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)"),
71 EnableARMCodePlacement("arm-code-placement", cl::Hidden,
72 cl::desc("Enable code placement pass for ARM"),
75 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
76 CCValAssign::LocInfo &LocInfo,
77 ISD::ArgFlagsTy &ArgFlags,
79 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
80 CCValAssign::LocInfo &LocInfo,
81 ISD::ArgFlagsTy &ArgFlags,
83 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
84 CCValAssign::LocInfo &LocInfo,
85 ISD::ArgFlagsTy &ArgFlags,
87 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
88 CCValAssign::LocInfo &LocInfo,
89 ISD::ArgFlagsTy &ArgFlags,
92 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
93 EVT PromotedBitwiseVT) {
94 if (VT != PromotedLdStVT) {
95 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
96 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
97 PromotedLdStVT.getSimpleVT());
99 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
100 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
101 PromotedLdStVT.getSimpleVT());
104 EVT ElemTy = VT.getVectorElementType();
105 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
106 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
107 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
108 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
109 if (ElemTy != MVT::i32) {
110 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
111 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
112 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
116 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
117 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
118 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
119 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
121 if (VT.isInteger()) {
122 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
123 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
124 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
127 // Promote all bit-wise operations.
128 if (VT.isInteger() && VT != PromotedBitwiseVT) {
129 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
130 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
131 PromotedBitwiseVT.getSimpleVT());
132 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
133 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
134 PromotedBitwiseVT.getSimpleVT());
135 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
136 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
137 PromotedBitwiseVT.getSimpleVT());
140 // Neon does not support vector divide/remainder operations.
141 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
142 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
143 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
144 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
145 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
146 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
149 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
150 addRegisterClass(VT, ARM::DPRRegisterClass);
151 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
154 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
155 addRegisterClass(VT, ARM::QPRRegisterClass);
156 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
159 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
160 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
161 return new TargetLoweringObjectFileMachO();
163 return new ARMElfTargetObjectFile();
166 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
167 : TargetLowering(TM, createTLOF(TM)) {
168 Subtarget = &TM.getSubtarget<ARMSubtarget>();
170 if (Subtarget->isTargetDarwin()) {
171 // Uses VFP for Thumb libfuncs if available.
172 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
173 // Single-precision floating-point arithmetic.
174 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
175 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
176 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
177 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
179 // Double-precision floating-point arithmetic.
180 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
181 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
182 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
183 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
185 // Single-precision comparisons.
186 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
187 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
188 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
189 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
190 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
191 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
192 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
193 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
195 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
204 // Double-precision comparisons.
205 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
206 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
207 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
208 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
209 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
210 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
211 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
212 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
214 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
223 // Floating-point to integer conversions.
224 // i64 conversions are done via library routines even when generating VFP
225 // instructions, so use the same ones.
226 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
227 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
228 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
229 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
231 // Conversions between floating types.
232 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
233 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
235 // Integer to floating-point conversions.
236 // i64 conversions are done via library routines even when generating VFP
237 // instructions, so use the same ones.
238 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
239 // e.g., __floatunsidf vs. __floatunssidfvfp.
240 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
241 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
242 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
243 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
247 // These libcalls are not available in 32-bit.
248 setLibcallName(RTLIB::SHL_I128, 0);
249 setLibcallName(RTLIB::SRL_I128, 0);
250 setLibcallName(RTLIB::SRA_I128, 0);
252 // Libcalls should use the AAPCS base standard ABI, even if hard float
253 // is in effect, as per the ARM RTABI specification, section 4.1.2.
254 if (Subtarget->isAAPCS_ABI()) {
255 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
256 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
257 CallingConv::ARM_AAPCS);
261 if (Subtarget->isThumb1Only())
262 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
264 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
265 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
266 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
267 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
269 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
272 if (Subtarget->hasNEON()) {
273 addDRTypeForNEON(MVT::v2f32);
274 addDRTypeForNEON(MVT::v8i8);
275 addDRTypeForNEON(MVT::v4i16);
276 addDRTypeForNEON(MVT::v2i32);
277 addDRTypeForNEON(MVT::v1i64);
279 addQRTypeForNEON(MVT::v4f32);
280 addQRTypeForNEON(MVT::v2f64);
281 addQRTypeForNEON(MVT::v16i8);
282 addQRTypeForNEON(MVT::v8i16);
283 addQRTypeForNEON(MVT::v4i32);
284 addQRTypeForNEON(MVT::v2i64);
286 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
287 // neither Neon nor VFP support any arithmetic operations on it.
288 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
289 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
290 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
291 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
292 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
293 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
294 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
295 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
296 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
297 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
298 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
299 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
300 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
301 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
302 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
303 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
304 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
305 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
306 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
307 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
308 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
309 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
310 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
311 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
313 // Neon does not support some operations on v1i64 and v2i64 types.
314 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
315 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
316 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
317 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
319 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
320 setTargetDAGCombine(ISD::SHL);
321 setTargetDAGCombine(ISD::SRL);
322 setTargetDAGCombine(ISD::SRA);
323 setTargetDAGCombine(ISD::SIGN_EXTEND);
324 setTargetDAGCombine(ISD::ZERO_EXTEND);
325 setTargetDAGCombine(ISD::ANY_EXTEND);
326 setTargetDAGCombine(ISD::SELECT_CC);
329 computeRegisterProperties();
331 // ARM does not have f32 extending load.
332 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
334 // ARM does not have i1 sign extending load.
335 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
337 // ARM supports all 4 flavors of integer indexed load / store.
338 if (!Subtarget->isThumb1Only()) {
339 for (unsigned im = (unsigned)ISD::PRE_INC;
340 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
341 setIndexedLoadAction(im, MVT::i1, Legal);
342 setIndexedLoadAction(im, MVT::i8, Legal);
343 setIndexedLoadAction(im, MVT::i16, Legal);
344 setIndexedLoadAction(im, MVT::i32, Legal);
345 setIndexedStoreAction(im, MVT::i1, Legal);
346 setIndexedStoreAction(im, MVT::i8, Legal);
347 setIndexedStoreAction(im, MVT::i16, Legal);
348 setIndexedStoreAction(im, MVT::i32, Legal);
352 // i64 operation support.
353 if (Subtarget->isThumb1Only()) {
354 setOperationAction(ISD::MUL, MVT::i64, Expand);
355 setOperationAction(ISD::MULHU, MVT::i32, Expand);
356 setOperationAction(ISD::MULHS, MVT::i32, Expand);
357 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
358 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
360 setOperationAction(ISD::MUL, MVT::i64, Expand);
361 setOperationAction(ISD::MULHU, MVT::i32, Expand);
362 if (!Subtarget->hasV6Ops())
363 setOperationAction(ISD::MULHS, MVT::i32, Expand);
365 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
366 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
367 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
368 setOperationAction(ISD::SRL, MVT::i64, Custom);
369 setOperationAction(ISD::SRA, MVT::i64, Custom);
371 // ARM does not have ROTL.
372 setOperationAction(ISD::ROTL, MVT::i32, Expand);
373 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
374 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
375 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
376 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
378 // Only ARMv6 has BSWAP.
379 if (!Subtarget->hasV6Ops())
380 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
382 // These are expanded into libcalls.
383 if (!Subtarget->hasDivide()) {
384 // v7M has a hardware divider
385 setOperationAction(ISD::SDIV, MVT::i32, Expand);
386 setOperationAction(ISD::UDIV, MVT::i32, Expand);
388 setOperationAction(ISD::SREM, MVT::i32, Expand);
389 setOperationAction(ISD::UREM, MVT::i32, Expand);
390 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
391 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
393 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
394 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
395 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
396 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
397 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
399 setOperationAction(ISD::TRAP, MVT::Other, Legal);
401 // Use the default implementation.
402 setOperationAction(ISD::VASTART, MVT::Other, Custom);
403 setOperationAction(ISD::VAARG, MVT::Other, Expand);
404 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
405 setOperationAction(ISD::VAEND, MVT::Other, Expand);
406 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
407 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
408 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
409 // FIXME: Shouldn't need this, since no register is used, but the legalizer
410 // doesn't yet know how to not do that for SjLj.
411 setExceptionSelectorRegister(ARM::R0);
412 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
413 // Handle atomics directly for ARMv[67] (except for Thumb1), otherwise
414 // use the default expansion.
415 bool canHandleAtomics =
416 (Subtarget->hasV7Ops() ||
417 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only()));
418 if (canHandleAtomics) {
419 // membarrier needs custom lowering; the rest are legal and handled
421 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
423 // Set them all for expansion, which will force libcalls.
424 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
425 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Expand);
426 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Expand);
427 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
428 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, Expand);
429 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, Expand);
430 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
431 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, Expand);
432 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, Expand);
433 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
434 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Expand);
435 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Expand);
436 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
437 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i8, Expand);
438 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i16, Expand);
439 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
440 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, Expand);
441 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, Expand);
442 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
443 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, Expand);
444 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, Expand);
445 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
446 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i8, Expand);
447 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i16, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
449 // Since the libcalls include locking, fold in the fences
450 setShouldFoldAtomicFences(true);
452 // 64-bit versions are always libcalls (for now)
453 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Expand);
454 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
455 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Expand);
456 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
457 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Expand);
458 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Expand);
459 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Expand);
460 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Expand);
462 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
463 if (!Subtarget->hasV6Ops()) {
464 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
465 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
467 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
469 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
470 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
471 // iff target supports vfp2.
472 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
474 // We want to custom lower some of our intrinsics.
475 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
476 if (Subtarget->isTargetDarwin()) {
477 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
478 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
481 setOperationAction(ISD::SETCC, MVT::i32, Expand);
482 setOperationAction(ISD::SETCC, MVT::f32, Expand);
483 setOperationAction(ISD::SETCC, MVT::f64, Expand);
484 setOperationAction(ISD::SELECT, MVT::i32, Expand);
485 setOperationAction(ISD::SELECT, MVT::f32, Expand);
486 setOperationAction(ISD::SELECT, MVT::f64, Expand);
487 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
488 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
489 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
491 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
492 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
493 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
494 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
495 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
497 // We don't support sin/cos/fmod/copysign/pow
498 setOperationAction(ISD::FSIN, MVT::f64, Expand);
499 setOperationAction(ISD::FSIN, MVT::f32, Expand);
500 setOperationAction(ISD::FCOS, MVT::f32, Expand);
501 setOperationAction(ISD::FCOS, MVT::f64, Expand);
502 setOperationAction(ISD::FREM, MVT::f64, Expand);
503 setOperationAction(ISD::FREM, MVT::f32, Expand);
504 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
505 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
506 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
508 setOperationAction(ISD::FPOW, MVT::f64, Expand);
509 setOperationAction(ISD::FPOW, MVT::f32, Expand);
511 // Various VFP goodness
512 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
513 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
514 if (Subtarget->hasVFP2()) {
515 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
516 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
517 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
518 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
520 // Special handling for half-precision FP.
521 if (!Subtarget->hasFP16()) {
522 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
523 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
527 // We have target-specific dag combine patterns for the following nodes:
528 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
529 setTargetDAGCombine(ISD::ADD);
530 setTargetDAGCombine(ISD::SUB);
531 setTargetDAGCombine(ISD::MUL);
533 setStackPointerRegisterToSaveRestore(ARM::SP);
535 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
536 setSchedulingPreference(Sched::RegPressure);
538 setSchedulingPreference(Sched::Hybrid);
540 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
542 // On ARM arguments smaller than 4 bytes are extended, so all arguments
543 // are at least 4 bytes aligned.
544 setMinStackArgumentAlignment(4);
546 if (EnableARMCodePlacement)
547 benefitFromCodePlacementOpt = true;
550 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
553 case ARMISD::Wrapper: return "ARMISD::Wrapper";
554 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
555 case ARMISD::CALL: return "ARMISD::CALL";
556 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
557 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
558 case ARMISD::tCALL: return "ARMISD::tCALL";
559 case ARMISD::BRCOND: return "ARMISD::BRCOND";
560 case ARMISD::BR_JT: return "ARMISD::BR_JT";
561 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
562 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
563 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
564 case ARMISD::CMP: return "ARMISD::CMP";
565 case ARMISD::CMPZ: return "ARMISD::CMPZ";
566 case ARMISD::CMPFP: return "ARMISD::CMPFP";
567 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
568 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
569 case ARMISD::CMOV: return "ARMISD::CMOV";
570 case ARMISD::CNEG: return "ARMISD::CNEG";
572 case ARMISD::RBIT: return "ARMISD::RBIT";
574 case ARMISD::FTOSI: return "ARMISD::FTOSI";
575 case ARMISD::FTOUI: return "ARMISD::FTOUI";
576 case ARMISD::SITOF: return "ARMISD::SITOF";
577 case ARMISD::UITOF: return "ARMISD::UITOF";
579 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
580 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
581 case ARMISD::RRX: return "ARMISD::RRX";
583 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
584 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
586 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
587 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
589 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
591 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
593 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
595 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
596 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
598 case ARMISD::VCEQ: return "ARMISD::VCEQ";
599 case ARMISD::VCGE: return "ARMISD::VCGE";
600 case ARMISD::VCGEU: return "ARMISD::VCGEU";
601 case ARMISD::VCGT: return "ARMISD::VCGT";
602 case ARMISD::VCGTU: return "ARMISD::VCGTU";
603 case ARMISD::VTST: return "ARMISD::VTST";
605 case ARMISD::VSHL: return "ARMISD::VSHL";
606 case ARMISD::VSHRs: return "ARMISD::VSHRs";
607 case ARMISD::VSHRu: return "ARMISD::VSHRu";
608 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
609 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
610 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
611 case ARMISD::VSHRN: return "ARMISD::VSHRN";
612 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
613 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
614 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
615 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
616 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
617 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
618 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
619 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
620 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
621 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
622 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
623 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
624 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
625 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
626 case ARMISD::VDUP: return "ARMISD::VDUP";
627 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
628 case ARMISD::VEXT: return "ARMISD::VEXT";
629 case ARMISD::VREV64: return "ARMISD::VREV64";
630 case ARMISD::VREV32: return "ARMISD::VREV32";
631 case ARMISD::VREV16: return "ARMISD::VREV16";
632 case ARMISD::VZIP: return "ARMISD::VZIP";
633 case ARMISD::VUZP: return "ARMISD::VUZP";
634 case ARMISD::VTRN: return "ARMISD::VTRN";
635 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
636 case ARMISD::FMAX: return "ARMISD::FMAX";
637 case ARMISD::FMIN: return "ARMISD::FMIN";
641 /// getRegClassFor - Return the register class that should be used for the
642 /// specified value type.
643 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
644 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
645 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
646 // load / store 4 to 8 consecutive D registers.
647 if (Subtarget->hasNEON()) {
648 if (VT == MVT::v4i64)
649 return ARM::QQPRRegisterClass;
650 else if (VT == MVT::v8i64)
651 return ARM::QQQQPRRegisterClass;
653 return TargetLowering::getRegClassFor(VT);
656 /// getFunctionAlignment - Return the Log2 alignment of this function.
657 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
658 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
661 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
662 unsigned NumVals = N->getNumValues();
664 return Sched::RegPressure;
666 for (unsigned i = 0; i != NumVals; ++i) {
667 EVT VT = N->getValueType(i);
668 if (VT.isFloatingPoint() || VT.isVector())
669 return Sched::Latency;
672 if (!N->isMachineOpcode())
673 return Sched::RegPressure;
675 // Load are scheduled for latency even if there instruction itinerary
677 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
678 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
680 return Sched::Latency;
682 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
683 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
684 return Sched::Latency;
685 return Sched::RegPressure;
688 //===----------------------------------------------------------------------===//
690 //===----------------------------------------------------------------------===//
692 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
693 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
695 default: llvm_unreachable("Unknown condition code!");
696 case ISD::SETNE: return ARMCC::NE;
697 case ISD::SETEQ: return ARMCC::EQ;
698 case ISD::SETGT: return ARMCC::GT;
699 case ISD::SETGE: return ARMCC::GE;
700 case ISD::SETLT: return ARMCC::LT;
701 case ISD::SETLE: return ARMCC::LE;
702 case ISD::SETUGT: return ARMCC::HI;
703 case ISD::SETUGE: return ARMCC::HS;
704 case ISD::SETULT: return ARMCC::LO;
705 case ISD::SETULE: return ARMCC::LS;
709 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
710 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
711 ARMCC::CondCodes &CondCode2) {
712 CondCode2 = ARMCC::AL;
714 default: llvm_unreachable("Unknown FP condition!");
716 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
718 case ISD::SETOGT: CondCode = ARMCC::GT; break;
720 case ISD::SETOGE: CondCode = ARMCC::GE; break;
721 case ISD::SETOLT: CondCode = ARMCC::MI; break;
722 case ISD::SETOLE: CondCode = ARMCC::LS; break;
723 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
724 case ISD::SETO: CondCode = ARMCC::VC; break;
725 case ISD::SETUO: CondCode = ARMCC::VS; break;
726 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
727 case ISD::SETUGT: CondCode = ARMCC::HI; break;
728 case ISD::SETUGE: CondCode = ARMCC::PL; break;
730 case ISD::SETULT: CondCode = ARMCC::LT; break;
732 case ISD::SETULE: CondCode = ARMCC::LE; break;
734 case ISD::SETUNE: CondCode = ARMCC::NE; break;
738 //===----------------------------------------------------------------------===//
739 // Calling Convention Implementation
740 //===----------------------------------------------------------------------===//
742 #include "ARMGenCallingConv.inc"
744 // APCS f64 is in register pairs, possibly split to stack
745 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
746 CCValAssign::LocInfo &LocInfo,
747 CCState &State, bool CanFail) {
748 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
750 // Try to get the first register.
751 if (unsigned Reg = State.AllocateReg(RegList, 4))
752 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
754 // For the 2nd half of a v2f64, do not fail.
758 // Put the whole thing on the stack.
759 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
760 State.AllocateStack(8, 4),
765 // Try to get the second register.
766 if (unsigned Reg = State.AllocateReg(RegList, 4))
767 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
769 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
770 State.AllocateStack(4, 4),
775 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
776 CCValAssign::LocInfo &LocInfo,
777 ISD::ArgFlagsTy &ArgFlags,
779 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
781 if (LocVT == MVT::v2f64 &&
782 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
784 return true; // we handled it
787 // AAPCS f64 is in aligned register pairs
788 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
789 CCValAssign::LocInfo &LocInfo,
790 CCState &State, bool CanFail) {
791 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
792 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
794 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
796 // For the 2nd half of a v2f64, do not just fail.
800 // Put the whole thing on the stack.
801 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
802 State.AllocateStack(8, 8),
808 for (i = 0; i < 2; ++i)
809 if (HiRegList[i] == Reg)
812 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
813 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
818 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
819 CCValAssign::LocInfo &LocInfo,
820 ISD::ArgFlagsTy &ArgFlags,
822 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
824 if (LocVT == MVT::v2f64 &&
825 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
827 return true; // we handled it
830 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
831 CCValAssign::LocInfo &LocInfo, CCState &State) {
832 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
833 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
835 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
837 return false; // we didn't handle it
840 for (i = 0; i < 2; ++i)
841 if (HiRegList[i] == Reg)
844 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
845 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
850 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
851 CCValAssign::LocInfo &LocInfo,
852 ISD::ArgFlagsTy &ArgFlags,
854 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
856 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
858 return true; // we handled it
861 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
862 CCValAssign::LocInfo &LocInfo,
863 ISD::ArgFlagsTy &ArgFlags,
865 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
869 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
870 /// given CallingConvention value.
871 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
873 bool isVarArg) const {
876 llvm_unreachable("Unsupported calling convention");
878 case CallingConv::Fast:
879 // Use target triple & subtarget features to do actual dispatch.
880 if (Subtarget->isAAPCS_ABI()) {
881 if (Subtarget->hasVFP2() &&
882 FloatABIType == FloatABI::Hard && !isVarArg)
883 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
885 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
887 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
888 case CallingConv::ARM_AAPCS_VFP:
889 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
890 case CallingConv::ARM_AAPCS:
891 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
892 case CallingConv::ARM_APCS:
893 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
897 /// LowerCallResult - Lower the result values of a call into the
898 /// appropriate copies out of appropriate physical registers.
900 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
901 CallingConv::ID CallConv, bool isVarArg,
902 const SmallVectorImpl<ISD::InputArg> &Ins,
903 DebugLoc dl, SelectionDAG &DAG,
904 SmallVectorImpl<SDValue> &InVals) const {
906 // Assign locations to each value returned by this call.
907 SmallVector<CCValAssign, 16> RVLocs;
908 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
909 RVLocs, *DAG.getContext());
910 CCInfo.AnalyzeCallResult(Ins,
911 CCAssignFnForNode(CallConv, /* Return*/ true,
914 // Copy all of the result registers out of their specified physreg.
915 for (unsigned i = 0; i != RVLocs.size(); ++i) {
916 CCValAssign VA = RVLocs[i];
919 if (VA.needsCustom()) {
920 // Handle f64 or half of a v2f64.
921 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
923 Chain = Lo.getValue(1);
924 InFlag = Lo.getValue(2);
925 VA = RVLocs[++i]; // skip ahead to next loc
926 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
928 Chain = Hi.getValue(1);
929 InFlag = Hi.getValue(2);
930 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
932 if (VA.getLocVT() == MVT::v2f64) {
933 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
934 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
935 DAG.getConstant(0, MVT::i32));
937 VA = RVLocs[++i]; // skip ahead to next loc
938 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
939 Chain = Lo.getValue(1);
940 InFlag = Lo.getValue(2);
941 VA = RVLocs[++i]; // skip ahead to next loc
942 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
943 Chain = Hi.getValue(1);
944 InFlag = Hi.getValue(2);
945 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
946 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
947 DAG.getConstant(1, MVT::i32));
950 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
952 Chain = Val.getValue(1);
953 InFlag = Val.getValue(2);
956 switch (VA.getLocInfo()) {
957 default: llvm_unreachable("Unknown loc info!");
958 case CCValAssign::Full: break;
959 case CCValAssign::BCvt:
960 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
964 InVals.push_back(Val);
970 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
971 /// by "Src" to address "Dst" of size "Size". Alignment information is
972 /// specified by the specific parameter attribute. The copy will be passed as
973 /// a byval function parameter.
974 /// Sometimes what we are copying is the end of a larger object, the part that
975 /// does not fit in registers.
977 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
978 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
980 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
981 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
982 /*isVolatile=*/false, /*AlwaysInline=*/false,
986 /// LowerMemOpCallTo - Store the argument to the stack.
988 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
989 SDValue StackPtr, SDValue Arg,
990 DebugLoc dl, SelectionDAG &DAG,
991 const CCValAssign &VA,
992 ISD::ArgFlagsTy Flags) const {
993 unsigned LocMemOffset = VA.getLocMemOffset();
994 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
995 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
996 if (Flags.isByVal()) {
997 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
999 return DAG.getStore(Chain, dl, Arg, PtrOff,
1000 PseudoSourceValue::getStack(), LocMemOffset,
1004 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1005 SDValue Chain, SDValue &Arg,
1006 RegsToPassVector &RegsToPass,
1007 CCValAssign &VA, CCValAssign &NextVA,
1009 SmallVector<SDValue, 8> &MemOpChains,
1010 ISD::ArgFlagsTy Flags) const {
1012 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1013 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1014 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1016 if (NextVA.isRegLoc())
1017 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1019 assert(NextVA.isMemLoc());
1020 if (StackPtr.getNode() == 0)
1021 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1023 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1029 /// LowerCall - Lowering a call into a callseq_start <-
1030 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1033 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1034 CallingConv::ID CallConv, bool isVarArg,
1036 const SmallVectorImpl<ISD::OutputArg> &Outs,
1037 const SmallVectorImpl<SDValue> &OutVals,
1038 const SmallVectorImpl<ISD::InputArg> &Ins,
1039 DebugLoc dl, SelectionDAG &DAG,
1040 SmallVectorImpl<SDValue> &InVals) const {
1041 MachineFunction &MF = DAG.getMachineFunction();
1042 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1043 bool IsSibCall = false;
1044 // Temporarily disable tail calls so things don't break.
1045 if (!EnableARMTailCalls)
1048 // Check if it's really possible to do a tail call.
1049 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1050 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1051 Outs, OutVals, Ins, DAG);
1052 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1053 // detected sibcalls.
1060 // Analyze operands of the call, assigning locations to each operand.
1061 SmallVector<CCValAssign, 16> ArgLocs;
1062 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1064 CCInfo.AnalyzeCallOperands(Outs,
1065 CCAssignFnForNode(CallConv, /* Return*/ false,
1068 // Get a count of how many bytes are to be pushed on the stack.
1069 unsigned NumBytes = CCInfo.getNextStackOffset();
1071 // For tail calls, memory operands are available in our caller's stack.
1075 // Adjust the stack pointer for the new arguments...
1076 // These operations are automatically eliminated by the prolog/epilog pass
1078 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1080 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1082 RegsToPassVector RegsToPass;
1083 SmallVector<SDValue, 8> MemOpChains;
1085 // Walk the register/memloc assignments, inserting copies/loads. In the case
1086 // of tail call optimization, arguments are handled later.
1087 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1089 ++i, ++realArgIdx) {
1090 CCValAssign &VA = ArgLocs[i];
1091 SDValue Arg = OutVals[realArgIdx];
1092 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1094 // Promote the value if needed.
1095 switch (VA.getLocInfo()) {
1096 default: llvm_unreachable("Unknown loc info!");
1097 case CCValAssign::Full: break;
1098 case CCValAssign::SExt:
1099 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1101 case CCValAssign::ZExt:
1102 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1104 case CCValAssign::AExt:
1105 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1107 case CCValAssign::BCvt:
1108 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1112 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1113 if (VA.needsCustom()) {
1114 if (VA.getLocVT() == MVT::v2f64) {
1115 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1116 DAG.getConstant(0, MVT::i32));
1117 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1118 DAG.getConstant(1, MVT::i32));
1120 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1121 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1123 VA = ArgLocs[++i]; // skip ahead to next loc
1124 if (VA.isRegLoc()) {
1125 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1126 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1128 assert(VA.isMemLoc());
1130 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1131 dl, DAG, VA, Flags));
1134 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1135 StackPtr, MemOpChains, Flags);
1137 } else if (VA.isRegLoc()) {
1138 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1139 } else if (!IsSibCall) {
1140 assert(VA.isMemLoc());
1142 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1143 dl, DAG, VA, Flags));
1147 if (!MemOpChains.empty())
1148 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1149 &MemOpChains[0], MemOpChains.size());
1151 // Build a sequence of copy-to-reg nodes chained together with token chain
1152 // and flag operands which copy the outgoing args into the appropriate regs.
1154 // Tail call byval lowering might overwrite argument registers so in case of
1155 // tail call optimization the copies to registers are lowered later.
1157 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1158 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1159 RegsToPass[i].second, InFlag);
1160 InFlag = Chain.getValue(1);
1163 // For tail calls lower the arguments to the 'real' stack slot.
1165 // Force all the incoming stack arguments to be loaded from the stack
1166 // before any new outgoing arguments are stored to the stack, because the
1167 // outgoing stack slots may alias the incoming argument stack slots, and
1168 // the alias isn't otherwise explicit. This is slightly more conservative
1169 // than necessary, because it means that each store effectively depends
1170 // on every argument instead of just those arguments it would clobber.
1172 // Do not flag preceeding copytoreg stuff together with the following stuff.
1174 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1175 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1176 RegsToPass[i].second, InFlag);
1177 InFlag = Chain.getValue(1);
1182 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1183 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1184 // node so that legalize doesn't hack it.
1185 bool isDirect = false;
1186 bool isARMFunc = false;
1187 bool isLocalARMFunc = false;
1188 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1190 if (EnableARMLongCalls) {
1191 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1192 && "long-calls with non-static relocation model!");
1193 // Handle a global address or an external symbol. If it's not one of
1194 // those, the target's already in a register, so we don't need to do
1196 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1197 const GlobalValue *GV = G->getGlobal();
1198 // Create a constant pool entry for the callee address
1199 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1200 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1203 // Get the address of the callee into a register
1204 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1205 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1206 Callee = DAG.getLoad(getPointerTy(), dl,
1207 DAG.getEntryNode(), CPAddr,
1208 PseudoSourceValue::getConstantPool(), 0,
1210 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1211 const char *Sym = S->getSymbol();
1213 // Create a constant pool entry for the callee address
1214 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1215 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1216 Sym, ARMPCLabelIndex, 0);
1217 // Get the address of the callee into a register
1218 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1219 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1220 Callee = DAG.getLoad(getPointerTy(), dl,
1221 DAG.getEntryNode(), CPAddr,
1222 PseudoSourceValue::getConstantPool(), 0,
1225 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1226 const GlobalValue *GV = G->getGlobal();
1228 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1229 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1230 getTargetMachine().getRelocationModel() != Reloc::Static;
1231 isARMFunc = !Subtarget->isThumb() || isStub;
1232 // ARM call to a local ARM function is predicable.
1233 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1234 // tBX takes a register source operand.
1235 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1236 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1237 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1240 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1241 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1242 Callee = DAG.getLoad(getPointerTy(), dl,
1243 DAG.getEntryNode(), CPAddr,
1244 PseudoSourceValue::getConstantPool(), 0,
1246 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1247 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1248 getPointerTy(), Callee, PICLabel);
1250 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1251 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1253 bool isStub = Subtarget->isTargetDarwin() &&
1254 getTargetMachine().getRelocationModel() != Reloc::Static;
1255 isARMFunc = !Subtarget->isThumb() || isStub;
1256 // tBX takes a register source operand.
1257 const char *Sym = S->getSymbol();
1258 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1259 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1260 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1261 Sym, ARMPCLabelIndex, 4);
1262 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1263 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1264 Callee = DAG.getLoad(getPointerTy(), dl,
1265 DAG.getEntryNode(), CPAddr,
1266 PseudoSourceValue::getConstantPool(), 0,
1268 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1269 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1270 getPointerTy(), Callee, PICLabel);
1272 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1275 // FIXME: handle tail calls differently.
1277 if (Subtarget->isThumb()) {
1278 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1279 CallOpc = ARMISD::CALL_NOLINK;
1281 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1283 CallOpc = (isDirect || Subtarget->hasV5TOps())
1284 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1285 : ARMISD::CALL_NOLINK;
1287 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1288 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1289 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1290 InFlag = Chain.getValue(1);
1293 std::vector<SDValue> Ops;
1294 Ops.push_back(Chain);
1295 Ops.push_back(Callee);
1297 // Add argument registers to the end of the list so that they are known live
1299 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1300 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1301 RegsToPass[i].second.getValueType()));
1303 if (InFlag.getNode())
1304 Ops.push_back(InFlag);
1306 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1308 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1310 // Returns a chain and a flag for retval copy to use.
1311 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1312 InFlag = Chain.getValue(1);
1314 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1315 DAG.getIntPtrConstant(0, true), InFlag);
1317 InFlag = Chain.getValue(1);
1319 // Handle result values, copying them out of physregs into vregs that we
1321 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1325 /// MatchingStackOffset - Return true if the given stack call argument is
1326 /// already available in the same position (relatively) of the caller's
1327 /// incoming argument stack.
1329 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1330 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1331 const ARMInstrInfo *TII) {
1332 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1334 if (Arg.getOpcode() == ISD::CopyFromReg) {
1335 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1336 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1338 MachineInstr *Def = MRI->getVRegDef(VR);
1341 if (!Flags.isByVal()) {
1342 if (!TII->isLoadFromStackSlot(Def, FI))
1347 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1348 if (Flags.isByVal())
1349 // ByVal argument is passed in as a pointer but it's now being
1350 // dereferenced. e.g.
1351 // define @foo(%struct.X* %A) {
1352 // tail call @bar(%struct.X* byval %A)
1355 SDValue Ptr = Ld->getBasePtr();
1356 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1359 FI = FINode->getIndex();
1363 assert(FI != INT_MAX);
1364 if (!MFI->isFixedObjectIndex(FI))
1366 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1369 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1370 /// for tail call optimization. Targets which want to do tail call
1371 /// optimization should implement this function.
1373 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1374 CallingConv::ID CalleeCC,
1376 bool isCalleeStructRet,
1377 bool isCallerStructRet,
1378 const SmallVectorImpl<ISD::OutputArg> &Outs,
1379 const SmallVectorImpl<SDValue> &OutVals,
1380 const SmallVectorImpl<ISD::InputArg> &Ins,
1381 SelectionDAG& DAG) const {
1382 const Function *CallerF = DAG.getMachineFunction().getFunction();
1383 CallingConv::ID CallerCC = CallerF->getCallingConv();
1384 bool CCMatch = CallerCC == CalleeCC;
1386 // Look for obvious safe cases to perform tail call optimization that do not
1387 // require ABI changes. This is what gcc calls sibcall.
1389 // Do not sibcall optimize vararg calls unless the call site is not passing
1391 if (isVarArg && !Outs.empty())
1394 // Also avoid sibcall optimization if either caller or callee uses struct
1395 // return semantics.
1396 if (isCalleeStructRet || isCallerStructRet)
1399 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1400 // emitEpilogue is not ready for them.
1401 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1402 // LR. This means if we need to reload LR, it takes an extra instructions,
1403 // which outweighs the value of the tail call; but here we don't know yet
1404 // whether LR is going to be used. Probably the right approach is to
1405 // generate the tail call here and turn it back into CALL/RET in
1406 // emitEpilogue if LR is used.
1407 if (Subtarget->isThumb1Only())
1410 // 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,
1412 // or vice versa, is not easily fixed up in the linker unlike BL.
1413 // (We could do this by loading the address of the callee into a register;
1414 // that is an extra instruction over the direct call and burns a register
1415 // as well, so is not likely to be a win.)
1417 // It might be safe to remove this restriction on non-Darwin.
1419 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1420 // but we need to make sure there are enough registers; the only valid
1421 // registers are the 4 used for parameters. We don't currently do this
1423 if (isa<ExternalSymbolSDNode>(Callee))
1426 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1427 const GlobalValue *GV = G->getGlobal();
1428 if (GV->isDeclaration() || GV->isWeakForLinker())
1432 // If the calling conventions do not match, then we'd better make sure the
1433 // results are returned in the same way as what the caller expects.
1435 SmallVector<CCValAssign, 16> RVLocs1;
1436 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1437 RVLocs1, *DAG.getContext());
1438 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1440 SmallVector<CCValAssign, 16> RVLocs2;
1441 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1442 RVLocs2, *DAG.getContext());
1443 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1445 if (RVLocs1.size() != RVLocs2.size())
1447 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1448 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1450 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1452 if (RVLocs1[i].isRegLoc()) {
1453 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1456 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1462 // If the callee takes no arguments then go on to check the results of the
1464 if (!Outs.empty()) {
1465 // Check if stack adjustment is needed. For now, do not do this if any
1466 // argument is passed on the stack.
1467 SmallVector<CCValAssign, 16> ArgLocs;
1468 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1469 ArgLocs, *DAG.getContext());
1470 CCInfo.AnalyzeCallOperands(Outs,
1471 CCAssignFnForNode(CalleeCC, false, isVarArg));
1472 if (CCInfo.getNextStackOffset()) {
1473 MachineFunction &MF = DAG.getMachineFunction();
1475 // Check if the arguments are already laid out in the right way as
1476 // the caller's fixed stack objects.
1477 MachineFrameInfo *MFI = MF.getFrameInfo();
1478 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1479 const ARMInstrInfo *TII =
1480 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1481 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1483 ++i, ++realArgIdx) {
1484 CCValAssign &VA = ArgLocs[i];
1485 EVT RegVT = VA.getLocVT();
1486 SDValue Arg = OutVals[realArgIdx];
1487 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1488 if (VA.getLocInfo() == CCValAssign::Indirect)
1490 if (VA.needsCustom()) {
1491 // f64 and vector types are split into multiple registers or
1492 // register/stack-slot combinations. The types will not match
1493 // the registers; give up on memory f64 refs until we figure
1494 // out what to do about this.
1497 if (!ArgLocs[++i].isRegLoc())
1499 if (RegVT == MVT::v2f64) {
1500 if (!ArgLocs[++i].isRegLoc())
1502 if (!ArgLocs[++i].isRegLoc())
1505 } else if (!VA.isRegLoc()) {
1506 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1518 ARMTargetLowering::LowerReturn(SDValue Chain,
1519 CallingConv::ID CallConv, bool isVarArg,
1520 const SmallVectorImpl<ISD::OutputArg> &Outs,
1521 const SmallVectorImpl<SDValue> &OutVals,
1522 DebugLoc dl, SelectionDAG &DAG) const {
1524 // CCValAssign - represent the assignment of the return value to a location.
1525 SmallVector<CCValAssign, 16> RVLocs;
1527 // CCState - Info about the registers and stack slots.
1528 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1531 // Analyze outgoing return values.
1532 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1535 // If this is the first return lowered for this function, add
1536 // the regs to the liveout set for the function.
1537 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1538 for (unsigned i = 0; i != RVLocs.size(); ++i)
1539 if (RVLocs[i].isRegLoc())
1540 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1545 // Copy the result values into the output registers.
1546 for (unsigned i = 0, realRVLocIdx = 0;
1548 ++i, ++realRVLocIdx) {
1549 CCValAssign &VA = RVLocs[i];
1550 assert(VA.isRegLoc() && "Can only return in registers!");
1552 SDValue Arg = OutVals[realRVLocIdx];
1554 switch (VA.getLocInfo()) {
1555 default: llvm_unreachable("Unknown loc info!");
1556 case CCValAssign::Full: break;
1557 case CCValAssign::BCvt:
1558 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1562 if (VA.needsCustom()) {
1563 if (VA.getLocVT() == MVT::v2f64) {
1564 // Extract the first half and return it in two registers.
1565 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1566 DAG.getConstant(0, MVT::i32));
1567 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1568 DAG.getVTList(MVT::i32, MVT::i32), Half);
1570 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1571 Flag = Chain.getValue(1);
1572 VA = RVLocs[++i]; // skip ahead to next loc
1573 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1574 HalfGPRs.getValue(1), Flag);
1575 Flag = Chain.getValue(1);
1576 VA = RVLocs[++i]; // skip ahead to next loc
1578 // Extract the 2nd half and fall through to handle it as an f64 value.
1579 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1580 DAG.getConstant(1, MVT::i32));
1582 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1584 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1585 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1586 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1587 Flag = Chain.getValue(1);
1588 VA = RVLocs[++i]; // skip ahead to next loc
1589 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1592 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1594 // Guarantee that all emitted copies are
1595 // stuck together, avoiding something bad.
1596 Flag = Chain.getValue(1);
1601 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1603 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1608 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1609 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1610 // one of the above mentioned nodes. It has to be wrapped because otherwise
1611 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1612 // be used to form addressing mode. These wrapped nodes will be selected
1614 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1615 EVT PtrVT = Op.getValueType();
1616 // FIXME there is no actual debug info here
1617 DebugLoc dl = Op.getDebugLoc();
1618 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1620 if (CP->isMachineConstantPoolEntry())
1621 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1622 CP->getAlignment());
1624 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1625 CP->getAlignment());
1626 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1629 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1630 SelectionDAG &DAG) const {
1631 MachineFunction &MF = DAG.getMachineFunction();
1632 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1633 unsigned ARMPCLabelIndex = 0;
1634 DebugLoc DL = Op.getDebugLoc();
1635 EVT PtrVT = getPointerTy();
1636 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1637 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1639 if (RelocM == Reloc::Static) {
1640 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1642 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1643 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1644 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1645 ARMCP::CPBlockAddress,
1647 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1649 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1650 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1651 PseudoSourceValue::getConstantPool(), 0,
1653 if (RelocM == Reloc::Static)
1655 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1656 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1659 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1661 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1662 SelectionDAG &DAG) const {
1663 DebugLoc dl = GA->getDebugLoc();
1664 EVT PtrVT = getPointerTy();
1665 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1666 MachineFunction &MF = DAG.getMachineFunction();
1667 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1668 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1669 ARMConstantPoolValue *CPV =
1670 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1671 ARMCP::CPValue, PCAdj, "tlsgd", true);
1672 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1673 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1674 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1675 PseudoSourceValue::getConstantPool(), 0,
1677 SDValue Chain = Argument.getValue(1);
1679 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1680 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1682 // call __tls_get_addr.
1685 Entry.Node = Argument;
1686 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1687 Args.push_back(Entry);
1688 // FIXME: is there useful debug info available here?
1689 std::pair<SDValue, SDValue> CallResult =
1690 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1691 false, false, false, false,
1692 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1693 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1694 return CallResult.first;
1697 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1698 // "local exec" model.
1700 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1701 SelectionDAG &DAG) const {
1702 const GlobalValue *GV = GA->getGlobal();
1703 DebugLoc dl = GA->getDebugLoc();
1705 SDValue Chain = DAG.getEntryNode();
1706 EVT PtrVT = getPointerTy();
1707 // Get the Thread Pointer
1708 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1710 if (GV->isDeclaration()) {
1711 MachineFunction &MF = DAG.getMachineFunction();
1712 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1713 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1714 // Initial exec model.
1715 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1716 ARMConstantPoolValue *CPV =
1717 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1718 ARMCP::CPValue, PCAdj, "gottpoff", true);
1719 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1720 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1721 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1722 PseudoSourceValue::getConstantPool(), 0,
1724 Chain = Offset.getValue(1);
1726 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1727 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1729 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1730 PseudoSourceValue::getConstantPool(), 0,
1734 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1735 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1736 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1737 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1738 PseudoSourceValue::getConstantPool(), 0,
1742 // The address of the thread local variable is the add of the thread
1743 // pointer with the offset of the variable.
1744 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1748 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1749 // TODO: implement the "local dynamic" model
1750 assert(Subtarget->isTargetELF() &&
1751 "TLS not implemented for non-ELF targets");
1752 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1753 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1754 // otherwise use the "Local Exec" TLS Model
1755 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1756 return LowerToTLSGeneralDynamicModel(GA, DAG);
1758 return LowerToTLSExecModels(GA, DAG);
1761 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1762 SelectionDAG &DAG) const {
1763 EVT PtrVT = getPointerTy();
1764 DebugLoc dl = Op.getDebugLoc();
1765 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1766 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1767 if (RelocM == Reloc::PIC_) {
1768 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1769 ARMConstantPoolValue *CPV =
1770 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1771 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1772 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1773 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1775 PseudoSourceValue::getConstantPool(), 0,
1777 SDValue Chain = Result.getValue(1);
1778 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1779 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1781 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1782 PseudoSourceValue::getGOT(), 0,
1786 // If we have T2 ops, we can materialize the address directly via movt/movw
1787 // pair. This is always cheaper.
1788 if (Subtarget->useMovt()) {
1789 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1790 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1792 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1793 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1794 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1795 PseudoSourceValue::getConstantPool(), 0,
1801 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1802 SelectionDAG &DAG) const {
1803 MachineFunction &MF = DAG.getMachineFunction();
1804 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1805 unsigned ARMPCLabelIndex = 0;
1806 EVT PtrVT = getPointerTy();
1807 DebugLoc dl = Op.getDebugLoc();
1808 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1809 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1811 if (RelocM == Reloc::Static)
1812 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1814 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1815 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1816 ARMConstantPoolValue *CPV =
1817 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1818 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1820 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1822 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1823 PseudoSourceValue::getConstantPool(), 0,
1825 SDValue Chain = Result.getValue(1);
1827 if (RelocM == Reloc::PIC_) {
1828 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1829 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1832 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1833 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1834 PseudoSourceValue::getGOT(), 0,
1840 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1841 SelectionDAG &DAG) const {
1842 assert(Subtarget->isTargetELF() &&
1843 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1844 MachineFunction &MF = DAG.getMachineFunction();
1845 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1846 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1847 EVT PtrVT = getPointerTy();
1848 DebugLoc dl = Op.getDebugLoc();
1849 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1850 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1851 "_GLOBAL_OFFSET_TABLE_",
1852 ARMPCLabelIndex, PCAdj);
1853 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1854 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1855 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1856 PseudoSourceValue::getConstantPool(), 0,
1858 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1859 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1863 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1864 DebugLoc dl = Op.getDebugLoc();
1865 SDValue Val = DAG.getConstant(0, MVT::i32);
1866 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1867 Op.getOperand(1), Val);
1871 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1872 DebugLoc dl = Op.getDebugLoc();
1873 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1874 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1878 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1879 const ARMSubtarget *Subtarget) const {
1880 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1881 DebugLoc dl = Op.getDebugLoc();
1883 default: return SDValue(); // Don't custom lower most intrinsics.
1884 case Intrinsic::arm_thread_pointer: {
1885 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1886 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1888 case Intrinsic::eh_sjlj_lsda: {
1889 MachineFunction &MF = DAG.getMachineFunction();
1890 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1891 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1892 EVT PtrVT = getPointerTy();
1893 DebugLoc dl = Op.getDebugLoc();
1894 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1896 unsigned PCAdj = (RelocM != Reloc::PIC_)
1897 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1898 ARMConstantPoolValue *CPV =
1899 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1900 ARMCP::CPLSDA, PCAdj);
1901 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1902 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1904 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1905 PseudoSourceValue::getConstantPool(), 0,
1908 if (RelocM == Reloc::PIC_) {
1909 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1910 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1917 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1918 const ARMSubtarget *Subtarget) {
1919 DebugLoc dl = Op.getDebugLoc();
1920 SDValue Op5 = Op.getOperand(5);
1921 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1922 // v6 and v7 can both handle barriers directly, but need handled a bit
1923 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1925 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1926 if (Subtarget->hasV7Ops())
1927 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1928 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1929 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1930 DAG.getConstant(0, MVT::i32));
1931 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1935 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1936 MachineFunction &MF = DAG.getMachineFunction();
1937 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1939 // vastart just stores the address of the VarArgsFrameIndex slot into the
1940 // memory location argument.
1941 DebugLoc dl = Op.getDebugLoc();
1942 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1943 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1944 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1945 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1950 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1951 SelectionDAG &DAG) const {
1952 SDNode *Node = Op.getNode();
1953 DebugLoc dl = Node->getDebugLoc();
1954 EVT VT = Node->getValueType(0);
1955 SDValue Chain = Op.getOperand(0);
1956 SDValue Size = Op.getOperand(1);
1957 SDValue Align = Op.getOperand(2);
1959 // Chain the dynamic stack allocation so that it doesn't modify the stack
1960 // pointer when other instructions are using the stack.
1961 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1963 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1964 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1965 if (AlignVal > StackAlign)
1966 // Do this now since selection pass cannot introduce new target
1967 // independent node.
1968 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1970 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1971 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1972 // do even more horrible hack later.
1973 MachineFunction &MF = DAG.getMachineFunction();
1974 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1975 if (AFI->isThumb1OnlyFunction()) {
1977 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1979 uint32_t Val = C->getZExtValue();
1980 if (Val <= 508 && ((Val & 3) == 0))
1984 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1987 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1988 SDValue Ops1[] = { Chain, Size, Align };
1989 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1990 Chain = Res.getValue(1);
1991 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1992 DAG.getIntPtrConstant(0, true), SDValue());
1993 SDValue Ops2[] = { Res, Chain };
1994 return DAG.getMergeValues(Ops2, 2, dl);
1998 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1999 SDValue &Root, SelectionDAG &DAG,
2000 DebugLoc dl) const {
2001 MachineFunction &MF = DAG.getMachineFunction();
2002 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2004 TargetRegisterClass *RC;
2005 if (AFI->isThumb1OnlyFunction())
2006 RC = ARM::tGPRRegisterClass;
2008 RC = ARM::GPRRegisterClass;
2010 // Transform the arguments stored in physical registers into virtual ones.
2011 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2012 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2015 if (NextVA.isMemLoc()) {
2016 MachineFrameInfo *MFI = MF.getFrameInfo();
2017 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2019 // Create load node to retrieve arguments from the stack.
2020 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2021 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2022 PseudoSourceValue::getFixedStack(FI), 0,
2025 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2026 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2029 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2033 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2034 CallingConv::ID CallConv, bool isVarArg,
2035 const SmallVectorImpl<ISD::InputArg>
2037 DebugLoc dl, SelectionDAG &DAG,
2038 SmallVectorImpl<SDValue> &InVals)
2041 MachineFunction &MF = DAG.getMachineFunction();
2042 MachineFrameInfo *MFI = MF.getFrameInfo();
2044 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2046 // Assign locations to all of the incoming arguments.
2047 SmallVector<CCValAssign, 16> ArgLocs;
2048 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2050 CCInfo.AnalyzeFormalArguments(Ins,
2051 CCAssignFnForNode(CallConv, /* Return*/ false,
2054 SmallVector<SDValue, 16> ArgValues;
2056 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2057 CCValAssign &VA = ArgLocs[i];
2059 // Arguments stored in registers.
2060 if (VA.isRegLoc()) {
2061 EVT RegVT = VA.getLocVT();
2064 if (VA.needsCustom()) {
2065 // f64 and vector types are split up into multiple registers or
2066 // combinations of registers and stack slots.
2067 if (VA.getLocVT() == MVT::v2f64) {
2068 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2070 VA = ArgLocs[++i]; // skip ahead to next loc
2072 if (VA.isMemLoc()) {
2073 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2074 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2075 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2076 PseudoSourceValue::getFixedStack(FI), 0,
2079 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2082 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2083 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2084 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2085 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2086 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2088 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2091 TargetRegisterClass *RC;
2093 if (RegVT == MVT::f32)
2094 RC = ARM::SPRRegisterClass;
2095 else if (RegVT == MVT::f64)
2096 RC = ARM::DPRRegisterClass;
2097 else if (RegVT == MVT::v2f64)
2098 RC = ARM::QPRRegisterClass;
2099 else if (RegVT == MVT::i32)
2100 RC = (AFI->isThumb1OnlyFunction() ?
2101 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2103 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2105 // Transform the arguments in physical registers into virtual ones.
2106 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2107 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2110 // If this is an 8 or 16-bit value, it is really passed promoted
2111 // to 32 bits. Insert an assert[sz]ext to capture this, then
2112 // truncate to the right size.
2113 switch (VA.getLocInfo()) {
2114 default: llvm_unreachable("Unknown loc info!");
2115 case CCValAssign::Full: break;
2116 case CCValAssign::BCvt:
2117 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2119 case CCValAssign::SExt:
2120 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2121 DAG.getValueType(VA.getValVT()));
2122 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2124 case CCValAssign::ZExt:
2125 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2126 DAG.getValueType(VA.getValVT()));
2127 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2131 InVals.push_back(ArgValue);
2133 } else { // VA.isRegLoc()
2136 assert(VA.isMemLoc());
2137 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2139 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2140 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2142 // Create load nodes to retrieve arguments from the stack.
2143 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2144 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2145 PseudoSourceValue::getFixedStack(FI), 0,
2152 static const unsigned GPRArgRegs[] = {
2153 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2156 unsigned NumGPRs = CCInfo.getFirstUnallocated
2157 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2159 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2160 unsigned VARegSize = (4 - NumGPRs) * 4;
2161 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2162 unsigned ArgOffset = CCInfo.getNextStackOffset();
2163 if (VARegSaveSize) {
2164 // If this function is vararg, store any remaining integer argument regs
2165 // to their spots on the stack so that they may be loaded by deferencing
2166 // the result of va_next.
2167 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2168 AFI->setVarArgsFrameIndex(
2169 MFI->CreateFixedObject(VARegSaveSize,
2170 ArgOffset + VARegSaveSize - VARegSize,
2172 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2175 SmallVector<SDValue, 4> MemOps;
2176 for (; NumGPRs < 4; ++NumGPRs) {
2177 TargetRegisterClass *RC;
2178 if (AFI->isThumb1OnlyFunction())
2179 RC = ARM::tGPRRegisterClass;
2181 RC = ARM::GPRRegisterClass;
2183 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2184 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2186 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2187 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2188 0, false, false, 0);
2189 MemOps.push_back(Store);
2190 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2191 DAG.getConstant(4, getPointerTy()));
2193 if (!MemOps.empty())
2194 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2195 &MemOps[0], MemOps.size());
2197 // This will point to the next argument passed via stack.
2198 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2204 /// isFloatingPointZero - Return true if this is +0.0.
2205 static bool isFloatingPointZero(SDValue Op) {
2206 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2207 return CFP->getValueAPF().isPosZero();
2208 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2209 // Maybe this has already been legalized into the constant pool?
2210 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2211 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2212 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2213 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2214 return CFP->getValueAPF().isPosZero();
2220 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2221 /// the given operands.
2223 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2224 SDValue &ARMCC, SelectionDAG &DAG,
2225 DebugLoc dl) const {
2226 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2227 unsigned C = RHSC->getZExtValue();
2228 if (!isLegalICmpImmediate(C)) {
2229 // Constant does not fit, try adjusting it by one?
2234 if (isLegalICmpImmediate(C-1)) {
2235 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2236 RHS = DAG.getConstant(C-1, MVT::i32);
2241 if (C > 0 && isLegalICmpImmediate(C-1)) {
2242 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2243 RHS = DAG.getConstant(C-1, MVT::i32);
2248 if (isLegalICmpImmediate(C+1)) {
2249 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2250 RHS = DAG.getConstant(C+1, MVT::i32);
2255 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2256 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2257 RHS = DAG.getConstant(C+1, MVT::i32);
2264 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2265 ARMISD::NodeType CompareType;
2268 CompareType = ARMISD::CMP;
2273 CompareType = ARMISD::CMPZ;
2276 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2277 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2280 static bool canBitcastToInt(SDNode *Op) {
2281 return Op->hasOneUse() &&
2282 ISD::isNormalLoad(Op) &&
2283 Op->getValueType(0) == MVT::f32;
2286 static SDValue bitcastToInt(SDValue Op, SelectionDAG &DAG) {
2287 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2288 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2289 Ld->getChain(), Ld->getBasePtr(),
2290 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2291 Ld->isVolatile(), Ld->isNonTemporal(),
2292 Ld->getAlignment());
2294 llvm_unreachable("Unknown VFP cmp argument!");
2297 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2299 ARMTargetLowering::getVFPCmp(SDValue &LHS, SDValue &RHS, ISD::CondCode CC,
2300 SDValue &ARMCC, SelectionDAG &DAG,
2301 DebugLoc dl) const {
2302 if (UnsafeFPMath && FiniteOnlyFPMath() &&
2303 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2304 CC == ISD::SETNE || CC == ISD::SETUNE) &&
2305 canBitcastToInt(LHS.getNode()) && canBitcastToInt(RHS.getNode())) {
2306 // If unsafe fp math optimization is enabled and there are no othter uses of
2307 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2308 // to an integer comparison.
2309 if (CC == ISD::SETOEQ)
2311 else if (CC == ISD::SETUNE)
2313 LHS = bitcastToInt(LHS, DAG);
2314 RHS = bitcastToInt(RHS, DAG);
2315 return getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2319 if (!isFloatingPointZero(RHS))
2320 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2322 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2323 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2326 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2327 EVT VT = Op.getValueType();
2328 SDValue LHS = Op.getOperand(0);
2329 SDValue RHS = Op.getOperand(1);
2330 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2331 SDValue TrueVal = Op.getOperand(2);
2332 SDValue FalseVal = Op.getOperand(3);
2333 DebugLoc dl = Op.getDebugLoc();
2335 if (LHS.getValueType() == MVT::i32) {
2337 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2338 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2339 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2342 ARMCC::CondCodes CondCode, CondCode2;
2343 FPCCToARMCC(CC, CondCode, CondCode2);
2345 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2346 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2347 SDValue Cmp = getVFPCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2348 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2350 if (CondCode2 != ARMCC::AL) {
2351 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2352 // FIXME: Needs another CMP because flag can have but one use.
2353 SDValue Cmp2 = getVFPCmp(LHS, RHS, CC, ARMCC2, DAG, dl);
2354 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2355 Result, TrueVal, ARMCC2, CCR, Cmp2);
2360 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2361 SDValue Chain = Op.getOperand(0);
2362 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2363 SDValue LHS = Op.getOperand(2);
2364 SDValue RHS = Op.getOperand(3);
2365 SDValue Dest = Op.getOperand(4);
2366 DebugLoc dl = Op.getDebugLoc();
2368 if (LHS.getValueType() == MVT::i32) {
2370 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2371 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2372 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2373 Chain, Dest, ARMCC, CCR,Cmp);
2376 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2377 ARMCC::CondCodes CondCode, CondCode2;
2378 FPCCToARMCC(CC, CondCode, CondCode2);
2380 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2381 SDValue Cmp = getVFPCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2382 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2383 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2384 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2385 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2386 if (CondCode2 != ARMCC::AL) {
2387 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2388 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2389 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2394 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2395 SDValue Chain = Op.getOperand(0);
2396 SDValue Table = Op.getOperand(1);
2397 SDValue Index = Op.getOperand(2);
2398 DebugLoc dl = Op.getDebugLoc();
2400 EVT PTy = getPointerTy();
2401 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2402 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2403 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2404 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2405 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2406 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2407 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2408 if (Subtarget->isThumb2()) {
2409 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2410 // which does another jump to the destination. This also makes it easier
2411 // to translate it to TBB / TBH later.
2412 // FIXME: This might not work if the function is extremely large.
2413 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2414 Addr, Op.getOperand(2), JTI, UId);
2416 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2417 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2418 PseudoSourceValue::getJumpTable(), 0,
2420 Chain = Addr.getValue(1);
2421 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2422 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2424 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2425 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2426 Chain = Addr.getValue(1);
2427 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2431 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2432 DebugLoc dl = Op.getDebugLoc();
2435 switch (Op.getOpcode()) {
2437 assert(0 && "Invalid opcode!");
2438 case ISD::FP_TO_SINT:
2439 Opc = ARMISD::FTOSI;
2441 case ISD::FP_TO_UINT:
2442 Opc = ARMISD::FTOUI;
2445 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2446 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2449 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2450 EVT VT = Op.getValueType();
2451 DebugLoc dl = Op.getDebugLoc();
2454 switch (Op.getOpcode()) {
2456 assert(0 && "Invalid opcode!");
2457 case ISD::SINT_TO_FP:
2458 Opc = ARMISD::SITOF;
2460 case ISD::UINT_TO_FP:
2461 Opc = ARMISD::UITOF;
2465 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2466 return DAG.getNode(Opc, dl, VT, Op);
2469 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2470 // Implement fcopysign with a fabs and a conditional fneg.
2471 SDValue Tmp0 = Op.getOperand(0);
2472 SDValue Tmp1 = Op.getOperand(1);
2473 DebugLoc dl = Op.getDebugLoc();
2474 EVT VT = Op.getValueType();
2475 EVT SrcVT = Tmp1.getValueType();
2476 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2477 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2478 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2479 SDValue Cmp = getVFPCmp(Tmp1, FP0,
2480 ISD::SETLT, ARMCC, DAG, dl);
2481 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2482 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2485 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2486 MachineFunction &MF = DAG.getMachineFunction();
2487 MachineFrameInfo *MFI = MF.getFrameInfo();
2488 MFI->setReturnAddressIsTaken(true);
2490 EVT VT = Op.getValueType();
2491 DebugLoc dl = Op.getDebugLoc();
2492 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2494 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2495 SDValue Offset = DAG.getConstant(4, MVT::i32);
2496 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2497 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2498 NULL, 0, false, false, 0);
2501 // Return LR, which contains the return address. Mark it an implicit live-in.
2502 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2503 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2506 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2507 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2508 MFI->setFrameAddressIsTaken(true);
2510 EVT VT = Op.getValueType();
2511 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2512 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2513 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2514 ? ARM::R7 : ARM::R11;
2515 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2517 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2522 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2523 /// expand a bit convert where either the source or destination type is i64 to
2524 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2525 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2526 /// vectors), since the legalizer won't know what to do with that.
2527 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2528 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2529 DebugLoc dl = N->getDebugLoc();
2530 SDValue Op = N->getOperand(0);
2532 // This function is only supposed to be called for i64 types, either as the
2533 // source or destination of the bit convert.
2534 EVT SrcVT = Op.getValueType();
2535 EVT DstVT = N->getValueType(0);
2536 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2537 "ExpandBIT_CONVERT called for non-i64 type");
2539 // Turn i64->f64 into VMOVDRR.
2540 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2541 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2542 DAG.getConstant(0, MVT::i32));
2543 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2544 DAG.getConstant(1, MVT::i32));
2545 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2546 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2549 // Turn f64->i64 into VMOVRRD.
2550 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2551 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2552 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2553 // Merge the pieces into a single i64 value.
2554 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2560 /// getZeroVector - Returns a vector of specified type with all zero elements.
2562 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2563 assert(VT.isVector() && "Expected a vector type");
2565 // Zero vectors are used to represent vector negation and in those cases
2566 // will be implemented with the NEON VNEG instruction. However, VNEG does
2567 // not support i64 elements, so sometimes the zero vectors will need to be
2568 // explicitly constructed. For those cases, and potentially other uses in
2569 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2570 // to their dest type. This ensures they get CSE'd.
2572 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2573 SmallVector<SDValue, 8> Ops;
2576 if (VT.getSizeInBits() == 64) {
2577 Ops.assign(8, Cst); TVT = MVT::v8i8;
2579 Ops.assign(16, Cst); TVT = MVT::v16i8;
2581 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2583 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2586 /// getOnesVector - Returns a vector of specified type with all bits set.
2588 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2589 assert(VT.isVector() && "Expected a vector type");
2591 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2592 // dest type. This ensures they get CSE'd.
2594 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2595 SmallVector<SDValue, 8> Ops;
2598 if (VT.getSizeInBits() == 64) {
2599 Ops.assign(8, Cst); TVT = MVT::v8i8;
2601 Ops.assign(16, Cst); TVT = MVT::v16i8;
2603 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2605 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2608 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2609 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2610 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2611 SelectionDAG &DAG) const {
2612 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2613 EVT VT = Op.getValueType();
2614 unsigned VTBits = VT.getSizeInBits();
2615 DebugLoc dl = Op.getDebugLoc();
2616 SDValue ShOpLo = Op.getOperand(0);
2617 SDValue ShOpHi = Op.getOperand(1);
2618 SDValue ShAmt = Op.getOperand(2);
2620 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2622 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2624 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2625 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2626 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2627 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2628 DAG.getConstant(VTBits, MVT::i32));
2629 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2630 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2631 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2633 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2634 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2636 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2637 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2640 SDValue Ops[2] = { Lo, Hi };
2641 return DAG.getMergeValues(Ops, 2, dl);
2644 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2645 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2646 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2647 SelectionDAG &DAG) const {
2648 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2649 EVT VT = Op.getValueType();
2650 unsigned VTBits = VT.getSizeInBits();
2651 DebugLoc dl = Op.getDebugLoc();
2652 SDValue ShOpLo = Op.getOperand(0);
2653 SDValue ShOpHi = Op.getOperand(1);
2654 SDValue ShAmt = Op.getOperand(2);
2657 assert(Op.getOpcode() == ISD::SHL_PARTS);
2658 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2659 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2660 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2661 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2662 DAG.getConstant(VTBits, MVT::i32));
2663 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2664 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2666 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2667 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2668 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2670 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2671 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2674 SDValue Ops[2] = { Lo, Hi };
2675 return DAG.getMergeValues(Ops, 2, dl);
2678 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2679 const ARMSubtarget *ST) {
2680 EVT VT = N->getValueType(0);
2681 DebugLoc dl = N->getDebugLoc();
2683 if (!ST->hasV6T2Ops())
2686 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2687 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2690 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2691 const ARMSubtarget *ST) {
2692 EVT VT = N->getValueType(0);
2693 DebugLoc dl = N->getDebugLoc();
2695 // Lower vector shifts on NEON to use VSHL.
2696 if (VT.isVector()) {
2697 assert(ST->hasNEON() && "unexpected vector shift");
2699 // Left shifts translate directly to the vshiftu intrinsic.
2700 if (N->getOpcode() == ISD::SHL)
2701 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2702 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2703 N->getOperand(0), N->getOperand(1));
2705 assert((N->getOpcode() == ISD::SRA ||
2706 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2708 // NEON uses the same intrinsics for both left and right shifts. For
2709 // right shifts, the shift amounts are negative, so negate the vector of
2711 EVT ShiftVT = N->getOperand(1).getValueType();
2712 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2713 getZeroVector(ShiftVT, DAG, dl),
2715 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2716 Intrinsic::arm_neon_vshifts :
2717 Intrinsic::arm_neon_vshiftu);
2718 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2719 DAG.getConstant(vshiftInt, MVT::i32),
2720 N->getOperand(0), NegatedCount);
2723 // We can get here for a node like i32 = ISD::SHL i32, i64
2727 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2728 "Unknown shift to lower!");
2730 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2731 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2732 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2735 // If we are in thumb mode, we don't have RRX.
2736 if (ST->isThumb1Only()) return SDValue();
2738 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2739 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2740 DAG.getConstant(0, MVT::i32));
2741 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2742 DAG.getConstant(1, MVT::i32));
2744 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2745 // captures the result into a carry flag.
2746 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2747 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2749 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2750 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2752 // Merge the pieces into a single i64 value.
2753 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2756 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2757 SDValue TmpOp0, TmpOp1;
2758 bool Invert = false;
2762 SDValue Op0 = Op.getOperand(0);
2763 SDValue Op1 = Op.getOperand(1);
2764 SDValue CC = Op.getOperand(2);
2765 EVT VT = Op.getValueType();
2766 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2767 DebugLoc dl = Op.getDebugLoc();
2769 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2770 switch (SetCCOpcode) {
2771 default: llvm_unreachable("Illegal FP comparison"); break;
2773 case ISD::SETNE: Invert = true; // Fallthrough
2775 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2777 case ISD::SETLT: Swap = true; // Fallthrough
2779 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2781 case ISD::SETLE: Swap = true; // Fallthrough
2783 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2784 case ISD::SETUGE: Swap = true; // Fallthrough
2785 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2786 case ISD::SETUGT: Swap = true; // Fallthrough
2787 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2788 case ISD::SETUEQ: Invert = true; // Fallthrough
2790 // Expand this to (OLT | OGT).
2794 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2795 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2797 case ISD::SETUO: Invert = true; // Fallthrough
2799 // Expand this to (OLT | OGE).
2803 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2804 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2808 // Integer comparisons.
2809 switch (SetCCOpcode) {
2810 default: llvm_unreachable("Illegal integer comparison"); break;
2811 case ISD::SETNE: Invert = true;
2812 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2813 case ISD::SETLT: Swap = true;
2814 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2815 case ISD::SETLE: Swap = true;
2816 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2817 case ISD::SETULT: Swap = true;
2818 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2819 case ISD::SETULE: Swap = true;
2820 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2823 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2824 if (Opc == ARMISD::VCEQ) {
2827 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2829 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2832 // Ignore bitconvert.
2833 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2834 AndOp = AndOp.getOperand(0);
2836 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2838 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2839 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2846 std::swap(Op0, Op1);
2848 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2851 Result = DAG.getNOT(dl, Result, VT);
2856 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2857 /// valid vector constant for a NEON instruction with a "modified immediate"
2858 /// operand (e.g., VMOV). If so, return either the constant being
2859 /// splatted or the encoded value, depending on the DoEncode parameter. The
2860 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2861 /// bits7-0=Immediate.
2862 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2863 unsigned SplatBitSize, SelectionDAG &DAG,
2864 bool isVMOV, bool DoEncode) {
2865 unsigned Op, Cmode, Imm;
2868 // SplatBitSize is set to the smallest size that splats the vector, so a
2869 // zero vector will always have SplatBitSize == 8. However, NEON modified
2870 // immediate instructions others than VMOV do not support the 8-bit encoding
2871 // of a zero vector, and the default encoding of zero is supposed to be the
2877 switch (SplatBitSize) {
2879 // Any 1-byte value is OK. Op=0, Cmode=1110.
2880 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2887 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2889 if ((SplatBits & ~0xff) == 0) {
2890 // Value = 0x00nn: Op=x, Cmode=100x.
2895 if ((SplatBits & ~0xff00) == 0) {
2896 // Value = 0xnn00: Op=x, Cmode=101x.
2898 Imm = SplatBits >> 8;
2904 // NEON's 32-bit VMOV supports splat values where:
2905 // * only one byte is nonzero, or
2906 // * the least significant byte is 0xff and the second byte is nonzero, or
2907 // * the least significant 2 bytes are 0xff and the third is nonzero.
2909 if ((SplatBits & ~0xff) == 0) {
2910 // Value = 0x000000nn: Op=x, Cmode=000x.
2915 if ((SplatBits & ~0xff00) == 0) {
2916 // Value = 0x0000nn00: Op=x, Cmode=001x.
2918 Imm = SplatBits >> 8;
2921 if ((SplatBits & ~0xff0000) == 0) {
2922 // Value = 0x00nn0000: Op=x, Cmode=010x.
2924 Imm = SplatBits >> 16;
2927 if ((SplatBits & ~0xff000000) == 0) {
2928 // Value = 0xnn000000: Op=x, Cmode=011x.
2930 Imm = SplatBits >> 24;
2934 if ((SplatBits & ~0xffff) == 0 &&
2935 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2936 // Value = 0x0000nnff: Op=x, Cmode=1100.
2938 Imm = SplatBits >> 8;
2943 if ((SplatBits & ~0xffffff) == 0 &&
2944 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2945 // Value = 0x00nnffff: Op=x, Cmode=1101.
2947 Imm = SplatBits >> 16;
2948 SplatBits |= 0xffff;
2952 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2953 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2954 // VMOV.I32. A (very) minor optimization would be to replicate the value
2955 // and fall through here to test for a valid 64-bit splat. But, then the
2956 // caller would also need to check and handle the change in size.
2960 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2963 uint64_t BitMask = 0xff;
2965 unsigned ImmMask = 1;
2967 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2968 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2971 } else if ((SplatBits & BitMask) != 0) {
2977 // Op=1, Cmode=1110.
2986 llvm_unreachable("unexpected size for isNEONModifiedImm");
2991 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2992 return DAG.getTargetConstant(SplatBits, VT);
2996 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2997 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2998 /// size, return the encoded value for that immediate. The ByteSize field
2999 /// indicates the number of bytes of each element [1248].
3000 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
3001 SelectionDAG &DAG) {
3002 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
3003 APInt SplatBits, SplatUndef;
3004 unsigned SplatBitSize;
3006 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3007 HasAnyUndefs, ByteSize * 8))
3010 if (SplatBitSize > ByteSize * 8)
3013 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
3014 SplatBitSize, DAG, isVMOV, true);
3017 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3018 bool &ReverseVEXT, unsigned &Imm) {
3019 unsigned NumElts = VT.getVectorNumElements();
3020 ReverseVEXT = false;
3023 // If this is a VEXT shuffle, the immediate value is the index of the first
3024 // element. The other shuffle indices must be the successive elements after
3026 unsigned ExpectedElt = Imm;
3027 for (unsigned i = 1; i < NumElts; ++i) {
3028 // Increment the expected index. If it wraps around, it may still be
3029 // a VEXT but the source vectors must be swapped.
3031 if (ExpectedElt == NumElts * 2) {
3036 if (ExpectedElt != static_cast<unsigned>(M[i]))
3040 // Adjust the index value if the source operands will be swapped.
3047 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3048 /// instruction with the specified blocksize. (The order of the elements
3049 /// within each block of the vector is reversed.)
3050 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3051 unsigned BlockSize) {
3052 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3053 "Only possible block sizes for VREV are: 16, 32, 64");
3055 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3059 unsigned NumElts = VT.getVectorNumElements();
3060 unsigned BlockElts = M[0] + 1;
3062 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3065 for (unsigned i = 0; i < NumElts; ++i) {
3066 if ((unsigned) M[i] !=
3067 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3074 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3075 unsigned &WhichResult) {
3076 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3080 unsigned NumElts = VT.getVectorNumElements();
3081 WhichResult = (M[0] == 0 ? 0 : 1);
3082 for (unsigned i = 0; i < NumElts; i += 2) {
3083 if ((unsigned) M[i] != i + WhichResult ||
3084 (unsigned) M[i+1] != i + NumElts + WhichResult)
3090 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3091 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3092 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3093 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3094 unsigned &WhichResult) {
3095 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3099 unsigned NumElts = VT.getVectorNumElements();
3100 WhichResult = (M[0] == 0 ? 0 : 1);
3101 for (unsigned i = 0; i < NumElts; i += 2) {
3102 if ((unsigned) M[i] != i + WhichResult ||
3103 (unsigned) M[i+1] != i + WhichResult)
3109 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3110 unsigned &WhichResult) {
3111 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3115 unsigned NumElts = VT.getVectorNumElements();
3116 WhichResult = (M[0] == 0 ? 0 : 1);
3117 for (unsigned i = 0; i != NumElts; ++i) {
3118 if ((unsigned) M[i] != 2 * i + WhichResult)
3122 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3123 if (VT.is64BitVector() && EltSz == 32)
3129 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3130 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3131 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3132 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3133 unsigned &WhichResult) {
3134 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3138 unsigned Half = VT.getVectorNumElements() / 2;
3139 WhichResult = (M[0] == 0 ? 0 : 1);
3140 for (unsigned j = 0; j != 2; ++j) {
3141 unsigned Idx = WhichResult;
3142 for (unsigned i = 0; i != Half; ++i) {
3143 if ((unsigned) M[i + j * Half] != Idx)
3149 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3150 if (VT.is64BitVector() && EltSz == 32)
3156 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3157 unsigned &WhichResult) {
3158 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3162 unsigned NumElts = VT.getVectorNumElements();
3163 WhichResult = (M[0] == 0 ? 0 : 1);
3164 unsigned Idx = WhichResult * NumElts / 2;
3165 for (unsigned i = 0; i != NumElts; i += 2) {
3166 if ((unsigned) M[i] != Idx ||
3167 (unsigned) M[i+1] != Idx + NumElts)
3172 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3173 if (VT.is64BitVector() && EltSz == 32)
3179 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3180 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3181 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3182 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3183 unsigned &WhichResult) {
3184 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3188 unsigned NumElts = VT.getVectorNumElements();
3189 WhichResult = (M[0] == 0 ? 0 : 1);
3190 unsigned Idx = WhichResult * NumElts / 2;
3191 for (unsigned i = 0; i != NumElts; i += 2) {
3192 if ((unsigned) M[i] != Idx ||
3193 (unsigned) M[i+1] != Idx)
3198 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3199 if (VT.is64BitVector() && EltSz == 32)
3206 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3207 // Canonicalize all-zeros and all-ones vectors.
3208 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3209 if (ConstVal->isNullValue())
3210 return getZeroVector(VT, DAG, dl);
3211 if (ConstVal->isAllOnesValue())
3212 return getOnesVector(VT, DAG, dl);
3215 if (VT.is64BitVector()) {
3216 switch (Val.getValueType().getSizeInBits()) {
3217 case 8: CanonicalVT = MVT::v8i8; break;
3218 case 16: CanonicalVT = MVT::v4i16; break;
3219 case 32: CanonicalVT = MVT::v2i32; break;
3220 case 64: CanonicalVT = MVT::v1i64; break;
3221 default: llvm_unreachable("unexpected splat element type"); break;
3224 assert(VT.is128BitVector() && "unknown splat vector size");
3225 switch (Val.getValueType().getSizeInBits()) {
3226 case 8: CanonicalVT = MVT::v16i8; break;
3227 case 16: CanonicalVT = MVT::v8i16; break;
3228 case 32: CanonicalVT = MVT::v4i32; break;
3229 case 64: CanonicalVT = MVT::v2i64; break;
3230 default: llvm_unreachable("unexpected splat element type"); break;
3234 // Build a canonical splat for this value.
3235 SmallVector<SDValue, 8> Ops;
3236 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3237 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3239 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3242 // If this is a case we can't handle, return null and let the default
3243 // expansion code take care of it.
3244 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3245 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3246 DebugLoc dl = Op.getDebugLoc();
3247 EVT VT = Op.getValueType();
3249 APInt SplatBits, SplatUndef;
3250 unsigned SplatBitSize;
3252 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3253 if (SplatBitSize <= 64) {
3254 // Check if an immediate VMOV works.
3255 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3256 SplatUndef.getZExtValue(),
3257 SplatBitSize, DAG, true, false);
3259 return BuildSplat(Val, VT, DAG, dl);
3263 // Scan through the operands to see if only one value is used.
3264 unsigned NumElts = VT.getVectorNumElements();
3265 bool isOnlyLowElement = true;
3266 bool usesOnlyOneValue = true;
3267 bool isConstant = true;
3269 for (unsigned i = 0; i < NumElts; ++i) {
3270 SDValue V = Op.getOperand(i);
3271 if (V.getOpcode() == ISD::UNDEF)
3274 isOnlyLowElement = false;
3275 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3278 if (!Value.getNode())
3280 else if (V != Value)
3281 usesOnlyOneValue = false;
3284 if (!Value.getNode())
3285 return DAG.getUNDEF(VT);
3287 if (isOnlyLowElement)
3288 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3290 // If all elements are constants, fall back to the default expansion, which
3291 // will generate a load from the constant pool.
3295 // Use VDUP for non-constant splats.
3296 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3297 if (usesOnlyOneValue && EltSize <= 32)
3298 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3300 // Vectors with 32- or 64-bit elements can be built by directly assigning
3301 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3302 // will be legalized.
3303 if (EltSize >= 32) {
3304 // Do the expansion with floating-point types, since that is what the VFP
3305 // registers are defined to use, and since i64 is not legal.
3306 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3307 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3308 SmallVector<SDValue, 8> Ops;
3309 for (unsigned i = 0; i < NumElts; ++i)
3310 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3311 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3312 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3318 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3319 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3320 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3321 /// are assumed to be legal.
3323 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3325 if (VT.getVectorNumElements() == 4 &&
3326 (VT.is128BitVector() || VT.is64BitVector())) {
3327 unsigned PFIndexes[4];
3328 for (unsigned i = 0; i != 4; ++i) {
3332 PFIndexes[i] = M[i];
3335 // Compute the index in the perfect shuffle table.
3336 unsigned PFTableIndex =
3337 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3338 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3339 unsigned Cost = (PFEntry >> 30);
3346 unsigned Imm, WhichResult;
3348 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3349 return (EltSize >= 32 ||
3350 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3351 isVREVMask(M, VT, 64) ||
3352 isVREVMask(M, VT, 32) ||
3353 isVREVMask(M, VT, 16) ||
3354 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3355 isVTRNMask(M, VT, WhichResult) ||
3356 isVUZPMask(M, VT, WhichResult) ||
3357 isVZIPMask(M, VT, WhichResult) ||
3358 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3359 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3360 isVZIP_v_undef_Mask(M, VT, WhichResult));
3363 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3364 /// the specified operations to build the shuffle.
3365 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3366 SDValue RHS, SelectionDAG &DAG,
3368 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3369 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3370 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3373 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3382 OP_VUZPL, // VUZP, left result
3383 OP_VUZPR, // VUZP, right result
3384 OP_VZIPL, // VZIP, left result
3385 OP_VZIPR, // VZIP, right result
3386 OP_VTRNL, // VTRN, left result
3387 OP_VTRNR // VTRN, right result
3390 if (OpNum == OP_COPY) {
3391 if (LHSID == (1*9+2)*9+3) return LHS;
3392 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3396 SDValue OpLHS, OpRHS;
3397 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3398 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3399 EVT VT = OpLHS.getValueType();
3402 default: llvm_unreachable("Unknown shuffle opcode!");
3404 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3409 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3410 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3414 return DAG.getNode(ARMISD::VEXT, dl, VT,
3416 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3419 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3420 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3423 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3424 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3427 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3428 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3432 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3433 SDValue V1 = Op.getOperand(0);
3434 SDValue V2 = Op.getOperand(1);
3435 DebugLoc dl = Op.getDebugLoc();
3436 EVT VT = Op.getValueType();
3437 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3438 SmallVector<int, 8> ShuffleMask;
3440 // Convert shuffles that are directly supported on NEON to target-specific
3441 // DAG nodes, instead of keeping them as shuffles and matching them again
3442 // during code selection. This is more efficient and avoids the possibility
3443 // of inconsistencies between legalization and selection.
3444 // FIXME: floating-point vectors should be canonicalized to integer vectors
3445 // of the same time so that they get CSEd properly.
3446 SVN->getMask(ShuffleMask);
3448 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3449 if (EltSize <= 32) {
3450 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3451 int Lane = SVN->getSplatIndex();
3452 // If this is undef splat, generate it via "just" vdup, if possible.
3453 if (Lane == -1) Lane = 0;
3455 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3456 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3458 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3459 DAG.getConstant(Lane, MVT::i32));
3464 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3467 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3468 DAG.getConstant(Imm, MVT::i32));
3471 if (isVREVMask(ShuffleMask, VT, 64))
3472 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3473 if (isVREVMask(ShuffleMask, VT, 32))
3474 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3475 if (isVREVMask(ShuffleMask, VT, 16))
3476 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3478 // Check for Neon shuffles that modify both input vectors in place.
3479 // If both results are used, i.e., if there are two shuffles with the same
3480 // source operands and with masks corresponding to both results of one of
3481 // these operations, DAG memoization will ensure that a single node is
3482 // used for both shuffles.
3483 unsigned WhichResult;
3484 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3485 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3486 V1, V2).getValue(WhichResult);
3487 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3488 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3489 V1, V2).getValue(WhichResult);
3490 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3491 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3492 V1, V2).getValue(WhichResult);
3494 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3495 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3496 V1, V1).getValue(WhichResult);
3497 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3498 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3499 V1, V1).getValue(WhichResult);
3500 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3501 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3502 V1, V1).getValue(WhichResult);
3505 // If the shuffle is not directly supported and it has 4 elements, use
3506 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3507 unsigned NumElts = VT.getVectorNumElements();
3509 unsigned PFIndexes[4];
3510 for (unsigned i = 0; i != 4; ++i) {
3511 if (ShuffleMask[i] < 0)
3514 PFIndexes[i] = ShuffleMask[i];
3517 // Compute the index in the perfect shuffle table.
3518 unsigned PFTableIndex =
3519 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3520 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3521 unsigned Cost = (PFEntry >> 30);
3524 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3527 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3528 if (EltSize >= 32) {
3529 // Do the expansion with floating-point types, since that is what the VFP
3530 // registers are defined to use, and since i64 is not legal.
3531 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3532 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3533 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3534 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3535 SmallVector<SDValue, 8> Ops;
3536 for (unsigned i = 0; i < NumElts; ++i) {
3537 if (ShuffleMask[i] < 0)
3538 Ops.push_back(DAG.getUNDEF(EltVT));
3540 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3541 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3542 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3545 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3546 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3552 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3553 EVT VT = Op.getValueType();
3554 DebugLoc dl = Op.getDebugLoc();
3555 SDValue Vec = Op.getOperand(0);
3556 SDValue Lane = Op.getOperand(1);
3557 assert(VT == MVT::i32 &&
3558 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3559 "unexpected type for custom-lowering vector extract");
3560 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3563 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3564 // The only time a CONCAT_VECTORS operation can have legal types is when
3565 // two 64-bit vectors are concatenated to a 128-bit vector.
3566 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3567 "unexpected CONCAT_VECTORS");
3568 DebugLoc dl = Op.getDebugLoc();
3569 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3570 SDValue Op0 = Op.getOperand(0);
3571 SDValue Op1 = Op.getOperand(1);
3572 if (Op0.getOpcode() != ISD::UNDEF)
3573 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3574 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3575 DAG.getIntPtrConstant(0));
3576 if (Op1.getOpcode() != ISD::UNDEF)
3577 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3578 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3579 DAG.getIntPtrConstant(1));
3580 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3583 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3584 switch (Op.getOpcode()) {
3585 default: llvm_unreachable("Don't know how to custom lower this!");
3586 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3587 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3588 case ISD::GlobalAddress:
3589 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3590 LowerGlobalAddressELF(Op, DAG);
3591 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3592 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3593 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3594 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3595 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3596 case ISD::VASTART: return LowerVASTART(Op, DAG);
3597 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3598 case ISD::SINT_TO_FP:
3599 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3600 case ISD::FP_TO_SINT:
3601 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3602 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3603 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3604 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3605 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3606 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3607 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3608 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3610 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3613 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3614 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3615 case ISD::SRL_PARTS:
3616 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3617 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3618 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3619 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3620 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3621 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3622 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3627 /// ReplaceNodeResults - Replace the results of node with an illegal result
3628 /// type with new values built out of custom code.
3629 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3630 SmallVectorImpl<SDValue>&Results,
3631 SelectionDAG &DAG) const {
3633 switch (N->getOpcode()) {
3635 llvm_unreachable("Don't know how to custom expand this!");
3637 case ISD::BIT_CONVERT:
3638 Res = ExpandBIT_CONVERT(N, DAG);
3642 Res = LowerShift(N, DAG, Subtarget);
3646 Results.push_back(Res);
3649 //===----------------------------------------------------------------------===//
3650 // ARM Scheduler Hooks
3651 //===----------------------------------------------------------------------===//
3654 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3655 MachineBasicBlock *BB,
3656 unsigned Size) const {
3657 unsigned dest = MI->getOperand(0).getReg();
3658 unsigned ptr = MI->getOperand(1).getReg();
3659 unsigned oldval = MI->getOperand(2).getReg();
3660 unsigned newval = MI->getOperand(3).getReg();
3661 unsigned scratch = BB->getParent()->getRegInfo()
3662 .createVirtualRegister(ARM::GPRRegisterClass);
3663 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3664 DebugLoc dl = MI->getDebugLoc();
3665 bool isThumb2 = Subtarget->isThumb2();
3667 unsigned ldrOpc, strOpc;
3669 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3671 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3672 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3675 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3676 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3679 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3680 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3684 MachineFunction *MF = BB->getParent();
3685 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3686 MachineFunction::iterator It = BB;
3687 ++It; // insert the new blocks after the current block
3689 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3690 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3691 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3692 MF->insert(It, loop1MBB);
3693 MF->insert(It, loop2MBB);
3694 MF->insert(It, exitMBB);
3696 // Transfer the remainder of BB and its successor edges to exitMBB.
3697 exitMBB->splice(exitMBB->begin(), BB,
3698 llvm::next(MachineBasicBlock::iterator(MI)),
3700 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3704 // fallthrough --> loop1MBB
3705 BB->addSuccessor(loop1MBB);
3708 // ldrex dest, [ptr]
3712 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3713 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3714 .addReg(dest).addReg(oldval));
3715 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3716 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3717 BB->addSuccessor(loop2MBB);
3718 BB->addSuccessor(exitMBB);
3721 // strex scratch, newval, [ptr]
3725 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3727 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3728 .addReg(scratch).addImm(0));
3729 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3730 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3731 BB->addSuccessor(loop1MBB);
3732 BB->addSuccessor(exitMBB);
3738 MI->eraseFromParent(); // The instruction is gone now.
3744 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3745 unsigned Size, unsigned BinOpcode) const {
3746 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3747 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3749 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3750 MachineFunction *MF = BB->getParent();
3751 MachineFunction::iterator It = BB;
3754 unsigned dest = MI->getOperand(0).getReg();
3755 unsigned ptr = MI->getOperand(1).getReg();
3756 unsigned incr = MI->getOperand(2).getReg();
3757 DebugLoc dl = MI->getDebugLoc();
3759 bool isThumb2 = Subtarget->isThumb2();
3760 unsigned ldrOpc, strOpc;
3762 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3764 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3765 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3768 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3769 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3772 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3773 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3777 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3778 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3779 MF->insert(It, loopMBB);
3780 MF->insert(It, exitMBB);
3782 // Transfer the remainder of BB and its successor edges to exitMBB.
3783 exitMBB->splice(exitMBB->begin(), BB,
3784 llvm::next(MachineBasicBlock::iterator(MI)),
3786 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3788 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3789 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3790 unsigned scratch2 = (!BinOpcode) ? incr :
3791 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3795 // fallthrough --> loopMBB
3796 BB->addSuccessor(loopMBB);
3800 // <binop> scratch2, dest, incr
3801 // strex scratch, scratch2, ptr
3804 // fallthrough --> exitMBB
3806 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3808 // operand order needs to go the other way for NAND
3809 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3810 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3811 addReg(incr).addReg(dest)).addReg(0);
3813 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3814 addReg(dest).addReg(incr)).addReg(0);
3817 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3819 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3820 .addReg(scratch).addImm(0));
3821 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3822 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3824 BB->addSuccessor(loopMBB);
3825 BB->addSuccessor(exitMBB);
3831 MI->eraseFromParent(); // The instruction is gone now.
3837 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3838 MachineBasicBlock *BB) const {
3839 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3840 DebugLoc dl = MI->getDebugLoc();
3841 bool isThumb2 = Subtarget->isThumb2();
3842 switch (MI->getOpcode()) {
3845 llvm_unreachable("Unexpected instr type to insert");
3847 case ARM::ATOMIC_LOAD_ADD_I8:
3848 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3849 case ARM::ATOMIC_LOAD_ADD_I16:
3850 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3851 case ARM::ATOMIC_LOAD_ADD_I32:
3852 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3854 case ARM::ATOMIC_LOAD_AND_I8:
3855 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3856 case ARM::ATOMIC_LOAD_AND_I16:
3857 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3858 case ARM::ATOMIC_LOAD_AND_I32:
3859 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3861 case ARM::ATOMIC_LOAD_OR_I8:
3862 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3863 case ARM::ATOMIC_LOAD_OR_I16:
3864 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3865 case ARM::ATOMIC_LOAD_OR_I32:
3866 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3868 case ARM::ATOMIC_LOAD_XOR_I8:
3869 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3870 case ARM::ATOMIC_LOAD_XOR_I16:
3871 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3872 case ARM::ATOMIC_LOAD_XOR_I32:
3873 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3875 case ARM::ATOMIC_LOAD_NAND_I8:
3876 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3877 case ARM::ATOMIC_LOAD_NAND_I16:
3878 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3879 case ARM::ATOMIC_LOAD_NAND_I32:
3880 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3882 case ARM::ATOMIC_LOAD_SUB_I8:
3883 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3884 case ARM::ATOMIC_LOAD_SUB_I16:
3885 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3886 case ARM::ATOMIC_LOAD_SUB_I32:
3887 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3889 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3890 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3891 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3893 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3894 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3895 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3897 case ARM::tMOVCCr_pseudo: {
3898 // To "insert" a SELECT_CC instruction, we actually have to insert the
3899 // diamond control-flow pattern. The incoming instruction knows the
3900 // destination vreg to set, the condition code register to branch on, the
3901 // true/false values to select between, and a branch opcode to use.
3902 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3903 MachineFunction::iterator It = BB;
3909 // cmpTY ccX, r1, r2
3911 // fallthrough --> copy0MBB
3912 MachineBasicBlock *thisMBB = BB;
3913 MachineFunction *F = BB->getParent();
3914 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3915 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3916 F->insert(It, copy0MBB);
3917 F->insert(It, sinkMBB);
3919 // Transfer the remainder of BB and its successor edges to sinkMBB.
3920 sinkMBB->splice(sinkMBB->begin(), BB,
3921 llvm::next(MachineBasicBlock::iterator(MI)),
3923 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
3925 BB->addSuccessor(copy0MBB);
3926 BB->addSuccessor(sinkMBB);
3928 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3929 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3932 // %FalseValue = ...
3933 // # fallthrough to sinkMBB
3936 // Update machine-CFG edges
3937 BB->addSuccessor(sinkMBB);
3940 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3943 BuildMI(*BB, BB->begin(), dl,
3944 TII->get(ARM::PHI), MI->getOperand(0).getReg())
3945 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3946 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3948 MI->eraseFromParent(); // The pseudo instruction is gone now.
3955 case ARM::t2SUBrSPi_:
3956 case ARM::t2SUBrSPi12_:
3957 case ARM::t2SUBrSPs_: {
3958 MachineFunction *MF = BB->getParent();
3959 unsigned DstReg = MI->getOperand(0).getReg();
3960 unsigned SrcReg = MI->getOperand(1).getReg();
3961 bool DstIsDead = MI->getOperand(0).isDead();
3962 bool SrcIsKill = MI->getOperand(1).isKill();
3964 if (SrcReg != ARM::SP) {
3965 // Copy the source to SP from virtual register.
3966 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3967 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3968 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3969 BuildMI(*BB, MI, dl, TII->get(CopyOpc), ARM::SP)
3970 .addReg(SrcReg, getKillRegState(SrcIsKill));
3974 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3975 switch (MI->getOpcode()) {
3977 llvm_unreachable("Unexpected pseudo instruction!");
3983 OpOpc = ARM::tADDspr;
3986 OpOpc = ARM::tSUBspi;
3988 case ARM::t2SUBrSPi_:
3989 OpOpc = ARM::t2SUBrSPi;
3990 NeedPred = true; NeedCC = true;
3992 case ARM::t2SUBrSPi12_:
3993 OpOpc = ARM::t2SUBrSPi12;
3996 case ARM::t2SUBrSPs_:
3997 OpOpc = ARM::t2SUBrSPs;
3998 NeedPred = true; NeedCC = true; NeedOp3 = true;
4001 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(OpOpc), ARM::SP);
4002 if (OpOpc == ARM::tAND)
4003 AddDefaultT1CC(MIB);
4004 MIB.addReg(ARM::SP);
4005 MIB.addOperand(MI->getOperand(2));
4007 MIB.addOperand(MI->getOperand(3));
4009 AddDefaultPred(MIB);
4013 // Copy the result from SP to virtual register.
4014 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
4015 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4016 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
4017 BuildMI(*BB, MI, dl, TII->get(CopyOpc))
4018 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
4020 MI->eraseFromParent(); // The pseudo instruction is gone now.
4026 //===----------------------------------------------------------------------===//
4027 // ARM Optimization Hooks
4028 //===----------------------------------------------------------------------===//
4031 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4032 TargetLowering::DAGCombinerInfo &DCI) {
4033 SelectionDAG &DAG = DCI.DAG;
4034 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4035 EVT VT = N->getValueType(0);
4036 unsigned Opc = N->getOpcode();
4037 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4038 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4039 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4040 ISD::CondCode CC = ISD::SETCC_INVALID;
4043 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4045 SDValue CCOp = Slct.getOperand(0);
4046 if (CCOp.getOpcode() == ISD::SETCC)
4047 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4050 bool DoXform = false;
4052 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4055 if (LHS.getOpcode() == ISD::Constant &&
4056 cast<ConstantSDNode>(LHS)->isNullValue()) {
4058 } else if (CC != ISD::SETCC_INVALID &&
4059 RHS.getOpcode() == ISD::Constant &&
4060 cast<ConstantSDNode>(RHS)->isNullValue()) {
4061 std::swap(LHS, RHS);
4062 SDValue Op0 = Slct.getOperand(0);
4063 EVT OpVT = isSlctCC ? Op0.getValueType() :
4064 Op0.getOperand(0).getValueType();
4065 bool isInt = OpVT.isInteger();
4066 CC = ISD::getSetCCInverse(CC, isInt);
4068 if (!TLI.isCondCodeLegal(CC, OpVT))
4069 return SDValue(); // Inverse operator isn't legal.
4076 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4078 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4079 Slct.getOperand(0), Slct.getOperand(1), CC);
4080 SDValue CCOp = Slct.getOperand(0);
4082 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4083 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4084 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4085 CCOp, OtherOp, Result);
4090 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4091 static SDValue PerformADDCombine(SDNode *N,
4092 TargetLowering::DAGCombinerInfo &DCI) {
4093 // added by evan in r37685 with no testcase.
4094 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4096 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4097 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4098 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4099 if (Result.getNode()) return Result;
4101 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4102 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4103 if (Result.getNode()) return Result;
4109 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4110 static SDValue PerformSUBCombine(SDNode *N,
4111 TargetLowering::DAGCombinerInfo &DCI) {
4112 // added by evan in r37685 with no testcase.
4113 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4115 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4116 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4117 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4118 if (Result.getNode()) return Result;
4124 static SDValue PerformMULCombine(SDNode *N,
4125 TargetLowering::DAGCombinerInfo &DCI,
4126 const ARMSubtarget *Subtarget) {
4127 SelectionDAG &DAG = DCI.DAG;
4129 if (Subtarget->isThumb1Only())
4132 if (DAG.getMachineFunction().
4133 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4136 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4139 EVT VT = N->getValueType(0);
4143 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4147 uint64_t MulAmt = C->getZExtValue();
4148 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4149 ShiftAmt = ShiftAmt & (32 - 1);
4150 SDValue V = N->getOperand(0);
4151 DebugLoc DL = N->getDebugLoc();
4154 MulAmt >>= ShiftAmt;
4155 if (isPowerOf2_32(MulAmt - 1)) {
4156 // (mul x, 2^N + 1) => (add (shl x, N), x)
4157 Res = DAG.getNode(ISD::ADD, DL, VT,
4158 V, DAG.getNode(ISD::SHL, DL, VT,
4159 V, DAG.getConstant(Log2_32(MulAmt-1),
4161 } else if (isPowerOf2_32(MulAmt + 1)) {
4162 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4163 Res = DAG.getNode(ISD::SUB, DL, VT,
4164 DAG.getNode(ISD::SHL, DL, VT,
4165 V, DAG.getConstant(Log2_32(MulAmt+1),
4172 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4173 DAG.getConstant(ShiftAmt, MVT::i32));
4175 // Do not add new nodes to DAG combiner worklist.
4176 DCI.CombineTo(N, Res, false);
4180 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4181 /// ARMISD::VMOVRRD.
4182 static SDValue PerformVMOVRRDCombine(SDNode *N,
4183 TargetLowering::DAGCombinerInfo &DCI) {
4184 // fmrrd(fmdrr x, y) -> x,y
4185 SDValue InDouble = N->getOperand(0);
4186 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4187 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4191 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4192 /// operand of a vector shift operation, where all the elements of the
4193 /// build_vector must have the same constant integer value.
4194 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4195 // Ignore bit_converts.
4196 while (Op.getOpcode() == ISD::BIT_CONVERT)
4197 Op = Op.getOperand(0);
4198 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4199 APInt SplatBits, SplatUndef;
4200 unsigned SplatBitSize;
4202 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4203 HasAnyUndefs, ElementBits) ||
4204 SplatBitSize > ElementBits)
4206 Cnt = SplatBits.getSExtValue();
4210 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4211 /// operand of a vector shift left operation. That value must be in the range:
4212 /// 0 <= Value < ElementBits for a left shift; or
4213 /// 0 <= Value <= ElementBits for a long left shift.
4214 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4215 assert(VT.isVector() && "vector shift count is not a vector type");
4216 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4217 if (! getVShiftImm(Op, ElementBits, Cnt))
4219 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4222 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4223 /// operand of a vector shift right operation. For a shift opcode, the value
4224 /// is positive, but for an intrinsic the value count must be negative. The
4225 /// absolute value must be in the range:
4226 /// 1 <= |Value| <= ElementBits for a right shift; or
4227 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4228 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4230 assert(VT.isVector() && "vector shift count is not a vector type");
4231 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4232 if (! getVShiftImm(Op, ElementBits, Cnt))
4236 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4239 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4240 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4241 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4244 // Don't do anything for most intrinsics.
4247 // Vector shifts: check for immediate versions and lower them.
4248 // Note: This is done during DAG combining instead of DAG legalizing because
4249 // the build_vectors for 64-bit vector element shift counts are generally
4250 // not legal, and it is hard to see their values after they get legalized to
4251 // loads from a constant pool.
4252 case Intrinsic::arm_neon_vshifts:
4253 case Intrinsic::arm_neon_vshiftu:
4254 case Intrinsic::arm_neon_vshiftls:
4255 case Intrinsic::arm_neon_vshiftlu:
4256 case Intrinsic::arm_neon_vshiftn:
4257 case Intrinsic::arm_neon_vrshifts:
4258 case Intrinsic::arm_neon_vrshiftu:
4259 case Intrinsic::arm_neon_vrshiftn:
4260 case Intrinsic::arm_neon_vqshifts:
4261 case Intrinsic::arm_neon_vqshiftu:
4262 case Intrinsic::arm_neon_vqshiftsu:
4263 case Intrinsic::arm_neon_vqshiftns:
4264 case Intrinsic::arm_neon_vqshiftnu:
4265 case Intrinsic::arm_neon_vqshiftnsu:
4266 case Intrinsic::arm_neon_vqrshiftns:
4267 case Intrinsic::arm_neon_vqrshiftnu:
4268 case Intrinsic::arm_neon_vqrshiftnsu: {
4269 EVT VT = N->getOperand(1).getValueType();
4271 unsigned VShiftOpc = 0;
4274 case Intrinsic::arm_neon_vshifts:
4275 case Intrinsic::arm_neon_vshiftu:
4276 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4277 VShiftOpc = ARMISD::VSHL;
4280 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4281 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4282 ARMISD::VSHRs : ARMISD::VSHRu);
4287 case Intrinsic::arm_neon_vshiftls:
4288 case Intrinsic::arm_neon_vshiftlu:
4289 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4291 llvm_unreachable("invalid shift count for vshll intrinsic");
4293 case Intrinsic::arm_neon_vrshifts:
4294 case Intrinsic::arm_neon_vrshiftu:
4295 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4299 case Intrinsic::arm_neon_vqshifts:
4300 case Intrinsic::arm_neon_vqshiftu:
4301 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4305 case Intrinsic::arm_neon_vqshiftsu:
4306 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4308 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4310 case Intrinsic::arm_neon_vshiftn:
4311 case Intrinsic::arm_neon_vrshiftn:
4312 case Intrinsic::arm_neon_vqshiftns:
4313 case Intrinsic::arm_neon_vqshiftnu:
4314 case Intrinsic::arm_neon_vqshiftnsu:
4315 case Intrinsic::arm_neon_vqrshiftns:
4316 case Intrinsic::arm_neon_vqrshiftnu:
4317 case Intrinsic::arm_neon_vqrshiftnsu:
4318 // Narrowing shifts require an immediate right shift.
4319 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4321 llvm_unreachable("invalid shift count for narrowing vector shift "
4325 llvm_unreachable("unhandled vector shift");
4329 case Intrinsic::arm_neon_vshifts:
4330 case Intrinsic::arm_neon_vshiftu:
4331 // Opcode already set above.
4333 case Intrinsic::arm_neon_vshiftls:
4334 case Intrinsic::arm_neon_vshiftlu:
4335 if (Cnt == VT.getVectorElementType().getSizeInBits())
4336 VShiftOpc = ARMISD::VSHLLi;
4338 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4339 ARMISD::VSHLLs : ARMISD::VSHLLu);
4341 case Intrinsic::arm_neon_vshiftn:
4342 VShiftOpc = ARMISD::VSHRN; break;
4343 case Intrinsic::arm_neon_vrshifts:
4344 VShiftOpc = ARMISD::VRSHRs; break;
4345 case Intrinsic::arm_neon_vrshiftu:
4346 VShiftOpc = ARMISD::VRSHRu; break;
4347 case Intrinsic::arm_neon_vrshiftn:
4348 VShiftOpc = ARMISD::VRSHRN; break;
4349 case Intrinsic::arm_neon_vqshifts:
4350 VShiftOpc = ARMISD::VQSHLs; break;
4351 case Intrinsic::arm_neon_vqshiftu:
4352 VShiftOpc = ARMISD::VQSHLu; break;
4353 case Intrinsic::arm_neon_vqshiftsu:
4354 VShiftOpc = ARMISD::VQSHLsu; break;
4355 case Intrinsic::arm_neon_vqshiftns:
4356 VShiftOpc = ARMISD::VQSHRNs; break;
4357 case Intrinsic::arm_neon_vqshiftnu:
4358 VShiftOpc = ARMISD::VQSHRNu; break;
4359 case Intrinsic::arm_neon_vqshiftnsu:
4360 VShiftOpc = ARMISD::VQSHRNsu; break;
4361 case Intrinsic::arm_neon_vqrshiftns:
4362 VShiftOpc = ARMISD::VQRSHRNs; break;
4363 case Intrinsic::arm_neon_vqrshiftnu:
4364 VShiftOpc = ARMISD::VQRSHRNu; break;
4365 case Intrinsic::arm_neon_vqrshiftnsu:
4366 VShiftOpc = ARMISD::VQRSHRNsu; break;
4369 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4370 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4373 case Intrinsic::arm_neon_vshiftins: {
4374 EVT VT = N->getOperand(1).getValueType();
4376 unsigned VShiftOpc = 0;
4378 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4379 VShiftOpc = ARMISD::VSLI;
4380 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4381 VShiftOpc = ARMISD::VSRI;
4383 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4386 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4387 N->getOperand(1), N->getOperand(2),
4388 DAG.getConstant(Cnt, MVT::i32));
4391 case Intrinsic::arm_neon_vqrshifts:
4392 case Intrinsic::arm_neon_vqrshiftu:
4393 // No immediate versions of these to check for.
4400 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4401 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4402 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4403 /// vector element shift counts are generally not legal, and it is hard to see
4404 /// their values after they get legalized to loads from a constant pool.
4405 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4406 const ARMSubtarget *ST) {
4407 EVT VT = N->getValueType(0);
4409 // Nothing to be done for scalar shifts.
4410 if (! VT.isVector())
4413 assert(ST->hasNEON() && "unexpected vector shift");
4416 switch (N->getOpcode()) {
4417 default: llvm_unreachable("unexpected shift opcode");
4420 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4421 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4422 DAG.getConstant(Cnt, MVT::i32));
4427 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4428 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4429 ARMISD::VSHRs : ARMISD::VSHRu);
4430 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4431 DAG.getConstant(Cnt, MVT::i32));
4437 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4438 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4439 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4440 const ARMSubtarget *ST) {
4441 SDValue N0 = N->getOperand(0);
4443 // Check for sign- and zero-extensions of vector extract operations of 8-
4444 // and 16-bit vector elements. NEON supports these directly. They are
4445 // handled during DAG combining because type legalization will promote them
4446 // to 32-bit types and it is messy to recognize the operations after that.
4447 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4448 SDValue Vec = N0.getOperand(0);
4449 SDValue Lane = N0.getOperand(1);
4450 EVT VT = N->getValueType(0);
4451 EVT EltVT = N0.getValueType();
4452 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4454 if (VT == MVT::i32 &&
4455 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4456 TLI.isTypeLegal(Vec.getValueType())) {
4459 switch (N->getOpcode()) {
4460 default: llvm_unreachable("unexpected opcode");
4461 case ISD::SIGN_EXTEND:
4462 Opc = ARMISD::VGETLANEs;
4464 case ISD::ZERO_EXTEND:
4465 case ISD::ANY_EXTEND:
4466 Opc = ARMISD::VGETLANEu;
4469 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4476 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4477 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4478 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4479 const ARMSubtarget *ST) {
4480 // If the target supports NEON, try to use vmax/vmin instructions for f32
4481 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4482 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4483 // a NaN; only do the transformation when it matches that behavior.
4485 // For now only do this when using NEON for FP operations; if using VFP, it
4486 // is not obvious that the benefit outweighs the cost of switching to the
4488 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4489 N->getValueType(0) != MVT::f32)
4492 SDValue CondLHS = N->getOperand(0);
4493 SDValue CondRHS = N->getOperand(1);
4494 SDValue LHS = N->getOperand(2);
4495 SDValue RHS = N->getOperand(3);
4496 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4498 unsigned Opcode = 0;
4500 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4501 IsReversed = false; // x CC y ? x : y
4502 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4503 IsReversed = true ; // x CC y ? y : x
4517 // If LHS is NaN, an ordered comparison will be false and the result will
4518 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4519 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4520 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4521 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4523 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4524 // will return -0, so vmin can only be used for unsafe math or if one of
4525 // the operands is known to be nonzero.
4526 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4528 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4530 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4539 // If LHS is NaN, an ordered comparison will be false and the result will
4540 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4541 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4542 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4543 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4545 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4546 // will return +0, so vmax can only be used for unsafe math or if one of
4547 // the operands is known to be nonzero.
4548 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4550 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4552 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4558 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4561 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4562 DAGCombinerInfo &DCI) const {
4563 switch (N->getOpcode()) {
4565 case ISD::ADD: return PerformADDCombine(N, DCI);
4566 case ISD::SUB: return PerformSUBCombine(N, DCI);
4567 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4568 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4569 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4572 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4573 case ISD::SIGN_EXTEND:
4574 case ISD::ZERO_EXTEND:
4575 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4576 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4581 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4582 if (!Subtarget->hasV6Ops())
4583 // Pre-v6 does not support unaligned mem access.
4586 // v6+ may or may not support unaligned mem access depending on the system
4588 // FIXME: This is pretty conservative. Should we provide cmdline option to
4589 // control the behaviour?
4590 if (!Subtarget->isTargetDarwin())
4593 switch (VT.getSimpleVT().SimpleTy) {
4600 // FIXME: VLD1 etc with standard alignment is legal.
4604 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4609 switch (VT.getSimpleVT().SimpleTy) {
4610 default: return false;
4625 if ((V & (Scale - 1)) != 0)
4628 return V == (V & ((1LL << 5) - 1));
4631 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4632 const ARMSubtarget *Subtarget) {
4639 switch (VT.getSimpleVT().SimpleTy) {
4640 default: return false;
4645 // + imm12 or - imm8
4647 return V == (V & ((1LL << 8) - 1));
4648 return V == (V & ((1LL << 12) - 1));
4651 // Same as ARM mode. FIXME: NEON?
4652 if (!Subtarget->hasVFP2())
4657 return V == (V & ((1LL << 8) - 1));
4661 /// isLegalAddressImmediate - Return true if the integer value can be used
4662 /// as the offset of the target addressing mode for load / store of the
4664 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4665 const ARMSubtarget *Subtarget) {
4672 if (Subtarget->isThumb1Only())
4673 return isLegalT1AddressImmediate(V, VT);
4674 else if (Subtarget->isThumb2())
4675 return isLegalT2AddressImmediate(V, VT, Subtarget);
4680 switch (VT.getSimpleVT().SimpleTy) {
4681 default: return false;
4686 return V == (V & ((1LL << 12) - 1));
4689 return V == (V & ((1LL << 8) - 1));
4692 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4697 return V == (V & ((1LL << 8) - 1));
4701 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4703 int Scale = AM.Scale;
4707 switch (VT.getSimpleVT().SimpleTy) {
4708 default: return false;
4717 return Scale == 2 || Scale == 4 || Scale == 8;
4720 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4724 // Note, we allow "void" uses (basically, uses that aren't loads or
4725 // stores), because arm allows folding a scale into many arithmetic
4726 // operations. This should be made more precise and revisited later.
4728 // Allow r << imm, but the imm has to be a multiple of two.
4729 if (Scale & 1) return false;
4730 return isPowerOf2_32(Scale);
4734 /// isLegalAddressingMode - Return true if the addressing mode represented
4735 /// by AM is legal for this target, for a load/store of the specified type.
4736 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4737 const Type *Ty) const {
4738 EVT VT = getValueType(Ty, true);
4739 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4742 // Can never fold addr of global into load/store.
4747 case 0: // no scale reg, must be "r+i" or "r", or "i".
4750 if (Subtarget->isThumb1Only())
4754 // ARM doesn't support any R+R*scale+imm addr modes.
4761 if (Subtarget->isThumb2())
4762 return isLegalT2ScaledAddressingMode(AM, VT);
4764 int Scale = AM.Scale;
4765 switch (VT.getSimpleVT().SimpleTy) {
4766 default: return false;
4770 if (Scale < 0) Scale = -Scale;
4774 return isPowerOf2_32(Scale & ~1);
4778 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4783 // Note, we allow "void" uses (basically, uses that aren't loads or
4784 // stores), because arm allows folding a scale into many arithmetic
4785 // operations. This should be made more precise and revisited later.
4787 // Allow r << imm, but the imm has to be a multiple of two.
4788 if (Scale & 1) return false;
4789 return isPowerOf2_32(Scale);
4796 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4797 /// icmp immediate, that is the target has icmp instructions which can compare
4798 /// a register against the immediate without having to materialize the
4799 /// immediate into a register.
4800 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4801 if (!Subtarget->isThumb())
4802 return ARM_AM::getSOImmVal(Imm) != -1;
4803 if (Subtarget->isThumb2())
4804 return ARM_AM::getT2SOImmVal(Imm) != -1;
4805 return Imm >= 0 && Imm <= 255;
4808 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4809 bool isSEXTLoad, SDValue &Base,
4810 SDValue &Offset, bool &isInc,
4811 SelectionDAG &DAG) {
4812 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4815 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4817 Base = Ptr->getOperand(0);
4818 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4819 int RHSC = (int)RHS->getZExtValue();
4820 if (RHSC < 0 && RHSC > -256) {
4821 assert(Ptr->getOpcode() == ISD::ADD);
4823 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4827 isInc = (Ptr->getOpcode() == ISD::ADD);
4828 Offset = Ptr->getOperand(1);
4830 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4832 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4833 int RHSC = (int)RHS->getZExtValue();
4834 if (RHSC < 0 && RHSC > -0x1000) {
4835 assert(Ptr->getOpcode() == ISD::ADD);
4837 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4838 Base = Ptr->getOperand(0);
4843 if (Ptr->getOpcode() == ISD::ADD) {
4845 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4846 if (ShOpcVal != ARM_AM::no_shift) {
4847 Base = Ptr->getOperand(1);
4848 Offset = Ptr->getOperand(0);
4850 Base = Ptr->getOperand(0);
4851 Offset = Ptr->getOperand(1);
4856 isInc = (Ptr->getOpcode() == ISD::ADD);
4857 Base = Ptr->getOperand(0);
4858 Offset = Ptr->getOperand(1);
4862 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4866 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4867 bool isSEXTLoad, SDValue &Base,
4868 SDValue &Offset, bool &isInc,
4869 SelectionDAG &DAG) {
4870 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4873 Base = Ptr->getOperand(0);
4874 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4875 int RHSC = (int)RHS->getZExtValue();
4876 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4877 assert(Ptr->getOpcode() == ISD::ADD);
4879 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4881 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4882 isInc = Ptr->getOpcode() == ISD::ADD;
4883 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4891 /// getPreIndexedAddressParts - returns true by value, base pointer and
4892 /// offset pointer and addressing mode by reference if the node's address
4893 /// can be legally represented as pre-indexed load / store address.
4895 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4897 ISD::MemIndexedMode &AM,
4898 SelectionDAG &DAG) const {
4899 if (Subtarget->isThumb1Only())
4904 bool isSEXTLoad = false;
4905 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4906 Ptr = LD->getBasePtr();
4907 VT = LD->getMemoryVT();
4908 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4909 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4910 Ptr = ST->getBasePtr();
4911 VT = ST->getMemoryVT();
4916 bool isLegal = false;
4917 if (Subtarget->isThumb2())
4918 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4919 Offset, isInc, DAG);
4921 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4922 Offset, isInc, DAG);
4926 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4930 /// getPostIndexedAddressParts - returns true by value, base pointer and
4931 /// offset pointer and addressing mode by reference if this node can be
4932 /// combined with a load / store to form a post-indexed load / store.
4933 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4936 ISD::MemIndexedMode &AM,
4937 SelectionDAG &DAG) const {
4938 if (Subtarget->isThumb1Only())
4943 bool isSEXTLoad = false;
4944 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4945 VT = LD->getMemoryVT();
4946 Ptr = LD->getBasePtr();
4947 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4948 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4949 VT = ST->getMemoryVT();
4950 Ptr = ST->getBasePtr();
4955 bool isLegal = false;
4956 if (Subtarget->isThumb2())
4957 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4960 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4966 // Swap base ptr and offset to catch more post-index load / store when
4967 // it's legal. In Thumb2 mode, offset must be an immediate.
4968 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4969 !Subtarget->isThumb2())
4970 std::swap(Base, Offset);
4972 // Post-indexed load / store update the base pointer.
4977 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4981 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4985 const SelectionDAG &DAG,
4986 unsigned Depth) const {
4987 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4988 switch (Op.getOpcode()) {
4990 case ARMISD::CMOV: {
4991 // Bits are known zero/one if known on the LHS and RHS.
4992 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4993 if (KnownZero == 0 && KnownOne == 0) return;
4995 APInt KnownZeroRHS, KnownOneRHS;
4996 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4997 KnownZeroRHS, KnownOneRHS, Depth+1);
4998 KnownZero &= KnownZeroRHS;
4999 KnownOne &= KnownOneRHS;
5005 //===----------------------------------------------------------------------===//
5006 // ARM Inline Assembly Support
5007 //===----------------------------------------------------------------------===//
5009 /// getConstraintType - Given a constraint letter, return the type of
5010 /// constraint it is for this target.
5011 ARMTargetLowering::ConstraintType
5012 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5013 if (Constraint.size() == 1) {
5014 switch (Constraint[0]) {
5016 case 'l': return C_RegisterClass;
5017 case 'w': return C_RegisterClass;
5020 return TargetLowering::getConstraintType(Constraint);
5023 std::pair<unsigned, const TargetRegisterClass*>
5024 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5026 if (Constraint.size() == 1) {
5027 // GCC ARM Constraint Letters
5028 switch (Constraint[0]) {
5030 if (Subtarget->isThumb())
5031 return std::make_pair(0U, ARM::tGPRRegisterClass);
5033 return std::make_pair(0U, ARM::GPRRegisterClass);
5035 return std::make_pair(0U, ARM::GPRRegisterClass);
5038 return std::make_pair(0U, ARM::SPRRegisterClass);
5039 if (VT.getSizeInBits() == 64)
5040 return std::make_pair(0U, ARM::DPRRegisterClass);
5041 if (VT.getSizeInBits() == 128)
5042 return std::make_pair(0U, ARM::QPRRegisterClass);
5046 if (StringRef("{cc}").equals_lower(Constraint))
5047 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5049 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5052 std::vector<unsigned> ARMTargetLowering::
5053 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5055 if (Constraint.size() != 1)
5056 return std::vector<unsigned>();
5058 switch (Constraint[0]) { // GCC ARM Constraint Letters
5061 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5062 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5065 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5066 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5067 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5068 ARM::R12, ARM::LR, 0);
5071 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5072 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5073 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5074 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5075 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5076 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5077 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5078 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5079 if (VT.getSizeInBits() == 64)
5080 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5081 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5082 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5083 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5084 if (VT.getSizeInBits() == 128)
5085 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5086 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5090 return std::vector<unsigned>();
5093 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5094 /// vector. If it is invalid, don't add anything to Ops.
5095 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5097 std::vector<SDValue>&Ops,
5098 SelectionDAG &DAG) const {
5099 SDValue Result(0, 0);
5101 switch (Constraint) {
5103 case 'I': case 'J': case 'K': case 'L':
5104 case 'M': case 'N': case 'O':
5105 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5109 int64_t CVal64 = C->getSExtValue();
5110 int CVal = (int) CVal64;
5111 // None of these constraints allow values larger than 32 bits. Check
5112 // that the value fits in an int.
5116 switch (Constraint) {
5118 if (Subtarget->isThumb1Only()) {
5119 // This must be a constant between 0 and 255, for ADD
5121 if (CVal >= 0 && CVal <= 255)
5123 } else if (Subtarget->isThumb2()) {
5124 // A constant that can be used as an immediate value in a
5125 // data-processing instruction.
5126 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5129 // A constant that can be used as an immediate value in a
5130 // data-processing instruction.
5131 if (ARM_AM::getSOImmVal(CVal) != -1)
5137 if (Subtarget->isThumb()) { // FIXME thumb2
5138 // This must be a constant between -255 and -1, for negated ADD
5139 // immediates. This can be used in GCC with an "n" modifier that
5140 // prints the negated value, for use with SUB instructions. It is
5141 // not useful otherwise but is implemented for compatibility.
5142 if (CVal >= -255 && CVal <= -1)
5145 // This must be a constant between -4095 and 4095. It is not clear
5146 // what this constraint is intended for. Implemented for
5147 // compatibility with GCC.
5148 if (CVal >= -4095 && CVal <= 4095)
5154 if (Subtarget->isThumb1Only()) {
5155 // A 32-bit value where only one byte has a nonzero value. Exclude
5156 // zero to match GCC. This constraint is used by GCC internally for
5157 // constants that can be loaded with a move/shift combination.
5158 // It is not useful otherwise but is implemented for compatibility.
5159 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5161 } else if (Subtarget->isThumb2()) {
5162 // A constant whose bitwise inverse can be used as an immediate
5163 // value in a data-processing instruction. This can be used in GCC
5164 // with a "B" modifier that prints the inverted value, for use with
5165 // BIC and MVN instructions. It is not useful otherwise but is
5166 // implemented for compatibility.
5167 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5170 // A constant whose bitwise inverse can be used as an immediate
5171 // value in a data-processing instruction. This can be used in GCC
5172 // with a "B" modifier that prints the inverted value, for use with
5173 // BIC and MVN instructions. It is not useful otherwise but is
5174 // implemented for compatibility.
5175 if (ARM_AM::getSOImmVal(~CVal) != -1)
5181 if (Subtarget->isThumb1Only()) {
5182 // This must be a constant between -7 and 7,
5183 // for 3-operand ADD/SUB immediate instructions.
5184 if (CVal >= -7 && CVal < 7)
5186 } else if (Subtarget->isThumb2()) {
5187 // A constant whose negation can be used as an immediate value in a
5188 // data-processing instruction. This can be used in GCC with an "n"
5189 // modifier that prints the negated value, for use with SUB
5190 // instructions. It is not useful otherwise but is implemented for
5192 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5195 // A constant whose negation can be used as an immediate value in a
5196 // data-processing instruction. This can be used in GCC with an "n"
5197 // modifier that prints the negated value, for use with SUB
5198 // instructions. It is not useful otherwise but is implemented for
5200 if (ARM_AM::getSOImmVal(-CVal) != -1)
5206 if (Subtarget->isThumb()) { // FIXME thumb2
5207 // This must be a multiple of 4 between 0 and 1020, for
5208 // ADD sp + immediate.
5209 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5212 // A power of two or a constant between 0 and 32. This is used in
5213 // GCC for the shift amount on shifted register operands, but it is
5214 // useful in general for any shift amounts.
5215 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5221 if (Subtarget->isThumb()) { // FIXME thumb2
5222 // This must be a constant between 0 and 31, for shift amounts.
5223 if (CVal >= 0 && CVal <= 31)
5229 if (Subtarget->isThumb()) { // FIXME thumb2
5230 // This must be a multiple of 4 between -508 and 508, for
5231 // ADD/SUB sp = sp + immediate.
5232 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5237 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5241 if (Result.getNode()) {
5242 Ops.push_back(Result);
5245 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5249 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5250 // The ARM target isn't yet aware of offsets.
5254 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5255 APInt Imm = FPImm.bitcastToAPInt();
5256 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5257 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5258 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5260 // We can handle 4 bits of mantissa.
5261 // mantissa = (16+UInt(e:f:g:h))/16.
5262 if (Mantissa & 0x7ffff)
5265 if ((Mantissa & 0xf) != Mantissa)
5268 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5269 if (Exp < -3 || Exp > 4)
5271 Exp = ((Exp+3) & 0x7) ^ 4;
5273 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5276 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5277 APInt Imm = FPImm.bitcastToAPInt();
5278 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5279 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5280 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5282 // We can handle 4 bits of mantissa.
5283 // mantissa = (16+UInt(e:f:g:h))/16.
5284 if (Mantissa & 0xffffffffffffLL)
5287 if ((Mantissa & 0xf) != Mantissa)
5290 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5291 if (Exp < -3 || Exp > 4)
5293 Exp = ((Exp+3) & 0x7) ^ 4;
5295 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5298 /// isFPImmLegal - Returns true if the target can instruction select the
5299 /// specified FP immediate natively. If false, the legalizer will
5300 /// materialize the FP immediate as a load from a constant pool.
5301 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5302 if (!Subtarget->hasVFP3())
5305 return ARM::getVFPf32Imm(Imm) != -1;
5307 return ARM::getVFPf64Imm(Imm) != -1;