1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMAddressingModes.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Type.h"
33 #include "llvm/CodeGen/CallingConvLower.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineInstrBuilder.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/CodeGen/PseudoSourceValue.h"
40 #include "llvm/CodeGen/SelectionDAG.h"
41 #include "llvm/MC/MCSectionMachO.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/ADT/VectorExtras.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
52 STATISTIC(NumTailCalls, "Number of tail calls");
54 // This option should go away when tail calls fully work.
56 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
57 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions."),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
70 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
71 CCValAssign::LocInfo &LocInfo,
72 ISD::ArgFlagsTy &ArgFlags,
74 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
75 CCValAssign::LocInfo &LocInfo,
76 ISD::ArgFlagsTy &ArgFlags,
78 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
79 CCValAssign::LocInfo &LocInfo,
80 ISD::ArgFlagsTy &ArgFlags,
82 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
83 CCValAssign::LocInfo &LocInfo,
84 ISD::ArgFlagsTy &ArgFlags,
87 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
88 EVT PromotedBitwiseVT) {
89 if (VT != PromotedLdStVT) {
90 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
91 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
92 PromotedLdStVT.getSimpleVT());
94 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
95 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
96 PromotedLdStVT.getSimpleVT());
99 EVT ElemTy = VT.getVectorElementType();
100 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
101 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
102 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
103 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
104 if (ElemTy != MVT::i32) {
105 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
106 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
108 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
110 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
112 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
113 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
116 if (VT.isInteger()) {
117 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
119 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
122 // Promote all bit-wise operations.
123 if (VT.isInteger() && VT != PromotedBitwiseVT) {
124 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
125 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
126 PromotedBitwiseVT.getSimpleVT());
127 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
128 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
129 PromotedBitwiseVT.getSimpleVT());
130 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
131 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
132 PromotedBitwiseVT.getSimpleVT());
135 // Neon does not support vector divide/remainder operations.
136 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
137 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
138 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
139 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
141 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
144 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
145 addRegisterClass(VT, ARM::DPRRegisterClass);
146 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
149 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
150 addRegisterClass(VT, ARM::QPRRegisterClass);
151 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
154 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
155 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
156 return new TargetLoweringObjectFileMachO();
158 return new ARMElfTargetObjectFile();
161 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
162 : TargetLowering(TM, createTLOF(TM)) {
163 Subtarget = &TM.getSubtarget<ARMSubtarget>();
165 if (Subtarget->isTargetDarwin()) {
166 // Uses VFP for Thumb libfuncs if available.
167 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
168 // Single-precision floating-point arithmetic.
169 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
170 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
171 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
172 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
174 // Double-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
176 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
177 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
178 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
180 // Single-precision comparisons.
181 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
182 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
183 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
184 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
185 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
186 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
187 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
188 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
190 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
199 // Double-precision comparisons.
200 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
201 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
202 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
203 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
204 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
205 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
206 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
207 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
209 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
218 // Floating-point to integer conversions.
219 // i64 conversions are done via library routines even when generating VFP
220 // instructions, so use the same ones.
221 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
222 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
223 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
224 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
226 // Conversions between floating types.
227 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
228 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
230 // Integer to floating-point conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
234 // e.g., __floatunsidf vs. __floatunssidfvfp.
235 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
236 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
237 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
238 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
242 // These libcalls are not available in 32-bit.
243 setLibcallName(RTLIB::SHL_I128, 0);
244 setLibcallName(RTLIB::SRL_I128, 0);
245 setLibcallName(RTLIB::SRA_I128, 0);
247 // Libcalls should use the AAPCS base standard ABI, even if hard float
248 // is in effect, as per the ARM RTABI specification, section 4.1.2.
249 if (Subtarget->isAAPCS_ABI()) {
250 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
251 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
252 CallingConv::ARM_AAPCS);
256 if (Subtarget->isThumb1Only())
257 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
259 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
260 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
261 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
262 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
264 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
267 if (Subtarget->hasNEON()) {
268 addDRTypeForNEON(MVT::v2f32);
269 addDRTypeForNEON(MVT::v8i8);
270 addDRTypeForNEON(MVT::v4i16);
271 addDRTypeForNEON(MVT::v2i32);
272 addDRTypeForNEON(MVT::v1i64);
274 addQRTypeForNEON(MVT::v4f32);
275 addQRTypeForNEON(MVT::v2f64);
276 addQRTypeForNEON(MVT::v16i8);
277 addQRTypeForNEON(MVT::v8i16);
278 addQRTypeForNEON(MVT::v4i32);
279 addQRTypeForNEON(MVT::v2i64);
281 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
282 // neither Neon nor VFP support any arithmetic operations on it.
283 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
284 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
285 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
286 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
287 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
288 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
289 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
290 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
291 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
292 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
293 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
294 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
295 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
296 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
297 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
298 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
299 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
300 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
301 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
302 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
303 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
304 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
305 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
306 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
308 // Neon does not support some operations on v1i64 and v2i64 types.
309 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
310 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
311 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
312 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
314 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
315 setTargetDAGCombine(ISD::SHL);
316 setTargetDAGCombine(ISD::SRL);
317 setTargetDAGCombine(ISD::SRA);
318 setTargetDAGCombine(ISD::SIGN_EXTEND);
319 setTargetDAGCombine(ISD::ZERO_EXTEND);
320 setTargetDAGCombine(ISD::ANY_EXTEND);
321 setTargetDAGCombine(ISD::SELECT_CC);
324 computeRegisterProperties();
326 // ARM does not have f32 extending load.
327 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
329 // ARM does not have i1 sign extending load.
330 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
332 // ARM supports all 4 flavors of integer indexed load / store.
333 if (!Subtarget->isThumb1Only()) {
334 for (unsigned im = (unsigned)ISD::PRE_INC;
335 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
336 setIndexedLoadAction(im, MVT::i1, Legal);
337 setIndexedLoadAction(im, MVT::i8, Legal);
338 setIndexedLoadAction(im, MVT::i16, Legal);
339 setIndexedLoadAction(im, MVT::i32, Legal);
340 setIndexedStoreAction(im, MVT::i1, Legal);
341 setIndexedStoreAction(im, MVT::i8, Legal);
342 setIndexedStoreAction(im, MVT::i16, Legal);
343 setIndexedStoreAction(im, MVT::i32, Legal);
347 // i64 operation support.
348 if (Subtarget->isThumb1Only()) {
349 setOperationAction(ISD::MUL, MVT::i64, Expand);
350 setOperationAction(ISD::MULHU, MVT::i32, Expand);
351 setOperationAction(ISD::MULHS, MVT::i32, Expand);
352 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
353 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
355 setOperationAction(ISD::MUL, MVT::i64, Expand);
356 setOperationAction(ISD::MULHU, MVT::i32, Expand);
357 if (!Subtarget->hasV6Ops())
358 setOperationAction(ISD::MULHS, MVT::i32, Expand);
360 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
361 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
362 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
363 setOperationAction(ISD::SRL, MVT::i64, Custom);
364 setOperationAction(ISD::SRA, MVT::i64, Custom);
366 // ARM does not have ROTL.
367 setOperationAction(ISD::ROTL, MVT::i32, Expand);
368 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
369 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
370 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
371 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
373 // Only ARMv6 has BSWAP.
374 if (!Subtarget->hasV6Ops())
375 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
377 // These are expanded into libcalls.
378 if (!Subtarget->hasDivide()) {
379 // v7M has a hardware divider
380 setOperationAction(ISD::SDIV, MVT::i32, Expand);
381 setOperationAction(ISD::UDIV, MVT::i32, Expand);
383 setOperationAction(ISD::SREM, MVT::i32, Expand);
384 setOperationAction(ISD::UREM, MVT::i32, Expand);
385 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
386 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
388 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
389 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
390 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
391 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
392 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
394 setOperationAction(ISD::TRAP, MVT::Other, Legal);
396 // Use the default implementation.
397 setOperationAction(ISD::VASTART, MVT::Other, Custom);
398 setOperationAction(ISD::VAARG, MVT::Other, Expand);
399 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
400 setOperationAction(ISD::VAEND, MVT::Other, Expand);
401 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
402 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
403 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
404 // FIXME: Shouldn't need this, since no register is used, but the legalizer
405 // doesn't yet know how to not do that for SjLj.
406 setExceptionSelectorRegister(ARM::R0);
407 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
408 // Handle atomics directly for ARMv[67] (except for Thumb1), otherwise
409 // use the default expansion.
410 bool canHandleAtomics =
411 (Subtarget->hasV7Ops() ||
412 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only()));
413 if (canHandleAtomics) {
414 // membarrier needs custom lowering; the rest are legal and handled
416 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
418 // Set them all for expansion, which will force libcalls.
419 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
420 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
421 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
422 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
423 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
424 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
425 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
426 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
427 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
428 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
429 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
430 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
431 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
432 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
445 // 64-bit versions are always libcalls (for now)
446 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
447 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
449 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
450 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
451 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
452 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
453 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
455 // If the subtarget does not have extract instructions, sign_extend_inreg
456 // needs to be expanded. Extract is available in ARM mode on v6 and up,
457 // and on most Thumb2 implementations.
458 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
459 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
460 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
461 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
463 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
465 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
466 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
467 // iff target supports vfp2.
468 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
470 // We want to custom lower some of our intrinsics.
471 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
473 setOperationAction(ISD::SETCC, MVT::i32, Expand);
474 setOperationAction(ISD::SETCC, MVT::f32, Expand);
475 setOperationAction(ISD::SETCC, MVT::f64, Expand);
476 setOperationAction(ISD::SELECT, MVT::i32, Expand);
477 setOperationAction(ISD::SELECT, MVT::f32, Expand);
478 setOperationAction(ISD::SELECT, MVT::f64, Expand);
479 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
480 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
481 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
483 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
484 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
485 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
486 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
487 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
489 // We don't support sin/cos/fmod/copysign/pow
490 setOperationAction(ISD::FSIN, MVT::f64, Expand);
491 setOperationAction(ISD::FSIN, MVT::f32, Expand);
492 setOperationAction(ISD::FCOS, MVT::f32, Expand);
493 setOperationAction(ISD::FCOS, MVT::f64, Expand);
494 setOperationAction(ISD::FREM, MVT::f64, Expand);
495 setOperationAction(ISD::FREM, MVT::f32, Expand);
496 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
497 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
498 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
500 setOperationAction(ISD::FPOW, MVT::f64, Expand);
501 setOperationAction(ISD::FPOW, MVT::f32, Expand);
503 // Various VFP goodness
504 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
505 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
506 if (Subtarget->hasVFP2()) {
507 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
508 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
509 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
510 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
512 // Special handling for half-precision FP.
513 if (!Subtarget->hasFP16()) {
514 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
515 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
519 // We have target-specific dag combine patterns for the following nodes:
520 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
521 setTargetDAGCombine(ISD::ADD);
522 setTargetDAGCombine(ISD::SUB);
523 setTargetDAGCombine(ISD::MUL);
525 setStackPointerRegisterToSaveRestore(ARM::SP);
527 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
528 setSchedulingPreference(Sched::RegPressure);
530 setSchedulingPreference(Sched::Hybrid);
532 // FIXME: If-converter should use instruction latency to determine
533 // profitability rather than relying on fixed limits.
534 if (Subtarget->getCPUString() == "generic") {
535 // Generic (and overly aggressive) if-conversion limits.
536 setIfCvtBlockSizeLimit(10);
537 setIfCvtDupBlockSizeLimit(2);
538 } else if (Subtarget->hasV7Ops()) {
539 setIfCvtBlockSizeLimit(3);
540 setIfCvtDupBlockSizeLimit(1);
541 } else if (Subtarget->hasV6Ops()) {
542 setIfCvtBlockSizeLimit(2);
543 setIfCvtDupBlockSizeLimit(1);
545 setIfCvtBlockSizeLimit(3);
546 setIfCvtDupBlockSizeLimit(2);
549 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
550 // Do not enable CodePlacementOpt for now: it currently runs after the
551 // ARMConstantIslandPass and messes up branch relaxation and placement
552 // of constant islands.
553 // benefitFromCodePlacementOpt = true;
556 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
559 case ARMISD::Wrapper: return "ARMISD::Wrapper";
560 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
561 case ARMISD::CALL: return "ARMISD::CALL";
562 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
563 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
564 case ARMISD::tCALL: return "ARMISD::tCALL";
565 case ARMISD::BRCOND: return "ARMISD::BRCOND";
566 case ARMISD::BR_JT: return "ARMISD::BR_JT";
567 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
568 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
569 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
570 case ARMISD::CMP: return "ARMISD::CMP";
571 case ARMISD::CMPZ: return "ARMISD::CMPZ";
572 case ARMISD::CMPFP: return "ARMISD::CMPFP";
573 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
574 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
575 case ARMISD::CMOV: return "ARMISD::CMOV";
576 case ARMISD::CNEG: return "ARMISD::CNEG";
578 case ARMISD::RBIT: return "ARMISD::RBIT";
580 case ARMISD::FTOSI: return "ARMISD::FTOSI";
581 case ARMISD::FTOUI: return "ARMISD::FTOUI";
582 case ARMISD::SITOF: return "ARMISD::SITOF";
583 case ARMISD::UITOF: return "ARMISD::UITOF";
585 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
586 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
587 case ARMISD::RRX: return "ARMISD::RRX";
589 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
590 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
592 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
593 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
595 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
597 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
599 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
601 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
602 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
604 case ARMISD::VCEQ: return "ARMISD::VCEQ";
605 case ARMISD::VCGE: return "ARMISD::VCGE";
606 case ARMISD::VCGEU: return "ARMISD::VCGEU";
607 case ARMISD::VCGT: return "ARMISD::VCGT";
608 case ARMISD::VCGTU: return "ARMISD::VCGTU";
609 case ARMISD::VTST: return "ARMISD::VTST";
611 case ARMISD::VSHL: return "ARMISD::VSHL";
612 case ARMISD::VSHRs: return "ARMISD::VSHRs";
613 case ARMISD::VSHRu: return "ARMISD::VSHRu";
614 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
615 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
616 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
617 case ARMISD::VSHRN: return "ARMISD::VSHRN";
618 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
619 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
620 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
621 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
622 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
623 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
624 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
625 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
626 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
627 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
628 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
629 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
630 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
631 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
632 case ARMISD::VDUP: return "ARMISD::VDUP";
633 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
634 case ARMISD::VEXT: return "ARMISD::VEXT";
635 case ARMISD::VREV64: return "ARMISD::VREV64";
636 case ARMISD::VREV32: return "ARMISD::VREV32";
637 case ARMISD::VREV16: return "ARMISD::VREV16";
638 case ARMISD::VZIP: return "ARMISD::VZIP";
639 case ARMISD::VUZP: return "ARMISD::VUZP";
640 case ARMISD::VTRN: return "ARMISD::VTRN";
641 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
642 case ARMISD::FMAX: return "ARMISD::FMAX";
643 case ARMISD::FMIN: return "ARMISD::FMIN";
647 /// getRegClassFor - Return the register class that should be used for the
648 /// specified value type.
649 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
650 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
651 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
652 // load / store 4 to 8 consecutive D registers.
653 if (Subtarget->hasNEON()) {
654 if (VT == MVT::v4i64)
655 return ARM::QQPRRegisterClass;
656 else if (VT == MVT::v8i64)
657 return ARM::QQQQPRRegisterClass;
659 return TargetLowering::getRegClassFor(VT);
662 /// getFunctionAlignment - Return the Log2 alignment of this function.
663 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
664 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
667 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
668 unsigned NumVals = N->getNumValues();
670 return Sched::RegPressure;
672 for (unsigned i = 0; i != NumVals; ++i) {
673 EVT VT = N->getValueType(i);
674 if (VT.isFloatingPoint() || VT.isVector())
675 return Sched::Latency;
678 if (!N->isMachineOpcode())
679 return Sched::RegPressure;
681 // Load are scheduled for latency even if there instruction itinerary
683 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
684 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
686 return Sched::Latency;
688 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
689 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
690 return Sched::Latency;
691 return Sched::RegPressure;
694 //===----------------------------------------------------------------------===//
696 //===----------------------------------------------------------------------===//
698 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
699 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
701 default: llvm_unreachable("Unknown condition code!");
702 case ISD::SETNE: return ARMCC::NE;
703 case ISD::SETEQ: return ARMCC::EQ;
704 case ISD::SETGT: return ARMCC::GT;
705 case ISD::SETGE: return ARMCC::GE;
706 case ISD::SETLT: return ARMCC::LT;
707 case ISD::SETLE: return ARMCC::LE;
708 case ISD::SETUGT: return ARMCC::HI;
709 case ISD::SETUGE: return ARMCC::HS;
710 case ISD::SETULT: return ARMCC::LO;
711 case ISD::SETULE: return ARMCC::LS;
715 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
716 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
717 ARMCC::CondCodes &CondCode2) {
718 CondCode2 = ARMCC::AL;
720 default: llvm_unreachable("Unknown FP condition!");
722 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
724 case ISD::SETOGT: CondCode = ARMCC::GT; break;
726 case ISD::SETOGE: CondCode = ARMCC::GE; break;
727 case ISD::SETOLT: CondCode = ARMCC::MI; break;
728 case ISD::SETOLE: CondCode = ARMCC::LS; break;
729 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
730 case ISD::SETO: CondCode = ARMCC::VC; break;
731 case ISD::SETUO: CondCode = ARMCC::VS; break;
732 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
733 case ISD::SETUGT: CondCode = ARMCC::HI; break;
734 case ISD::SETUGE: CondCode = ARMCC::PL; break;
736 case ISD::SETULT: CondCode = ARMCC::LT; break;
738 case ISD::SETULE: CondCode = ARMCC::LE; break;
740 case ISD::SETUNE: CondCode = ARMCC::NE; break;
744 //===----------------------------------------------------------------------===//
745 // Calling Convention Implementation
746 //===----------------------------------------------------------------------===//
748 #include "ARMGenCallingConv.inc"
750 // APCS f64 is in register pairs, possibly split to stack
751 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
752 CCValAssign::LocInfo &LocInfo,
753 CCState &State, bool CanFail) {
754 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
756 // Try to get the first register.
757 if (unsigned Reg = State.AllocateReg(RegList, 4))
758 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
760 // For the 2nd half of a v2f64, do not fail.
764 // Put the whole thing on the stack.
765 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
766 State.AllocateStack(8, 4),
771 // Try to get the second register.
772 if (unsigned Reg = State.AllocateReg(RegList, 4))
773 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
775 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
776 State.AllocateStack(4, 4),
781 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
782 CCValAssign::LocInfo &LocInfo,
783 ISD::ArgFlagsTy &ArgFlags,
785 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
787 if (LocVT == MVT::v2f64 &&
788 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
790 return true; // we handled it
793 // AAPCS f64 is in aligned register pairs
794 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
795 CCValAssign::LocInfo &LocInfo,
796 CCState &State, bool CanFail) {
797 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
798 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
800 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
802 // For the 2nd half of a v2f64, do not just fail.
806 // Put the whole thing on the stack.
807 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
808 State.AllocateStack(8, 8),
814 for (i = 0; i < 2; ++i)
815 if (HiRegList[i] == Reg)
818 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
819 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
824 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
825 CCValAssign::LocInfo &LocInfo,
826 ISD::ArgFlagsTy &ArgFlags,
828 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
830 if (LocVT == MVT::v2f64 &&
831 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
833 return true; // we handled it
836 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
837 CCValAssign::LocInfo &LocInfo, CCState &State) {
838 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
839 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
841 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
843 return false; // we didn't handle it
846 for (i = 0; i < 2; ++i)
847 if (HiRegList[i] == Reg)
850 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
851 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
856 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
857 CCValAssign::LocInfo &LocInfo,
858 ISD::ArgFlagsTy &ArgFlags,
860 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
862 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
864 return true; // we handled it
867 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
868 CCValAssign::LocInfo &LocInfo,
869 ISD::ArgFlagsTy &ArgFlags,
871 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
875 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
876 /// given CallingConvention value.
877 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
879 bool isVarArg) const {
882 llvm_unreachable("Unsupported calling convention");
884 case CallingConv::Fast:
885 // Use target triple & subtarget features to do actual dispatch.
886 if (Subtarget->isAAPCS_ABI()) {
887 if (Subtarget->hasVFP2() &&
888 FloatABIType == FloatABI::Hard && !isVarArg)
889 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
891 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
893 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
894 case CallingConv::ARM_AAPCS_VFP:
895 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
896 case CallingConv::ARM_AAPCS:
897 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
898 case CallingConv::ARM_APCS:
899 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
903 /// LowerCallResult - Lower the result values of a call into the
904 /// appropriate copies out of appropriate physical registers.
906 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
907 CallingConv::ID CallConv, bool isVarArg,
908 const SmallVectorImpl<ISD::InputArg> &Ins,
909 DebugLoc dl, SelectionDAG &DAG,
910 SmallVectorImpl<SDValue> &InVals) const {
912 // Assign locations to each value returned by this call.
913 SmallVector<CCValAssign, 16> RVLocs;
914 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
915 RVLocs, *DAG.getContext());
916 CCInfo.AnalyzeCallResult(Ins,
917 CCAssignFnForNode(CallConv, /* Return*/ true,
920 // Copy all of the result registers out of their specified physreg.
921 for (unsigned i = 0; i != RVLocs.size(); ++i) {
922 CCValAssign VA = RVLocs[i];
925 if (VA.needsCustom()) {
926 // Handle f64 or half of a v2f64.
927 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
929 Chain = Lo.getValue(1);
930 InFlag = Lo.getValue(2);
931 VA = RVLocs[++i]; // skip ahead to next loc
932 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
934 Chain = Hi.getValue(1);
935 InFlag = Hi.getValue(2);
936 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
938 if (VA.getLocVT() == MVT::v2f64) {
939 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
940 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
941 DAG.getConstant(0, MVT::i32));
943 VA = RVLocs[++i]; // skip ahead to next loc
944 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
945 Chain = Lo.getValue(1);
946 InFlag = Lo.getValue(2);
947 VA = RVLocs[++i]; // skip ahead to next loc
948 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
949 Chain = Hi.getValue(1);
950 InFlag = Hi.getValue(2);
951 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
952 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
953 DAG.getConstant(1, MVT::i32));
956 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
958 Chain = Val.getValue(1);
959 InFlag = Val.getValue(2);
962 switch (VA.getLocInfo()) {
963 default: llvm_unreachable("Unknown loc info!");
964 case CCValAssign::Full: break;
965 case CCValAssign::BCvt:
966 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
970 InVals.push_back(Val);
976 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
977 /// by "Src" to address "Dst" of size "Size". Alignment information is
978 /// specified by the specific parameter attribute. The copy will be passed as
979 /// a byval function parameter.
980 /// Sometimes what we are copying is the end of a larger object, the part that
981 /// does not fit in registers.
983 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
984 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
986 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
987 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
988 /*isVolatile=*/false, /*AlwaysInline=*/false,
992 /// LowerMemOpCallTo - Store the argument to the stack.
994 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
995 SDValue StackPtr, SDValue Arg,
996 DebugLoc dl, SelectionDAG &DAG,
997 const CCValAssign &VA,
998 ISD::ArgFlagsTy Flags) const {
999 unsigned LocMemOffset = VA.getLocMemOffset();
1000 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1001 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1002 if (Flags.isByVal()) {
1003 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1005 return DAG.getStore(Chain, dl, Arg, PtrOff,
1006 PseudoSourceValue::getStack(), LocMemOffset,
1010 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1011 SDValue Chain, SDValue &Arg,
1012 RegsToPassVector &RegsToPass,
1013 CCValAssign &VA, CCValAssign &NextVA,
1015 SmallVector<SDValue, 8> &MemOpChains,
1016 ISD::ArgFlagsTy Flags) const {
1018 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1019 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1020 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1022 if (NextVA.isRegLoc())
1023 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1025 assert(NextVA.isMemLoc());
1026 if (StackPtr.getNode() == 0)
1027 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1029 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1035 /// LowerCall - Lowering a call into a callseq_start <-
1036 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1039 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1040 CallingConv::ID CallConv, bool isVarArg,
1042 const SmallVectorImpl<ISD::OutputArg> &Outs,
1043 const SmallVectorImpl<ISD::InputArg> &Ins,
1044 DebugLoc dl, SelectionDAG &DAG,
1045 SmallVectorImpl<SDValue> &InVals) const {
1046 MachineFunction &MF = DAG.getMachineFunction();
1047 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1048 bool IsSibCall = false;
1049 // Temporarily disable tail calls so things don't break.
1050 if (!EnableARMTailCalls)
1053 // Check if it's really possible to do a tail call.
1054 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1055 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1057 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1058 // detected sibcalls.
1065 // Analyze operands of the call, assigning locations to each operand.
1066 SmallVector<CCValAssign, 16> ArgLocs;
1067 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1069 CCInfo.AnalyzeCallOperands(Outs,
1070 CCAssignFnForNode(CallConv, /* Return*/ false,
1073 // Get a count of how many bytes are to be pushed on the stack.
1074 unsigned NumBytes = CCInfo.getNextStackOffset();
1076 // For tail calls, memory operands are available in our caller's stack.
1080 // Adjust the stack pointer for the new arguments...
1081 // These operations are automatically eliminated by the prolog/epilog pass
1083 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1085 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1087 RegsToPassVector RegsToPass;
1088 SmallVector<SDValue, 8> MemOpChains;
1090 // Walk the register/memloc assignments, inserting copies/loads. In the case
1091 // of tail call optimization, arguments are handled later.
1092 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1094 ++i, ++realArgIdx) {
1095 CCValAssign &VA = ArgLocs[i];
1096 SDValue Arg = Outs[realArgIdx].Val;
1097 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1099 // Promote the value if needed.
1100 switch (VA.getLocInfo()) {
1101 default: llvm_unreachable("Unknown loc info!");
1102 case CCValAssign::Full: break;
1103 case CCValAssign::SExt:
1104 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1106 case CCValAssign::ZExt:
1107 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1109 case CCValAssign::AExt:
1110 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1112 case CCValAssign::BCvt:
1113 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1117 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1118 if (VA.needsCustom()) {
1119 if (VA.getLocVT() == MVT::v2f64) {
1120 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1121 DAG.getConstant(0, MVT::i32));
1122 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1123 DAG.getConstant(1, MVT::i32));
1125 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1126 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1128 VA = ArgLocs[++i]; // skip ahead to next loc
1129 if (VA.isRegLoc()) {
1130 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1131 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1133 assert(VA.isMemLoc());
1135 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1136 dl, DAG, VA, Flags));
1139 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1140 StackPtr, MemOpChains, Flags);
1142 } else if (VA.isRegLoc()) {
1143 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1144 } else if (!IsSibCall) {
1145 assert(VA.isMemLoc());
1147 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1148 dl, DAG, VA, Flags));
1152 if (!MemOpChains.empty())
1153 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1154 &MemOpChains[0], MemOpChains.size());
1156 // Build a sequence of copy-to-reg nodes chained together with token chain
1157 // and flag operands which copy the outgoing args into the appropriate regs.
1159 // Tail call byval lowering might overwrite argument registers so in case of
1160 // tail call optimization the copies to registers are lowered later.
1162 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1163 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1164 RegsToPass[i].second, InFlag);
1165 InFlag = Chain.getValue(1);
1168 // For tail calls lower the arguments to the 'real' stack slot.
1170 // Force all the incoming stack arguments to be loaded from the stack
1171 // before any new outgoing arguments are stored to the stack, because the
1172 // outgoing stack slots may alias the incoming argument stack slots, and
1173 // the alias isn't otherwise explicit. This is slightly more conservative
1174 // than necessary, because it means that each store effectively depends
1175 // on every argument instead of just those arguments it would clobber.
1177 // Do not flag preceeding copytoreg stuff together with the following stuff.
1179 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1180 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1181 RegsToPass[i].second, InFlag);
1182 InFlag = Chain.getValue(1);
1187 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1188 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1189 // node so that legalize doesn't hack it.
1190 bool isDirect = false;
1191 bool isARMFunc = false;
1192 bool isLocalARMFunc = false;
1193 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1195 if (EnableARMLongCalls) {
1196 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1197 && "long-calls with non-static relocation model!");
1198 // Handle a global address or an external symbol. If it's not one of
1199 // those, the target's already in a register, so we don't need to do
1201 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1202 const GlobalValue *GV = G->getGlobal();
1203 // Create a constant pool entry for the callee address
1204 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1205 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1208 // Get the address of the callee into a register
1209 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1210 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1211 Callee = DAG.getLoad(getPointerTy(), dl,
1212 DAG.getEntryNode(), CPAddr,
1213 PseudoSourceValue::getConstantPool(), 0,
1215 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1216 const char *Sym = S->getSymbol();
1218 // Create a constant pool entry for the callee address
1219 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1220 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1221 Sym, ARMPCLabelIndex, 0);
1222 // Get the address of the callee into a register
1223 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1224 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1225 Callee = DAG.getLoad(getPointerTy(), dl,
1226 DAG.getEntryNode(), CPAddr,
1227 PseudoSourceValue::getConstantPool(), 0,
1230 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1231 const GlobalValue *GV = G->getGlobal();
1233 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1234 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1235 getTargetMachine().getRelocationModel() != Reloc::Static;
1236 isARMFunc = !Subtarget->isThumb() || isStub;
1237 // ARM call to a local ARM function is predicable.
1238 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1239 // tBX takes a register source operand.
1240 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1241 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1242 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1245 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1246 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1247 Callee = DAG.getLoad(getPointerTy(), dl,
1248 DAG.getEntryNode(), CPAddr,
1249 PseudoSourceValue::getConstantPool(), 0,
1251 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1252 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1253 getPointerTy(), Callee, PICLabel);
1255 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1256 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1258 bool isStub = Subtarget->isTargetDarwin() &&
1259 getTargetMachine().getRelocationModel() != Reloc::Static;
1260 isARMFunc = !Subtarget->isThumb() || isStub;
1261 // tBX takes a register source operand.
1262 const char *Sym = S->getSymbol();
1263 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1264 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1265 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1266 Sym, ARMPCLabelIndex, 4);
1267 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1268 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1269 Callee = DAG.getLoad(getPointerTy(), dl,
1270 DAG.getEntryNode(), CPAddr,
1271 PseudoSourceValue::getConstantPool(), 0,
1273 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1274 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1275 getPointerTy(), Callee, PICLabel);
1277 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1280 // FIXME: handle tail calls differently.
1282 if (Subtarget->isThumb()) {
1283 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1284 CallOpc = ARMISD::CALL_NOLINK;
1286 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1288 CallOpc = (isDirect || Subtarget->hasV5TOps())
1289 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1290 : ARMISD::CALL_NOLINK;
1292 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1293 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1294 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1295 InFlag = Chain.getValue(1);
1298 std::vector<SDValue> Ops;
1299 Ops.push_back(Chain);
1300 Ops.push_back(Callee);
1302 // Add argument registers to the end of the list so that they are known live
1304 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1305 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1306 RegsToPass[i].second.getValueType()));
1308 if (InFlag.getNode())
1309 Ops.push_back(InFlag);
1311 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1313 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1315 // Returns a chain and a flag for retval copy to use.
1316 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1317 InFlag = Chain.getValue(1);
1319 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1320 DAG.getIntPtrConstant(0, true), InFlag);
1322 InFlag = Chain.getValue(1);
1324 // Handle result values, copying them out of physregs into vregs that we
1326 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1330 /// MatchingStackOffset - Return true if the given stack call argument is
1331 /// already available in the same position (relatively) of the caller's
1332 /// incoming argument stack.
1334 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1335 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1336 const ARMInstrInfo *TII) {
1337 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1339 if (Arg.getOpcode() == ISD::CopyFromReg) {
1340 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1341 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1343 MachineInstr *Def = MRI->getVRegDef(VR);
1346 if (!Flags.isByVal()) {
1347 if (!TII->isLoadFromStackSlot(Def, FI))
1350 // unsigned Opcode = Def->getOpcode();
1351 // if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
1352 // Def->getOperand(1).isFI()) {
1353 // FI = Def->getOperand(1).getIndex();
1354 // Bytes = Flags.getByValSize();
1358 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1359 if (Flags.isByVal())
1360 // ByVal argument is passed in as a pointer but it's now being
1361 // dereferenced. e.g.
1362 // define @foo(%struct.X* %A) {
1363 // tail call @bar(%struct.X* byval %A)
1366 SDValue Ptr = Ld->getBasePtr();
1367 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1370 FI = FINode->getIndex();
1374 assert(FI != INT_MAX);
1375 if (!MFI->isFixedObjectIndex(FI))
1377 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1380 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1381 /// for tail call optimization. Targets which want to do tail call
1382 /// optimization should implement this function.
1384 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1385 CallingConv::ID CalleeCC,
1387 bool isCalleeStructRet,
1388 bool isCallerStructRet,
1389 const SmallVectorImpl<ISD::OutputArg> &Outs,
1390 const SmallVectorImpl<ISD::InputArg> &Ins,
1391 SelectionDAG& DAG) const {
1393 const Function *CallerF = DAG.getMachineFunction().getFunction();
1394 CallingConv::ID CallerCC = CallerF->getCallingConv();
1395 bool CCMatch = CallerCC == CalleeCC;
1397 // Look for obvious safe cases to perform tail call optimization that do not
1398 // require ABI changes. This is what gcc calls sibcall.
1400 // Do not sibcall optimize vararg calls unless the call site is not passing
1402 if (isVarArg && !Outs.empty())
1405 // Also avoid sibcall optimization if either caller or callee uses struct
1406 // return semantics.
1407 if (isCalleeStructRet || isCallerStructRet)
1410 // On Thumb, for the moment, we can only do this to functions defined in this
1411 // compilation, or to indirect calls. A Thumb B to an ARM function is not
1412 // easily fixed up in the linker, unlike BL.
1413 if (Subtarget->isThumb()) {
1414 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1415 const GlobalValue *GV = G->getGlobal();
1416 if (GV->isDeclaration() || GV->isWeakForLinker())
1418 } else if (isa<ExternalSymbolSDNode>(Callee)) {
1423 // If the calling conventions do not match, then we'd better make sure the
1424 // results are returned in the same way as what the caller expects.
1426 SmallVector<CCValAssign, 16> RVLocs1;
1427 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1428 RVLocs1, *DAG.getContext());
1429 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1431 SmallVector<CCValAssign, 16> RVLocs2;
1432 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1433 RVLocs2, *DAG.getContext());
1434 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1436 if (RVLocs1.size() != RVLocs2.size())
1438 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1439 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1441 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1443 if (RVLocs1[i].isRegLoc()) {
1444 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1447 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1453 // If the callee takes no arguments then go on to check the results of the
1455 if (!Outs.empty()) {
1456 // Check if stack adjustment is needed. For now, do not do this if any
1457 // argument is passed on the stack.
1458 SmallVector<CCValAssign, 16> ArgLocs;
1459 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1460 ArgLocs, *DAG.getContext());
1461 CCInfo.AnalyzeCallOperands(Outs,
1462 CCAssignFnForNode(CalleeCC, false, isVarArg));
1463 if (CCInfo.getNextStackOffset()) {
1464 MachineFunction &MF = DAG.getMachineFunction();
1466 // Check if the arguments are already laid out in the right way as
1467 // the caller's fixed stack objects.
1468 MachineFrameInfo *MFI = MF.getFrameInfo();
1469 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1470 const ARMInstrInfo *TII =
1471 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1472 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1474 ++i, ++realArgIdx) {
1475 CCValAssign &VA = ArgLocs[i];
1476 EVT RegVT = VA.getLocVT();
1477 SDValue Arg = Outs[realArgIdx].Val;
1478 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1479 if (VA.getLocInfo() == CCValAssign::Indirect)
1481 if (VA.needsCustom()) {
1482 // f64 and vector types are split into multiple registers or
1483 // register/stack-slot combinations. The types will not match
1484 // the registers; give up on memory f64 refs until we figure
1485 // out what to do about this.
1488 if (!ArgLocs[++i].isRegLoc())
1490 if (RegVT == MVT::v2f64) {
1491 if (!ArgLocs[++i].isRegLoc())
1493 if (!ArgLocs[++i].isRegLoc())
1496 } else if (!VA.isRegLoc()) {
1497 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1509 ARMTargetLowering::LowerReturn(SDValue Chain,
1510 CallingConv::ID CallConv, bool isVarArg,
1511 const SmallVectorImpl<ISD::OutputArg> &Outs,
1512 DebugLoc dl, SelectionDAG &DAG) const {
1514 // CCValAssign - represent the assignment of the return value to a location.
1515 SmallVector<CCValAssign, 16> RVLocs;
1517 // CCState - Info about the registers and stack slots.
1518 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1521 // Analyze outgoing return values.
1522 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1525 // If this is the first return lowered for this function, add
1526 // the regs to the liveout set for the function.
1527 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1528 for (unsigned i = 0; i != RVLocs.size(); ++i)
1529 if (RVLocs[i].isRegLoc())
1530 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1535 // Copy the result values into the output registers.
1536 for (unsigned i = 0, realRVLocIdx = 0;
1538 ++i, ++realRVLocIdx) {
1539 CCValAssign &VA = RVLocs[i];
1540 assert(VA.isRegLoc() && "Can only return in registers!");
1542 SDValue Arg = Outs[realRVLocIdx].Val;
1544 switch (VA.getLocInfo()) {
1545 default: llvm_unreachable("Unknown loc info!");
1546 case CCValAssign::Full: break;
1547 case CCValAssign::BCvt:
1548 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1552 if (VA.needsCustom()) {
1553 if (VA.getLocVT() == MVT::v2f64) {
1554 // Extract the first half and return it in two registers.
1555 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1556 DAG.getConstant(0, MVT::i32));
1557 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1558 DAG.getVTList(MVT::i32, MVT::i32), Half);
1560 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1561 Flag = Chain.getValue(1);
1562 VA = RVLocs[++i]; // skip ahead to next loc
1563 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1564 HalfGPRs.getValue(1), Flag);
1565 Flag = Chain.getValue(1);
1566 VA = RVLocs[++i]; // skip ahead to next loc
1568 // Extract the 2nd half and fall through to handle it as an f64 value.
1569 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1570 DAG.getConstant(1, MVT::i32));
1572 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1574 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1575 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1576 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1577 Flag = Chain.getValue(1);
1578 VA = RVLocs[++i]; // skip ahead to next loc
1579 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1582 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1584 // Guarantee that all emitted copies are
1585 // stuck together, avoiding something bad.
1586 Flag = Chain.getValue(1);
1591 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1593 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1598 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1599 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1600 // one of the above mentioned nodes. It has to be wrapped because otherwise
1601 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1602 // be used to form addressing mode. These wrapped nodes will be selected
1604 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1605 EVT PtrVT = Op.getValueType();
1606 // FIXME there is no actual debug info here
1607 DebugLoc dl = Op.getDebugLoc();
1608 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1610 if (CP->isMachineConstantPoolEntry())
1611 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1612 CP->getAlignment());
1614 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1615 CP->getAlignment());
1616 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1619 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1620 SelectionDAG &DAG) const {
1621 MachineFunction &MF = DAG.getMachineFunction();
1622 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1623 unsigned ARMPCLabelIndex = 0;
1624 DebugLoc DL = Op.getDebugLoc();
1625 EVT PtrVT = getPointerTy();
1626 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1627 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1629 if (RelocM == Reloc::Static) {
1630 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1632 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1633 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1634 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1635 ARMCP::CPBlockAddress,
1637 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1639 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1640 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1641 PseudoSourceValue::getConstantPool(), 0,
1643 if (RelocM == Reloc::Static)
1645 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1646 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1649 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1651 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1652 SelectionDAG &DAG) const {
1653 DebugLoc dl = GA->getDebugLoc();
1654 EVT PtrVT = getPointerTy();
1655 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1656 MachineFunction &MF = DAG.getMachineFunction();
1657 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1658 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1659 ARMConstantPoolValue *CPV =
1660 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1661 ARMCP::CPValue, PCAdj, "tlsgd", true);
1662 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1663 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1664 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1665 PseudoSourceValue::getConstantPool(), 0,
1667 SDValue Chain = Argument.getValue(1);
1669 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1670 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1672 // call __tls_get_addr.
1675 Entry.Node = Argument;
1676 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1677 Args.push_back(Entry);
1678 // FIXME: is there useful debug info available here?
1679 std::pair<SDValue, SDValue> CallResult =
1680 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1681 false, false, false, false,
1682 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1683 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1684 return CallResult.first;
1687 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1688 // "local exec" model.
1690 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1691 SelectionDAG &DAG) const {
1692 const GlobalValue *GV = GA->getGlobal();
1693 DebugLoc dl = GA->getDebugLoc();
1695 SDValue Chain = DAG.getEntryNode();
1696 EVT PtrVT = getPointerTy();
1697 // Get the Thread Pointer
1698 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1700 if (GV->isDeclaration()) {
1701 MachineFunction &MF = DAG.getMachineFunction();
1702 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1703 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1704 // Initial exec model.
1705 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1706 ARMConstantPoolValue *CPV =
1707 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1708 ARMCP::CPValue, PCAdj, "gottpoff", true);
1709 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1710 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1711 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1712 PseudoSourceValue::getConstantPool(), 0,
1714 Chain = Offset.getValue(1);
1716 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1717 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1719 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1720 PseudoSourceValue::getConstantPool(), 0,
1724 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1725 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1726 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1727 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1728 PseudoSourceValue::getConstantPool(), 0,
1732 // The address of the thread local variable is the add of the thread
1733 // pointer with the offset of the variable.
1734 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1738 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1739 // TODO: implement the "local dynamic" model
1740 assert(Subtarget->isTargetELF() &&
1741 "TLS not implemented for non-ELF targets");
1742 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1743 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1744 // otherwise use the "Local Exec" TLS Model
1745 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1746 return LowerToTLSGeneralDynamicModel(GA, DAG);
1748 return LowerToTLSExecModels(GA, DAG);
1751 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1752 SelectionDAG &DAG) const {
1753 EVT PtrVT = getPointerTy();
1754 DebugLoc dl = Op.getDebugLoc();
1755 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1756 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1757 if (RelocM == Reloc::PIC_) {
1758 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1759 ARMConstantPoolValue *CPV =
1760 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1761 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1762 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1763 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1765 PseudoSourceValue::getConstantPool(), 0,
1767 SDValue Chain = Result.getValue(1);
1768 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1769 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1771 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1772 PseudoSourceValue::getGOT(), 0,
1776 // If we have T2 ops, we can materialize the address directly via movt/movw
1777 // pair. This is always cheaper.
1778 if (Subtarget->useMovt()) {
1779 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1780 DAG.getTargetGlobalAddress(GV, PtrVT));
1782 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1783 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1784 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1785 PseudoSourceValue::getConstantPool(), 0,
1791 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1792 SelectionDAG &DAG) const {
1793 MachineFunction &MF = DAG.getMachineFunction();
1794 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1795 unsigned ARMPCLabelIndex = 0;
1796 EVT PtrVT = getPointerTy();
1797 DebugLoc dl = Op.getDebugLoc();
1798 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1799 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1801 if (RelocM == Reloc::Static)
1802 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1804 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1805 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1806 ARMConstantPoolValue *CPV =
1807 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1808 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1810 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1812 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1813 PseudoSourceValue::getConstantPool(), 0,
1815 SDValue Chain = Result.getValue(1);
1817 if (RelocM == Reloc::PIC_) {
1818 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1819 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1822 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1823 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1824 PseudoSourceValue::getGOT(), 0,
1830 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1831 SelectionDAG &DAG) const {
1832 assert(Subtarget->isTargetELF() &&
1833 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1834 MachineFunction &MF = DAG.getMachineFunction();
1835 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1836 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1837 EVT PtrVT = getPointerTy();
1838 DebugLoc dl = Op.getDebugLoc();
1839 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1840 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1841 "_GLOBAL_OFFSET_TABLE_",
1842 ARMPCLabelIndex, PCAdj);
1843 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1844 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1845 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1846 PseudoSourceValue::getConstantPool(), 0,
1848 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1849 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1853 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1854 DebugLoc dl = Op.getDebugLoc();
1855 SDValue Val = DAG.getConstant(0, MVT::i32);
1856 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1857 Op.getOperand(1), Val);
1861 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1862 DebugLoc dl = Op.getDebugLoc();
1863 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1864 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1868 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1869 const ARMSubtarget *Subtarget) const {
1870 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1871 DebugLoc dl = Op.getDebugLoc();
1873 default: return SDValue(); // Don't custom lower most intrinsics.
1874 case Intrinsic::arm_thread_pointer: {
1875 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1876 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1878 case Intrinsic::eh_sjlj_lsda: {
1879 MachineFunction &MF = DAG.getMachineFunction();
1880 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1881 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1882 EVT PtrVT = getPointerTy();
1883 DebugLoc dl = Op.getDebugLoc();
1884 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1886 unsigned PCAdj = (RelocM != Reloc::PIC_)
1887 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1888 ARMConstantPoolValue *CPV =
1889 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1890 ARMCP::CPLSDA, PCAdj);
1891 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1892 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1894 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1895 PseudoSourceValue::getConstantPool(), 0,
1897 SDValue Chain = Result.getValue(1);
1899 if (RelocM == Reloc::PIC_) {
1900 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1901 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1908 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1909 const ARMSubtarget *Subtarget) {
1910 DebugLoc dl = Op.getDebugLoc();
1911 SDValue Op5 = Op.getOperand(5);
1912 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1913 // v6 and v7 can both handle barriers directly, but need handled a bit
1914 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1916 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1917 if (Subtarget->hasV7Ops())
1918 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1919 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1920 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1921 DAG.getConstant(0, MVT::i32));
1922 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1926 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1927 MachineFunction &MF = DAG.getMachineFunction();
1928 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1930 // vastart just stores the address of the VarArgsFrameIndex slot into the
1931 // memory location argument.
1932 DebugLoc dl = Op.getDebugLoc();
1933 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1934 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1935 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1936 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1941 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1942 SelectionDAG &DAG) const {
1943 SDNode *Node = Op.getNode();
1944 DebugLoc dl = Node->getDebugLoc();
1945 EVT VT = Node->getValueType(0);
1946 SDValue Chain = Op.getOperand(0);
1947 SDValue Size = Op.getOperand(1);
1948 SDValue Align = Op.getOperand(2);
1950 // Chain the dynamic stack allocation so that it doesn't modify the stack
1951 // pointer when other instructions are using the stack.
1952 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1954 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1955 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1956 if (AlignVal > StackAlign)
1957 // Do this now since selection pass cannot introduce new target
1958 // independent node.
1959 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1961 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1962 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1963 // do even more horrible hack later.
1964 MachineFunction &MF = DAG.getMachineFunction();
1965 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1966 if (AFI->isThumb1OnlyFunction()) {
1968 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1970 uint32_t Val = C->getZExtValue();
1971 if (Val <= 508 && ((Val & 3) == 0))
1975 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1978 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1979 SDValue Ops1[] = { Chain, Size, Align };
1980 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1981 Chain = Res.getValue(1);
1982 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1983 DAG.getIntPtrConstant(0, true), SDValue());
1984 SDValue Ops2[] = { Res, Chain };
1985 return DAG.getMergeValues(Ops2, 2, dl);
1989 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1990 SDValue &Root, SelectionDAG &DAG,
1991 DebugLoc dl) const {
1992 MachineFunction &MF = DAG.getMachineFunction();
1993 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1995 TargetRegisterClass *RC;
1996 if (AFI->isThumb1OnlyFunction())
1997 RC = ARM::tGPRRegisterClass;
1999 RC = ARM::GPRRegisterClass;
2001 // Transform the arguments stored in physical registers into virtual ones.
2002 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2003 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2006 if (NextVA.isMemLoc()) {
2007 MachineFrameInfo *MFI = MF.getFrameInfo();
2008 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
2010 // Create load node to retrieve arguments from the stack.
2011 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2012 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2013 PseudoSourceValue::getFixedStack(FI), 0,
2016 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2017 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2020 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2024 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2025 CallingConv::ID CallConv, bool isVarArg,
2026 const SmallVectorImpl<ISD::InputArg>
2028 DebugLoc dl, SelectionDAG &DAG,
2029 SmallVectorImpl<SDValue> &InVals)
2032 MachineFunction &MF = DAG.getMachineFunction();
2033 MachineFrameInfo *MFI = MF.getFrameInfo();
2035 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2037 // Assign locations to all of the incoming arguments.
2038 SmallVector<CCValAssign, 16> ArgLocs;
2039 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2041 CCInfo.AnalyzeFormalArguments(Ins,
2042 CCAssignFnForNode(CallConv, /* Return*/ false,
2045 SmallVector<SDValue, 16> ArgValues;
2047 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2048 CCValAssign &VA = ArgLocs[i];
2050 // Arguments stored in registers.
2051 if (VA.isRegLoc()) {
2052 EVT RegVT = VA.getLocVT();
2055 if (VA.needsCustom()) {
2056 // f64 and vector types are split up into multiple registers or
2057 // combinations of registers and stack slots.
2058 if (VA.getLocVT() == MVT::v2f64) {
2059 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2061 VA = ArgLocs[++i]; // skip ahead to next loc
2063 if (VA.isMemLoc()) {
2064 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
2066 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2067 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2068 PseudoSourceValue::getFixedStack(FI), 0,
2071 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2074 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2075 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2076 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2077 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2078 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2080 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2083 TargetRegisterClass *RC;
2085 if (RegVT == MVT::f32)
2086 RC = ARM::SPRRegisterClass;
2087 else if (RegVT == MVT::f64)
2088 RC = ARM::DPRRegisterClass;
2089 else if (RegVT == MVT::v2f64)
2090 RC = ARM::QPRRegisterClass;
2091 else if (RegVT == MVT::i32)
2092 RC = (AFI->isThumb1OnlyFunction() ?
2093 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2095 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2097 // Transform the arguments in physical registers into virtual ones.
2098 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2099 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2102 // If this is an 8 or 16-bit value, it is really passed promoted
2103 // to 32 bits. Insert an assert[sz]ext to capture this, then
2104 // truncate to the right size.
2105 switch (VA.getLocInfo()) {
2106 default: llvm_unreachable("Unknown loc info!");
2107 case CCValAssign::Full: break;
2108 case CCValAssign::BCvt:
2109 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2111 case CCValAssign::SExt:
2112 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2113 DAG.getValueType(VA.getValVT()));
2114 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2116 case CCValAssign::ZExt:
2117 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2118 DAG.getValueType(VA.getValVT()));
2119 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2123 InVals.push_back(ArgValue);
2125 } else { // VA.isRegLoc()
2128 assert(VA.isMemLoc());
2129 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2131 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2132 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2135 // Create load nodes to retrieve arguments from the stack.
2136 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2137 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2138 PseudoSourceValue::getFixedStack(FI), 0,
2145 static const unsigned GPRArgRegs[] = {
2146 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2149 unsigned NumGPRs = CCInfo.getFirstUnallocated
2150 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2152 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2153 unsigned VARegSize = (4 - NumGPRs) * 4;
2154 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2155 unsigned ArgOffset = CCInfo.getNextStackOffset();
2156 if (VARegSaveSize) {
2157 // If this function is vararg, store any remaining integer argument regs
2158 // to their spots on the stack so that they may be loaded by deferencing
2159 // the result of va_next.
2160 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2161 AFI->setVarArgsFrameIndex(
2162 MFI->CreateFixedObject(VARegSaveSize,
2163 ArgOffset + VARegSaveSize - VARegSize,
2165 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2168 SmallVector<SDValue, 4> MemOps;
2169 for (; NumGPRs < 4; ++NumGPRs) {
2170 TargetRegisterClass *RC;
2171 if (AFI->isThumb1OnlyFunction())
2172 RC = ARM::tGPRRegisterClass;
2174 RC = ARM::GPRRegisterClass;
2176 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2177 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2179 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2180 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2181 0, false, false, 0);
2182 MemOps.push_back(Store);
2183 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2184 DAG.getConstant(4, getPointerTy()));
2186 if (!MemOps.empty())
2187 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2188 &MemOps[0], MemOps.size());
2190 // This will point to the next argument passed via stack.
2191 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
2198 /// isFloatingPointZero - Return true if this is +0.0.
2199 static bool isFloatingPointZero(SDValue Op) {
2200 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2201 return CFP->getValueAPF().isPosZero();
2202 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2203 // Maybe this has already been legalized into the constant pool?
2204 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2205 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2206 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2207 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2208 return CFP->getValueAPF().isPosZero();
2214 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2215 /// the given operands.
2217 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2218 SDValue &ARMCC, SelectionDAG &DAG,
2219 DebugLoc dl) const {
2220 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2221 unsigned C = RHSC->getZExtValue();
2222 if (!isLegalICmpImmediate(C)) {
2223 // Constant does not fit, try adjusting it by one?
2228 if (isLegalICmpImmediate(C-1)) {
2229 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2230 RHS = DAG.getConstant(C-1, MVT::i32);
2235 if (C > 0 && isLegalICmpImmediate(C-1)) {
2236 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2237 RHS = DAG.getConstant(C-1, MVT::i32);
2242 if (isLegalICmpImmediate(C+1)) {
2243 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2244 RHS = DAG.getConstant(C+1, MVT::i32);
2249 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2250 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2251 RHS = DAG.getConstant(C+1, MVT::i32);
2258 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2259 ARMISD::NodeType CompareType;
2262 CompareType = ARMISD::CMP;
2267 CompareType = ARMISD::CMPZ;
2270 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2271 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2274 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2275 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2278 if (!isFloatingPointZero(RHS))
2279 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2281 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2282 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2285 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2286 EVT VT = Op.getValueType();
2287 SDValue LHS = Op.getOperand(0);
2288 SDValue RHS = Op.getOperand(1);
2289 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2290 SDValue TrueVal = Op.getOperand(2);
2291 SDValue FalseVal = Op.getOperand(3);
2292 DebugLoc dl = Op.getDebugLoc();
2294 if (LHS.getValueType() == MVT::i32) {
2296 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2297 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2298 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2301 ARMCC::CondCodes CondCode, CondCode2;
2302 FPCCToARMCC(CC, CondCode, CondCode2);
2304 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2305 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2306 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2307 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2309 if (CondCode2 != ARMCC::AL) {
2310 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2311 // FIXME: Needs another CMP because flag can have but one use.
2312 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2313 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2314 Result, TrueVal, ARMCC2, CCR, Cmp2);
2319 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2320 SDValue Chain = Op.getOperand(0);
2321 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2322 SDValue LHS = Op.getOperand(2);
2323 SDValue RHS = Op.getOperand(3);
2324 SDValue Dest = Op.getOperand(4);
2325 DebugLoc dl = Op.getDebugLoc();
2327 if (LHS.getValueType() == MVT::i32) {
2329 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2330 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2331 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2332 Chain, Dest, ARMCC, CCR,Cmp);
2335 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2336 ARMCC::CondCodes CondCode, CondCode2;
2337 FPCCToARMCC(CC, CondCode, CondCode2);
2339 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2340 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2341 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2342 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2343 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2344 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2345 if (CondCode2 != ARMCC::AL) {
2346 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2347 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2348 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2353 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2354 SDValue Chain = Op.getOperand(0);
2355 SDValue Table = Op.getOperand(1);
2356 SDValue Index = Op.getOperand(2);
2357 DebugLoc dl = Op.getDebugLoc();
2359 EVT PTy = getPointerTy();
2360 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2361 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2362 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2363 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2364 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2365 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2366 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2367 if (Subtarget->isThumb2()) {
2368 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2369 // which does another jump to the destination. This also makes it easier
2370 // to translate it to TBB / TBH later.
2371 // FIXME: This might not work if the function is extremely large.
2372 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2373 Addr, Op.getOperand(2), JTI, UId);
2375 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2376 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2377 PseudoSourceValue::getJumpTable(), 0,
2379 Chain = Addr.getValue(1);
2380 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2381 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2383 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2384 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2385 Chain = Addr.getValue(1);
2386 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2390 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2391 DebugLoc dl = Op.getDebugLoc();
2394 switch (Op.getOpcode()) {
2396 assert(0 && "Invalid opcode!");
2397 case ISD::FP_TO_SINT:
2398 Opc = ARMISD::FTOSI;
2400 case ISD::FP_TO_UINT:
2401 Opc = ARMISD::FTOUI;
2404 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2405 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2408 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2409 EVT VT = Op.getValueType();
2410 DebugLoc dl = Op.getDebugLoc();
2413 switch (Op.getOpcode()) {
2415 assert(0 && "Invalid opcode!");
2416 case ISD::SINT_TO_FP:
2417 Opc = ARMISD::SITOF;
2419 case ISD::UINT_TO_FP:
2420 Opc = ARMISD::UITOF;
2424 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2425 return DAG.getNode(Opc, dl, VT, Op);
2428 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2429 // Implement fcopysign with a fabs and a conditional fneg.
2430 SDValue Tmp0 = Op.getOperand(0);
2431 SDValue Tmp1 = Op.getOperand(1);
2432 DebugLoc dl = Op.getDebugLoc();
2433 EVT VT = Op.getValueType();
2434 EVT SrcVT = Tmp1.getValueType();
2435 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2436 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2437 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2438 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2439 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2442 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2443 MachineFunction &MF = DAG.getMachineFunction();
2444 MachineFrameInfo *MFI = MF.getFrameInfo();
2445 MFI->setReturnAddressIsTaken(true);
2447 EVT VT = Op.getValueType();
2448 DebugLoc dl = Op.getDebugLoc();
2449 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2451 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2452 SDValue Offset = DAG.getConstant(4, MVT::i32);
2453 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2454 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2455 NULL, 0, false, false, 0);
2458 // Return LR, which contains the return address. Mark it an implicit live-in.
2459 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2460 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2463 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2464 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2465 MFI->setFrameAddressIsTaken(true);
2467 EVT VT = Op.getValueType();
2468 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2469 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2470 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2471 ? ARM::R7 : ARM::R11;
2472 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2474 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2479 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2480 /// expand a bit convert where either the source or destination type is i64 to
2481 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2482 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2483 /// vectors), since the legalizer won't know what to do with that.
2484 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2485 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2486 DebugLoc dl = N->getDebugLoc();
2487 SDValue Op = N->getOperand(0);
2489 // This function is only supposed to be called for i64 types, either as the
2490 // source or destination of the bit convert.
2491 EVT SrcVT = Op.getValueType();
2492 EVT DstVT = N->getValueType(0);
2493 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2494 "ExpandBIT_CONVERT called for non-i64 type");
2496 // Turn i64->f64 into VMOVDRR.
2497 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2498 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2499 DAG.getConstant(0, MVT::i32));
2500 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2501 DAG.getConstant(1, MVT::i32));
2502 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2503 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2506 // Turn f64->i64 into VMOVRRD.
2507 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2508 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2509 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2510 // Merge the pieces into a single i64 value.
2511 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2517 /// getZeroVector - Returns a vector of specified type with all zero elements.
2519 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2520 assert(VT.isVector() && "Expected a vector type");
2522 // Zero vectors are used to represent vector negation and in those cases
2523 // will be implemented with the NEON VNEG instruction. However, VNEG does
2524 // not support i64 elements, so sometimes the zero vectors will need to be
2525 // explicitly constructed. For those cases, and potentially other uses in
2526 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2527 // to their dest type. This ensures they get CSE'd.
2529 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2530 SmallVector<SDValue, 8> Ops;
2533 if (VT.getSizeInBits() == 64) {
2534 Ops.assign(8, Cst); TVT = MVT::v8i8;
2536 Ops.assign(16, Cst); TVT = MVT::v16i8;
2538 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2540 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2543 /// getOnesVector - Returns a vector of specified type with all bits set.
2545 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2546 assert(VT.isVector() && "Expected a vector type");
2548 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2549 // dest type. This ensures they get CSE'd.
2551 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2552 SmallVector<SDValue, 8> Ops;
2555 if (VT.getSizeInBits() == 64) {
2556 Ops.assign(8, Cst); TVT = MVT::v8i8;
2558 Ops.assign(16, Cst); TVT = MVT::v16i8;
2560 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2562 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2565 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2566 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2567 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2568 SelectionDAG &DAG) const {
2569 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2570 EVT VT = Op.getValueType();
2571 unsigned VTBits = VT.getSizeInBits();
2572 DebugLoc dl = Op.getDebugLoc();
2573 SDValue ShOpLo = Op.getOperand(0);
2574 SDValue ShOpHi = Op.getOperand(1);
2575 SDValue ShAmt = Op.getOperand(2);
2577 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2579 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2581 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2582 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2583 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2584 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2585 DAG.getConstant(VTBits, MVT::i32));
2586 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2587 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2588 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2590 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2591 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2593 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2594 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2597 SDValue Ops[2] = { Lo, Hi };
2598 return DAG.getMergeValues(Ops, 2, dl);
2601 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2602 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2603 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2604 SelectionDAG &DAG) const {
2605 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2606 EVT VT = Op.getValueType();
2607 unsigned VTBits = VT.getSizeInBits();
2608 DebugLoc dl = Op.getDebugLoc();
2609 SDValue ShOpLo = Op.getOperand(0);
2610 SDValue ShOpHi = Op.getOperand(1);
2611 SDValue ShAmt = Op.getOperand(2);
2614 assert(Op.getOpcode() == ISD::SHL_PARTS);
2615 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2616 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2617 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2618 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2619 DAG.getConstant(VTBits, MVT::i32));
2620 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2621 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2623 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2624 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2625 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2627 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2628 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2631 SDValue Ops[2] = { Lo, Hi };
2632 return DAG.getMergeValues(Ops, 2, dl);
2635 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2636 const ARMSubtarget *ST) {
2637 EVT VT = N->getValueType(0);
2638 DebugLoc dl = N->getDebugLoc();
2640 if (!ST->hasV6T2Ops())
2643 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2644 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2647 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2648 const ARMSubtarget *ST) {
2649 EVT VT = N->getValueType(0);
2650 DebugLoc dl = N->getDebugLoc();
2652 // Lower vector shifts on NEON to use VSHL.
2653 if (VT.isVector()) {
2654 assert(ST->hasNEON() && "unexpected vector shift");
2656 // Left shifts translate directly to the vshiftu intrinsic.
2657 if (N->getOpcode() == ISD::SHL)
2658 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2659 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2660 N->getOperand(0), N->getOperand(1));
2662 assert((N->getOpcode() == ISD::SRA ||
2663 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2665 // NEON uses the same intrinsics for both left and right shifts. For
2666 // right shifts, the shift amounts are negative, so negate the vector of
2668 EVT ShiftVT = N->getOperand(1).getValueType();
2669 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2670 getZeroVector(ShiftVT, DAG, dl),
2672 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2673 Intrinsic::arm_neon_vshifts :
2674 Intrinsic::arm_neon_vshiftu);
2675 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2676 DAG.getConstant(vshiftInt, MVT::i32),
2677 N->getOperand(0), NegatedCount);
2680 // We can get here for a node like i32 = ISD::SHL i32, i64
2684 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2685 "Unknown shift to lower!");
2687 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2688 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2689 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2692 // If we are in thumb mode, we don't have RRX.
2693 if (ST->isThumb1Only()) return SDValue();
2695 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2696 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2697 DAG.getConstant(0, MVT::i32));
2698 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2699 DAG.getConstant(1, MVT::i32));
2701 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2702 // captures the result into a carry flag.
2703 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2704 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2706 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2707 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2709 // Merge the pieces into a single i64 value.
2710 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2713 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2714 SDValue TmpOp0, TmpOp1;
2715 bool Invert = false;
2719 SDValue Op0 = Op.getOperand(0);
2720 SDValue Op1 = Op.getOperand(1);
2721 SDValue CC = Op.getOperand(2);
2722 EVT VT = Op.getValueType();
2723 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2724 DebugLoc dl = Op.getDebugLoc();
2726 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2727 switch (SetCCOpcode) {
2728 default: llvm_unreachable("Illegal FP comparison"); break;
2730 case ISD::SETNE: Invert = true; // Fallthrough
2732 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2734 case ISD::SETLT: Swap = true; // Fallthrough
2736 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2738 case ISD::SETLE: Swap = true; // Fallthrough
2740 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2741 case ISD::SETUGE: Swap = true; // Fallthrough
2742 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2743 case ISD::SETUGT: Swap = true; // Fallthrough
2744 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2745 case ISD::SETUEQ: Invert = true; // Fallthrough
2747 // Expand this to (OLT | OGT).
2751 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2752 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2754 case ISD::SETUO: Invert = true; // Fallthrough
2756 // Expand this to (OLT | OGE).
2760 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2761 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2765 // Integer comparisons.
2766 switch (SetCCOpcode) {
2767 default: llvm_unreachable("Illegal integer comparison"); break;
2768 case ISD::SETNE: Invert = true;
2769 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2770 case ISD::SETLT: Swap = true;
2771 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2772 case ISD::SETLE: Swap = true;
2773 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2774 case ISD::SETULT: Swap = true;
2775 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2776 case ISD::SETULE: Swap = true;
2777 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2780 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2781 if (Opc == ARMISD::VCEQ) {
2784 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2786 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2789 // Ignore bitconvert.
2790 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2791 AndOp = AndOp.getOperand(0);
2793 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2795 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2796 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2803 std::swap(Op0, Op1);
2805 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2808 Result = DAG.getNOT(dl, Result, VT);
2813 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2814 /// valid vector constant for a NEON instruction with a "modified immediate"
2815 /// operand (e.g., VMOV). If so, return either the constant being
2816 /// splatted or the encoded value, depending on the DoEncode parameter. The
2817 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2818 /// bits7-0=Immediate.
2819 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2820 unsigned SplatBitSize, SelectionDAG &DAG,
2821 bool isVMOV, bool DoEncode) {
2822 unsigned Op, Cmode, Imm;
2825 // SplatBitSize is set to the smallest size that splats the vector, so a
2826 // zero vector will always have SplatBitSize == 8. However, NEON modified
2827 // immediate instructions others than VMOV do not support the 8-bit encoding
2828 // of a zero vector, and the default encoding of zero is supposed to be the
2834 switch (SplatBitSize) {
2836 // Any 1-byte value is OK. Op=0, Cmode=1110.
2837 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2844 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2846 if ((SplatBits & ~0xff) == 0) {
2847 // Value = 0x00nn: Op=x, Cmode=100x.
2852 if ((SplatBits & ~0xff00) == 0) {
2853 // Value = 0xnn00: Op=x, Cmode=101x.
2855 Imm = SplatBits >> 8;
2861 // NEON's 32-bit VMOV supports splat values where:
2862 // * only one byte is nonzero, or
2863 // * the least significant byte is 0xff and the second byte is nonzero, or
2864 // * the least significant 2 bytes are 0xff and the third is nonzero.
2866 if ((SplatBits & ~0xff) == 0) {
2867 // Value = 0x000000nn: Op=x, Cmode=000x.
2872 if ((SplatBits & ~0xff00) == 0) {
2873 // Value = 0x0000nn00: Op=x, Cmode=001x.
2875 Imm = SplatBits >> 8;
2878 if ((SplatBits & ~0xff0000) == 0) {
2879 // Value = 0x00nn0000: Op=x, Cmode=010x.
2881 Imm = SplatBits >> 16;
2884 if ((SplatBits & ~0xff000000) == 0) {
2885 // Value = 0xnn000000: Op=x, Cmode=011x.
2887 Imm = SplatBits >> 24;
2891 if ((SplatBits & ~0xffff) == 0 &&
2892 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2893 // Value = 0x0000nnff: Op=x, Cmode=1100.
2895 Imm = SplatBits >> 8;
2900 if ((SplatBits & ~0xffffff) == 0 &&
2901 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2902 // Value = 0x00nnffff: Op=x, Cmode=1101.
2904 Imm = SplatBits >> 16;
2905 SplatBits |= 0xffff;
2909 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2910 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2911 // VMOV.I32. A (very) minor optimization would be to replicate the value
2912 // and fall through here to test for a valid 64-bit splat. But, then the
2913 // caller would also need to check and handle the change in size.
2917 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2920 uint64_t BitMask = 0xff;
2922 unsigned ImmMask = 1;
2924 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2925 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2928 } else if ((SplatBits & BitMask) != 0) {
2934 // Op=1, Cmode=1110.
2943 llvm_unreachable("unexpected size for EncodeNEONModImm");
2948 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2949 return DAG.getTargetConstant(SplatBits, VT);
2953 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2954 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2955 /// size, return the encoded value for that immediate. The ByteSize field
2956 /// indicates the number of bytes of each element [1248].
2957 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2958 SelectionDAG &DAG) {
2959 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2960 APInt SplatBits, SplatUndef;
2961 unsigned SplatBitSize;
2963 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2964 HasAnyUndefs, ByteSize * 8))
2967 if (SplatBitSize > ByteSize * 8)
2970 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2971 SplatBitSize, DAG, isVMOV, true);
2974 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2975 bool &ReverseVEXT, unsigned &Imm) {
2976 unsigned NumElts = VT.getVectorNumElements();
2977 ReverseVEXT = false;
2980 // If this is a VEXT shuffle, the immediate value is the index of the first
2981 // element. The other shuffle indices must be the successive elements after
2983 unsigned ExpectedElt = Imm;
2984 for (unsigned i = 1; i < NumElts; ++i) {
2985 // Increment the expected index. If it wraps around, it may still be
2986 // a VEXT but the source vectors must be swapped.
2988 if (ExpectedElt == NumElts * 2) {
2993 if (ExpectedElt != static_cast<unsigned>(M[i]))
2997 // Adjust the index value if the source operands will be swapped.
3004 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3005 /// instruction with the specified blocksize. (The order of the elements
3006 /// within each block of the vector is reversed.)
3007 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3008 unsigned BlockSize) {
3009 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3010 "Only possible block sizes for VREV are: 16, 32, 64");
3012 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3016 unsigned NumElts = VT.getVectorNumElements();
3017 unsigned BlockElts = M[0] + 1;
3019 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3022 for (unsigned i = 0; i < NumElts; ++i) {
3023 if ((unsigned) M[i] !=
3024 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3031 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3032 unsigned &WhichResult) {
3033 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3037 unsigned NumElts = VT.getVectorNumElements();
3038 WhichResult = (M[0] == 0 ? 0 : 1);
3039 for (unsigned i = 0; i < NumElts; i += 2) {
3040 if ((unsigned) M[i] != i + WhichResult ||
3041 (unsigned) M[i+1] != i + NumElts + WhichResult)
3047 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3048 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3049 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3050 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3051 unsigned &WhichResult) {
3052 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3056 unsigned NumElts = VT.getVectorNumElements();
3057 WhichResult = (M[0] == 0 ? 0 : 1);
3058 for (unsigned i = 0; i < NumElts; i += 2) {
3059 if ((unsigned) M[i] != i + WhichResult ||
3060 (unsigned) M[i+1] != i + WhichResult)
3066 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3067 unsigned &WhichResult) {
3068 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3072 unsigned NumElts = VT.getVectorNumElements();
3073 WhichResult = (M[0] == 0 ? 0 : 1);
3074 for (unsigned i = 0; i != NumElts; ++i) {
3075 if ((unsigned) M[i] != 2 * i + WhichResult)
3079 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3080 if (VT.is64BitVector() && EltSz == 32)
3086 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3087 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3088 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3089 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3090 unsigned &WhichResult) {
3091 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3095 unsigned Half = VT.getVectorNumElements() / 2;
3096 WhichResult = (M[0] == 0 ? 0 : 1);
3097 for (unsigned j = 0; j != 2; ++j) {
3098 unsigned Idx = WhichResult;
3099 for (unsigned i = 0; i != Half; ++i) {
3100 if ((unsigned) M[i + j * Half] != Idx)
3106 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3107 if (VT.is64BitVector() && EltSz == 32)
3113 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3114 unsigned &WhichResult) {
3115 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3119 unsigned NumElts = VT.getVectorNumElements();
3120 WhichResult = (M[0] == 0 ? 0 : 1);
3121 unsigned Idx = WhichResult * NumElts / 2;
3122 for (unsigned i = 0; i != NumElts; i += 2) {
3123 if ((unsigned) M[i] != Idx ||
3124 (unsigned) M[i+1] != Idx + NumElts)
3129 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3130 if (VT.is64BitVector() && EltSz == 32)
3136 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3137 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3138 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3139 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3140 unsigned &WhichResult) {
3141 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3145 unsigned NumElts = VT.getVectorNumElements();
3146 WhichResult = (M[0] == 0 ? 0 : 1);
3147 unsigned Idx = WhichResult * NumElts / 2;
3148 for (unsigned i = 0; i != NumElts; i += 2) {
3149 if ((unsigned) M[i] != Idx ||
3150 (unsigned) M[i+1] != Idx)
3155 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3156 if (VT.is64BitVector() && EltSz == 32)
3163 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3164 // Canonicalize all-zeros and all-ones vectors.
3165 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3166 if (ConstVal->isNullValue())
3167 return getZeroVector(VT, DAG, dl);
3168 if (ConstVal->isAllOnesValue())
3169 return getOnesVector(VT, DAG, dl);
3172 if (VT.is64BitVector()) {
3173 switch (Val.getValueType().getSizeInBits()) {
3174 case 8: CanonicalVT = MVT::v8i8; break;
3175 case 16: CanonicalVT = MVT::v4i16; break;
3176 case 32: CanonicalVT = MVT::v2i32; break;
3177 case 64: CanonicalVT = MVT::v1i64; break;
3178 default: llvm_unreachable("unexpected splat element type"); break;
3181 assert(VT.is128BitVector() && "unknown splat vector size");
3182 switch (Val.getValueType().getSizeInBits()) {
3183 case 8: CanonicalVT = MVT::v16i8; break;
3184 case 16: CanonicalVT = MVT::v8i16; break;
3185 case 32: CanonicalVT = MVT::v4i32; break;
3186 case 64: CanonicalVT = MVT::v2i64; break;
3187 default: llvm_unreachable("unexpected splat element type"); break;
3191 // Build a canonical splat for this value.
3192 SmallVector<SDValue, 8> Ops;
3193 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3194 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3196 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3199 // If this is a case we can't handle, return null and let the default
3200 // expansion code take care of it.
3201 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3202 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3203 DebugLoc dl = Op.getDebugLoc();
3204 EVT VT = Op.getValueType();
3206 APInt SplatBits, SplatUndef;
3207 unsigned SplatBitSize;
3209 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3210 if (SplatBitSize <= 64) {
3211 // Check if an immediate VMOV works.
3212 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3213 SplatUndef.getZExtValue(),
3214 SplatBitSize, DAG, true, false);
3216 return BuildSplat(Val, VT, DAG, dl);
3220 // Scan through the operands to see if only one value is used.
3221 unsigned NumElts = VT.getVectorNumElements();
3222 bool isOnlyLowElement = true;
3223 bool usesOnlyOneValue = true;
3224 bool isConstant = true;
3226 for (unsigned i = 0; i < NumElts; ++i) {
3227 SDValue V = Op.getOperand(i);
3228 if (V.getOpcode() == ISD::UNDEF)
3231 isOnlyLowElement = false;
3232 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3235 if (!Value.getNode())
3237 else if (V != Value)
3238 usesOnlyOneValue = false;
3241 if (!Value.getNode())
3242 return DAG.getUNDEF(VT);
3244 if (isOnlyLowElement)
3245 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3247 // If all elements are constants, fall back to the default expansion, which
3248 // will generate a load from the constant pool.
3252 // Use VDUP for non-constant splats.
3253 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3254 if (usesOnlyOneValue && EltSize <= 32)
3255 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3257 // Vectors with 32- or 64-bit elements can be built by directly assigning
3258 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3259 // will be legalized.
3260 if (EltSize >= 32) {
3261 // Do the expansion with floating-point types, since that is what the VFP
3262 // registers are defined to use, and since i64 is not legal.
3263 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3264 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3265 SmallVector<SDValue, 8> Ops;
3266 for (unsigned i = 0; i < NumElts; ++i)
3267 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3268 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3269 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3275 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3276 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3277 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3278 /// are assumed to be legal.
3280 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3282 if (VT.getVectorNumElements() == 4 &&
3283 (VT.is128BitVector() || VT.is64BitVector())) {
3284 unsigned PFIndexes[4];
3285 for (unsigned i = 0; i != 4; ++i) {
3289 PFIndexes[i] = M[i];
3292 // Compute the index in the perfect shuffle table.
3293 unsigned PFTableIndex =
3294 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3295 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3296 unsigned Cost = (PFEntry >> 30);
3303 unsigned Imm, WhichResult;
3305 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3306 return (EltSize >= 32 ||
3307 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3308 isVREVMask(M, VT, 64) ||
3309 isVREVMask(M, VT, 32) ||
3310 isVREVMask(M, VT, 16) ||
3311 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3312 isVTRNMask(M, VT, WhichResult) ||
3313 isVUZPMask(M, VT, WhichResult) ||
3314 isVZIPMask(M, VT, WhichResult) ||
3315 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3316 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3317 isVZIP_v_undef_Mask(M, VT, WhichResult));
3320 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3321 /// the specified operations to build the shuffle.
3322 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3323 SDValue RHS, SelectionDAG &DAG,
3325 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3326 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3327 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3330 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3339 OP_VUZPL, // VUZP, left result
3340 OP_VUZPR, // VUZP, right result
3341 OP_VZIPL, // VZIP, left result
3342 OP_VZIPR, // VZIP, right result
3343 OP_VTRNL, // VTRN, left result
3344 OP_VTRNR // VTRN, right result
3347 if (OpNum == OP_COPY) {
3348 if (LHSID == (1*9+2)*9+3) return LHS;
3349 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3353 SDValue OpLHS, OpRHS;
3354 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3355 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3356 EVT VT = OpLHS.getValueType();
3359 default: llvm_unreachable("Unknown shuffle opcode!");
3361 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3366 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3367 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3371 return DAG.getNode(ARMISD::VEXT, dl, VT,
3373 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3376 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3377 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3380 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3381 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3384 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3385 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3389 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3390 SDValue V1 = Op.getOperand(0);
3391 SDValue V2 = Op.getOperand(1);
3392 DebugLoc dl = Op.getDebugLoc();
3393 EVT VT = Op.getValueType();
3394 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3395 SmallVector<int, 8> ShuffleMask;
3397 // Convert shuffles that are directly supported on NEON to target-specific
3398 // DAG nodes, instead of keeping them as shuffles and matching them again
3399 // during code selection. This is more efficient and avoids the possibility
3400 // of inconsistencies between legalization and selection.
3401 // FIXME: floating-point vectors should be canonicalized to integer vectors
3402 // of the same time so that they get CSEd properly.
3403 SVN->getMask(ShuffleMask);
3405 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3406 if (EltSize <= 32) {
3407 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3408 int Lane = SVN->getSplatIndex();
3409 // If this is undef splat, generate it via "just" vdup, if possible.
3410 if (Lane == -1) Lane = 0;
3412 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3413 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3415 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3416 DAG.getConstant(Lane, MVT::i32));
3421 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3424 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3425 DAG.getConstant(Imm, MVT::i32));
3428 if (isVREVMask(ShuffleMask, VT, 64))
3429 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3430 if (isVREVMask(ShuffleMask, VT, 32))
3431 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3432 if (isVREVMask(ShuffleMask, VT, 16))
3433 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3435 // Check for Neon shuffles that modify both input vectors in place.
3436 // If both results are used, i.e., if there are two shuffles with the same
3437 // source operands and with masks corresponding to both results of one of
3438 // these operations, DAG memoization will ensure that a single node is
3439 // used for both shuffles.
3440 unsigned WhichResult;
3441 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3442 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3443 V1, V2).getValue(WhichResult);
3444 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3445 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3446 V1, V2).getValue(WhichResult);
3447 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3448 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3449 V1, V2).getValue(WhichResult);
3451 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3452 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3453 V1, V1).getValue(WhichResult);
3454 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3455 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3456 V1, V1).getValue(WhichResult);
3457 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3458 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3459 V1, V1).getValue(WhichResult);
3462 // If the shuffle is not directly supported and it has 4 elements, use
3463 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3464 unsigned NumElts = VT.getVectorNumElements();
3466 unsigned PFIndexes[4];
3467 for (unsigned i = 0; i != 4; ++i) {
3468 if (ShuffleMask[i] < 0)
3471 PFIndexes[i] = ShuffleMask[i];
3474 // Compute the index in the perfect shuffle table.
3475 unsigned PFTableIndex =
3476 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3477 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3478 unsigned Cost = (PFEntry >> 30);
3481 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3484 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3485 if (EltSize >= 32) {
3486 // Do the expansion with floating-point types, since that is what the VFP
3487 // registers are defined to use, and since i64 is not legal.
3488 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3489 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3490 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3491 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3492 SmallVector<SDValue, 8> Ops;
3493 for (unsigned i = 0; i < NumElts; ++i) {
3494 if (ShuffleMask[i] < 0)
3495 Ops.push_back(DAG.getUNDEF(EltVT));
3497 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3498 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3499 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3502 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3503 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3509 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3510 EVT VT = Op.getValueType();
3511 DebugLoc dl = Op.getDebugLoc();
3512 SDValue Vec = Op.getOperand(0);
3513 SDValue Lane = Op.getOperand(1);
3514 assert(VT == MVT::i32 &&
3515 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3516 "unexpected type for custom-lowering vector extract");
3517 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3520 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3521 // The only time a CONCAT_VECTORS operation can have legal types is when
3522 // two 64-bit vectors are concatenated to a 128-bit vector.
3523 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3524 "unexpected CONCAT_VECTORS");
3525 DebugLoc dl = Op.getDebugLoc();
3526 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3527 SDValue Op0 = Op.getOperand(0);
3528 SDValue Op1 = Op.getOperand(1);
3529 if (Op0.getOpcode() != ISD::UNDEF)
3530 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3531 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3532 DAG.getIntPtrConstant(0));
3533 if (Op1.getOpcode() != ISD::UNDEF)
3534 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3535 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3536 DAG.getIntPtrConstant(1));
3537 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3540 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3541 switch (Op.getOpcode()) {
3542 default: llvm_unreachable("Don't know how to custom lower this!");
3543 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3544 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3545 case ISD::GlobalAddress:
3546 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3547 LowerGlobalAddressELF(Op, DAG);
3548 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3549 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3550 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3551 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3552 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3553 case ISD::VASTART: return LowerVASTART(Op, DAG);
3554 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3555 case ISD::SINT_TO_FP:
3556 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3557 case ISD::FP_TO_SINT:
3558 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3559 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3560 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3561 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3562 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3563 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3564 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3565 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3567 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3570 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3571 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3572 case ISD::SRL_PARTS:
3573 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3574 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3575 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3576 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3577 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3578 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3579 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3584 /// ReplaceNodeResults - Replace the results of node with an illegal result
3585 /// type with new values built out of custom code.
3586 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3587 SmallVectorImpl<SDValue>&Results,
3588 SelectionDAG &DAG) const {
3590 switch (N->getOpcode()) {
3592 llvm_unreachable("Don't know how to custom expand this!");
3594 case ISD::BIT_CONVERT:
3595 Res = ExpandBIT_CONVERT(N, DAG);
3599 Res = LowerShift(N, DAG, Subtarget);
3603 Results.push_back(Res);
3606 //===----------------------------------------------------------------------===//
3607 // ARM Scheduler Hooks
3608 //===----------------------------------------------------------------------===//
3611 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3612 MachineBasicBlock *BB,
3613 unsigned Size) const {
3614 unsigned dest = MI->getOperand(0).getReg();
3615 unsigned ptr = MI->getOperand(1).getReg();
3616 unsigned oldval = MI->getOperand(2).getReg();
3617 unsigned newval = MI->getOperand(3).getReg();
3618 unsigned scratch = BB->getParent()->getRegInfo()
3619 .createVirtualRegister(ARM::GPRRegisterClass);
3620 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3621 DebugLoc dl = MI->getDebugLoc();
3622 bool isThumb2 = Subtarget->isThumb2();
3624 unsigned ldrOpc, strOpc;
3626 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3628 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3629 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3632 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3633 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3636 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3637 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3641 MachineFunction *MF = BB->getParent();
3642 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3643 MachineFunction::iterator It = BB;
3644 ++It; // insert the new blocks after the current block
3646 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3647 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3648 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3649 MF->insert(It, loop1MBB);
3650 MF->insert(It, loop2MBB);
3651 MF->insert(It, exitMBB);
3652 exitMBB->transferSuccessors(BB);
3656 // fallthrough --> loop1MBB
3657 BB->addSuccessor(loop1MBB);
3660 // ldrex dest, [ptr]
3664 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3665 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3666 .addReg(dest).addReg(oldval));
3667 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3668 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3669 BB->addSuccessor(loop2MBB);
3670 BB->addSuccessor(exitMBB);
3673 // strex scratch, newval, [ptr]
3677 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3679 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3680 .addReg(scratch).addImm(0));
3681 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3682 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3683 BB->addSuccessor(loop1MBB);
3684 BB->addSuccessor(exitMBB);
3690 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3696 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3697 unsigned Size, unsigned BinOpcode) const {
3698 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3699 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3701 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3702 MachineFunction *MF = BB->getParent();
3703 MachineFunction::iterator It = BB;
3706 unsigned dest = MI->getOperand(0).getReg();
3707 unsigned ptr = MI->getOperand(1).getReg();
3708 unsigned incr = MI->getOperand(2).getReg();
3709 DebugLoc dl = MI->getDebugLoc();
3711 bool isThumb2 = Subtarget->isThumb2();
3712 unsigned ldrOpc, strOpc;
3714 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3716 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3717 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3720 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3721 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3724 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3725 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3729 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3730 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3731 MF->insert(It, loopMBB);
3732 MF->insert(It, exitMBB);
3733 exitMBB->transferSuccessors(BB);
3735 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3736 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3737 unsigned scratch2 = (!BinOpcode) ? incr :
3738 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3742 // fallthrough --> loopMBB
3743 BB->addSuccessor(loopMBB);
3747 // <binop> scratch2, dest, incr
3748 // strex scratch, scratch2, ptr
3751 // fallthrough --> exitMBB
3753 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3755 // operand order needs to go the other way for NAND
3756 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3757 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3758 addReg(incr).addReg(dest)).addReg(0);
3760 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3761 addReg(dest).addReg(incr)).addReg(0);
3764 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3766 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3767 .addReg(scratch).addImm(0));
3768 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3769 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3771 BB->addSuccessor(loopMBB);
3772 BB->addSuccessor(exitMBB);
3778 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3784 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3785 MachineBasicBlock *BB) const {
3786 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3787 DebugLoc dl = MI->getDebugLoc();
3788 bool isThumb2 = Subtarget->isThumb2();
3789 switch (MI->getOpcode()) {
3792 llvm_unreachable("Unexpected instr type to insert");
3794 case ARM::ATOMIC_LOAD_ADD_I8:
3795 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3796 case ARM::ATOMIC_LOAD_ADD_I16:
3797 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3798 case ARM::ATOMIC_LOAD_ADD_I32:
3799 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3801 case ARM::ATOMIC_LOAD_AND_I8:
3802 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3803 case ARM::ATOMIC_LOAD_AND_I16:
3804 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3805 case ARM::ATOMIC_LOAD_AND_I32:
3806 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3808 case ARM::ATOMIC_LOAD_OR_I8:
3809 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3810 case ARM::ATOMIC_LOAD_OR_I16:
3811 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3812 case ARM::ATOMIC_LOAD_OR_I32:
3813 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3815 case ARM::ATOMIC_LOAD_XOR_I8:
3816 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3817 case ARM::ATOMIC_LOAD_XOR_I16:
3818 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3819 case ARM::ATOMIC_LOAD_XOR_I32:
3820 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3822 case ARM::ATOMIC_LOAD_NAND_I8:
3823 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3824 case ARM::ATOMIC_LOAD_NAND_I16:
3825 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3826 case ARM::ATOMIC_LOAD_NAND_I32:
3827 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3829 case ARM::ATOMIC_LOAD_SUB_I8:
3830 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3831 case ARM::ATOMIC_LOAD_SUB_I16:
3832 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3833 case ARM::ATOMIC_LOAD_SUB_I32:
3834 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3836 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3837 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3838 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3840 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3841 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3842 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3844 case ARM::tMOVCCr_pseudo: {
3845 // To "insert" a SELECT_CC instruction, we actually have to insert the
3846 // diamond control-flow pattern. The incoming instruction knows the
3847 // destination vreg to set, the condition code register to branch on, the
3848 // true/false values to select between, and a branch opcode to use.
3849 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3850 MachineFunction::iterator It = BB;
3856 // cmpTY ccX, r1, r2
3858 // fallthrough --> copy0MBB
3859 MachineBasicBlock *thisMBB = BB;
3860 MachineFunction *F = BB->getParent();
3861 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3862 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3863 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3864 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3865 F->insert(It, copy0MBB);
3866 F->insert(It, sinkMBB);
3867 // Update machine-CFG edges by first adding all successors of the current
3868 // block to the new block which will contain the Phi node for the select.
3869 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3870 E = BB->succ_end(); I != E; ++I)
3871 sinkMBB->addSuccessor(*I);
3872 // Next, remove all successors of the current block, and add the true
3873 // and fallthrough blocks as its successors.
3874 while (!BB->succ_empty())
3875 BB->removeSuccessor(BB->succ_begin());
3876 BB->addSuccessor(copy0MBB);
3877 BB->addSuccessor(sinkMBB);
3880 // %FalseValue = ...
3881 // # fallthrough to sinkMBB
3884 // Update machine-CFG edges
3885 BB->addSuccessor(sinkMBB);
3888 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3891 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3892 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3893 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3895 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3902 case ARM::t2SUBrSPi_:
3903 case ARM::t2SUBrSPi12_:
3904 case ARM::t2SUBrSPs_: {
3905 MachineFunction *MF = BB->getParent();
3906 unsigned DstReg = MI->getOperand(0).getReg();
3907 unsigned SrcReg = MI->getOperand(1).getReg();
3908 bool DstIsDead = MI->getOperand(0).isDead();
3909 bool SrcIsKill = MI->getOperand(1).isKill();
3911 if (SrcReg != ARM::SP) {
3912 // Copy the source to SP from virtual register.
3913 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3914 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3915 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3916 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3917 .addReg(SrcReg, getKillRegState(SrcIsKill));
3921 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3922 switch (MI->getOpcode()) {
3924 llvm_unreachable("Unexpected pseudo instruction!");
3930 OpOpc = ARM::tADDspr;
3933 OpOpc = ARM::tSUBspi;
3935 case ARM::t2SUBrSPi_:
3936 OpOpc = ARM::t2SUBrSPi;
3937 NeedPred = true; NeedCC = true;
3939 case ARM::t2SUBrSPi12_:
3940 OpOpc = ARM::t2SUBrSPi12;
3943 case ARM::t2SUBrSPs_:
3944 OpOpc = ARM::t2SUBrSPs;
3945 NeedPred = true; NeedCC = true; NeedOp3 = true;
3948 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3949 if (OpOpc == ARM::tAND)
3950 AddDefaultT1CC(MIB);
3951 MIB.addReg(ARM::SP);
3952 MIB.addOperand(MI->getOperand(2));
3954 MIB.addOperand(MI->getOperand(3));
3956 AddDefaultPred(MIB);
3960 // Copy the result from SP to virtual register.
3961 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3962 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3963 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3964 BuildMI(BB, dl, TII->get(CopyOpc))
3965 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3967 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3973 //===----------------------------------------------------------------------===//
3974 // ARM Optimization Hooks
3975 //===----------------------------------------------------------------------===//
3978 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3979 TargetLowering::DAGCombinerInfo &DCI) {
3980 SelectionDAG &DAG = DCI.DAG;
3981 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3982 EVT VT = N->getValueType(0);
3983 unsigned Opc = N->getOpcode();
3984 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3985 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3986 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3987 ISD::CondCode CC = ISD::SETCC_INVALID;
3990 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3992 SDValue CCOp = Slct.getOperand(0);
3993 if (CCOp.getOpcode() == ISD::SETCC)
3994 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3997 bool DoXform = false;
3999 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4002 if (LHS.getOpcode() == ISD::Constant &&
4003 cast<ConstantSDNode>(LHS)->isNullValue()) {
4005 } else if (CC != ISD::SETCC_INVALID &&
4006 RHS.getOpcode() == ISD::Constant &&
4007 cast<ConstantSDNode>(RHS)->isNullValue()) {
4008 std::swap(LHS, RHS);
4009 SDValue Op0 = Slct.getOperand(0);
4010 EVT OpVT = isSlctCC ? Op0.getValueType() :
4011 Op0.getOperand(0).getValueType();
4012 bool isInt = OpVT.isInteger();
4013 CC = ISD::getSetCCInverse(CC, isInt);
4015 if (!TLI.isCondCodeLegal(CC, OpVT))
4016 return SDValue(); // Inverse operator isn't legal.
4023 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4025 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4026 Slct.getOperand(0), Slct.getOperand(1), CC);
4027 SDValue CCOp = Slct.getOperand(0);
4029 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4030 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4031 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4032 CCOp, OtherOp, Result);
4037 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4038 static SDValue PerformADDCombine(SDNode *N,
4039 TargetLowering::DAGCombinerInfo &DCI) {
4040 // added by evan in r37685 with no testcase.
4041 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4043 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4044 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4045 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4046 if (Result.getNode()) return Result;
4048 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4049 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4050 if (Result.getNode()) return Result;
4056 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4057 static SDValue PerformSUBCombine(SDNode *N,
4058 TargetLowering::DAGCombinerInfo &DCI) {
4059 // added by evan in r37685 with no testcase.
4060 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4062 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4063 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4064 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4065 if (Result.getNode()) return Result;
4071 static SDValue PerformMULCombine(SDNode *N,
4072 TargetLowering::DAGCombinerInfo &DCI,
4073 const ARMSubtarget *Subtarget) {
4074 SelectionDAG &DAG = DCI.DAG;
4076 if (Subtarget->isThumb1Only())
4079 if (DAG.getMachineFunction().
4080 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4083 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4086 EVT VT = N->getValueType(0);
4090 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4094 uint64_t MulAmt = C->getZExtValue();
4095 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4096 ShiftAmt = ShiftAmt & (32 - 1);
4097 SDValue V = N->getOperand(0);
4098 DebugLoc DL = N->getDebugLoc();
4101 MulAmt >>= ShiftAmt;
4102 if (isPowerOf2_32(MulAmt - 1)) {
4103 // (mul x, 2^N + 1) => (add (shl x, N), x)
4104 Res = DAG.getNode(ISD::ADD, DL, VT,
4105 V, DAG.getNode(ISD::SHL, DL, VT,
4106 V, DAG.getConstant(Log2_32(MulAmt-1),
4108 } else if (isPowerOf2_32(MulAmt + 1)) {
4109 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4110 Res = DAG.getNode(ISD::SUB, DL, VT,
4111 DAG.getNode(ISD::SHL, DL, VT,
4112 V, DAG.getConstant(Log2_32(MulAmt+1),
4119 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4120 DAG.getConstant(ShiftAmt, MVT::i32));
4122 // Do not add new nodes to DAG combiner worklist.
4123 DCI.CombineTo(N, Res, false);
4127 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4128 /// ARMISD::VMOVRRD.
4129 static SDValue PerformVMOVRRDCombine(SDNode *N,
4130 TargetLowering::DAGCombinerInfo &DCI) {
4131 // fmrrd(fmdrr x, y) -> x,y
4132 SDValue InDouble = N->getOperand(0);
4133 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4134 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4138 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4139 /// operand of a vector shift operation, where all the elements of the
4140 /// build_vector must have the same constant integer value.
4141 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4142 // Ignore bit_converts.
4143 while (Op.getOpcode() == ISD::BIT_CONVERT)
4144 Op = Op.getOperand(0);
4145 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4146 APInt SplatBits, SplatUndef;
4147 unsigned SplatBitSize;
4149 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4150 HasAnyUndefs, ElementBits) ||
4151 SplatBitSize > ElementBits)
4153 Cnt = SplatBits.getSExtValue();
4157 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4158 /// operand of a vector shift left operation. That value must be in the range:
4159 /// 0 <= Value < ElementBits for a left shift; or
4160 /// 0 <= Value <= ElementBits for a long left shift.
4161 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4162 assert(VT.isVector() && "vector shift count is not a vector type");
4163 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4164 if (! getVShiftImm(Op, ElementBits, Cnt))
4166 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4169 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4170 /// operand of a vector shift right operation. For a shift opcode, the value
4171 /// is positive, but for an intrinsic the value count must be negative. The
4172 /// absolute value must be in the range:
4173 /// 1 <= |Value| <= ElementBits for a right shift; or
4174 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4175 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4177 assert(VT.isVector() && "vector shift count is not a vector type");
4178 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4179 if (! getVShiftImm(Op, ElementBits, Cnt))
4183 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4186 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4187 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4188 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4191 // Don't do anything for most intrinsics.
4194 // Vector shifts: check for immediate versions and lower them.
4195 // Note: This is done during DAG combining instead of DAG legalizing because
4196 // the build_vectors for 64-bit vector element shift counts are generally
4197 // not legal, and it is hard to see their values after they get legalized to
4198 // loads from a constant pool.
4199 case Intrinsic::arm_neon_vshifts:
4200 case Intrinsic::arm_neon_vshiftu:
4201 case Intrinsic::arm_neon_vshiftls:
4202 case Intrinsic::arm_neon_vshiftlu:
4203 case Intrinsic::arm_neon_vshiftn:
4204 case Intrinsic::arm_neon_vrshifts:
4205 case Intrinsic::arm_neon_vrshiftu:
4206 case Intrinsic::arm_neon_vrshiftn:
4207 case Intrinsic::arm_neon_vqshifts:
4208 case Intrinsic::arm_neon_vqshiftu:
4209 case Intrinsic::arm_neon_vqshiftsu:
4210 case Intrinsic::arm_neon_vqshiftns:
4211 case Intrinsic::arm_neon_vqshiftnu:
4212 case Intrinsic::arm_neon_vqshiftnsu:
4213 case Intrinsic::arm_neon_vqrshiftns:
4214 case Intrinsic::arm_neon_vqrshiftnu:
4215 case Intrinsic::arm_neon_vqrshiftnsu: {
4216 EVT VT = N->getOperand(1).getValueType();
4218 unsigned VShiftOpc = 0;
4221 case Intrinsic::arm_neon_vshifts:
4222 case Intrinsic::arm_neon_vshiftu:
4223 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4224 VShiftOpc = ARMISD::VSHL;
4227 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4228 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4229 ARMISD::VSHRs : ARMISD::VSHRu);
4234 case Intrinsic::arm_neon_vshiftls:
4235 case Intrinsic::arm_neon_vshiftlu:
4236 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4238 llvm_unreachable("invalid shift count for vshll intrinsic");
4240 case Intrinsic::arm_neon_vrshifts:
4241 case Intrinsic::arm_neon_vrshiftu:
4242 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4246 case Intrinsic::arm_neon_vqshifts:
4247 case Intrinsic::arm_neon_vqshiftu:
4248 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4252 case Intrinsic::arm_neon_vqshiftsu:
4253 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4255 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4257 case Intrinsic::arm_neon_vshiftn:
4258 case Intrinsic::arm_neon_vrshiftn:
4259 case Intrinsic::arm_neon_vqshiftns:
4260 case Intrinsic::arm_neon_vqshiftnu:
4261 case Intrinsic::arm_neon_vqshiftnsu:
4262 case Intrinsic::arm_neon_vqrshiftns:
4263 case Intrinsic::arm_neon_vqrshiftnu:
4264 case Intrinsic::arm_neon_vqrshiftnsu:
4265 // Narrowing shifts require an immediate right shift.
4266 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4268 llvm_unreachable("invalid shift count for narrowing vector shift "
4272 llvm_unreachable("unhandled vector shift");
4276 case Intrinsic::arm_neon_vshifts:
4277 case Intrinsic::arm_neon_vshiftu:
4278 // Opcode already set above.
4280 case Intrinsic::arm_neon_vshiftls:
4281 case Intrinsic::arm_neon_vshiftlu:
4282 if (Cnt == VT.getVectorElementType().getSizeInBits())
4283 VShiftOpc = ARMISD::VSHLLi;
4285 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4286 ARMISD::VSHLLs : ARMISD::VSHLLu);
4288 case Intrinsic::arm_neon_vshiftn:
4289 VShiftOpc = ARMISD::VSHRN; break;
4290 case Intrinsic::arm_neon_vrshifts:
4291 VShiftOpc = ARMISD::VRSHRs; break;
4292 case Intrinsic::arm_neon_vrshiftu:
4293 VShiftOpc = ARMISD::VRSHRu; break;
4294 case Intrinsic::arm_neon_vrshiftn:
4295 VShiftOpc = ARMISD::VRSHRN; break;
4296 case Intrinsic::arm_neon_vqshifts:
4297 VShiftOpc = ARMISD::VQSHLs; break;
4298 case Intrinsic::arm_neon_vqshiftu:
4299 VShiftOpc = ARMISD::VQSHLu; break;
4300 case Intrinsic::arm_neon_vqshiftsu:
4301 VShiftOpc = ARMISD::VQSHLsu; break;
4302 case Intrinsic::arm_neon_vqshiftns:
4303 VShiftOpc = ARMISD::VQSHRNs; break;
4304 case Intrinsic::arm_neon_vqshiftnu:
4305 VShiftOpc = ARMISD::VQSHRNu; break;
4306 case Intrinsic::arm_neon_vqshiftnsu:
4307 VShiftOpc = ARMISD::VQSHRNsu; break;
4308 case Intrinsic::arm_neon_vqrshiftns:
4309 VShiftOpc = ARMISD::VQRSHRNs; break;
4310 case Intrinsic::arm_neon_vqrshiftnu:
4311 VShiftOpc = ARMISD::VQRSHRNu; break;
4312 case Intrinsic::arm_neon_vqrshiftnsu:
4313 VShiftOpc = ARMISD::VQRSHRNsu; break;
4316 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4317 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4320 case Intrinsic::arm_neon_vshiftins: {
4321 EVT VT = N->getOperand(1).getValueType();
4323 unsigned VShiftOpc = 0;
4325 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4326 VShiftOpc = ARMISD::VSLI;
4327 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4328 VShiftOpc = ARMISD::VSRI;
4330 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4333 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4334 N->getOperand(1), N->getOperand(2),
4335 DAG.getConstant(Cnt, MVT::i32));
4338 case Intrinsic::arm_neon_vqrshifts:
4339 case Intrinsic::arm_neon_vqrshiftu:
4340 // No immediate versions of these to check for.
4347 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4348 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4349 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4350 /// vector element shift counts are generally not legal, and it is hard to see
4351 /// their values after they get legalized to loads from a constant pool.
4352 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4353 const ARMSubtarget *ST) {
4354 EVT VT = N->getValueType(0);
4356 // Nothing to be done for scalar shifts.
4357 if (! VT.isVector())
4360 assert(ST->hasNEON() && "unexpected vector shift");
4363 switch (N->getOpcode()) {
4364 default: llvm_unreachable("unexpected shift opcode");
4367 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4368 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4369 DAG.getConstant(Cnt, MVT::i32));
4374 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4375 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4376 ARMISD::VSHRs : ARMISD::VSHRu);
4377 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4378 DAG.getConstant(Cnt, MVT::i32));
4384 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4385 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4386 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4387 const ARMSubtarget *ST) {
4388 SDValue N0 = N->getOperand(0);
4390 // Check for sign- and zero-extensions of vector extract operations of 8-
4391 // and 16-bit vector elements. NEON supports these directly. They are
4392 // handled during DAG combining because type legalization will promote them
4393 // to 32-bit types and it is messy to recognize the operations after that.
4394 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4395 SDValue Vec = N0.getOperand(0);
4396 SDValue Lane = N0.getOperand(1);
4397 EVT VT = N->getValueType(0);
4398 EVT EltVT = N0.getValueType();
4399 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4401 if (VT == MVT::i32 &&
4402 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4403 TLI.isTypeLegal(Vec.getValueType())) {
4406 switch (N->getOpcode()) {
4407 default: llvm_unreachable("unexpected opcode");
4408 case ISD::SIGN_EXTEND:
4409 Opc = ARMISD::VGETLANEs;
4411 case ISD::ZERO_EXTEND:
4412 case ISD::ANY_EXTEND:
4413 Opc = ARMISD::VGETLANEu;
4416 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4423 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4424 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4425 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4426 const ARMSubtarget *ST) {
4427 // If the target supports NEON, try to use vmax/vmin instructions for f32
4428 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4429 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4430 // a NaN; only do the transformation when it matches that behavior.
4432 // For now only do this when using NEON for FP operations; if using VFP, it
4433 // is not obvious that the benefit outweighs the cost of switching to the
4435 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4436 N->getValueType(0) != MVT::f32)
4439 SDValue CondLHS = N->getOperand(0);
4440 SDValue CondRHS = N->getOperand(1);
4441 SDValue LHS = N->getOperand(2);
4442 SDValue RHS = N->getOperand(3);
4443 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4445 unsigned Opcode = 0;
4447 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4448 IsReversed = false; // x CC y ? x : y
4449 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4450 IsReversed = true ; // x CC y ? y : x
4464 // If LHS is NaN, an ordered comparison will be false and the result will
4465 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4466 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4467 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4468 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4470 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4471 // will return -0, so vmin can only be used for unsafe math or if one of
4472 // the operands is known to be nonzero.
4473 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4475 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4477 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4486 // If LHS is NaN, an ordered comparison will be false and the result will
4487 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4488 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4489 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4490 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4492 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4493 // will return +0, so vmax can only be used for unsafe math or if one of
4494 // the operands is known to be nonzero.
4495 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4497 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4499 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4505 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4508 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4509 DAGCombinerInfo &DCI) const {
4510 switch (N->getOpcode()) {
4512 case ISD::ADD: return PerformADDCombine(N, DCI);
4513 case ISD::SUB: return PerformSUBCombine(N, DCI);
4514 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4515 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4516 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4519 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4520 case ISD::SIGN_EXTEND:
4521 case ISD::ZERO_EXTEND:
4522 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4523 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4528 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4529 if (!Subtarget->hasV6Ops())
4530 // Pre-v6 does not support unaligned mem access.
4533 // v6+ may or may not support unaligned mem access depending on the system
4535 // FIXME: This is pretty conservative. Should we provide cmdline option to
4536 // control the behaviour?
4537 if (!Subtarget->isTargetDarwin())
4541 switch (VT.getSimpleVT().SimpleTy) {
4548 // FIXME: VLD1 etc with standard alignment is legal.
4552 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4557 switch (VT.getSimpleVT().SimpleTy) {
4558 default: return false;
4573 if ((V & (Scale - 1)) != 0)
4576 return V == (V & ((1LL << 5) - 1));
4579 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4580 const ARMSubtarget *Subtarget) {
4587 switch (VT.getSimpleVT().SimpleTy) {
4588 default: return false;
4593 // + imm12 or - imm8
4595 return V == (V & ((1LL << 8) - 1));
4596 return V == (V & ((1LL << 12) - 1));
4599 // Same as ARM mode. FIXME: NEON?
4600 if (!Subtarget->hasVFP2())
4605 return V == (V & ((1LL << 8) - 1));
4609 /// isLegalAddressImmediate - Return true if the integer value can be used
4610 /// as the offset of the target addressing mode for load / store of the
4612 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4613 const ARMSubtarget *Subtarget) {
4620 if (Subtarget->isThumb1Only())
4621 return isLegalT1AddressImmediate(V, VT);
4622 else if (Subtarget->isThumb2())
4623 return isLegalT2AddressImmediate(V, VT, Subtarget);
4628 switch (VT.getSimpleVT().SimpleTy) {
4629 default: return false;
4634 return V == (V & ((1LL << 12) - 1));
4637 return V == (V & ((1LL << 8) - 1));
4640 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4645 return V == (V & ((1LL << 8) - 1));
4649 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4651 int Scale = AM.Scale;
4655 switch (VT.getSimpleVT().SimpleTy) {
4656 default: return false;
4665 return Scale == 2 || Scale == 4 || Scale == 8;
4668 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4672 // Note, we allow "void" uses (basically, uses that aren't loads or
4673 // stores), because arm allows folding a scale into many arithmetic
4674 // operations. This should be made more precise and revisited later.
4676 // Allow r << imm, but the imm has to be a multiple of two.
4677 if (Scale & 1) return false;
4678 return isPowerOf2_32(Scale);
4682 /// isLegalAddressingMode - Return true if the addressing mode represented
4683 /// by AM is legal for this target, for a load/store of the specified type.
4684 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4685 const Type *Ty) const {
4686 EVT VT = getValueType(Ty, true);
4687 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4690 // Can never fold addr of global into load/store.
4695 case 0: // no scale reg, must be "r+i" or "r", or "i".
4698 if (Subtarget->isThumb1Only())
4702 // ARM doesn't support any R+R*scale+imm addr modes.
4709 if (Subtarget->isThumb2())
4710 return isLegalT2ScaledAddressingMode(AM, VT);
4712 int Scale = AM.Scale;
4713 switch (VT.getSimpleVT().SimpleTy) {
4714 default: return false;
4718 if (Scale < 0) Scale = -Scale;
4722 return isPowerOf2_32(Scale & ~1);
4726 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4731 // Note, we allow "void" uses (basically, uses that aren't loads or
4732 // stores), because arm allows folding a scale into many arithmetic
4733 // operations. This should be made more precise and revisited later.
4735 // Allow r << imm, but the imm has to be a multiple of two.
4736 if (Scale & 1) return false;
4737 return isPowerOf2_32(Scale);
4744 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4745 /// icmp immediate, that is the target has icmp instructions which can compare
4746 /// a register against the immediate without having to materialize the
4747 /// immediate into a register.
4748 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4749 if (!Subtarget->isThumb())
4750 return ARM_AM::getSOImmVal(Imm) != -1;
4751 if (Subtarget->isThumb2())
4752 return ARM_AM::getT2SOImmVal(Imm) != -1;
4753 return Imm >= 0 && Imm <= 255;
4756 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4757 bool isSEXTLoad, SDValue &Base,
4758 SDValue &Offset, bool &isInc,
4759 SelectionDAG &DAG) {
4760 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4763 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4765 Base = Ptr->getOperand(0);
4766 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4767 int RHSC = (int)RHS->getZExtValue();
4768 if (RHSC < 0 && RHSC > -256) {
4769 assert(Ptr->getOpcode() == ISD::ADD);
4771 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4775 isInc = (Ptr->getOpcode() == ISD::ADD);
4776 Offset = Ptr->getOperand(1);
4778 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4780 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4781 int RHSC = (int)RHS->getZExtValue();
4782 if (RHSC < 0 && RHSC > -0x1000) {
4783 assert(Ptr->getOpcode() == ISD::ADD);
4785 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4786 Base = Ptr->getOperand(0);
4791 if (Ptr->getOpcode() == ISD::ADD) {
4793 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4794 if (ShOpcVal != ARM_AM::no_shift) {
4795 Base = Ptr->getOperand(1);
4796 Offset = Ptr->getOperand(0);
4798 Base = Ptr->getOperand(0);
4799 Offset = Ptr->getOperand(1);
4804 isInc = (Ptr->getOpcode() == ISD::ADD);
4805 Base = Ptr->getOperand(0);
4806 Offset = Ptr->getOperand(1);
4810 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4814 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4815 bool isSEXTLoad, SDValue &Base,
4816 SDValue &Offset, bool &isInc,
4817 SelectionDAG &DAG) {
4818 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4821 Base = Ptr->getOperand(0);
4822 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4823 int RHSC = (int)RHS->getZExtValue();
4824 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4825 assert(Ptr->getOpcode() == ISD::ADD);
4827 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4829 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4830 isInc = Ptr->getOpcode() == ISD::ADD;
4831 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4839 /// getPreIndexedAddressParts - returns true by value, base pointer and
4840 /// offset pointer and addressing mode by reference if the node's address
4841 /// can be legally represented as pre-indexed load / store address.
4843 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4845 ISD::MemIndexedMode &AM,
4846 SelectionDAG &DAG) const {
4847 if (Subtarget->isThumb1Only())
4852 bool isSEXTLoad = false;
4853 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4854 Ptr = LD->getBasePtr();
4855 VT = LD->getMemoryVT();
4856 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4857 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4858 Ptr = ST->getBasePtr();
4859 VT = ST->getMemoryVT();
4864 bool isLegal = false;
4865 if (Subtarget->isThumb2())
4866 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4867 Offset, isInc, DAG);
4869 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4870 Offset, isInc, DAG);
4874 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4878 /// getPostIndexedAddressParts - returns true by value, base pointer and
4879 /// offset pointer and addressing mode by reference if this node can be
4880 /// combined with a load / store to form a post-indexed load / store.
4881 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4884 ISD::MemIndexedMode &AM,
4885 SelectionDAG &DAG) const {
4886 if (Subtarget->isThumb1Only())
4891 bool isSEXTLoad = false;
4892 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4893 VT = LD->getMemoryVT();
4894 Ptr = LD->getBasePtr();
4895 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4896 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4897 VT = ST->getMemoryVT();
4898 Ptr = ST->getBasePtr();
4903 bool isLegal = false;
4904 if (Subtarget->isThumb2())
4905 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4908 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4914 // Swap base ptr and offset to catch more post-index load / store when
4915 // it's legal. In Thumb2 mode, offset must be an immediate.
4916 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4917 !Subtarget->isThumb2())
4918 std::swap(Base, Offset);
4920 // Post-indexed load / store update the base pointer.
4925 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4929 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4933 const SelectionDAG &DAG,
4934 unsigned Depth) const {
4935 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4936 switch (Op.getOpcode()) {
4938 case ARMISD::CMOV: {
4939 // Bits are known zero/one if known on the LHS and RHS.
4940 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4941 if (KnownZero == 0 && KnownOne == 0) return;
4943 APInt KnownZeroRHS, KnownOneRHS;
4944 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4945 KnownZeroRHS, KnownOneRHS, Depth+1);
4946 KnownZero &= KnownZeroRHS;
4947 KnownOne &= KnownOneRHS;
4953 //===----------------------------------------------------------------------===//
4954 // ARM Inline Assembly Support
4955 //===----------------------------------------------------------------------===//
4957 /// getConstraintType - Given a constraint letter, return the type of
4958 /// constraint it is for this target.
4959 ARMTargetLowering::ConstraintType
4960 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4961 if (Constraint.size() == 1) {
4962 switch (Constraint[0]) {
4964 case 'l': return C_RegisterClass;
4965 case 'w': return C_RegisterClass;
4968 return TargetLowering::getConstraintType(Constraint);
4971 std::pair<unsigned, const TargetRegisterClass*>
4972 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4974 if (Constraint.size() == 1) {
4975 // GCC ARM Constraint Letters
4976 switch (Constraint[0]) {
4978 if (Subtarget->isThumb())
4979 return std::make_pair(0U, ARM::tGPRRegisterClass);
4981 return std::make_pair(0U, ARM::GPRRegisterClass);
4983 return std::make_pair(0U, ARM::GPRRegisterClass);
4986 return std::make_pair(0U, ARM::SPRRegisterClass);
4987 if (VT.getSizeInBits() == 64)
4988 return std::make_pair(0U, ARM::DPRRegisterClass);
4989 if (VT.getSizeInBits() == 128)
4990 return std::make_pair(0U, ARM::QPRRegisterClass);
4994 if (StringRef("{cc}").equals_lower(Constraint))
4995 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
4997 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5000 std::vector<unsigned> ARMTargetLowering::
5001 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5003 if (Constraint.size() != 1)
5004 return std::vector<unsigned>();
5006 switch (Constraint[0]) { // GCC ARM Constraint Letters
5009 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5010 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5013 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5014 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5015 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5016 ARM::R12, ARM::LR, 0);
5019 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5020 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5021 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5022 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5023 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5024 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5025 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5026 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5027 if (VT.getSizeInBits() == 64)
5028 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5029 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5030 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5031 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5032 if (VT.getSizeInBits() == 128)
5033 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5034 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5038 return std::vector<unsigned>();
5041 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5042 /// vector. If it is invalid, don't add anything to Ops.
5043 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5046 std::vector<SDValue>&Ops,
5047 SelectionDAG &DAG) const {
5048 SDValue Result(0, 0);
5050 switch (Constraint) {
5052 case 'I': case 'J': case 'K': case 'L':
5053 case 'M': case 'N': case 'O':
5054 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5058 int64_t CVal64 = C->getSExtValue();
5059 int CVal = (int) CVal64;
5060 // None of these constraints allow values larger than 32 bits. Check
5061 // that the value fits in an int.
5065 switch (Constraint) {
5067 if (Subtarget->isThumb1Only()) {
5068 // This must be a constant between 0 and 255, for ADD
5070 if (CVal >= 0 && CVal <= 255)
5072 } else if (Subtarget->isThumb2()) {
5073 // A constant that can be used as an immediate value in a
5074 // data-processing instruction.
5075 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5078 // A constant that can be used as an immediate value in a
5079 // data-processing instruction.
5080 if (ARM_AM::getSOImmVal(CVal) != -1)
5086 if (Subtarget->isThumb()) { // FIXME thumb2
5087 // This must be a constant between -255 and -1, for negated ADD
5088 // immediates. This can be used in GCC with an "n" modifier that
5089 // prints the negated value, for use with SUB instructions. It is
5090 // not useful otherwise but is implemented for compatibility.
5091 if (CVal >= -255 && CVal <= -1)
5094 // This must be a constant between -4095 and 4095. It is not clear
5095 // what this constraint is intended for. Implemented for
5096 // compatibility with GCC.
5097 if (CVal >= -4095 && CVal <= 4095)
5103 if (Subtarget->isThumb1Only()) {
5104 // A 32-bit value where only one byte has a nonzero value. Exclude
5105 // zero to match GCC. This constraint is used by GCC internally for
5106 // constants that can be loaded with a move/shift combination.
5107 // It is not useful otherwise but is implemented for compatibility.
5108 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5110 } else if (Subtarget->isThumb2()) {
5111 // A constant whose bitwise inverse can be used as an immediate
5112 // value in a data-processing instruction. This can be used in GCC
5113 // with a "B" modifier that prints the inverted value, for use with
5114 // BIC and MVN instructions. It is not useful otherwise but is
5115 // implemented for compatibility.
5116 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5119 // A constant whose bitwise inverse can be used as an immediate
5120 // value in a data-processing instruction. This can be used in GCC
5121 // with a "B" modifier that prints the inverted value, for use with
5122 // BIC and MVN instructions. It is not useful otherwise but is
5123 // implemented for compatibility.
5124 if (ARM_AM::getSOImmVal(~CVal) != -1)
5130 if (Subtarget->isThumb1Only()) {
5131 // This must be a constant between -7 and 7,
5132 // for 3-operand ADD/SUB immediate instructions.
5133 if (CVal >= -7 && CVal < 7)
5135 } else if (Subtarget->isThumb2()) {
5136 // A constant whose negation can be used as an immediate value in a
5137 // data-processing instruction. This can be used in GCC with an "n"
5138 // modifier that prints the negated value, for use with SUB
5139 // instructions. It is not useful otherwise but is implemented for
5141 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5144 // A constant whose negation can be used as an immediate value in a
5145 // data-processing instruction. This can be used in GCC with an "n"
5146 // modifier that prints the negated value, for use with SUB
5147 // instructions. It is not useful otherwise but is implemented for
5149 if (ARM_AM::getSOImmVal(-CVal) != -1)
5155 if (Subtarget->isThumb()) { // FIXME thumb2
5156 // This must be a multiple of 4 between 0 and 1020, for
5157 // ADD sp + immediate.
5158 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5161 // A power of two or a constant between 0 and 32. This is used in
5162 // GCC for the shift amount on shifted register operands, but it is
5163 // useful in general for any shift amounts.
5164 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5170 if (Subtarget->isThumb()) { // FIXME thumb2
5171 // This must be a constant between 0 and 31, for shift amounts.
5172 if (CVal >= 0 && CVal <= 31)
5178 if (Subtarget->isThumb()) { // FIXME thumb2
5179 // This must be a multiple of 4 between -508 and 508, for
5180 // ADD/SUB sp = sp + immediate.
5181 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5186 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5190 if (Result.getNode()) {
5191 Ops.push_back(Result);
5194 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
5199 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5200 // The ARM target isn't yet aware of offsets.
5204 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5205 APInt Imm = FPImm.bitcastToAPInt();
5206 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5207 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5208 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5210 // We can handle 4 bits of mantissa.
5211 // mantissa = (16+UInt(e:f:g:h))/16.
5212 if (Mantissa & 0x7ffff)
5215 if ((Mantissa & 0xf) != Mantissa)
5218 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5219 if (Exp < -3 || Exp > 4)
5221 Exp = ((Exp+3) & 0x7) ^ 4;
5223 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5226 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5227 APInt Imm = FPImm.bitcastToAPInt();
5228 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5229 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5230 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5232 // We can handle 4 bits of mantissa.
5233 // mantissa = (16+UInt(e:f:g:h))/16.
5234 if (Mantissa & 0xffffffffffffLL)
5237 if ((Mantissa & 0xf) != Mantissa)
5240 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5241 if (Exp < -3 || Exp > 4)
5243 Exp = ((Exp+3) & 0x7) ^ 4;
5245 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5248 /// isFPImmLegal - Returns true if the target can instruction select the
5249 /// specified FP immediate natively. If false, the legalizer will
5250 /// materialize the FP immediate as a load from a constant pool.
5251 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5252 if (!Subtarget->hasVFP3())
5255 return ARM::getVFPf32Imm(Imm) != -1;
5257 return ARM::getVFPf64Imm(Imm) != -1;