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 if (Subtarget->hasV6T2Ops())
534 setTargetDAGCombine(ISD::OR);
536 setStackPointerRegisterToSaveRestore(ARM::SP);
538 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
539 setSchedulingPreference(Sched::RegPressure);
541 setSchedulingPreference(Sched::Hybrid);
543 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
545 // On ARM arguments smaller than 4 bytes are extended, so all arguments
546 // are at least 4 bytes aligned.
547 setMinStackArgumentAlignment(4);
549 if (EnableARMCodePlacement)
550 benefitFromCodePlacementOpt = true;
553 std::pair<const TargetRegisterClass*, uint8_t>
554 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
555 const TargetRegisterClass *RRC = 0;
557 switch (VT.getSimpleVT().SimpleTy) {
559 return TargetLowering::findRepresentativeClass(VT);
560 // Use SPR as representative register class for all floating point
563 RRC = ARM::SPRRegisterClass;
565 case MVT::f64: case MVT::v8i8: case MVT::v4i16:
566 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
567 RRC = ARM::SPRRegisterClass;
570 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
571 case MVT::v4f32: case MVT::v2f64:
572 RRC = ARM::SPRRegisterClass;
576 RRC = ARM::SPRRegisterClass;
580 RRC = ARM::SPRRegisterClass;
584 return std::make_pair(RRC, Cost);
587 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
590 case ARMISD::Wrapper: return "ARMISD::Wrapper";
591 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
592 case ARMISD::CALL: return "ARMISD::CALL";
593 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
594 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
595 case ARMISD::tCALL: return "ARMISD::tCALL";
596 case ARMISD::BRCOND: return "ARMISD::BRCOND";
597 case ARMISD::BR_JT: return "ARMISD::BR_JT";
598 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
599 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
600 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
601 case ARMISD::CMP: return "ARMISD::CMP";
602 case ARMISD::CMPZ: return "ARMISD::CMPZ";
603 case ARMISD::CMPFP: return "ARMISD::CMPFP";
604 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
605 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
606 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
607 case ARMISD::CMOV: return "ARMISD::CMOV";
608 case ARMISD::CNEG: return "ARMISD::CNEG";
610 case ARMISD::RBIT: return "ARMISD::RBIT";
612 case ARMISD::FTOSI: return "ARMISD::FTOSI";
613 case ARMISD::FTOUI: return "ARMISD::FTOUI";
614 case ARMISD::SITOF: return "ARMISD::SITOF";
615 case ARMISD::UITOF: return "ARMISD::UITOF";
617 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
618 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
619 case ARMISD::RRX: return "ARMISD::RRX";
621 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
622 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
624 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
625 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
627 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
629 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
631 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
633 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
634 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
636 case ARMISD::VCEQ: return "ARMISD::VCEQ";
637 case ARMISD::VCGE: return "ARMISD::VCGE";
638 case ARMISD::VCGEU: return "ARMISD::VCGEU";
639 case ARMISD::VCGT: return "ARMISD::VCGT";
640 case ARMISD::VCGTU: return "ARMISD::VCGTU";
641 case ARMISD::VTST: return "ARMISD::VTST";
643 case ARMISD::VSHL: return "ARMISD::VSHL";
644 case ARMISD::VSHRs: return "ARMISD::VSHRs";
645 case ARMISD::VSHRu: return "ARMISD::VSHRu";
646 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
647 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
648 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
649 case ARMISD::VSHRN: return "ARMISD::VSHRN";
650 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
651 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
652 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
653 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
654 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
655 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
656 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
657 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
658 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
659 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
660 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
661 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
662 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
663 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
664 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
665 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
666 case ARMISD::VDUP: return "ARMISD::VDUP";
667 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
668 case ARMISD::VEXT: return "ARMISD::VEXT";
669 case ARMISD::VREV64: return "ARMISD::VREV64";
670 case ARMISD::VREV32: return "ARMISD::VREV32";
671 case ARMISD::VREV16: return "ARMISD::VREV16";
672 case ARMISD::VZIP: return "ARMISD::VZIP";
673 case ARMISD::VUZP: return "ARMISD::VUZP";
674 case ARMISD::VTRN: return "ARMISD::VTRN";
675 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
676 case ARMISD::FMAX: return "ARMISD::FMAX";
677 case ARMISD::FMIN: return "ARMISD::FMIN";
678 case ARMISD::BFI: return "ARMISD::BFI";
682 /// getRegClassFor - Return the register class that should be used for the
683 /// specified value type.
684 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
685 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
686 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
687 // load / store 4 to 8 consecutive D registers.
688 if (Subtarget->hasNEON()) {
689 if (VT == MVT::v4i64)
690 return ARM::QQPRRegisterClass;
691 else if (VT == MVT::v8i64)
692 return ARM::QQQQPRRegisterClass;
694 return TargetLowering::getRegClassFor(VT);
697 /// getFunctionAlignment - Return the Log2 alignment of this function.
698 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
699 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
702 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
703 unsigned NumVals = N->getNumValues();
705 return Sched::RegPressure;
707 for (unsigned i = 0; i != NumVals; ++i) {
708 EVT VT = N->getValueType(i);
709 if (VT.isFloatingPoint() || VT.isVector())
710 return Sched::Latency;
713 if (!N->isMachineOpcode())
714 return Sched::RegPressure;
716 // Load are scheduled for latency even if there instruction itinerary
718 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
719 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
721 return Sched::Latency;
723 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
724 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
725 return Sched::Latency;
726 return Sched::RegPressure;
729 //===----------------------------------------------------------------------===//
731 //===----------------------------------------------------------------------===//
733 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
734 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
736 default: llvm_unreachable("Unknown condition code!");
737 case ISD::SETNE: return ARMCC::NE;
738 case ISD::SETEQ: return ARMCC::EQ;
739 case ISD::SETGT: return ARMCC::GT;
740 case ISD::SETGE: return ARMCC::GE;
741 case ISD::SETLT: return ARMCC::LT;
742 case ISD::SETLE: return ARMCC::LE;
743 case ISD::SETUGT: return ARMCC::HI;
744 case ISD::SETUGE: return ARMCC::HS;
745 case ISD::SETULT: return ARMCC::LO;
746 case ISD::SETULE: return ARMCC::LS;
750 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
751 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
752 ARMCC::CondCodes &CondCode2) {
753 CondCode2 = ARMCC::AL;
755 default: llvm_unreachable("Unknown FP condition!");
757 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
759 case ISD::SETOGT: CondCode = ARMCC::GT; break;
761 case ISD::SETOGE: CondCode = ARMCC::GE; break;
762 case ISD::SETOLT: CondCode = ARMCC::MI; break;
763 case ISD::SETOLE: CondCode = ARMCC::LS; break;
764 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
765 case ISD::SETO: CondCode = ARMCC::VC; break;
766 case ISD::SETUO: CondCode = ARMCC::VS; break;
767 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
768 case ISD::SETUGT: CondCode = ARMCC::HI; break;
769 case ISD::SETUGE: CondCode = ARMCC::PL; break;
771 case ISD::SETULT: CondCode = ARMCC::LT; break;
773 case ISD::SETULE: CondCode = ARMCC::LE; break;
775 case ISD::SETUNE: CondCode = ARMCC::NE; break;
779 //===----------------------------------------------------------------------===//
780 // Calling Convention Implementation
781 //===----------------------------------------------------------------------===//
783 #include "ARMGenCallingConv.inc"
785 // APCS f64 is in register pairs, possibly split to stack
786 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
787 CCValAssign::LocInfo &LocInfo,
788 CCState &State, bool CanFail) {
789 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
791 // Try to get the first register.
792 if (unsigned Reg = State.AllocateReg(RegList, 4))
793 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
795 // For the 2nd half of a v2f64, do not fail.
799 // Put the whole thing on the stack.
800 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
801 State.AllocateStack(8, 4),
806 // Try to get the second register.
807 if (unsigned Reg = State.AllocateReg(RegList, 4))
808 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
810 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
811 State.AllocateStack(4, 4),
816 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
817 CCValAssign::LocInfo &LocInfo,
818 ISD::ArgFlagsTy &ArgFlags,
820 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
822 if (LocVT == MVT::v2f64 &&
823 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
825 return true; // we handled it
828 // AAPCS f64 is in aligned register pairs
829 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
830 CCValAssign::LocInfo &LocInfo,
831 CCState &State, bool CanFail) {
832 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
833 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
835 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
837 // For the 2nd half of a v2f64, do not just fail.
841 // Put the whole thing on the stack.
842 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
843 State.AllocateStack(8, 8),
849 for (i = 0; i < 2; ++i)
850 if (HiRegList[i] == Reg)
853 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
854 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
859 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
860 CCValAssign::LocInfo &LocInfo,
861 ISD::ArgFlagsTy &ArgFlags,
863 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
865 if (LocVT == MVT::v2f64 &&
866 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
868 return true; // we handled it
871 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
872 CCValAssign::LocInfo &LocInfo, CCState &State) {
873 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
874 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
876 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
878 return false; // we didn't handle it
881 for (i = 0; i < 2; ++i)
882 if (HiRegList[i] == Reg)
885 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
886 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
891 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
892 CCValAssign::LocInfo &LocInfo,
893 ISD::ArgFlagsTy &ArgFlags,
895 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
897 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
899 return true; // we handled it
902 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
903 CCValAssign::LocInfo &LocInfo,
904 ISD::ArgFlagsTy &ArgFlags,
906 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
910 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
911 /// given CallingConvention value.
912 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
914 bool isVarArg) const {
917 llvm_unreachable("Unsupported calling convention");
919 case CallingConv::Fast:
920 // Use target triple & subtarget features to do actual dispatch.
921 if (Subtarget->isAAPCS_ABI()) {
922 if (Subtarget->hasVFP2() &&
923 FloatABIType == FloatABI::Hard && !isVarArg)
924 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
926 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
928 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
929 case CallingConv::ARM_AAPCS_VFP:
930 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
931 case CallingConv::ARM_AAPCS:
932 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
933 case CallingConv::ARM_APCS:
934 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
938 /// LowerCallResult - Lower the result values of a call into the
939 /// appropriate copies out of appropriate physical registers.
941 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
942 CallingConv::ID CallConv, bool isVarArg,
943 const SmallVectorImpl<ISD::InputArg> &Ins,
944 DebugLoc dl, SelectionDAG &DAG,
945 SmallVectorImpl<SDValue> &InVals) const {
947 // Assign locations to each value returned by this call.
948 SmallVector<CCValAssign, 16> RVLocs;
949 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
950 RVLocs, *DAG.getContext());
951 CCInfo.AnalyzeCallResult(Ins,
952 CCAssignFnForNode(CallConv, /* Return*/ true,
955 // Copy all of the result registers out of their specified physreg.
956 for (unsigned i = 0; i != RVLocs.size(); ++i) {
957 CCValAssign VA = RVLocs[i];
960 if (VA.needsCustom()) {
961 // Handle f64 or half of a v2f64.
962 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
964 Chain = Lo.getValue(1);
965 InFlag = Lo.getValue(2);
966 VA = RVLocs[++i]; // skip ahead to next loc
967 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
969 Chain = Hi.getValue(1);
970 InFlag = Hi.getValue(2);
971 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
973 if (VA.getLocVT() == MVT::v2f64) {
974 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
975 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
976 DAG.getConstant(0, MVT::i32));
978 VA = RVLocs[++i]; // skip ahead to next loc
979 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
980 Chain = Lo.getValue(1);
981 InFlag = Lo.getValue(2);
982 VA = RVLocs[++i]; // skip ahead to next loc
983 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
984 Chain = Hi.getValue(1);
985 InFlag = Hi.getValue(2);
986 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
987 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
988 DAG.getConstant(1, MVT::i32));
991 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
993 Chain = Val.getValue(1);
994 InFlag = Val.getValue(2);
997 switch (VA.getLocInfo()) {
998 default: llvm_unreachable("Unknown loc info!");
999 case CCValAssign::Full: break;
1000 case CCValAssign::BCvt:
1001 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1005 InVals.push_back(Val);
1011 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1012 /// by "Src" to address "Dst" of size "Size". Alignment information is
1013 /// specified by the specific parameter attribute. The copy will be passed as
1014 /// a byval function parameter.
1015 /// Sometimes what we are copying is the end of a larger object, the part that
1016 /// does not fit in registers.
1018 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1019 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1021 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1022 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1023 /*isVolatile=*/false, /*AlwaysInline=*/false,
1027 /// LowerMemOpCallTo - Store the argument to the stack.
1029 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1030 SDValue StackPtr, SDValue Arg,
1031 DebugLoc dl, SelectionDAG &DAG,
1032 const CCValAssign &VA,
1033 ISD::ArgFlagsTy Flags) const {
1034 unsigned LocMemOffset = VA.getLocMemOffset();
1035 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1036 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1037 if (Flags.isByVal()) {
1038 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1040 return DAG.getStore(Chain, dl, Arg, PtrOff,
1041 PseudoSourceValue::getStack(), LocMemOffset,
1045 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1046 SDValue Chain, SDValue &Arg,
1047 RegsToPassVector &RegsToPass,
1048 CCValAssign &VA, CCValAssign &NextVA,
1050 SmallVector<SDValue, 8> &MemOpChains,
1051 ISD::ArgFlagsTy Flags) const {
1053 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1054 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1055 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1057 if (NextVA.isRegLoc())
1058 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1060 assert(NextVA.isMemLoc());
1061 if (StackPtr.getNode() == 0)
1062 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1064 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1070 /// LowerCall - Lowering a call into a callseq_start <-
1071 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1074 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1075 CallingConv::ID CallConv, bool isVarArg,
1077 const SmallVectorImpl<ISD::OutputArg> &Outs,
1078 const SmallVectorImpl<SDValue> &OutVals,
1079 const SmallVectorImpl<ISD::InputArg> &Ins,
1080 DebugLoc dl, SelectionDAG &DAG,
1081 SmallVectorImpl<SDValue> &InVals) const {
1082 MachineFunction &MF = DAG.getMachineFunction();
1083 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1084 bool IsSibCall = false;
1085 // Temporarily disable tail calls so things don't break.
1086 if (!EnableARMTailCalls)
1089 // Check if it's really possible to do a tail call.
1090 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1091 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1092 Outs, OutVals, Ins, DAG);
1093 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1094 // detected sibcalls.
1101 // Analyze operands of the call, assigning locations to each operand.
1102 SmallVector<CCValAssign, 16> ArgLocs;
1103 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1105 CCInfo.AnalyzeCallOperands(Outs,
1106 CCAssignFnForNode(CallConv, /* Return*/ false,
1109 // Get a count of how many bytes are to be pushed on the stack.
1110 unsigned NumBytes = CCInfo.getNextStackOffset();
1112 // For tail calls, memory operands are available in our caller's stack.
1116 // Adjust the stack pointer for the new arguments...
1117 // These operations are automatically eliminated by the prolog/epilog pass
1119 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1121 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1123 RegsToPassVector RegsToPass;
1124 SmallVector<SDValue, 8> MemOpChains;
1126 // Walk the register/memloc assignments, inserting copies/loads. In the case
1127 // of tail call optimization, arguments are handled later.
1128 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1130 ++i, ++realArgIdx) {
1131 CCValAssign &VA = ArgLocs[i];
1132 SDValue Arg = OutVals[realArgIdx];
1133 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1135 // Promote the value if needed.
1136 switch (VA.getLocInfo()) {
1137 default: llvm_unreachable("Unknown loc info!");
1138 case CCValAssign::Full: break;
1139 case CCValAssign::SExt:
1140 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1142 case CCValAssign::ZExt:
1143 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1145 case CCValAssign::AExt:
1146 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1148 case CCValAssign::BCvt:
1149 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1153 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1154 if (VA.needsCustom()) {
1155 if (VA.getLocVT() == MVT::v2f64) {
1156 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1157 DAG.getConstant(0, MVT::i32));
1158 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1159 DAG.getConstant(1, MVT::i32));
1161 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1162 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1164 VA = ArgLocs[++i]; // skip ahead to next loc
1165 if (VA.isRegLoc()) {
1166 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1167 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1169 assert(VA.isMemLoc());
1171 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1172 dl, DAG, VA, Flags));
1175 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1176 StackPtr, MemOpChains, Flags);
1178 } else if (VA.isRegLoc()) {
1179 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1180 } else if (!IsSibCall) {
1181 assert(VA.isMemLoc());
1183 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1184 dl, DAG, VA, Flags));
1188 if (!MemOpChains.empty())
1189 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1190 &MemOpChains[0], MemOpChains.size());
1192 // Build a sequence of copy-to-reg nodes chained together with token chain
1193 // and flag operands which copy the outgoing args into the appropriate regs.
1195 // Tail call byval lowering might overwrite argument registers so in case of
1196 // tail call optimization the copies to registers are lowered later.
1198 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1199 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1200 RegsToPass[i].second, InFlag);
1201 InFlag = Chain.getValue(1);
1204 // For tail calls lower the arguments to the 'real' stack slot.
1206 // Force all the incoming stack arguments to be loaded from the stack
1207 // before any new outgoing arguments are stored to the stack, because the
1208 // outgoing stack slots may alias the incoming argument stack slots, and
1209 // the alias isn't otherwise explicit. This is slightly more conservative
1210 // than necessary, because it means that each store effectively depends
1211 // on every argument instead of just those arguments it would clobber.
1213 // Do not flag preceeding copytoreg stuff together with the following stuff.
1215 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1216 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1217 RegsToPass[i].second, InFlag);
1218 InFlag = Chain.getValue(1);
1223 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1224 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1225 // node so that legalize doesn't hack it.
1226 bool isDirect = false;
1227 bool isARMFunc = false;
1228 bool isLocalARMFunc = false;
1229 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1231 if (EnableARMLongCalls) {
1232 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1233 && "long-calls with non-static relocation model!");
1234 // Handle a global address or an external symbol. If it's not one of
1235 // those, the target's already in a register, so we don't need to do
1237 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1238 const GlobalValue *GV = G->getGlobal();
1239 // Create a constant pool entry for the callee address
1240 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1241 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1244 // Get the address of the callee into a register
1245 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1246 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1247 Callee = DAG.getLoad(getPointerTy(), dl,
1248 DAG.getEntryNode(), CPAddr,
1249 PseudoSourceValue::getConstantPool(), 0,
1251 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1252 const char *Sym = S->getSymbol();
1254 // Create a constant pool entry for the callee address
1255 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1256 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1257 Sym, ARMPCLabelIndex, 0);
1258 // Get the address of the callee into a register
1259 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1260 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1261 Callee = DAG.getLoad(getPointerTy(), dl,
1262 DAG.getEntryNode(), CPAddr,
1263 PseudoSourceValue::getConstantPool(), 0,
1266 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1267 const GlobalValue *GV = G->getGlobal();
1269 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1270 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1271 getTargetMachine().getRelocationModel() != Reloc::Static;
1272 isARMFunc = !Subtarget->isThumb() || isStub;
1273 // ARM call to a local ARM function is predicable.
1274 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1275 // tBX takes a register source operand.
1276 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1277 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1278 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1281 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1282 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1283 Callee = DAG.getLoad(getPointerTy(), dl,
1284 DAG.getEntryNode(), CPAddr,
1285 PseudoSourceValue::getConstantPool(), 0,
1287 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1288 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1289 getPointerTy(), Callee, PICLabel);
1291 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1292 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1294 bool isStub = Subtarget->isTargetDarwin() &&
1295 getTargetMachine().getRelocationModel() != Reloc::Static;
1296 isARMFunc = !Subtarget->isThumb() || isStub;
1297 // tBX takes a register source operand.
1298 const char *Sym = S->getSymbol();
1299 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1300 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1301 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1302 Sym, ARMPCLabelIndex, 4);
1303 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1304 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1305 Callee = DAG.getLoad(getPointerTy(), dl,
1306 DAG.getEntryNode(), CPAddr,
1307 PseudoSourceValue::getConstantPool(), 0,
1309 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1310 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1311 getPointerTy(), Callee, PICLabel);
1313 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1316 // FIXME: handle tail calls differently.
1318 if (Subtarget->isThumb()) {
1319 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1320 CallOpc = ARMISD::CALL_NOLINK;
1322 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1324 CallOpc = (isDirect || Subtarget->hasV5TOps())
1325 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1326 : ARMISD::CALL_NOLINK;
1329 std::vector<SDValue> Ops;
1330 Ops.push_back(Chain);
1331 Ops.push_back(Callee);
1333 // Add argument registers to the end of the list so that they are known live
1335 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1336 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1337 RegsToPass[i].second.getValueType()));
1339 if (InFlag.getNode())
1340 Ops.push_back(InFlag);
1342 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1344 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1346 // Returns a chain and a flag for retval copy to use.
1347 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1348 InFlag = Chain.getValue(1);
1350 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1351 DAG.getIntPtrConstant(0, true), InFlag);
1353 InFlag = Chain.getValue(1);
1355 // Handle result values, copying them out of physregs into vregs that we
1357 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1361 /// MatchingStackOffset - Return true if the given stack call argument is
1362 /// already available in the same position (relatively) of the caller's
1363 /// incoming argument stack.
1365 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1366 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1367 const ARMInstrInfo *TII) {
1368 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1370 if (Arg.getOpcode() == ISD::CopyFromReg) {
1371 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1372 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1374 MachineInstr *Def = MRI->getVRegDef(VR);
1377 if (!Flags.isByVal()) {
1378 if (!TII->isLoadFromStackSlot(Def, FI))
1383 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1384 if (Flags.isByVal())
1385 // ByVal argument is passed in as a pointer but it's now being
1386 // dereferenced. e.g.
1387 // define @foo(%struct.X* %A) {
1388 // tail call @bar(%struct.X* byval %A)
1391 SDValue Ptr = Ld->getBasePtr();
1392 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1395 FI = FINode->getIndex();
1399 assert(FI != INT_MAX);
1400 if (!MFI->isFixedObjectIndex(FI))
1402 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1405 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1406 /// for tail call optimization. Targets which want to do tail call
1407 /// optimization should implement this function.
1409 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1410 CallingConv::ID CalleeCC,
1412 bool isCalleeStructRet,
1413 bool isCallerStructRet,
1414 const SmallVectorImpl<ISD::OutputArg> &Outs,
1415 const SmallVectorImpl<SDValue> &OutVals,
1416 const SmallVectorImpl<ISD::InputArg> &Ins,
1417 SelectionDAG& DAG) const {
1418 const Function *CallerF = DAG.getMachineFunction().getFunction();
1419 CallingConv::ID CallerCC = CallerF->getCallingConv();
1420 bool CCMatch = CallerCC == CalleeCC;
1422 // Look for obvious safe cases to perform tail call optimization that do not
1423 // require ABI changes. This is what gcc calls sibcall.
1425 // Do not sibcall optimize vararg calls unless the call site is not passing
1427 if (isVarArg && !Outs.empty())
1430 // Also avoid sibcall optimization if either caller or callee uses struct
1431 // return semantics.
1432 if (isCalleeStructRet || isCallerStructRet)
1435 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1436 // emitEpilogue is not ready for them.
1437 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1438 // LR. This means if we need to reload LR, it takes an extra instructions,
1439 // which outweighs the value of the tail call; but here we don't know yet
1440 // whether LR is going to be used. Probably the right approach is to
1441 // generate the tail call here and turn it back into CALL/RET in
1442 // emitEpilogue if LR is used.
1443 if (Subtarget->isThumb1Only())
1446 // For the moment, we can only do this to functions defined in this
1447 // compilation, or to indirect calls. A Thumb B to an ARM function,
1448 // or vice versa, is not easily fixed up in the linker unlike BL.
1449 // (We could do this by loading the address of the callee into a register;
1450 // that is an extra instruction over the direct call and burns a register
1451 // as well, so is not likely to be a win.)
1453 // It might be safe to remove this restriction on non-Darwin.
1455 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1456 // but we need to make sure there are enough registers; the only valid
1457 // registers are the 4 used for parameters. We don't currently do this
1459 if (isa<ExternalSymbolSDNode>(Callee))
1462 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1463 const GlobalValue *GV = G->getGlobal();
1464 if (GV->isDeclaration() || GV->isWeakForLinker())
1468 // If the calling conventions do not match, then we'd better make sure the
1469 // results are returned in the same way as what the caller expects.
1471 SmallVector<CCValAssign, 16> RVLocs1;
1472 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1473 RVLocs1, *DAG.getContext());
1474 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1476 SmallVector<CCValAssign, 16> RVLocs2;
1477 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1478 RVLocs2, *DAG.getContext());
1479 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1481 if (RVLocs1.size() != RVLocs2.size())
1483 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1484 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1486 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1488 if (RVLocs1[i].isRegLoc()) {
1489 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1492 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1498 // If the callee takes no arguments then go on to check the results of the
1500 if (!Outs.empty()) {
1501 // Check if stack adjustment is needed. For now, do not do this if any
1502 // argument is passed on the stack.
1503 SmallVector<CCValAssign, 16> ArgLocs;
1504 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1505 ArgLocs, *DAG.getContext());
1506 CCInfo.AnalyzeCallOperands(Outs,
1507 CCAssignFnForNode(CalleeCC, false, isVarArg));
1508 if (CCInfo.getNextStackOffset()) {
1509 MachineFunction &MF = DAG.getMachineFunction();
1511 // Check if the arguments are already laid out in the right way as
1512 // the caller's fixed stack objects.
1513 MachineFrameInfo *MFI = MF.getFrameInfo();
1514 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1515 const ARMInstrInfo *TII =
1516 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1517 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1519 ++i, ++realArgIdx) {
1520 CCValAssign &VA = ArgLocs[i];
1521 EVT RegVT = VA.getLocVT();
1522 SDValue Arg = OutVals[realArgIdx];
1523 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1524 if (VA.getLocInfo() == CCValAssign::Indirect)
1526 if (VA.needsCustom()) {
1527 // f64 and vector types are split into multiple registers or
1528 // register/stack-slot combinations. The types will not match
1529 // the registers; give up on memory f64 refs until we figure
1530 // out what to do about this.
1533 if (!ArgLocs[++i].isRegLoc())
1535 if (RegVT == MVT::v2f64) {
1536 if (!ArgLocs[++i].isRegLoc())
1538 if (!ArgLocs[++i].isRegLoc())
1541 } else if (!VA.isRegLoc()) {
1542 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1554 ARMTargetLowering::LowerReturn(SDValue Chain,
1555 CallingConv::ID CallConv, bool isVarArg,
1556 const SmallVectorImpl<ISD::OutputArg> &Outs,
1557 const SmallVectorImpl<SDValue> &OutVals,
1558 DebugLoc dl, SelectionDAG &DAG) const {
1560 // CCValAssign - represent the assignment of the return value to a location.
1561 SmallVector<CCValAssign, 16> RVLocs;
1563 // CCState - Info about the registers and stack slots.
1564 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1567 // Analyze outgoing return values.
1568 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1571 // If this is the first return lowered for this function, add
1572 // the regs to the liveout set for the function.
1573 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1574 for (unsigned i = 0; i != RVLocs.size(); ++i)
1575 if (RVLocs[i].isRegLoc())
1576 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1581 // Copy the result values into the output registers.
1582 for (unsigned i = 0, realRVLocIdx = 0;
1584 ++i, ++realRVLocIdx) {
1585 CCValAssign &VA = RVLocs[i];
1586 assert(VA.isRegLoc() && "Can only return in registers!");
1588 SDValue Arg = OutVals[realRVLocIdx];
1590 switch (VA.getLocInfo()) {
1591 default: llvm_unreachable("Unknown loc info!");
1592 case CCValAssign::Full: break;
1593 case CCValAssign::BCvt:
1594 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1598 if (VA.needsCustom()) {
1599 if (VA.getLocVT() == MVT::v2f64) {
1600 // Extract the first half and return it in two registers.
1601 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1602 DAG.getConstant(0, MVT::i32));
1603 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1604 DAG.getVTList(MVT::i32, MVT::i32), Half);
1606 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1607 Flag = Chain.getValue(1);
1608 VA = RVLocs[++i]; // skip ahead to next loc
1609 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1610 HalfGPRs.getValue(1), Flag);
1611 Flag = Chain.getValue(1);
1612 VA = RVLocs[++i]; // skip ahead to next loc
1614 // Extract the 2nd half and fall through to handle it as an f64 value.
1615 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1616 DAG.getConstant(1, MVT::i32));
1618 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1620 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1621 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1622 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1623 Flag = Chain.getValue(1);
1624 VA = RVLocs[++i]; // skip ahead to next loc
1625 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1628 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1630 // Guarantee that all emitted copies are
1631 // stuck together, avoiding something bad.
1632 Flag = Chain.getValue(1);
1637 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1639 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1644 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1645 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1646 // one of the above mentioned nodes. It has to be wrapped because otherwise
1647 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1648 // be used to form addressing mode. These wrapped nodes will be selected
1650 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1651 EVT PtrVT = Op.getValueType();
1652 // FIXME there is no actual debug info here
1653 DebugLoc dl = Op.getDebugLoc();
1654 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1656 if (CP->isMachineConstantPoolEntry())
1657 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1658 CP->getAlignment());
1660 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1661 CP->getAlignment());
1662 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1665 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1666 return MachineJumpTableInfo::EK_Inline;
1669 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1670 SelectionDAG &DAG) const {
1671 MachineFunction &MF = DAG.getMachineFunction();
1672 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1673 unsigned ARMPCLabelIndex = 0;
1674 DebugLoc DL = Op.getDebugLoc();
1675 EVT PtrVT = getPointerTy();
1676 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1677 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1679 if (RelocM == Reloc::Static) {
1680 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1682 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1683 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1684 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1685 ARMCP::CPBlockAddress,
1687 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1689 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1690 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1691 PseudoSourceValue::getConstantPool(), 0,
1693 if (RelocM == Reloc::Static)
1695 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1696 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1699 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1701 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1702 SelectionDAG &DAG) const {
1703 DebugLoc dl = GA->getDebugLoc();
1704 EVT PtrVT = getPointerTy();
1705 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1706 MachineFunction &MF = DAG.getMachineFunction();
1707 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1708 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1709 ARMConstantPoolValue *CPV =
1710 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1711 ARMCP::CPValue, PCAdj, "tlsgd", true);
1712 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1713 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1714 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1715 PseudoSourceValue::getConstantPool(), 0,
1717 SDValue Chain = Argument.getValue(1);
1719 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1720 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1722 // call __tls_get_addr.
1725 Entry.Node = Argument;
1726 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1727 Args.push_back(Entry);
1728 // FIXME: is there useful debug info available here?
1729 std::pair<SDValue, SDValue> CallResult =
1730 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1731 false, false, false, false,
1732 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1733 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1734 return CallResult.first;
1737 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1738 // "local exec" model.
1740 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1741 SelectionDAG &DAG) const {
1742 const GlobalValue *GV = GA->getGlobal();
1743 DebugLoc dl = GA->getDebugLoc();
1745 SDValue Chain = DAG.getEntryNode();
1746 EVT PtrVT = getPointerTy();
1747 // Get the Thread Pointer
1748 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1750 if (GV->isDeclaration()) {
1751 MachineFunction &MF = DAG.getMachineFunction();
1752 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1753 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1754 // Initial exec model.
1755 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1756 ARMConstantPoolValue *CPV =
1757 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1758 ARMCP::CPValue, PCAdj, "gottpoff", true);
1759 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1760 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1761 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1762 PseudoSourceValue::getConstantPool(), 0,
1764 Chain = Offset.getValue(1);
1766 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1767 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1769 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1770 PseudoSourceValue::getConstantPool(), 0,
1774 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1775 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1776 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1777 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1778 PseudoSourceValue::getConstantPool(), 0,
1782 // The address of the thread local variable is the add of the thread
1783 // pointer with the offset of the variable.
1784 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1788 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1789 // TODO: implement the "local dynamic" model
1790 assert(Subtarget->isTargetELF() &&
1791 "TLS not implemented for non-ELF targets");
1792 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1793 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1794 // otherwise use the "Local Exec" TLS Model
1795 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1796 return LowerToTLSGeneralDynamicModel(GA, DAG);
1798 return LowerToTLSExecModels(GA, DAG);
1801 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1802 SelectionDAG &DAG) const {
1803 EVT PtrVT = getPointerTy();
1804 DebugLoc dl = Op.getDebugLoc();
1805 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1806 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1807 if (RelocM == Reloc::PIC_) {
1808 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1809 ARMConstantPoolValue *CPV =
1810 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1811 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1812 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1813 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1815 PseudoSourceValue::getConstantPool(), 0,
1817 SDValue Chain = Result.getValue(1);
1818 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1819 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1821 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1822 PseudoSourceValue::getGOT(), 0,
1826 // If we have T2 ops, we can materialize the address directly via movt/movw
1827 // pair. This is always cheaper.
1828 if (Subtarget->useMovt()) {
1829 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1830 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1832 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1833 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1834 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1835 PseudoSourceValue::getConstantPool(), 0,
1841 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1842 SelectionDAG &DAG) const {
1843 MachineFunction &MF = DAG.getMachineFunction();
1844 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1845 unsigned ARMPCLabelIndex = 0;
1846 EVT PtrVT = getPointerTy();
1847 DebugLoc dl = Op.getDebugLoc();
1848 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1849 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1851 if (RelocM == Reloc::Static)
1852 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1854 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1855 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1856 ARMConstantPoolValue *CPV =
1857 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1858 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1860 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1862 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1863 PseudoSourceValue::getConstantPool(), 0,
1865 SDValue Chain = Result.getValue(1);
1867 if (RelocM == Reloc::PIC_) {
1868 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1869 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1872 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1873 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1874 PseudoSourceValue::getGOT(), 0,
1880 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1881 SelectionDAG &DAG) const {
1882 assert(Subtarget->isTargetELF() &&
1883 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1884 MachineFunction &MF = DAG.getMachineFunction();
1885 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1886 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1887 EVT PtrVT = getPointerTy();
1888 DebugLoc dl = Op.getDebugLoc();
1889 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1890 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1891 "_GLOBAL_OFFSET_TABLE_",
1892 ARMPCLabelIndex, PCAdj);
1893 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1894 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1895 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1896 PseudoSourceValue::getConstantPool(), 0,
1898 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1899 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1903 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1904 DebugLoc dl = Op.getDebugLoc();
1905 SDValue Val = DAG.getConstant(0, MVT::i32);
1906 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1907 Op.getOperand(1), Val);
1911 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1912 DebugLoc dl = Op.getDebugLoc();
1913 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1914 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1918 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1919 const ARMSubtarget *Subtarget) const {
1920 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1921 DebugLoc dl = Op.getDebugLoc();
1923 default: return SDValue(); // Don't custom lower most intrinsics.
1924 case Intrinsic::arm_thread_pointer: {
1925 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1926 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1928 case Intrinsic::eh_sjlj_lsda: {
1929 MachineFunction &MF = DAG.getMachineFunction();
1930 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1931 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1932 EVT PtrVT = getPointerTy();
1933 DebugLoc dl = Op.getDebugLoc();
1934 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1936 unsigned PCAdj = (RelocM != Reloc::PIC_)
1937 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1938 ARMConstantPoolValue *CPV =
1939 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1940 ARMCP::CPLSDA, PCAdj);
1941 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1942 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1944 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1945 PseudoSourceValue::getConstantPool(), 0,
1948 if (RelocM == Reloc::PIC_) {
1949 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1950 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1957 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1958 const ARMSubtarget *Subtarget) {
1959 DebugLoc dl = Op.getDebugLoc();
1960 SDValue Op5 = Op.getOperand(5);
1961 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1962 // v6 and v7 can both handle barriers directly, but need handled a bit
1963 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1965 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1966 if (Subtarget->hasV7Ops())
1967 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1968 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1969 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1970 DAG.getConstant(0, MVT::i32));
1971 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1975 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1976 MachineFunction &MF = DAG.getMachineFunction();
1977 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1979 // vastart just stores the address of the VarArgsFrameIndex slot into the
1980 // memory location argument.
1981 DebugLoc dl = Op.getDebugLoc();
1982 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1983 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1984 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1985 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1990 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1991 SDValue &Root, SelectionDAG &DAG,
1992 DebugLoc dl) const {
1993 MachineFunction &MF = DAG.getMachineFunction();
1994 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1996 TargetRegisterClass *RC;
1997 if (AFI->isThumb1OnlyFunction())
1998 RC = ARM::tGPRRegisterClass;
2000 RC = ARM::GPRRegisterClass;
2002 // Transform the arguments stored in physical registers into virtual ones.
2003 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2004 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2007 if (NextVA.isMemLoc()) {
2008 MachineFrameInfo *MFI = MF.getFrameInfo();
2009 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2011 // Create load node to retrieve arguments from the stack.
2012 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2013 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2014 PseudoSourceValue::getFixedStack(FI), 0,
2017 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2018 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2021 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2025 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2026 CallingConv::ID CallConv, bool isVarArg,
2027 const SmallVectorImpl<ISD::InputArg>
2029 DebugLoc dl, SelectionDAG &DAG,
2030 SmallVectorImpl<SDValue> &InVals)
2033 MachineFunction &MF = DAG.getMachineFunction();
2034 MachineFrameInfo *MFI = MF.getFrameInfo();
2036 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2038 // Assign locations to all of the incoming arguments.
2039 SmallVector<CCValAssign, 16> ArgLocs;
2040 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2042 CCInfo.AnalyzeFormalArguments(Ins,
2043 CCAssignFnForNode(CallConv, /* Return*/ false,
2046 SmallVector<SDValue, 16> ArgValues;
2048 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2049 CCValAssign &VA = ArgLocs[i];
2051 // Arguments stored in registers.
2052 if (VA.isRegLoc()) {
2053 EVT RegVT = VA.getLocVT();
2056 if (VA.needsCustom()) {
2057 // f64 and vector types are split up into multiple registers or
2058 // combinations of registers and stack slots.
2059 if (VA.getLocVT() == MVT::v2f64) {
2060 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2062 VA = ArgLocs[++i]; // skip ahead to next loc
2064 if (VA.isMemLoc()) {
2065 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2066 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2067 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2068 PseudoSourceValue::getFixedStack(FI), 0,
2071 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2074 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2075 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2076 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2077 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2078 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2080 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2083 TargetRegisterClass *RC;
2085 if (RegVT == MVT::f32)
2086 RC = ARM::SPRRegisterClass;
2087 else if (RegVT == MVT::f64)
2088 RC = ARM::DPRRegisterClass;
2089 else if (RegVT == MVT::v2f64)
2090 RC = ARM::QPRRegisterClass;
2091 else if (RegVT == MVT::i32)
2092 RC = (AFI->isThumb1OnlyFunction() ?
2093 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2095 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2097 // Transform the arguments in physical registers into virtual ones.
2098 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2099 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2102 // If this is an 8 or 16-bit value, it is really passed promoted
2103 // to 32 bits. Insert an assert[sz]ext to capture this, then
2104 // truncate to the right size.
2105 switch (VA.getLocInfo()) {
2106 default: llvm_unreachable("Unknown loc info!");
2107 case CCValAssign::Full: break;
2108 case CCValAssign::BCvt:
2109 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2111 case CCValAssign::SExt:
2112 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2113 DAG.getValueType(VA.getValVT()));
2114 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2116 case CCValAssign::ZExt:
2117 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2118 DAG.getValueType(VA.getValVT()));
2119 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2123 InVals.push_back(ArgValue);
2125 } else { // VA.isRegLoc()
2128 assert(VA.isMemLoc());
2129 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2131 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2132 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2134 // Create load nodes to retrieve arguments from the stack.
2135 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2136 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2137 PseudoSourceValue::getFixedStack(FI), 0,
2144 static const unsigned GPRArgRegs[] = {
2145 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2148 unsigned NumGPRs = CCInfo.getFirstUnallocated
2149 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2151 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2152 unsigned VARegSize = (4 - NumGPRs) * 4;
2153 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2154 unsigned ArgOffset = CCInfo.getNextStackOffset();
2155 if (VARegSaveSize) {
2156 // If this function is vararg, store any remaining integer argument regs
2157 // to their spots on the stack so that they may be loaded by deferencing
2158 // the result of va_next.
2159 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2160 AFI->setVarArgsFrameIndex(
2161 MFI->CreateFixedObject(VARegSaveSize,
2162 ArgOffset + VARegSaveSize - VARegSize,
2164 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2167 SmallVector<SDValue, 4> MemOps;
2168 for (; NumGPRs < 4; ++NumGPRs) {
2169 TargetRegisterClass *RC;
2170 if (AFI->isThumb1OnlyFunction())
2171 RC = ARM::tGPRRegisterClass;
2173 RC = ARM::GPRRegisterClass;
2175 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2176 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2178 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2179 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2180 0, false, false, 0);
2181 MemOps.push_back(Store);
2182 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2183 DAG.getConstant(4, getPointerTy()));
2185 if (!MemOps.empty())
2186 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2187 &MemOps[0], MemOps.size());
2189 // This will point to the next argument passed via stack.
2190 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2196 /// isFloatingPointZero - Return true if this is +0.0.
2197 static bool isFloatingPointZero(SDValue Op) {
2198 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2199 return CFP->getValueAPF().isPosZero();
2200 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2201 // Maybe this has already been legalized into the constant pool?
2202 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2203 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2204 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2205 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2206 return CFP->getValueAPF().isPosZero();
2212 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2213 /// the given operands.
2215 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2216 SDValue &ARMcc, SelectionDAG &DAG,
2217 DebugLoc dl) const {
2218 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2219 unsigned C = RHSC->getZExtValue();
2220 if (!isLegalICmpImmediate(C)) {
2221 // Constant does not fit, try adjusting it by one?
2226 if (isLegalICmpImmediate(C-1)) {
2227 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2228 RHS = DAG.getConstant(C-1, MVT::i32);
2233 if (C > 0 && isLegalICmpImmediate(C-1)) {
2234 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2235 RHS = DAG.getConstant(C-1, MVT::i32);
2240 if (isLegalICmpImmediate(C+1)) {
2241 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2242 RHS = DAG.getConstant(C+1, MVT::i32);
2247 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2248 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2249 RHS = DAG.getConstant(C+1, MVT::i32);
2256 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2257 ARMISD::NodeType CompareType;
2260 CompareType = ARMISD::CMP;
2265 CompareType = ARMISD::CMPZ;
2268 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2269 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2272 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2274 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2275 DebugLoc dl) const {
2277 if (!isFloatingPointZero(RHS))
2278 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2280 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2281 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2284 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2285 EVT VT = Op.getValueType();
2286 SDValue LHS = Op.getOperand(0);
2287 SDValue RHS = Op.getOperand(1);
2288 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2289 SDValue TrueVal = Op.getOperand(2);
2290 SDValue FalseVal = Op.getOperand(3);
2291 DebugLoc dl = Op.getDebugLoc();
2293 if (LHS.getValueType() == MVT::i32) {
2295 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2296 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2297 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2300 ARMCC::CondCodes CondCode, CondCode2;
2301 FPCCToARMCC(CC, CondCode, CondCode2);
2303 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2304 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2305 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2306 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2308 if (CondCode2 != ARMCC::AL) {
2309 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2310 // FIXME: Needs another CMP because flag can have but one use.
2311 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2312 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2313 Result, TrueVal, ARMcc2, CCR, Cmp2);
2318 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2319 /// to morph to an integer compare sequence.
2320 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2321 const ARMSubtarget *Subtarget) {
2322 SDNode *N = Op.getNode();
2323 if (!N->hasOneUse())
2324 // Otherwise it requires moving the value from fp to integer registers.
2326 if (!N->getNumValues())
2328 EVT VT = Op.getValueType();
2329 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2330 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2331 // vmrs are very slow, e.g. cortex-a8.
2334 if (isFloatingPointZero(Op)) {
2338 return ISD::isNormalLoad(N);
2341 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2342 if (isFloatingPointZero(Op))
2343 return DAG.getConstant(0, MVT::i32);
2345 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2346 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2347 Ld->getChain(), Ld->getBasePtr(),
2348 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2349 Ld->isVolatile(), Ld->isNonTemporal(),
2350 Ld->getAlignment());
2352 llvm_unreachable("Unknown VFP cmp argument!");
2355 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2356 SDValue &RetVal1, SDValue &RetVal2) {
2357 if (isFloatingPointZero(Op)) {
2358 RetVal1 = DAG.getConstant(0, MVT::i32);
2359 RetVal2 = DAG.getConstant(0, MVT::i32);
2363 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2364 SDValue Ptr = Ld->getBasePtr();
2365 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2366 Ld->getChain(), Ptr,
2367 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2368 Ld->isVolatile(), Ld->isNonTemporal(),
2369 Ld->getAlignment());
2371 EVT PtrType = Ptr.getValueType();
2372 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2373 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2374 PtrType, Ptr, DAG.getConstant(4, PtrType));
2375 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2376 Ld->getChain(), NewPtr,
2377 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2378 Ld->isVolatile(), Ld->isNonTemporal(),
2383 llvm_unreachable("Unknown VFP cmp argument!");
2386 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2387 /// f32 and even f64 comparisons to integer ones.
2389 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2390 SDValue Chain = Op.getOperand(0);
2391 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2392 SDValue LHS = Op.getOperand(2);
2393 SDValue RHS = Op.getOperand(3);
2394 SDValue Dest = Op.getOperand(4);
2395 DebugLoc dl = Op.getDebugLoc();
2397 bool SeenZero = false;
2398 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2399 canChangeToInt(RHS, SeenZero, Subtarget) &&
2400 // If one of the operand is zero, it's safe to ignore the NaN case since
2401 // we only care about equality comparisons.
2402 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2403 // If unsafe fp math optimization is enabled and there are no othter uses of
2404 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2405 // to an integer comparison.
2406 if (CC == ISD::SETOEQ)
2408 else if (CC == ISD::SETUNE)
2412 if (LHS.getValueType() == MVT::f32) {
2413 LHS = bitcastf32Toi32(LHS, DAG);
2414 RHS = bitcastf32Toi32(RHS, DAG);
2415 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2416 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2417 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2418 Chain, Dest, ARMcc, CCR, Cmp);
2423 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2424 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2425 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2426 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2427 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2428 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2429 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2435 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2436 SDValue Chain = Op.getOperand(0);
2437 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2438 SDValue LHS = Op.getOperand(2);
2439 SDValue RHS = Op.getOperand(3);
2440 SDValue Dest = Op.getOperand(4);
2441 DebugLoc dl = Op.getDebugLoc();
2443 if (LHS.getValueType() == MVT::i32) {
2445 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2446 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2447 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2448 Chain, Dest, ARMcc, CCR, Cmp);
2451 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2454 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2455 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2456 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2457 if (Result.getNode())
2461 ARMCC::CondCodes CondCode, CondCode2;
2462 FPCCToARMCC(CC, CondCode, CondCode2);
2464 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2465 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2466 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2467 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2468 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2469 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2470 if (CondCode2 != ARMCC::AL) {
2471 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2472 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2473 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2478 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2479 SDValue Chain = Op.getOperand(0);
2480 SDValue Table = Op.getOperand(1);
2481 SDValue Index = Op.getOperand(2);
2482 DebugLoc dl = Op.getDebugLoc();
2484 EVT PTy = getPointerTy();
2485 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2486 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2487 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2488 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2489 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2490 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2491 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2492 if (Subtarget->isThumb2()) {
2493 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2494 // which does another jump to the destination. This also makes it easier
2495 // to translate it to TBB / TBH later.
2496 // FIXME: This might not work if the function is extremely large.
2497 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2498 Addr, Op.getOperand(2), JTI, UId);
2500 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2501 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2502 PseudoSourceValue::getJumpTable(), 0,
2504 Chain = Addr.getValue(1);
2505 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2506 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2508 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2509 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2510 Chain = Addr.getValue(1);
2511 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2515 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2516 DebugLoc dl = Op.getDebugLoc();
2519 switch (Op.getOpcode()) {
2521 assert(0 && "Invalid opcode!");
2522 case ISD::FP_TO_SINT:
2523 Opc = ARMISD::FTOSI;
2525 case ISD::FP_TO_UINT:
2526 Opc = ARMISD::FTOUI;
2529 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2530 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2533 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2534 EVT VT = Op.getValueType();
2535 DebugLoc dl = Op.getDebugLoc();
2538 switch (Op.getOpcode()) {
2540 assert(0 && "Invalid opcode!");
2541 case ISD::SINT_TO_FP:
2542 Opc = ARMISD::SITOF;
2544 case ISD::UINT_TO_FP:
2545 Opc = ARMISD::UITOF;
2549 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2550 return DAG.getNode(Opc, dl, VT, Op);
2553 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2554 // Implement fcopysign with a fabs and a conditional fneg.
2555 SDValue Tmp0 = Op.getOperand(0);
2556 SDValue Tmp1 = Op.getOperand(1);
2557 DebugLoc dl = Op.getDebugLoc();
2558 EVT VT = Op.getValueType();
2559 EVT SrcVT = Tmp1.getValueType();
2560 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2561 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2562 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2563 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2564 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2565 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2568 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2569 MachineFunction &MF = DAG.getMachineFunction();
2570 MachineFrameInfo *MFI = MF.getFrameInfo();
2571 MFI->setReturnAddressIsTaken(true);
2573 EVT VT = Op.getValueType();
2574 DebugLoc dl = Op.getDebugLoc();
2575 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2577 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2578 SDValue Offset = DAG.getConstant(4, MVT::i32);
2579 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2580 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2581 NULL, 0, false, false, 0);
2584 // Return LR, which contains the return address. Mark it an implicit live-in.
2585 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2586 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2589 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2590 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2591 MFI->setFrameAddressIsTaken(true);
2593 EVT VT = Op.getValueType();
2594 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2595 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2596 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2597 ? ARM::R7 : ARM::R11;
2598 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2600 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2605 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2606 /// expand a bit convert where either the source or destination type is i64 to
2607 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2608 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2609 /// vectors), since the legalizer won't know what to do with that.
2610 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2611 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2612 DebugLoc dl = N->getDebugLoc();
2613 SDValue Op = N->getOperand(0);
2615 // This function is only supposed to be called for i64 types, either as the
2616 // source or destination of the bit convert.
2617 EVT SrcVT = Op.getValueType();
2618 EVT DstVT = N->getValueType(0);
2619 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2620 "ExpandBIT_CONVERT called for non-i64 type");
2622 // Turn i64->f64 into VMOVDRR.
2623 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2624 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2625 DAG.getConstant(0, MVT::i32));
2626 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2627 DAG.getConstant(1, MVT::i32));
2628 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2629 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2632 // Turn f64->i64 into VMOVRRD.
2633 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2634 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2635 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2636 // Merge the pieces into a single i64 value.
2637 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2643 /// getZeroVector - Returns a vector of specified type with all zero elements.
2644 /// Zero vectors are used to represent vector negation and in those cases
2645 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2646 /// not support i64 elements, so sometimes the zero vectors will need to be
2647 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2649 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2650 assert(VT.isVector() && "Expected a vector type");
2651 // The canonical modified immediate encoding of a zero vector is....0!
2652 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2653 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2654 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2655 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2658 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2659 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2660 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2661 SelectionDAG &DAG) const {
2662 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2663 EVT VT = Op.getValueType();
2664 unsigned VTBits = VT.getSizeInBits();
2665 DebugLoc dl = Op.getDebugLoc();
2666 SDValue ShOpLo = Op.getOperand(0);
2667 SDValue ShOpHi = Op.getOperand(1);
2668 SDValue ShAmt = Op.getOperand(2);
2670 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2672 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2674 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2675 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2676 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2677 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2678 DAG.getConstant(VTBits, MVT::i32));
2679 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2680 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2681 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2683 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2684 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2686 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2687 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2690 SDValue Ops[2] = { Lo, Hi };
2691 return DAG.getMergeValues(Ops, 2, dl);
2694 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2695 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2696 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2697 SelectionDAG &DAG) const {
2698 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2699 EVT VT = Op.getValueType();
2700 unsigned VTBits = VT.getSizeInBits();
2701 DebugLoc dl = Op.getDebugLoc();
2702 SDValue ShOpLo = Op.getOperand(0);
2703 SDValue ShOpHi = Op.getOperand(1);
2704 SDValue ShAmt = Op.getOperand(2);
2707 assert(Op.getOpcode() == ISD::SHL_PARTS);
2708 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2709 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2710 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2711 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2712 DAG.getConstant(VTBits, MVT::i32));
2713 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2714 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2716 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2717 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2718 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2720 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2721 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2724 SDValue Ops[2] = { Lo, Hi };
2725 return DAG.getMergeValues(Ops, 2, dl);
2728 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2729 const ARMSubtarget *ST) {
2730 EVT VT = N->getValueType(0);
2731 DebugLoc dl = N->getDebugLoc();
2733 if (!ST->hasV6T2Ops())
2736 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2737 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2740 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2741 const ARMSubtarget *ST) {
2742 EVT VT = N->getValueType(0);
2743 DebugLoc dl = N->getDebugLoc();
2745 // Lower vector shifts on NEON to use VSHL.
2746 if (VT.isVector()) {
2747 assert(ST->hasNEON() && "unexpected vector shift");
2749 // Left shifts translate directly to the vshiftu intrinsic.
2750 if (N->getOpcode() == ISD::SHL)
2751 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2752 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2753 N->getOperand(0), N->getOperand(1));
2755 assert((N->getOpcode() == ISD::SRA ||
2756 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2758 // NEON uses the same intrinsics for both left and right shifts. For
2759 // right shifts, the shift amounts are negative, so negate the vector of
2761 EVT ShiftVT = N->getOperand(1).getValueType();
2762 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2763 getZeroVector(ShiftVT, DAG, dl),
2765 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2766 Intrinsic::arm_neon_vshifts :
2767 Intrinsic::arm_neon_vshiftu);
2768 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2769 DAG.getConstant(vshiftInt, MVT::i32),
2770 N->getOperand(0), NegatedCount);
2773 // We can get here for a node like i32 = ISD::SHL i32, i64
2777 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2778 "Unknown shift to lower!");
2780 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2781 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2782 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2785 // If we are in thumb mode, we don't have RRX.
2786 if (ST->isThumb1Only()) return SDValue();
2788 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2789 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2790 DAG.getConstant(0, MVT::i32));
2791 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2792 DAG.getConstant(1, MVT::i32));
2794 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2795 // captures the result into a carry flag.
2796 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2797 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2799 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2800 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2802 // Merge the pieces into a single i64 value.
2803 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2806 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2807 SDValue TmpOp0, TmpOp1;
2808 bool Invert = false;
2812 SDValue Op0 = Op.getOperand(0);
2813 SDValue Op1 = Op.getOperand(1);
2814 SDValue CC = Op.getOperand(2);
2815 EVT VT = Op.getValueType();
2816 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2817 DebugLoc dl = Op.getDebugLoc();
2819 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2820 switch (SetCCOpcode) {
2821 default: llvm_unreachable("Illegal FP comparison"); break;
2823 case ISD::SETNE: Invert = true; // Fallthrough
2825 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2827 case ISD::SETLT: Swap = true; // Fallthrough
2829 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2831 case ISD::SETLE: Swap = true; // Fallthrough
2833 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2834 case ISD::SETUGE: Swap = true; // Fallthrough
2835 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2836 case ISD::SETUGT: Swap = true; // Fallthrough
2837 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2838 case ISD::SETUEQ: Invert = true; // Fallthrough
2840 // Expand this to (OLT | OGT).
2844 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2845 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2847 case ISD::SETUO: Invert = true; // Fallthrough
2849 // Expand this to (OLT | OGE).
2853 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2854 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2858 // Integer comparisons.
2859 switch (SetCCOpcode) {
2860 default: llvm_unreachable("Illegal integer comparison"); break;
2861 case ISD::SETNE: Invert = true;
2862 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2863 case ISD::SETLT: Swap = true;
2864 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2865 case ISD::SETLE: Swap = true;
2866 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2867 case ISD::SETULT: Swap = true;
2868 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2869 case ISD::SETULE: Swap = true;
2870 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2873 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2874 if (Opc == ARMISD::VCEQ) {
2877 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2879 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2882 // Ignore bitconvert.
2883 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2884 AndOp = AndOp.getOperand(0);
2886 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2888 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2889 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2896 std::swap(Op0, Op1);
2898 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2901 Result = DAG.getNOT(dl, Result, VT);
2906 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2907 /// valid vector constant for a NEON instruction with a "modified immediate"
2908 /// operand (e.g., VMOV). If so, return the encoded value.
2909 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2910 unsigned SplatBitSize, SelectionDAG &DAG,
2911 EVT &VT, bool is128Bits, bool isVMOV) {
2912 unsigned OpCmode, Imm;
2914 // SplatBitSize is set to the smallest size that splats the vector, so a
2915 // zero vector will always have SplatBitSize == 8. However, NEON modified
2916 // immediate instructions others than VMOV do not support the 8-bit encoding
2917 // of a zero vector, and the default encoding of zero is supposed to be the
2922 switch (SplatBitSize) {
2926 // Any 1-byte value is OK. Op=0, Cmode=1110.
2927 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2930 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2934 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2935 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2936 if ((SplatBits & ~0xff) == 0) {
2937 // Value = 0x00nn: Op=x, Cmode=100x.
2942 if ((SplatBits & ~0xff00) == 0) {
2943 // Value = 0xnn00: Op=x, Cmode=101x.
2945 Imm = SplatBits >> 8;
2951 // NEON's 32-bit VMOV supports splat values where:
2952 // * only one byte is nonzero, or
2953 // * the least significant byte is 0xff and the second byte is nonzero, or
2954 // * the least significant 2 bytes are 0xff and the third is nonzero.
2955 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
2956 if ((SplatBits & ~0xff) == 0) {
2957 // Value = 0x000000nn: Op=x, Cmode=000x.
2962 if ((SplatBits & ~0xff00) == 0) {
2963 // Value = 0x0000nn00: Op=x, Cmode=001x.
2965 Imm = SplatBits >> 8;
2968 if ((SplatBits & ~0xff0000) == 0) {
2969 // Value = 0x00nn0000: Op=x, Cmode=010x.
2971 Imm = SplatBits >> 16;
2974 if ((SplatBits & ~0xff000000) == 0) {
2975 // Value = 0xnn000000: Op=x, Cmode=011x.
2977 Imm = SplatBits >> 24;
2981 if ((SplatBits & ~0xffff) == 0 &&
2982 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2983 // Value = 0x0000nnff: Op=x, Cmode=1100.
2985 Imm = SplatBits >> 8;
2990 if ((SplatBits & ~0xffffff) == 0 &&
2991 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2992 // Value = 0x00nnffff: Op=x, Cmode=1101.
2994 Imm = SplatBits >> 16;
2995 SplatBits |= 0xffff;
2999 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3000 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3001 // VMOV.I32. A (very) minor optimization would be to replicate the value
3002 // and fall through here to test for a valid 64-bit splat. But, then the
3003 // caller would also need to check and handle the change in size.
3009 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3010 uint64_t BitMask = 0xff;
3012 unsigned ImmMask = 1;
3014 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3015 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3018 } else if ((SplatBits & BitMask) != 0) {
3024 // Op=1, Cmode=1110.
3027 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3032 llvm_unreachable("unexpected size for isNEONModifiedImm");
3036 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3037 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3040 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3041 bool &ReverseVEXT, unsigned &Imm) {
3042 unsigned NumElts = VT.getVectorNumElements();
3043 ReverseVEXT = false;
3046 // If this is a VEXT shuffle, the immediate value is the index of the first
3047 // element. The other shuffle indices must be the successive elements after
3049 unsigned ExpectedElt = Imm;
3050 for (unsigned i = 1; i < NumElts; ++i) {
3051 // Increment the expected index. If it wraps around, it may still be
3052 // a VEXT but the source vectors must be swapped.
3054 if (ExpectedElt == NumElts * 2) {
3059 if (ExpectedElt != static_cast<unsigned>(M[i]))
3063 // Adjust the index value if the source operands will be swapped.
3070 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3071 /// instruction with the specified blocksize. (The order of the elements
3072 /// within each block of the vector is reversed.)
3073 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3074 unsigned BlockSize) {
3075 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3076 "Only possible block sizes for VREV are: 16, 32, 64");
3078 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3082 unsigned NumElts = VT.getVectorNumElements();
3083 unsigned BlockElts = M[0] + 1;
3085 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3088 for (unsigned i = 0; i < NumElts; ++i) {
3089 if ((unsigned) M[i] !=
3090 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3097 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3098 unsigned &WhichResult) {
3099 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3103 unsigned NumElts = VT.getVectorNumElements();
3104 WhichResult = (M[0] == 0 ? 0 : 1);
3105 for (unsigned i = 0; i < NumElts; i += 2) {
3106 if ((unsigned) M[i] != i + WhichResult ||
3107 (unsigned) M[i+1] != i + NumElts + WhichResult)
3113 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3114 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3115 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3116 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3117 unsigned &WhichResult) {
3118 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3122 unsigned NumElts = VT.getVectorNumElements();
3123 WhichResult = (M[0] == 0 ? 0 : 1);
3124 for (unsigned i = 0; i < NumElts; i += 2) {
3125 if ((unsigned) M[i] != i + WhichResult ||
3126 (unsigned) M[i+1] != i + WhichResult)
3132 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3133 unsigned &WhichResult) {
3134 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3138 unsigned NumElts = VT.getVectorNumElements();
3139 WhichResult = (M[0] == 0 ? 0 : 1);
3140 for (unsigned i = 0; i != NumElts; ++i) {
3141 if ((unsigned) M[i] != 2 * i + WhichResult)
3145 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3146 if (VT.is64BitVector() && EltSz == 32)
3152 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3153 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3154 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3155 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3156 unsigned &WhichResult) {
3157 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3161 unsigned Half = VT.getVectorNumElements() / 2;
3162 WhichResult = (M[0] == 0 ? 0 : 1);
3163 for (unsigned j = 0; j != 2; ++j) {
3164 unsigned Idx = WhichResult;
3165 for (unsigned i = 0; i != Half; ++i) {
3166 if ((unsigned) M[i + j * Half] != Idx)
3172 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3173 if (VT.is64BitVector() && EltSz == 32)
3179 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3180 unsigned &WhichResult) {
3181 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3185 unsigned NumElts = VT.getVectorNumElements();
3186 WhichResult = (M[0] == 0 ? 0 : 1);
3187 unsigned Idx = WhichResult * NumElts / 2;
3188 for (unsigned i = 0; i != NumElts; i += 2) {
3189 if ((unsigned) M[i] != Idx ||
3190 (unsigned) M[i+1] != Idx + NumElts)
3195 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3196 if (VT.is64BitVector() && EltSz == 32)
3202 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3203 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3204 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3205 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3206 unsigned &WhichResult) {
3207 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3211 unsigned NumElts = VT.getVectorNumElements();
3212 WhichResult = (M[0] == 0 ? 0 : 1);
3213 unsigned Idx = WhichResult * NumElts / 2;
3214 for (unsigned i = 0; i != NumElts; i += 2) {
3215 if ((unsigned) M[i] != Idx ||
3216 (unsigned) M[i+1] != Idx)
3221 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3222 if (VT.is64BitVector() && EltSz == 32)
3228 // If this is a case we can't handle, return null and let the default
3229 // expansion code take care of it.
3230 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3231 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3232 DebugLoc dl = Op.getDebugLoc();
3233 EVT VT = Op.getValueType();
3235 APInt SplatBits, SplatUndef;
3236 unsigned SplatBitSize;
3238 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3239 if (SplatBitSize <= 64) {
3240 // Check if an immediate VMOV works.
3242 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3243 SplatUndef.getZExtValue(), SplatBitSize,
3244 DAG, VmovVT, VT.is128BitVector(), true);
3245 if (Val.getNode()) {
3246 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3247 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3250 // Try an immediate VMVN.
3251 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3252 ((1LL << SplatBitSize) - 1));
3253 Val = isNEONModifiedImm(NegatedImm,
3254 SplatUndef.getZExtValue(), SplatBitSize,
3255 DAG, VmovVT, VT.is128BitVector(), false);
3256 if (Val.getNode()) {
3257 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3258 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3263 // Scan through the operands to see if only one value is used.
3264 unsigned NumElts = VT.getVectorNumElements();
3265 bool isOnlyLowElement = true;
3266 bool usesOnlyOneValue = true;
3267 bool isConstant = true;
3269 for (unsigned i = 0; i < NumElts; ++i) {
3270 SDValue V = Op.getOperand(i);
3271 if (V.getOpcode() == ISD::UNDEF)
3274 isOnlyLowElement = false;
3275 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3278 if (!Value.getNode())
3280 else if (V != Value)
3281 usesOnlyOneValue = false;
3284 if (!Value.getNode())
3285 return DAG.getUNDEF(VT);
3287 if (isOnlyLowElement)
3288 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3290 // If all elements are constants, fall back to the default expansion, which
3291 // will generate a load from the constant pool.
3295 // Use VDUP for non-constant splats.
3296 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3297 if (usesOnlyOneValue && EltSize <= 32)
3298 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3300 // Vectors with 32- or 64-bit elements can be built by directly assigning
3301 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3302 // will be legalized.
3303 if (EltSize >= 32) {
3304 // Do the expansion with floating-point types, since that is what the VFP
3305 // registers are defined to use, and since i64 is not legal.
3306 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3307 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3308 SmallVector<SDValue, 8> Ops;
3309 for (unsigned i = 0; i < NumElts; ++i)
3310 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3311 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3312 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3318 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3319 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3320 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3321 /// are assumed to be legal.
3323 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3325 if (VT.getVectorNumElements() == 4 &&
3326 (VT.is128BitVector() || VT.is64BitVector())) {
3327 unsigned PFIndexes[4];
3328 for (unsigned i = 0; i != 4; ++i) {
3332 PFIndexes[i] = M[i];
3335 // Compute the index in the perfect shuffle table.
3336 unsigned PFTableIndex =
3337 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3338 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3339 unsigned Cost = (PFEntry >> 30);
3346 unsigned Imm, WhichResult;
3348 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3349 return (EltSize >= 32 ||
3350 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3351 isVREVMask(M, VT, 64) ||
3352 isVREVMask(M, VT, 32) ||
3353 isVREVMask(M, VT, 16) ||
3354 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3355 isVTRNMask(M, VT, WhichResult) ||
3356 isVUZPMask(M, VT, WhichResult) ||
3357 isVZIPMask(M, VT, WhichResult) ||
3358 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3359 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3360 isVZIP_v_undef_Mask(M, VT, WhichResult));
3363 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3364 /// the specified operations to build the shuffle.
3365 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3366 SDValue RHS, SelectionDAG &DAG,
3368 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3369 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3370 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3373 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3382 OP_VUZPL, // VUZP, left result
3383 OP_VUZPR, // VUZP, right result
3384 OP_VZIPL, // VZIP, left result
3385 OP_VZIPR, // VZIP, right result
3386 OP_VTRNL, // VTRN, left result
3387 OP_VTRNR // VTRN, right result
3390 if (OpNum == OP_COPY) {
3391 if (LHSID == (1*9+2)*9+3) return LHS;
3392 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3396 SDValue OpLHS, OpRHS;
3397 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3398 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3399 EVT VT = OpLHS.getValueType();
3402 default: llvm_unreachable("Unknown shuffle opcode!");
3404 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3409 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3410 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3414 return DAG.getNode(ARMISD::VEXT, dl, VT,
3416 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3419 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3420 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3423 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3424 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3427 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3428 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3432 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3433 SDValue V1 = Op.getOperand(0);
3434 SDValue V2 = Op.getOperand(1);
3435 DebugLoc dl = Op.getDebugLoc();
3436 EVT VT = Op.getValueType();
3437 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3438 SmallVector<int, 8> ShuffleMask;
3440 // Convert shuffles that are directly supported on NEON to target-specific
3441 // DAG nodes, instead of keeping them as shuffles and matching them again
3442 // during code selection. This is more efficient and avoids the possibility
3443 // of inconsistencies between legalization and selection.
3444 // FIXME: floating-point vectors should be canonicalized to integer vectors
3445 // of the same time so that they get CSEd properly.
3446 SVN->getMask(ShuffleMask);
3448 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3449 if (EltSize <= 32) {
3450 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3451 int Lane = SVN->getSplatIndex();
3452 // If this is undef splat, generate it via "just" vdup, if possible.
3453 if (Lane == -1) Lane = 0;
3455 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3456 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3458 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3459 DAG.getConstant(Lane, MVT::i32));
3464 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3467 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3468 DAG.getConstant(Imm, MVT::i32));
3471 if (isVREVMask(ShuffleMask, VT, 64))
3472 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3473 if (isVREVMask(ShuffleMask, VT, 32))
3474 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3475 if (isVREVMask(ShuffleMask, VT, 16))
3476 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3478 // Check for Neon shuffles that modify both input vectors in place.
3479 // If both results are used, i.e., if there are two shuffles with the same
3480 // source operands and with masks corresponding to both results of one of
3481 // these operations, DAG memoization will ensure that a single node is
3482 // used for both shuffles.
3483 unsigned WhichResult;
3484 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3485 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3486 V1, V2).getValue(WhichResult);
3487 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3488 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3489 V1, V2).getValue(WhichResult);
3490 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3491 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3492 V1, V2).getValue(WhichResult);
3494 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3495 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3496 V1, V1).getValue(WhichResult);
3497 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3498 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3499 V1, V1).getValue(WhichResult);
3500 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3501 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3502 V1, V1).getValue(WhichResult);
3505 // If the shuffle is not directly supported and it has 4 elements, use
3506 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3507 unsigned NumElts = VT.getVectorNumElements();
3509 unsigned PFIndexes[4];
3510 for (unsigned i = 0; i != 4; ++i) {
3511 if (ShuffleMask[i] < 0)
3514 PFIndexes[i] = ShuffleMask[i];
3517 // Compute the index in the perfect shuffle table.
3518 unsigned PFTableIndex =
3519 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3520 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3521 unsigned Cost = (PFEntry >> 30);
3524 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3527 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3528 if (EltSize >= 32) {
3529 // Do the expansion with floating-point types, since that is what the VFP
3530 // registers are defined to use, and since i64 is not legal.
3531 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3532 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3533 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3534 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3535 SmallVector<SDValue, 8> Ops;
3536 for (unsigned i = 0; i < NumElts; ++i) {
3537 if (ShuffleMask[i] < 0)
3538 Ops.push_back(DAG.getUNDEF(EltVT));
3540 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3541 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3542 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3545 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3546 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3552 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3553 EVT VT = Op.getValueType();
3554 DebugLoc dl = Op.getDebugLoc();
3555 SDValue Vec = Op.getOperand(0);
3556 SDValue Lane = Op.getOperand(1);
3557 assert(VT == MVT::i32 &&
3558 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3559 "unexpected type for custom-lowering vector extract");
3560 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3563 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3564 // The only time a CONCAT_VECTORS operation can have legal types is when
3565 // two 64-bit vectors are concatenated to a 128-bit vector.
3566 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3567 "unexpected CONCAT_VECTORS");
3568 DebugLoc dl = Op.getDebugLoc();
3569 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3570 SDValue Op0 = Op.getOperand(0);
3571 SDValue Op1 = Op.getOperand(1);
3572 if (Op0.getOpcode() != ISD::UNDEF)
3573 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3574 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3575 DAG.getIntPtrConstant(0));
3576 if (Op1.getOpcode() != ISD::UNDEF)
3577 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3578 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3579 DAG.getIntPtrConstant(1));
3580 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3583 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3584 switch (Op.getOpcode()) {
3585 default: llvm_unreachable("Don't know how to custom lower this!");
3586 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3587 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3588 case ISD::GlobalAddress:
3589 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3590 LowerGlobalAddressELF(Op, DAG);
3591 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3592 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3593 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3594 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3595 case ISD::VASTART: return LowerVASTART(Op, DAG);
3596 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3597 case ISD::SINT_TO_FP:
3598 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3599 case ISD::FP_TO_SINT:
3600 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3601 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3602 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3603 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3604 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3605 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3606 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3607 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3609 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3612 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3613 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3614 case ISD::SRL_PARTS:
3615 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3616 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3617 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3618 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3619 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3620 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3621 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3626 /// ReplaceNodeResults - Replace the results of node with an illegal result
3627 /// type with new values built out of custom code.
3628 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3629 SmallVectorImpl<SDValue>&Results,
3630 SelectionDAG &DAG) const {
3632 switch (N->getOpcode()) {
3634 llvm_unreachable("Don't know how to custom expand this!");
3636 case ISD::BIT_CONVERT:
3637 Res = ExpandBIT_CONVERT(N, DAG);
3641 Res = LowerShift(N, DAG, Subtarget);
3645 Results.push_back(Res);
3648 //===----------------------------------------------------------------------===//
3649 // ARM Scheduler Hooks
3650 //===----------------------------------------------------------------------===//
3653 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3654 MachineBasicBlock *BB,
3655 unsigned Size) const {
3656 unsigned dest = MI->getOperand(0).getReg();
3657 unsigned ptr = MI->getOperand(1).getReg();
3658 unsigned oldval = MI->getOperand(2).getReg();
3659 unsigned newval = MI->getOperand(3).getReg();
3660 unsigned scratch = BB->getParent()->getRegInfo()
3661 .createVirtualRegister(ARM::GPRRegisterClass);
3662 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3663 DebugLoc dl = MI->getDebugLoc();
3664 bool isThumb2 = Subtarget->isThumb2();
3666 unsigned ldrOpc, strOpc;
3668 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3670 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3671 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3674 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3675 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3678 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3679 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3683 MachineFunction *MF = BB->getParent();
3684 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3685 MachineFunction::iterator It = BB;
3686 ++It; // insert the new blocks after the current block
3688 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3689 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3690 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3691 MF->insert(It, loop1MBB);
3692 MF->insert(It, loop2MBB);
3693 MF->insert(It, exitMBB);
3695 // Transfer the remainder of BB and its successor edges to exitMBB.
3696 exitMBB->splice(exitMBB->begin(), BB,
3697 llvm::next(MachineBasicBlock::iterator(MI)),
3699 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3703 // fallthrough --> loop1MBB
3704 BB->addSuccessor(loop1MBB);
3707 // ldrex dest, [ptr]
3711 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3712 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3713 .addReg(dest).addReg(oldval));
3714 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3715 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3716 BB->addSuccessor(loop2MBB);
3717 BB->addSuccessor(exitMBB);
3720 // strex scratch, newval, [ptr]
3724 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3726 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3727 .addReg(scratch).addImm(0));
3728 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3729 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3730 BB->addSuccessor(loop1MBB);
3731 BB->addSuccessor(exitMBB);
3737 MI->eraseFromParent(); // The instruction is gone now.
3743 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3744 unsigned Size, unsigned BinOpcode) const {
3745 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3746 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3748 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3749 MachineFunction *MF = BB->getParent();
3750 MachineFunction::iterator It = BB;
3753 unsigned dest = MI->getOperand(0).getReg();
3754 unsigned ptr = MI->getOperand(1).getReg();
3755 unsigned incr = MI->getOperand(2).getReg();
3756 DebugLoc dl = MI->getDebugLoc();
3758 bool isThumb2 = Subtarget->isThumb2();
3759 unsigned ldrOpc, strOpc;
3761 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3763 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3764 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3767 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3768 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3771 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3772 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3776 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3777 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3778 MF->insert(It, loopMBB);
3779 MF->insert(It, exitMBB);
3781 // Transfer the remainder of BB and its successor edges to exitMBB.
3782 exitMBB->splice(exitMBB->begin(), BB,
3783 llvm::next(MachineBasicBlock::iterator(MI)),
3785 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3787 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3788 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3789 unsigned scratch2 = (!BinOpcode) ? incr :
3790 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3794 // fallthrough --> loopMBB
3795 BB->addSuccessor(loopMBB);
3799 // <binop> scratch2, dest, incr
3800 // strex scratch, scratch2, ptr
3803 // fallthrough --> exitMBB
3805 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3807 // operand order needs to go the other way for NAND
3808 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3809 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3810 addReg(incr).addReg(dest)).addReg(0);
3812 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3813 addReg(dest).addReg(incr)).addReg(0);
3816 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3818 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3819 .addReg(scratch).addImm(0));
3820 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3821 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3823 BB->addSuccessor(loopMBB);
3824 BB->addSuccessor(exitMBB);
3830 MI->eraseFromParent(); // The instruction is gone now.
3836 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3837 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3838 E = MBB->succ_end(); I != E; ++I)
3841 llvm_unreachable("Expecting a BB with two successors!");
3845 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3846 MachineBasicBlock *BB) const {
3847 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3848 DebugLoc dl = MI->getDebugLoc();
3849 bool isThumb2 = Subtarget->isThumb2();
3850 switch (MI->getOpcode()) {
3853 llvm_unreachable("Unexpected instr type to insert");
3855 case ARM::ATOMIC_LOAD_ADD_I8:
3856 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3857 case ARM::ATOMIC_LOAD_ADD_I16:
3858 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3859 case ARM::ATOMIC_LOAD_ADD_I32:
3860 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3862 case ARM::ATOMIC_LOAD_AND_I8:
3863 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3864 case ARM::ATOMIC_LOAD_AND_I16:
3865 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3866 case ARM::ATOMIC_LOAD_AND_I32:
3867 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3869 case ARM::ATOMIC_LOAD_OR_I8:
3870 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3871 case ARM::ATOMIC_LOAD_OR_I16:
3872 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3873 case ARM::ATOMIC_LOAD_OR_I32:
3874 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3876 case ARM::ATOMIC_LOAD_XOR_I8:
3877 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3878 case ARM::ATOMIC_LOAD_XOR_I16:
3879 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3880 case ARM::ATOMIC_LOAD_XOR_I32:
3881 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3883 case ARM::ATOMIC_LOAD_NAND_I8:
3884 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3885 case ARM::ATOMIC_LOAD_NAND_I16:
3886 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3887 case ARM::ATOMIC_LOAD_NAND_I32:
3888 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3890 case ARM::ATOMIC_LOAD_SUB_I8:
3891 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3892 case ARM::ATOMIC_LOAD_SUB_I16:
3893 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3894 case ARM::ATOMIC_LOAD_SUB_I32:
3895 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3897 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3898 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3899 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3901 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3902 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3903 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3905 case ARM::tMOVCCr_pseudo: {
3906 // To "insert" a SELECT_CC instruction, we actually have to insert the
3907 // diamond control-flow pattern. The incoming instruction knows the
3908 // destination vreg to set, the condition code register to branch on, the
3909 // true/false values to select between, and a branch opcode to use.
3910 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3911 MachineFunction::iterator It = BB;
3917 // cmpTY ccX, r1, r2
3919 // fallthrough --> copy0MBB
3920 MachineBasicBlock *thisMBB = BB;
3921 MachineFunction *F = BB->getParent();
3922 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3923 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3924 F->insert(It, copy0MBB);
3925 F->insert(It, sinkMBB);
3927 // Transfer the remainder of BB and its successor edges to sinkMBB.
3928 sinkMBB->splice(sinkMBB->begin(), BB,
3929 llvm::next(MachineBasicBlock::iterator(MI)),
3931 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
3933 BB->addSuccessor(copy0MBB);
3934 BB->addSuccessor(sinkMBB);
3936 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3937 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3940 // %FalseValue = ...
3941 // # fallthrough to sinkMBB
3944 // Update machine-CFG edges
3945 BB->addSuccessor(sinkMBB);
3948 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3951 BuildMI(*BB, BB->begin(), dl,
3952 TII->get(ARM::PHI), MI->getOperand(0).getReg())
3953 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3954 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3956 MI->eraseFromParent(); // The pseudo instruction is gone now.
3961 case ARM::BCCZi64: {
3962 // Compare both parts that make up the double comparison separately for
3964 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
3966 unsigned LHS1 = MI->getOperand(1).getReg();
3967 unsigned LHS2 = MI->getOperand(2).getReg();
3969 AddDefaultPred(BuildMI(BB, dl,
3970 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3971 .addReg(LHS1).addImm(0));
3972 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3973 .addReg(LHS2).addImm(0)
3974 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
3976 unsigned RHS1 = MI->getOperand(3).getReg();
3977 unsigned RHS2 = MI->getOperand(4).getReg();
3978 AddDefaultPred(BuildMI(BB, dl,
3979 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3980 .addReg(LHS1).addReg(RHS1));
3981 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3982 .addReg(LHS2).addReg(RHS2)
3983 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
3986 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
3987 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
3988 if (MI->getOperand(0).getImm() == ARMCC::NE)
3989 std::swap(destMBB, exitMBB);
3991 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3992 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
3993 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
3996 MI->eraseFromParent(); // The pseudo instruction is gone now.
4003 case ARM::t2SUBrSPi_:
4004 case ARM::t2SUBrSPi12_:
4005 case ARM::t2SUBrSPs_: {
4006 MachineFunction *MF = BB->getParent();
4007 unsigned DstReg = MI->getOperand(0).getReg();
4008 unsigned SrcReg = MI->getOperand(1).getReg();
4009 bool DstIsDead = MI->getOperand(0).isDead();
4010 bool SrcIsKill = MI->getOperand(1).isKill();
4012 if (SrcReg != ARM::SP) {
4013 // Copy the source to SP from virtual register.
4014 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
4015 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4016 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
4017 BuildMI(*BB, MI, dl, TII->get(CopyOpc), ARM::SP)
4018 .addReg(SrcReg, getKillRegState(SrcIsKill));
4022 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
4023 switch (MI->getOpcode()) {
4025 llvm_unreachable("Unexpected pseudo instruction!");
4031 OpOpc = ARM::tADDspr;
4034 OpOpc = ARM::tSUBspi;
4036 case ARM::t2SUBrSPi_:
4037 OpOpc = ARM::t2SUBrSPi;
4038 NeedPred = true; NeedCC = true;
4040 case ARM::t2SUBrSPi12_:
4041 OpOpc = ARM::t2SUBrSPi12;
4044 case ARM::t2SUBrSPs_:
4045 OpOpc = ARM::t2SUBrSPs;
4046 NeedPred = true; NeedCC = true; NeedOp3 = true;
4049 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(OpOpc), ARM::SP);
4050 if (OpOpc == ARM::tAND)
4051 AddDefaultT1CC(MIB);
4052 MIB.addReg(ARM::SP);
4053 MIB.addOperand(MI->getOperand(2));
4055 MIB.addOperand(MI->getOperand(3));
4057 AddDefaultPred(MIB);
4061 // Copy the result from SP to virtual register.
4062 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
4063 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4064 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
4065 BuildMI(*BB, MI, dl, TII->get(CopyOpc))
4066 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
4068 MI->eraseFromParent(); // The pseudo instruction is gone now.
4074 //===----------------------------------------------------------------------===//
4075 // ARM Optimization Hooks
4076 //===----------------------------------------------------------------------===//
4079 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4080 TargetLowering::DAGCombinerInfo &DCI) {
4081 SelectionDAG &DAG = DCI.DAG;
4082 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4083 EVT VT = N->getValueType(0);
4084 unsigned Opc = N->getOpcode();
4085 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4086 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4087 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4088 ISD::CondCode CC = ISD::SETCC_INVALID;
4091 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4093 SDValue CCOp = Slct.getOperand(0);
4094 if (CCOp.getOpcode() == ISD::SETCC)
4095 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4098 bool DoXform = false;
4100 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4103 if (LHS.getOpcode() == ISD::Constant &&
4104 cast<ConstantSDNode>(LHS)->isNullValue()) {
4106 } else if (CC != ISD::SETCC_INVALID &&
4107 RHS.getOpcode() == ISD::Constant &&
4108 cast<ConstantSDNode>(RHS)->isNullValue()) {
4109 std::swap(LHS, RHS);
4110 SDValue Op0 = Slct.getOperand(0);
4111 EVT OpVT = isSlctCC ? Op0.getValueType() :
4112 Op0.getOperand(0).getValueType();
4113 bool isInt = OpVT.isInteger();
4114 CC = ISD::getSetCCInverse(CC, isInt);
4116 if (!TLI.isCondCodeLegal(CC, OpVT))
4117 return SDValue(); // Inverse operator isn't legal.
4124 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4126 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4127 Slct.getOperand(0), Slct.getOperand(1), CC);
4128 SDValue CCOp = Slct.getOperand(0);
4130 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4131 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4132 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4133 CCOp, OtherOp, Result);
4138 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4139 static SDValue PerformADDCombine(SDNode *N,
4140 TargetLowering::DAGCombinerInfo &DCI) {
4141 // added by evan in r37685 with no testcase.
4142 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4144 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4145 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4146 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4147 if (Result.getNode()) return Result;
4149 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4150 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4151 if (Result.getNode()) return Result;
4157 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4158 static SDValue PerformSUBCombine(SDNode *N,
4159 TargetLowering::DAGCombinerInfo &DCI) {
4160 // added by evan in r37685 with no testcase.
4161 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4163 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4164 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4165 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4166 if (Result.getNode()) return Result;
4172 static SDValue PerformMULCombine(SDNode *N,
4173 TargetLowering::DAGCombinerInfo &DCI,
4174 const ARMSubtarget *Subtarget) {
4175 SelectionDAG &DAG = DCI.DAG;
4177 if (Subtarget->isThumb1Only())
4180 if (DAG.getMachineFunction().
4181 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4184 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4187 EVT VT = N->getValueType(0);
4191 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4195 uint64_t MulAmt = C->getZExtValue();
4196 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4197 ShiftAmt = ShiftAmt & (32 - 1);
4198 SDValue V = N->getOperand(0);
4199 DebugLoc DL = N->getDebugLoc();
4202 MulAmt >>= ShiftAmt;
4203 if (isPowerOf2_32(MulAmt - 1)) {
4204 // (mul x, 2^N + 1) => (add (shl x, N), x)
4205 Res = DAG.getNode(ISD::ADD, DL, VT,
4206 V, DAG.getNode(ISD::SHL, DL, VT,
4207 V, DAG.getConstant(Log2_32(MulAmt-1),
4209 } else if (isPowerOf2_32(MulAmt + 1)) {
4210 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4211 Res = DAG.getNode(ISD::SUB, DL, VT,
4212 DAG.getNode(ISD::SHL, DL, VT,
4213 V, DAG.getConstant(Log2_32(MulAmt+1),
4220 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4221 DAG.getConstant(ShiftAmt, MVT::i32));
4223 // Do not add new nodes to DAG combiner worklist.
4224 DCI.CombineTo(N, Res, false);
4228 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4229 static SDValue PerformORCombine(SDNode *N,
4230 TargetLowering::DAGCombinerInfo &DCI,
4231 const ARMSubtarget *Subtarget) {
4232 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4235 // BFI is only available on V6T2+
4236 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4239 SelectionDAG &DAG = DCI.DAG;
4240 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4241 DebugLoc DL = N->getDebugLoc();
4242 // 1) or (and A, mask), val => ARMbfi A, val, mask
4243 // iff (val & mask) == val
4245 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4246 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4247 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4248 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4249 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4250 // (i.e., copy a bitfield value into another bitfield of the same width)
4251 if (N0.getOpcode() != ISD::AND)
4254 EVT VT = N->getValueType(0);
4259 // The value and the mask need to be constants so we can verify this is
4260 // actually a bitfield set. If the mask is 0xffff, we can do better
4261 // via a movt instruction, so don't use BFI in that case.
4262 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4265 unsigned Mask = C->getZExtValue();
4269 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4270 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4271 unsigned Val = C->getZExtValue();
4272 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4274 Val >>= CountTrailingZeros_32(~Mask);
4276 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4277 DAG.getConstant(Val, MVT::i32),
4278 DAG.getConstant(Mask, MVT::i32));
4280 // Do not add new nodes to DAG combiner worklist.
4281 DCI.CombineTo(N, Res, false);
4282 } else if (N1.getOpcode() == ISD::AND) {
4283 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4284 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4287 unsigned Mask2 = C->getZExtValue();
4289 if (ARM::isBitFieldInvertedMask(Mask) &&
4290 ARM::isBitFieldInvertedMask(~Mask2) &&
4291 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4292 // The pack halfword instruction works better for masks that fit it,
4293 // so use that when it's available.
4294 if (Subtarget->hasT2ExtractPack() &&
4295 (Mask == 0xffff || Mask == 0xffff0000))
4298 unsigned lsb = CountTrailingZeros_32(Mask2);
4299 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4300 DAG.getConstant(lsb, MVT::i32));
4301 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4302 DAG.getConstant(Mask, MVT::i32));
4303 // Do not add new nodes to DAG combiner worklist.
4304 DCI.CombineTo(N, Res, false);
4305 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4306 ARM::isBitFieldInvertedMask(Mask2) &&
4307 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4308 // The pack halfword instruction works better for masks that fit it,
4309 // so use that when it's available.
4310 if (Subtarget->hasT2ExtractPack() &&
4311 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4314 unsigned lsb = CountTrailingZeros_32(Mask);
4315 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4316 DAG.getConstant(lsb, MVT::i32));
4317 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4318 DAG.getConstant(Mask2, MVT::i32));
4319 // Do not add new nodes to DAG combiner worklist.
4320 DCI.CombineTo(N, Res, false);
4327 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4328 /// ARMISD::VMOVRRD.
4329 static SDValue PerformVMOVRRDCombine(SDNode *N,
4330 TargetLowering::DAGCombinerInfo &DCI) {
4331 // fmrrd(fmdrr x, y) -> x,y
4332 SDValue InDouble = N->getOperand(0);
4333 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4334 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4338 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4339 /// ARMISD::VDUPLANE.
4340 static SDValue PerformVDUPLANECombine(SDNode *N,
4341 TargetLowering::DAGCombinerInfo &DCI) {
4342 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4344 SDValue Op = N->getOperand(0);
4345 EVT VT = N->getValueType(0);
4347 // Ignore bit_converts.
4348 while (Op.getOpcode() == ISD::BIT_CONVERT)
4349 Op = Op.getOperand(0);
4350 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4353 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4354 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4355 // The canonical VMOV for a zero vector uses a 32-bit element size.
4356 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4358 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4360 if (EltSize > VT.getVectorElementType().getSizeInBits())
4363 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4364 return DCI.CombineTo(N, Res, false);
4367 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4368 /// operand of a vector shift operation, where all the elements of the
4369 /// build_vector must have the same constant integer value.
4370 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4371 // Ignore bit_converts.
4372 while (Op.getOpcode() == ISD::BIT_CONVERT)
4373 Op = Op.getOperand(0);
4374 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4375 APInt SplatBits, SplatUndef;
4376 unsigned SplatBitSize;
4378 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4379 HasAnyUndefs, ElementBits) ||
4380 SplatBitSize > ElementBits)
4382 Cnt = SplatBits.getSExtValue();
4386 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4387 /// operand of a vector shift left operation. That value must be in the range:
4388 /// 0 <= Value < ElementBits for a left shift; or
4389 /// 0 <= Value <= ElementBits for a long left shift.
4390 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4391 assert(VT.isVector() && "vector shift count is not a vector type");
4392 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4393 if (! getVShiftImm(Op, ElementBits, Cnt))
4395 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4398 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4399 /// operand of a vector shift right operation. For a shift opcode, the value
4400 /// is positive, but for an intrinsic the value count must be negative. The
4401 /// absolute value must be in the range:
4402 /// 1 <= |Value| <= ElementBits for a right shift; or
4403 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4404 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4406 assert(VT.isVector() && "vector shift count is not a vector type");
4407 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4408 if (! getVShiftImm(Op, ElementBits, Cnt))
4412 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4415 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4416 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4417 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4420 // Don't do anything for most intrinsics.
4423 // Vector shifts: check for immediate versions and lower them.
4424 // Note: This is done during DAG combining instead of DAG legalizing because
4425 // the build_vectors for 64-bit vector element shift counts are generally
4426 // not legal, and it is hard to see their values after they get legalized to
4427 // loads from a constant pool.
4428 case Intrinsic::arm_neon_vshifts:
4429 case Intrinsic::arm_neon_vshiftu:
4430 case Intrinsic::arm_neon_vshiftls:
4431 case Intrinsic::arm_neon_vshiftlu:
4432 case Intrinsic::arm_neon_vshiftn:
4433 case Intrinsic::arm_neon_vrshifts:
4434 case Intrinsic::arm_neon_vrshiftu:
4435 case Intrinsic::arm_neon_vrshiftn:
4436 case Intrinsic::arm_neon_vqshifts:
4437 case Intrinsic::arm_neon_vqshiftu:
4438 case Intrinsic::arm_neon_vqshiftsu:
4439 case Intrinsic::arm_neon_vqshiftns:
4440 case Intrinsic::arm_neon_vqshiftnu:
4441 case Intrinsic::arm_neon_vqshiftnsu:
4442 case Intrinsic::arm_neon_vqrshiftns:
4443 case Intrinsic::arm_neon_vqrshiftnu:
4444 case Intrinsic::arm_neon_vqrshiftnsu: {
4445 EVT VT = N->getOperand(1).getValueType();
4447 unsigned VShiftOpc = 0;
4450 case Intrinsic::arm_neon_vshifts:
4451 case Intrinsic::arm_neon_vshiftu:
4452 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4453 VShiftOpc = ARMISD::VSHL;
4456 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4457 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4458 ARMISD::VSHRs : ARMISD::VSHRu);
4463 case Intrinsic::arm_neon_vshiftls:
4464 case Intrinsic::arm_neon_vshiftlu:
4465 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4467 llvm_unreachable("invalid shift count for vshll intrinsic");
4469 case Intrinsic::arm_neon_vrshifts:
4470 case Intrinsic::arm_neon_vrshiftu:
4471 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4475 case Intrinsic::arm_neon_vqshifts:
4476 case Intrinsic::arm_neon_vqshiftu:
4477 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4481 case Intrinsic::arm_neon_vqshiftsu:
4482 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4484 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4486 case Intrinsic::arm_neon_vshiftn:
4487 case Intrinsic::arm_neon_vrshiftn:
4488 case Intrinsic::arm_neon_vqshiftns:
4489 case Intrinsic::arm_neon_vqshiftnu:
4490 case Intrinsic::arm_neon_vqshiftnsu:
4491 case Intrinsic::arm_neon_vqrshiftns:
4492 case Intrinsic::arm_neon_vqrshiftnu:
4493 case Intrinsic::arm_neon_vqrshiftnsu:
4494 // Narrowing shifts require an immediate right shift.
4495 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4497 llvm_unreachable("invalid shift count for narrowing vector shift "
4501 llvm_unreachable("unhandled vector shift");
4505 case Intrinsic::arm_neon_vshifts:
4506 case Intrinsic::arm_neon_vshiftu:
4507 // Opcode already set above.
4509 case Intrinsic::arm_neon_vshiftls:
4510 case Intrinsic::arm_neon_vshiftlu:
4511 if (Cnt == VT.getVectorElementType().getSizeInBits())
4512 VShiftOpc = ARMISD::VSHLLi;
4514 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4515 ARMISD::VSHLLs : ARMISD::VSHLLu);
4517 case Intrinsic::arm_neon_vshiftn:
4518 VShiftOpc = ARMISD::VSHRN; break;
4519 case Intrinsic::arm_neon_vrshifts:
4520 VShiftOpc = ARMISD::VRSHRs; break;
4521 case Intrinsic::arm_neon_vrshiftu:
4522 VShiftOpc = ARMISD::VRSHRu; break;
4523 case Intrinsic::arm_neon_vrshiftn:
4524 VShiftOpc = ARMISD::VRSHRN; break;
4525 case Intrinsic::arm_neon_vqshifts:
4526 VShiftOpc = ARMISD::VQSHLs; break;
4527 case Intrinsic::arm_neon_vqshiftu:
4528 VShiftOpc = ARMISD::VQSHLu; break;
4529 case Intrinsic::arm_neon_vqshiftsu:
4530 VShiftOpc = ARMISD::VQSHLsu; break;
4531 case Intrinsic::arm_neon_vqshiftns:
4532 VShiftOpc = ARMISD::VQSHRNs; break;
4533 case Intrinsic::arm_neon_vqshiftnu:
4534 VShiftOpc = ARMISD::VQSHRNu; break;
4535 case Intrinsic::arm_neon_vqshiftnsu:
4536 VShiftOpc = ARMISD::VQSHRNsu; break;
4537 case Intrinsic::arm_neon_vqrshiftns:
4538 VShiftOpc = ARMISD::VQRSHRNs; break;
4539 case Intrinsic::arm_neon_vqrshiftnu:
4540 VShiftOpc = ARMISD::VQRSHRNu; break;
4541 case Intrinsic::arm_neon_vqrshiftnsu:
4542 VShiftOpc = ARMISD::VQRSHRNsu; break;
4545 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4546 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4549 case Intrinsic::arm_neon_vshiftins: {
4550 EVT VT = N->getOperand(1).getValueType();
4552 unsigned VShiftOpc = 0;
4554 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4555 VShiftOpc = ARMISD::VSLI;
4556 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4557 VShiftOpc = ARMISD::VSRI;
4559 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4562 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4563 N->getOperand(1), N->getOperand(2),
4564 DAG.getConstant(Cnt, MVT::i32));
4567 case Intrinsic::arm_neon_vqrshifts:
4568 case Intrinsic::arm_neon_vqrshiftu:
4569 // No immediate versions of these to check for.
4576 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4577 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4578 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4579 /// vector element shift counts are generally not legal, and it is hard to see
4580 /// their values after they get legalized to loads from a constant pool.
4581 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4582 const ARMSubtarget *ST) {
4583 EVT VT = N->getValueType(0);
4585 // Nothing to be done for scalar shifts.
4586 if (! VT.isVector())
4589 assert(ST->hasNEON() && "unexpected vector shift");
4592 switch (N->getOpcode()) {
4593 default: llvm_unreachable("unexpected shift opcode");
4596 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4597 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4598 DAG.getConstant(Cnt, MVT::i32));
4603 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4604 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4605 ARMISD::VSHRs : ARMISD::VSHRu);
4606 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4607 DAG.getConstant(Cnt, MVT::i32));
4613 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4614 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4615 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4616 const ARMSubtarget *ST) {
4617 SDValue N0 = N->getOperand(0);
4619 // Check for sign- and zero-extensions of vector extract operations of 8-
4620 // and 16-bit vector elements. NEON supports these directly. They are
4621 // handled during DAG combining because type legalization will promote them
4622 // to 32-bit types and it is messy to recognize the operations after that.
4623 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4624 SDValue Vec = N0.getOperand(0);
4625 SDValue Lane = N0.getOperand(1);
4626 EVT VT = N->getValueType(0);
4627 EVT EltVT = N0.getValueType();
4628 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4630 if (VT == MVT::i32 &&
4631 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4632 TLI.isTypeLegal(Vec.getValueType())) {
4635 switch (N->getOpcode()) {
4636 default: llvm_unreachable("unexpected opcode");
4637 case ISD::SIGN_EXTEND:
4638 Opc = ARMISD::VGETLANEs;
4640 case ISD::ZERO_EXTEND:
4641 case ISD::ANY_EXTEND:
4642 Opc = ARMISD::VGETLANEu;
4645 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4652 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4653 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4654 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4655 const ARMSubtarget *ST) {
4656 // If the target supports NEON, try to use vmax/vmin instructions for f32
4657 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4658 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4659 // a NaN; only do the transformation when it matches that behavior.
4661 // For now only do this when using NEON for FP operations; if using VFP, it
4662 // is not obvious that the benefit outweighs the cost of switching to the
4664 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4665 N->getValueType(0) != MVT::f32)
4668 SDValue CondLHS = N->getOperand(0);
4669 SDValue CondRHS = N->getOperand(1);
4670 SDValue LHS = N->getOperand(2);
4671 SDValue RHS = N->getOperand(3);
4672 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4674 unsigned Opcode = 0;
4676 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4677 IsReversed = false; // x CC y ? x : y
4678 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4679 IsReversed = true ; // x CC y ? y : x
4693 // If LHS is NaN, an ordered comparison will be false and the result will
4694 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4695 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4696 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4697 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4699 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4700 // will return -0, so vmin can only be used for unsafe math or if one of
4701 // the operands is known to be nonzero.
4702 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4704 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4706 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4715 // If LHS is NaN, an ordered comparison will be false and the result will
4716 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4717 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4718 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4719 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4721 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4722 // will return +0, so vmax can only be used for unsafe math or if one of
4723 // the operands is known to be nonzero.
4724 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4726 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4728 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4734 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4737 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4738 DAGCombinerInfo &DCI) const {
4739 switch (N->getOpcode()) {
4741 case ISD::ADD: return PerformADDCombine(N, DCI);
4742 case ISD::SUB: return PerformSUBCombine(N, DCI);
4743 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4744 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4745 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4746 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4747 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4750 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4751 case ISD::SIGN_EXTEND:
4752 case ISD::ZERO_EXTEND:
4753 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4754 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4759 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4760 if (!Subtarget->hasV6Ops())
4761 // Pre-v6 does not support unaligned mem access.
4764 // v6+ may or may not support unaligned mem access depending on the system
4766 // FIXME: This is pretty conservative. Should we provide cmdline option to
4767 // control the behaviour?
4768 if (!Subtarget->isTargetDarwin())
4771 switch (VT.getSimpleVT().SimpleTy) {
4778 // FIXME: VLD1 etc with standard alignment is legal.
4782 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4787 switch (VT.getSimpleVT().SimpleTy) {
4788 default: return false;
4803 if ((V & (Scale - 1)) != 0)
4806 return V == (V & ((1LL << 5) - 1));
4809 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4810 const ARMSubtarget *Subtarget) {
4817 switch (VT.getSimpleVT().SimpleTy) {
4818 default: return false;
4823 // + imm12 or - imm8
4825 return V == (V & ((1LL << 8) - 1));
4826 return V == (V & ((1LL << 12) - 1));
4829 // Same as ARM mode. FIXME: NEON?
4830 if (!Subtarget->hasVFP2())
4835 return V == (V & ((1LL << 8) - 1));
4839 /// isLegalAddressImmediate - Return true if the integer value can be used
4840 /// as the offset of the target addressing mode for load / store of the
4842 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4843 const ARMSubtarget *Subtarget) {
4850 if (Subtarget->isThumb1Only())
4851 return isLegalT1AddressImmediate(V, VT);
4852 else if (Subtarget->isThumb2())
4853 return isLegalT2AddressImmediate(V, VT, Subtarget);
4858 switch (VT.getSimpleVT().SimpleTy) {
4859 default: return false;
4864 return V == (V & ((1LL << 12) - 1));
4867 return V == (V & ((1LL << 8) - 1));
4870 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4875 return V == (V & ((1LL << 8) - 1));
4879 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4881 int Scale = AM.Scale;
4885 switch (VT.getSimpleVT().SimpleTy) {
4886 default: return false;
4895 return Scale == 2 || Scale == 4 || Scale == 8;
4898 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4902 // Note, we allow "void" uses (basically, uses that aren't loads or
4903 // stores), because arm allows folding a scale into many arithmetic
4904 // operations. This should be made more precise and revisited later.
4906 // Allow r << imm, but the imm has to be a multiple of two.
4907 if (Scale & 1) return false;
4908 return isPowerOf2_32(Scale);
4912 /// isLegalAddressingMode - Return true if the addressing mode represented
4913 /// by AM is legal for this target, for a load/store of the specified type.
4914 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4915 const Type *Ty) const {
4916 EVT VT = getValueType(Ty, true);
4917 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4920 // Can never fold addr of global into load/store.
4925 case 0: // no scale reg, must be "r+i" or "r", or "i".
4928 if (Subtarget->isThumb1Only())
4932 // ARM doesn't support any R+R*scale+imm addr modes.
4939 if (Subtarget->isThumb2())
4940 return isLegalT2ScaledAddressingMode(AM, VT);
4942 int Scale = AM.Scale;
4943 switch (VT.getSimpleVT().SimpleTy) {
4944 default: return false;
4948 if (Scale < 0) Scale = -Scale;
4952 return isPowerOf2_32(Scale & ~1);
4956 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4961 // Note, we allow "void" uses (basically, uses that aren't loads or
4962 // stores), because arm allows folding a scale into many arithmetic
4963 // operations. This should be made more precise and revisited later.
4965 // Allow r << imm, but the imm has to be a multiple of two.
4966 if (Scale & 1) return false;
4967 return isPowerOf2_32(Scale);
4974 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4975 /// icmp immediate, that is the target has icmp instructions which can compare
4976 /// a register against the immediate without having to materialize the
4977 /// immediate into a register.
4978 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4979 if (!Subtarget->isThumb())
4980 return ARM_AM::getSOImmVal(Imm) != -1;
4981 if (Subtarget->isThumb2())
4982 return ARM_AM::getT2SOImmVal(Imm) != -1;
4983 return Imm >= 0 && Imm <= 255;
4986 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4987 bool isSEXTLoad, SDValue &Base,
4988 SDValue &Offset, bool &isInc,
4989 SelectionDAG &DAG) {
4990 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4993 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4995 Base = Ptr->getOperand(0);
4996 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4997 int RHSC = (int)RHS->getZExtValue();
4998 if (RHSC < 0 && RHSC > -256) {
4999 assert(Ptr->getOpcode() == ISD::ADD);
5001 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5005 isInc = (Ptr->getOpcode() == ISD::ADD);
5006 Offset = Ptr->getOperand(1);
5008 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5010 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5011 int RHSC = (int)RHS->getZExtValue();
5012 if (RHSC < 0 && RHSC > -0x1000) {
5013 assert(Ptr->getOpcode() == ISD::ADD);
5015 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5016 Base = Ptr->getOperand(0);
5021 if (Ptr->getOpcode() == ISD::ADD) {
5023 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5024 if (ShOpcVal != ARM_AM::no_shift) {
5025 Base = Ptr->getOperand(1);
5026 Offset = Ptr->getOperand(0);
5028 Base = Ptr->getOperand(0);
5029 Offset = Ptr->getOperand(1);
5034 isInc = (Ptr->getOpcode() == ISD::ADD);
5035 Base = Ptr->getOperand(0);
5036 Offset = Ptr->getOperand(1);
5040 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5044 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5045 bool isSEXTLoad, SDValue &Base,
5046 SDValue &Offset, bool &isInc,
5047 SelectionDAG &DAG) {
5048 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5051 Base = Ptr->getOperand(0);
5052 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5053 int RHSC = (int)RHS->getZExtValue();
5054 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5055 assert(Ptr->getOpcode() == ISD::ADD);
5057 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5059 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5060 isInc = Ptr->getOpcode() == ISD::ADD;
5061 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5069 /// getPreIndexedAddressParts - returns true by value, base pointer and
5070 /// offset pointer and addressing mode by reference if the node's address
5071 /// can be legally represented as pre-indexed load / store address.
5073 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5075 ISD::MemIndexedMode &AM,
5076 SelectionDAG &DAG) const {
5077 if (Subtarget->isThumb1Only())
5082 bool isSEXTLoad = false;
5083 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5084 Ptr = LD->getBasePtr();
5085 VT = LD->getMemoryVT();
5086 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5087 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5088 Ptr = ST->getBasePtr();
5089 VT = ST->getMemoryVT();
5094 bool isLegal = false;
5095 if (Subtarget->isThumb2())
5096 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5097 Offset, isInc, DAG);
5099 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5100 Offset, isInc, DAG);
5104 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5108 /// getPostIndexedAddressParts - returns true by value, base pointer and
5109 /// offset pointer and addressing mode by reference if this node can be
5110 /// combined with a load / store to form a post-indexed load / store.
5111 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5114 ISD::MemIndexedMode &AM,
5115 SelectionDAG &DAG) const {
5116 if (Subtarget->isThumb1Only())
5121 bool isSEXTLoad = false;
5122 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5123 VT = LD->getMemoryVT();
5124 Ptr = LD->getBasePtr();
5125 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5126 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5127 VT = ST->getMemoryVT();
5128 Ptr = ST->getBasePtr();
5133 bool isLegal = false;
5134 if (Subtarget->isThumb2())
5135 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5138 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5144 // Swap base ptr and offset to catch more post-index load / store when
5145 // it's legal. In Thumb2 mode, offset must be an immediate.
5146 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5147 !Subtarget->isThumb2())
5148 std::swap(Base, Offset);
5150 // Post-indexed load / store update the base pointer.
5155 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5159 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5163 const SelectionDAG &DAG,
5164 unsigned Depth) const {
5165 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5166 switch (Op.getOpcode()) {
5168 case ARMISD::CMOV: {
5169 // Bits are known zero/one if known on the LHS and RHS.
5170 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5171 if (KnownZero == 0 && KnownOne == 0) return;
5173 APInt KnownZeroRHS, KnownOneRHS;
5174 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5175 KnownZeroRHS, KnownOneRHS, Depth+1);
5176 KnownZero &= KnownZeroRHS;
5177 KnownOne &= KnownOneRHS;
5183 //===----------------------------------------------------------------------===//
5184 // ARM Inline Assembly Support
5185 //===----------------------------------------------------------------------===//
5187 /// getConstraintType - Given a constraint letter, return the type of
5188 /// constraint it is for this target.
5189 ARMTargetLowering::ConstraintType
5190 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5191 if (Constraint.size() == 1) {
5192 switch (Constraint[0]) {
5194 case 'l': return C_RegisterClass;
5195 case 'w': return C_RegisterClass;
5198 return TargetLowering::getConstraintType(Constraint);
5201 std::pair<unsigned, const TargetRegisterClass*>
5202 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5204 if (Constraint.size() == 1) {
5205 // GCC ARM Constraint Letters
5206 switch (Constraint[0]) {
5208 if (Subtarget->isThumb())
5209 return std::make_pair(0U, ARM::tGPRRegisterClass);
5211 return std::make_pair(0U, ARM::GPRRegisterClass);
5213 return std::make_pair(0U, ARM::GPRRegisterClass);
5216 return std::make_pair(0U, ARM::SPRRegisterClass);
5217 if (VT.getSizeInBits() == 64)
5218 return std::make_pair(0U, ARM::DPRRegisterClass);
5219 if (VT.getSizeInBits() == 128)
5220 return std::make_pair(0U, ARM::QPRRegisterClass);
5224 if (StringRef("{cc}").equals_lower(Constraint))
5225 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5227 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5230 std::vector<unsigned> ARMTargetLowering::
5231 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5233 if (Constraint.size() != 1)
5234 return std::vector<unsigned>();
5236 switch (Constraint[0]) { // GCC ARM Constraint Letters
5239 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5240 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5243 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5244 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5245 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5246 ARM::R12, ARM::LR, 0);
5249 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5250 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5251 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5252 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5253 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5254 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5255 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5256 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5257 if (VT.getSizeInBits() == 64)
5258 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5259 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5260 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5261 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5262 if (VT.getSizeInBits() == 128)
5263 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5264 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5268 return std::vector<unsigned>();
5271 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5272 /// vector. If it is invalid, don't add anything to Ops.
5273 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5275 std::vector<SDValue>&Ops,
5276 SelectionDAG &DAG) const {
5277 SDValue Result(0, 0);
5279 switch (Constraint) {
5281 case 'I': case 'J': case 'K': case 'L':
5282 case 'M': case 'N': case 'O':
5283 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5287 int64_t CVal64 = C->getSExtValue();
5288 int CVal = (int) CVal64;
5289 // None of these constraints allow values larger than 32 bits. Check
5290 // that the value fits in an int.
5294 switch (Constraint) {
5296 if (Subtarget->isThumb1Only()) {
5297 // This must be a constant between 0 and 255, for ADD
5299 if (CVal >= 0 && CVal <= 255)
5301 } else if (Subtarget->isThumb2()) {
5302 // A constant that can be used as an immediate value in a
5303 // data-processing instruction.
5304 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5307 // A constant that can be used as an immediate value in a
5308 // data-processing instruction.
5309 if (ARM_AM::getSOImmVal(CVal) != -1)
5315 if (Subtarget->isThumb()) { // FIXME thumb2
5316 // This must be a constant between -255 and -1, for negated ADD
5317 // immediates. This can be used in GCC with an "n" modifier that
5318 // prints the negated value, for use with SUB instructions. It is
5319 // not useful otherwise but is implemented for compatibility.
5320 if (CVal >= -255 && CVal <= -1)
5323 // This must be a constant between -4095 and 4095. It is not clear
5324 // what this constraint is intended for. Implemented for
5325 // compatibility with GCC.
5326 if (CVal >= -4095 && CVal <= 4095)
5332 if (Subtarget->isThumb1Only()) {
5333 // A 32-bit value where only one byte has a nonzero value. Exclude
5334 // zero to match GCC. This constraint is used by GCC internally for
5335 // constants that can be loaded with a move/shift combination.
5336 // It is not useful otherwise but is implemented for compatibility.
5337 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5339 } else if (Subtarget->isThumb2()) {
5340 // A constant whose bitwise inverse can be used as an immediate
5341 // value in a data-processing instruction. This can be used in GCC
5342 // with a "B" modifier that prints the inverted value, for use with
5343 // BIC and MVN instructions. It is not useful otherwise but is
5344 // implemented for compatibility.
5345 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5348 // A constant whose bitwise inverse can be used as an immediate
5349 // value in a data-processing instruction. This can be used in GCC
5350 // with a "B" modifier that prints the inverted value, for use with
5351 // BIC and MVN instructions. It is not useful otherwise but is
5352 // implemented for compatibility.
5353 if (ARM_AM::getSOImmVal(~CVal) != -1)
5359 if (Subtarget->isThumb1Only()) {
5360 // This must be a constant between -7 and 7,
5361 // for 3-operand ADD/SUB immediate instructions.
5362 if (CVal >= -7 && CVal < 7)
5364 } else if (Subtarget->isThumb2()) {
5365 // A constant whose negation can be used as an immediate value in a
5366 // data-processing instruction. This can be used in GCC with an "n"
5367 // modifier that prints the negated value, for use with SUB
5368 // instructions. It is not useful otherwise but is implemented for
5370 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5373 // A constant whose negation can be used as an immediate value in a
5374 // data-processing instruction. This can be used in GCC with an "n"
5375 // modifier that prints the negated value, for use with SUB
5376 // instructions. It is not useful otherwise but is implemented for
5378 if (ARM_AM::getSOImmVal(-CVal) != -1)
5384 if (Subtarget->isThumb()) { // FIXME thumb2
5385 // This must be a multiple of 4 between 0 and 1020, for
5386 // ADD sp + immediate.
5387 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5390 // A power of two or a constant between 0 and 32. This is used in
5391 // GCC for the shift amount on shifted register operands, but it is
5392 // useful in general for any shift amounts.
5393 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5399 if (Subtarget->isThumb()) { // FIXME thumb2
5400 // This must be a constant between 0 and 31, for shift amounts.
5401 if (CVal >= 0 && CVal <= 31)
5407 if (Subtarget->isThumb()) { // FIXME thumb2
5408 // This must be a multiple of 4 between -508 and 508, for
5409 // ADD/SUB sp = sp + immediate.
5410 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5415 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5419 if (Result.getNode()) {
5420 Ops.push_back(Result);
5423 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5427 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5428 // The ARM target isn't yet aware of offsets.
5432 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5433 APInt Imm = FPImm.bitcastToAPInt();
5434 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5435 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5436 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5438 // We can handle 4 bits of mantissa.
5439 // mantissa = (16+UInt(e:f:g:h))/16.
5440 if (Mantissa & 0x7ffff)
5443 if ((Mantissa & 0xf) != Mantissa)
5446 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5447 if (Exp < -3 || Exp > 4)
5449 Exp = ((Exp+3) & 0x7) ^ 4;
5451 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5454 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5455 APInt Imm = FPImm.bitcastToAPInt();
5456 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5457 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5458 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5460 // We can handle 4 bits of mantissa.
5461 // mantissa = (16+UInt(e:f:g:h))/16.
5462 if (Mantissa & 0xffffffffffffLL)
5465 if ((Mantissa & 0xf) != Mantissa)
5468 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5469 if (Exp < -3 || Exp > 4)
5471 Exp = ((Exp+3) & 0x7) ^ 4;
5473 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5476 bool ARM::isBitFieldInvertedMask(unsigned v) {
5477 if (v == 0xffffffff)
5479 // there can be 1's on either or both "outsides", all the "inside"
5481 unsigned int lsb = 0, msb = 31;
5482 while (v & (1 << msb)) --msb;
5483 while (v & (1 << lsb)) ++lsb;
5484 for (unsigned int i = lsb; i <= msb; ++i) {
5491 /// isFPImmLegal - Returns true if the target can instruction select the
5492 /// specified FP immediate natively. If false, the legalizer will
5493 /// materialize the FP immediate as a load from a constant pool.
5494 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5495 if (!Subtarget->hasVFP3())
5498 return ARM::getVFPf32Imm(Imm) != -1;
5500 return ARM::getVFPf64Imm(Imm) != -1;