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 if (EnableARMCodePlacement)
543 benefitFromCodePlacementOpt = true;
546 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
549 case ARMISD::Wrapper: return "ARMISD::Wrapper";
550 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
551 case ARMISD::CALL: return "ARMISD::CALL";
552 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
553 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
554 case ARMISD::tCALL: return "ARMISD::tCALL";
555 case ARMISD::BRCOND: return "ARMISD::BRCOND";
556 case ARMISD::BR_JT: return "ARMISD::BR_JT";
557 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
558 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
559 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
560 case ARMISD::CMP: return "ARMISD::CMP";
561 case ARMISD::CMPZ: return "ARMISD::CMPZ";
562 case ARMISD::CMPFP: return "ARMISD::CMPFP";
563 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
564 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
565 case ARMISD::CMOV: return "ARMISD::CMOV";
566 case ARMISD::CNEG: return "ARMISD::CNEG";
568 case ARMISD::RBIT: return "ARMISD::RBIT";
570 case ARMISD::FTOSI: return "ARMISD::FTOSI";
571 case ARMISD::FTOUI: return "ARMISD::FTOUI";
572 case ARMISD::SITOF: return "ARMISD::SITOF";
573 case ARMISD::UITOF: return "ARMISD::UITOF";
575 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
576 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
577 case ARMISD::RRX: return "ARMISD::RRX";
579 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
580 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
582 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
583 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
585 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
587 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
589 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
591 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
592 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
594 case ARMISD::VCEQ: return "ARMISD::VCEQ";
595 case ARMISD::VCGE: return "ARMISD::VCGE";
596 case ARMISD::VCGEU: return "ARMISD::VCGEU";
597 case ARMISD::VCGT: return "ARMISD::VCGT";
598 case ARMISD::VCGTU: return "ARMISD::VCGTU";
599 case ARMISD::VTST: return "ARMISD::VTST";
601 case ARMISD::VSHL: return "ARMISD::VSHL";
602 case ARMISD::VSHRs: return "ARMISD::VSHRs";
603 case ARMISD::VSHRu: return "ARMISD::VSHRu";
604 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
605 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
606 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
607 case ARMISD::VSHRN: return "ARMISD::VSHRN";
608 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
609 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
610 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
611 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
612 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
613 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
614 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
615 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
616 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
617 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
618 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
619 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
620 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
621 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
622 case ARMISD::VDUP: return "ARMISD::VDUP";
623 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
624 case ARMISD::VEXT: return "ARMISD::VEXT";
625 case ARMISD::VREV64: return "ARMISD::VREV64";
626 case ARMISD::VREV32: return "ARMISD::VREV32";
627 case ARMISD::VREV16: return "ARMISD::VREV16";
628 case ARMISD::VZIP: return "ARMISD::VZIP";
629 case ARMISD::VUZP: return "ARMISD::VUZP";
630 case ARMISD::VTRN: return "ARMISD::VTRN";
631 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
632 case ARMISD::FMAX: return "ARMISD::FMAX";
633 case ARMISD::FMIN: return "ARMISD::FMIN";
637 /// getRegClassFor - Return the register class that should be used for the
638 /// specified value type.
639 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
640 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
641 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
642 // load / store 4 to 8 consecutive D registers.
643 if (Subtarget->hasNEON()) {
644 if (VT == MVT::v4i64)
645 return ARM::QQPRRegisterClass;
646 else if (VT == MVT::v8i64)
647 return ARM::QQQQPRRegisterClass;
649 return TargetLowering::getRegClassFor(VT);
652 /// getFunctionAlignment - Return the Log2 alignment of this function.
653 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
654 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
657 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
658 unsigned NumVals = N->getNumValues();
660 return Sched::RegPressure;
662 for (unsigned i = 0; i != NumVals; ++i) {
663 EVT VT = N->getValueType(i);
664 if (VT.isFloatingPoint() || VT.isVector())
665 return Sched::Latency;
668 if (!N->isMachineOpcode())
669 return Sched::RegPressure;
671 // Load are scheduled for latency even if there instruction itinerary
673 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
674 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
676 return Sched::Latency;
678 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
679 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
680 return Sched::Latency;
681 return Sched::RegPressure;
684 //===----------------------------------------------------------------------===//
686 //===----------------------------------------------------------------------===//
688 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
689 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
691 default: llvm_unreachable("Unknown condition code!");
692 case ISD::SETNE: return ARMCC::NE;
693 case ISD::SETEQ: return ARMCC::EQ;
694 case ISD::SETGT: return ARMCC::GT;
695 case ISD::SETGE: return ARMCC::GE;
696 case ISD::SETLT: return ARMCC::LT;
697 case ISD::SETLE: return ARMCC::LE;
698 case ISD::SETUGT: return ARMCC::HI;
699 case ISD::SETUGE: return ARMCC::HS;
700 case ISD::SETULT: return ARMCC::LO;
701 case ISD::SETULE: return ARMCC::LS;
705 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
706 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
707 ARMCC::CondCodes &CondCode2) {
708 CondCode2 = ARMCC::AL;
710 default: llvm_unreachable("Unknown FP condition!");
712 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
714 case ISD::SETOGT: CondCode = ARMCC::GT; break;
716 case ISD::SETOGE: CondCode = ARMCC::GE; break;
717 case ISD::SETOLT: CondCode = ARMCC::MI; break;
718 case ISD::SETOLE: CondCode = ARMCC::LS; break;
719 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
720 case ISD::SETO: CondCode = ARMCC::VC; break;
721 case ISD::SETUO: CondCode = ARMCC::VS; break;
722 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
723 case ISD::SETUGT: CondCode = ARMCC::HI; break;
724 case ISD::SETUGE: CondCode = ARMCC::PL; break;
726 case ISD::SETULT: CondCode = ARMCC::LT; break;
728 case ISD::SETULE: CondCode = ARMCC::LE; break;
730 case ISD::SETUNE: CondCode = ARMCC::NE; break;
734 //===----------------------------------------------------------------------===//
735 // Calling Convention Implementation
736 //===----------------------------------------------------------------------===//
738 #include "ARMGenCallingConv.inc"
740 // APCS f64 is in register pairs, possibly split to stack
741 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
742 CCValAssign::LocInfo &LocInfo,
743 CCState &State, bool CanFail) {
744 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
746 // Try to get the first register.
747 if (unsigned Reg = State.AllocateReg(RegList, 4))
748 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
750 // For the 2nd half of a v2f64, do not fail.
754 // Put the whole thing on the stack.
755 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
756 State.AllocateStack(8, 4),
761 // Try to get the second register.
762 if (unsigned Reg = State.AllocateReg(RegList, 4))
763 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
765 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
766 State.AllocateStack(4, 4),
771 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
772 CCValAssign::LocInfo &LocInfo,
773 ISD::ArgFlagsTy &ArgFlags,
775 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
777 if (LocVT == MVT::v2f64 &&
778 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
780 return true; // we handled it
783 // AAPCS f64 is in aligned register pairs
784 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
785 CCValAssign::LocInfo &LocInfo,
786 CCState &State, bool CanFail) {
787 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
788 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
790 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
792 // For the 2nd half of a v2f64, do not just fail.
796 // Put the whole thing on the stack.
797 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
798 State.AllocateStack(8, 8),
804 for (i = 0; i < 2; ++i)
805 if (HiRegList[i] == Reg)
808 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
809 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
814 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
815 CCValAssign::LocInfo &LocInfo,
816 ISD::ArgFlagsTy &ArgFlags,
818 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
820 if (LocVT == MVT::v2f64 &&
821 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
823 return true; // we handled it
826 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
827 CCValAssign::LocInfo &LocInfo, CCState &State) {
828 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
829 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
831 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
833 return false; // we didn't handle it
836 for (i = 0; i < 2; ++i)
837 if (HiRegList[i] == Reg)
840 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
841 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
846 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
847 CCValAssign::LocInfo &LocInfo,
848 ISD::ArgFlagsTy &ArgFlags,
850 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
852 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
854 return true; // we handled it
857 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
858 CCValAssign::LocInfo &LocInfo,
859 ISD::ArgFlagsTy &ArgFlags,
861 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
865 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
866 /// given CallingConvention value.
867 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
869 bool isVarArg) const {
872 llvm_unreachable("Unsupported calling convention");
874 case CallingConv::Fast:
875 // Use target triple & subtarget features to do actual dispatch.
876 if (Subtarget->isAAPCS_ABI()) {
877 if (Subtarget->hasVFP2() &&
878 FloatABIType == FloatABI::Hard && !isVarArg)
879 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
881 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
883 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
884 case CallingConv::ARM_AAPCS_VFP:
885 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
886 case CallingConv::ARM_AAPCS:
887 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
888 case CallingConv::ARM_APCS:
889 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
893 /// LowerCallResult - Lower the result values of a call into the
894 /// appropriate copies out of appropriate physical registers.
896 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
897 CallingConv::ID CallConv, bool isVarArg,
898 const SmallVectorImpl<ISD::InputArg> &Ins,
899 DebugLoc dl, SelectionDAG &DAG,
900 SmallVectorImpl<SDValue> &InVals) const {
902 // Assign locations to each value returned by this call.
903 SmallVector<CCValAssign, 16> RVLocs;
904 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
905 RVLocs, *DAG.getContext());
906 CCInfo.AnalyzeCallResult(Ins,
907 CCAssignFnForNode(CallConv, /* Return*/ true,
910 // Copy all of the result registers out of their specified physreg.
911 for (unsigned i = 0; i != RVLocs.size(); ++i) {
912 CCValAssign VA = RVLocs[i];
915 if (VA.needsCustom()) {
916 // Handle f64 or half of a v2f64.
917 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
919 Chain = Lo.getValue(1);
920 InFlag = Lo.getValue(2);
921 VA = RVLocs[++i]; // skip ahead to next loc
922 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
924 Chain = Hi.getValue(1);
925 InFlag = Hi.getValue(2);
926 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
928 if (VA.getLocVT() == MVT::v2f64) {
929 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
930 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
931 DAG.getConstant(0, MVT::i32));
933 VA = RVLocs[++i]; // skip ahead to next loc
934 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
935 Chain = Lo.getValue(1);
936 InFlag = Lo.getValue(2);
937 VA = RVLocs[++i]; // skip ahead to next loc
938 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
939 Chain = Hi.getValue(1);
940 InFlag = Hi.getValue(2);
941 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
942 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
943 DAG.getConstant(1, MVT::i32));
946 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
948 Chain = Val.getValue(1);
949 InFlag = Val.getValue(2);
952 switch (VA.getLocInfo()) {
953 default: llvm_unreachable("Unknown loc info!");
954 case CCValAssign::Full: break;
955 case CCValAssign::BCvt:
956 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
960 InVals.push_back(Val);
966 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
967 /// by "Src" to address "Dst" of size "Size". Alignment information is
968 /// specified by the specific parameter attribute. The copy will be passed as
969 /// a byval function parameter.
970 /// Sometimes what we are copying is the end of a larger object, the part that
971 /// does not fit in registers.
973 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
974 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
976 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
977 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
978 /*isVolatile=*/false, /*AlwaysInline=*/false,
982 /// LowerMemOpCallTo - Store the argument to the stack.
984 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
985 SDValue StackPtr, SDValue Arg,
986 DebugLoc dl, SelectionDAG &DAG,
987 const CCValAssign &VA,
988 ISD::ArgFlagsTy Flags) const {
989 unsigned LocMemOffset = VA.getLocMemOffset();
990 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
991 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
992 if (Flags.isByVal()) {
993 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
995 return DAG.getStore(Chain, dl, Arg, PtrOff,
996 PseudoSourceValue::getStack(), LocMemOffset,
1000 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1001 SDValue Chain, SDValue &Arg,
1002 RegsToPassVector &RegsToPass,
1003 CCValAssign &VA, CCValAssign &NextVA,
1005 SmallVector<SDValue, 8> &MemOpChains,
1006 ISD::ArgFlagsTy Flags) const {
1008 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1009 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1010 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1012 if (NextVA.isRegLoc())
1013 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1015 assert(NextVA.isMemLoc());
1016 if (StackPtr.getNode() == 0)
1017 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1019 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1025 /// LowerCall - Lowering a call into a callseq_start <-
1026 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1029 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1030 CallingConv::ID CallConv, bool isVarArg,
1032 const SmallVectorImpl<ISD::OutputArg> &Outs,
1033 const SmallVectorImpl<SDValue> &OutVals,
1034 const SmallVectorImpl<ISD::InputArg> &Ins,
1035 DebugLoc dl, SelectionDAG &DAG,
1036 SmallVectorImpl<SDValue> &InVals) const {
1037 MachineFunction &MF = DAG.getMachineFunction();
1038 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1039 bool IsSibCall = false;
1040 // Temporarily disable tail calls so things don't break.
1041 if (!EnableARMTailCalls)
1044 // Check if it's really possible to do a tail call.
1045 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1046 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1047 Outs, OutVals, Ins, DAG);
1048 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1049 // detected sibcalls.
1056 // Analyze operands of the call, assigning locations to each operand.
1057 SmallVector<CCValAssign, 16> ArgLocs;
1058 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1060 CCInfo.AnalyzeCallOperands(Outs,
1061 CCAssignFnForNode(CallConv, /* Return*/ false,
1064 // Get a count of how many bytes are to be pushed on the stack.
1065 unsigned NumBytes = CCInfo.getNextStackOffset();
1067 // For tail calls, memory operands are available in our caller's stack.
1071 // Adjust the stack pointer for the new arguments...
1072 // These operations are automatically eliminated by the prolog/epilog pass
1074 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1076 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1078 RegsToPassVector RegsToPass;
1079 SmallVector<SDValue, 8> MemOpChains;
1081 // Walk the register/memloc assignments, inserting copies/loads. In the case
1082 // of tail call optimization, arguments are handled later.
1083 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1085 ++i, ++realArgIdx) {
1086 CCValAssign &VA = ArgLocs[i];
1087 SDValue Arg = OutVals[realArgIdx];
1088 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1090 // Promote the value if needed.
1091 switch (VA.getLocInfo()) {
1092 default: llvm_unreachable("Unknown loc info!");
1093 case CCValAssign::Full: break;
1094 case CCValAssign::SExt:
1095 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1097 case CCValAssign::ZExt:
1098 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1100 case CCValAssign::AExt:
1101 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1103 case CCValAssign::BCvt:
1104 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1108 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1109 if (VA.needsCustom()) {
1110 if (VA.getLocVT() == MVT::v2f64) {
1111 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1112 DAG.getConstant(0, MVT::i32));
1113 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1114 DAG.getConstant(1, MVT::i32));
1116 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1117 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1119 VA = ArgLocs[++i]; // skip ahead to next loc
1120 if (VA.isRegLoc()) {
1121 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1122 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1124 assert(VA.isMemLoc());
1126 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1127 dl, DAG, VA, Flags));
1130 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1131 StackPtr, MemOpChains, Flags);
1133 } else if (VA.isRegLoc()) {
1134 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1135 } else if (!IsSibCall) {
1136 assert(VA.isMemLoc());
1138 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1139 dl, DAG, VA, Flags));
1143 if (!MemOpChains.empty())
1144 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1145 &MemOpChains[0], MemOpChains.size());
1147 // Build a sequence of copy-to-reg nodes chained together with token chain
1148 // and flag operands which copy the outgoing args into the appropriate regs.
1150 // Tail call byval lowering might overwrite argument registers so in case of
1151 // tail call optimization the copies to registers are lowered later.
1153 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1154 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1155 RegsToPass[i].second, InFlag);
1156 InFlag = Chain.getValue(1);
1159 // For tail calls lower the arguments to the 'real' stack slot.
1161 // Force all the incoming stack arguments to be loaded from the stack
1162 // before any new outgoing arguments are stored to the stack, because the
1163 // outgoing stack slots may alias the incoming argument stack slots, and
1164 // the alias isn't otherwise explicit. This is slightly more conservative
1165 // than necessary, because it means that each store effectively depends
1166 // on every argument instead of just those arguments it would clobber.
1168 // Do not flag preceeding copytoreg stuff together with the following stuff.
1170 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1171 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1172 RegsToPass[i].second, InFlag);
1173 InFlag = Chain.getValue(1);
1178 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1179 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1180 // node so that legalize doesn't hack it.
1181 bool isDirect = false;
1182 bool isARMFunc = false;
1183 bool isLocalARMFunc = false;
1184 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1186 if (EnableARMLongCalls) {
1187 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1188 && "long-calls with non-static relocation model!");
1189 // Handle a global address or an external symbol. If it's not one of
1190 // those, the target's already in a register, so we don't need to do
1192 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1193 const GlobalValue *GV = G->getGlobal();
1194 // Create a constant pool entry for the callee address
1195 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1196 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1199 // Get the address of the callee into a register
1200 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1201 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1202 Callee = DAG.getLoad(getPointerTy(), dl,
1203 DAG.getEntryNode(), CPAddr,
1204 PseudoSourceValue::getConstantPool(), 0,
1206 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1207 const char *Sym = S->getSymbol();
1209 // Create a constant pool entry for the callee address
1210 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1211 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1212 Sym, ARMPCLabelIndex, 0);
1213 // Get the address of the callee into a register
1214 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1215 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1216 Callee = DAG.getLoad(getPointerTy(), dl,
1217 DAG.getEntryNode(), CPAddr,
1218 PseudoSourceValue::getConstantPool(), 0,
1221 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1222 const GlobalValue *GV = G->getGlobal();
1224 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1225 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1226 getTargetMachine().getRelocationModel() != Reloc::Static;
1227 isARMFunc = !Subtarget->isThumb() || isStub;
1228 // ARM call to a local ARM function is predicable.
1229 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1230 // tBX takes a register source operand.
1231 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1232 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1233 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1236 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1237 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1238 Callee = DAG.getLoad(getPointerTy(), dl,
1239 DAG.getEntryNode(), CPAddr,
1240 PseudoSourceValue::getConstantPool(), 0,
1242 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1243 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1244 getPointerTy(), Callee, PICLabel);
1246 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1247 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1249 bool isStub = Subtarget->isTargetDarwin() &&
1250 getTargetMachine().getRelocationModel() != Reloc::Static;
1251 isARMFunc = !Subtarget->isThumb() || isStub;
1252 // tBX takes a register source operand.
1253 const char *Sym = S->getSymbol();
1254 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1255 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1256 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1257 Sym, ARMPCLabelIndex, 4);
1258 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1259 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1260 Callee = DAG.getLoad(getPointerTy(), dl,
1261 DAG.getEntryNode(), CPAddr,
1262 PseudoSourceValue::getConstantPool(), 0,
1264 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1265 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1266 getPointerTy(), Callee, PICLabel);
1268 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1271 // FIXME: handle tail calls differently.
1273 if (Subtarget->isThumb()) {
1274 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1275 CallOpc = ARMISD::CALL_NOLINK;
1277 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1279 CallOpc = (isDirect || Subtarget->hasV5TOps())
1280 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1281 : ARMISD::CALL_NOLINK;
1283 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1284 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1285 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1286 InFlag = Chain.getValue(1);
1289 std::vector<SDValue> Ops;
1290 Ops.push_back(Chain);
1291 Ops.push_back(Callee);
1293 // Add argument registers to the end of the list so that they are known live
1295 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1296 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1297 RegsToPass[i].second.getValueType()));
1299 if (InFlag.getNode())
1300 Ops.push_back(InFlag);
1302 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1304 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1306 // Returns a chain and a flag for retval copy to use.
1307 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1308 InFlag = Chain.getValue(1);
1310 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1311 DAG.getIntPtrConstant(0, true), InFlag);
1313 InFlag = Chain.getValue(1);
1315 // Handle result values, copying them out of physregs into vregs that we
1317 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1321 /// MatchingStackOffset - Return true if the given stack call argument is
1322 /// already available in the same position (relatively) of the caller's
1323 /// incoming argument stack.
1325 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1326 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1327 const ARMInstrInfo *TII) {
1328 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1330 if (Arg.getOpcode() == ISD::CopyFromReg) {
1331 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1332 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1334 MachineInstr *Def = MRI->getVRegDef(VR);
1337 if (!Flags.isByVal()) {
1338 if (!TII->isLoadFromStackSlot(Def, FI))
1343 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1344 if (Flags.isByVal())
1345 // ByVal argument is passed in as a pointer but it's now being
1346 // dereferenced. e.g.
1347 // define @foo(%struct.X* %A) {
1348 // tail call @bar(%struct.X* byval %A)
1351 SDValue Ptr = Ld->getBasePtr();
1352 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1355 FI = FINode->getIndex();
1359 assert(FI != INT_MAX);
1360 if (!MFI->isFixedObjectIndex(FI))
1362 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1365 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1366 /// for tail call optimization. Targets which want to do tail call
1367 /// optimization should implement this function.
1369 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1370 CallingConv::ID CalleeCC,
1372 bool isCalleeStructRet,
1373 bool isCallerStructRet,
1374 const SmallVectorImpl<ISD::OutputArg> &Outs,
1375 const SmallVectorImpl<SDValue> &OutVals,
1376 const SmallVectorImpl<ISD::InputArg> &Ins,
1377 SelectionDAG& DAG) const {
1378 const Function *CallerF = DAG.getMachineFunction().getFunction();
1379 CallingConv::ID CallerCC = CallerF->getCallingConv();
1380 bool CCMatch = CallerCC == CalleeCC;
1382 // Look for obvious safe cases to perform tail call optimization that do not
1383 // require ABI changes. This is what gcc calls sibcall.
1385 // Do not sibcall optimize vararg calls unless the call site is not passing
1387 if (isVarArg && !Outs.empty())
1390 // Also avoid sibcall optimization if either caller or callee uses struct
1391 // return semantics.
1392 if (isCalleeStructRet || isCallerStructRet)
1395 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1396 // emitEpilogue is not ready for them.
1397 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1398 // LR. This means if we need to reload LR, it takes an extra instructions,
1399 // which outweighs the value of the tail call; but here we don't know yet
1400 // whether LR is going to be used. Probably the right approach is to
1401 // generate the tail call here and turn it back into CALL/RET in
1402 // emitEpilogue if LR is used.
1403 if (Subtarget->isThumb1Only())
1406 // For the moment, we can only do this to functions defined in this
1407 // compilation, or to indirect calls. A Thumb B to an ARM function,
1408 // or vice versa, is not easily fixed up in the linker unlike BL.
1409 // (We could do this by loading the address of the callee into a register;
1410 // that is an extra instruction over the direct call and burns a register
1411 // as well, so is not likely to be a win.)
1413 // It might be safe to remove this restriction on non-Darwin.
1415 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1416 // but we need to make sure there are enough registers; the only valid
1417 // registers are the 4 used for parameters. We don't currently do this
1419 if (isa<ExternalSymbolSDNode>(Callee))
1422 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1423 const GlobalValue *GV = G->getGlobal();
1424 if (GV->isDeclaration() || GV->isWeakForLinker())
1428 // If the calling conventions do not match, then we'd better make sure the
1429 // results are returned in the same way as what the caller expects.
1431 SmallVector<CCValAssign, 16> RVLocs1;
1432 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1433 RVLocs1, *DAG.getContext());
1434 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1436 SmallVector<CCValAssign, 16> RVLocs2;
1437 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1438 RVLocs2, *DAG.getContext());
1439 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1441 if (RVLocs1.size() != RVLocs2.size())
1443 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1444 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1446 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1448 if (RVLocs1[i].isRegLoc()) {
1449 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1452 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1458 // If the callee takes no arguments then go on to check the results of the
1460 if (!Outs.empty()) {
1461 // Check if stack adjustment is needed. For now, do not do this if any
1462 // argument is passed on the stack.
1463 SmallVector<CCValAssign, 16> ArgLocs;
1464 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1465 ArgLocs, *DAG.getContext());
1466 CCInfo.AnalyzeCallOperands(Outs,
1467 CCAssignFnForNode(CalleeCC, false, isVarArg));
1468 if (CCInfo.getNextStackOffset()) {
1469 MachineFunction &MF = DAG.getMachineFunction();
1471 // Check if the arguments are already laid out in the right way as
1472 // the caller's fixed stack objects.
1473 MachineFrameInfo *MFI = MF.getFrameInfo();
1474 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1475 const ARMInstrInfo *TII =
1476 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1477 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1479 ++i, ++realArgIdx) {
1480 CCValAssign &VA = ArgLocs[i];
1481 EVT RegVT = VA.getLocVT();
1482 SDValue Arg = OutVals[realArgIdx];
1483 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1484 if (VA.getLocInfo() == CCValAssign::Indirect)
1486 if (VA.needsCustom()) {
1487 // f64 and vector types are split into multiple registers or
1488 // register/stack-slot combinations. The types will not match
1489 // the registers; give up on memory f64 refs until we figure
1490 // out what to do about this.
1493 if (!ArgLocs[++i].isRegLoc())
1495 if (RegVT == MVT::v2f64) {
1496 if (!ArgLocs[++i].isRegLoc())
1498 if (!ArgLocs[++i].isRegLoc())
1501 } else if (!VA.isRegLoc()) {
1502 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1514 ARMTargetLowering::LowerReturn(SDValue Chain,
1515 CallingConv::ID CallConv, bool isVarArg,
1516 const SmallVectorImpl<ISD::OutputArg> &Outs,
1517 const SmallVectorImpl<SDValue> &OutVals,
1518 DebugLoc dl, SelectionDAG &DAG) const {
1520 // CCValAssign - represent the assignment of the return value to a location.
1521 SmallVector<CCValAssign, 16> RVLocs;
1523 // CCState - Info about the registers and stack slots.
1524 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1527 // Analyze outgoing return values.
1528 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1531 // If this is the first return lowered for this function, add
1532 // the regs to the liveout set for the function.
1533 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1534 for (unsigned i = 0; i != RVLocs.size(); ++i)
1535 if (RVLocs[i].isRegLoc())
1536 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1541 // Copy the result values into the output registers.
1542 for (unsigned i = 0, realRVLocIdx = 0;
1544 ++i, ++realRVLocIdx) {
1545 CCValAssign &VA = RVLocs[i];
1546 assert(VA.isRegLoc() && "Can only return in registers!");
1548 SDValue Arg = OutVals[realRVLocIdx];
1550 switch (VA.getLocInfo()) {
1551 default: llvm_unreachable("Unknown loc info!");
1552 case CCValAssign::Full: break;
1553 case CCValAssign::BCvt:
1554 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1558 if (VA.needsCustom()) {
1559 if (VA.getLocVT() == MVT::v2f64) {
1560 // Extract the first half and return it in two registers.
1561 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1562 DAG.getConstant(0, MVT::i32));
1563 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1564 DAG.getVTList(MVT::i32, MVT::i32), Half);
1566 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1567 Flag = Chain.getValue(1);
1568 VA = RVLocs[++i]; // skip ahead to next loc
1569 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1570 HalfGPRs.getValue(1), Flag);
1571 Flag = Chain.getValue(1);
1572 VA = RVLocs[++i]; // skip ahead to next loc
1574 // Extract the 2nd half and fall through to handle it as an f64 value.
1575 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1576 DAG.getConstant(1, MVT::i32));
1578 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1580 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1581 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1582 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1583 Flag = Chain.getValue(1);
1584 VA = RVLocs[++i]; // skip ahead to next loc
1585 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1588 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1590 // Guarantee that all emitted copies are
1591 // stuck together, avoiding something bad.
1592 Flag = Chain.getValue(1);
1597 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1599 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1604 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1605 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1606 // one of the above mentioned nodes. It has to be wrapped because otherwise
1607 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1608 // be used to form addressing mode. These wrapped nodes will be selected
1610 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1611 EVT PtrVT = Op.getValueType();
1612 // FIXME there is no actual debug info here
1613 DebugLoc dl = Op.getDebugLoc();
1614 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1616 if (CP->isMachineConstantPoolEntry())
1617 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1618 CP->getAlignment());
1620 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1621 CP->getAlignment());
1622 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1625 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1626 SelectionDAG &DAG) const {
1627 MachineFunction &MF = DAG.getMachineFunction();
1628 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1629 unsigned ARMPCLabelIndex = 0;
1630 DebugLoc DL = Op.getDebugLoc();
1631 EVT PtrVT = getPointerTy();
1632 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1633 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1635 if (RelocM == Reloc::Static) {
1636 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1638 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1639 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1640 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1641 ARMCP::CPBlockAddress,
1643 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1645 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1646 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1647 PseudoSourceValue::getConstantPool(), 0,
1649 if (RelocM == Reloc::Static)
1651 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1652 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1655 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1657 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1658 SelectionDAG &DAG) const {
1659 DebugLoc dl = GA->getDebugLoc();
1660 EVT PtrVT = getPointerTy();
1661 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1662 MachineFunction &MF = DAG.getMachineFunction();
1663 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1664 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1665 ARMConstantPoolValue *CPV =
1666 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1667 ARMCP::CPValue, PCAdj, "tlsgd", true);
1668 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1669 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1670 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1671 PseudoSourceValue::getConstantPool(), 0,
1673 SDValue Chain = Argument.getValue(1);
1675 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1676 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1678 // call __tls_get_addr.
1681 Entry.Node = Argument;
1682 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1683 Args.push_back(Entry);
1684 // FIXME: is there useful debug info available here?
1685 std::pair<SDValue, SDValue> CallResult =
1686 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1687 false, false, false, false,
1688 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1689 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1690 return CallResult.first;
1693 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1694 // "local exec" model.
1696 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1697 SelectionDAG &DAG) const {
1698 const GlobalValue *GV = GA->getGlobal();
1699 DebugLoc dl = GA->getDebugLoc();
1701 SDValue Chain = DAG.getEntryNode();
1702 EVT PtrVT = getPointerTy();
1703 // Get the Thread Pointer
1704 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1706 if (GV->isDeclaration()) {
1707 MachineFunction &MF = DAG.getMachineFunction();
1708 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1709 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1710 // Initial exec model.
1711 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1712 ARMConstantPoolValue *CPV =
1713 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1714 ARMCP::CPValue, PCAdj, "gottpoff", true);
1715 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1716 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1717 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1718 PseudoSourceValue::getConstantPool(), 0,
1720 Chain = Offset.getValue(1);
1722 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1723 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1725 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1726 PseudoSourceValue::getConstantPool(), 0,
1730 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1731 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1732 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1733 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1734 PseudoSourceValue::getConstantPool(), 0,
1738 // The address of the thread local variable is the add of the thread
1739 // pointer with the offset of the variable.
1740 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1744 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1745 // TODO: implement the "local dynamic" model
1746 assert(Subtarget->isTargetELF() &&
1747 "TLS not implemented for non-ELF targets");
1748 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1749 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1750 // otherwise use the "Local Exec" TLS Model
1751 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1752 return LowerToTLSGeneralDynamicModel(GA, DAG);
1754 return LowerToTLSExecModels(GA, DAG);
1757 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1758 SelectionDAG &DAG) const {
1759 EVT PtrVT = getPointerTy();
1760 DebugLoc dl = Op.getDebugLoc();
1761 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1762 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1763 if (RelocM == Reloc::PIC_) {
1764 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1765 ARMConstantPoolValue *CPV =
1766 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1767 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1768 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1769 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1771 PseudoSourceValue::getConstantPool(), 0,
1773 SDValue Chain = Result.getValue(1);
1774 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1775 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1777 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1778 PseudoSourceValue::getGOT(), 0,
1782 // If we have T2 ops, we can materialize the address directly via movt/movw
1783 // pair. This is always cheaper.
1784 if (Subtarget->useMovt()) {
1785 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1786 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1788 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1789 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1790 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1791 PseudoSourceValue::getConstantPool(), 0,
1797 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1798 SelectionDAG &DAG) const {
1799 MachineFunction &MF = DAG.getMachineFunction();
1800 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1801 unsigned ARMPCLabelIndex = 0;
1802 EVT PtrVT = getPointerTy();
1803 DebugLoc dl = Op.getDebugLoc();
1804 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1805 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1807 if (RelocM == Reloc::Static)
1808 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1810 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1811 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1812 ARMConstantPoolValue *CPV =
1813 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1814 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1816 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1818 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1819 PseudoSourceValue::getConstantPool(), 0,
1821 SDValue Chain = Result.getValue(1);
1823 if (RelocM == Reloc::PIC_) {
1824 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1825 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1828 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1829 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1830 PseudoSourceValue::getGOT(), 0,
1836 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1837 SelectionDAG &DAG) const {
1838 assert(Subtarget->isTargetELF() &&
1839 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1840 MachineFunction &MF = DAG.getMachineFunction();
1841 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1842 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1843 EVT PtrVT = getPointerTy();
1844 DebugLoc dl = Op.getDebugLoc();
1845 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1846 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1847 "_GLOBAL_OFFSET_TABLE_",
1848 ARMPCLabelIndex, PCAdj);
1849 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1850 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1851 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1852 PseudoSourceValue::getConstantPool(), 0,
1854 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1855 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1859 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1860 DebugLoc dl = Op.getDebugLoc();
1861 SDValue Val = DAG.getConstant(0, MVT::i32);
1862 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1863 Op.getOperand(1), Val);
1867 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1868 DebugLoc dl = Op.getDebugLoc();
1869 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1870 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1874 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1875 const ARMSubtarget *Subtarget) const {
1876 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1877 DebugLoc dl = Op.getDebugLoc();
1879 default: return SDValue(); // Don't custom lower most intrinsics.
1880 case Intrinsic::arm_thread_pointer: {
1881 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1882 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1884 case Intrinsic::eh_sjlj_lsda: {
1885 MachineFunction &MF = DAG.getMachineFunction();
1886 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1887 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1888 EVT PtrVT = getPointerTy();
1889 DebugLoc dl = Op.getDebugLoc();
1890 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1892 unsigned PCAdj = (RelocM != Reloc::PIC_)
1893 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1894 ARMConstantPoolValue *CPV =
1895 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1896 ARMCP::CPLSDA, PCAdj);
1897 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1898 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1900 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1901 PseudoSourceValue::getConstantPool(), 0,
1904 if (RelocM == Reloc::PIC_) {
1905 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1906 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1913 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1914 const ARMSubtarget *Subtarget) {
1915 DebugLoc dl = Op.getDebugLoc();
1916 SDValue Op5 = Op.getOperand(5);
1917 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1918 // v6 and v7 can both handle barriers directly, but need handled a bit
1919 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1921 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1922 if (Subtarget->hasV7Ops())
1923 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1924 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1925 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1926 DAG.getConstant(0, MVT::i32));
1927 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1931 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1932 MachineFunction &MF = DAG.getMachineFunction();
1933 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1935 // vastart just stores the address of the VarArgsFrameIndex slot into the
1936 // memory location argument.
1937 DebugLoc dl = Op.getDebugLoc();
1938 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1939 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1940 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1941 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1946 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1947 SelectionDAG &DAG) const {
1948 SDNode *Node = Op.getNode();
1949 DebugLoc dl = Node->getDebugLoc();
1950 EVT VT = Node->getValueType(0);
1951 SDValue Chain = Op.getOperand(0);
1952 SDValue Size = Op.getOperand(1);
1953 SDValue Align = Op.getOperand(2);
1955 // Chain the dynamic stack allocation so that it doesn't modify the stack
1956 // pointer when other instructions are using the stack.
1957 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1959 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1960 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1961 if (AlignVal > StackAlign)
1962 // Do this now since selection pass cannot introduce new target
1963 // independent node.
1964 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1966 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1967 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1968 // do even more horrible hack later.
1969 MachineFunction &MF = DAG.getMachineFunction();
1970 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1971 if (AFI->isThumb1OnlyFunction()) {
1973 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1975 uint32_t Val = C->getZExtValue();
1976 if (Val <= 508 && ((Val & 3) == 0))
1980 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1983 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1984 SDValue Ops1[] = { Chain, Size, Align };
1985 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1986 Chain = Res.getValue(1);
1987 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1988 DAG.getIntPtrConstant(0, true), SDValue());
1989 SDValue Ops2[] = { Res, Chain };
1990 return DAG.getMergeValues(Ops2, 2, dl);
1994 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1995 SDValue &Root, SelectionDAG &DAG,
1996 DebugLoc dl) const {
1997 MachineFunction &MF = DAG.getMachineFunction();
1998 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2000 TargetRegisterClass *RC;
2001 if (AFI->isThumb1OnlyFunction())
2002 RC = ARM::tGPRRegisterClass;
2004 RC = ARM::GPRRegisterClass;
2006 // Transform the arguments stored in physical registers into virtual ones.
2007 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2008 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2011 if (NextVA.isMemLoc()) {
2012 MachineFrameInfo *MFI = MF.getFrameInfo();
2013 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2015 // Create load node to retrieve arguments from the stack.
2016 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2017 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2018 PseudoSourceValue::getFixedStack(FI), 0,
2021 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2022 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2025 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2029 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2030 CallingConv::ID CallConv, bool isVarArg,
2031 const SmallVectorImpl<ISD::InputArg>
2033 DebugLoc dl, SelectionDAG &DAG,
2034 SmallVectorImpl<SDValue> &InVals)
2037 MachineFunction &MF = DAG.getMachineFunction();
2038 MachineFrameInfo *MFI = MF.getFrameInfo();
2040 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2042 // Assign locations to all of the incoming arguments.
2043 SmallVector<CCValAssign, 16> ArgLocs;
2044 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2046 CCInfo.AnalyzeFormalArguments(Ins,
2047 CCAssignFnForNode(CallConv, /* Return*/ false,
2050 SmallVector<SDValue, 16> ArgValues;
2052 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2053 CCValAssign &VA = ArgLocs[i];
2055 // Arguments stored in registers.
2056 if (VA.isRegLoc()) {
2057 EVT RegVT = VA.getLocVT();
2060 if (VA.needsCustom()) {
2061 // f64 and vector types are split up into multiple registers or
2062 // combinations of registers and stack slots.
2063 if (VA.getLocVT() == MVT::v2f64) {
2064 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2066 VA = ArgLocs[++i]; // skip ahead to next loc
2068 if (VA.isMemLoc()) {
2069 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2070 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2071 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2072 PseudoSourceValue::getFixedStack(FI), 0,
2075 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2078 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2079 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2080 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2081 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2082 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2084 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2087 TargetRegisterClass *RC;
2089 if (RegVT == MVT::f32)
2090 RC = ARM::SPRRegisterClass;
2091 else if (RegVT == MVT::f64)
2092 RC = ARM::DPRRegisterClass;
2093 else if (RegVT == MVT::v2f64)
2094 RC = ARM::QPRRegisterClass;
2095 else if (RegVT == MVT::i32)
2096 RC = (AFI->isThumb1OnlyFunction() ?
2097 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2099 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2101 // Transform the arguments in physical registers into virtual ones.
2102 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2103 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2106 // If this is an 8 or 16-bit value, it is really passed promoted
2107 // to 32 bits. Insert an assert[sz]ext to capture this, then
2108 // truncate to the right size.
2109 switch (VA.getLocInfo()) {
2110 default: llvm_unreachable("Unknown loc info!");
2111 case CCValAssign::Full: break;
2112 case CCValAssign::BCvt:
2113 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2115 case CCValAssign::SExt:
2116 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2117 DAG.getValueType(VA.getValVT()));
2118 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2120 case CCValAssign::ZExt:
2121 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2122 DAG.getValueType(VA.getValVT()));
2123 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2127 InVals.push_back(ArgValue);
2129 } else { // VA.isRegLoc()
2132 assert(VA.isMemLoc());
2133 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2135 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2136 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2138 // Create load nodes to retrieve arguments from the stack.
2139 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2140 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2141 PseudoSourceValue::getFixedStack(FI), 0,
2148 static const unsigned GPRArgRegs[] = {
2149 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2152 unsigned NumGPRs = CCInfo.getFirstUnallocated
2153 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2155 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2156 unsigned VARegSize = (4 - NumGPRs) * 4;
2157 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2158 unsigned ArgOffset = CCInfo.getNextStackOffset();
2159 if (VARegSaveSize) {
2160 // If this function is vararg, store any remaining integer argument regs
2161 // to their spots on the stack so that they may be loaded by deferencing
2162 // the result of va_next.
2163 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2164 AFI->setVarArgsFrameIndex(
2165 MFI->CreateFixedObject(VARegSaveSize,
2166 ArgOffset + VARegSaveSize - VARegSize,
2168 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2171 SmallVector<SDValue, 4> MemOps;
2172 for (; NumGPRs < 4; ++NumGPRs) {
2173 TargetRegisterClass *RC;
2174 if (AFI->isThumb1OnlyFunction())
2175 RC = ARM::tGPRRegisterClass;
2177 RC = ARM::GPRRegisterClass;
2179 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2180 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2182 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2183 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2184 0, false, false, 0);
2185 MemOps.push_back(Store);
2186 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2187 DAG.getConstant(4, getPointerTy()));
2189 if (!MemOps.empty())
2190 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2191 &MemOps[0], MemOps.size());
2193 // This will point to the next argument passed via stack.
2194 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2200 /// isFloatingPointZero - Return true if this is +0.0.
2201 static bool isFloatingPointZero(SDValue Op) {
2202 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2203 return CFP->getValueAPF().isPosZero();
2204 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2205 // Maybe this has already been legalized into the constant pool?
2206 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2207 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2208 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2209 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2210 return CFP->getValueAPF().isPosZero();
2216 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2217 /// the given operands.
2219 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2220 SDValue &ARMCC, SelectionDAG &DAG,
2221 DebugLoc dl) const {
2222 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2223 unsigned C = RHSC->getZExtValue();
2224 if (!isLegalICmpImmediate(C)) {
2225 // Constant does not fit, try adjusting it by one?
2230 if (isLegalICmpImmediate(C-1)) {
2231 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2232 RHS = DAG.getConstant(C-1, MVT::i32);
2237 if (C > 0 && isLegalICmpImmediate(C-1)) {
2238 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2239 RHS = DAG.getConstant(C-1, MVT::i32);
2244 if (isLegalICmpImmediate(C+1)) {
2245 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2246 RHS = DAG.getConstant(C+1, MVT::i32);
2251 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2252 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2253 RHS = DAG.getConstant(C+1, MVT::i32);
2260 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2261 ARMISD::NodeType CompareType;
2264 CompareType = ARMISD::CMP;
2269 CompareType = ARMISD::CMPZ;
2272 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2273 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2276 static bool canBitcastToInt(SDNode *Op) {
2277 return Op->hasOneUse() &&
2278 ISD::isNormalLoad(Op) &&
2279 Op->getValueType(0) == MVT::f32;
2282 static SDValue bitcastToInt(SDValue Op, SelectionDAG &DAG) {
2283 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2284 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2285 Ld->getChain(), Ld->getBasePtr(),
2286 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2287 Ld->isVolatile(), Ld->isNonTemporal(),
2288 Ld->getAlignment());
2290 llvm_unreachable("Unknown VFP cmp argument!");
2293 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2295 ARMTargetLowering::getVFPCmp(SDValue &LHS, SDValue &RHS, ISD::CondCode CC,
2296 SDValue &ARMCC, SelectionDAG &DAG,
2297 DebugLoc dl) const {
2298 if (UnsafeFPMath && FiniteOnlyFPMath() &&
2299 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2300 CC == ISD::SETNE || CC == ISD::SETUNE) &&
2301 canBitcastToInt(LHS.getNode()) && canBitcastToInt(RHS.getNode())) {
2302 // If unsafe fp math optimization is enabled and there are no othter uses of
2303 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2304 // to an integer comparison.
2305 if (CC == ISD::SETOEQ)
2307 else if (CC == ISD::SETUNE)
2309 LHS = bitcastToInt(LHS, DAG);
2310 RHS = bitcastToInt(RHS, DAG);
2311 return getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2315 if (!isFloatingPointZero(RHS))
2316 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2318 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2319 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2322 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2323 EVT VT = Op.getValueType();
2324 SDValue LHS = Op.getOperand(0);
2325 SDValue RHS = Op.getOperand(1);
2326 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2327 SDValue TrueVal = Op.getOperand(2);
2328 SDValue FalseVal = Op.getOperand(3);
2329 DebugLoc dl = Op.getDebugLoc();
2331 if (LHS.getValueType() == MVT::i32) {
2333 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2334 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2335 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2338 ARMCC::CondCodes CondCode, CondCode2;
2339 FPCCToARMCC(CC, CondCode, CondCode2);
2341 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2342 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2343 SDValue Cmp = getVFPCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2344 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2346 if (CondCode2 != ARMCC::AL) {
2347 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2348 // FIXME: Needs another CMP because flag can have but one use.
2349 SDValue Cmp2 = getVFPCmp(LHS, RHS, CC, ARMCC2, DAG, dl);
2350 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2351 Result, TrueVal, ARMCC2, CCR, Cmp2);
2356 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2357 SDValue Chain = Op.getOperand(0);
2358 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2359 SDValue LHS = Op.getOperand(2);
2360 SDValue RHS = Op.getOperand(3);
2361 SDValue Dest = Op.getOperand(4);
2362 DebugLoc dl = Op.getDebugLoc();
2364 if (LHS.getValueType() == MVT::i32) {
2366 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2367 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2368 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2369 Chain, Dest, ARMCC, CCR,Cmp);
2372 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2373 ARMCC::CondCodes CondCode, CondCode2;
2374 FPCCToARMCC(CC, CondCode, CondCode2);
2376 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2377 SDValue Cmp = getVFPCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2378 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2379 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2380 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2381 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2382 if (CondCode2 != ARMCC::AL) {
2383 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2384 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2385 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2390 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2391 SDValue Chain = Op.getOperand(0);
2392 SDValue Table = Op.getOperand(1);
2393 SDValue Index = Op.getOperand(2);
2394 DebugLoc dl = Op.getDebugLoc();
2396 EVT PTy = getPointerTy();
2397 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2398 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2399 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2400 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2401 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2402 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2403 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2404 if (Subtarget->isThumb2()) {
2405 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2406 // which does another jump to the destination. This also makes it easier
2407 // to translate it to TBB / TBH later.
2408 // FIXME: This might not work if the function is extremely large.
2409 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2410 Addr, Op.getOperand(2), JTI, UId);
2412 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2413 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2414 PseudoSourceValue::getJumpTable(), 0,
2416 Chain = Addr.getValue(1);
2417 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2418 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2420 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2421 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2422 Chain = Addr.getValue(1);
2423 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2427 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2428 DebugLoc dl = Op.getDebugLoc();
2431 switch (Op.getOpcode()) {
2433 assert(0 && "Invalid opcode!");
2434 case ISD::FP_TO_SINT:
2435 Opc = ARMISD::FTOSI;
2437 case ISD::FP_TO_UINT:
2438 Opc = ARMISD::FTOUI;
2441 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2442 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2445 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2446 EVT VT = Op.getValueType();
2447 DebugLoc dl = Op.getDebugLoc();
2450 switch (Op.getOpcode()) {
2452 assert(0 && "Invalid opcode!");
2453 case ISD::SINT_TO_FP:
2454 Opc = ARMISD::SITOF;
2456 case ISD::UINT_TO_FP:
2457 Opc = ARMISD::UITOF;
2461 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2462 return DAG.getNode(Opc, dl, VT, Op);
2465 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2466 // Implement fcopysign with a fabs and a conditional fneg.
2467 SDValue Tmp0 = Op.getOperand(0);
2468 SDValue Tmp1 = Op.getOperand(1);
2469 DebugLoc dl = Op.getDebugLoc();
2470 EVT VT = Op.getValueType();
2471 EVT SrcVT = Tmp1.getValueType();
2472 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2473 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2474 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2475 SDValue Cmp = getVFPCmp(Tmp1, FP0,
2476 ISD::SETLT, ARMCC, DAG, dl);
2477 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2478 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2481 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2482 MachineFunction &MF = DAG.getMachineFunction();
2483 MachineFrameInfo *MFI = MF.getFrameInfo();
2484 MFI->setReturnAddressIsTaken(true);
2486 EVT VT = Op.getValueType();
2487 DebugLoc dl = Op.getDebugLoc();
2488 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2490 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2491 SDValue Offset = DAG.getConstant(4, MVT::i32);
2492 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2493 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2494 NULL, 0, false, false, 0);
2497 // Return LR, which contains the return address. Mark it an implicit live-in.
2498 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2499 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2502 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2503 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2504 MFI->setFrameAddressIsTaken(true);
2506 EVT VT = Op.getValueType();
2507 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2508 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2509 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2510 ? ARM::R7 : ARM::R11;
2511 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2513 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2518 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2519 /// expand a bit convert where either the source or destination type is i64 to
2520 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2521 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2522 /// vectors), since the legalizer won't know what to do with that.
2523 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2524 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2525 DebugLoc dl = N->getDebugLoc();
2526 SDValue Op = N->getOperand(0);
2528 // This function is only supposed to be called for i64 types, either as the
2529 // source or destination of the bit convert.
2530 EVT SrcVT = Op.getValueType();
2531 EVT DstVT = N->getValueType(0);
2532 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2533 "ExpandBIT_CONVERT called for non-i64 type");
2535 // Turn i64->f64 into VMOVDRR.
2536 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2537 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2538 DAG.getConstant(0, MVT::i32));
2539 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2540 DAG.getConstant(1, MVT::i32));
2541 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2542 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2545 // Turn f64->i64 into VMOVRRD.
2546 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2547 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2548 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2549 // Merge the pieces into a single i64 value.
2550 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2556 /// getZeroVector - Returns a vector of specified type with all zero elements.
2558 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2559 assert(VT.isVector() && "Expected a vector type");
2561 // Zero vectors are used to represent vector negation and in those cases
2562 // will be implemented with the NEON VNEG instruction. However, VNEG does
2563 // not support i64 elements, so sometimes the zero vectors will need to be
2564 // explicitly constructed. For those cases, and potentially other uses in
2565 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2566 // to their dest type. This ensures they get CSE'd.
2568 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2569 SmallVector<SDValue, 8> Ops;
2572 if (VT.getSizeInBits() == 64) {
2573 Ops.assign(8, Cst); TVT = MVT::v8i8;
2575 Ops.assign(16, Cst); TVT = MVT::v16i8;
2577 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2579 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2582 /// getOnesVector - Returns a vector of specified type with all bits set.
2584 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2585 assert(VT.isVector() && "Expected a vector type");
2587 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2588 // dest type. This ensures they get CSE'd.
2590 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2591 SmallVector<SDValue, 8> Ops;
2594 if (VT.getSizeInBits() == 64) {
2595 Ops.assign(8, Cst); TVT = MVT::v8i8;
2597 Ops.assign(16, Cst); TVT = MVT::v16i8;
2599 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2601 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2604 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2605 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2606 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2607 SelectionDAG &DAG) const {
2608 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2609 EVT VT = Op.getValueType();
2610 unsigned VTBits = VT.getSizeInBits();
2611 DebugLoc dl = Op.getDebugLoc();
2612 SDValue ShOpLo = Op.getOperand(0);
2613 SDValue ShOpHi = Op.getOperand(1);
2614 SDValue ShAmt = Op.getOperand(2);
2616 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2618 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2620 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2621 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2622 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2623 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2624 DAG.getConstant(VTBits, MVT::i32));
2625 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2626 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2627 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2629 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2630 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2632 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2633 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2636 SDValue Ops[2] = { Lo, Hi };
2637 return DAG.getMergeValues(Ops, 2, dl);
2640 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2641 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2642 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2643 SelectionDAG &DAG) const {
2644 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2645 EVT VT = Op.getValueType();
2646 unsigned VTBits = VT.getSizeInBits();
2647 DebugLoc dl = Op.getDebugLoc();
2648 SDValue ShOpLo = Op.getOperand(0);
2649 SDValue ShOpHi = Op.getOperand(1);
2650 SDValue ShAmt = Op.getOperand(2);
2653 assert(Op.getOpcode() == ISD::SHL_PARTS);
2654 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2655 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2656 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2657 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2658 DAG.getConstant(VTBits, MVT::i32));
2659 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2660 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2662 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2663 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2664 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2666 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2667 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2670 SDValue Ops[2] = { Lo, Hi };
2671 return DAG.getMergeValues(Ops, 2, dl);
2674 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2675 const ARMSubtarget *ST) {
2676 EVT VT = N->getValueType(0);
2677 DebugLoc dl = N->getDebugLoc();
2679 if (!ST->hasV6T2Ops())
2682 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2683 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2686 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2687 const ARMSubtarget *ST) {
2688 EVT VT = N->getValueType(0);
2689 DebugLoc dl = N->getDebugLoc();
2691 // Lower vector shifts on NEON to use VSHL.
2692 if (VT.isVector()) {
2693 assert(ST->hasNEON() && "unexpected vector shift");
2695 // Left shifts translate directly to the vshiftu intrinsic.
2696 if (N->getOpcode() == ISD::SHL)
2697 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2698 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2699 N->getOperand(0), N->getOperand(1));
2701 assert((N->getOpcode() == ISD::SRA ||
2702 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2704 // NEON uses the same intrinsics for both left and right shifts. For
2705 // right shifts, the shift amounts are negative, so negate the vector of
2707 EVT ShiftVT = N->getOperand(1).getValueType();
2708 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2709 getZeroVector(ShiftVT, DAG, dl),
2711 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2712 Intrinsic::arm_neon_vshifts :
2713 Intrinsic::arm_neon_vshiftu);
2714 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2715 DAG.getConstant(vshiftInt, MVT::i32),
2716 N->getOperand(0), NegatedCount);
2719 // We can get here for a node like i32 = ISD::SHL i32, i64
2723 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2724 "Unknown shift to lower!");
2726 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2727 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2728 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2731 // If we are in thumb mode, we don't have RRX.
2732 if (ST->isThumb1Only()) return SDValue();
2734 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2735 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2736 DAG.getConstant(0, MVT::i32));
2737 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2738 DAG.getConstant(1, MVT::i32));
2740 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2741 // captures the result into a carry flag.
2742 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2743 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2745 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2746 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2748 // Merge the pieces into a single i64 value.
2749 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2752 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2753 SDValue TmpOp0, TmpOp1;
2754 bool Invert = false;
2758 SDValue Op0 = Op.getOperand(0);
2759 SDValue Op1 = Op.getOperand(1);
2760 SDValue CC = Op.getOperand(2);
2761 EVT VT = Op.getValueType();
2762 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2763 DebugLoc dl = Op.getDebugLoc();
2765 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2766 switch (SetCCOpcode) {
2767 default: llvm_unreachable("Illegal FP comparison"); break;
2769 case ISD::SETNE: Invert = true; // Fallthrough
2771 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2773 case ISD::SETLT: Swap = true; // Fallthrough
2775 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2777 case ISD::SETLE: Swap = true; // Fallthrough
2779 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2780 case ISD::SETUGE: Swap = true; // Fallthrough
2781 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2782 case ISD::SETUGT: Swap = true; // Fallthrough
2783 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2784 case ISD::SETUEQ: Invert = true; // Fallthrough
2786 // Expand this to (OLT | OGT).
2790 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2791 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2793 case ISD::SETUO: Invert = true; // Fallthrough
2795 // Expand this to (OLT | OGE).
2799 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2800 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2804 // Integer comparisons.
2805 switch (SetCCOpcode) {
2806 default: llvm_unreachable("Illegal integer comparison"); break;
2807 case ISD::SETNE: Invert = true;
2808 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2809 case ISD::SETLT: Swap = true;
2810 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2811 case ISD::SETLE: Swap = true;
2812 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2813 case ISD::SETULT: Swap = true;
2814 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2815 case ISD::SETULE: Swap = true;
2816 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2819 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2820 if (Opc == ARMISD::VCEQ) {
2823 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2825 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2828 // Ignore bitconvert.
2829 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2830 AndOp = AndOp.getOperand(0);
2832 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2834 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2835 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2842 std::swap(Op0, Op1);
2844 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2847 Result = DAG.getNOT(dl, Result, VT);
2852 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2853 /// valid vector constant for a NEON instruction with a "modified immediate"
2854 /// operand (e.g., VMOV). If so, return either the constant being
2855 /// splatted or the encoded value, depending on the DoEncode parameter. The
2856 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2857 /// bits7-0=Immediate.
2858 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2859 unsigned SplatBitSize, SelectionDAG &DAG,
2860 bool isVMOV, bool DoEncode) {
2861 unsigned Op, Cmode, Imm;
2864 // SplatBitSize is set to the smallest size that splats the vector, so a
2865 // zero vector will always have SplatBitSize == 8. However, NEON modified
2866 // immediate instructions others than VMOV do not support the 8-bit encoding
2867 // of a zero vector, and the default encoding of zero is supposed to be the
2873 switch (SplatBitSize) {
2875 // Any 1-byte value is OK. Op=0, Cmode=1110.
2876 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2883 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2885 if ((SplatBits & ~0xff) == 0) {
2886 // Value = 0x00nn: Op=x, Cmode=100x.
2891 if ((SplatBits & ~0xff00) == 0) {
2892 // Value = 0xnn00: Op=x, Cmode=101x.
2894 Imm = SplatBits >> 8;
2900 // NEON's 32-bit VMOV supports splat values where:
2901 // * only one byte is nonzero, or
2902 // * the least significant byte is 0xff and the second byte is nonzero, or
2903 // * the least significant 2 bytes are 0xff and the third is nonzero.
2905 if ((SplatBits & ~0xff) == 0) {
2906 // Value = 0x000000nn: Op=x, Cmode=000x.
2911 if ((SplatBits & ~0xff00) == 0) {
2912 // Value = 0x0000nn00: Op=x, Cmode=001x.
2914 Imm = SplatBits >> 8;
2917 if ((SplatBits & ~0xff0000) == 0) {
2918 // Value = 0x00nn0000: Op=x, Cmode=010x.
2920 Imm = SplatBits >> 16;
2923 if ((SplatBits & ~0xff000000) == 0) {
2924 // Value = 0xnn000000: Op=x, Cmode=011x.
2926 Imm = SplatBits >> 24;
2930 if ((SplatBits & ~0xffff) == 0 &&
2931 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2932 // Value = 0x0000nnff: Op=x, Cmode=1100.
2934 Imm = SplatBits >> 8;
2939 if ((SplatBits & ~0xffffff) == 0 &&
2940 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2941 // Value = 0x00nnffff: Op=x, Cmode=1101.
2943 Imm = SplatBits >> 16;
2944 SplatBits |= 0xffff;
2948 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2949 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2950 // VMOV.I32. A (very) minor optimization would be to replicate the value
2951 // and fall through here to test for a valid 64-bit splat. But, then the
2952 // caller would also need to check and handle the change in size.
2956 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2959 uint64_t BitMask = 0xff;
2961 unsigned ImmMask = 1;
2963 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2964 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2967 } else if ((SplatBits & BitMask) != 0) {
2973 // Op=1, Cmode=1110.
2982 llvm_unreachable("unexpected size for isNEONModifiedImm");
2987 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2988 return DAG.getTargetConstant(SplatBits, VT);
2992 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2993 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2994 /// size, return the encoded value for that immediate. The ByteSize field
2995 /// indicates the number of bytes of each element [1248].
2996 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2997 SelectionDAG &DAG) {
2998 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2999 APInt SplatBits, SplatUndef;
3000 unsigned SplatBitSize;
3002 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3003 HasAnyUndefs, ByteSize * 8))
3006 if (SplatBitSize > ByteSize * 8)
3009 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
3010 SplatBitSize, DAG, isVMOV, true);
3013 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3014 bool &ReverseVEXT, unsigned &Imm) {
3015 unsigned NumElts = VT.getVectorNumElements();
3016 ReverseVEXT = false;
3019 // If this is a VEXT shuffle, the immediate value is the index of the first
3020 // element. The other shuffle indices must be the successive elements after
3022 unsigned ExpectedElt = Imm;
3023 for (unsigned i = 1; i < NumElts; ++i) {
3024 // Increment the expected index. If it wraps around, it may still be
3025 // a VEXT but the source vectors must be swapped.
3027 if (ExpectedElt == NumElts * 2) {
3032 if (ExpectedElt != static_cast<unsigned>(M[i]))
3036 // Adjust the index value if the source operands will be swapped.
3043 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3044 /// instruction with the specified blocksize. (The order of the elements
3045 /// within each block of the vector is reversed.)
3046 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3047 unsigned BlockSize) {
3048 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3049 "Only possible block sizes for VREV are: 16, 32, 64");
3051 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3055 unsigned NumElts = VT.getVectorNumElements();
3056 unsigned BlockElts = M[0] + 1;
3058 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3061 for (unsigned i = 0; i < NumElts; ++i) {
3062 if ((unsigned) M[i] !=
3063 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3070 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3071 unsigned &WhichResult) {
3072 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3076 unsigned NumElts = VT.getVectorNumElements();
3077 WhichResult = (M[0] == 0 ? 0 : 1);
3078 for (unsigned i = 0; i < NumElts; i += 2) {
3079 if ((unsigned) M[i] != i + WhichResult ||
3080 (unsigned) M[i+1] != i + NumElts + WhichResult)
3086 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3087 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3088 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3089 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3090 unsigned &WhichResult) {
3091 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3095 unsigned NumElts = VT.getVectorNumElements();
3096 WhichResult = (M[0] == 0 ? 0 : 1);
3097 for (unsigned i = 0; i < NumElts; i += 2) {
3098 if ((unsigned) M[i] != i + WhichResult ||
3099 (unsigned) M[i+1] != i + WhichResult)
3105 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3106 unsigned &WhichResult) {
3107 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3111 unsigned NumElts = VT.getVectorNumElements();
3112 WhichResult = (M[0] == 0 ? 0 : 1);
3113 for (unsigned i = 0; i != NumElts; ++i) {
3114 if ((unsigned) M[i] != 2 * i + WhichResult)
3118 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3119 if (VT.is64BitVector() && EltSz == 32)
3125 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3126 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3127 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3128 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3129 unsigned &WhichResult) {
3130 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3134 unsigned Half = VT.getVectorNumElements() / 2;
3135 WhichResult = (M[0] == 0 ? 0 : 1);
3136 for (unsigned j = 0; j != 2; ++j) {
3137 unsigned Idx = WhichResult;
3138 for (unsigned i = 0; i != Half; ++i) {
3139 if ((unsigned) M[i + j * Half] != Idx)
3145 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3146 if (VT.is64BitVector() && EltSz == 32)
3152 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3153 unsigned &WhichResult) {
3154 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3158 unsigned NumElts = VT.getVectorNumElements();
3159 WhichResult = (M[0] == 0 ? 0 : 1);
3160 unsigned Idx = WhichResult * NumElts / 2;
3161 for (unsigned i = 0; i != NumElts; i += 2) {
3162 if ((unsigned) M[i] != Idx ||
3163 (unsigned) M[i+1] != Idx + NumElts)
3168 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3169 if (VT.is64BitVector() && EltSz == 32)
3175 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3176 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3177 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3178 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3179 unsigned &WhichResult) {
3180 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3184 unsigned NumElts = VT.getVectorNumElements();
3185 WhichResult = (M[0] == 0 ? 0 : 1);
3186 unsigned Idx = WhichResult * NumElts / 2;
3187 for (unsigned i = 0; i != NumElts; i += 2) {
3188 if ((unsigned) M[i] != Idx ||
3189 (unsigned) M[i+1] != Idx)
3194 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3195 if (VT.is64BitVector() && EltSz == 32)
3202 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3203 // Canonicalize all-zeros and all-ones vectors.
3204 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3205 if (ConstVal->isNullValue())
3206 return getZeroVector(VT, DAG, dl);
3207 if (ConstVal->isAllOnesValue())
3208 return getOnesVector(VT, DAG, dl);
3211 if (VT.is64BitVector()) {
3212 switch (Val.getValueType().getSizeInBits()) {
3213 case 8: CanonicalVT = MVT::v8i8; break;
3214 case 16: CanonicalVT = MVT::v4i16; break;
3215 case 32: CanonicalVT = MVT::v2i32; break;
3216 case 64: CanonicalVT = MVT::v1i64; break;
3217 default: llvm_unreachable("unexpected splat element type"); break;
3220 assert(VT.is128BitVector() && "unknown splat vector size");
3221 switch (Val.getValueType().getSizeInBits()) {
3222 case 8: CanonicalVT = MVT::v16i8; break;
3223 case 16: CanonicalVT = MVT::v8i16; break;
3224 case 32: CanonicalVT = MVT::v4i32; break;
3225 case 64: CanonicalVT = MVT::v2i64; break;
3226 default: llvm_unreachable("unexpected splat element type"); break;
3230 // Build a canonical splat for this value.
3231 SmallVector<SDValue, 8> Ops;
3232 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3233 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3235 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3238 // If this is a case we can't handle, return null and let the default
3239 // expansion code take care of it.
3240 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3241 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3242 DebugLoc dl = Op.getDebugLoc();
3243 EVT VT = Op.getValueType();
3245 APInt SplatBits, SplatUndef;
3246 unsigned SplatBitSize;
3248 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3249 if (SplatBitSize <= 64) {
3250 // Check if an immediate VMOV works.
3251 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3252 SplatUndef.getZExtValue(),
3253 SplatBitSize, DAG, true, false);
3255 return BuildSplat(Val, VT, DAG, dl);
3259 // Scan through the operands to see if only one value is used.
3260 unsigned NumElts = VT.getVectorNumElements();
3261 bool isOnlyLowElement = true;
3262 bool usesOnlyOneValue = true;
3263 bool isConstant = true;
3265 for (unsigned i = 0; i < NumElts; ++i) {
3266 SDValue V = Op.getOperand(i);
3267 if (V.getOpcode() == ISD::UNDEF)
3270 isOnlyLowElement = false;
3271 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3274 if (!Value.getNode())
3276 else if (V != Value)
3277 usesOnlyOneValue = false;
3280 if (!Value.getNode())
3281 return DAG.getUNDEF(VT);
3283 if (isOnlyLowElement)
3284 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3286 // If all elements are constants, fall back to the default expansion, which
3287 // will generate a load from the constant pool.
3291 // Use VDUP for non-constant splats.
3292 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3293 if (usesOnlyOneValue && EltSize <= 32)
3294 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3296 // Vectors with 32- or 64-bit elements can be built by directly assigning
3297 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3298 // will be legalized.
3299 if (EltSize >= 32) {
3300 // Do the expansion with floating-point types, since that is what the VFP
3301 // registers are defined to use, and since i64 is not legal.
3302 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3303 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3304 SmallVector<SDValue, 8> Ops;
3305 for (unsigned i = 0; i < NumElts; ++i)
3306 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3307 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3308 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3314 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3315 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3316 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3317 /// are assumed to be legal.
3319 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3321 if (VT.getVectorNumElements() == 4 &&
3322 (VT.is128BitVector() || VT.is64BitVector())) {
3323 unsigned PFIndexes[4];
3324 for (unsigned i = 0; i != 4; ++i) {
3328 PFIndexes[i] = M[i];
3331 // Compute the index in the perfect shuffle table.
3332 unsigned PFTableIndex =
3333 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3334 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3335 unsigned Cost = (PFEntry >> 30);
3342 unsigned Imm, WhichResult;
3344 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3345 return (EltSize >= 32 ||
3346 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3347 isVREVMask(M, VT, 64) ||
3348 isVREVMask(M, VT, 32) ||
3349 isVREVMask(M, VT, 16) ||
3350 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3351 isVTRNMask(M, VT, WhichResult) ||
3352 isVUZPMask(M, VT, WhichResult) ||
3353 isVZIPMask(M, VT, WhichResult) ||
3354 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3355 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3356 isVZIP_v_undef_Mask(M, VT, WhichResult));
3359 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3360 /// the specified operations to build the shuffle.
3361 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3362 SDValue RHS, SelectionDAG &DAG,
3364 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3365 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3366 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3369 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3378 OP_VUZPL, // VUZP, left result
3379 OP_VUZPR, // VUZP, right result
3380 OP_VZIPL, // VZIP, left result
3381 OP_VZIPR, // VZIP, right result
3382 OP_VTRNL, // VTRN, left result
3383 OP_VTRNR // VTRN, right result
3386 if (OpNum == OP_COPY) {
3387 if (LHSID == (1*9+2)*9+3) return LHS;
3388 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3392 SDValue OpLHS, OpRHS;
3393 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3394 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3395 EVT VT = OpLHS.getValueType();
3398 default: llvm_unreachable("Unknown shuffle opcode!");
3400 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3405 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3406 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3410 return DAG.getNode(ARMISD::VEXT, dl, VT,
3412 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3415 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3416 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3419 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3420 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3423 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3424 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3428 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3429 SDValue V1 = Op.getOperand(0);
3430 SDValue V2 = Op.getOperand(1);
3431 DebugLoc dl = Op.getDebugLoc();
3432 EVT VT = Op.getValueType();
3433 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3434 SmallVector<int, 8> ShuffleMask;
3436 // Convert shuffles that are directly supported on NEON to target-specific
3437 // DAG nodes, instead of keeping them as shuffles and matching them again
3438 // during code selection. This is more efficient and avoids the possibility
3439 // of inconsistencies between legalization and selection.
3440 // FIXME: floating-point vectors should be canonicalized to integer vectors
3441 // of the same time so that they get CSEd properly.
3442 SVN->getMask(ShuffleMask);
3444 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3445 if (EltSize <= 32) {
3446 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3447 int Lane = SVN->getSplatIndex();
3448 // If this is undef splat, generate it via "just" vdup, if possible.
3449 if (Lane == -1) Lane = 0;
3451 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3452 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3454 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3455 DAG.getConstant(Lane, MVT::i32));
3460 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3463 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3464 DAG.getConstant(Imm, MVT::i32));
3467 if (isVREVMask(ShuffleMask, VT, 64))
3468 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3469 if (isVREVMask(ShuffleMask, VT, 32))
3470 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3471 if (isVREVMask(ShuffleMask, VT, 16))
3472 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3474 // Check for Neon shuffles that modify both input vectors in place.
3475 // If both results are used, i.e., if there are two shuffles with the same
3476 // source operands and with masks corresponding to both results of one of
3477 // these operations, DAG memoization will ensure that a single node is
3478 // used for both shuffles.
3479 unsigned WhichResult;
3480 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3481 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3482 V1, V2).getValue(WhichResult);
3483 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3484 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3485 V1, V2).getValue(WhichResult);
3486 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3487 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3488 V1, V2).getValue(WhichResult);
3490 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3491 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3492 V1, V1).getValue(WhichResult);
3493 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3494 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3495 V1, V1).getValue(WhichResult);
3496 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3497 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3498 V1, V1).getValue(WhichResult);
3501 // If the shuffle is not directly supported and it has 4 elements, use
3502 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3503 unsigned NumElts = VT.getVectorNumElements();
3505 unsigned PFIndexes[4];
3506 for (unsigned i = 0; i != 4; ++i) {
3507 if (ShuffleMask[i] < 0)
3510 PFIndexes[i] = ShuffleMask[i];
3513 // Compute the index in the perfect shuffle table.
3514 unsigned PFTableIndex =
3515 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3516 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3517 unsigned Cost = (PFEntry >> 30);
3520 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3523 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3524 if (EltSize >= 32) {
3525 // Do the expansion with floating-point types, since that is what the VFP
3526 // registers are defined to use, and since i64 is not legal.
3527 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3528 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3529 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3530 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3531 SmallVector<SDValue, 8> Ops;
3532 for (unsigned i = 0; i < NumElts; ++i) {
3533 if (ShuffleMask[i] < 0)
3534 Ops.push_back(DAG.getUNDEF(EltVT));
3536 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3537 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3538 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3541 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3542 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3548 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3549 EVT VT = Op.getValueType();
3550 DebugLoc dl = Op.getDebugLoc();
3551 SDValue Vec = Op.getOperand(0);
3552 SDValue Lane = Op.getOperand(1);
3553 assert(VT == MVT::i32 &&
3554 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3555 "unexpected type for custom-lowering vector extract");
3556 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3559 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3560 // The only time a CONCAT_VECTORS operation can have legal types is when
3561 // two 64-bit vectors are concatenated to a 128-bit vector.
3562 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3563 "unexpected CONCAT_VECTORS");
3564 DebugLoc dl = Op.getDebugLoc();
3565 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3566 SDValue Op0 = Op.getOperand(0);
3567 SDValue Op1 = Op.getOperand(1);
3568 if (Op0.getOpcode() != ISD::UNDEF)
3569 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3570 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3571 DAG.getIntPtrConstant(0));
3572 if (Op1.getOpcode() != ISD::UNDEF)
3573 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3574 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3575 DAG.getIntPtrConstant(1));
3576 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3579 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3580 switch (Op.getOpcode()) {
3581 default: llvm_unreachable("Don't know how to custom lower this!");
3582 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3583 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3584 case ISD::GlobalAddress:
3585 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3586 LowerGlobalAddressELF(Op, DAG);
3587 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3588 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3589 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3590 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3591 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3592 case ISD::VASTART: return LowerVASTART(Op, DAG);
3593 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3594 case ISD::SINT_TO_FP:
3595 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3596 case ISD::FP_TO_SINT:
3597 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3598 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3599 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3600 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3601 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3602 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3603 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3604 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3606 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3609 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3610 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3611 case ISD::SRL_PARTS:
3612 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3613 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3614 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3615 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3616 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3617 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3618 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3623 /// ReplaceNodeResults - Replace the results of node with an illegal result
3624 /// type with new values built out of custom code.
3625 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3626 SmallVectorImpl<SDValue>&Results,
3627 SelectionDAG &DAG) const {
3629 switch (N->getOpcode()) {
3631 llvm_unreachable("Don't know how to custom expand this!");
3633 case ISD::BIT_CONVERT:
3634 Res = ExpandBIT_CONVERT(N, DAG);
3638 Res = LowerShift(N, DAG, Subtarget);
3642 Results.push_back(Res);
3645 //===----------------------------------------------------------------------===//
3646 // ARM Scheduler Hooks
3647 //===----------------------------------------------------------------------===//
3650 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3651 MachineBasicBlock *BB,
3652 unsigned Size) const {
3653 unsigned dest = MI->getOperand(0).getReg();
3654 unsigned ptr = MI->getOperand(1).getReg();
3655 unsigned oldval = MI->getOperand(2).getReg();
3656 unsigned newval = MI->getOperand(3).getReg();
3657 unsigned scratch = BB->getParent()->getRegInfo()
3658 .createVirtualRegister(ARM::GPRRegisterClass);
3659 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3660 DebugLoc dl = MI->getDebugLoc();
3661 bool isThumb2 = Subtarget->isThumb2();
3663 unsigned ldrOpc, strOpc;
3665 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3667 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3668 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3671 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3672 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3675 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3676 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3680 MachineFunction *MF = BB->getParent();
3681 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3682 MachineFunction::iterator It = BB;
3683 ++It; // insert the new blocks after the current block
3685 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3686 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3687 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3688 MF->insert(It, loop1MBB);
3689 MF->insert(It, loop2MBB);
3690 MF->insert(It, exitMBB);
3692 // Transfer the remainder of BB and its successor edges to exitMBB.
3693 exitMBB->splice(exitMBB->begin(), BB,
3694 llvm::next(MachineBasicBlock::iterator(MI)),
3696 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3700 // fallthrough --> loop1MBB
3701 BB->addSuccessor(loop1MBB);
3704 // ldrex dest, [ptr]
3708 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3709 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3710 .addReg(dest).addReg(oldval));
3711 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3712 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3713 BB->addSuccessor(loop2MBB);
3714 BB->addSuccessor(exitMBB);
3717 // strex scratch, newval, [ptr]
3721 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3723 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3724 .addReg(scratch).addImm(0));
3725 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3726 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3727 BB->addSuccessor(loop1MBB);
3728 BB->addSuccessor(exitMBB);
3734 MI->eraseFromParent(); // The instruction is gone now.
3740 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3741 unsigned Size, unsigned BinOpcode) const {
3742 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3743 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3745 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3746 MachineFunction *MF = BB->getParent();
3747 MachineFunction::iterator It = BB;
3750 unsigned dest = MI->getOperand(0).getReg();
3751 unsigned ptr = MI->getOperand(1).getReg();
3752 unsigned incr = MI->getOperand(2).getReg();
3753 DebugLoc dl = MI->getDebugLoc();
3755 bool isThumb2 = Subtarget->isThumb2();
3756 unsigned ldrOpc, strOpc;
3758 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3760 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3761 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3764 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3765 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3768 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3769 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3773 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3774 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3775 MF->insert(It, loopMBB);
3776 MF->insert(It, exitMBB);
3778 // Transfer the remainder of BB and its successor edges to exitMBB.
3779 exitMBB->splice(exitMBB->begin(), BB,
3780 llvm::next(MachineBasicBlock::iterator(MI)),
3782 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3784 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3785 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3786 unsigned scratch2 = (!BinOpcode) ? incr :
3787 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3791 // fallthrough --> loopMBB
3792 BB->addSuccessor(loopMBB);
3796 // <binop> scratch2, dest, incr
3797 // strex scratch, scratch2, ptr
3800 // fallthrough --> exitMBB
3802 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3804 // operand order needs to go the other way for NAND
3805 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3806 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3807 addReg(incr).addReg(dest)).addReg(0);
3809 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3810 addReg(dest).addReg(incr)).addReg(0);
3813 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3815 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3816 .addReg(scratch).addImm(0));
3817 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3818 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3820 BB->addSuccessor(loopMBB);
3821 BB->addSuccessor(exitMBB);
3827 MI->eraseFromParent(); // The instruction is gone now.
3833 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3834 MachineBasicBlock *BB) const {
3835 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3836 DebugLoc dl = MI->getDebugLoc();
3837 bool isThumb2 = Subtarget->isThumb2();
3838 switch (MI->getOpcode()) {
3841 llvm_unreachable("Unexpected instr type to insert");
3843 case ARM::ATOMIC_LOAD_ADD_I8:
3844 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3845 case ARM::ATOMIC_LOAD_ADD_I16:
3846 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3847 case ARM::ATOMIC_LOAD_ADD_I32:
3848 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3850 case ARM::ATOMIC_LOAD_AND_I8:
3851 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3852 case ARM::ATOMIC_LOAD_AND_I16:
3853 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3854 case ARM::ATOMIC_LOAD_AND_I32:
3855 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3857 case ARM::ATOMIC_LOAD_OR_I8:
3858 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3859 case ARM::ATOMIC_LOAD_OR_I16:
3860 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3861 case ARM::ATOMIC_LOAD_OR_I32:
3862 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3864 case ARM::ATOMIC_LOAD_XOR_I8:
3865 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3866 case ARM::ATOMIC_LOAD_XOR_I16:
3867 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3868 case ARM::ATOMIC_LOAD_XOR_I32:
3869 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3871 case ARM::ATOMIC_LOAD_NAND_I8:
3872 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3873 case ARM::ATOMIC_LOAD_NAND_I16:
3874 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3875 case ARM::ATOMIC_LOAD_NAND_I32:
3876 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3878 case ARM::ATOMIC_LOAD_SUB_I8:
3879 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3880 case ARM::ATOMIC_LOAD_SUB_I16:
3881 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3882 case ARM::ATOMIC_LOAD_SUB_I32:
3883 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3885 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3886 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3887 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3889 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3890 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3891 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3893 case ARM::tMOVCCr_pseudo: {
3894 // To "insert" a SELECT_CC instruction, we actually have to insert the
3895 // diamond control-flow pattern. The incoming instruction knows the
3896 // destination vreg to set, the condition code register to branch on, the
3897 // true/false values to select between, and a branch opcode to use.
3898 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3899 MachineFunction::iterator It = BB;
3905 // cmpTY ccX, r1, r2
3907 // fallthrough --> copy0MBB
3908 MachineBasicBlock *thisMBB = BB;
3909 MachineFunction *F = BB->getParent();
3910 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3911 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3912 F->insert(It, copy0MBB);
3913 F->insert(It, sinkMBB);
3915 // Transfer the remainder of BB and its successor edges to sinkMBB.
3916 sinkMBB->splice(sinkMBB->begin(), BB,
3917 llvm::next(MachineBasicBlock::iterator(MI)),
3919 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
3921 BB->addSuccessor(copy0MBB);
3922 BB->addSuccessor(sinkMBB);
3924 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3925 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3928 // %FalseValue = ...
3929 // # fallthrough to sinkMBB
3932 // Update machine-CFG edges
3933 BB->addSuccessor(sinkMBB);
3936 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3939 BuildMI(*BB, BB->begin(), dl,
3940 TII->get(ARM::PHI), MI->getOperand(0).getReg())
3941 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3942 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3944 MI->eraseFromParent(); // The pseudo instruction is gone now.
3951 case ARM::t2SUBrSPi_:
3952 case ARM::t2SUBrSPi12_:
3953 case ARM::t2SUBrSPs_: {
3954 MachineFunction *MF = BB->getParent();
3955 unsigned DstReg = MI->getOperand(0).getReg();
3956 unsigned SrcReg = MI->getOperand(1).getReg();
3957 bool DstIsDead = MI->getOperand(0).isDead();
3958 bool SrcIsKill = MI->getOperand(1).isKill();
3960 if (SrcReg != ARM::SP) {
3961 // Copy the source to SP from virtual register.
3962 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3963 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3964 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3965 BuildMI(*BB, MI, dl, TII->get(CopyOpc), ARM::SP)
3966 .addReg(SrcReg, getKillRegState(SrcIsKill));
3970 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3971 switch (MI->getOpcode()) {
3973 llvm_unreachable("Unexpected pseudo instruction!");
3979 OpOpc = ARM::tADDspr;
3982 OpOpc = ARM::tSUBspi;
3984 case ARM::t2SUBrSPi_:
3985 OpOpc = ARM::t2SUBrSPi;
3986 NeedPred = true; NeedCC = true;
3988 case ARM::t2SUBrSPi12_:
3989 OpOpc = ARM::t2SUBrSPi12;
3992 case ARM::t2SUBrSPs_:
3993 OpOpc = ARM::t2SUBrSPs;
3994 NeedPred = true; NeedCC = true; NeedOp3 = true;
3997 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(OpOpc), ARM::SP);
3998 if (OpOpc == ARM::tAND)
3999 AddDefaultT1CC(MIB);
4000 MIB.addReg(ARM::SP);
4001 MIB.addOperand(MI->getOperand(2));
4003 MIB.addOperand(MI->getOperand(3));
4005 AddDefaultPred(MIB);
4009 // Copy the result from SP to virtual register.
4010 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
4011 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4012 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
4013 BuildMI(*BB, MI, dl, TII->get(CopyOpc))
4014 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
4016 MI->eraseFromParent(); // The pseudo instruction is gone now.
4022 //===----------------------------------------------------------------------===//
4023 // ARM Optimization Hooks
4024 //===----------------------------------------------------------------------===//
4027 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4028 TargetLowering::DAGCombinerInfo &DCI) {
4029 SelectionDAG &DAG = DCI.DAG;
4030 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4031 EVT VT = N->getValueType(0);
4032 unsigned Opc = N->getOpcode();
4033 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4034 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4035 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4036 ISD::CondCode CC = ISD::SETCC_INVALID;
4039 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4041 SDValue CCOp = Slct.getOperand(0);
4042 if (CCOp.getOpcode() == ISD::SETCC)
4043 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4046 bool DoXform = false;
4048 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4051 if (LHS.getOpcode() == ISD::Constant &&
4052 cast<ConstantSDNode>(LHS)->isNullValue()) {
4054 } else if (CC != ISD::SETCC_INVALID &&
4055 RHS.getOpcode() == ISD::Constant &&
4056 cast<ConstantSDNode>(RHS)->isNullValue()) {
4057 std::swap(LHS, RHS);
4058 SDValue Op0 = Slct.getOperand(0);
4059 EVT OpVT = isSlctCC ? Op0.getValueType() :
4060 Op0.getOperand(0).getValueType();
4061 bool isInt = OpVT.isInteger();
4062 CC = ISD::getSetCCInverse(CC, isInt);
4064 if (!TLI.isCondCodeLegal(CC, OpVT))
4065 return SDValue(); // Inverse operator isn't legal.
4072 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4074 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4075 Slct.getOperand(0), Slct.getOperand(1), CC);
4076 SDValue CCOp = Slct.getOperand(0);
4078 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4079 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4080 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4081 CCOp, OtherOp, Result);
4086 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4087 static SDValue PerformADDCombine(SDNode *N,
4088 TargetLowering::DAGCombinerInfo &DCI) {
4089 // added by evan in r37685 with no testcase.
4090 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4092 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4093 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4094 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4095 if (Result.getNode()) return Result;
4097 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4098 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4099 if (Result.getNode()) return Result;
4105 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4106 static SDValue PerformSUBCombine(SDNode *N,
4107 TargetLowering::DAGCombinerInfo &DCI) {
4108 // added by evan in r37685 with no testcase.
4109 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4111 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4112 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4113 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4114 if (Result.getNode()) return Result;
4120 static SDValue PerformMULCombine(SDNode *N,
4121 TargetLowering::DAGCombinerInfo &DCI,
4122 const ARMSubtarget *Subtarget) {
4123 SelectionDAG &DAG = DCI.DAG;
4125 if (Subtarget->isThumb1Only())
4128 if (DAG.getMachineFunction().
4129 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4132 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4135 EVT VT = N->getValueType(0);
4139 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4143 uint64_t MulAmt = C->getZExtValue();
4144 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4145 ShiftAmt = ShiftAmt & (32 - 1);
4146 SDValue V = N->getOperand(0);
4147 DebugLoc DL = N->getDebugLoc();
4150 MulAmt >>= ShiftAmt;
4151 if (isPowerOf2_32(MulAmt - 1)) {
4152 // (mul x, 2^N + 1) => (add (shl x, N), x)
4153 Res = DAG.getNode(ISD::ADD, DL, VT,
4154 V, DAG.getNode(ISD::SHL, DL, VT,
4155 V, DAG.getConstant(Log2_32(MulAmt-1),
4157 } else if (isPowerOf2_32(MulAmt + 1)) {
4158 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4159 Res = DAG.getNode(ISD::SUB, DL, VT,
4160 DAG.getNode(ISD::SHL, DL, VT,
4161 V, DAG.getConstant(Log2_32(MulAmt+1),
4168 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4169 DAG.getConstant(ShiftAmt, MVT::i32));
4171 // Do not add new nodes to DAG combiner worklist.
4172 DCI.CombineTo(N, Res, false);
4176 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4177 /// ARMISD::VMOVRRD.
4178 static SDValue PerformVMOVRRDCombine(SDNode *N,
4179 TargetLowering::DAGCombinerInfo &DCI) {
4180 // fmrrd(fmdrr x, y) -> x,y
4181 SDValue InDouble = N->getOperand(0);
4182 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4183 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4187 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4188 /// operand of a vector shift operation, where all the elements of the
4189 /// build_vector must have the same constant integer value.
4190 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4191 // Ignore bit_converts.
4192 while (Op.getOpcode() == ISD::BIT_CONVERT)
4193 Op = Op.getOperand(0);
4194 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4195 APInt SplatBits, SplatUndef;
4196 unsigned SplatBitSize;
4198 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4199 HasAnyUndefs, ElementBits) ||
4200 SplatBitSize > ElementBits)
4202 Cnt = SplatBits.getSExtValue();
4206 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4207 /// operand of a vector shift left operation. That value must be in the range:
4208 /// 0 <= Value < ElementBits for a left shift; or
4209 /// 0 <= Value <= ElementBits for a long left shift.
4210 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4211 assert(VT.isVector() && "vector shift count is not a vector type");
4212 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4213 if (! getVShiftImm(Op, ElementBits, Cnt))
4215 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4218 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4219 /// operand of a vector shift right operation. For a shift opcode, the value
4220 /// is positive, but for an intrinsic the value count must be negative. The
4221 /// absolute value must be in the range:
4222 /// 1 <= |Value| <= ElementBits for a right shift; or
4223 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4224 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4226 assert(VT.isVector() && "vector shift count is not a vector type");
4227 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4228 if (! getVShiftImm(Op, ElementBits, Cnt))
4232 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4235 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4236 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4237 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4240 // Don't do anything for most intrinsics.
4243 // Vector shifts: check for immediate versions and lower them.
4244 // Note: This is done during DAG combining instead of DAG legalizing because
4245 // the build_vectors for 64-bit vector element shift counts are generally
4246 // not legal, and it is hard to see their values after they get legalized to
4247 // loads from a constant pool.
4248 case Intrinsic::arm_neon_vshifts:
4249 case Intrinsic::arm_neon_vshiftu:
4250 case Intrinsic::arm_neon_vshiftls:
4251 case Intrinsic::arm_neon_vshiftlu:
4252 case Intrinsic::arm_neon_vshiftn:
4253 case Intrinsic::arm_neon_vrshifts:
4254 case Intrinsic::arm_neon_vrshiftu:
4255 case Intrinsic::arm_neon_vrshiftn:
4256 case Intrinsic::arm_neon_vqshifts:
4257 case Intrinsic::arm_neon_vqshiftu:
4258 case Intrinsic::arm_neon_vqshiftsu:
4259 case Intrinsic::arm_neon_vqshiftns:
4260 case Intrinsic::arm_neon_vqshiftnu:
4261 case Intrinsic::arm_neon_vqshiftnsu:
4262 case Intrinsic::arm_neon_vqrshiftns:
4263 case Intrinsic::arm_neon_vqrshiftnu:
4264 case Intrinsic::arm_neon_vqrshiftnsu: {
4265 EVT VT = N->getOperand(1).getValueType();
4267 unsigned VShiftOpc = 0;
4270 case Intrinsic::arm_neon_vshifts:
4271 case Intrinsic::arm_neon_vshiftu:
4272 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4273 VShiftOpc = ARMISD::VSHL;
4276 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4277 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4278 ARMISD::VSHRs : ARMISD::VSHRu);
4283 case Intrinsic::arm_neon_vshiftls:
4284 case Intrinsic::arm_neon_vshiftlu:
4285 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4287 llvm_unreachable("invalid shift count for vshll intrinsic");
4289 case Intrinsic::arm_neon_vrshifts:
4290 case Intrinsic::arm_neon_vrshiftu:
4291 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4295 case Intrinsic::arm_neon_vqshifts:
4296 case Intrinsic::arm_neon_vqshiftu:
4297 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4301 case Intrinsic::arm_neon_vqshiftsu:
4302 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4304 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4306 case Intrinsic::arm_neon_vshiftn:
4307 case Intrinsic::arm_neon_vrshiftn:
4308 case Intrinsic::arm_neon_vqshiftns:
4309 case Intrinsic::arm_neon_vqshiftnu:
4310 case Intrinsic::arm_neon_vqshiftnsu:
4311 case Intrinsic::arm_neon_vqrshiftns:
4312 case Intrinsic::arm_neon_vqrshiftnu:
4313 case Intrinsic::arm_neon_vqrshiftnsu:
4314 // Narrowing shifts require an immediate right shift.
4315 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4317 llvm_unreachable("invalid shift count for narrowing vector shift "
4321 llvm_unreachable("unhandled vector shift");
4325 case Intrinsic::arm_neon_vshifts:
4326 case Intrinsic::arm_neon_vshiftu:
4327 // Opcode already set above.
4329 case Intrinsic::arm_neon_vshiftls:
4330 case Intrinsic::arm_neon_vshiftlu:
4331 if (Cnt == VT.getVectorElementType().getSizeInBits())
4332 VShiftOpc = ARMISD::VSHLLi;
4334 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4335 ARMISD::VSHLLs : ARMISD::VSHLLu);
4337 case Intrinsic::arm_neon_vshiftn:
4338 VShiftOpc = ARMISD::VSHRN; break;
4339 case Intrinsic::arm_neon_vrshifts:
4340 VShiftOpc = ARMISD::VRSHRs; break;
4341 case Intrinsic::arm_neon_vrshiftu:
4342 VShiftOpc = ARMISD::VRSHRu; break;
4343 case Intrinsic::arm_neon_vrshiftn:
4344 VShiftOpc = ARMISD::VRSHRN; break;
4345 case Intrinsic::arm_neon_vqshifts:
4346 VShiftOpc = ARMISD::VQSHLs; break;
4347 case Intrinsic::arm_neon_vqshiftu:
4348 VShiftOpc = ARMISD::VQSHLu; break;
4349 case Intrinsic::arm_neon_vqshiftsu:
4350 VShiftOpc = ARMISD::VQSHLsu; break;
4351 case Intrinsic::arm_neon_vqshiftns:
4352 VShiftOpc = ARMISD::VQSHRNs; break;
4353 case Intrinsic::arm_neon_vqshiftnu:
4354 VShiftOpc = ARMISD::VQSHRNu; break;
4355 case Intrinsic::arm_neon_vqshiftnsu:
4356 VShiftOpc = ARMISD::VQSHRNsu; break;
4357 case Intrinsic::arm_neon_vqrshiftns:
4358 VShiftOpc = ARMISD::VQRSHRNs; break;
4359 case Intrinsic::arm_neon_vqrshiftnu:
4360 VShiftOpc = ARMISD::VQRSHRNu; break;
4361 case Intrinsic::arm_neon_vqrshiftnsu:
4362 VShiftOpc = ARMISD::VQRSHRNsu; break;
4365 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4366 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4369 case Intrinsic::arm_neon_vshiftins: {
4370 EVT VT = N->getOperand(1).getValueType();
4372 unsigned VShiftOpc = 0;
4374 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4375 VShiftOpc = ARMISD::VSLI;
4376 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4377 VShiftOpc = ARMISD::VSRI;
4379 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4382 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4383 N->getOperand(1), N->getOperand(2),
4384 DAG.getConstant(Cnt, MVT::i32));
4387 case Intrinsic::arm_neon_vqrshifts:
4388 case Intrinsic::arm_neon_vqrshiftu:
4389 // No immediate versions of these to check for.
4396 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4397 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4398 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4399 /// vector element shift counts are generally not legal, and it is hard to see
4400 /// their values after they get legalized to loads from a constant pool.
4401 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4402 const ARMSubtarget *ST) {
4403 EVT VT = N->getValueType(0);
4405 // Nothing to be done for scalar shifts.
4406 if (! VT.isVector())
4409 assert(ST->hasNEON() && "unexpected vector shift");
4412 switch (N->getOpcode()) {
4413 default: llvm_unreachable("unexpected shift opcode");
4416 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4417 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4418 DAG.getConstant(Cnt, MVT::i32));
4423 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4424 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4425 ARMISD::VSHRs : ARMISD::VSHRu);
4426 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4427 DAG.getConstant(Cnt, MVT::i32));
4433 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4434 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4435 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4436 const ARMSubtarget *ST) {
4437 SDValue N0 = N->getOperand(0);
4439 // Check for sign- and zero-extensions of vector extract operations of 8-
4440 // and 16-bit vector elements. NEON supports these directly. They are
4441 // handled during DAG combining because type legalization will promote them
4442 // to 32-bit types and it is messy to recognize the operations after that.
4443 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4444 SDValue Vec = N0.getOperand(0);
4445 SDValue Lane = N0.getOperand(1);
4446 EVT VT = N->getValueType(0);
4447 EVT EltVT = N0.getValueType();
4448 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4450 if (VT == MVT::i32 &&
4451 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4452 TLI.isTypeLegal(Vec.getValueType())) {
4455 switch (N->getOpcode()) {
4456 default: llvm_unreachable("unexpected opcode");
4457 case ISD::SIGN_EXTEND:
4458 Opc = ARMISD::VGETLANEs;
4460 case ISD::ZERO_EXTEND:
4461 case ISD::ANY_EXTEND:
4462 Opc = ARMISD::VGETLANEu;
4465 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4472 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4473 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4474 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4475 const ARMSubtarget *ST) {
4476 // If the target supports NEON, try to use vmax/vmin instructions for f32
4477 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4478 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4479 // a NaN; only do the transformation when it matches that behavior.
4481 // For now only do this when using NEON for FP operations; if using VFP, it
4482 // is not obvious that the benefit outweighs the cost of switching to the
4484 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4485 N->getValueType(0) != MVT::f32)
4488 SDValue CondLHS = N->getOperand(0);
4489 SDValue CondRHS = N->getOperand(1);
4490 SDValue LHS = N->getOperand(2);
4491 SDValue RHS = N->getOperand(3);
4492 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4494 unsigned Opcode = 0;
4496 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4497 IsReversed = false; // x CC y ? x : y
4498 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4499 IsReversed = true ; // x CC y ? y : x
4513 // If LHS is NaN, an ordered comparison will be false and the result will
4514 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4515 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4516 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4517 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4519 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4520 // will return -0, so vmin can only be used for unsafe math or if one of
4521 // the operands is known to be nonzero.
4522 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4524 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4526 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4535 // If LHS is NaN, an ordered comparison will be false and the result will
4536 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4537 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4538 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4539 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4541 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4542 // will return +0, so vmax can only be used for unsafe math or if one of
4543 // the operands is known to be nonzero.
4544 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4546 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4548 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4554 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4557 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4558 DAGCombinerInfo &DCI) const {
4559 switch (N->getOpcode()) {
4561 case ISD::ADD: return PerformADDCombine(N, DCI);
4562 case ISD::SUB: return PerformSUBCombine(N, DCI);
4563 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4564 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4565 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4568 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4569 case ISD::SIGN_EXTEND:
4570 case ISD::ZERO_EXTEND:
4571 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4572 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4577 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4578 if (!Subtarget->hasV6Ops())
4579 // Pre-v6 does not support unaligned mem access.
4582 // v6+ may or may not support unaligned mem access depending on the system
4584 // FIXME: This is pretty conservative. Should we provide cmdline option to
4585 // control the behaviour?
4586 if (!Subtarget->isTargetDarwin())
4589 switch (VT.getSimpleVT().SimpleTy) {
4596 // FIXME: VLD1 etc with standard alignment is legal.
4600 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4605 switch (VT.getSimpleVT().SimpleTy) {
4606 default: return false;
4621 if ((V & (Scale - 1)) != 0)
4624 return V == (V & ((1LL << 5) - 1));
4627 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4628 const ARMSubtarget *Subtarget) {
4635 switch (VT.getSimpleVT().SimpleTy) {
4636 default: return false;
4641 // + imm12 or - imm8
4643 return V == (V & ((1LL << 8) - 1));
4644 return V == (V & ((1LL << 12) - 1));
4647 // Same as ARM mode. FIXME: NEON?
4648 if (!Subtarget->hasVFP2())
4653 return V == (V & ((1LL << 8) - 1));
4657 /// isLegalAddressImmediate - Return true if the integer value can be used
4658 /// as the offset of the target addressing mode for load / store of the
4660 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4661 const ARMSubtarget *Subtarget) {
4668 if (Subtarget->isThumb1Only())
4669 return isLegalT1AddressImmediate(V, VT);
4670 else if (Subtarget->isThumb2())
4671 return isLegalT2AddressImmediate(V, VT, Subtarget);
4676 switch (VT.getSimpleVT().SimpleTy) {
4677 default: return false;
4682 return V == (V & ((1LL << 12) - 1));
4685 return V == (V & ((1LL << 8) - 1));
4688 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4693 return V == (V & ((1LL << 8) - 1));
4697 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4699 int Scale = AM.Scale;
4703 switch (VT.getSimpleVT().SimpleTy) {
4704 default: return false;
4713 return Scale == 2 || Scale == 4 || Scale == 8;
4716 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4720 // Note, we allow "void" uses (basically, uses that aren't loads or
4721 // stores), because arm allows folding a scale into many arithmetic
4722 // operations. This should be made more precise and revisited later.
4724 // Allow r << imm, but the imm has to be a multiple of two.
4725 if (Scale & 1) return false;
4726 return isPowerOf2_32(Scale);
4730 /// isLegalAddressingMode - Return true if the addressing mode represented
4731 /// by AM is legal for this target, for a load/store of the specified type.
4732 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4733 const Type *Ty) const {
4734 EVT VT = getValueType(Ty, true);
4735 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4738 // Can never fold addr of global into load/store.
4743 case 0: // no scale reg, must be "r+i" or "r", or "i".
4746 if (Subtarget->isThumb1Only())
4750 // ARM doesn't support any R+R*scale+imm addr modes.
4757 if (Subtarget->isThumb2())
4758 return isLegalT2ScaledAddressingMode(AM, VT);
4760 int Scale = AM.Scale;
4761 switch (VT.getSimpleVT().SimpleTy) {
4762 default: return false;
4766 if (Scale < 0) Scale = -Scale;
4770 return isPowerOf2_32(Scale & ~1);
4774 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4779 // Note, we allow "void" uses (basically, uses that aren't loads or
4780 // stores), because arm allows folding a scale into many arithmetic
4781 // operations. This should be made more precise and revisited later.
4783 // Allow r << imm, but the imm has to be a multiple of two.
4784 if (Scale & 1) return false;
4785 return isPowerOf2_32(Scale);
4792 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4793 /// icmp immediate, that is the target has icmp instructions which can compare
4794 /// a register against the immediate without having to materialize the
4795 /// immediate into a register.
4796 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4797 if (!Subtarget->isThumb())
4798 return ARM_AM::getSOImmVal(Imm) != -1;
4799 if (Subtarget->isThumb2())
4800 return ARM_AM::getT2SOImmVal(Imm) != -1;
4801 return Imm >= 0 && Imm <= 255;
4804 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4805 bool isSEXTLoad, SDValue &Base,
4806 SDValue &Offset, bool &isInc,
4807 SelectionDAG &DAG) {
4808 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4811 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4813 Base = Ptr->getOperand(0);
4814 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4815 int RHSC = (int)RHS->getZExtValue();
4816 if (RHSC < 0 && RHSC > -256) {
4817 assert(Ptr->getOpcode() == ISD::ADD);
4819 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4823 isInc = (Ptr->getOpcode() == ISD::ADD);
4824 Offset = Ptr->getOperand(1);
4826 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4828 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4829 int RHSC = (int)RHS->getZExtValue();
4830 if (RHSC < 0 && RHSC > -0x1000) {
4831 assert(Ptr->getOpcode() == ISD::ADD);
4833 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4834 Base = Ptr->getOperand(0);
4839 if (Ptr->getOpcode() == ISD::ADD) {
4841 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4842 if (ShOpcVal != ARM_AM::no_shift) {
4843 Base = Ptr->getOperand(1);
4844 Offset = Ptr->getOperand(0);
4846 Base = Ptr->getOperand(0);
4847 Offset = Ptr->getOperand(1);
4852 isInc = (Ptr->getOpcode() == ISD::ADD);
4853 Base = Ptr->getOperand(0);
4854 Offset = Ptr->getOperand(1);
4858 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4862 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4863 bool isSEXTLoad, SDValue &Base,
4864 SDValue &Offset, bool &isInc,
4865 SelectionDAG &DAG) {
4866 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4869 Base = Ptr->getOperand(0);
4870 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4871 int RHSC = (int)RHS->getZExtValue();
4872 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4873 assert(Ptr->getOpcode() == ISD::ADD);
4875 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4877 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4878 isInc = Ptr->getOpcode() == ISD::ADD;
4879 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4887 /// getPreIndexedAddressParts - returns true by value, base pointer and
4888 /// offset pointer and addressing mode by reference if the node's address
4889 /// can be legally represented as pre-indexed load / store address.
4891 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4893 ISD::MemIndexedMode &AM,
4894 SelectionDAG &DAG) const {
4895 if (Subtarget->isThumb1Only())
4900 bool isSEXTLoad = false;
4901 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4902 Ptr = LD->getBasePtr();
4903 VT = LD->getMemoryVT();
4904 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4905 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4906 Ptr = ST->getBasePtr();
4907 VT = ST->getMemoryVT();
4912 bool isLegal = false;
4913 if (Subtarget->isThumb2())
4914 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4915 Offset, isInc, DAG);
4917 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4918 Offset, isInc, DAG);
4922 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4926 /// getPostIndexedAddressParts - returns true by value, base pointer and
4927 /// offset pointer and addressing mode by reference if this node can be
4928 /// combined with a load / store to form a post-indexed load / store.
4929 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4932 ISD::MemIndexedMode &AM,
4933 SelectionDAG &DAG) const {
4934 if (Subtarget->isThumb1Only())
4939 bool isSEXTLoad = false;
4940 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4941 VT = LD->getMemoryVT();
4942 Ptr = LD->getBasePtr();
4943 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4944 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4945 VT = ST->getMemoryVT();
4946 Ptr = ST->getBasePtr();
4951 bool isLegal = false;
4952 if (Subtarget->isThumb2())
4953 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4956 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4962 // Swap base ptr and offset to catch more post-index load / store when
4963 // it's legal. In Thumb2 mode, offset must be an immediate.
4964 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4965 !Subtarget->isThumb2())
4966 std::swap(Base, Offset);
4968 // Post-indexed load / store update the base pointer.
4973 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4977 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4981 const SelectionDAG &DAG,
4982 unsigned Depth) const {
4983 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4984 switch (Op.getOpcode()) {
4986 case ARMISD::CMOV: {
4987 // Bits are known zero/one if known on the LHS and RHS.
4988 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4989 if (KnownZero == 0 && KnownOne == 0) return;
4991 APInt KnownZeroRHS, KnownOneRHS;
4992 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4993 KnownZeroRHS, KnownOneRHS, Depth+1);
4994 KnownZero &= KnownZeroRHS;
4995 KnownOne &= KnownOneRHS;
5001 //===----------------------------------------------------------------------===//
5002 // ARM Inline Assembly Support
5003 //===----------------------------------------------------------------------===//
5005 /// getConstraintType - Given a constraint letter, return the type of
5006 /// constraint it is for this target.
5007 ARMTargetLowering::ConstraintType
5008 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5009 if (Constraint.size() == 1) {
5010 switch (Constraint[0]) {
5012 case 'l': return C_RegisterClass;
5013 case 'w': return C_RegisterClass;
5016 return TargetLowering::getConstraintType(Constraint);
5019 std::pair<unsigned, const TargetRegisterClass*>
5020 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5022 if (Constraint.size() == 1) {
5023 // GCC ARM Constraint Letters
5024 switch (Constraint[0]) {
5026 if (Subtarget->isThumb())
5027 return std::make_pair(0U, ARM::tGPRRegisterClass);
5029 return std::make_pair(0U, ARM::GPRRegisterClass);
5031 return std::make_pair(0U, ARM::GPRRegisterClass);
5034 return std::make_pair(0U, ARM::SPRRegisterClass);
5035 if (VT.getSizeInBits() == 64)
5036 return std::make_pair(0U, ARM::DPRRegisterClass);
5037 if (VT.getSizeInBits() == 128)
5038 return std::make_pair(0U, ARM::QPRRegisterClass);
5042 if (StringRef("{cc}").equals_lower(Constraint))
5043 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5045 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5048 std::vector<unsigned> ARMTargetLowering::
5049 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5051 if (Constraint.size() != 1)
5052 return std::vector<unsigned>();
5054 switch (Constraint[0]) { // GCC ARM Constraint Letters
5057 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5058 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5061 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5062 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5063 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5064 ARM::R12, ARM::LR, 0);
5067 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5068 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5069 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5070 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5071 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5072 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5073 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5074 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5075 if (VT.getSizeInBits() == 64)
5076 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5077 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5078 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5079 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5080 if (VT.getSizeInBits() == 128)
5081 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5082 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5086 return std::vector<unsigned>();
5089 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5090 /// vector. If it is invalid, don't add anything to Ops.
5091 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5093 std::vector<SDValue>&Ops,
5094 SelectionDAG &DAG) const {
5095 SDValue Result(0, 0);
5097 switch (Constraint) {
5099 case 'I': case 'J': case 'K': case 'L':
5100 case 'M': case 'N': case 'O':
5101 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5105 int64_t CVal64 = C->getSExtValue();
5106 int CVal = (int) CVal64;
5107 // None of these constraints allow values larger than 32 bits. Check
5108 // that the value fits in an int.
5112 switch (Constraint) {
5114 if (Subtarget->isThumb1Only()) {
5115 // This must be a constant between 0 and 255, for ADD
5117 if (CVal >= 0 && CVal <= 255)
5119 } else if (Subtarget->isThumb2()) {
5120 // A constant that can be used as an immediate value in a
5121 // data-processing instruction.
5122 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5125 // A constant that can be used as an immediate value in a
5126 // data-processing instruction.
5127 if (ARM_AM::getSOImmVal(CVal) != -1)
5133 if (Subtarget->isThumb()) { // FIXME thumb2
5134 // This must be a constant between -255 and -1, for negated ADD
5135 // immediates. This can be used in GCC with an "n" modifier that
5136 // prints the negated value, for use with SUB instructions. It is
5137 // not useful otherwise but is implemented for compatibility.
5138 if (CVal >= -255 && CVal <= -1)
5141 // This must be a constant between -4095 and 4095. It is not clear
5142 // what this constraint is intended for. Implemented for
5143 // compatibility with GCC.
5144 if (CVal >= -4095 && CVal <= 4095)
5150 if (Subtarget->isThumb1Only()) {
5151 // A 32-bit value where only one byte has a nonzero value. Exclude
5152 // zero to match GCC. This constraint is used by GCC internally for
5153 // constants that can be loaded with a move/shift combination.
5154 // It is not useful otherwise but is implemented for compatibility.
5155 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5157 } else if (Subtarget->isThumb2()) {
5158 // A constant whose bitwise inverse can be used as an immediate
5159 // value in a data-processing instruction. This can be used in GCC
5160 // with a "B" modifier that prints the inverted value, for use with
5161 // BIC and MVN instructions. It is not useful otherwise but is
5162 // implemented for compatibility.
5163 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5166 // A constant whose bitwise inverse can be used as an immediate
5167 // value in a data-processing instruction. This can be used in GCC
5168 // with a "B" modifier that prints the inverted value, for use with
5169 // BIC and MVN instructions. It is not useful otherwise but is
5170 // implemented for compatibility.
5171 if (ARM_AM::getSOImmVal(~CVal) != -1)
5177 if (Subtarget->isThumb1Only()) {
5178 // This must be a constant between -7 and 7,
5179 // for 3-operand ADD/SUB immediate instructions.
5180 if (CVal >= -7 && CVal < 7)
5182 } else if (Subtarget->isThumb2()) {
5183 // A constant whose negation can be used as an immediate value in a
5184 // data-processing instruction. This can be used in GCC with an "n"
5185 // modifier that prints the negated value, for use with SUB
5186 // instructions. It is not useful otherwise but is implemented for
5188 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5191 // A constant whose negation can be used as an immediate value in a
5192 // data-processing instruction. This can be used in GCC with an "n"
5193 // modifier that prints the negated value, for use with SUB
5194 // instructions. It is not useful otherwise but is implemented for
5196 if (ARM_AM::getSOImmVal(-CVal) != -1)
5202 if (Subtarget->isThumb()) { // FIXME thumb2
5203 // This must be a multiple of 4 between 0 and 1020, for
5204 // ADD sp + immediate.
5205 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5208 // A power of two or a constant between 0 and 32. This is used in
5209 // GCC for the shift amount on shifted register operands, but it is
5210 // useful in general for any shift amounts.
5211 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5217 if (Subtarget->isThumb()) { // FIXME thumb2
5218 // This must be a constant between 0 and 31, for shift amounts.
5219 if (CVal >= 0 && CVal <= 31)
5225 if (Subtarget->isThumb()) { // FIXME thumb2
5226 // This must be a multiple of 4 between -508 and 508, for
5227 // ADD/SUB sp = sp + immediate.
5228 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5233 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5237 if (Result.getNode()) {
5238 Ops.push_back(Result);
5241 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5245 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5246 // The ARM target isn't yet aware of offsets.
5250 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5251 APInt Imm = FPImm.bitcastToAPInt();
5252 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5253 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5254 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5256 // We can handle 4 bits of mantissa.
5257 // mantissa = (16+UInt(e:f:g:h))/16.
5258 if (Mantissa & 0x7ffff)
5261 if ((Mantissa & 0xf) != Mantissa)
5264 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5265 if (Exp < -3 || Exp > 4)
5267 Exp = ((Exp+3) & 0x7) ^ 4;
5269 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5272 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5273 APInt Imm = FPImm.bitcastToAPInt();
5274 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5275 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5276 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5278 // We can handle 4 bits of mantissa.
5279 // mantissa = (16+UInt(e:f:g:h))/16.
5280 if (Mantissa & 0xffffffffffffLL)
5283 if ((Mantissa & 0xf) != Mantissa)
5286 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5287 if (Exp < -3 || Exp > 4)
5289 Exp = ((Exp+3) & 0x7) ^ 4;
5291 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5294 /// isFPImmLegal - Returns true if the target can instruction select the
5295 /// specified FP immediate natively. If false, the legalizer will
5296 /// materialize the FP immediate as a load from a constant pool.
5297 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5298 if (!Subtarget->hasVFP3())
5301 return ARM::getVFPf32Imm(Imm) != -1;
5303 return ARM::getVFPf64Imm(Imm) != -1;