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 sibcall for Thumb1 since Thumb1RegisterInfo::
1412 // emitEpilogue is not ready for them.
1413 if (Subtarget->isThumb1Only())
1416 // For the moment, we can only do this to functions defined in this
1417 // compilation, or to indirect calls. A Thumb B to an ARM function,
1418 // or vice versa, is not easily fixed up in the linker unlike BL.
1419 // (We could do this by loading the address of the callee into a register;
1420 // that is an extra instruction over the direct call and burns a register
1421 // as well, so is not likely to be a win.)
1422 if (isa<ExternalSymbolSDNode>(Callee))
1425 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1426 const GlobalValue *GV = G->getGlobal();
1427 if (GV->isDeclaration() || GV->isWeakForLinker())
1431 // If the calling conventions do not match, then we'd better make sure the
1432 // results are returned in the same way as what the caller expects.
1434 SmallVector<CCValAssign, 16> RVLocs1;
1435 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1436 RVLocs1, *DAG.getContext());
1437 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1439 SmallVector<CCValAssign, 16> RVLocs2;
1440 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1441 RVLocs2, *DAG.getContext());
1442 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1444 if (RVLocs1.size() != RVLocs2.size())
1446 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1447 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1449 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1451 if (RVLocs1[i].isRegLoc()) {
1452 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1455 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1461 // If the callee takes no arguments then go on to check the results of the
1463 if (!Outs.empty()) {
1464 // Check if stack adjustment is needed. For now, do not do this if any
1465 // argument is passed on the stack.
1466 SmallVector<CCValAssign, 16> ArgLocs;
1467 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1468 ArgLocs, *DAG.getContext());
1469 CCInfo.AnalyzeCallOperands(Outs,
1470 CCAssignFnForNode(CalleeCC, false, isVarArg));
1471 if (CCInfo.getNextStackOffset()) {
1472 MachineFunction &MF = DAG.getMachineFunction();
1474 // Check if the arguments are already laid out in the right way as
1475 // the caller's fixed stack objects.
1476 MachineFrameInfo *MFI = MF.getFrameInfo();
1477 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1478 const ARMInstrInfo *TII =
1479 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1480 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1482 ++i, ++realArgIdx) {
1483 CCValAssign &VA = ArgLocs[i];
1484 EVT RegVT = VA.getLocVT();
1485 SDValue Arg = Outs[realArgIdx].Val;
1486 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1487 if (VA.getLocInfo() == CCValAssign::Indirect)
1489 if (VA.needsCustom()) {
1490 // f64 and vector types are split into multiple registers or
1491 // register/stack-slot combinations. The types will not match
1492 // the registers; give up on memory f64 refs until we figure
1493 // out what to do about this.
1496 if (!ArgLocs[++i].isRegLoc())
1498 if (RegVT == MVT::v2f64) {
1499 if (!ArgLocs[++i].isRegLoc())
1501 if (!ArgLocs[++i].isRegLoc())
1504 } else if (!VA.isRegLoc()) {
1505 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1517 ARMTargetLowering::LowerReturn(SDValue Chain,
1518 CallingConv::ID CallConv, bool isVarArg,
1519 const SmallVectorImpl<ISD::OutputArg> &Outs,
1520 DebugLoc dl, SelectionDAG &DAG) const {
1522 // CCValAssign - represent the assignment of the return value to a location.
1523 SmallVector<CCValAssign, 16> RVLocs;
1525 // CCState - Info about the registers and stack slots.
1526 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1529 // Analyze outgoing return values.
1530 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1533 // If this is the first return lowered for this function, add
1534 // the regs to the liveout set for the function.
1535 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1536 for (unsigned i = 0; i != RVLocs.size(); ++i)
1537 if (RVLocs[i].isRegLoc())
1538 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1543 // Copy the result values into the output registers.
1544 for (unsigned i = 0, realRVLocIdx = 0;
1546 ++i, ++realRVLocIdx) {
1547 CCValAssign &VA = RVLocs[i];
1548 assert(VA.isRegLoc() && "Can only return in registers!");
1550 SDValue Arg = Outs[realRVLocIdx].Val;
1552 switch (VA.getLocInfo()) {
1553 default: llvm_unreachable("Unknown loc info!");
1554 case CCValAssign::Full: break;
1555 case CCValAssign::BCvt:
1556 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1560 if (VA.needsCustom()) {
1561 if (VA.getLocVT() == MVT::v2f64) {
1562 // Extract the first half and return it in two registers.
1563 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1564 DAG.getConstant(0, MVT::i32));
1565 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1566 DAG.getVTList(MVT::i32, MVT::i32), Half);
1568 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1569 Flag = Chain.getValue(1);
1570 VA = RVLocs[++i]; // skip ahead to next loc
1571 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1572 HalfGPRs.getValue(1), Flag);
1573 Flag = Chain.getValue(1);
1574 VA = RVLocs[++i]; // skip ahead to next loc
1576 // Extract the 2nd half and fall through to handle it as an f64 value.
1577 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1578 DAG.getConstant(1, MVT::i32));
1580 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1582 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1583 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1584 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1585 Flag = Chain.getValue(1);
1586 VA = RVLocs[++i]; // skip ahead to next loc
1587 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1590 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1592 // Guarantee that all emitted copies are
1593 // stuck together, avoiding something bad.
1594 Flag = Chain.getValue(1);
1599 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1601 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1606 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1607 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1608 // one of the above mentioned nodes. It has to be wrapped because otherwise
1609 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1610 // be used to form addressing mode. These wrapped nodes will be selected
1612 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1613 EVT PtrVT = Op.getValueType();
1614 // FIXME there is no actual debug info here
1615 DebugLoc dl = Op.getDebugLoc();
1616 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1618 if (CP->isMachineConstantPoolEntry())
1619 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1620 CP->getAlignment());
1622 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1623 CP->getAlignment());
1624 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1627 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1628 SelectionDAG &DAG) const {
1629 MachineFunction &MF = DAG.getMachineFunction();
1630 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1631 unsigned ARMPCLabelIndex = 0;
1632 DebugLoc DL = Op.getDebugLoc();
1633 EVT PtrVT = getPointerTy();
1634 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1635 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1637 if (RelocM == Reloc::Static) {
1638 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1640 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1641 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1642 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1643 ARMCP::CPBlockAddress,
1645 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1647 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1648 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1649 PseudoSourceValue::getConstantPool(), 0,
1651 if (RelocM == Reloc::Static)
1653 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1654 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1657 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1659 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1660 SelectionDAG &DAG) const {
1661 DebugLoc dl = GA->getDebugLoc();
1662 EVT PtrVT = getPointerTy();
1663 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1664 MachineFunction &MF = DAG.getMachineFunction();
1665 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1666 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1667 ARMConstantPoolValue *CPV =
1668 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1669 ARMCP::CPValue, PCAdj, "tlsgd", true);
1670 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1671 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1672 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1673 PseudoSourceValue::getConstantPool(), 0,
1675 SDValue Chain = Argument.getValue(1);
1677 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1678 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1680 // call __tls_get_addr.
1683 Entry.Node = Argument;
1684 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1685 Args.push_back(Entry);
1686 // FIXME: is there useful debug info available here?
1687 std::pair<SDValue, SDValue> CallResult =
1688 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1689 false, false, false, false,
1690 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1691 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1692 return CallResult.first;
1695 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1696 // "local exec" model.
1698 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1699 SelectionDAG &DAG) const {
1700 const GlobalValue *GV = GA->getGlobal();
1701 DebugLoc dl = GA->getDebugLoc();
1703 SDValue Chain = DAG.getEntryNode();
1704 EVT PtrVT = getPointerTy();
1705 // Get the Thread Pointer
1706 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1708 if (GV->isDeclaration()) {
1709 MachineFunction &MF = DAG.getMachineFunction();
1710 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1711 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1712 // Initial exec model.
1713 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1714 ARMConstantPoolValue *CPV =
1715 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1716 ARMCP::CPValue, PCAdj, "gottpoff", true);
1717 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1718 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1719 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1720 PseudoSourceValue::getConstantPool(), 0,
1722 Chain = Offset.getValue(1);
1724 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1725 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1727 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1728 PseudoSourceValue::getConstantPool(), 0,
1732 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1733 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1734 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1735 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1736 PseudoSourceValue::getConstantPool(), 0,
1740 // The address of the thread local variable is the add of the thread
1741 // pointer with the offset of the variable.
1742 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1746 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1747 // TODO: implement the "local dynamic" model
1748 assert(Subtarget->isTargetELF() &&
1749 "TLS not implemented for non-ELF targets");
1750 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1751 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1752 // otherwise use the "Local Exec" TLS Model
1753 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1754 return LowerToTLSGeneralDynamicModel(GA, DAG);
1756 return LowerToTLSExecModels(GA, DAG);
1759 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1760 SelectionDAG &DAG) const {
1761 EVT PtrVT = getPointerTy();
1762 DebugLoc dl = Op.getDebugLoc();
1763 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1764 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1765 if (RelocM == Reloc::PIC_) {
1766 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1767 ARMConstantPoolValue *CPV =
1768 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1769 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1770 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1771 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1773 PseudoSourceValue::getConstantPool(), 0,
1775 SDValue Chain = Result.getValue(1);
1776 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1777 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1779 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1780 PseudoSourceValue::getGOT(), 0,
1784 // If we have T2 ops, we can materialize the address directly via movt/movw
1785 // pair. This is always cheaper.
1786 if (Subtarget->useMovt()) {
1787 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1788 DAG.getTargetGlobalAddress(GV, PtrVT));
1790 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1791 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1792 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1793 PseudoSourceValue::getConstantPool(), 0,
1799 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1800 SelectionDAG &DAG) const {
1801 MachineFunction &MF = DAG.getMachineFunction();
1802 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1803 unsigned ARMPCLabelIndex = 0;
1804 EVT PtrVT = getPointerTy();
1805 DebugLoc dl = Op.getDebugLoc();
1806 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1807 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1809 if (RelocM == Reloc::Static)
1810 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1812 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1813 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1814 ARMConstantPoolValue *CPV =
1815 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1816 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1818 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1820 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1821 PseudoSourceValue::getConstantPool(), 0,
1823 SDValue Chain = Result.getValue(1);
1825 if (RelocM == Reloc::PIC_) {
1826 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1827 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1830 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1831 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1832 PseudoSourceValue::getGOT(), 0,
1838 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1839 SelectionDAG &DAG) const {
1840 assert(Subtarget->isTargetELF() &&
1841 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1842 MachineFunction &MF = DAG.getMachineFunction();
1843 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1844 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1845 EVT PtrVT = getPointerTy();
1846 DebugLoc dl = Op.getDebugLoc();
1847 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1848 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1849 "_GLOBAL_OFFSET_TABLE_",
1850 ARMPCLabelIndex, PCAdj);
1851 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1852 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1853 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1854 PseudoSourceValue::getConstantPool(), 0,
1856 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1857 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1861 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1862 DebugLoc dl = Op.getDebugLoc();
1863 SDValue Val = DAG.getConstant(0, MVT::i32);
1864 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1865 Op.getOperand(1), Val);
1869 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1870 DebugLoc dl = Op.getDebugLoc();
1871 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1872 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1876 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1877 const ARMSubtarget *Subtarget) const {
1878 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1879 DebugLoc dl = Op.getDebugLoc();
1881 default: return SDValue(); // Don't custom lower most intrinsics.
1882 case Intrinsic::arm_thread_pointer: {
1883 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1884 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1886 case Intrinsic::eh_sjlj_lsda: {
1887 MachineFunction &MF = DAG.getMachineFunction();
1888 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1889 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1890 EVT PtrVT = getPointerTy();
1891 DebugLoc dl = Op.getDebugLoc();
1892 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1894 unsigned PCAdj = (RelocM != Reloc::PIC_)
1895 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1896 ARMConstantPoolValue *CPV =
1897 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1898 ARMCP::CPLSDA, PCAdj);
1899 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1900 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1902 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1903 PseudoSourceValue::getConstantPool(), 0,
1905 SDValue Chain = Result.getValue(1);
1907 if (RelocM == Reloc::PIC_) {
1908 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1909 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1916 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1917 const ARMSubtarget *Subtarget) {
1918 DebugLoc dl = Op.getDebugLoc();
1919 SDValue Op5 = Op.getOperand(5);
1920 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1921 // v6 and v7 can both handle barriers directly, but need handled a bit
1922 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1924 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1925 if (Subtarget->hasV7Ops())
1926 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1927 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1928 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1929 DAG.getConstant(0, MVT::i32));
1930 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1934 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1935 MachineFunction &MF = DAG.getMachineFunction();
1936 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1938 // vastart just stores the address of the VarArgsFrameIndex slot into the
1939 // memory location argument.
1940 DebugLoc dl = Op.getDebugLoc();
1941 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1942 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1943 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1944 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1949 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1950 SelectionDAG &DAG) const {
1951 SDNode *Node = Op.getNode();
1952 DebugLoc dl = Node->getDebugLoc();
1953 EVT VT = Node->getValueType(0);
1954 SDValue Chain = Op.getOperand(0);
1955 SDValue Size = Op.getOperand(1);
1956 SDValue Align = Op.getOperand(2);
1958 // Chain the dynamic stack allocation so that it doesn't modify the stack
1959 // pointer when other instructions are using the stack.
1960 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1962 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1963 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1964 if (AlignVal > StackAlign)
1965 // Do this now since selection pass cannot introduce new target
1966 // independent node.
1967 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1969 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1970 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1971 // do even more horrible hack later.
1972 MachineFunction &MF = DAG.getMachineFunction();
1973 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1974 if (AFI->isThumb1OnlyFunction()) {
1976 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1978 uint32_t Val = C->getZExtValue();
1979 if (Val <= 508 && ((Val & 3) == 0))
1983 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1986 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1987 SDValue Ops1[] = { Chain, Size, Align };
1988 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1989 Chain = Res.getValue(1);
1990 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1991 DAG.getIntPtrConstant(0, true), SDValue());
1992 SDValue Ops2[] = { Res, Chain };
1993 return DAG.getMergeValues(Ops2, 2, dl);
1997 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1998 SDValue &Root, SelectionDAG &DAG,
1999 DebugLoc dl) const {
2000 MachineFunction &MF = DAG.getMachineFunction();
2001 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2003 TargetRegisterClass *RC;
2004 if (AFI->isThumb1OnlyFunction())
2005 RC = ARM::tGPRRegisterClass;
2007 RC = ARM::GPRRegisterClass;
2009 // Transform the arguments stored in physical registers into virtual ones.
2010 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2011 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2014 if (NextVA.isMemLoc()) {
2015 MachineFrameInfo *MFI = MF.getFrameInfo();
2016 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
2018 // Create load node to retrieve arguments from the stack.
2019 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2020 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2021 PseudoSourceValue::getFixedStack(FI), 0,
2024 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2025 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2028 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2032 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2033 CallingConv::ID CallConv, bool isVarArg,
2034 const SmallVectorImpl<ISD::InputArg>
2036 DebugLoc dl, SelectionDAG &DAG,
2037 SmallVectorImpl<SDValue> &InVals)
2040 MachineFunction &MF = DAG.getMachineFunction();
2041 MachineFrameInfo *MFI = MF.getFrameInfo();
2043 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2045 // Assign locations to all of the incoming arguments.
2046 SmallVector<CCValAssign, 16> ArgLocs;
2047 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2049 CCInfo.AnalyzeFormalArguments(Ins,
2050 CCAssignFnForNode(CallConv, /* Return*/ false,
2053 SmallVector<SDValue, 16> ArgValues;
2055 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2056 CCValAssign &VA = ArgLocs[i];
2058 // Arguments stored in registers.
2059 if (VA.isRegLoc()) {
2060 EVT RegVT = VA.getLocVT();
2063 if (VA.needsCustom()) {
2064 // f64 and vector types are split up into multiple registers or
2065 // combinations of registers and stack slots.
2066 if (VA.getLocVT() == MVT::v2f64) {
2067 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2069 VA = ArgLocs[++i]; // skip ahead to next loc
2071 if (VA.isMemLoc()) {
2072 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
2074 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2075 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2076 PseudoSourceValue::getFixedStack(FI), 0,
2079 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2082 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2083 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2084 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2085 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2086 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2088 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2091 TargetRegisterClass *RC;
2093 if (RegVT == MVT::f32)
2094 RC = ARM::SPRRegisterClass;
2095 else if (RegVT == MVT::f64)
2096 RC = ARM::DPRRegisterClass;
2097 else if (RegVT == MVT::v2f64)
2098 RC = ARM::QPRRegisterClass;
2099 else if (RegVT == MVT::i32)
2100 RC = (AFI->isThumb1OnlyFunction() ?
2101 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2103 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2105 // Transform the arguments in physical registers into virtual ones.
2106 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2107 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2110 // If this is an 8 or 16-bit value, it is really passed promoted
2111 // to 32 bits. Insert an assert[sz]ext to capture this, then
2112 // truncate to the right size.
2113 switch (VA.getLocInfo()) {
2114 default: llvm_unreachable("Unknown loc info!");
2115 case CCValAssign::Full: break;
2116 case CCValAssign::BCvt:
2117 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2119 case CCValAssign::SExt:
2120 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2121 DAG.getValueType(VA.getValVT()));
2122 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2124 case CCValAssign::ZExt:
2125 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2126 DAG.getValueType(VA.getValVT()));
2127 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2131 InVals.push_back(ArgValue);
2133 } else { // VA.isRegLoc()
2136 assert(VA.isMemLoc());
2137 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2139 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2140 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2143 // Create load nodes to retrieve arguments from the stack.
2144 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2145 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2146 PseudoSourceValue::getFixedStack(FI), 0,
2153 static const unsigned GPRArgRegs[] = {
2154 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2157 unsigned NumGPRs = CCInfo.getFirstUnallocated
2158 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2160 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2161 unsigned VARegSize = (4 - NumGPRs) * 4;
2162 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2163 unsigned ArgOffset = CCInfo.getNextStackOffset();
2164 if (VARegSaveSize) {
2165 // If this function is vararg, store any remaining integer argument regs
2166 // to their spots on the stack so that they may be loaded by deferencing
2167 // the result of va_next.
2168 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2169 AFI->setVarArgsFrameIndex(
2170 MFI->CreateFixedObject(VARegSaveSize,
2171 ArgOffset + VARegSaveSize - VARegSize,
2173 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2176 SmallVector<SDValue, 4> MemOps;
2177 for (; NumGPRs < 4; ++NumGPRs) {
2178 TargetRegisterClass *RC;
2179 if (AFI->isThumb1OnlyFunction())
2180 RC = ARM::tGPRRegisterClass;
2182 RC = ARM::GPRRegisterClass;
2184 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2185 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2187 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2188 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2189 0, false, false, 0);
2190 MemOps.push_back(Store);
2191 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2192 DAG.getConstant(4, getPointerTy()));
2194 if (!MemOps.empty())
2195 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2196 &MemOps[0], MemOps.size());
2198 // This will point to the next argument passed via stack.
2199 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
2206 /// isFloatingPointZero - Return true if this is +0.0.
2207 static bool isFloatingPointZero(SDValue Op) {
2208 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2209 return CFP->getValueAPF().isPosZero();
2210 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2211 // Maybe this has already been legalized into the constant pool?
2212 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2213 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2214 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2215 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2216 return CFP->getValueAPF().isPosZero();
2222 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2223 /// the given operands.
2225 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2226 SDValue &ARMCC, SelectionDAG &DAG,
2227 DebugLoc dl) const {
2228 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2229 unsigned C = RHSC->getZExtValue();
2230 if (!isLegalICmpImmediate(C)) {
2231 // Constant does not fit, try adjusting it by one?
2236 if (isLegalICmpImmediate(C-1)) {
2237 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2238 RHS = DAG.getConstant(C-1, MVT::i32);
2243 if (C > 0 && isLegalICmpImmediate(C-1)) {
2244 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2245 RHS = DAG.getConstant(C-1, MVT::i32);
2250 if (isLegalICmpImmediate(C+1)) {
2251 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2252 RHS = DAG.getConstant(C+1, MVT::i32);
2257 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2258 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2259 RHS = DAG.getConstant(C+1, MVT::i32);
2266 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2267 ARMISD::NodeType CompareType;
2270 CompareType = ARMISD::CMP;
2275 CompareType = ARMISD::CMPZ;
2278 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2279 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2282 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2283 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2286 if (!isFloatingPointZero(RHS))
2287 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2289 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2290 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2293 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2294 EVT VT = Op.getValueType();
2295 SDValue LHS = Op.getOperand(0);
2296 SDValue RHS = Op.getOperand(1);
2297 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2298 SDValue TrueVal = Op.getOperand(2);
2299 SDValue FalseVal = Op.getOperand(3);
2300 DebugLoc dl = Op.getDebugLoc();
2302 if (LHS.getValueType() == MVT::i32) {
2304 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2305 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2306 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2309 ARMCC::CondCodes CondCode, CondCode2;
2310 FPCCToARMCC(CC, CondCode, CondCode2);
2312 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2313 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2314 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2315 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2317 if (CondCode2 != ARMCC::AL) {
2318 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2319 // FIXME: Needs another CMP because flag can have but one use.
2320 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2321 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2322 Result, TrueVal, ARMCC2, CCR, Cmp2);
2327 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2328 SDValue Chain = Op.getOperand(0);
2329 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2330 SDValue LHS = Op.getOperand(2);
2331 SDValue RHS = Op.getOperand(3);
2332 SDValue Dest = Op.getOperand(4);
2333 DebugLoc dl = Op.getDebugLoc();
2335 if (LHS.getValueType() == MVT::i32) {
2337 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2338 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2339 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2340 Chain, Dest, ARMCC, CCR,Cmp);
2343 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2344 ARMCC::CondCodes CondCode, CondCode2;
2345 FPCCToARMCC(CC, CondCode, CondCode2);
2347 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2348 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2349 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2350 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2351 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2352 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2353 if (CondCode2 != ARMCC::AL) {
2354 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2355 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2356 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2361 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2362 SDValue Chain = Op.getOperand(0);
2363 SDValue Table = Op.getOperand(1);
2364 SDValue Index = Op.getOperand(2);
2365 DebugLoc dl = Op.getDebugLoc();
2367 EVT PTy = getPointerTy();
2368 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2369 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2370 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2371 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2372 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2373 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2374 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2375 if (Subtarget->isThumb2()) {
2376 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2377 // which does another jump to the destination. This also makes it easier
2378 // to translate it to TBB / TBH later.
2379 // FIXME: This might not work if the function is extremely large.
2380 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2381 Addr, Op.getOperand(2), JTI, UId);
2383 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2384 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2385 PseudoSourceValue::getJumpTable(), 0,
2387 Chain = Addr.getValue(1);
2388 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2389 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2391 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2392 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2393 Chain = Addr.getValue(1);
2394 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2398 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2399 DebugLoc dl = Op.getDebugLoc();
2402 switch (Op.getOpcode()) {
2404 assert(0 && "Invalid opcode!");
2405 case ISD::FP_TO_SINT:
2406 Opc = ARMISD::FTOSI;
2408 case ISD::FP_TO_UINT:
2409 Opc = ARMISD::FTOUI;
2412 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2413 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2416 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2417 EVT VT = Op.getValueType();
2418 DebugLoc dl = Op.getDebugLoc();
2421 switch (Op.getOpcode()) {
2423 assert(0 && "Invalid opcode!");
2424 case ISD::SINT_TO_FP:
2425 Opc = ARMISD::SITOF;
2427 case ISD::UINT_TO_FP:
2428 Opc = ARMISD::UITOF;
2432 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2433 return DAG.getNode(Opc, dl, VT, Op);
2436 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2437 // Implement fcopysign with a fabs and a conditional fneg.
2438 SDValue Tmp0 = Op.getOperand(0);
2439 SDValue Tmp1 = Op.getOperand(1);
2440 DebugLoc dl = Op.getDebugLoc();
2441 EVT VT = Op.getValueType();
2442 EVT SrcVT = Tmp1.getValueType();
2443 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2444 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2445 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2446 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2447 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2450 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2451 MachineFunction &MF = DAG.getMachineFunction();
2452 MachineFrameInfo *MFI = MF.getFrameInfo();
2453 MFI->setReturnAddressIsTaken(true);
2455 EVT VT = Op.getValueType();
2456 DebugLoc dl = Op.getDebugLoc();
2457 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2459 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2460 SDValue Offset = DAG.getConstant(4, MVT::i32);
2461 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2462 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2463 NULL, 0, false, false, 0);
2466 // Return LR, which contains the return address. Mark it an implicit live-in.
2467 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2468 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2471 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2472 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2473 MFI->setFrameAddressIsTaken(true);
2475 EVT VT = Op.getValueType();
2476 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2477 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2478 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2479 ? ARM::R7 : ARM::R11;
2480 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2482 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2487 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2488 /// expand a bit convert where either the source or destination type is i64 to
2489 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2490 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2491 /// vectors), since the legalizer won't know what to do with that.
2492 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2493 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2494 DebugLoc dl = N->getDebugLoc();
2495 SDValue Op = N->getOperand(0);
2497 // This function is only supposed to be called for i64 types, either as the
2498 // source or destination of the bit convert.
2499 EVT SrcVT = Op.getValueType();
2500 EVT DstVT = N->getValueType(0);
2501 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2502 "ExpandBIT_CONVERT called for non-i64 type");
2504 // Turn i64->f64 into VMOVDRR.
2505 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2506 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2507 DAG.getConstant(0, MVT::i32));
2508 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2509 DAG.getConstant(1, MVT::i32));
2510 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2511 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2514 // Turn f64->i64 into VMOVRRD.
2515 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2516 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2517 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2518 // Merge the pieces into a single i64 value.
2519 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2525 /// getZeroVector - Returns a vector of specified type with all zero elements.
2527 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2528 assert(VT.isVector() && "Expected a vector type");
2530 // Zero vectors are used to represent vector negation and in those cases
2531 // will be implemented with the NEON VNEG instruction. However, VNEG does
2532 // not support i64 elements, so sometimes the zero vectors will need to be
2533 // explicitly constructed. For those cases, and potentially other uses in
2534 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2535 // to their dest type. This ensures they get CSE'd.
2537 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2538 SmallVector<SDValue, 8> Ops;
2541 if (VT.getSizeInBits() == 64) {
2542 Ops.assign(8, Cst); TVT = MVT::v8i8;
2544 Ops.assign(16, Cst); TVT = MVT::v16i8;
2546 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2548 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2551 /// getOnesVector - Returns a vector of specified type with all bits set.
2553 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2554 assert(VT.isVector() && "Expected a vector type");
2556 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2557 // dest type. This ensures they get CSE'd.
2559 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2560 SmallVector<SDValue, 8> Ops;
2563 if (VT.getSizeInBits() == 64) {
2564 Ops.assign(8, Cst); TVT = MVT::v8i8;
2566 Ops.assign(16, Cst); TVT = MVT::v16i8;
2568 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2570 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2573 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2574 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2575 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2576 SelectionDAG &DAG) const {
2577 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2578 EVT VT = Op.getValueType();
2579 unsigned VTBits = VT.getSizeInBits();
2580 DebugLoc dl = Op.getDebugLoc();
2581 SDValue ShOpLo = Op.getOperand(0);
2582 SDValue ShOpHi = Op.getOperand(1);
2583 SDValue ShAmt = Op.getOperand(2);
2585 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2587 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2589 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2590 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2591 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2592 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2593 DAG.getConstant(VTBits, MVT::i32));
2594 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2595 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2596 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2598 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2599 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2601 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2602 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2605 SDValue Ops[2] = { Lo, Hi };
2606 return DAG.getMergeValues(Ops, 2, dl);
2609 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2610 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2611 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2612 SelectionDAG &DAG) const {
2613 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2614 EVT VT = Op.getValueType();
2615 unsigned VTBits = VT.getSizeInBits();
2616 DebugLoc dl = Op.getDebugLoc();
2617 SDValue ShOpLo = Op.getOperand(0);
2618 SDValue ShOpHi = Op.getOperand(1);
2619 SDValue ShAmt = Op.getOperand(2);
2622 assert(Op.getOpcode() == ISD::SHL_PARTS);
2623 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2624 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2625 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2626 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2627 DAG.getConstant(VTBits, MVT::i32));
2628 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2629 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2631 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2632 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2633 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2635 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2636 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2639 SDValue Ops[2] = { Lo, Hi };
2640 return DAG.getMergeValues(Ops, 2, dl);
2643 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2644 const ARMSubtarget *ST) {
2645 EVT VT = N->getValueType(0);
2646 DebugLoc dl = N->getDebugLoc();
2648 if (!ST->hasV6T2Ops())
2651 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2652 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2655 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2656 const ARMSubtarget *ST) {
2657 EVT VT = N->getValueType(0);
2658 DebugLoc dl = N->getDebugLoc();
2660 // Lower vector shifts on NEON to use VSHL.
2661 if (VT.isVector()) {
2662 assert(ST->hasNEON() && "unexpected vector shift");
2664 // Left shifts translate directly to the vshiftu intrinsic.
2665 if (N->getOpcode() == ISD::SHL)
2666 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2667 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2668 N->getOperand(0), N->getOperand(1));
2670 assert((N->getOpcode() == ISD::SRA ||
2671 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2673 // NEON uses the same intrinsics for both left and right shifts. For
2674 // right shifts, the shift amounts are negative, so negate the vector of
2676 EVT ShiftVT = N->getOperand(1).getValueType();
2677 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2678 getZeroVector(ShiftVT, DAG, dl),
2680 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2681 Intrinsic::arm_neon_vshifts :
2682 Intrinsic::arm_neon_vshiftu);
2683 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2684 DAG.getConstant(vshiftInt, MVT::i32),
2685 N->getOperand(0), NegatedCount);
2688 // We can get here for a node like i32 = ISD::SHL i32, i64
2692 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2693 "Unknown shift to lower!");
2695 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2696 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2697 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2700 // If we are in thumb mode, we don't have RRX.
2701 if (ST->isThumb1Only()) return SDValue();
2703 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2704 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2705 DAG.getConstant(0, MVT::i32));
2706 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2707 DAG.getConstant(1, MVT::i32));
2709 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2710 // captures the result into a carry flag.
2711 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2712 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2714 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2715 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2717 // Merge the pieces into a single i64 value.
2718 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2721 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2722 SDValue TmpOp0, TmpOp1;
2723 bool Invert = false;
2727 SDValue Op0 = Op.getOperand(0);
2728 SDValue Op1 = Op.getOperand(1);
2729 SDValue CC = Op.getOperand(2);
2730 EVT VT = Op.getValueType();
2731 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2732 DebugLoc dl = Op.getDebugLoc();
2734 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2735 switch (SetCCOpcode) {
2736 default: llvm_unreachable("Illegal FP comparison"); break;
2738 case ISD::SETNE: Invert = true; // Fallthrough
2740 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2742 case ISD::SETLT: Swap = true; // Fallthrough
2744 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2746 case ISD::SETLE: Swap = true; // Fallthrough
2748 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2749 case ISD::SETUGE: Swap = true; // Fallthrough
2750 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2751 case ISD::SETUGT: Swap = true; // Fallthrough
2752 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2753 case ISD::SETUEQ: Invert = true; // Fallthrough
2755 // Expand this to (OLT | OGT).
2759 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2760 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2762 case ISD::SETUO: Invert = true; // Fallthrough
2764 // Expand this to (OLT | OGE).
2768 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2769 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2773 // Integer comparisons.
2774 switch (SetCCOpcode) {
2775 default: llvm_unreachable("Illegal integer comparison"); break;
2776 case ISD::SETNE: Invert = true;
2777 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2778 case ISD::SETLT: Swap = true;
2779 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2780 case ISD::SETLE: Swap = true;
2781 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2782 case ISD::SETULT: Swap = true;
2783 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2784 case ISD::SETULE: Swap = true;
2785 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2788 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2789 if (Opc == ARMISD::VCEQ) {
2792 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2794 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2797 // Ignore bitconvert.
2798 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2799 AndOp = AndOp.getOperand(0);
2801 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2803 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2804 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2811 std::swap(Op0, Op1);
2813 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2816 Result = DAG.getNOT(dl, Result, VT);
2821 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2822 /// valid vector constant for a NEON instruction with a "modified immediate"
2823 /// operand (e.g., VMOV). If so, return either the constant being
2824 /// splatted or the encoded value, depending on the DoEncode parameter. The
2825 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2826 /// bits7-0=Immediate.
2827 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2828 unsigned SplatBitSize, SelectionDAG &DAG,
2829 bool isVMOV, bool DoEncode) {
2830 unsigned Op, Cmode, Imm;
2833 // SplatBitSize is set to the smallest size that splats the vector, so a
2834 // zero vector will always have SplatBitSize == 8. However, NEON modified
2835 // immediate instructions others than VMOV do not support the 8-bit encoding
2836 // of a zero vector, and the default encoding of zero is supposed to be the
2842 switch (SplatBitSize) {
2844 // Any 1-byte value is OK. Op=0, Cmode=1110.
2845 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2852 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2854 if ((SplatBits & ~0xff) == 0) {
2855 // Value = 0x00nn: Op=x, Cmode=100x.
2860 if ((SplatBits & ~0xff00) == 0) {
2861 // Value = 0xnn00: Op=x, Cmode=101x.
2863 Imm = SplatBits >> 8;
2869 // NEON's 32-bit VMOV supports splat values where:
2870 // * only one byte is nonzero, or
2871 // * the least significant byte is 0xff and the second byte is nonzero, or
2872 // * the least significant 2 bytes are 0xff and the third is nonzero.
2874 if ((SplatBits & ~0xff) == 0) {
2875 // Value = 0x000000nn: Op=x, Cmode=000x.
2880 if ((SplatBits & ~0xff00) == 0) {
2881 // Value = 0x0000nn00: Op=x, Cmode=001x.
2883 Imm = SplatBits >> 8;
2886 if ((SplatBits & ~0xff0000) == 0) {
2887 // Value = 0x00nn0000: Op=x, Cmode=010x.
2889 Imm = SplatBits >> 16;
2892 if ((SplatBits & ~0xff000000) == 0) {
2893 // Value = 0xnn000000: Op=x, Cmode=011x.
2895 Imm = SplatBits >> 24;
2899 if ((SplatBits & ~0xffff) == 0 &&
2900 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2901 // Value = 0x0000nnff: Op=x, Cmode=1100.
2903 Imm = SplatBits >> 8;
2908 if ((SplatBits & ~0xffffff) == 0 &&
2909 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2910 // Value = 0x00nnffff: Op=x, Cmode=1101.
2912 Imm = SplatBits >> 16;
2913 SplatBits |= 0xffff;
2917 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2918 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2919 // VMOV.I32. A (very) minor optimization would be to replicate the value
2920 // and fall through here to test for a valid 64-bit splat. But, then the
2921 // caller would also need to check and handle the change in size.
2925 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2928 uint64_t BitMask = 0xff;
2930 unsigned ImmMask = 1;
2932 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2933 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2936 } else if ((SplatBits & BitMask) != 0) {
2942 // Op=1, Cmode=1110.
2951 llvm_unreachable("unexpected size for isNEONModifiedImm");
2956 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2957 return DAG.getTargetConstant(SplatBits, VT);
2961 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2962 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2963 /// size, return the encoded value for that immediate. The ByteSize field
2964 /// indicates the number of bytes of each element [1248].
2965 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2966 SelectionDAG &DAG) {
2967 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2968 APInt SplatBits, SplatUndef;
2969 unsigned SplatBitSize;
2971 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2972 HasAnyUndefs, ByteSize * 8))
2975 if (SplatBitSize > ByteSize * 8)
2978 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2979 SplatBitSize, DAG, isVMOV, true);
2982 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2983 bool &ReverseVEXT, unsigned &Imm) {
2984 unsigned NumElts = VT.getVectorNumElements();
2985 ReverseVEXT = false;
2988 // If this is a VEXT shuffle, the immediate value is the index of the first
2989 // element. The other shuffle indices must be the successive elements after
2991 unsigned ExpectedElt = Imm;
2992 for (unsigned i = 1; i < NumElts; ++i) {
2993 // Increment the expected index. If it wraps around, it may still be
2994 // a VEXT but the source vectors must be swapped.
2996 if (ExpectedElt == NumElts * 2) {
3001 if (ExpectedElt != static_cast<unsigned>(M[i]))
3005 // Adjust the index value if the source operands will be swapped.
3012 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3013 /// instruction with the specified blocksize. (The order of the elements
3014 /// within each block of the vector is reversed.)
3015 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3016 unsigned BlockSize) {
3017 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3018 "Only possible block sizes for VREV are: 16, 32, 64");
3020 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3024 unsigned NumElts = VT.getVectorNumElements();
3025 unsigned BlockElts = M[0] + 1;
3027 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3030 for (unsigned i = 0; i < NumElts; ++i) {
3031 if ((unsigned) M[i] !=
3032 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3039 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3040 unsigned &WhichResult) {
3041 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3045 unsigned NumElts = VT.getVectorNumElements();
3046 WhichResult = (M[0] == 0 ? 0 : 1);
3047 for (unsigned i = 0; i < NumElts; i += 2) {
3048 if ((unsigned) M[i] != i + WhichResult ||
3049 (unsigned) M[i+1] != i + NumElts + WhichResult)
3055 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3056 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3057 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3058 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3059 unsigned &WhichResult) {
3060 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3064 unsigned NumElts = VT.getVectorNumElements();
3065 WhichResult = (M[0] == 0 ? 0 : 1);
3066 for (unsigned i = 0; i < NumElts; i += 2) {
3067 if ((unsigned) M[i] != i + WhichResult ||
3068 (unsigned) M[i+1] != i + WhichResult)
3074 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3075 unsigned &WhichResult) {
3076 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3080 unsigned NumElts = VT.getVectorNumElements();
3081 WhichResult = (M[0] == 0 ? 0 : 1);
3082 for (unsigned i = 0; i != NumElts; ++i) {
3083 if ((unsigned) M[i] != 2 * i + WhichResult)
3087 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3088 if (VT.is64BitVector() && EltSz == 32)
3094 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3095 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3096 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3097 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3098 unsigned &WhichResult) {
3099 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3103 unsigned Half = VT.getVectorNumElements() / 2;
3104 WhichResult = (M[0] == 0 ? 0 : 1);
3105 for (unsigned j = 0; j != 2; ++j) {
3106 unsigned Idx = WhichResult;
3107 for (unsigned i = 0; i != Half; ++i) {
3108 if ((unsigned) M[i + j * Half] != Idx)
3114 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3115 if (VT.is64BitVector() && EltSz == 32)
3121 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3122 unsigned &WhichResult) {
3123 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3127 unsigned NumElts = VT.getVectorNumElements();
3128 WhichResult = (M[0] == 0 ? 0 : 1);
3129 unsigned Idx = WhichResult * NumElts / 2;
3130 for (unsigned i = 0; i != NumElts; i += 2) {
3131 if ((unsigned) M[i] != Idx ||
3132 (unsigned) M[i+1] != Idx + NumElts)
3137 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3138 if (VT.is64BitVector() && EltSz == 32)
3144 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3145 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3146 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3147 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3148 unsigned &WhichResult) {
3149 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3153 unsigned NumElts = VT.getVectorNumElements();
3154 WhichResult = (M[0] == 0 ? 0 : 1);
3155 unsigned Idx = WhichResult * NumElts / 2;
3156 for (unsigned i = 0; i != NumElts; i += 2) {
3157 if ((unsigned) M[i] != Idx ||
3158 (unsigned) M[i+1] != Idx)
3163 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3164 if (VT.is64BitVector() && EltSz == 32)
3171 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3172 // Canonicalize all-zeros and all-ones vectors.
3173 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3174 if (ConstVal->isNullValue())
3175 return getZeroVector(VT, DAG, dl);
3176 if (ConstVal->isAllOnesValue())
3177 return getOnesVector(VT, DAG, dl);
3180 if (VT.is64BitVector()) {
3181 switch (Val.getValueType().getSizeInBits()) {
3182 case 8: CanonicalVT = MVT::v8i8; break;
3183 case 16: CanonicalVT = MVT::v4i16; break;
3184 case 32: CanonicalVT = MVT::v2i32; break;
3185 case 64: CanonicalVT = MVT::v1i64; break;
3186 default: llvm_unreachable("unexpected splat element type"); break;
3189 assert(VT.is128BitVector() && "unknown splat vector size");
3190 switch (Val.getValueType().getSizeInBits()) {
3191 case 8: CanonicalVT = MVT::v16i8; break;
3192 case 16: CanonicalVT = MVT::v8i16; break;
3193 case 32: CanonicalVT = MVT::v4i32; break;
3194 case 64: CanonicalVT = MVT::v2i64; break;
3195 default: llvm_unreachable("unexpected splat element type"); break;
3199 // Build a canonical splat for this value.
3200 SmallVector<SDValue, 8> Ops;
3201 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3202 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3204 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3207 // If this is a case we can't handle, return null and let the default
3208 // expansion code take care of it.
3209 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3210 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3211 DebugLoc dl = Op.getDebugLoc();
3212 EVT VT = Op.getValueType();
3214 APInt SplatBits, SplatUndef;
3215 unsigned SplatBitSize;
3217 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3218 if (SplatBitSize <= 64) {
3219 // Check if an immediate VMOV works.
3220 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3221 SplatUndef.getZExtValue(),
3222 SplatBitSize, DAG, true, false);
3224 return BuildSplat(Val, VT, DAG, dl);
3228 // Scan through the operands to see if only one value is used.
3229 unsigned NumElts = VT.getVectorNumElements();
3230 bool isOnlyLowElement = true;
3231 bool usesOnlyOneValue = true;
3232 bool isConstant = true;
3234 for (unsigned i = 0; i < NumElts; ++i) {
3235 SDValue V = Op.getOperand(i);
3236 if (V.getOpcode() == ISD::UNDEF)
3239 isOnlyLowElement = false;
3240 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3243 if (!Value.getNode())
3245 else if (V != Value)
3246 usesOnlyOneValue = false;
3249 if (!Value.getNode())
3250 return DAG.getUNDEF(VT);
3252 if (isOnlyLowElement)
3253 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3255 // If all elements are constants, fall back to the default expansion, which
3256 // will generate a load from the constant pool.
3260 // Use VDUP for non-constant splats.
3261 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3262 if (usesOnlyOneValue && EltSize <= 32)
3263 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3265 // Vectors with 32- or 64-bit elements can be built by directly assigning
3266 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3267 // will be legalized.
3268 if (EltSize >= 32) {
3269 // Do the expansion with floating-point types, since that is what the VFP
3270 // registers are defined to use, and since i64 is not legal.
3271 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3272 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3273 SmallVector<SDValue, 8> Ops;
3274 for (unsigned i = 0; i < NumElts; ++i)
3275 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3276 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3277 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3283 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3284 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3285 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3286 /// are assumed to be legal.
3288 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3290 if (VT.getVectorNumElements() == 4 &&
3291 (VT.is128BitVector() || VT.is64BitVector())) {
3292 unsigned PFIndexes[4];
3293 for (unsigned i = 0; i != 4; ++i) {
3297 PFIndexes[i] = M[i];
3300 // Compute the index in the perfect shuffle table.
3301 unsigned PFTableIndex =
3302 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3303 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3304 unsigned Cost = (PFEntry >> 30);
3311 unsigned Imm, WhichResult;
3313 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3314 return (EltSize >= 32 ||
3315 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3316 isVREVMask(M, VT, 64) ||
3317 isVREVMask(M, VT, 32) ||
3318 isVREVMask(M, VT, 16) ||
3319 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3320 isVTRNMask(M, VT, WhichResult) ||
3321 isVUZPMask(M, VT, WhichResult) ||
3322 isVZIPMask(M, VT, WhichResult) ||
3323 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3324 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3325 isVZIP_v_undef_Mask(M, VT, WhichResult));
3328 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3329 /// the specified operations to build the shuffle.
3330 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3331 SDValue RHS, SelectionDAG &DAG,
3333 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3334 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3335 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3338 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3347 OP_VUZPL, // VUZP, left result
3348 OP_VUZPR, // VUZP, right result
3349 OP_VZIPL, // VZIP, left result
3350 OP_VZIPR, // VZIP, right result
3351 OP_VTRNL, // VTRN, left result
3352 OP_VTRNR // VTRN, right result
3355 if (OpNum == OP_COPY) {
3356 if (LHSID == (1*9+2)*9+3) return LHS;
3357 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3361 SDValue OpLHS, OpRHS;
3362 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3363 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3364 EVT VT = OpLHS.getValueType();
3367 default: llvm_unreachable("Unknown shuffle opcode!");
3369 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3374 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3375 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3379 return DAG.getNode(ARMISD::VEXT, dl, VT,
3381 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3384 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3385 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3388 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3389 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3392 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3393 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3397 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3398 SDValue V1 = Op.getOperand(0);
3399 SDValue V2 = Op.getOperand(1);
3400 DebugLoc dl = Op.getDebugLoc();
3401 EVT VT = Op.getValueType();
3402 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3403 SmallVector<int, 8> ShuffleMask;
3405 // Convert shuffles that are directly supported on NEON to target-specific
3406 // DAG nodes, instead of keeping them as shuffles and matching them again
3407 // during code selection. This is more efficient and avoids the possibility
3408 // of inconsistencies between legalization and selection.
3409 // FIXME: floating-point vectors should be canonicalized to integer vectors
3410 // of the same time so that they get CSEd properly.
3411 SVN->getMask(ShuffleMask);
3413 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3414 if (EltSize <= 32) {
3415 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3416 int Lane = SVN->getSplatIndex();
3417 // If this is undef splat, generate it via "just" vdup, if possible.
3418 if (Lane == -1) Lane = 0;
3420 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3421 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3423 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3424 DAG.getConstant(Lane, MVT::i32));
3429 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3432 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3433 DAG.getConstant(Imm, MVT::i32));
3436 if (isVREVMask(ShuffleMask, VT, 64))
3437 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3438 if (isVREVMask(ShuffleMask, VT, 32))
3439 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3440 if (isVREVMask(ShuffleMask, VT, 16))
3441 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3443 // Check for Neon shuffles that modify both input vectors in place.
3444 // If both results are used, i.e., if there are two shuffles with the same
3445 // source operands and with masks corresponding to both results of one of
3446 // these operations, DAG memoization will ensure that a single node is
3447 // used for both shuffles.
3448 unsigned WhichResult;
3449 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3450 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3451 V1, V2).getValue(WhichResult);
3452 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3453 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3454 V1, V2).getValue(WhichResult);
3455 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3456 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3457 V1, V2).getValue(WhichResult);
3459 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3460 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3461 V1, V1).getValue(WhichResult);
3462 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3463 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3464 V1, V1).getValue(WhichResult);
3465 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3466 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3467 V1, V1).getValue(WhichResult);
3470 // If the shuffle is not directly supported and it has 4 elements, use
3471 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3472 unsigned NumElts = VT.getVectorNumElements();
3474 unsigned PFIndexes[4];
3475 for (unsigned i = 0; i != 4; ++i) {
3476 if (ShuffleMask[i] < 0)
3479 PFIndexes[i] = ShuffleMask[i];
3482 // Compute the index in the perfect shuffle table.
3483 unsigned PFTableIndex =
3484 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3485 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3486 unsigned Cost = (PFEntry >> 30);
3489 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3492 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3493 if (EltSize >= 32) {
3494 // Do the expansion with floating-point types, since that is what the VFP
3495 // registers are defined to use, and since i64 is not legal.
3496 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3497 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3498 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3499 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3500 SmallVector<SDValue, 8> Ops;
3501 for (unsigned i = 0; i < NumElts; ++i) {
3502 if (ShuffleMask[i] < 0)
3503 Ops.push_back(DAG.getUNDEF(EltVT));
3505 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3506 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3507 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3510 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3511 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3517 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3518 EVT VT = Op.getValueType();
3519 DebugLoc dl = Op.getDebugLoc();
3520 SDValue Vec = Op.getOperand(0);
3521 SDValue Lane = Op.getOperand(1);
3522 assert(VT == MVT::i32 &&
3523 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3524 "unexpected type for custom-lowering vector extract");
3525 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3528 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3529 // The only time a CONCAT_VECTORS operation can have legal types is when
3530 // two 64-bit vectors are concatenated to a 128-bit vector.
3531 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3532 "unexpected CONCAT_VECTORS");
3533 DebugLoc dl = Op.getDebugLoc();
3534 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3535 SDValue Op0 = Op.getOperand(0);
3536 SDValue Op1 = Op.getOperand(1);
3537 if (Op0.getOpcode() != ISD::UNDEF)
3538 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3539 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3540 DAG.getIntPtrConstant(0));
3541 if (Op1.getOpcode() != ISD::UNDEF)
3542 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3543 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3544 DAG.getIntPtrConstant(1));
3545 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3548 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3549 switch (Op.getOpcode()) {
3550 default: llvm_unreachable("Don't know how to custom lower this!");
3551 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3552 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3553 case ISD::GlobalAddress:
3554 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3555 LowerGlobalAddressELF(Op, DAG);
3556 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3557 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3558 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3559 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3560 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3561 case ISD::VASTART: return LowerVASTART(Op, DAG);
3562 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3563 case ISD::SINT_TO_FP:
3564 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3565 case ISD::FP_TO_SINT:
3566 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3567 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3568 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3569 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3570 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3571 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3572 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3573 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3575 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3578 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3579 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3580 case ISD::SRL_PARTS:
3581 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3582 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3583 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3584 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3585 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3586 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3587 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3592 /// ReplaceNodeResults - Replace the results of node with an illegal result
3593 /// type with new values built out of custom code.
3594 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3595 SmallVectorImpl<SDValue>&Results,
3596 SelectionDAG &DAG) const {
3598 switch (N->getOpcode()) {
3600 llvm_unreachable("Don't know how to custom expand this!");
3602 case ISD::BIT_CONVERT:
3603 Res = ExpandBIT_CONVERT(N, DAG);
3607 Res = LowerShift(N, DAG, Subtarget);
3611 Results.push_back(Res);
3614 //===----------------------------------------------------------------------===//
3615 // ARM Scheduler Hooks
3616 //===----------------------------------------------------------------------===//
3619 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3620 MachineBasicBlock *BB,
3621 unsigned Size) const {
3622 unsigned dest = MI->getOperand(0).getReg();
3623 unsigned ptr = MI->getOperand(1).getReg();
3624 unsigned oldval = MI->getOperand(2).getReg();
3625 unsigned newval = MI->getOperand(3).getReg();
3626 unsigned scratch = BB->getParent()->getRegInfo()
3627 .createVirtualRegister(ARM::GPRRegisterClass);
3628 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3629 DebugLoc dl = MI->getDebugLoc();
3630 bool isThumb2 = Subtarget->isThumb2();
3632 unsigned ldrOpc, strOpc;
3634 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3636 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3637 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3640 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3641 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3644 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3645 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3649 MachineFunction *MF = BB->getParent();
3650 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3651 MachineFunction::iterator It = BB;
3652 ++It; // insert the new blocks after the current block
3654 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3655 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3656 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3657 MF->insert(It, loop1MBB);
3658 MF->insert(It, loop2MBB);
3659 MF->insert(It, exitMBB);
3660 exitMBB->transferSuccessors(BB);
3664 // fallthrough --> loop1MBB
3665 BB->addSuccessor(loop1MBB);
3668 // ldrex dest, [ptr]
3672 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3673 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3674 .addReg(dest).addReg(oldval));
3675 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3676 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3677 BB->addSuccessor(loop2MBB);
3678 BB->addSuccessor(exitMBB);
3681 // strex scratch, newval, [ptr]
3685 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3687 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3688 .addReg(scratch).addImm(0));
3689 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3690 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3691 BB->addSuccessor(loop1MBB);
3692 BB->addSuccessor(exitMBB);
3698 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3704 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3705 unsigned Size, unsigned BinOpcode) const {
3706 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3707 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3709 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3710 MachineFunction *MF = BB->getParent();
3711 MachineFunction::iterator It = BB;
3714 unsigned dest = MI->getOperand(0).getReg();
3715 unsigned ptr = MI->getOperand(1).getReg();
3716 unsigned incr = MI->getOperand(2).getReg();
3717 DebugLoc dl = MI->getDebugLoc();
3719 bool isThumb2 = Subtarget->isThumb2();
3720 unsigned ldrOpc, strOpc;
3722 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3724 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3725 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3728 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3729 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3732 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3733 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3737 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3738 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3739 MF->insert(It, loopMBB);
3740 MF->insert(It, exitMBB);
3741 exitMBB->transferSuccessors(BB);
3743 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3744 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3745 unsigned scratch2 = (!BinOpcode) ? incr :
3746 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3750 // fallthrough --> loopMBB
3751 BB->addSuccessor(loopMBB);
3755 // <binop> scratch2, dest, incr
3756 // strex scratch, scratch2, ptr
3759 // fallthrough --> exitMBB
3761 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3763 // operand order needs to go the other way for NAND
3764 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3765 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3766 addReg(incr).addReg(dest)).addReg(0);
3768 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3769 addReg(dest).addReg(incr)).addReg(0);
3772 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3774 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3775 .addReg(scratch).addImm(0));
3776 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3777 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3779 BB->addSuccessor(loopMBB);
3780 BB->addSuccessor(exitMBB);
3786 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3792 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3793 MachineBasicBlock *BB) const {
3794 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3795 DebugLoc dl = MI->getDebugLoc();
3796 bool isThumb2 = Subtarget->isThumb2();
3797 switch (MI->getOpcode()) {
3800 llvm_unreachable("Unexpected instr type to insert");
3802 case ARM::ATOMIC_LOAD_ADD_I8:
3803 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3804 case ARM::ATOMIC_LOAD_ADD_I16:
3805 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3806 case ARM::ATOMIC_LOAD_ADD_I32:
3807 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3809 case ARM::ATOMIC_LOAD_AND_I8:
3810 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3811 case ARM::ATOMIC_LOAD_AND_I16:
3812 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3813 case ARM::ATOMIC_LOAD_AND_I32:
3814 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3816 case ARM::ATOMIC_LOAD_OR_I8:
3817 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3818 case ARM::ATOMIC_LOAD_OR_I16:
3819 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3820 case ARM::ATOMIC_LOAD_OR_I32:
3821 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3823 case ARM::ATOMIC_LOAD_XOR_I8:
3824 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3825 case ARM::ATOMIC_LOAD_XOR_I16:
3826 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3827 case ARM::ATOMIC_LOAD_XOR_I32:
3828 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3830 case ARM::ATOMIC_LOAD_NAND_I8:
3831 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3832 case ARM::ATOMIC_LOAD_NAND_I16:
3833 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3834 case ARM::ATOMIC_LOAD_NAND_I32:
3835 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3837 case ARM::ATOMIC_LOAD_SUB_I8:
3838 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3839 case ARM::ATOMIC_LOAD_SUB_I16:
3840 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3841 case ARM::ATOMIC_LOAD_SUB_I32:
3842 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3844 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3845 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3846 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3848 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3849 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3850 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3852 case ARM::tMOVCCr_pseudo: {
3853 // To "insert" a SELECT_CC instruction, we actually have to insert the
3854 // diamond control-flow pattern. The incoming instruction knows the
3855 // destination vreg to set, the condition code register to branch on, the
3856 // true/false values to select between, and a branch opcode to use.
3857 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3858 MachineFunction::iterator It = BB;
3864 // cmpTY ccX, r1, r2
3866 // fallthrough --> copy0MBB
3867 MachineBasicBlock *thisMBB = BB;
3868 MachineFunction *F = BB->getParent();
3869 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3870 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3871 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3872 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3873 F->insert(It, copy0MBB);
3874 F->insert(It, sinkMBB);
3875 // Update machine-CFG edges by first adding all successors of the current
3876 // block to the new block which will contain the Phi node for the select.
3877 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3878 E = BB->succ_end(); I != E; ++I)
3879 sinkMBB->addSuccessor(*I);
3880 // Next, remove all successors of the current block, and add the true
3881 // and fallthrough blocks as its successors.
3882 while (!BB->succ_empty())
3883 BB->removeSuccessor(BB->succ_begin());
3884 BB->addSuccessor(copy0MBB);
3885 BB->addSuccessor(sinkMBB);
3888 // %FalseValue = ...
3889 // # fallthrough to sinkMBB
3892 // Update machine-CFG edges
3893 BB->addSuccessor(sinkMBB);
3896 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3899 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3900 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3901 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3903 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3910 case ARM::t2SUBrSPi_:
3911 case ARM::t2SUBrSPi12_:
3912 case ARM::t2SUBrSPs_: {
3913 MachineFunction *MF = BB->getParent();
3914 unsigned DstReg = MI->getOperand(0).getReg();
3915 unsigned SrcReg = MI->getOperand(1).getReg();
3916 bool DstIsDead = MI->getOperand(0).isDead();
3917 bool SrcIsKill = MI->getOperand(1).isKill();
3919 if (SrcReg != ARM::SP) {
3920 // Copy the source to SP from virtual register.
3921 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3922 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3923 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3924 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3925 .addReg(SrcReg, getKillRegState(SrcIsKill));
3929 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3930 switch (MI->getOpcode()) {
3932 llvm_unreachable("Unexpected pseudo instruction!");
3938 OpOpc = ARM::tADDspr;
3941 OpOpc = ARM::tSUBspi;
3943 case ARM::t2SUBrSPi_:
3944 OpOpc = ARM::t2SUBrSPi;
3945 NeedPred = true; NeedCC = true;
3947 case ARM::t2SUBrSPi12_:
3948 OpOpc = ARM::t2SUBrSPi12;
3951 case ARM::t2SUBrSPs_:
3952 OpOpc = ARM::t2SUBrSPs;
3953 NeedPred = true; NeedCC = true; NeedOp3 = true;
3956 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3957 if (OpOpc == ARM::tAND)
3958 AddDefaultT1CC(MIB);
3959 MIB.addReg(ARM::SP);
3960 MIB.addOperand(MI->getOperand(2));
3962 MIB.addOperand(MI->getOperand(3));
3964 AddDefaultPred(MIB);
3968 // Copy the result from SP to virtual register.
3969 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3970 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3971 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3972 BuildMI(BB, dl, TII->get(CopyOpc))
3973 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3975 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3981 //===----------------------------------------------------------------------===//
3982 // ARM Optimization Hooks
3983 //===----------------------------------------------------------------------===//
3986 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3987 TargetLowering::DAGCombinerInfo &DCI) {
3988 SelectionDAG &DAG = DCI.DAG;
3989 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3990 EVT VT = N->getValueType(0);
3991 unsigned Opc = N->getOpcode();
3992 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3993 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3994 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3995 ISD::CondCode CC = ISD::SETCC_INVALID;
3998 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4000 SDValue CCOp = Slct.getOperand(0);
4001 if (CCOp.getOpcode() == ISD::SETCC)
4002 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4005 bool DoXform = false;
4007 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4010 if (LHS.getOpcode() == ISD::Constant &&
4011 cast<ConstantSDNode>(LHS)->isNullValue()) {
4013 } else if (CC != ISD::SETCC_INVALID &&
4014 RHS.getOpcode() == ISD::Constant &&
4015 cast<ConstantSDNode>(RHS)->isNullValue()) {
4016 std::swap(LHS, RHS);
4017 SDValue Op0 = Slct.getOperand(0);
4018 EVT OpVT = isSlctCC ? Op0.getValueType() :
4019 Op0.getOperand(0).getValueType();
4020 bool isInt = OpVT.isInteger();
4021 CC = ISD::getSetCCInverse(CC, isInt);
4023 if (!TLI.isCondCodeLegal(CC, OpVT))
4024 return SDValue(); // Inverse operator isn't legal.
4031 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4033 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4034 Slct.getOperand(0), Slct.getOperand(1), CC);
4035 SDValue CCOp = Slct.getOperand(0);
4037 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4038 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4039 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4040 CCOp, OtherOp, Result);
4045 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4046 static SDValue PerformADDCombine(SDNode *N,
4047 TargetLowering::DAGCombinerInfo &DCI) {
4048 // added by evan in r37685 with no testcase.
4049 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4051 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4052 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4053 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4054 if (Result.getNode()) return Result;
4056 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4057 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4058 if (Result.getNode()) return Result;
4064 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4065 static SDValue PerformSUBCombine(SDNode *N,
4066 TargetLowering::DAGCombinerInfo &DCI) {
4067 // added by evan in r37685 with no testcase.
4068 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4070 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4071 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4072 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4073 if (Result.getNode()) return Result;
4079 static SDValue PerformMULCombine(SDNode *N,
4080 TargetLowering::DAGCombinerInfo &DCI,
4081 const ARMSubtarget *Subtarget) {
4082 SelectionDAG &DAG = DCI.DAG;
4084 if (Subtarget->isThumb1Only())
4087 if (DAG.getMachineFunction().
4088 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4091 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4094 EVT VT = N->getValueType(0);
4098 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4102 uint64_t MulAmt = C->getZExtValue();
4103 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4104 ShiftAmt = ShiftAmt & (32 - 1);
4105 SDValue V = N->getOperand(0);
4106 DebugLoc DL = N->getDebugLoc();
4109 MulAmt >>= ShiftAmt;
4110 if (isPowerOf2_32(MulAmt - 1)) {
4111 // (mul x, 2^N + 1) => (add (shl x, N), x)
4112 Res = DAG.getNode(ISD::ADD, DL, VT,
4113 V, DAG.getNode(ISD::SHL, DL, VT,
4114 V, DAG.getConstant(Log2_32(MulAmt-1),
4116 } else if (isPowerOf2_32(MulAmt + 1)) {
4117 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4118 Res = DAG.getNode(ISD::SUB, DL, VT,
4119 DAG.getNode(ISD::SHL, DL, VT,
4120 V, DAG.getConstant(Log2_32(MulAmt+1),
4127 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4128 DAG.getConstant(ShiftAmt, MVT::i32));
4130 // Do not add new nodes to DAG combiner worklist.
4131 DCI.CombineTo(N, Res, false);
4135 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4136 /// ARMISD::VMOVRRD.
4137 static SDValue PerformVMOVRRDCombine(SDNode *N,
4138 TargetLowering::DAGCombinerInfo &DCI) {
4139 // fmrrd(fmdrr x, y) -> x,y
4140 SDValue InDouble = N->getOperand(0);
4141 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4142 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4146 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4147 /// operand of a vector shift operation, where all the elements of the
4148 /// build_vector must have the same constant integer value.
4149 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4150 // Ignore bit_converts.
4151 while (Op.getOpcode() == ISD::BIT_CONVERT)
4152 Op = Op.getOperand(0);
4153 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4154 APInt SplatBits, SplatUndef;
4155 unsigned SplatBitSize;
4157 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4158 HasAnyUndefs, ElementBits) ||
4159 SplatBitSize > ElementBits)
4161 Cnt = SplatBits.getSExtValue();
4165 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4166 /// operand of a vector shift left operation. That value must be in the range:
4167 /// 0 <= Value < ElementBits for a left shift; or
4168 /// 0 <= Value <= ElementBits for a long left shift.
4169 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4170 assert(VT.isVector() && "vector shift count is not a vector type");
4171 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4172 if (! getVShiftImm(Op, ElementBits, Cnt))
4174 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4177 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4178 /// operand of a vector shift right operation. For a shift opcode, the value
4179 /// is positive, but for an intrinsic the value count must be negative. The
4180 /// absolute value must be in the range:
4181 /// 1 <= |Value| <= ElementBits for a right shift; or
4182 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4183 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4185 assert(VT.isVector() && "vector shift count is not a vector type");
4186 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4187 if (! getVShiftImm(Op, ElementBits, Cnt))
4191 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4194 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4195 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4196 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4199 // Don't do anything for most intrinsics.
4202 // Vector shifts: check for immediate versions and lower them.
4203 // Note: This is done during DAG combining instead of DAG legalizing because
4204 // the build_vectors for 64-bit vector element shift counts are generally
4205 // not legal, and it is hard to see their values after they get legalized to
4206 // loads from a constant pool.
4207 case Intrinsic::arm_neon_vshifts:
4208 case Intrinsic::arm_neon_vshiftu:
4209 case Intrinsic::arm_neon_vshiftls:
4210 case Intrinsic::arm_neon_vshiftlu:
4211 case Intrinsic::arm_neon_vshiftn:
4212 case Intrinsic::arm_neon_vrshifts:
4213 case Intrinsic::arm_neon_vrshiftu:
4214 case Intrinsic::arm_neon_vrshiftn:
4215 case Intrinsic::arm_neon_vqshifts:
4216 case Intrinsic::arm_neon_vqshiftu:
4217 case Intrinsic::arm_neon_vqshiftsu:
4218 case Intrinsic::arm_neon_vqshiftns:
4219 case Intrinsic::arm_neon_vqshiftnu:
4220 case Intrinsic::arm_neon_vqshiftnsu:
4221 case Intrinsic::arm_neon_vqrshiftns:
4222 case Intrinsic::arm_neon_vqrshiftnu:
4223 case Intrinsic::arm_neon_vqrshiftnsu: {
4224 EVT VT = N->getOperand(1).getValueType();
4226 unsigned VShiftOpc = 0;
4229 case Intrinsic::arm_neon_vshifts:
4230 case Intrinsic::arm_neon_vshiftu:
4231 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4232 VShiftOpc = ARMISD::VSHL;
4235 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4236 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4237 ARMISD::VSHRs : ARMISD::VSHRu);
4242 case Intrinsic::arm_neon_vshiftls:
4243 case Intrinsic::arm_neon_vshiftlu:
4244 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4246 llvm_unreachable("invalid shift count for vshll intrinsic");
4248 case Intrinsic::arm_neon_vrshifts:
4249 case Intrinsic::arm_neon_vrshiftu:
4250 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4254 case Intrinsic::arm_neon_vqshifts:
4255 case Intrinsic::arm_neon_vqshiftu:
4256 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4260 case Intrinsic::arm_neon_vqshiftsu:
4261 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4263 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4265 case Intrinsic::arm_neon_vshiftn:
4266 case Intrinsic::arm_neon_vrshiftn:
4267 case Intrinsic::arm_neon_vqshiftns:
4268 case Intrinsic::arm_neon_vqshiftnu:
4269 case Intrinsic::arm_neon_vqshiftnsu:
4270 case Intrinsic::arm_neon_vqrshiftns:
4271 case Intrinsic::arm_neon_vqrshiftnu:
4272 case Intrinsic::arm_neon_vqrshiftnsu:
4273 // Narrowing shifts require an immediate right shift.
4274 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4276 llvm_unreachable("invalid shift count for narrowing vector shift "
4280 llvm_unreachable("unhandled vector shift");
4284 case Intrinsic::arm_neon_vshifts:
4285 case Intrinsic::arm_neon_vshiftu:
4286 // Opcode already set above.
4288 case Intrinsic::arm_neon_vshiftls:
4289 case Intrinsic::arm_neon_vshiftlu:
4290 if (Cnt == VT.getVectorElementType().getSizeInBits())
4291 VShiftOpc = ARMISD::VSHLLi;
4293 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4294 ARMISD::VSHLLs : ARMISD::VSHLLu);
4296 case Intrinsic::arm_neon_vshiftn:
4297 VShiftOpc = ARMISD::VSHRN; break;
4298 case Intrinsic::arm_neon_vrshifts:
4299 VShiftOpc = ARMISD::VRSHRs; break;
4300 case Intrinsic::arm_neon_vrshiftu:
4301 VShiftOpc = ARMISD::VRSHRu; break;
4302 case Intrinsic::arm_neon_vrshiftn:
4303 VShiftOpc = ARMISD::VRSHRN; break;
4304 case Intrinsic::arm_neon_vqshifts:
4305 VShiftOpc = ARMISD::VQSHLs; break;
4306 case Intrinsic::arm_neon_vqshiftu:
4307 VShiftOpc = ARMISD::VQSHLu; break;
4308 case Intrinsic::arm_neon_vqshiftsu:
4309 VShiftOpc = ARMISD::VQSHLsu; break;
4310 case Intrinsic::arm_neon_vqshiftns:
4311 VShiftOpc = ARMISD::VQSHRNs; break;
4312 case Intrinsic::arm_neon_vqshiftnu:
4313 VShiftOpc = ARMISD::VQSHRNu; break;
4314 case Intrinsic::arm_neon_vqshiftnsu:
4315 VShiftOpc = ARMISD::VQSHRNsu; break;
4316 case Intrinsic::arm_neon_vqrshiftns:
4317 VShiftOpc = ARMISD::VQRSHRNs; break;
4318 case Intrinsic::arm_neon_vqrshiftnu:
4319 VShiftOpc = ARMISD::VQRSHRNu; break;
4320 case Intrinsic::arm_neon_vqrshiftnsu:
4321 VShiftOpc = ARMISD::VQRSHRNsu; break;
4324 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4325 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4328 case Intrinsic::arm_neon_vshiftins: {
4329 EVT VT = N->getOperand(1).getValueType();
4331 unsigned VShiftOpc = 0;
4333 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4334 VShiftOpc = ARMISD::VSLI;
4335 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4336 VShiftOpc = ARMISD::VSRI;
4338 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4341 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4342 N->getOperand(1), N->getOperand(2),
4343 DAG.getConstant(Cnt, MVT::i32));
4346 case Intrinsic::arm_neon_vqrshifts:
4347 case Intrinsic::arm_neon_vqrshiftu:
4348 // No immediate versions of these to check for.
4355 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4356 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4357 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4358 /// vector element shift counts are generally not legal, and it is hard to see
4359 /// their values after they get legalized to loads from a constant pool.
4360 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4361 const ARMSubtarget *ST) {
4362 EVT VT = N->getValueType(0);
4364 // Nothing to be done for scalar shifts.
4365 if (! VT.isVector())
4368 assert(ST->hasNEON() && "unexpected vector shift");
4371 switch (N->getOpcode()) {
4372 default: llvm_unreachable("unexpected shift opcode");
4375 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4376 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4377 DAG.getConstant(Cnt, MVT::i32));
4382 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4383 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4384 ARMISD::VSHRs : ARMISD::VSHRu);
4385 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4386 DAG.getConstant(Cnt, MVT::i32));
4392 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4393 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4394 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4395 const ARMSubtarget *ST) {
4396 SDValue N0 = N->getOperand(0);
4398 // Check for sign- and zero-extensions of vector extract operations of 8-
4399 // and 16-bit vector elements. NEON supports these directly. They are
4400 // handled during DAG combining because type legalization will promote them
4401 // to 32-bit types and it is messy to recognize the operations after that.
4402 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4403 SDValue Vec = N0.getOperand(0);
4404 SDValue Lane = N0.getOperand(1);
4405 EVT VT = N->getValueType(0);
4406 EVT EltVT = N0.getValueType();
4407 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4409 if (VT == MVT::i32 &&
4410 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4411 TLI.isTypeLegal(Vec.getValueType())) {
4414 switch (N->getOpcode()) {
4415 default: llvm_unreachable("unexpected opcode");
4416 case ISD::SIGN_EXTEND:
4417 Opc = ARMISD::VGETLANEs;
4419 case ISD::ZERO_EXTEND:
4420 case ISD::ANY_EXTEND:
4421 Opc = ARMISD::VGETLANEu;
4424 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4431 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4432 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4433 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4434 const ARMSubtarget *ST) {
4435 // If the target supports NEON, try to use vmax/vmin instructions for f32
4436 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4437 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4438 // a NaN; only do the transformation when it matches that behavior.
4440 // For now only do this when using NEON for FP operations; if using VFP, it
4441 // is not obvious that the benefit outweighs the cost of switching to the
4443 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4444 N->getValueType(0) != MVT::f32)
4447 SDValue CondLHS = N->getOperand(0);
4448 SDValue CondRHS = N->getOperand(1);
4449 SDValue LHS = N->getOperand(2);
4450 SDValue RHS = N->getOperand(3);
4451 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4453 unsigned Opcode = 0;
4455 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4456 IsReversed = false; // x CC y ? x : y
4457 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4458 IsReversed = true ; // x CC y ? y : x
4472 // If LHS is NaN, an ordered comparison will be false and the result will
4473 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4474 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4475 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4476 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4478 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4479 // will return -0, so vmin can only be used for unsafe math or if one of
4480 // the operands is known to be nonzero.
4481 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4483 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4485 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4494 // If LHS is NaN, an ordered comparison will be false and the result will
4495 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4496 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4497 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4498 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4500 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4501 // will return +0, so vmax can only be used for unsafe math or if one of
4502 // the operands is known to be nonzero.
4503 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4505 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4507 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4513 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4516 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4517 DAGCombinerInfo &DCI) const {
4518 switch (N->getOpcode()) {
4520 case ISD::ADD: return PerformADDCombine(N, DCI);
4521 case ISD::SUB: return PerformSUBCombine(N, DCI);
4522 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4523 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4524 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4527 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4528 case ISD::SIGN_EXTEND:
4529 case ISD::ZERO_EXTEND:
4530 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4531 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4536 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4537 if (!Subtarget->hasV6Ops())
4538 // Pre-v6 does not support unaligned mem access.
4541 // v6+ may or may not support unaligned mem access depending on the system
4543 // FIXME: This is pretty conservative. Should we provide cmdline option to
4544 // control the behaviour?
4545 if (!Subtarget->isTargetDarwin())
4548 switch (VT.getSimpleVT().SimpleTy) {
4555 // FIXME: VLD1 etc with standard alignment is legal.
4559 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4564 switch (VT.getSimpleVT().SimpleTy) {
4565 default: return false;
4580 if ((V & (Scale - 1)) != 0)
4583 return V == (V & ((1LL << 5) - 1));
4586 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4587 const ARMSubtarget *Subtarget) {
4594 switch (VT.getSimpleVT().SimpleTy) {
4595 default: return false;
4600 // + imm12 or - imm8
4602 return V == (V & ((1LL << 8) - 1));
4603 return V == (V & ((1LL << 12) - 1));
4606 // Same as ARM mode. FIXME: NEON?
4607 if (!Subtarget->hasVFP2())
4612 return V == (V & ((1LL << 8) - 1));
4616 /// isLegalAddressImmediate - Return true if the integer value can be used
4617 /// as the offset of the target addressing mode for load / store of the
4619 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4620 const ARMSubtarget *Subtarget) {
4627 if (Subtarget->isThumb1Only())
4628 return isLegalT1AddressImmediate(V, VT);
4629 else if (Subtarget->isThumb2())
4630 return isLegalT2AddressImmediate(V, VT, Subtarget);
4635 switch (VT.getSimpleVT().SimpleTy) {
4636 default: return false;
4641 return V == (V & ((1LL << 12) - 1));
4644 return V == (V & ((1LL << 8) - 1));
4647 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4652 return V == (V & ((1LL << 8) - 1));
4656 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4658 int Scale = AM.Scale;
4662 switch (VT.getSimpleVT().SimpleTy) {
4663 default: return false;
4672 return Scale == 2 || Scale == 4 || Scale == 8;
4675 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4679 // Note, we allow "void" uses (basically, uses that aren't loads or
4680 // stores), because arm allows folding a scale into many arithmetic
4681 // operations. This should be made more precise and revisited later.
4683 // Allow r << imm, but the imm has to be a multiple of two.
4684 if (Scale & 1) return false;
4685 return isPowerOf2_32(Scale);
4689 /// isLegalAddressingMode - Return true if the addressing mode represented
4690 /// by AM is legal for this target, for a load/store of the specified type.
4691 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4692 const Type *Ty) const {
4693 EVT VT = getValueType(Ty, true);
4694 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4697 // Can never fold addr of global into load/store.
4702 case 0: // no scale reg, must be "r+i" or "r", or "i".
4705 if (Subtarget->isThumb1Only())
4709 // ARM doesn't support any R+R*scale+imm addr modes.
4716 if (Subtarget->isThumb2())
4717 return isLegalT2ScaledAddressingMode(AM, VT);
4719 int Scale = AM.Scale;
4720 switch (VT.getSimpleVT().SimpleTy) {
4721 default: return false;
4725 if (Scale < 0) Scale = -Scale;
4729 return isPowerOf2_32(Scale & ~1);
4733 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4738 // Note, we allow "void" uses (basically, uses that aren't loads or
4739 // stores), because arm allows folding a scale into many arithmetic
4740 // operations. This should be made more precise and revisited later.
4742 // Allow r << imm, but the imm has to be a multiple of two.
4743 if (Scale & 1) return false;
4744 return isPowerOf2_32(Scale);
4751 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4752 /// icmp immediate, that is the target has icmp instructions which can compare
4753 /// a register against the immediate without having to materialize the
4754 /// immediate into a register.
4755 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4756 if (!Subtarget->isThumb())
4757 return ARM_AM::getSOImmVal(Imm) != -1;
4758 if (Subtarget->isThumb2())
4759 return ARM_AM::getT2SOImmVal(Imm) != -1;
4760 return Imm >= 0 && Imm <= 255;
4763 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4764 bool isSEXTLoad, SDValue &Base,
4765 SDValue &Offset, bool &isInc,
4766 SelectionDAG &DAG) {
4767 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4770 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4772 Base = Ptr->getOperand(0);
4773 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4774 int RHSC = (int)RHS->getZExtValue();
4775 if (RHSC < 0 && RHSC > -256) {
4776 assert(Ptr->getOpcode() == ISD::ADD);
4778 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4782 isInc = (Ptr->getOpcode() == ISD::ADD);
4783 Offset = Ptr->getOperand(1);
4785 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4787 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4788 int RHSC = (int)RHS->getZExtValue();
4789 if (RHSC < 0 && RHSC > -0x1000) {
4790 assert(Ptr->getOpcode() == ISD::ADD);
4792 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4793 Base = Ptr->getOperand(0);
4798 if (Ptr->getOpcode() == ISD::ADD) {
4800 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4801 if (ShOpcVal != ARM_AM::no_shift) {
4802 Base = Ptr->getOperand(1);
4803 Offset = Ptr->getOperand(0);
4805 Base = Ptr->getOperand(0);
4806 Offset = Ptr->getOperand(1);
4811 isInc = (Ptr->getOpcode() == ISD::ADD);
4812 Base = Ptr->getOperand(0);
4813 Offset = Ptr->getOperand(1);
4817 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4821 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4822 bool isSEXTLoad, SDValue &Base,
4823 SDValue &Offset, bool &isInc,
4824 SelectionDAG &DAG) {
4825 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4828 Base = Ptr->getOperand(0);
4829 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4830 int RHSC = (int)RHS->getZExtValue();
4831 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4832 assert(Ptr->getOpcode() == ISD::ADD);
4834 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4836 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4837 isInc = Ptr->getOpcode() == ISD::ADD;
4838 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4846 /// getPreIndexedAddressParts - returns true by value, base pointer and
4847 /// offset pointer and addressing mode by reference if the node's address
4848 /// can be legally represented as pre-indexed load / store address.
4850 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4852 ISD::MemIndexedMode &AM,
4853 SelectionDAG &DAG) const {
4854 if (Subtarget->isThumb1Only())
4859 bool isSEXTLoad = false;
4860 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4861 Ptr = LD->getBasePtr();
4862 VT = LD->getMemoryVT();
4863 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4864 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4865 Ptr = ST->getBasePtr();
4866 VT = ST->getMemoryVT();
4871 bool isLegal = false;
4872 if (Subtarget->isThumb2())
4873 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4874 Offset, isInc, DAG);
4876 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4877 Offset, isInc, DAG);
4881 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4885 /// getPostIndexedAddressParts - returns true by value, base pointer and
4886 /// offset pointer and addressing mode by reference if this node can be
4887 /// combined with a load / store to form a post-indexed load / store.
4888 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4891 ISD::MemIndexedMode &AM,
4892 SelectionDAG &DAG) const {
4893 if (Subtarget->isThumb1Only())
4898 bool isSEXTLoad = false;
4899 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4900 VT = LD->getMemoryVT();
4901 Ptr = LD->getBasePtr();
4902 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4903 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4904 VT = ST->getMemoryVT();
4905 Ptr = ST->getBasePtr();
4910 bool isLegal = false;
4911 if (Subtarget->isThumb2())
4912 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4915 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4921 // Swap base ptr and offset to catch more post-index load / store when
4922 // it's legal. In Thumb2 mode, offset must be an immediate.
4923 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4924 !Subtarget->isThumb2())
4925 std::swap(Base, Offset);
4927 // Post-indexed load / store update the base pointer.
4932 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4936 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4940 const SelectionDAG &DAG,
4941 unsigned Depth) const {
4942 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4943 switch (Op.getOpcode()) {
4945 case ARMISD::CMOV: {
4946 // Bits are known zero/one if known on the LHS and RHS.
4947 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4948 if (KnownZero == 0 && KnownOne == 0) return;
4950 APInt KnownZeroRHS, KnownOneRHS;
4951 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4952 KnownZeroRHS, KnownOneRHS, Depth+1);
4953 KnownZero &= KnownZeroRHS;
4954 KnownOne &= KnownOneRHS;
4960 //===----------------------------------------------------------------------===//
4961 // ARM Inline Assembly Support
4962 //===----------------------------------------------------------------------===//
4964 /// getConstraintType - Given a constraint letter, return the type of
4965 /// constraint it is for this target.
4966 ARMTargetLowering::ConstraintType
4967 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4968 if (Constraint.size() == 1) {
4969 switch (Constraint[0]) {
4971 case 'l': return C_RegisterClass;
4972 case 'w': return C_RegisterClass;
4975 return TargetLowering::getConstraintType(Constraint);
4978 std::pair<unsigned, const TargetRegisterClass*>
4979 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4981 if (Constraint.size() == 1) {
4982 // GCC ARM Constraint Letters
4983 switch (Constraint[0]) {
4985 if (Subtarget->isThumb())
4986 return std::make_pair(0U, ARM::tGPRRegisterClass);
4988 return std::make_pair(0U, ARM::GPRRegisterClass);
4990 return std::make_pair(0U, ARM::GPRRegisterClass);
4993 return std::make_pair(0U, ARM::SPRRegisterClass);
4994 if (VT.getSizeInBits() == 64)
4995 return std::make_pair(0U, ARM::DPRRegisterClass);
4996 if (VT.getSizeInBits() == 128)
4997 return std::make_pair(0U, ARM::QPRRegisterClass);
5001 if (StringRef("{cc}").equals_lower(Constraint))
5002 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5004 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5007 std::vector<unsigned> ARMTargetLowering::
5008 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5010 if (Constraint.size() != 1)
5011 return std::vector<unsigned>();
5013 switch (Constraint[0]) { // GCC ARM Constraint Letters
5016 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5017 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5020 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5021 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5022 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5023 ARM::R12, ARM::LR, 0);
5026 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5027 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5028 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5029 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5030 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5031 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5032 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5033 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5034 if (VT.getSizeInBits() == 64)
5035 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5036 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5037 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5038 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5039 if (VT.getSizeInBits() == 128)
5040 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5041 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5045 return std::vector<unsigned>();
5048 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5049 /// vector. If it is invalid, don't add anything to Ops.
5050 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5052 std::vector<SDValue>&Ops,
5053 SelectionDAG &DAG) const {
5054 SDValue Result(0, 0);
5056 switch (Constraint) {
5058 case 'I': case 'J': case 'K': case 'L':
5059 case 'M': case 'N': case 'O':
5060 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5064 int64_t CVal64 = C->getSExtValue();
5065 int CVal = (int) CVal64;
5066 // None of these constraints allow values larger than 32 bits. Check
5067 // that the value fits in an int.
5071 switch (Constraint) {
5073 if (Subtarget->isThumb1Only()) {
5074 // This must be a constant between 0 and 255, for ADD
5076 if (CVal >= 0 && CVal <= 255)
5078 } else if (Subtarget->isThumb2()) {
5079 // A constant that can be used as an immediate value in a
5080 // data-processing instruction.
5081 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5084 // A constant that can be used as an immediate value in a
5085 // data-processing instruction.
5086 if (ARM_AM::getSOImmVal(CVal) != -1)
5092 if (Subtarget->isThumb()) { // FIXME thumb2
5093 // This must be a constant between -255 and -1, for negated ADD
5094 // immediates. This can be used in GCC with an "n" modifier that
5095 // prints the negated value, for use with SUB instructions. It is
5096 // not useful otherwise but is implemented for compatibility.
5097 if (CVal >= -255 && CVal <= -1)
5100 // This must be a constant between -4095 and 4095. It is not clear
5101 // what this constraint is intended for. Implemented for
5102 // compatibility with GCC.
5103 if (CVal >= -4095 && CVal <= 4095)
5109 if (Subtarget->isThumb1Only()) {
5110 // A 32-bit value where only one byte has a nonzero value. Exclude
5111 // zero to match GCC. This constraint is used by GCC internally for
5112 // constants that can be loaded with a move/shift combination.
5113 // It is not useful otherwise but is implemented for compatibility.
5114 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5116 } else if (Subtarget->isThumb2()) {
5117 // A constant whose bitwise inverse can be used as an immediate
5118 // value in a data-processing instruction. This can be used in GCC
5119 // with a "B" modifier that prints the inverted value, for use with
5120 // BIC and MVN instructions. It is not useful otherwise but is
5121 // implemented for compatibility.
5122 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5125 // A constant whose bitwise inverse can be used as an immediate
5126 // value in a data-processing instruction. This can be used in GCC
5127 // with a "B" modifier that prints the inverted value, for use with
5128 // BIC and MVN instructions. It is not useful otherwise but is
5129 // implemented for compatibility.
5130 if (ARM_AM::getSOImmVal(~CVal) != -1)
5136 if (Subtarget->isThumb1Only()) {
5137 // This must be a constant between -7 and 7,
5138 // for 3-operand ADD/SUB immediate instructions.
5139 if (CVal >= -7 && CVal < 7)
5141 } else if (Subtarget->isThumb2()) {
5142 // A constant whose negation can be used as an immediate value in a
5143 // data-processing instruction. This can be used in GCC with an "n"
5144 // modifier that prints the negated value, for use with SUB
5145 // instructions. It is not useful otherwise but is implemented for
5147 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5150 // A constant whose negation can be used as an immediate value in a
5151 // data-processing instruction. This can be used in GCC with an "n"
5152 // modifier that prints the negated value, for use with SUB
5153 // instructions. It is not useful otherwise but is implemented for
5155 if (ARM_AM::getSOImmVal(-CVal) != -1)
5161 if (Subtarget->isThumb()) { // FIXME thumb2
5162 // This must be a multiple of 4 between 0 and 1020, for
5163 // ADD sp + immediate.
5164 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5167 // A power of two or a constant between 0 and 32. This is used in
5168 // GCC for the shift amount on shifted register operands, but it is
5169 // useful in general for any shift amounts.
5170 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5176 if (Subtarget->isThumb()) { // FIXME thumb2
5177 // This must be a constant between 0 and 31, for shift amounts.
5178 if (CVal >= 0 && CVal <= 31)
5184 if (Subtarget->isThumb()) { // FIXME thumb2
5185 // This must be a multiple of 4 between -508 and 508, for
5186 // ADD/SUB sp = sp + immediate.
5187 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5192 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5196 if (Result.getNode()) {
5197 Ops.push_back(Result);
5200 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5204 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5205 // The ARM target isn't yet aware of offsets.
5209 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5210 APInt Imm = FPImm.bitcastToAPInt();
5211 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5212 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5213 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5215 // We can handle 4 bits of mantissa.
5216 // mantissa = (16+UInt(e:f:g:h))/16.
5217 if (Mantissa & 0x7ffff)
5220 if ((Mantissa & 0xf) != Mantissa)
5223 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5224 if (Exp < -3 || Exp > 4)
5226 Exp = ((Exp+3) & 0x7) ^ 4;
5228 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5231 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5232 APInt Imm = FPImm.bitcastToAPInt();
5233 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5234 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5235 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5237 // We can handle 4 bits of mantissa.
5238 // mantissa = (16+UInt(e:f:g:h))/16.
5239 if (Mantissa & 0xffffffffffffLL)
5242 if ((Mantissa & 0xf) != Mantissa)
5245 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5246 if (Exp < -3 || Exp > 4)
5248 Exp = ((Exp+3) & 0x7) ^ 4;
5250 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5253 /// isFPImmLegal - Returns true if the target can instruction select the
5254 /// specified FP immediate natively. If false, the legalizer will
5255 /// materialize the FP immediate as a load from a constant pool.
5256 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5257 if (!Subtarget->hasVFP3())
5260 return ARM::getVFPf32Imm(Imm) != -1;
5262 return ARM::getVFPf64Imm(Imm) != -1;