f64)
2101 RC = ARM::QPRRegisterClass;
2102 else if (RegVT == MVT::i32)
2103 RC = (AFI->isThumb1OnlyFunction() ?
2104 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2106 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2108 // Transform the arguments in physical registers into virtual ones.
2109 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2110 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2113 // If this is an 8 or 16-bit value, it is really passed promoted
2114 // to 32 bits. Insert an assert[sz]ext to capture this, then
2115 // truncate to the right size.
2116 switch (VA.getLocInfo()) {
2117 default: llvm_unreachable("Unknown loc info!");
2118 case CCValAssign::Full: break;
2119 case CCValAssign::BCvt:
2120 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2122 case CCValAssign::SExt:
2123 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2124 DAG.getValueType(VA.getValVT()));
2125 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2127 case CCValAssign::ZExt:
2128 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2129 DAG.getValueType(VA.getValVT()));
2130 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2134 InVals.push_back(ArgValue);
2136 } else { // VA.isRegLoc()
2139 assert(VA.isMemLoc());
2140 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2142 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2143 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2146 // Create load nodes to retrieve arguments from the stack.
2147 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2148 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2149 PseudoSourceValue::getFixedStack(FI), 0,
2156 static const unsigned GPRArgRegs[] = {
2157 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2160 unsigned NumGPRs = CCInfo.getFirstUnallocated
2161 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2163 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2164 unsigned VARegSize = (4 - NumGPRs) * 4;
2165 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2166 unsigned ArgOffset = CCInfo.getNextStackOffset();
2167 if (VARegSaveSize) {
2168 // If this function is vararg, store any remaining integer argument regs
2169 // to their spots on the stack so that they may be loaded by deferencing
2170 // the result of va_next.
2171 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2172 AFI->setVarArgsFrameIndex(
2173 MFI->CreateFixedObject(VARegSaveSize,
2174 ArgOffset + VARegSaveSize - VARegSize,
2176 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2179 SmallVector<SDValue, 4> MemOps;
2180 for (; NumGPRs < 4; ++NumGPRs) {
2181 TargetRegisterClass *RC;
2182 if (AFI->isThumb1OnlyFunction())
2183 RC = ARM::tGPRRegisterClass;
2185 RC = ARM::GPRRegisterClass;
2187 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2188 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2190 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2191 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2192 0, false, false, 0);
2193 MemOps.push_back(Store);
2194 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2195 DAG.getConstant(4, getPointerTy()));
2197 if (!MemOps.empty())
2198 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2199 &MemOps[0], MemOps.size());
2201 // This will point to the next argument passed via stack.
2202 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
2209 /// isFloatingPointZero - Return true if this is +0.0.
2210 static bool isFloatingPointZero(SDValue Op) {
2211 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2212 return CFP->getValueAPF().isPosZero();
2213 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2214 // Maybe this has already been legalized into the constant pool?
2215 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2216 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2217 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2218 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2219 return CFP->getValueAPF().isPosZero();
2225 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2226 /// the given operands.
2228 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2229 SDValue &ARMCC, SelectionDAG &DAG,
2230 DebugLoc dl) const {
2231 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2232 unsigned C = RHSC->getZExtValue();
2233 if (!isLegalICmpImmediate(C)) {
2234 // Constant does not fit, try adjusting it by one?
2239 if (isLegalICmpImmediate(C-1)) {
2240 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2241 RHS = DAG.getConstant(C-1, MVT::i32);
2246 if (C > 0 && isLegalICmpImmediate(C-1)) {
2247 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2248 RHS = DAG.getConstant(C-1, MVT::i32);
2253 if (isLegalICmpImmediate(C+1)) {
2254 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2255 RHS = DAG.getConstant(C+1, MVT::i32);
2260 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2261 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2262 RHS = DAG.getConstant(C+1, MVT::i32);
2269 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2270 ARMISD::NodeType CompareType;
2273 CompareType = ARMISD::CMP;
2278 CompareType = ARMISD::CMPZ;
2281 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2282 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2285 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2286 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2289 if (!isFloatingPointZero(RHS))
2290 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2292 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2293 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2296 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2297 EVT VT = Op.getValueType();
2298 SDValue LHS = Op.getOperand(0);
2299 SDValue RHS = Op.getOperand(1);
2300 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2301 SDValue TrueVal = Op.getOperand(2);
2302 SDValue FalseVal = Op.getOperand(3);
2303 DebugLoc dl = Op.getDebugLoc();
2305 if (LHS.getValueType() == MVT::i32) {
2307 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2308 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2309 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2312 ARMCC::CondCodes CondCode, CondCode2;
2313 FPCCToARMCC(CC, CondCode, CondCode2);
2315 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2316 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2317 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2318 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2320 if (CondCode2 != ARMCC::AL) {
2321 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2322 // FIXME: Needs another CMP because flag can have but one use.
2323 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2324 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2325 Result, TrueVal, ARMCC2, CCR, Cmp2);
2330 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2331 SDValue Chain = Op.getOperand(0);
2332 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2333 SDValue LHS = Op.getOperand(2);
2334 SDValue RHS = Op.getOperand(3);
2335 SDValue Dest = Op.getOperand(4);
2336 DebugLoc dl = Op.getDebugLoc();
2338 if (LHS.getValueType() == MVT::i32) {
2340 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2341 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2342 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2343 Chain, Dest, ARMCC, CCR,Cmp);
2346 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2347 ARMCC::CondCodes CondCode, CondCode2;
2348 FPCCToARMCC(CC, CondCode, CondCode2);
2350 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2351 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2352 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2353 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2354 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2355 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2356 if (CondCode2 != ARMCC::AL) {
2357 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2358 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2359 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2364 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2365 SDValue Chain = Op.getOperand(0);
2366 SDValue Table = Op.getOperand(1);
2367 SDValue Index = Op.getOperand(2);
2368 DebugLoc dl = Op.getDebugLoc();
2370 EVT PTy = getPointerTy();
2371 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2372 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2373 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2374 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2375 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2376 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2377 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2378 if (Subtarget->isThumb2()) {
2379 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2380 // which does another jump to the destination. This also makes it easier
2381 // to translate it to TBB / TBH later.
2382 // FIXME: This might not work if the function is extremely large.
2383 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2384 Addr, Op.getOperand(2), JTI, UId);
2386 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2387 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2388 PseudoSourceValue::getJumpTable(), 0,
2390 Chain = Addr.getValue(1);
2391 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2392 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2394 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2395 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2396 Chain = Addr.getValue(1);
2397 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2401 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2402 DebugLoc dl = Op.getDebugLoc();
2405 switch (Op.getOpcode()) {
2407 assert(0 && "Invalid opcode!");
2408 case ISD::FP_TO_SINT:
2409 Opc = ARMISD::FTOSI;
2411 case ISD::FP_TO_UINT:
2412 Opc = ARMISD::FTOUI;
2415 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2416 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2419 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2420 EVT VT = Op.getValueType();
2421 DebugLoc dl = Op.getDebugLoc();
2424 switch (Op.getOpcode()) {
2426 assert(0 && "Invalid opcode!");
2427 case ISD::SINT_TO_FP:
2428 Opc = ARMISD::SITOF;
2430 case ISD::UINT_TO_FP:
2431 Opc = ARMISD::UITOF;
2435 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2436 return DAG.getNode(Opc, dl, VT, Op);
2439 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2440 // Implement fcopysign with a fabs and a conditional fneg.
2441 SDValue Tmp0 = Op.getOperand(0);
2442 SDValue Tmp1 = Op.getOperand(1);
2443 DebugLoc dl = Op.getDebugLoc();
2444 EVT VT = Op.getValueType();
2445 EVT SrcVT = Tmp1.getValueType();
2446 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2447 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2448 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2449 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2450 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2453 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2454 MachineFunction &MF = DAG.getMachineFunction();
2455 MachineFrameInfo *MFI = MF.getFrameInfo();
2456 MFI->setReturnAddressIsTaken(true);
2458 EVT VT = Op.getValueType();
2459 DebugLoc dl = Op.getDebugLoc();
2460 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2462 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2463 SDValue Offset = DAG.getConstant(4, MVT::i32);
2464 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2465 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2466 NULL, 0, false, false, 0);
2469 // Return LR, which contains the return address. Mark it an implicit live-in.
2470 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2471 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2474 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2475 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2476 MFI->setFrameAddressIsTaken(true);
2478 EVT VT = Op.getValueType();
2479 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2480 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2481 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2482 ? ARM::R7 : ARM::R11;
2483 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2485 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2490 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2491 /// expand a bit convert where either the source or destination type is i64 to
2492 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2493 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2494 /// vectors), since the legalizer won't know what to do with that.
2495 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2496 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2497 DebugLoc dl = N->getDebugLoc();
2498 SDValue Op = N->getOperand(0);
2500 // This function is only supposed to be called for i64 types, either as the
2501 // source or destination of the bit convert.
2502 EVT SrcVT = Op.getValueType();
2503 EVT DstVT = N->getValueType(0);
2504 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2505 "ExpandBIT_CONVERT called for non-i64 type");
2507 // Turn i64->f64 into VMOVDRR.
2508 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2509 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2510 DAG.getConstant(0, MVT::i32));
2511 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2512 DAG.getConstant(1, MVT::i32));
2513 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2514 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2517 // Turn f64->i64 into VMOVRRD.
2518 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2519 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2520 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2521 // Merge the pieces into a single i64 value.
2522 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2528 /// getZeroVector - Returns a vector of specified type with all zero elements.
2530 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2531 assert(VT.isVector() && "Expected a vector type");
2533 // Zero vectors are used to represent vector negation and in those cases
2534 // will be implemented with the NEON VNEG instruction. However, VNEG does
2535 // not support i64 elements, so sometimes the zero vectors will need to be
2536 // explicitly constructed. For those cases, and potentially other uses in
2537 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2538 // to their dest type. This ensures they get CSE'd.
2540 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2541 SmallVector<SDValue, 8> Ops;
2544 if (VT.getSizeInBits() == 64) {
2545 Ops.assign(8, Cst); TVT = MVT::v8i8;
2547 Ops.assign(16, Cst); TVT = MVT::v16i8;
2549 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2551 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2554 /// getOnesVector - Returns a vector of specified type with all bits set.
2556 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2557 assert(VT.isVector() && "Expected a vector type");
2559 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2560 // dest type. This ensures they get CSE'd.
2562 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2563 SmallVector<SDValue, 8> Ops;
2566 if (VT.getSizeInBits() == 64) {
2567 Ops.assign(8, Cst); TVT = MVT::v8i8;
2569 Ops.assign(16, Cst); TVT = MVT::v16i8;
2571 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2573 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2576 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2577 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2578 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2579 SelectionDAG &DAG) const {
2580 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2581 EVT VT = Op.getValueType();
2582 unsigned VTBits = VT.getSizeInBits();
2583 DebugLoc dl = Op.getDebugLoc();
2584 SDValue ShOpLo = Op.getOperand(0);
2585 SDValue ShOpHi = Op.getOperand(1);
2586 SDValue ShAmt = Op.getOperand(2);
2588 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2590 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2592 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2593 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2594 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2595 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2596 DAG.getConstant(VTBits, MVT::i32));
2597 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2598 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2599 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2601 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2602 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2604 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2605 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2608 SDValue Ops[2] = { Lo, Hi };
2609 return DAG.getMergeValues(Ops, 2, dl);
2612 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2613 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2614 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2615 SelectionDAG &DAG) const {
2616 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2617 EVT VT = Op.getValueType();
2618 unsigned VTBits = VT.getSizeInBits();
2619 DebugLoc dl = Op.getDebugLoc();
2620 SDValue ShOpLo = Op.getOperand(0);
2621 SDValue ShOpHi = Op.getOperand(1);
2622 SDValue ShAmt = Op.getOperand(2);
2625 assert(Op.getOpcode() == ISD::SHL_PARTS);
2626 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2627 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2628 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2629 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2630 DAG.getConstant(VTBits, MVT::i32));
2631 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2632 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2634 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2635 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2636 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2638 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2639 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2642 SDValue Ops[2] = { Lo, Hi };
2643 return DAG.getMergeValues(Ops, 2, dl);
2646 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2647 const ARMSubtarget *ST) {
2648 EVT VT = N->getValueType(0);
2649 DebugLoc dl = N->getDebugLoc();
2651 if (!ST->hasV6T2Ops())
2654 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2655 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2658 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2659 const ARMSubtarget *ST) {
2660 EVT VT = N->getValueType(0);
2661 DebugLoc dl = N->getDebugLoc();
2663 // Lower vector shifts on NEON to use VSHL.
2664 if (VT.isVector()) {
2665 assert(ST->hasNEON() && "unexpected vector shift");
2667 // Left shifts translate directly to the vshiftu intrinsic.
2668 if (N->getOpcode() == ISD::SHL)
2669 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2670 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2671 N->getOperand(0), N->getOperand(1));
2673 assert((N->getOpcode() == ISD::SRA ||
2674 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2676 // NEON uses the same intrinsics for both left and right shifts. For
2677 // right shifts, the shift amounts are negative, so negate the vector of
2679 EVT ShiftVT = N->getOperand(1).getValueType();
2680 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2681 getZeroVector(ShiftVT, DAG, dl),
2683 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2684 Intrinsic::arm_neon_vshifts :
2685 Intrinsic::arm_neon_vshiftu);
2686 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2687 DAG.getConstant(vshiftInt, MVT::i32),
2688 N->getOperand(0), NegatedCount);
2691 // We can get here for a node like i32 = ISD::SHL i32, i64
2695 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2696 "Unknown shift to lower!");
2698 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2699 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2700 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2703 // If we are in thumb mode, we don't have RRX.
2704 if (ST->isThumb1Only()) return SDValue();
2706 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2707 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2708 DAG.getConstant(0, MVT::i32));
2709 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2710 DAG.getConstant(1, MVT::i32));
2712 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2713 // captures the result into a carry flag.
2714 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2715 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2717 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2718 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2720 // Merge the pieces into a single i64 value.
2721 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2724 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2725 SDValue TmpOp0, TmpOp1;
2726 bool Invert = false;
2730 SDValue Op0 = Op.getOperand(0);
2731 SDValue Op1 = Op.getOperand(1);
2732 SDValue CC = Op.getOperand(2);
2733 EVT VT = Op.getValueType();
2734 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2735 DebugLoc dl = Op.getDebugLoc();
2737 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2738 switch (SetCCOpcode) {
2739 default: llvm_unreachable("Illegal FP comparison"); break;
2741 case ISD::SETNE: Invert = true; // Fallthrough
2743 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2745 case ISD::SETLT: Swap = true; // Fallthrough
2747 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2749 case ISD::SETLE: Swap = true; // Fallthrough
2751 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2752 case ISD::SETUGE: Swap = true; // Fallthrough
2753 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2754 case ISD::SETUGT: Swap = true; // Fallthrough
2755 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2756 case ISD::SETUEQ: Invert = true; // Fallthrough
2758 // Expand this to (OLT | OGT).
2762 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2763 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2765 case ISD::SETUO: Invert = true; // Fallthrough
2767 // Expand this to (OLT | OGE).
2771 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2772 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2776 // Integer comparisons.
2777 switch (SetCCOpcode) {
2778 default: llvm_unreachable("Illegal integer comparison"); break;
2779 case ISD::SETNE: Invert = true;
2780 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2781 case ISD::SETLT: Swap = true;
2782 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2783 case ISD::SETLE: Swap = true;
2784 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2785 case ISD::SETULT: Swap = true;
2786 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2787 case ISD::SETULE: Swap = true;
2788 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2791 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2792 if (Opc == ARMISD::VCEQ) {
2795 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2797 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2800 // Ignore bitconvert.
2801 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2802 AndOp = AndOp.getOperand(0);
2804 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2806 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2807 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2814 std::swap(Op0, Op1);
2816 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2819 Result = DAG.getNOT(dl, Result, VT);
2824 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2825 /// valid vector constant for a NEON instruction with a "modified immediate"
2826 /// operand (e.g., VMOV). If so, return either the constant being
2827 /// splatted or the encoded value, depending on the DoEncode parameter. The
2828 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2829 /// bits7-0=Immediate.
2830 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2831 unsigned SplatBitSize, SelectionDAG &DAG,
2832 bool isVMOV, bool DoEncode) {
2833 unsigned Op, Cmode, Imm;
2836 // SplatBitSize is set to the smallest size that splats the vector, so a
2837 // zero vector will always have SplatBitSize == 8. However, NEON modified
2838 // immediate instructions others than VMOV do not support the 8-bit encoding
2839 // of a zero vector, and the default encoding of zero is supposed to be the
2845 switch (SplatBitSize) {
2847 // Any 1-byte value is OK. Op=0, Cmode=1110.
2848 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2855 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2857 if ((SplatBits & ~0xff) == 0) {
2858 // Value = 0x00nn: Op=x, Cmode=100x.
2863 if ((SplatBits & ~0xff00) == 0) {
2864 // Value = 0xnn00: Op=x, Cmode=101x.
2866 Imm = SplatBits >> 8;
2872 // NEON's 32-bit VMOV supports splat values where:
2873 // * only one byte is nonzero, or
2874 // * the least significant byte is 0xff and the second byte is nonzero, or
2875 // * the least significant 2 bytes are 0xff and the third is nonzero.
2877 if ((SplatBits & ~0xff) == 0) {
2878 // Value = 0x000000nn: Op=x, Cmode=000x.
2883 if ((SplatBits & ~0xff00) == 0) {
2884 // Value = 0x0000nn00: Op=x, Cmode=001x.
2886 Imm = SplatBits >> 8;
2889 if ((SplatBits & ~0xff0000) == 0) {
2890 // Value = 0x00nn0000: Op=x, Cmode=010x.
2892 Imm = SplatBits >> 16;
2895 if ((SplatBits & ~0xff000000) == 0) {
2896 // Value = 0xnn000000: Op=x, Cmode=011x.
2898 Imm = SplatBits >> 24;
2902 if ((SplatBits & ~0xffff) == 0 &&
2903 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2904 // Value = 0x0000nnff: Op=x, Cmode=1100.
2906 Imm = SplatBits >> 8;
2911 if ((SplatBits & ~0xffffff) == 0 &&
2912 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2913 // Value = 0x00nnffff: Op=x, Cmode=1101.
2915 Imm = SplatBits >> 16;
2916 SplatBits |= 0xffff;
2920 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2921 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2922 // VMOV.I32. A (very) minor optimization would be to replicate the value
2923 // and fall through here to test for a valid 64-bit splat. But, then the
2924 // caller would also need to check and handle the change in size.
2928 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2931 uint64_t BitMask = 0xff;
2933 unsigned ImmMask = 1;
2935 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2936 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2939 } else if ((SplatBits & BitMask) != 0) {
2945 // Op=1, Cmode=1110.
2954 llvm_unreachable("unexpected size for isNEONModifiedImm");
2959 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2960 return DAG.getTargetConstant(SplatBits, VT);
2964 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2965 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2966 /// size, return the encoded value for that immediate. The ByteSize field
2967 /// indicates the number of bytes of each element [1248].
2968 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2969 SelectionDAG &DAG) {
2970 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2971 APInt SplatBits, SplatUndef;
2972 unsigned SplatBitSize;
2974 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2975 HasAnyUndefs, ByteSize * 8))
2978 if (SplatBitSize > ByteSize * 8)
2981 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2982 SplatBitSize, DAG, isVMOV, true);
2985 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2986 bool &ReverseVEXT, unsigned &Imm) {
2987 unsigned NumElts = VT.getVectorNumElements();
2988 ReverseVEXT = false;
2991 // If this is a VEXT shuffle, the immediate value is the index of the first
2992 // element. The other shuffle indices must be the successive elements after
2994 unsigned ExpectedElt = Imm;
2995 for (unsigned i = 1; i < NumElts; ++i) {
2996 // Increment the expected index. If it wraps around, it may still be
2997 // a VEXT but the source vectors must be swapped.
2999 if (ExpectedElt == NumElts * 2) {
3004 if (ExpectedElt != static_cast<unsigned>(M[i]))
3008 // Adjust the index value if the source operands will be swapped.
3015 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3016 /// instruction with the specified blocksize. (The order of the elements
3017 /// within each block of the vector is reversed.)
3018 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3019 unsigned BlockSize) {
3020 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3021 "Only possible block sizes for VREV are: 16, 32, 64");
3023 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3027 unsigned NumElts = VT.getVectorNumElements();
3028 unsigned BlockElts = M[0] + 1;
3030 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3033 for (unsigned i = 0; i < NumElts; ++i) {
3034 if ((unsigned) M[i] !=
3035 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3042 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3043 unsigned &WhichResult) {
3044 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3048 unsigned NumElts = VT.getVectorNumElements();
3049 WhichResult = (M[0] == 0 ? 0 : 1);
3050 for (unsigned i = 0; i < NumElts; i += 2) {
3051 if ((unsigned) M[i] != i + WhichResult ||
3052 (unsigned) M[i+1] != i + NumElts + WhichResult)
3058 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3059 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3060 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3061 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3062 unsigned &WhichResult) {
3063 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3067 unsigned NumElts = VT.getVectorNumElements();
3068 WhichResult = (M[0] == 0 ? 0 : 1);
3069 for (unsigned i = 0; i < NumElts; i += 2) {
3070 if ((unsigned) M[i] != i + WhichResult ||
3071 (unsigned) M[i+1] != i + WhichResult)
3077 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3078 unsigned &WhichResult) {
3079 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3083 unsigned NumElts = VT.getVectorNumElements();
3084 WhichResult = (M[0] == 0 ? 0 : 1);
3085 for (unsigned i = 0; i != NumElts; ++i) {
3086 if ((unsigned) M[i] != 2 * i + WhichResult)
3090 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3091 if (VT.is64BitVector() && EltSz == 32)
3097 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3098 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3099 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3100 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3101 unsigned &WhichResult) {
3102 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3106 unsigned Half = VT.getVectorNumElements() / 2;
3107 WhichResult = (M[0] == 0 ? 0 : 1);
3108 for (unsigned j = 0; j != 2; ++j) {
3109 unsigned Idx = WhichResult;
3110 for (unsigned i = 0; i != Half; ++i) {
3111 if ((unsigned) M[i + j * Half] != Idx)
3117 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3118 if (VT.is64BitVector() && EltSz == 32)
3124 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3125 unsigned &WhichResult) {
3126 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3130 unsigned NumElts = VT.getVectorNumElements();
3131 WhichResult = (M[0] == 0 ? 0 : 1);
3132 unsigned Idx = WhichResult * NumElts / 2;
3133 for (unsigned i = 0; i != NumElts; i += 2) {
3134 if ((unsigned) M[i] != Idx ||
3135 (unsigned) M[i+1] != Idx + NumElts)
3140 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3141 if (VT.is64BitVector() && EltSz == 32)
3147 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3148 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3149 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3150 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3151 unsigned &WhichResult) {
3152 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3156 unsigned NumElts = VT.getVectorNumElements();
3157 WhichResult = (M[0] == 0 ? 0 : 1);
3158 unsigned Idx = WhichResult * NumElts / 2;
3159 for (unsigned i = 0; i != NumElts; i += 2) {
3160 if ((unsigned) M[i] != Idx ||
3161 (unsigned) M[i+1] != Idx)
3166 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3167 if (VT.is64BitVector() && EltSz == 32)
3174 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3175 // Canonicalize all-zeros and all-ones vectors.
3176 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3177 if (ConstVal->isNullValue())
3178 return getZeroVector(VT, DAG, dl);
3179 if (ConstVal->isAllOnesValue())
3180 return getOnesVector(VT, DAG, dl);
3183 if (VT.is64BitVector()) {
3184 switch (Val.getValueType().getSizeInBits()) {
3185 case 8: CanonicalVT = MVT::v8i8; break;
3186 case 16: CanonicalVT = MVT::v4i16; break;
3187 case 32: CanonicalVT = MVT::v2i32; break;
3188 case 64: CanonicalVT = MVT::v1i64; break;
3189 default: llvm_unreachable("unexpected splat element type"); break;
3192 assert(VT.is128BitVector() && "unknown splat vector size");
3193 switch (Val.getValueType().getSizeInBits()) {
3194 case 8: CanonicalVT = MVT::v16i8; break;
3195 case 16: CanonicalVT = MVT::v8i16; break;
3196 case 32: CanonicalVT = MVT::v4i32; break;
3197 case 64: CanonicalVT = MVT::v2i64; break;
3198 default: llvm_unreachable("unexpected splat element type"); break;
3202 // Build a canonical splat for this value.
3203 SmallVector<SDValue, 8> Ops;
3204 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3205 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3207 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3210 // If this is a case we can't handle, return null and let the default
3211 // expansion code take care of it.
3212 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3213 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3214 DebugLoc dl = Op.getDebugLoc();
3215 EVT VT = Op.getValueType();
3217 APInt SplatBits, SplatUndef;
3218 unsigned SplatBitSize;
3220 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3221 if (SplatBitSize <= 64) {
3222 // Check if an immediate VMOV works.
3223 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3224 SplatUndef.getZExtValue(),
3225 SplatBitSize, DAG, true, false);
3227 return BuildSplat(Val, VT, DAG, dl);
3231 // Scan through the operands to see if only one value is used.
3232 unsigned NumElts = VT.getVectorNumElements();
3233 bool isOnlyLowElement = true;
3234 bool usesOnlyOneValue = true;
3235 bool isConstant = true;
3237 for (unsigned i = 0; i < NumElts; ++i) {
3238 SDValue V = Op.getOperand(i);
3239 if (V.getOpcode() == ISD::UNDEF)
3242 isOnlyLowElement = false;
3243 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3246 if (!Value.getNode())
3248 else if (V != Value)
3249 usesOnlyOneValue = false;
3252 if (!Value.getNode())
3253 return DAG.getUNDEF(VT);
3255 if (isOnlyLowElement)
3256 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3258 // If all elements are constants, fall back to the default expansion, which
3259 // will generate a load from the constant pool.
3263 // Use VDUP for non-constant splats.
3264 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3265 if (usesOnlyOneValue && EltSize <= 32)
3266 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3268 // Vectors with 32- or 64-bit elements can be built by directly assigning
3269 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3270 // will be legalized.
3271 if (EltSize >= 32) {
3272 // Do the expansion with floating-point types, since that is what the VFP
3273 // registers are defined to use, and since i64 is not legal.
3274 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3275 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3276 SmallVector<SDValue, 8> Ops;
3277 for (unsigned i = 0; i < NumElts; ++i)
3278 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3279 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3280 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3286 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3287 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3288 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3289 /// are assumed to be legal.
3291 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3293 if (VT.getVectorNumElements() == 4 &&
3294 (VT.is128BitVector() || VT.is64BitVector())) {
3295 unsigned PFIndexes[4];
3296 for (unsigned i = 0; i != 4; ++i) {
3300 PFIndexes[i] = M[i];
3303 // Compute the index in the perfect shuffle table.
3304 unsigned PFTableIndex =
3305 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3306 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3307 unsigned Cost = (PFEntry >> 30);
3314 unsigned Imm, WhichResult;
3316 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3317 return (EltSize >= 32 ||
3318 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3319 isVREVMask(M, VT, 64) ||
3320 isVREVMask(M, VT, 32) ||
3321 isVREVMask(M, VT, 16) ||
3322 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3323 isVTRNMask(M, VT, WhichResult) ||
3324 isVUZPMask(M, VT, WhichResult) ||
3325 isVZIPMask(M, VT, WhichResult) ||
3326 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3327 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3328 isVZIP_v_undef_Mask(M, VT, WhichResult));
3331 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3332 /// the specified operations to build the shuffle.
3333 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3334 SDValue RHS, SelectionDAG &DAG,
3336 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3337 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3338 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3341 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3350 OP_VUZPL, // VUZP, left result
3351 OP_VUZPR, // VUZP, right result
3352 OP_VZIPL, // VZIP, left result
3353 OP_VZIPR, // VZIP, right result
3354 OP_VTRNL, // VTRN, left result
3355 OP_VTRNR // VTRN, right result
3358 if (OpNum == OP_COPY) {
3359 if (LHSID == (1*9+2)*9+3) return LHS;
3360 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3364 SDValue OpLHS, OpRHS;
3365 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3366 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3367 EVT VT = OpLHS.getValueType();
3370 default: llvm_unreachable("Unknown shuffle opcode!");
3372 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3377 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3378 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3382 return DAG.getNode(ARMISD::VEXT, dl, VT,
3384 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3387 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3388 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3391 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3392 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3395 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3396 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3400 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3401 SDValue V1 = Op.getOperand(0);
3402 SDValue V2 = Op.getOperand(1);
3403 DebugLoc dl = Op.getDebugLoc();
3404 EVT VT = Op.getValueType();
3405 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3406 SmallVector<int, 8> ShuffleMask;
3408 // Convert shuffles that are directly supported on NEON to target-specific
3409 // DAG nodes, instead of keeping them as shuffles and matching them again
3410 // during code selection. This is more efficient and avoids the possibility
3411 // of inconsistencies between legalization and selection.
3412 // FIXME: floating-point vectors should be canonicalized to integer vectors
3413 // of the same time so that they get CSEd properly.
3414 SVN->getMask(ShuffleMask);
3416 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3417 if (EltSize <= 32) {
3418 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3419 int Lane = SVN->getSplatIndex();
3420 // If this is undef splat, generate it via "just" vdup, if possible.
3421 if (Lane == -1) Lane = 0;
3423 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3424 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3426 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3427 DAG.getConstant(Lane, MVT::i32));
3432 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3435 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3436 DAG.getConstant(Imm, MVT::i32));
3439 if (isVREVMask(ShuffleMask, VT, 64))
3440 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3441 if (isVREVMask(ShuffleMask, VT, 32))
3442 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3443 if (isVREVMask(ShuffleMask, VT, 16))
3444 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3446 // Check for Neon shuffles that modify both input vectors in place.
3447 // If both results are used, i.e., if there are two shuffles with the same
3448 // source operands and with masks corresponding to both results of one of
3449 // these operations, DAG memoization will ensure that a single node is
3450 // used for both shuffles.
3451 unsigned WhichResult;
3452 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3453 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3454 V1, V2).getValue(WhichResult);
3455 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3456 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3457 V1, V2).getValue(WhichResult);
3458 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3459 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3460 V1, V2).getValue(WhichResult);
3462 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3463 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3464 V1, V1).getValue(WhichResult);
3465 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3466 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3467 V1, V1).getValue(WhichResult);
3468 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3469 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3470 V1, V1).getValue(WhichResult);
3473 // If the shuffle is not directly supported and it has 4 elements, use
3474 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3475 unsigned NumElts = VT.getVectorNumElements();
3477 unsigned PFIndexes[4];
3478 for (unsigned i = 0; i != 4; ++i) {
3479 if (ShuffleMask[i] < 0)
3482 PFIndexes[i] = ShuffleMask[i];
3485 // Compute the index in the perfect shuffle table.
3486 unsigned PFTableIndex =
3487 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3488 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3489 unsigned Cost = (PFEntry >> 30);
3492 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3495 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3496 if (EltSize >= 32) {
3497 // Do the expansion with floating-point types, since that is what the VFP
3498 // registers are defined to use, and since i64 is not legal.
3499 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3500 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3501 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3502 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3503 SmallVector<SDValue, 8> Ops;
3504 for (unsigned i = 0; i < NumElts; ++i) {
3505 if (ShuffleMask[i] < 0)
3506 Ops.push_back(DAG.getUNDEF(EltVT));
3508 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3509 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3510 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3513 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3514 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3520 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3521 EVT VT = Op.getValueType();
3522 DebugLoc dl = Op.getDebugLoc();
3523 SDValue Vec = Op.getOperand(0);
3524 SDValue Lane = Op.getOperand(1);
3525 assert(VT == MVT::i32 &&
3526 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3527 "unexpected type for custom-lowering vector extract");
3528 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3531 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3532 // The only time a CONCAT_VECTORS operation can have legal types is when
3533 // two 64-bit vectors are concatenated to a 128-bit vector.
3534 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3535 "unexpected CONCAT_VECTORS");
3536 DebugLoc dl = Op.getDebugLoc();
3537 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3538 SDValue Op0 = Op.getOperand(0);
3539 SDValue Op1 = Op.getOperand(1);
3540 if (Op0.getOpcode() != ISD::UNDEF)
3541 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3542 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3543 DAG.getIntPtrConstant(0));
3544 if (Op1.getOpcode() != ISD::UNDEF)
3545 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3546 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3547 DAG.getIntPtrConstant(1));
3548 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3551 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3552 switch (Op.getOpcode()) {
3553 default: llvm_unreachable("Don't know how to custom lower this!");
3554 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3555 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3556 case ISD::GlobalAddress:
3557 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3558 LowerGlobalAddressELF(Op, DAG);
3559 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3560 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3561 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3562 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3563 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3564 case ISD::VASTART: return LowerVASTART(Op, DAG);
3565 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3566 case ISD::SINT_TO_FP:
3567 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3568 case ISD::FP_TO_SINT:
3569 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3570 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3571 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3572 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3573 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3574 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3575 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3576 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3578 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3581 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3582 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3583 case ISD::SRL_PARTS:
3584 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3585 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3586 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3587 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3588 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3589 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3590 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3595 /// ReplaceNodeResults - Replace the results of node with an illegal result
3596 /// type with new values built out of custom code.
3597 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3598 SmallVectorImpl<SDValue>&Results,
3599 SelectionDAG &DAG) const {
3601 switch (N->getOpcode()) {
3603 llvm_unreachable("Don't know how to custom expand this!");
3605 case ISD::BIT_CONVERT:
3606 Res = ExpandBIT_CONVERT(N, DAG);
3610 Res = LowerShift(N, DAG, Subtarget);
3614 Results.push_back(Res);
3617 //===----------------------------------------------------------------------===//
3618 // ARM Scheduler Hooks
3619 //===----------------------------------------------------------------------===//
3622 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3623 MachineBasicBlock *BB,
3624 unsigned Size) const {
3625 unsigned dest = MI->getOperand(0).getReg();
3626 unsigned ptr = MI->getOperand(1).getReg();
3627 unsigned oldval = MI->getOperand(2).getReg();
3628 unsigned newval = MI->getOperand(3).getReg();
3629 unsigned scratch = BB->getParent()->getRegInfo()
3630 .createVirtualRegister(ARM::GPRRegisterClass);
3631 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3632 DebugLoc dl = MI->getDebugLoc();
3633 bool isThumb2 = Subtarget->isThumb2();
3635 unsigned ldrOpc, strOpc;
3637 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3639 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3640 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3643 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3644 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3647 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3648 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3652 MachineFunction *MF = BB->getParent();
3653 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3654 MachineFunction::iterator It = BB;
3655 ++It; // insert the new blocks after the current block
3657 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3658 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3659 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3660 MF->insert(It, loop1MBB);
3661 MF->insert(It, loop2MBB);
3662 MF->insert(It, exitMBB);
3663 exitMBB->transferSuccessors(BB);
3667 // fallthrough --> loop1MBB
3668 BB->addSuccessor(loop1MBB);
3671 // ldrex dest, [ptr]
3675 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3676 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3677 .addReg(dest).addReg(oldval));
3678 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3679 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3680 BB->addSuccessor(loop2MBB);
3681 BB->addSuccessor(exitMBB);
3684 // strex scratch, newval, [ptr]
3688 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3690 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3691 .addReg(scratch).addImm(0));
3692 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3693 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3694 BB->addSuccessor(loop1MBB);
3695 BB->addSuccessor(exitMBB);
3701 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3707 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3708 unsigned Size, unsigned BinOpcode) const {
3709 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3710 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3712 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3713 MachineFunction *MF = BB->getParent();
3714 MachineFunction::iterator It = BB;
3717 unsigned dest = MI->getOperand(0).getReg();
3718 unsigned ptr = MI->getOperand(1).getReg();
3719 unsigned incr = MI->getOperand(2).getReg();
3720 DebugLoc dl = MI->getDebugLoc();
3722 bool isThumb2 = Subtarget->isThumb2();
3723 unsigned ldrOpc, strOpc;
3725 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3727 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3728 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3731 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3732 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3735 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3736 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3740 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3741 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3742 MF->insert(It, loopMBB);
3743 MF->insert(It, exitMBB);
3744 exitMBB->transferSuccessors(BB);
3746 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3747 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3748 unsigned scratch2 = (!BinOpcode) ? incr :
3749 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3753 // fallthrough --> loopMBB
3754 BB->addSuccessor(loopMBB);
3758 // <binop> scratch2, dest, incr
3759 // strex scratch, scratch2, ptr
3762 // fallthrough --> exitMBB
3764 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3766 // operand order needs to go the other way for NAND
3767 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3768 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3769 addReg(incr).addReg(dest)).addReg(0);
3771 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3772 addReg(dest).addReg(incr)).addReg(0);
3775 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3777 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3778 .addReg(scratch).addImm(0));
3779 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3780 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3782 BB->addSuccessor(loopMBB);
3783 BB->addSuccessor(exitMBB);
3789 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3795 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3796 MachineBasicBlock *BB) const {
3797 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3798 DebugLoc dl = MI->getDebugLoc();
3799 bool isThumb2 = Subtarget->isThumb2();
3800 switch (MI->getOpcode()) {
3803 llvm_unreachable("Unexpected instr type to insert");
3805 case ARM::ATOMIC_LOAD_ADD_I8:
3806 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3807 case ARM::ATOMIC_LOAD_ADD_I16:
3808 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3809 case ARM::ATOMIC_LOAD_ADD_I32:
3810 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3812 case ARM::ATOMIC_LOAD_AND_I8:
3813 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3814 case ARM::ATOMIC_LOAD_AND_I16:
3815 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3816 case ARM::ATOMIC_LOAD_AND_I32:
3817 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3819 case ARM::ATOMIC_LOAD_OR_I8:
3820 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3821 case ARM::ATOMIC_LOAD_OR_I16:
3822 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3823 case ARM::ATOMIC_LOAD_OR_I32:
3824 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3826 case ARM::ATOMIC_LOAD_XOR_I8:
3827 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3828 case ARM::ATOMIC_LOAD_XOR_I16:
3829 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3830 case ARM::ATOMIC_LOAD_XOR_I32:
3831 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3833 case ARM::ATOMIC_LOAD_NAND_I8:
3834 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3835 case ARM::ATOMIC_LOAD_NAND_I16:
3836 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3837 case ARM::ATOMIC_LOAD_NAND_I32:
3838 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3840 case ARM::ATOMIC_LOAD_SUB_I8:
3841 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3842 case ARM::ATOMIC_LOAD_SUB_I16:
3843 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3844 case ARM::ATOMIC_LOAD_SUB_I32:
3845 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3847 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3848 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3849 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3851 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3852 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3853 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3855 case ARM::tMOVCCr_pseudo: {
3856 // To "insert" a SELECT_CC instruction, we actually have to insert the
3857 // diamond control-flow pattern. The incoming instruction knows the
3858 // destination vreg to set, the condition code register to branch on, the
3859 // true/false values to select between, and a branch opcode to use.
3860 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3861 MachineFunction::iterator It = BB;
3867 // cmpTY ccX, r1, r2
3869 // fallthrough --> copy0MBB
3870 MachineBasicBlock *thisMBB = BB;
3871 MachineFunction *F = BB->getParent();
3872 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3873 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3874 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3875 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3876 F->insert(It, copy0MBB);
3877 F->insert(It, sinkMBB);
3878 // Update machine-CFG edges by first adding all successors of the current
3879 // block to the new block which will contain the Phi node for the select.
3880 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3881 E = BB->succ_end(); I != E; ++I)
3882 sinkMBB->addSuccessor(*I);
3883 // Next, remove all successors of the current block, and add the true
3884 // and fallthrough blocks as its successors.
3885 while (!BB->succ_empty())
3886 BB->removeSuccessor(BB->succ_begin());
3887 BB->addSuccessor(copy0MBB);
3888 BB->addSuccessor(sinkMBB);
3891 // %FalseValue = ...
3892 // # fallthrough to sinkMBB
3895 // Update machine-CFG edges
3896 BB->addSuccessor(sinkMBB);
3899 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3902 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3903 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3904 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3906 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3913 case ARM::t2SUBrSPi_:
3914 case ARM::t2SUBrSPi12_:
3915 case ARM::t2SUBrSPs_: {
3916 MachineFunction *MF = BB->getParent();
3917 unsigned DstReg = MI->getOperand(0).getReg();
3918 unsigned SrcReg = MI->getOperand(1).getReg();
3919 bool DstIsDead = MI->getOperand(0).isDead();
3920 bool SrcIsKill = MI->getOperand(1).isKill();
3922 if (SrcReg != ARM::SP) {
3923 // Copy the source to SP from virtual register.
3924 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3925 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3926 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3927 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3928 .addReg(SrcReg, getKillRegState(SrcIsKill));
3932 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3933 switch (MI->getOpcode()) {
3935 llvm_unreachable("Unexpected pseudo instruction!");
3941 OpOpc = ARM::tADDspr;
3944 OpOpc = ARM::tSUBspi;
3946 case ARM::t2SUBrSPi_:
3947 OpOpc = ARM::t2SUBrSPi;
3948 NeedPred = true; NeedCC = true;
3950 case ARM::t2SUBrSPi12_:
3951 OpOpc = ARM::t2SUBrSPi12;
3954 case ARM::t2SUBrSPs_:
3955 OpOpc = ARM::t2SUBrSPs;
3956 NeedPred = true; NeedCC = true; NeedOp3 = true;
3959 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3960 if (OpOpc == ARM::tAND)
3961 AddDefaultT1CC(MIB);
3962 MIB.addReg(ARM::SP);
3963 MIB.addOperand(MI->getOperand(2));
3965 MIB.addOperand(MI->getOperand(3));
3967 AddDefaultPred(MIB);
3971 // Copy the result from SP to virtual register.
3972 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3973 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3974 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3975 BuildMI(BB, dl, TII->get(CopyOpc))
3976 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3978 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3984 //===----------------------------------------------------------------------===//
3985 // ARM Optimization Hooks
3986 //===----------------------------------------------------------------------===//
3989 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3990 TargetLowering::DAGCombinerInfo &DCI) {
3991 SelectionDAG &DAG = DCI.DAG;
3992 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3993 EVT VT = N->getValueType(0);
3994 unsigned Opc = N->getOpcode();
3995 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3996 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3997 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3998 ISD::CondCode CC = ISD::SETCC_INVALID;
4001 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4003 SDValue CCOp = Slct.getOperand(0);
4004 if (CCOp.getOpcode() == ISD::SETCC)
4005 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4008 bool DoXform = false;
4010 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4013 if (LHS.getOpcode() == ISD::Constant &&
4014 cast<ConstantSDNode>(LHS)->isNullValue()) {
4016 } else if (CC != ISD::SETCC_INVALID &&
4017 RHS.getOpcode() == ISD::Constant &&
4018 cast<ConstantSDNode>(RHS)->isNullValue()) {
4019 std::swap(LHS, RHS);
4020 SDValue Op0 = Slct.getOperand(0);
4021 EVT OpVT = isSlctCC ? Op0.getValueType() :
4022 Op0.getOperand(0).getValueType();
4023 bool isInt = OpVT.isInteger();
4024 CC = ISD::getSetCCInverse(CC, isInt);
4026 if (!TLI.isCondCodeLegal(CC, OpVT))
4027 return SDValue(); // Inverse operator isn't legal.
4034 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4036 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4037 Slct.getOperand(0), Slct.getOperand(1), CC);
4038 SDValue CCOp = Slct.getOperand(0);
4040 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4041 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4042 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4043 CCOp, OtherOp, Result);
4048 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4049 static SDValue PerformADDCombine(SDNode *N,
4050 TargetLowering::DAGCombinerInfo &DCI) {
4051 // added by evan in r37685 with no testcase.
4052 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4054 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4055 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4056 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4057 if (Result.getNode()) return Result;
4059 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4060 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4061 if (Result.getNode()) return Result;
4067 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4068 static SDValue PerformSUBCombine(SDNode *N,
4069 TargetLowering::DAGCombinerInfo &DCI) {
4070 // added by evan in r37685 with no testcase.
4071 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4073 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4074 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4075 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4076 if (Result.getNode()) return Result;
4082 static SDValue PerformMULCombine(SDNode *N,
4083 TargetLowering::DAGCombinerInfo &DCI,
4084 const ARMSubtarget *Subtarget) {
4085 SelectionDAG &DAG = DCI.DAG;
4087 if (Subtarget->isThumb1Only())
4090 if (DAG.getMachineFunction().
4091 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4094 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4097 EVT VT = N->getValueType(0);
4101 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4105 uint64_t MulAmt = C->getZExtValue();
4106 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4107 ShiftAmt = ShiftAmt & (32 - 1);
4108 SDValue V = N->getOperand(0);
4109 DebugLoc DL = N->getDebugLoc();
4112 MulAmt >>= ShiftAmt;
4113 if (isPowerOf2_32(MulAmt - 1)) {
4114 // (mul x, 2^N + 1) => (add (shl x, N), x)
4115 Res = DAG.getNode(ISD::ADD, DL, VT,
4116 V, DAG.getNode(ISD::SHL, DL, VT,
4117 V, DAG.getConstant(Log2_32(MulAmt-1),
4119 } else if (isPowerOf2_32(MulAmt + 1)) {
4120 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4121 Res = DAG.getNode(ISD::SUB, DL, VT,
4122 DAG.getNode(ISD::SHL, DL, VT,
4123 V, DAG.getConstant(Log2_32(MulAmt+1),
4130 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4131 DAG.getConstant(ShiftAmt, MVT::i32));
4133 // Do not add new nodes to DAG combiner worklist.
4134 DCI.CombineTo(N, Res, false);
4138 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4139 /// ARMISD::VMOVRRD.
4140 static SDValue PerformVMOVRRDCombine(SDNode *N,
4141 TargetLowering::DAGCombinerInfo &DCI) {
4142 // fmrrd(fmdrr x, y) -> x,y
4143 SDValue InDouble = N->getOperand(0);
4144 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4145 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4149 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4150 /// operand of a vector shift operation, where all the elements of the
4151 /// build_vector must have the same constant integer value.
4152 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4153 // Ignore bit_converts.
4154 while (Op.getOpcode() == ISD::BIT_CONVERT)
4155 Op = Op.getOperand(0);
4156 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4157 APInt SplatBits, SplatUndef;
4158 unsigned SplatBitSize;
4160 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4161 HasAnyUndefs, ElementBits) ||
4162 SplatBitSize > ElementBits)
4164 Cnt = SplatBits.getSExtValue();
4168 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4169 /// operand of a vector shift left operation. That value must be in the range:
4170 /// 0 <= Value < ElementBits for a left shift; or
4171 /// 0 <= Value <= ElementBits for a long left shift.
4172 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4173 assert(VT.isVector() && "vector shift count is not a vector type");
4174 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4175 if (! getVShiftImm(Op, ElementBits, Cnt))
4177 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4180 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4181 /// operand of a vector shift right operation. For a shift opcode, the value
4182 /// is positive, but for an intrinsic the value count must be negative. The
4183 /// absolute value must be in the range:
4184 /// 1 <= |Value| <= ElementBits for a right shift; or
4185 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4186 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4188 assert(VT.isVector() && "vector shift count is not a vector type");
4189 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4190 if (! getVShiftImm(Op, ElementBits, Cnt))
4194 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4197 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4198 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4199 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4202 // Don't do anything for most intrinsics.
4205 // Vector shifts: check for immediate versions and lower them.
4206 // Note: This is done during DAG combining instead of DAG legalizing because
4207 // the build_vectors for 64-bit vector element shift counts are generally
4208 // not legal, and it is hard to see their values after they get legalized to
4209 // loads from a constant pool.
4210 case Intrinsic::arm_neon_vshifts:
4211 case Intrinsic::arm_neon_vshiftu:
4212 case Intrinsic::arm_neon_vshiftls:
4213 case Intrinsic::arm_neon_vshiftlu:
4214 case Intrinsic::arm_neon_vshiftn:
4215 case Intrinsic::arm_neon_vrshifts:
4216 case Intrinsic::arm_neon_vrshiftu:
4217 case Intrinsic::arm_neon_vrshiftn:
4218 case Intrinsic::arm_neon_vqshifts:
4219 case Intrinsic::arm_neon_vqshiftu:
4220 case Intrinsic::arm_neon_vqshiftsu:
4221 case Intrinsic::arm_neon_vqshiftns:
4222 case Intrinsic::arm_neon_vqshiftnu:
4223 case Intrinsic::arm_neon_vqshiftnsu:
4224 case Intrinsic::arm_neon_vqrshiftns:
4225 case Intrinsic::arm_neon_vqrshiftnu:
4226 case Intrinsic::arm_neon_vqrshiftnsu: {
4227 EVT VT = N->getOperand(1).getValueType();
4229 unsigned VShiftOpc = 0;
4232 case Intrinsic::arm_neon_vshifts:
4233 case Intrinsic::arm_neon_vshiftu:
4234 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4235 VShiftOpc = ARMISD::VSHL;
4238 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4239 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4240 ARMISD::VSHRs : ARMISD::VSHRu);
4245 case Intrinsic::arm_neon_vshiftls:
4246 case Intrinsic::arm_neon_vshiftlu:
4247 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4249 llvm_unreachable("invalid shift count for vshll intrinsic");
4251 case Intrinsic::arm_neon_vrshifts:
4252 case Intrinsic::arm_neon_vrshiftu:
4253 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4257 case Intrinsic::arm_neon_vqshifts:
4258 case Intrinsic::arm_neon_vqshiftu:
4259 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4263 case Intrinsic::arm_neon_vqshiftsu:
4264 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4266 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4268 case Intrinsic::arm_neon_vshiftn:
4269 case Intrinsic::arm_neon_vrshiftn:
4270 case Intrinsic::arm_neon_vqshiftns:
4271 case Intrinsic::arm_neon_vqshiftnu:
4272 case Intrinsic::arm_neon_vqshiftnsu:
4273 case Intrinsic::arm_neon_vqrshiftns:
4274 case Intrinsic::arm_neon_vqrshiftnu:
4275 case Intrinsic::arm_neon_vqrshiftnsu:
4276 // Narrowing shifts require an immediate right shift.
4277 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4279 llvm_unreachable("invalid shift count for narrowing vector shift "
4283 llvm_unreachable("unhandled vector shift");
4287 case Intrinsic::arm_neon_vshifts:
4288 case Intrinsic::arm_neon_vshiftu:
4289 // Opcode already set above.
4291 case Intrinsic::arm_neon_vshiftls:
4292 case Intrinsic::arm_neon_vshiftlu:
4293 if (Cnt == VT.getVectorElementType().getSizeInBits())
4294 VShiftOpc = ARMISD::VSHLLi;
4296 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4297 ARMISD::VSHLLs : ARMISD::VSHLLu);
4299 case Intrinsic::arm_neon_vshiftn:
4300 VShiftOpc = ARMISD::VSHRN; break;
4301 case Intrinsic::arm_neon_vrshifts:
4302 VShiftOpc = ARMISD::VRSHRs; break;
4303 case Intrinsic::arm_neon_vrshiftu:
4304 VShiftOpc = ARMISD::VRSHRu; break;
4305 case Intrinsic::arm_neon_vrshiftn:
4306 VShiftOpc = ARMISD::VRSHRN; break;
4307 case Intrinsic::arm_neon_vqshifts:
4308 VShiftOpc = ARMISD::VQSHLs; break;
4309 case Intrinsic::arm_neon_vqshiftu:
4310 VShiftOpc = ARMISD::VQSHLu; break;
4311 case Intrinsic::arm_neon_vqshiftsu:
4312 VShiftOpc = ARMISD::VQSHLsu; break;
4313 case Intrinsic::arm_neon_vqshiftns:
4314 VShiftOpc = ARMISD::VQSHRNs; break;
4315 case Intrinsic::arm_neon_vqshiftnu:
4316 VShiftOpc = ARMISD::VQSHRNu; break;
4317 case Intrinsic::arm_neon_vqshiftnsu:
4318 VShiftOpc = ARMISD::VQSHRNsu; break;
4319 case Intrinsic::arm_neon_vqrshiftns:
4320 VShiftOpc = ARMISD::VQRSHRNs; break;
4321 case Intrinsic::arm_neon_vqrshiftnu:
4322 VShiftOpc = ARMISD::VQRSHRNu; break;
4323 case Intrinsic::arm_neon_vqrshiftnsu:
4324 VShiftOpc = ARMISD::VQRSHRNsu; break;
4327 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4328 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4331 case Intrinsic::arm_neon_vshiftins: {
4332 EVT VT = N->getOperand(1).getValueType();
4334 unsigned VShiftOpc = 0;
4336 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4337 VShiftOpc = ARMISD::VSLI;
4338 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4339 VShiftOpc = ARMISD::VSRI;
4341 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4344 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4345 N->getOperand(1), N->getOperand(2),
4346 DAG.getConstant(Cnt, MVT::i32));
4349 case Intrinsic::arm_neon_vqrshifts:
4350 case Intrinsic::arm_neon_vqrshiftu:
4351 // No immediate versions of these to check for.
4358 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4359 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4360 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4361 /// vector element shift counts are generally not legal, and it is hard to see
4362 /// their values after they get legalized to loads from a constant pool.
4363 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4364 const ARMSubtarget *ST) {
4365 EVT VT = N->getValueType(0);
4367 // Nothing to be done for scalar shifts.
4368 if (! VT.isVector())
4371 assert(ST->hasNEON() && "unexpected vector shift");
4374 switch (N->getOpcode()) {
4375 default: llvm_unreachable("unexpected shift opcode");
4378 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4379 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4380 DAG.getConstant(Cnt, MVT::i32));
4385 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4386 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4387 ARMISD::VSHRs : ARMISD::VSHRu);
4388 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4389 DAG.getConstant(Cnt, MVT::i32));
4395 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4396 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4397 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4398 const ARMSubtarget *ST) {
4399 SDValue N0 = N->getOperand(0);
4401 // Check for sign- and zero-extensions of vector extract operations of 8-
4402 // and 16-bit vector elements. NEON supports these directly. They are
4403 // handled during DAG combining because type legalization will promote them
4404 // to 32-bit types and it is messy to recognize the operations after that.
4405 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4406 SDValue Vec = N0.getOperand(0);
4407 SDValue Lane = N0.getOperand(1);
4408 EVT VT = N->getValueType(0);
4409 EVT EltVT = N0.getValueType();
4410 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4412 if (VT == MVT::i32 &&
4413 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4414 TLI.isTypeLegal(Vec.getValueType())) {
4417 switch (N->getOpcode()) {
4418 default: llvm_unreachable("unexpected opcode");
4419 case ISD::SIGN_EXTEND:
4420 Opc = ARMISD::VGETLANEs;
4422 case ISD::ZERO_EXTEND:
4423 case ISD::ANY_EXTEND:
4424 Opc = ARMISD::VGETLANEu;
4427 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4434 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4435 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4436 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4437 const ARMSubtarget *ST) {
4438 // If the target supports NEON, try to use vmax/vmin instructions for f32
4439 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4440 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4441 // a NaN; only do the transformation when it matches that behavior.
4443 // For now only do this when using NEON for FP operations; if using VFP, it
4444 // is not obvious that the benefit outweighs the cost of switching to the
4446 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4447 N->getValueType(0) != MVT::f32)
4450 SDValue CondLHS = N->getOperand(0);
4451 SDValue CondRHS = N->getOperand(1);
4452 SDValue LHS = N->getOperand(2);
4453 SDValue RHS = N->getOperand(3);
4454 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4456 unsigned Opcode = 0;
4458 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4459 IsReversed = false; // x CC y ? x : y
4460 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4461 IsReversed = true ; // x CC y ? y : x
4475 // If LHS is NaN, an ordered comparison will be false and the result will
4476 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4477 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4478 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4479 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4481 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4482 // will return -0, so vmin can only be used for unsafe math or if one of
4483 // the operands is known to be nonzero.
4484 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4486 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4488 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4497 // If LHS is NaN, an ordered comparison will be false and the result will
4498 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4499 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4500 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4501 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4503 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4504 // will return +0, so vmax can only be used for unsafe math or if one of
4505 // the operands is known to be nonzero.
4506 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4508 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4510 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4516 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4519 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4520 DAGCombinerInfo &DCI) const {
4521 switch (N->getOpcode()) {
4523 case ISD::ADD: return PerformADDCombine(N, DCI);
4524 case ISD::SUB: return PerformSUBCombine(N, DCI);
4525 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4526 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4527 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4530 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4531 case ISD::SIGN_EXTEND:
4532 case ISD::ZERO_EXTEND:
4533 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4534 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4539 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4540 if (!Subtarget->hasV6Ops())
4541 // Pre-v6 does not support unaligned mem access.
4544 // v6+ may or may not support unaligned mem access depending on the system
4546 // FIXME: This is pretty conservative. Should we provide cmdline option to
4547 // control the behaviour?
4548 if (!Subtarget->isTargetDarwin())
4552 switch (VT.getSimpleVT().SimpleTy) {
4559 // FIXME: VLD1 etc with standard alignment is legal.
4563 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4568 switch (VT.getSimpleVT().SimpleTy) {
4569 default: return false;
4584 if ((V & (Scale - 1)) != 0)
4587 return V == (V & ((1LL << 5) - 1));
4590 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4591 const ARMSubtarget *Subtarget) {
4598 switch (VT.getSimpleVT().SimpleTy) {
4599 default: return false;
4604 // + imm12 or - imm8
4606 return V == (V & ((1LL << 8) - 1));
4607 return V == (V & ((1LL << 12) - 1));
4610 // Same as ARM mode. FIXME: NEON?
4611 if (!Subtarget->hasVFP2())
4616 return V == (V & ((1LL << 8) - 1));
4620 /// isLegalAddressImmediate - Return true if the integer value can be used
4621 /// as the offset of the target addressing mode for load / store of the
4623 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4624 const ARMSubtarget *Subtarget) {
4631 if (Subtarget->isThumb1Only())
4632 return isLegalT1AddressImmediate(V, VT);
4633 else if (Subtarget->isThumb2())
4634 return isLegalT2AddressImmediate(V, VT, Subtarget);
4639 switch (VT.getSimpleVT().SimpleTy) {
4640 default: return false;
4645 return V == (V & ((1LL << 12) - 1));
4648 return V == (V & ((1LL << 8) - 1));
4651 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4656 return V == (V & ((1LL << 8) - 1));
4660 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4662 int Scale = AM.Scale;
4666 switch (VT.getSimpleVT().SimpleTy) {
4667 default: return false;
4676 return Scale == 2 || Scale == 4 || Scale == 8;
4679 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4683 // Note, we allow "void" uses (basically, uses that aren't loads or
4684 // stores), because arm allows folding a scale into many arithmetic
4685 // operations. This should be made more precise and revisited later.
4687 // Allow r << imm, but the imm has to be a multiple of two.
4688 if (Scale & 1) return false;
4689 return isPowerOf2_32(Scale);
4693 /// isLegalAddressingMode - Return true if the addressing mode represented
4694 /// by AM is legal for this target, for a load/store of the specified type.
4695 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4696 const Type *Ty) const {
4697 EVT VT = getValueType(Ty, true);
4698 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4701 // Can never fold addr of global into load/store.
4706 case 0: // no scale reg, must be "r+i" or "r", or "i".
4709 if (Subtarget->isThumb1Only())
4713 // ARM doesn't support any R+R*scale+imm addr modes.
4720 if (Subtarget->isThumb2())
4721 return isLegalT2ScaledAddressingMode(AM, VT);
4723 int Scale = AM.Scale;
4724 switch (VT.getSimpleVT().SimpleTy) {
4725 default: return false;
4729 if (Scale < 0) Scale = -Scale;
4733 return isPowerOf2_32(Scale & ~1);
4737 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4742 // Note, we allow "void" uses (basically, uses that aren't loads or
4743 // stores), because arm allows folding a scale into many arithmetic
4744 // operations. This should be made more precise and revisited later.
4746 // Allow r << imm, but the imm has to be a multiple of two.
4747 if (Scale & 1) return false;
4748 return isPowerOf2_32(Scale);
4755 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4756 /// icmp immediate, that is the target has icmp instructions which can compare
4757 /// a register against the immediate without having to materialize the
4758 /// immediate into a register.
4759 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4760 if (!Subtarget->isThumb())
4761 return ARM_AM::getSOImmVal(Imm) != -1;
4762 if (Subtarget->isThumb2())
4763 return ARM_AM::getT2SOImmVal(Imm) != -1;
4764 return Imm >= 0 && Imm <= 255;
4767 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4768 bool isSEXTLoad, SDValue &Base,
4769 SDValue &Offset, bool &isInc,
4770 SelectionDAG &DAG) {
4771 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4774 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4776 Base = Ptr->getOperand(0);
4777 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4778 int RHSC = (int)RHS->getZExtValue();
4779 if (RHSC < 0 && RHSC > -256) {
4780 assert(Ptr->getOpcode() == ISD::ADD);
4782 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4786 isInc = (Ptr->getOpcode() == ISD::ADD);
4787 Offset = Ptr->getOperand(1);
4789 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4791 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4792 int RHSC = (int)RHS->getZExtValue();
4793 if (RHSC < 0 && RHSC > -0x1000) {
4794 assert(Ptr->getOpcode() == ISD::ADD);
4796 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4797 Base = Ptr->getOperand(0);
4802 if (Ptr->getOpcode() == ISD::ADD) {
4804 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4805 if (ShOpcVal != ARM_AM::no_shift) {
4806 Base = Ptr->getOperand(1);
4807 Offset = Ptr->getOperand(0);
4809 Base = Ptr->getOperand(0);
4810 Offset = Ptr->getOperand(1);
4815 isInc = (Ptr->getOpcode() == ISD::ADD);
4816 Base = Ptr->getOperand(0);
4817 Offset = Ptr->getOperand(1);
4821 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4825 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4826 bool isSEXTLoad, SDValue &Base,
4827 SDValue &Offset, bool &isInc,
4828 SelectionDAG &DAG) {
4829 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4832 Base = Ptr->getOperand(0);
4833 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4834 int RHSC = (int)RHS->getZExtValue();
4835 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4836 assert(Ptr->getOpcode() == ISD::ADD);
4838 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4840 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4841 isInc = Ptr->getOpcode() == ISD::ADD;
4842 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4850 /// getPreIndexedAddressParts - returns true by value, base pointer and
4851 /// offset pointer and addressing mode by reference if the node's address
4852 /// can be legally represented as pre-indexed load / store address.
4854 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4856 ISD::MemIndexedMode &AM,
4857 SelectionDAG &DAG) const {
4858 if (Subtarget->isThumb1Only())
4863 bool isSEXTLoad = false;
4864 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4865 Ptr = LD->getBasePtr();
4866 VT = LD->getMemoryVT();
4867 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4868 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4869 Ptr = ST->getBasePtr();
4870 VT = ST->getMemoryVT();
4875 bool isLegal = false;
4876 if (Subtarget->isThumb2())
4877 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4878 Offset, isInc, DAG);
4880 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4881 Offset, isInc, DAG);
4885 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4889 /// getPostIndexedAddressParts - returns true by value, base pointer and
4890 /// offset pointer and addressing mode by reference if this node can be
4891 /// combined with a load / store to form a post-indexed load / store.
4892 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4895 ISD::MemIndexedMode &AM,
4896 SelectionDAG &DAG) const {
4897 if (Subtarget->isThumb1Only())
4902 bool isSEXTLoad = false;
4903 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4904 VT = LD->getMemoryVT();
4905 Ptr = LD->getBasePtr();
4906 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4907 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4908 VT = ST->getMemoryVT();
4909 Ptr = ST->getBasePtr();
4914 bool isLegal = false;
4915 if (Subtarget->isThumb2())
4916 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4919 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4925 // Swap base ptr and offset to catch more post-index load / store when
4926 // it's legal. In Thumb2 mode, offset must be an immediate.
4927 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4928 !Subtarget->isThumb2())
4929 std::swap(Base, Offset);
4931 // Post-indexed load / store update the base pointer.
4936 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4940 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4944 const SelectionDAG &DAG,
4945 unsigned Depth) const {
4946 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4947 switch (Op.getOpcode()) {
4949 case ARMISD::CMOV: {
4950 // Bits are known zero/one if known on the LHS and RHS.
4951 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4952 if (KnownZero == 0 && KnownOne == 0) return;
4954 APInt KnownZeroRHS, KnownOneRHS;
4955 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4956 KnownZeroRHS, KnownOneRHS, Depth+1);
4957 KnownZero &= KnownZeroRHS;
4958 KnownOne &= KnownOneRHS;
4964 //===----------------------------------------------------------------------===//
4965 // ARM Inline Assembly Support
4966 //===----------------------------------------------------------------------===//
4968 /// getConstraintType - Given a constraint letter, return the type of
4969 /// constraint it is for this target.
4970 ARMTargetLowering::ConstraintType
4971 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4972 if (Constraint.size() == 1) {
4973 switch (Constraint[0]) {
4975 case 'l': return C_RegisterClass;
4976 case 'w': return C_RegisterClass;
4979 return TargetLowering::getConstraintType(Constraint);
4982 std::pair<unsigned, const TargetRegisterClass*>
4983 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4985 if (Constraint.size() == 1) {
4986 // GCC ARM Constraint Letters
4987 switch (Constraint[0]) {
4989 if (Subtarget->isThumb())
4990 return std::make_pair(0U, ARM::tGPRRegisterClass);
4992 return std::make_pair(0U, ARM::GPRRegisterClass);
4994 return std::make_pair(0U, ARM::GPRRegisterClass);
4997 return std::make_pair(0U, ARM::SPRRegisterClass);
4998 if (VT.getSizeInBits() == 64)
4999 return std::make_pair(0U, ARM::DPRRegisterClass);
5000 if (VT.getSizeInBits() == 128)
5001 return std::make_pair(0U, ARM::QPRRegisterClass);
5005 if (StringRef("{cc}").equals_lower(Constraint))
5006 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5008 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5011 std::vector<unsigned> ARMTargetLowering::
5012 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5014 if (Constraint.size() != 1)
5015 return std::vector<unsigned>();
5017 switch (Constraint[0]) { // GCC ARM Constraint Letters
5020 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5021 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5024 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5025 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5026 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5027 ARM::R12, ARM::LR, 0);
5030 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5031 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5032 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5033 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5034 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5035 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5036 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5037 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5038 if (VT.getSizeInBits() == 64)
5039 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5040 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5041 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5042 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5043 if (VT.getSizeInBits() == 128)
5044 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5045 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5049 return std::vector<unsigned>();
5052 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5053 /// vector. If it is invalid, don't add anything to Ops.
5054 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5057 std::vector<SDValue>&Ops,
5058 SelectionDAG &DAG) const {
5059 SDValue Result(0, 0);
5061 switch (Constraint) {
5063 case 'I': case 'J': case 'K': case 'L':
5064 case 'M': case 'N': case 'O':
5065 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5069 int64_t CVal64 = C->getSExtValue();
5070 int CVal = (int) CVal64;
5071 // None of these constraints allow values larger than 32 bits. Check
5072 // that the value fits in an int.
5076 switch (Constraint) {
5078 if (Subtarget->isThumb1Only()) {
5079 // This must be a constant between 0 and 255, for ADD
5081 if (CVal >= 0 && CVal <= 255)
5083 } else if (Subtarget->isThumb2()) {
5084 // A constant that can be used as an immediate value in a
5085 // data-processing instruction.
5086 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5089 // A constant that can be used as an immediate value in a
5090 // data-processing instruction.
5091 if (ARM_AM::getSOImmVal(CVal) != -1)
5097 if (Subtarget->isThumb()) { // FIXME thumb2
5098 // This must be a constant between -255 and -1, for negated ADD
5099 // immediates. This can be used in GCC with an "n" modifier that
5100 // prints the negated value, for use with SUB instructions. It is
5101 // not useful otherwise but is implemented for compatibility.
5102 if (CVal >= -255 && CVal <= -1)
5105 // This must be a constant between -4095 and 4095. It is not clear
5106 // what this constraint is intended for. Implemented for
5107 // compatibility with GCC.
5108 if (CVal >= -4095 && CVal <= 4095)
5114 if (Subtarget->isThumb1Only()) {
5115 // A 32-bit value where only one byte has a nonzero value. Exclude
5116 // zero to match GCC. This constraint is used by GCC internally for
5117 // constants that can be loaded with a move/shift combination.
5118 // It is not useful otherwise but is implemented for compatibility.
5119 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5121 } else if (Subtarget->isThumb2()) {
5122 // A constant whose bitwise inverse can be used as an immediate
5123 // value in a data-processing instruction. This can be used in GCC
5124 // with a "B" modifier that prints the inverted value, for use with
5125 // BIC and MVN instructions. It is not useful otherwise but is
5126 // implemented for compatibility.
5127 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5130 // A constant whose bitwise inverse can be used as an immediate
5131 // value in a data-processing instruction. This can be used in GCC
5132 // with a "B" modifier that prints the inverted value, for use with
5133 // BIC and MVN instructions. It is not useful otherwise but is
5134 // implemented for compatibility.
5135 if (ARM_AM::getSOImmVal(~CVal) != -1)
5141 if (Subtarget->isThumb1Only()) {
5142 // This must be a constant between -7 and 7,
5143 // for 3-operand ADD/SUB immediate instructions.
5144 if (CVal >= -7 && CVal < 7)
5146 } else if (Subtarget->isThumb2()) {
5147 // A constant whose negation can be used as an immediate value in a
5148 // data-processing instruction. This can be used in GCC with an "n"
5149 // modifier that prints the negated value, for use with SUB
5150 // instructions. It is not useful otherwise but is implemented for
5152 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5155 // A constant whose negation can be used as an immediate value in a
5156 // data-processing instruction. This can be used in GCC with an "n"
5157 // modifier that prints the negated value, for use with SUB
5158 // instructions. It is not useful otherwise but is implemented for
5160 if (ARM_AM::getSOImmVal(-CVal) != -1)
5166 if (Subtarget->isThumb()) { // FIXME thumb2
5167 // This must be a multiple of 4 between 0 and 1020, for
5168 // ADD sp + immediate.
5169 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5172 // A power of two or a constant between 0 and 32. This is used in
5173 // GCC for the shift amount on shifted register operands, but it is
5174 // useful in general for any shift amounts.
5175 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5181 if (Subtarget->isThumb()) { // FIXME thumb2
5182 // This must be a constant between 0 and 31, for shift amounts.
5183 if (CVal >= 0 && CVal <= 31)
5189 if (Subtarget->isThumb()) { // FIXME thumb2
5190 // This must be a multiple of 4 between -508 and 508, for
5191 // ADD/SUB sp = sp + immediate.
5192 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5197 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5201 if (Result.getNode()) {
5202 Ops.push_back(Result);
5205 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
5210 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5211 // The ARM target isn't yet aware of offsets.
5215 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5216 APInt Imm = FPImm.bitcastToAPInt();
5217 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5218 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5219 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5221 // We can handle 4 bits of mantissa.
5222 // mantissa = (16+UInt(e:f:g:h))/16.
5223 if (Mantissa & 0x7ffff)
5226 if ((Mantissa & 0xf) != Mantissa)
5229 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5230 if (Exp < -3 || Exp > 4)
5232 Exp = ((Exp+3) & 0x7) ^ 4;
5234 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5237 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5238 APInt Imm = FPImm.bitcastToAPInt();
5239 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5240 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5241 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5243 // We can handle 4 bits of mantissa.
5244 // mantissa = (16+UInt(e:f:g:h))/16.
5245 if (Mantissa & 0xffffffffffffLL)
5248 if ((Mantissa & 0xf) != Mantissa)
5251 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5252 if (Exp < -3 || Exp > 4)
5254 Exp = ((Exp+3) & 0x7) ^ 4;
5256 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5259 /// isFPImmLegal - Returns true if the target can instruction select the
5260 /// specified FP immediate natively. If false, the legalizer will
5261 /// materialize the FP immediate as a load from a constant pool.
5262 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5263 if (!Subtarget->hasVFP3())
5266 return ARM::getVFPf32Imm(Imm) != -1;
5268 return ARM::getVFPf64Imm(Imm) != -1;