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 };
834 static const unsigned ShadowRegList[] = { ARM::R0, ARM::R1 };
836 unsigned Reg = State.AllocateReg(HiRegList, ShadowRegList, 2);
838 // For the 2nd half of a v2f64, do not just fail.
842 // Put the whole thing on the stack.
843 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
844 State.AllocateStack(8, 8),
850 for (i = 0; i < 2; ++i)
851 if (HiRegList[i] == Reg)
854 unsigned T = State.AllocateReg(LoRegList[i]);
855 assert(T == LoRegList[i] && "Could not allocate register");
857 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
858 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
863 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
864 CCValAssign::LocInfo &LocInfo,
865 ISD::ArgFlagsTy &ArgFlags,
867 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
869 if (LocVT == MVT::v2f64 &&
870 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
872 return true; // we handled it
875 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
876 CCValAssign::LocInfo &LocInfo, CCState &State) {
877 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
878 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
880 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
882 return false; // we didn't handle it
885 for (i = 0; i < 2; ++i)
886 if (HiRegList[i] == Reg)
889 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
890 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
895 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
896 CCValAssign::LocInfo &LocInfo,
897 ISD::ArgFlagsTy &ArgFlags,
899 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
901 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
903 return true; // we handled it
906 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
907 CCValAssign::LocInfo &LocInfo,
908 ISD::ArgFlagsTy &ArgFlags,
910 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
914 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
915 /// given CallingConvention value.
916 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
918 bool isVarArg) const {
921 llvm_unreachable("Unsupported calling convention");
923 case CallingConv::Fast:
924 // Use target triple & subtarget features to do actual dispatch.
925 if (Subtarget->isAAPCS_ABI()) {
926 if (Subtarget->hasVFP2() &&
927 FloatABIType == FloatABI::Hard && !isVarArg)
928 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
930 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
932 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
933 case CallingConv::ARM_AAPCS_VFP:
934 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
935 case CallingConv::ARM_AAPCS:
936 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
937 case CallingConv::ARM_APCS:
938 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
942 /// LowerCallResult - Lower the result values of a call into the
943 /// appropriate copies out of appropriate physical registers.
945 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
946 CallingConv::ID CallConv, bool isVarArg,
947 const SmallVectorImpl<ISD::InputArg> &Ins,
948 DebugLoc dl, SelectionDAG &DAG,
949 SmallVectorImpl<SDValue> &InVals) const {
951 // Assign locations to each value returned by this call.
952 SmallVector<CCValAssign, 16> RVLocs;
953 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
954 RVLocs, *DAG.getContext());
955 CCInfo.AnalyzeCallResult(Ins,
956 CCAssignFnForNode(CallConv, /* Return*/ true,
959 // Copy all of the result registers out of their specified physreg.
960 for (unsigned i = 0; i != RVLocs.size(); ++i) {
961 CCValAssign VA = RVLocs[i];
964 if (VA.needsCustom()) {
965 // Handle f64 or half of a v2f64.
966 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
968 Chain = Lo.getValue(1);
969 InFlag = Lo.getValue(2);
970 VA = RVLocs[++i]; // skip ahead to next loc
971 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
973 Chain = Hi.getValue(1);
974 InFlag = Hi.getValue(2);
975 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
977 if (VA.getLocVT() == MVT::v2f64) {
978 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
979 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
980 DAG.getConstant(0, MVT::i32));
982 VA = RVLocs[++i]; // skip ahead to next loc
983 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
984 Chain = Lo.getValue(1);
985 InFlag = Lo.getValue(2);
986 VA = RVLocs[++i]; // skip ahead to next loc
987 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
988 Chain = Hi.getValue(1);
989 InFlag = Hi.getValue(2);
990 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
991 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
992 DAG.getConstant(1, MVT::i32));
995 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
997 Chain = Val.getValue(1);
998 InFlag = Val.getValue(2);
1001 switch (VA.getLocInfo()) {
1002 default: llvm_unreachable("Unknown loc info!");
1003 case CCValAssign::Full: break;
1004 case CCValAssign::BCvt:
1005 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
1009 InVals.push_back(Val);
1015 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1016 /// by "Src" to address "Dst" of size "Size". Alignment information is
1017 /// specified by the specific parameter attribute. The copy will be passed as
1018 /// a byval function parameter.
1019 /// Sometimes what we are copying is the end of a larger object, the part that
1020 /// does not fit in registers.
1022 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1023 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1025 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1026 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1027 /*isVolatile=*/false, /*AlwaysInline=*/false,
1031 /// LowerMemOpCallTo - Store the argument to the stack.
1033 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1034 SDValue StackPtr, SDValue Arg,
1035 DebugLoc dl, SelectionDAG &DAG,
1036 const CCValAssign &VA,
1037 ISD::ArgFlagsTy Flags) const {
1038 unsigned LocMemOffset = VA.getLocMemOffset();
1039 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1040 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1041 if (Flags.isByVal()) {
1042 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1044 return DAG.getStore(Chain, dl, Arg, PtrOff,
1045 PseudoSourceValue::getStack(), LocMemOffset,
1049 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1050 SDValue Chain, SDValue &Arg,
1051 RegsToPassVector &RegsToPass,
1052 CCValAssign &VA, CCValAssign &NextVA,
1054 SmallVector<SDValue, 8> &MemOpChains,
1055 ISD::ArgFlagsTy Flags) const {
1057 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1058 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1059 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1061 if (NextVA.isRegLoc())
1062 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1064 assert(NextVA.isMemLoc());
1065 if (StackPtr.getNode() == 0)
1066 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1068 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1074 /// LowerCall - Lowering a call into a callseq_start <-
1075 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1078 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1079 CallingConv::ID CallConv, bool isVarArg,
1081 const SmallVectorImpl<ISD::OutputArg> &Outs,
1082 const SmallVectorImpl<SDValue> &OutVals,
1083 const SmallVectorImpl<ISD::InputArg> &Ins,
1084 DebugLoc dl, SelectionDAG &DAG,
1085 SmallVectorImpl<SDValue> &InVals) const {
1086 MachineFunction &MF = DAG.getMachineFunction();
1087 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1088 bool IsSibCall = false;
1089 // Temporarily disable tail calls so things don't break.
1090 if (!EnableARMTailCalls)
1093 // Check if it's really possible to do a tail call.
1094 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1095 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1096 Outs, OutVals, Ins, DAG);
1097 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1098 // detected sibcalls.
1105 // Analyze operands of the call, assigning locations to each operand.
1106 SmallVector<CCValAssign, 16> ArgLocs;
1107 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1109 CCInfo.AnalyzeCallOperands(Outs,
1110 CCAssignFnForNode(CallConv, /* Return*/ false,
1113 // Get a count of how many bytes are to be pushed on the stack.
1114 unsigned NumBytes = CCInfo.getNextStackOffset();
1116 // For tail calls, memory operands are available in our caller's stack.
1120 // Adjust the stack pointer for the new arguments...
1121 // These operations are automatically eliminated by the prolog/epilog pass
1123 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1125 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1127 RegsToPassVector RegsToPass;
1128 SmallVector<SDValue, 8> MemOpChains;
1130 // Walk the register/memloc assignments, inserting copies/loads. In the case
1131 // of tail call optimization, arguments are handled later.
1132 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1134 ++i, ++realArgIdx) {
1135 CCValAssign &VA = ArgLocs[i];
1136 SDValue Arg = OutVals[realArgIdx];
1137 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1139 // Promote the value if needed.
1140 switch (VA.getLocInfo()) {
1141 default: llvm_unreachable("Unknown loc info!");
1142 case CCValAssign::Full: break;
1143 case CCValAssign::SExt:
1144 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1146 case CCValAssign::ZExt:
1147 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1149 case CCValAssign::AExt:
1150 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1152 case CCValAssign::BCvt:
1153 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1157 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1158 if (VA.needsCustom()) {
1159 if (VA.getLocVT() == MVT::v2f64) {
1160 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1161 DAG.getConstant(0, MVT::i32));
1162 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1163 DAG.getConstant(1, MVT::i32));
1165 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1166 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1168 VA = ArgLocs[++i]; // skip ahead to next loc
1169 if (VA.isRegLoc()) {
1170 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1171 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1173 assert(VA.isMemLoc());
1175 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1176 dl, DAG, VA, Flags));
1179 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1180 StackPtr, MemOpChains, Flags);
1182 } else if (VA.isRegLoc()) {
1183 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1184 } else if (!IsSibCall) {
1185 assert(VA.isMemLoc());
1187 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1188 dl, DAG, VA, Flags));
1192 if (!MemOpChains.empty())
1193 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1194 &MemOpChains[0], MemOpChains.size());
1196 // Build a sequence of copy-to-reg nodes chained together with token chain
1197 // and flag operands which copy the outgoing args into the appropriate regs.
1199 // Tail call byval lowering might overwrite argument registers so in case of
1200 // tail call optimization the copies to registers are lowered later.
1202 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1203 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1204 RegsToPass[i].second, InFlag);
1205 InFlag = Chain.getValue(1);
1208 // For tail calls lower the arguments to the 'real' stack slot.
1210 // Force all the incoming stack arguments to be loaded from the stack
1211 // before any new outgoing arguments are stored to the stack, because the
1212 // outgoing stack slots may alias the incoming argument stack slots, and
1213 // the alias isn't otherwise explicit. This is slightly more conservative
1214 // than necessary, because it means that each store effectively depends
1215 // on every argument instead of just those arguments it would clobber.
1217 // Do not flag preceeding copytoreg stuff together with the following stuff.
1219 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1220 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1221 RegsToPass[i].second, InFlag);
1222 InFlag = Chain.getValue(1);
1227 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1228 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1229 // node so that legalize doesn't hack it.
1230 bool isDirect = false;
1231 bool isARMFunc = false;
1232 bool isLocalARMFunc = false;
1233 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1235 if (EnableARMLongCalls) {
1236 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1237 && "long-calls with non-static relocation model!");
1238 // Handle a global address or an external symbol. If it's not one of
1239 // those, the target's already in a register, so we don't need to do
1241 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1242 const GlobalValue *GV = G->getGlobal();
1243 // Create a constant pool entry for the callee address
1244 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1245 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1248 // Get the address of the callee into a register
1249 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1250 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1251 Callee = DAG.getLoad(getPointerTy(), dl,
1252 DAG.getEntryNode(), CPAddr,
1253 PseudoSourceValue::getConstantPool(), 0,
1255 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1256 const char *Sym = S->getSymbol();
1258 // Create a constant pool entry for the callee address
1259 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1260 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1261 Sym, ARMPCLabelIndex, 0);
1262 // Get the address of the callee into a register
1263 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1264 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1265 Callee = DAG.getLoad(getPointerTy(), dl,
1266 DAG.getEntryNode(), CPAddr,
1267 PseudoSourceValue::getConstantPool(), 0,
1270 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1271 const GlobalValue *GV = G->getGlobal();
1273 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1274 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1275 getTargetMachine().getRelocationModel() != Reloc::Static;
1276 isARMFunc = !Subtarget->isThumb() || isStub;
1277 // ARM call to a local ARM function is predicable.
1278 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1279 // tBX takes a register source operand.
1280 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1281 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1282 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1285 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1286 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1287 Callee = DAG.getLoad(getPointerTy(), dl,
1288 DAG.getEntryNode(), CPAddr,
1289 PseudoSourceValue::getConstantPool(), 0,
1291 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1292 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1293 getPointerTy(), Callee, PICLabel);
1295 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1296 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1298 bool isStub = Subtarget->isTargetDarwin() &&
1299 getTargetMachine().getRelocationModel() != Reloc::Static;
1300 isARMFunc = !Subtarget->isThumb() || isStub;
1301 // tBX takes a register source operand.
1302 const char *Sym = S->getSymbol();
1303 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1304 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1305 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1306 Sym, ARMPCLabelIndex, 4);
1307 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1308 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1309 Callee = DAG.getLoad(getPointerTy(), dl,
1310 DAG.getEntryNode(), CPAddr,
1311 PseudoSourceValue::getConstantPool(), 0,
1313 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1314 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1315 getPointerTy(), Callee, PICLabel);
1317 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1320 // FIXME: handle tail calls differently.
1322 if (Subtarget->isThumb()) {
1323 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1324 CallOpc = ARMISD::CALL_NOLINK;
1326 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1328 CallOpc = (isDirect || Subtarget->hasV5TOps())
1329 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1330 : ARMISD::CALL_NOLINK;
1333 std::vector<SDValue> Ops;
1334 Ops.push_back(Chain);
1335 Ops.push_back(Callee);
1337 // Add argument registers to the end of the list so that they are known live
1339 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1340 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1341 RegsToPass[i].second.getValueType()));
1343 if (InFlag.getNode())
1344 Ops.push_back(InFlag);
1346 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1348 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1350 // Returns a chain and a flag for retval copy to use.
1351 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1352 InFlag = Chain.getValue(1);
1354 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1355 DAG.getIntPtrConstant(0, true), InFlag);
1357 InFlag = Chain.getValue(1);
1359 // Handle result values, copying them out of physregs into vregs that we
1361 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1365 /// MatchingStackOffset - Return true if the given stack call argument is
1366 /// already available in the same position (relatively) of the caller's
1367 /// incoming argument stack.
1369 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1370 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1371 const ARMInstrInfo *TII) {
1372 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1374 if (Arg.getOpcode() == ISD::CopyFromReg) {
1375 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1376 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1378 MachineInstr *Def = MRI->getVRegDef(VR);
1381 if (!Flags.isByVal()) {
1382 if (!TII->isLoadFromStackSlot(Def, FI))
1387 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1388 if (Flags.isByVal())
1389 // ByVal argument is passed in as a pointer but it's now being
1390 // dereferenced. e.g.
1391 // define @foo(%struct.X* %A) {
1392 // tail call @bar(%struct.X* byval %A)
1395 SDValue Ptr = Ld->getBasePtr();
1396 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1399 FI = FINode->getIndex();
1403 assert(FI != INT_MAX);
1404 if (!MFI->isFixedObjectIndex(FI))
1406 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1409 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1410 /// for tail call optimization. Targets which want to do tail call
1411 /// optimization should implement this function.
1413 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1414 CallingConv::ID CalleeCC,
1416 bool isCalleeStructRet,
1417 bool isCallerStructRet,
1418 const SmallVectorImpl<ISD::OutputArg> &Outs,
1419 const SmallVectorImpl<SDValue> &OutVals,
1420 const SmallVectorImpl<ISD::InputArg> &Ins,
1421 SelectionDAG& DAG) const {
1422 const Function *CallerF = DAG.getMachineFunction().getFunction();
1423 CallingConv::ID CallerCC = CallerF->getCallingConv();
1424 bool CCMatch = CallerCC == CalleeCC;
1426 // Look for obvious safe cases to perform tail call optimization that do not
1427 // require ABI changes. This is what gcc calls sibcall.
1429 // Do not sibcall optimize vararg calls unless the call site is not passing
1431 if (isVarArg && !Outs.empty())
1434 // Also avoid sibcall optimization if either caller or callee uses struct
1435 // return semantics.
1436 if (isCalleeStructRet || isCallerStructRet)
1439 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1440 // emitEpilogue is not ready for them.
1441 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1442 // LR. This means if we need to reload LR, it takes an extra instructions,
1443 // which outweighs the value of the tail call; but here we don't know yet
1444 // whether LR is going to be used. Probably the right approach is to
1445 // generate the tail call here and turn it back into CALL/RET in
1446 // emitEpilogue if LR is used.
1447 if (Subtarget->isThumb1Only())
1450 // For the moment, we can only do this to functions defined in this
1451 // compilation, or to indirect calls. A Thumb B to an ARM function,
1452 // or vice versa, is not easily fixed up in the linker unlike BL.
1453 // (We could do this by loading the address of the callee into a register;
1454 // that is an extra instruction over the direct call and burns a register
1455 // as well, so is not likely to be a win.)
1457 // It might be safe to remove this restriction on non-Darwin.
1459 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1460 // but we need to make sure there are enough registers; the only valid
1461 // registers are the 4 used for parameters. We don't currently do this
1463 if (isa<ExternalSymbolSDNode>(Callee))
1466 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1467 const GlobalValue *GV = G->getGlobal();
1468 if (GV->isDeclaration() || GV->isWeakForLinker())
1472 // If the calling conventions do not match, then we'd better make sure the
1473 // results are returned in the same way as what the caller expects.
1475 SmallVector<CCValAssign, 16> RVLocs1;
1476 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1477 RVLocs1, *DAG.getContext());
1478 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1480 SmallVector<CCValAssign, 16> RVLocs2;
1481 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1482 RVLocs2, *DAG.getContext());
1483 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1485 if (RVLocs1.size() != RVLocs2.size())
1487 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1488 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1490 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1492 if (RVLocs1[i].isRegLoc()) {
1493 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1496 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1502 // If the callee takes no arguments then go on to check the results of the
1504 if (!Outs.empty()) {
1505 // Check if stack adjustment is needed. For now, do not do this if any
1506 // argument is passed on the stack.
1507 SmallVector<CCValAssign, 16> ArgLocs;
1508 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1509 ArgLocs, *DAG.getContext());
1510 CCInfo.AnalyzeCallOperands(Outs,
1511 CCAssignFnForNode(CalleeCC, false, isVarArg));
1512 if (CCInfo.getNextStackOffset()) {
1513 MachineFunction &MF = DAG.getMachineFunction();
1515 // Check if the arguments are already laid out in the right way as
1516 // the caller's fixed stack objects.
1517 MachineFrameInfo *MFI = MF.getFrameInfo();
1518 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1519 const ARMInstrInfo *TII =
1520 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1521 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1523 ++i, ++realArgIdx) {
1524 CCValAssign &VA = ArgLocs[i];
1525 EVT RegVT = VA.getLocVT();
1526 SDValue Arg = OutVals[realArgIdx];
1527 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1528 if (VA.getLocInfo() == CCValAssign::Indirect)
1530 if (VA.needsCustom()) {
1531 // f64 and vector types are split into multiple registers or
1532 // register/stack-slot combinations. The types will not match
1533 // the registers; give up on memory f64 refs until we figure
1534 // out what to do about this.
1537 if (!ArgLocs[++i].isRegLoc())
1539 if (RegVT == MVT::v2f64) {
1540 if (!ArgLocs[++i].isRegLoc())
1542 if (!ArgLocs[++i].isRegLoc())
1545 } else if (!VA.isRegLoc()) {
1546 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1558 ARMTargetLowering::LowerReturn(SDValue Chain,
1559 CallingConv::ID CallConv, bool isVarArg,
1560 const SmallVectorImpl<ISD::OutputArg> &Outs,
1561 const SmallVectorImpl<SDValue> &OutVals,
1562 DebugLoc dl, SelectionDAG &DAG) const {
1564 // CCValAssign - represent the assignment of the return value to a location.
1565 SmallVector<CCValAssign, 16> RVLocs;
1567 // CCState - Info about the registers and stack slots.
1568 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1571 // Analyze outgoing return values.
1572 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1575 // If this is the first return lowered for this function, add
1576 // the regs to the liveout set for the function.
1577 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1578 for (unsigned i = 0; i != RVLocs.size(); ++i)
1579 if (RVLocs[i].isRegLoc())
1580 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1585 // Copy the result values into the output registers.
1586 for (unsigned i = 0, realRVLocIdx = 0;
1588 ++i, ++realRVLocIdx) {
1589 CCValAssign &VA = RVLocs[i];
1590 assert(VA.isRegLoc() && "Can only return in registers!");
1592 SDValue Arg = OutVals[realRVLocIdx];
1594 switch (VA.getLocInfo()) {
1595 default: llvm_unreachable("Unknown loc info!");
1596 case CCValAssign::Full: break;
1597 case CCValAssign::BCvt:
1598 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1602 if (VA.needsCustom()) {
1603 if (VA.getLocVT() == MVT::v2f64) {
1604 // Extract the first half and return it in two registers.
1605 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1606 DAG.getConstant(0, MVT::i32));
1607 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1608 DAG.getVTList(MVT::i32, MVT::i32), Half);
1610 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1611 Flag = Chain.getValue(1);
1612 VA = RVLocs[++i]; // skip ahead to next loc
1613 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1614 HalfGPRs.getValue(1), Flag);
1615 Flag = Chain.getValue(1);
1616 VA = RVLocs[++i]; // skip ahead to next loc
1618 // Extract the 2nd half and fall through to handle it as an f64 value.
1619 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1620 DAG.getConstant(1, MVT::i32));
1622 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1624 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1625 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1626 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1627 Flag = Chain.getValue(1);
1628 VA = RVLocs[++i]; // skip ahead to next loc
1629 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1632 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1634 // Guarantee that all emitted copies are
1635 // stuck together, avoiding something bad.
1636 Flag = Chain.getValue(1);
1641 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1643 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1648 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1649 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1650 // one of the above mentioned nodes. It has to be wrapped because otherwise
1651 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1652 // be used to form addressing mode. These wrapped nodes will be selected
1654 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1655 EVT PtrVT = Op.getValueType();
1656 // FIXME there is no actual debug info here
1657 DebugLoc dl = Op.getDebugLoc();
1658 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1660 if (CP->isMachineConstantPoolEntry())
1661 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1662 CP->getAlignment());
1664 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1665 CP->getAlignment());
1666 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1669 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1670 return MachineJumpTableInfo::EK_Inline;
1673 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1674 SelectionDAG &DAG) const {
1675 MachineFunction &MF = DAG.getMachineFunction();
1676 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1677 unsigned ARMPCLabelIndex = 0;
1678 DebugLoc DL = Op.getDebugLoc();
1679 EVT PtrVT = getPointerTy();
1680 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1681 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1683 if (RelocM == Reloc::Static) {
1684 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1686 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1687 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1688 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1689 ARMCP::CPBlockAddress,
1691 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1693 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1694 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1695 PseudoSourceValue::getConstantPool(), 0,
1697 if (RelocM == Reloc::Static)
1699 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1700 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1703 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1705 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1706 SelectionDAG &DAG) const {
1707 DebugLoc dl = GA->getDebugLoc();
1708 EVT PtrVT = getPointerTy();
1709 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1710 MachineFunction &MF = DAG.getMachineFunction();
1711 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1712 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1713 ARMConstantPoolValue *CPV =
1714 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1715 ARMCP::CPValue, PCAdj, "tlsgd", true);
1716 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1717 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1718 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1719 PseudoSourceValue::getConstantPool(), 0,
1721 SDValue Chain = Argument.getValue(1);
1723 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1724 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1726 // call __tls_get_addr.
1729 Entry.Node = Argument;
1730 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1731 Args.push_back(Entry);
1732 // FIXME: is there useful debug info available here?
1733 std::pair<SDValue, SDValue> CallResult =
1734 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1735 false, false, false, false,
1736 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1737 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1738 return CallResult.first;
1741 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1742 // "local exec" model.
1744 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1745 SelectionDAG &DAG) const {
1746 const GlobalValue *GV = GA->getGlobal();
1747 DebugLoc dl = GA->getDebugLoc();
1749 SDValue Chain = DAG.getEntryNode();
1750 EVT PtrVT = getPointerTy();
1751 // Get the Thread Pointer
1752 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1754 if (GV->isDeclaration()) {
1755 MachineFunction &MF = DAG.getMachineFunction();
1756 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1757 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1758 // Initial exec model.
1759 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1760 ARMConstantPoolValue *CPV =
1761 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1762 ARMCP::CPValue, PCAdj, "gottpoff", true);
1763 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1764 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1765 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1766 PseudoSourceValue::getConstantPool(), 0,
1768 Chain = Offset.getValue(1);
1770 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1771 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1773 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1774 PseudoSourceValue::getConstantPool(), 0,
1778 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1779 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1780 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1781 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1782 PseudoSourceValue::getConstantPool(), 0,
1786 // The address of the thread local variable is the add of the thread
1787 // pointer with the offset of the variable.
1788 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1792 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1793 // TODO: implement the "local dynamic" model
1794 assert(Subtarget->isTargetELF() &&
1795 "TLS not implemented for non-ELF targets");
1796 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1797 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1798 // otherwise use the "Local Exec" TLS Model
1799 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1800 return LowerToTLSGeneralDynamicModel(GA, DAG);
1802 return LowerToTLSExecModels(GA, DAG);
1805 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1806 SelectionDAG &DAG) const {
1807 EVT PtrVT = getPointerTy();
1808 DebugLoc dl = Op.getDebugLoc();
1809 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1810 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1811 if (RelocM == Reloc::PIC_) {
1812 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1813 ARMConstantPoolValue *CPV =
1814 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1815 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1816 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1817 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1819 PseudoSourceValue::getConstantPool(), 0,
1821 SDValue Chain = Result.getValue(1);
1822 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1823 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1825 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1826 PseudoSourceValue::getGOT(), 0,
1830 // If we have T2 ops, we can materialize the address directly via movt/movw
1831 // pair. This is always cheaper.
1832 if (Subtarget->useMovt()) {
1833 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1834 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
1836 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1837 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1838 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1839 PseudoSourceValue::getConstantPool(), 0,
1845 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1846 SelectionDAG &DAG) const {
1847 MachineFunction &MF = DAG.getMachineFunction();
1848 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1849 unsigned ARMPCLabelIndex = 0;
1850 EVT PtrVT = getPointerTy();
1851 DebugLoc dl = Op.getDebugLoc();
1852 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1853 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1855 if (RelocM == Reloc::Static)
1856 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1858 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1859 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1860 ARMConstantPoolValue *CPV =
1861 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1862 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1864 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1866 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1867 PseudoSourceValue::getConstantPool(), 0,
1869 SDValue Chain = Result.getValue(1);
1871 if (RelocM == Reloc::PIC_) {
1872 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1873 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1876 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1877 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1878 PseudoSourceValue::getGOT(), 0,
1884 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1885 SelectionDAG &DAG) const {
1886 assert(Subtarget->isTargetELF() &&
1887 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1888 MachineFunction &MF = DAG.getMachineFunction();
1889 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1890 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1891 EVT PtrVT = getPointerTy();
1892 DebugLoc dl = Op.getDebugLoc();
1893 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1894 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1895 "_GLOBAL_OFFSET_TABLE_",
1896 ARMPCLabelIndex, PCAdj);
1897 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1898 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1899 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1900 PseudoSourceValue::getConstantPool(), 0,
1902 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1903 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1907 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1908 DebugLoc dl = Op.getDebugLoc();
1909 SDValue Val = DAG.getConstant(0, MVT::i32);
1910 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1911 Op.getOperand(1), Val);
1915 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1916 DebugLoc dl = Op.getDebugLoc();
1917 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1918 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1922 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1923 const ARMSubtarget *Subtarget) const {
1924 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1925 DebugLoc dl = Op.getDebugLoc();
1927 default: return SDValue(); // Don't custom lower most intrinsics.
1928 case Intrinsic::arm_thread_pointer: {
1929 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1930 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1932 case Intrinsic::eh_sjlj_lsda: {
1933 MachineFunction &MF = DAG.getMachineFunction();
1934 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1935 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1936 EVT PtrVT = getPointerTy();
1937 DebugLoc dl = Op.getDebugLoc();
1938 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1940 unsigned PCAdj = (RelocM != Reloc::PIC_)
1941 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1942 ARMConstantPoolValue *CPV =
1943 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1944 ARMCP::CPLSDA, PCAdj);
1945 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1946 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1948 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1949 PseudoSourceValue::getConstantPool(), 0,
1952 if (RelocM == Reloc::PIC_) {
1953 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1954 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1961 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1962 const ARMSubtarget *Subtarget) {
1963 DebugLoc dl = Op.getDebugLoc();
1964 SDValue Op5 = Op.getOperand(5);
1965 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1966 // v6 and v7 can both handle barriers directly, but need handled a bit
1967 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1969 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1970 if (Subtarget->hasV7Ops())
1971 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1972 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1973 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1974 DAG.getConstant(0, MVT::i32));
1975 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1979 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1980 MachineFunction &MF = DAG.getMachineFunction();
1981 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1983 // vastart just stores the address of the VarArgsFrameIndex slot into the
1984 // memory location argument.
1985 DebugLoc dl = Op.getDebugLoc();
1986 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1987 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1988 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1989 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1994 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1995 SDValue &Root, SelectionDAG &DAG,
1996 DebugLoc dl) const {
1997 MachineFunction &MF = DAG.getMachineFunction();
1998 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2000 TargetRegisterClass *RC;
2001 if (AFI->isThumb1OnlyFunction())
2002 RC = ARM::tGPRRegisterClass;
2004 RC = ARM::GPRRegisterClass;
2006 // Transform the arguments stored in physical registers into virtual ones.
2007 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2008 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2011 if (NextVA.isMemLoc()) {
2012 MachineFrameInfo *MFI = MF.getFrameInfo();
2013 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2015 // Create load node to retrieve arguments from the stack.
2016 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2017 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2018 PseudoSourceValue::getFixedStack(FI), 0,
2021 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2022 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2025 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2029 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2030 CallingConv::ID CallConv, bool isVarArg,
2031 const SmallVectorImpl<ISD::InputArg>
2033 DebugLoc dl, SelectionDAG &DAG,
2034 SmallVectorImpl<SDValue> &InVals)
2037 MachineFunction &MF = DAG.getMachineFunction();
2038 MachineFrameInfo *MFI = MF.getFrameInfo();
2040 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2042 // Assign locations to all of the incoming arguments.
2043 SmallVector<CCValAssign, 16> ArgLocs;
2044 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2046 CCInfo.AnalyzeFormalArguments(Ins,
2047 CCAssignFnForNode(CallConv, /* Return*/ false,
2050 SmallVector<SDValue, 16> ArgValues;
2052 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2053 CCValAssign &VA = ArgLocs[i];
2055 // Arguments stored in registers.
2056 if (VA.isRegLoc()) {
2057 EVT RegVT = VA.getLocVT();
2060 if (VA.needsCustom()) {
2061 // f64 and vector types are split up into multiple registers or
2062 // combinations of registers and stack slots.
2063 if (VA.getLocVT() == MVT::v2f64) {
2064 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2066 VA = ArgLocs[++i]; // skip ahead to next loc
2068 if (VA.isMemLoc()) {
2069 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2070 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2071 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2072 PseudoSourceValue::getFixedStack(FI), 0,
2075 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2078 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2079 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2080 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2081 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2082 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2084 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2087 TargetRegisterClass *RC;
2089 if (RegVT == MVT::f32)
2090 RC = ARM::SPRRegisterClass;
2091 else if (RegVT == MVT::f64)
2092 RC = ARM::DPRRegisterClass;
2093 else if (RegVT == MVT::v2f64)
2094 RC = ARM::QPRRegisterClass;
2095 else if (RegVT == MVT::i32)
2096 RC = (AFI->isThumb1OnlyFunction() ?
2097 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2099 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2101 // Transform the arguments in physical registers into virtual ones.
2102 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2103 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2106 // If this is an 8 or 16-bit value, it is really passed promoted
2107 // to 32 bits. Insert an assert[sz]ext to capture this, then
2108 // truncate to the right size.
2109 switch (VA.getLocInfo()) {
2110 default: llvm_unreachable("Unknown loc info!");
2111 case CCValAssign::Full: break;
2112 case CCValAssign::BCvt:
2113 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2115 case CCValAssign::SExt:
2116 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2117 DAG.getValueType(VA.getValVT()));
2118 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2120 case CCValAssign::ZExt:
2121 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2122 DAG.getValueType(VA.getValVT()));
2123 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2127 InVals.push_back(ArgValue);
2129 } else { // VA.isRegLoc()
2132 assert(VA.isMemLoc());
2133 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2135 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2136 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
2138 // Create load nodes to retrieve arguments from the stack.
2139 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2140 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2141 PseudoSourceValue::getFixedStack(FI), 0,
2148 static const unsigned GPRArgRegs[] = {
2149 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2152 unsigned NumGPRs = CCInfo.getFirstUnallocated
2153 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2155 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2156 unsigned VARegSize = (4 - NumGPRs) * 4;
2157 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2158 unsigned ArgOffset = CCInfo.getNextStackOffset();
2159 if (VARegSaveSize) {
2160 // If this function is vararg, store any remaining integer argument regs
2161 // to their spots on the stack so that they may be loaded by deferencing
2162 // the result of va_next.
2163 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2164 AFI->setVarArgsFrameIndex(
2165 MFI->CreateFixedObject(VARegSaveSize,
2166 ArgOffset + VARegSaveSize - VARegSize,
2168 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2171 SmallVector<SDValue, 4> MemOps;
2172 for (; NumGPRs < 4; ++NumGPRs) {
2173 TargetRegisterClass *RC;
2174 if (AFI->isThumb1OnlyFunction())
2175 RC = ARM::tGPRRegisterClass;
2177 RC = ARM::GPRRegisterClass;
2179 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2180 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2182 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2183 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2184 0, false, false, 0);
2185 MemOps.push_back(Store);
2186 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2187 DAG.getConstant(4, getPointerTy()));
2189 if (!MemOps.empty())
2190 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2191 &MemOps[0], MemOps.size());
2193 // This will point to the next argument passed via stack.
2194 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2200 /// isFloatingPointZero - Return true if this is +0.0.
2201 static bool isFloatingPointZero(SDValue Op) {
2202 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2203 return CFP->getValueAPF().isPosZero();
2204 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2205 // Maybe this has already been legalized into the constant pool?
2206 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2207 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2208 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2209 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2210 return CFP->getValueAPF().isPosZero();
2216 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2217 /// the given operands.
2219 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2220 SDValue &ARMcc, SelectionDAG &DAG,
2221 DebugLoc dl) const {
2222 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2223 unsigned C = RHSC->getZExtValue();
2224 if (!isLegalICmpImmediate(C)) {
2225 // Constant does not fit, try adjusting it by one?
2230 if (isLegalICmpImmediate(C-1)) {
2231 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2232 RHS = DAG.getConstant(C-1, MVT::i32);
2237 if (C > 0 && isLegalICmpImmediate(C-1)) {
2238 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2239 RHS = DAG.getConstant(C-1, MVT::i32);
2244 if (isLegalICmpImmediate(C+1)) {
2245 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2246 RHS = DAG.getConstant(C+1, MVT::i32);
2251 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2252 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2253 RHS = DAG.getConstant(C+1, MVT::i32);
2260 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2261 ARMISD::NodeType CompareType;
2264 CompareType = ARMISD::CMP;
2269 CompareType = ARMISD::CMPZ;
2272 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2273 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2276 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2278 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2279 DebugLoc dl) const {
2281 if (!isFloatingPointZero(RHS))
2282 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2284 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2285 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2288 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2289 EVT VT = Op.getValueType();
2290 SDValue LHS = Op.getOperand(0);
2291 SDValue RHS = Op.getOperand(1);
2292 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2293 SDValue TrueVal = Op.getOperand(2);
2294 SDValue FalseVal = Op.getOperand(3);
2295 DebugLoc dl = Op.getDebugLoc();
2297 if (LHS.getValueType() == MVT::i32) {
2299 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2300 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2301 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2304 ARMCC::CondCodes CondCode, CondCode2;
2305 FPCCToARMCC(CC, CondCode, CondCode2);
2307 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2308 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2309 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2310 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2312 if (CondCode2 != ARMCC::AL) {
2313 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2314 // FIXME: Needs another CMP because flag can have but one use.
2315 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2316 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2317 Result, TrueVal, ARMcc2, CCR, Cmp2);
2322 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2323 /// to morph to an integer compare sequence.
2324 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2325 const ARMSubtarget *Subtarget) {
2326 SDNode *N = Op.getNode();
2327 if (!N->hasOneUse())
2328 // Otherwise it requires moving the value from fp to integer registers.
2330 if (!N->getNumValues())
2332 EVT VT = Op.getValueType();
2333 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2334 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2335 // vmrs are very slow, e.g. cortex-a8.
2338 if (isFloatingPointZero(Op)) {
2342 return ISD::isNormalLoad(N);
2345 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2346 if (isFloatingPointZero(Op))
2347 return DAG.getConstant(0, MVT::i32);
2349 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2350 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2351 Ld->getChain(), Ld->getBasePtr(),
2352 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2353 Ld->isVolatile(), Ld->isNonTemporal(),
2354 Ld->getAlignment());
2356 llvm_unreachable("Unknown VFP cmp argument!");
2359 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2360 SDValue &RetVal1, SDValue &RetVal2) {
2361 if (isFloatingPointZero(Op)) {
2362 RetVal1 = DAG.getConstant(0, MVT::i32);
2363 RetVal2 = DAG.getConstant(0, MVT::i32);
2367 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2368 SDValue Ptr = Ld->getBasePtr();
2369 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2370 Ld->getChain(), Ptr,
2371 Ld->getSrcValue(), Ld->getSrcValueOffset(),
2372 Ld->isVolatile(), Ld->isNonTemporal(),
2373 Ld->getAlignment());
2375 EVT PtrType = Ptr.getValueType();
2376 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2377 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2378 PtrType, Ptr, DAG.getConstant(4, PtrType));
2379 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2380 Ld->getChain(), NewPtr,
2381 Ld->getSrcValue(), Ld->getSrcValueOffset() + 4,
2382 Ld->isVolatile(), Ld->isNonTemporal(),
2387 llvm_unreachable("Unknown VFP cmp argument!");
2390 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2391 /// f32 and even f64 comparisons to integer ones.
2393 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2394 SDValue Chain = Op.getOperand(0);
2395 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2396 SDValue LHS = Op.getOperand(2);
2397 SDValue RHS = Op.getOperand(3);
2398 SDValue Dest = Op.getOperand(4);
2399 DebugLoc dl = Op.getDebugLoc();
2401 bool SeenZero = false;
2402 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2403 canChangeToInt(RHS, SeenZero, Subtarget) &&
2404 // If one of the operand is zero, it's safe to ignore the NaN case since
2405 // we only care about equality comparisons.
2406 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2407 // If unsafe fp math optimization is enabled and there are no othter uses of
2408 // the CMP operands, and the condition code is EQ oe NE, we can optimize it
2409 // to an integer comparison.
2410 if (CC == ISD::SETOEQ)
2412 else if (CC == ISD::SETUNE)
2416 if (LHS.getValueType() == MVT::f32) {
2417 LHS = bitcastf32Toi32(LHS, DAG);
2418 RHS = bitcastf32Toi32(RHS, DAG);
2419 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2420 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2421 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2422 Chain, Dest, ARMcc, CCR, Cmp);
2427 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2428 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2429 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2430 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2431 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2432 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2433 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2439 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2440 SDValue Chain = Op.getOperand(0);
2441 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2442 SDValue LHS = Op.getOperand(2);
2443 SDValue RHS = Op.getOperand(3);
2444 SDValue Dest = Op.getOperand(4);
2445 DebugLoc dl = Op.getDebugLoc();
2447 if (LHS.getValueType() == MVT::i32) {
2449 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2450 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2451 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2452 Chain, Dest, ARMcc, CCR, Cmp);
2455 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2458 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2459 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2460 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2461 if (Result.getNode())
2465 ARMCC::CondCodes CondCode, CondCode2;
2466 FPCCToARMCC(CC, CondCode, CondCode2);
2468 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2469 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2470 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2471 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2472 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2473 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2474 if (CondCode2 != ARMCC::AL) {
2475 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2476 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2477 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2482 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2483 SDValue Chain = Op.getOperand(0);
2484 SDValue Table = Op.getOperand(1);
2485 SDValue Index = Op.getOperand(2);
2486 DebugLoc dl = Op.getDebugLoc();
2488 EVT PTy = getPointerTy();
2489 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2490 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2491 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2492 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2493 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2494 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2495 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2496 if (Subtarget->isThumb2()) {
2497 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2498 // which does another jump to the destination. This also makes it easier
2499 // to translate it to TBB / TBH later.
2500 // FIXME: This might not work if the function is extremely large.
2501 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2502 Addr, Op.getOperand(2), JTI, UId);
2504 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2505 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2506 PseudoSourceValue::getJumpTable(), 0,
2508 Chain = Addr.getValue(1);
2509 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2510 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2512 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2513 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2514 Chain = Addr.getValue(1);
2515 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2519 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2520 DebugLoc dl = Op.getDebugLoc();
2523 switch (Op.getOpcode()) {
2525 assert(0 && "Invalid opcode!");
2526 case ISD::FP_TO_SINT:
2527 Opc = ARMISD::FTOSI;
2529 case ISD::FP_TO_UINT:
2530 Opc = ARMISD::FTOUI;
2533 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2534 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2537 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2538 EVT VT = Op.getValueType();
2539 DebugLoc dl = Op.getDebugLoc();
2542 switch (Op.getOpcode()) {
2544 assert(0 && "Invalid opcode!");
2545 case ISD::SINT_TO_FP:
2546 Opc = ARMISD::SITOF;
2548 case ISD::UINT_TO_FP:
2549 Opc = ARMISD::UITOF;
2553 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2554 return DAG.getNode(Opc, dl, VT, Op);
2557 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2558 // Implement fcopysign with a fabs and a conditional fneg.
2559 SDValue Tmp0 = Op.getOperand(0);
2560 SDValue Tmp1 = Op.getOperand(1);
2561 DebugLoc dl = Op.getDebugLoc();
2562 EVT VT = Op.getValueType();
2563 EVT SrcVT = Tmp1.getValueType();
2564 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2565 SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
2566 SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
2567 SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
2568 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2569 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
2572 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2573 MachineFunction &MF = DAG.getMachineFunction();
2574 MachineFrameInfo *MFI = MF.getFrameInfo();
2575 MFI->setReturnAddressIsTaken(true);
2577 EVT VT = Op.getValueType();
2578 DebugLoc dl = Op.getDebugLoc();
2579 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2581 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2582 SDValue Offset = DAG.getConstant(4, MVT::i32);
2583 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2584 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2585 NULL, 0, false, false, 0);
2588 // Return LR, which contains the return address. Mark it an implicit live-in.
2589 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2590 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2593 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2594 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2595 MFI->setFrameAddressIsTaken(true);
2597 EVT VT = Op.getValueType();
2598 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2599 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2600 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2601 ? ARM::R7 : ARM::R11;
2602 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2604 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2609 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2610 /// expand a bit convert where either the source or destination type is i64 to
2611 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2612 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2613 /// vectors), since the legalizer won't know what to do with that.
2614 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2615 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2616 DebugLoc dl = N->getDebugLoc();
2617 SDValue Op = N->getOperand(0);
2619 // This function is only supposed to be called for i64 types, either as the
2620 // source or destination of the bit convert.
2621 EVT SrcVT = Op.getValueType();
2622 EVT DstVT = N->getValueType(0);
2623 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2624 "ExpandBIT_CONVERT called for non-i64 type");
2626 // Turn i64->f64 into VMOVDRR.
2627 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2628 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2629 DAG.getConstant(0, MVT::i32));
2630 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2631 DAG.getConstant(1, MVT::i32));
2632 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2633 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2636 // Turn f64->i64 into VMOVRRD.
2637 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2638 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2639 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2640 // Merge the pieces into a single i64 value.
2641 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2647 /// getZeroVector - Returns a vector of specified type with all zero elements.
2648 /// Zero vectors are used to represent vector negation and in those cases
2649 /// will be implemented with the NEON VNEG instruction. However, VNEG does
2650 /// not support i64 elements, so sometimes the zero vectors will need to be
2651 /// explicitly constructed. Regardless, use a canonical VMOV to create the
2653 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2654 assert(VT.isVector() && "Expected a vector type");
2655 // The canonical modified immediate encoding of a zero vector is....0!
2656 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
2657 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
2658 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
2659 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
2662 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2663 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2664 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2665 SelectionDAG &DAG) const {
2666 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2667 EVT VT = Op.getValueType();
2668 unsigned VTBits = VT.getSizeInBits();
2669 DebugLoc dl = Op.getDebugLoc();
2670 SDValue ShOpLo = Op.getOperand(0);
2671 SDValue ShOpHi = Op.getOperand(1);
2672 SDValue ShAmt = Op.getOperand(2);
2674 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2676 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2678 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2679 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2680 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2681 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2682 DAG.getConstant(VTBits, MVT::i32));
2683 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2684 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2685 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2687 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2688 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2690 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2691 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
2694 SDValue Ops[2] = { Lo, Hi };
2695 return DAG.getMergeValues(Ops, 2, dl);
2698 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2699 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2700 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2701 SelectionDAG &DAG) const {
2702 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2703 EVT VT = Op.getValueType();
2704 unsigned VTBits = VT.getSizeInBits();
2705 DebugLoc dl = Op.getDebugLoc();
2706 SDValue ShOpLo = Op.getOperand(0);
2707 SDValue ShOpHi = Op.getOperand(1);
2708 SDValue ShAmt = Op.getOperand(2);
2711 assert(Op.getOpcode() == ISD::SHL_PARTS);
2712 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2713 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2714 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2715 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2716 DAG.getConstant(VTBits, MVT::i32));
2717 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2718 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2720 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2721 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2722 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2724 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2725 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
2728 SDValue Ops[2] = { Lo, Hi };
2729 return DAG.getMergeValues(Ops, 2, dl);
2732 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2733 const ARMSubtarget *ST) {
2734 EVT VT = N->getValueType(0);
2735 DebugLoc dl = N->getDebugLoc();
2737 if (!ST->hasV6T2Ops())
2740 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2741 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2744 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2745 const ARMSubtarget *ST) {
2746 EVT VT = N->getValueType(0);
2747 DebugLoc dl = N->getDebugLoc();
2749 // Lower vector shifts on NEON to use VSHL.
2750 if (VT.isVector()) {
2751 assert(ST->hasNEON() && "unexpected vector shift");
2753 // Left shifts translate directly to the vshiftu intrinsic.
2754 if (N->getOpcode() == ISD::SHL)
2755 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2756 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2757 N->getOperand(0), N->getOperand(1));
2759 assert((N->getOpcode() == ISD::SRA ||
2760 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2762 // NEON uses the same intrinsics for both left and right shifts. For
2763 // right shifts, the shift amounts are negative, so negate the vector of
2765 EVT ShiftVT = N->getOperand(1).getValueType();
2766 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2767 getZeroVector(ShiftVT, DAG, dl),
2769 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2770 Intrinsic::arm_neon_vshifts :
2771 Intrinsic::arm_neon_vshiftu);
2772 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2773 DAG.getConstant(vshiftInt, MVT::i32),
2774 N->getOperand(0), NegatedCount);
2777 // We can get here for a node like i32 = ISD::SHL i32, i64
2781 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2782 "Unknown shift to lower!");
2784 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2785 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2786 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2789 // If we are in thumb mode, we don't have RRX.
2790 if (ST->isThumb1Only()) return SDValue();
2792 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2793 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2794 DAG.getConstant(0, MVT::i32));
2795 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2796 DAG.getConstant(1, MVT::i32));
2798 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2799 // captures the result into a carry flag.
2800 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2801 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2803 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2804 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2806 // Merge the pieces into a single i64 value.
2807 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2810 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2811 SDValue TmpOp0, TmpOp1;
2812 bool Invert = false;
2816 SDValue Op0 = Op.getOperand(0);
2817 SDValue Op1 = Op.getOperand(1);
2818 SDValue CC = Op.getOperand(2);
2819 EVT VT = Op.getValueType();
2820 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2821 DebugLoc dl = Op.getDebugLoc();
2823 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2824 switch (SetCCOpcode) {
2825 default: llvm_unreachable("Illegal FP comparison"); break;
2827 case ISD::SETNE: Invert = true; // Fallthrough
2829 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2831 case ISD::SETLT: Swap = true; // Fallthrough
2833 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2835 case ISD::SETLE: Swap = true; // Fallthrough
2837 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2838 case ISD::SETUGE: Swap = true; // Fallthrough
2839 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2840 case ISD::SETUGT: Swap = true; // Fallthrough
2841 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2842 case ISD::SETUEQ: Invert = true; // Fallthrough
2844 // Expand this to (OLT | OGT).
2848 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2849 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2851 case ISD::SETUO: Invert = true; // Fallthrough
2853 // Expand this to (OLT | OGE).
2857 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2858 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2862 // Integer comparisons.
2863 switch (SetCCOpcode) {
2864 default: llvm_unreachable("Illegal integer comparison"); break;
2865 case ISD::SETNE: Invert = true;
2866 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2867 case ISD::SETLT: Swap = true;
2868 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2869 case ISD::SETLE: Swap = true;
2870 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2871 case ISD::SETULT: Swap = true;
2872 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2873 case ISD::SETULE: Swap = true;
2874 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2877 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2878 if (Opc == ARMISD::VCEQ) {
2881 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2883 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2886 // Ignore bitconvert.
2887 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2888 AndOp = AndOp.getOperand(0);
2890 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2892 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2893 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2900 std::swap(Op0, Op1);
2902 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2905 Result = DAG.getNOT(dl, Result, VT);
2910 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2911 /// valid vector constant for a NEON instruction with a "modified immediate"
2912 /// operand (e.g., VMOV). If so, return the encoded value.
2913 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2914 unsigned SplatBitSize, SelectionDAG &DAG,
2915 EVT &VT, bool is128Bits, bool isVMOV) {
2916 unsigned OpCmode, Imm;
2918 // SplatBitSize is set to the smallest size that splats the vector, so a
2919 // zero vector will always have SplatBitSize == 8. However, NEON modified
2920 // immediate instructions others than VMOV do not support the 8-bit encoding
2921 // of a zero vector, and the default encoding of zero is supposed to be the
2926 switch (SplatBitSize) {
2930 // Any 1-byte value is OK. Op=0, Cmode=1110.
2931 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2934 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
2938 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2939 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
2940 if ((SplatBits & ~0xff) == 0) {
2941 // Value = 0x00nn: Op=x, Cmode=100x.
2946 if ((SplatBits & ~0xff00) == 0) {
2947 // Value = 0xnn00: Op=x, Cmode=101x.
2949 Imm = SplatBits >> 8;
2955 // NEON's 32-bit VMOV supports splat values where:
2956 // * only one byte is nonzero, or
2957 // * the least significant byte is 0xff and the second byte is nonzero, or
2958 // * the least significant 2 bytes are 0xff and the third is nonzero.
2959 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
2960 if ((SplatBits & ~0xff) == 0) {
2961 // Value = 0x000000nn: Op=x, Cmode=000x.
2966 if ((SplatBits & ~0xff00) == 0) {
2967 // Value = 0x0000nn00: Op=x, Cmode=001x.
2969 Imm = SplatBits >> 8;
2972 if ((SplatBits & ~0xff0000) == 0) {
2973 // Value = 0x00nn0000: Op=x, Cmode=010x.
2975 Imm = SplatBits >> 16;
2978 if ((SplatBits & ~0xff000000) == 0) {
2979 // Value = 0xnn000000: Op=x, Cmode=011x.
2981 Imm = SplatBits >> 24;
2985 if ((SplatBits & ~0xffff) == 0 &&
2986 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2987 // Value = 0x0000nnff: Op=x, Cmode=1100.
2989 Imm = SplatBits >> 8;
2994 if ((SplatBits & ~0xffffff) == 0 &&
2995 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2996 // Value = 0x00nnffff: Op=x, Cmode=1101.
2998 Imm = SplatBits >> 16;
2999 SplatBits |= 0xffff;
3003 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3004 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3005 // VMOV.I32. A (very) minor optimization would be to replicate the value
3006 // and fall through here to test for a valid 64-bit splat. But, then the
3007 // caller would also need to check and handle the change in size.
3013 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3014 uint64_t BitMask = 0xff;
3016 unsigned ImmMask = 1;
3018 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3019 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3022 } else if ((SplatBits & BitMask) != 0) {
3028 // Op=1, Cmode=1110.
3031 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3036 llvm_unreachable("unexpected size for isNEONModifiedImm");
3040 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3041 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3044 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3045 bool &ReverseVEXT, unsigned &Imm) {
3046 unsigned NumElts = VT.getVectorNumElements();
3047 ReverseVEXT = false;
3050 // If this is a VEXT shuffle, the immediate value is the index of the first
3051 // element. The other shuffle indices must be the successive elements after
3053 unsigned ExpectedElt = Imm;
3054 for (unsigned i = 1; i < NumElts; ++i) {
3055 // Increment the expected index. If it wraps around, it may still be
3056 // a VEXT but the source vectors must be swapped.
3058 if (ExpectedElt == NumElts * 2) {
3063 if (ExpectedElt != static_cast<unsigned>(M[i]))
3067 // Adjust the index value if the source operands will be swapped.
3074 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3075 /// instruction with the specified blocksize. (The order of the elements
3076 /// within each block of the vector is reversed.)
3077 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3078 unsigned BlockSize) {
3079 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3080 "Only possible block sizes for VREV are: 16, 32, 64");
3082 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3086 unsigned NumElts = VT.getVectorNumElements();
3087 unsigned BlockElts = M[0] + 1;
3089 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3092 for (unsigned i = 0; i < NumElts; ++i) {
3093 if ((unsigned) M[i] !=
3094 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3101 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3102 unsigned &WhichResult) {
3103 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3107 unsigned NumElts = VT.getVectorNumElements();
3108 WhichResult = (M[0] == 0 ? 0 : 1);
3109 for (unsigned i = 0; i < NumElts; i += 2) {
3110 if ((unsigned) M[i] != i + WhichResult ||
3111 (unsigned) M[i+1] != i + NumElts + WhichResult)
3117 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3118 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3119 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3120 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3121 unsigned &WhichResult) {
3122 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3126 unsigned NumElts = VT.getVectorNumElements();
3127 WhichResult = (M[0] == 0 ? 0 : 1);
3128 for (unsigned i = 0; i < NumElts; i += 2) {
3129 if ((unsigned) M[i] != i + WhichResult ||
3130 (unsigned) M[i+1] != i + WhichResult)
3136 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3137 unsigned &WhichResult) {
3138 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3142 unsigned NumElts = VT.getVectorNumElements();
3143 WhichResult = (M[0] == 0 ? 0 : 1);
3144 for (unsigned i = 0; i != NumElts; ++i) {
3145 if ((unsigned) M[i] != 2 * i + WhichResult)
3149 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3150 if (VT.is64BitVector() && EltSz == 32)
3156 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3157 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3158 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3159 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3160 unsigned &WhichResult) {
3161 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3165 unsigned Half = VT.getVectorNumElements() / 2;
3166 WhichResult = (M[0] == 0 ? 0 : 1);
3167 for (unsigned j = 0; j != 2; ++j) {
3168 unsigned Idx = WhichResult;
3169 for (unsigned i = 0; i != Half; ++i) {
3170 if ((unsigned) M[i + j * Half] != Idx)
3176 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3177 if (VT.is64BitVector() && EltSz == 32)
3183 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3184 unsigned &WhichResult) {
3185 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3189 unsigned NumElts = VT.getVectorNumElements();
3190 WhichResult = (M[0] == 0 ? 0 : 1);
3191 unsigned Idx = WhichResult * NumElts / 2;
3192 for (unsigned i = 0; i != NumElts; i += 2) {
3193 if ((unsigned) M[i] != Idx ||
3194 (unsigned) M[i+1] != Idx + NumElts)
3199 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3200 if (VT.is64BitVector() && EltSz == 32)
3206 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3207 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3208 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3209 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3210 unsigned &WhichResult) {
3211 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3215 unsigned NumElts = VT.getVectorNumElements();
3216 WhichResult = (M[0] == 0 ? 0 : 1);
3217 unsigned Idx = WhichResult * NumElts / 2;
3218 for (unsigned i = 0; i != NumElts; i += 2) {
3219 if ((unsigned) M[i] != Idx ||
3220 (unsigned) M[i+1] != Idx)
3225 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3226 if (VT.is64BitVector() && EltSz == 32)
3232 // If this is a case we can't handle, return null and let the default
3233 // expansion code take care of it.
3234 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3235 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3236 DebugLoc dl = Op.getDebugLoc();
3237 EVT VT = Op.getValueType();
3239 APInt SplatBits, SplatUndef;
3240 unsigned SplatBitSize;
3242 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3243 if (SplatBitSize <= 64) {
3244 // Check if an immediate VMOV works.
3246 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3247 SplatUndef.getZExtValue(), SplatBitSize,
3248 DAG, VmovVT, VT.is128BitVector(), true);
3249 if (Val.getNode()) {
3250 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3251 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3254 // Try an immediate VMVN.
3255 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3256 ((1LL << SplatBitSize) - 1));
3257 Val = isNEONModifiedImm(NegatedImm,
3258 SplatUndef.getZExtValue(), SplatBitSize,
3259 DAG, VmovVT, VT.is128BitVector(), false);
3260 if (Val.getNode()) {
3261 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3262 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
3267 // Scan through the operands to see if only one value is used.
3268 unsigned NumElts = VT.getVectorNumElements();
3269 bool isOnlyLowElement = true;
3270 bool usesOnlyOneValue = true;
3271 bool isConstant = true;
3273 for (unsigned i = 0; i < NumElts; ++i) {
3274 SDValue V = Op.getOperand(i);
3275 if (V.getOpcode() == ISD::UNDEF)
3278 isOnlyLowElement = false;
3279 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3282 if (!Value.getNode())
3284 else if (V != Value)
3285 usesOnlyOneValue = false;
3288 if (!Value.getNode())
3289 return DAG.getUNDEF(VT);
3291 if (isOnlyLowElement)
3292 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3294 // If all elements are constants, fall back to the default expansion, which
3295 // will generate a load from the constant pool.
3299 // Use VDUP for non-constant splats.
3300 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3301 if (usesOnlyOneValue && EltSize <= 32)
3302 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3304 // Vectors with 32- or 64-bit elements can be built by directly assigning
3305 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3306 // will be legalized.
3307 if (EltSize >= 32) {
3308 // Do the expansion with floating-point types, since that is what the VFP
3309 // registers are defined to use, and since i64 is not legal.
3310 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3311 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3312 SmallVector<SDValue, 8> Ops;
3313 for (unsigned i = 0; i < NumElts; ++i)
3314 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3315 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3316 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3322 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3323 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3324 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3325 /// are assumed to be legal.
3327 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3329 if (VT.getVectorNumElements() == 4 &&
3330 (VT.is128BitVector() || VT.is64BitVector())) {
3331 unsigned PFIndexes[4];
3332 for (unsigned i = 0; i != 4; ++i) {
3336 PFIndexes[i] = M[i];
3339 // Compute the index in the perfect shuffle table.
3340 unsigned PFTableIndex =
3341 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3342 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3343 unsigned Cost = (PFEntry >> 30);
3350 unsigned Imm, WhichResult;
3352 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3353 return (EltSize >= 32 ||
3354 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3355 isVREVMask(M, VT, 64) ||
3356 isVREVMask(M, VT, 32) ||
3357 isVREVMask(M, VT, 16) ||
3358 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3359 isVTRNMask(M, VT, WhichResult) ||
3360 isVUZPMask(M, VT, WhichResult) ||
3361 isVZIPMask(M, VT, WhichResult) ||
3362 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3363 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3364 isVZIP_v_undef_Mask(M, VT, WhichResult));
3367 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3368 /// the specified operations to build the shuffle.
3369 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3370 SDValue RHS, SelectionDAG &DAG,
3372 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3373 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3374 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3377 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3386 OP_VUZPL, // VUZP, left result
3387 OP_VUZPR, // VUZP, right result
3388 OP_VZIPL, // VZIP, left result
3389 OP_VZIPR, // VZIP, right result
3390 OP_VTRNL, // VTRN, left result
3391 OP_VTRNR // VTRN, right result
3394 if (OpNum == OP_COPY) {
3395 if (LHSID == (1*9+2)*9+3) return LHS;
3396 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3400 SDValue OpLHS, OpRHS;
3401 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3402 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3403 EVT VT = OpLHS.getValueType();
3406 default: llvm_unreachable("Unknown shuffle opcode!");
3408 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3413 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3414 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3418 return DAG.getNode(ARMISD::VEXT, dl, VT,
3420 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3423 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3424 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3427 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3428 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3431 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3432 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3436 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3437 SDValue V1 = Op.getOperand(0);
3438 SDValue V2 = Op.getOperand(1);
3439 DebugLoc dl = Op.getDebugLoc();
3440 EVT VT = Op.getValueType();
3441 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3442 SmallVector<int, 8> ShuffleMask;
3444 // Convert shuffles that are directly supported on NEON to target-specific
3445 // DAG nodes, instead of keeping them as shuffles and matching them again
3446 // during code selection. This is more efficient and avoids the possibility
3447 // of inconsistencies between legalization and selection.
3448 // FIXME: floating-point vectors should be canonicalized to integer vectors
3449 // of the same time so that they get CSEd properly.
3450 SVN->getMask(ShuffleMask);
3452 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3453 if (EltSize <= 32) {
3454 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3455 int Lane = SVN->getSplatIndex();
3456 // If this is undef splat, generate it via "just" vdup, if possible.
3457 if (Lane == -1) Lane = 0;
3459 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3460 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3462 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3463 DAG.getConstant(Lane, MVT::i32));
3468 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3471 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3472 DAG.getConstant(Imm, MVT::i32));
3475 if (isVREVMask(ShuffleMask, VT, 64))
3476 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3477 if (isVREVMask(ShuffleMask, VT, 32))
3478 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3479 if (isVREVMask(ShuffleMask, VT, 16))
3480 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3482 // Check for Neon shuffles that modify both input vectors in place.
3483 // If both results are used, i.e., if there are two shuffles with the same
3484 // source operands and with masks corresponding to both results of one of
3485 // these operations, DAG memoization will ensure that a single node is
3486 // used for both shuffles.
3487 unsigned WhichResult;
3488 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3489 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3490 V1, V2).getValue(WhichResult);
3491 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3492 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3493 V1, V2).getValue(WhichResult);
3494 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3495 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3496 V1, V2).getValue(WhichResult);
3498 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3499 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3500 V1, V1).getValue(WhichResult);
3501 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3502 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3503 V1, V1).getValue(WhichResult);
3504 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3505 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3506 V1, V1).getValue(WhichResult);
3509 // If the shuffle is not directly supported and it has 4 elements, use
3510 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3511 unsigned NumElts = VT.getVectorNumElements();
3513 unsigned PFIndexes[4];
3514 for (unsigned i = 0; i != 4; ++i) {
3515 if (ShuffleMask[i] < 0)
3518 PFIndexes[i] = ShuffleMask[i];
3521 // Compute the index in the perfect shuffle table.
3522 unsigned PFTableIndex =
3523 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3524 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3525 unsigned Cost = (PFEntry >> 30);
3528 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3531 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3532 if (EltSize >= 32) {
3533 // Do the expansion with floating-point types, since that is what the VFP
3534 // registers are defined to use, and since i64 is not legal.
3535 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3536 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3537 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3538 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3539 SmallVector<SDValue, 8> Ops;
3540 for (unsigned i = 0; i < NumElts; ++i) {
3541 if (ShuffleMask[i] < 0)
3542 Ops.push_back(DAG.getUNDEF(EltVT));
3544 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3545 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3546 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3549 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3550 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3556 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3557 EVT VT = Op.getValueType();
3558 DebugLoc dl = Op.getDebugLoc();
3559 SDValue Vec = Op.getOperand(0);
3560 SDValue Lane = Op.getOperand(1);
3561 assert(VT == MVT::i32 &&
3562 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3563 "unexpected type for custom-lowering vector extract");
3564 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3567 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3568 // The only time a CONCAT_VECTORS operation can have legal types is when
3569 // two 64-bit vectors are concatenated to a 128-bit vector.
3570 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3571 "unexpected CONCAT_VECTORS");
3572 DebugLoc dl = Op.getDebugLoc();
3573 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3574 SDValue Op0 = Op.getOperand(0);
3575 SDValue Op1 = Op.getOperand(1);
3576 if (Op0.getOpcode() != ISD::UNDEF)
3577 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3578 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3579 DAG.getIntPtrConstant(0));
3580 if (Op1.getOpcode() != ISD::UNDEF)
3581 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3582 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3583 DAG.getIntPtrConstant(1));
3584 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3587 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3588 switch (Op.getOpcode()) {
3589 default: llvm_unreachable("Don't know how to custom lower this!");
3590 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3591 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3592 case ISD::GlobalAddress:
3593 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3594 LowerGlobalAddressELF(Op, DAG);
3595 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3596 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3597 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3598 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3599 case ISD::VASTART: return LowerVASTART(Op, DAG);
3600 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3601 case ISD::SINT_TO_FP:
3602 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3603 case ISD::FP_TO_SINT:
3604 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3605 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3606 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3607 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3608 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3609 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3610 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3611 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3613 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3616 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3617 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3618 case ISD::SRL_PARTS:
3619 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3620 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3621 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3622 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3623 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3624 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3625 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3630 /// ReplaceNodeResults - Replace the results of node with an illegal result
3631 /// type with new values built out of custom code.
3632 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3633 SmallVectorImpl<SDValue>&Results,
3634 SelectionDAG &DAG) const {
3636 switch (N->getOpcode()) {
3638 llvm_unreachable("Don't know how to custom expand this!");
3640 case ISD::BIT_CONVERT:
3641 Res = ExpandBIT_CONVERT(N, DAG);
3645 Res = LowerShift(N, DAG, Subtarget);
3649 Results.push_back(Res);
3652 //===----------------------------------------------------------------------===//
3653 // ARM Scheduler Hooks
3654 //===----------------------------------------------------------------------===//
3657 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3658 MachineBasicBlock *BB,
3659 unsigned Size) const {
3660 unsigned dest = MI->getOperand(0).getReg();
3661 unsigned ptr = MI->getOperand(1).getReg();
3662 unsigned oldval = MI->getOperand(2).getReg();
3663 unsigned newval = MI->getOperand(3).getReg();
3664 unsigned scratch = BB->getParent()->getRegInfo()
3665 .createVirtualRegister(ARM::GPRRegisterClass);
3666 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3667 DebugLoc dl = MI->getDebugLoc();
3668 bool isThumb2 = Subtarget->isThumb2();
3670 unsigned ldrOpc, strOpc;
3672 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3674 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3675 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3678 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3679 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3682 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3683 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3687 MachineFunction *MF = BB->getParent();
3688 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3689 MachineFunction::iterator It = BB;
3690 ++It; // insert the new blocks after the current block
3692 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3693 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3694 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3695 MF->insert(It, loop1MBB);
3696 MF->insert(It, loop2MBB);
3697 MF->insert(It, exitMBB);
3699 // Transfer the remainder of BB and its successor edges to exitMBB.
3700 exitMBB->splice(exitMBB->begin(), BB,
3701 llvm::next(MachineBasicBlock::iterator(MI)),
3703 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3707 // fallthrough --> loop1MBB
3708 BB->addSuccessor(loop1MBB);
3711 // ldrex dest, [ptr]
3715 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3716 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3717 .addReg(dest).addReg(oldval));
3718 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3719 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3720 BB->addSuccessor(loop2MBB);
3721 BB->addSuccessor(exitMBB);
3724 // strex scratch, newval, [ptr]
3728 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3730 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3731 .addReg(scratch).addImm(0));
3732 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3733 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3734 BB->addSuccessor(loop1MBB);
3735 BB->addSuccessor(exitMBB);
3741 MI->eraseFromParent(); // The instruction is gone now.
3747 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3748 unsigned Size, unsigned BinOpcode) const {
3749 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3750 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3752 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3753 MachineFunction *MF = BB->getParent();
3754 MachineFunction::iterator It = BB;
3757 unsigned dest = MI->getOperand(0).getReg();
3758 unsigned ptr = MI->getOperand(1).getReg();
3759 unsigned incr = MI->getOperand(2).getReg();
3760 DebugLoc dl = MI->getDebugLoc();
3762 bool isThumb2 = Subtarget->isThumb2();
3763 unsigned ldrOpc, strOpc;
3765 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3767 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3768 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3771 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3772 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3775 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3776 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3780 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3781 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3782 MF->insert(It, loopMBB);
3783 MF->insert(It, exitMBB);
3785 // Transfer the remainder of BB and its successor edges to exitMBB.
3786 exitMBB->splice(exitMBB->begin(), BB,
3787 llvm::next(MachineBasicBlock::iterator(MI)),
3789 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
3791 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3792 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3793 unsigned scratch2 = (!BinOpcode) ? incr :
3794 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3798 // fallthrough --> loopMBB
3799 BB->addSuccessor(loopMBB);
3803 // <binop> scratch2, dest, incr
3804 // strex scratch, scratch2, ptr
3807 // fallthrough --> exitMBB
3809 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3811 // operand order needs to go the other way for NAND
3812 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3813 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3814 addReg(incr).addReg(dest)).addReg(0);
3816 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3817 addReg(dest).addReg(incr)).addReg(0);
3820 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3822 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3823 .addReg(scratch).addImm(0));
3824 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3825 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3827 BB->addSuccessor(loopMBB);
3828 BB->addSuccessor(exitMBB);
3834 MI->eraseFromParent(); // The instruction is gone now.
3840 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
3841 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
3842 E = MBB->succ_end(); I != E; ++I)
3845 llvm_unreachable("Expecting a BB with two successors!");
3849 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3850 MachineBasicBlock *BB) const {
3851 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3852 DebugLoc dl = MI->getDebugLoc();
3853 bool isThumb2 = Subtarget->isThumb2();
3854 switch (MI->getOpcode()) {
3857 llvm_unreachable("Unexpected instr type to insert");
3859 case ARM::ATOMIC_LOAD_ADD_I8:
3860 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3861 case ARM::ATOMIC_LOAD_ADD_I16:
3862 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3863 case ARM::ATOMIC_LOAD_ADD_I32:
3864 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3866 case ARM::ATOMIC_LOAD_AND_I8:
3867 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3868 case ARM::ATOMIC_LOAD_AND_I16:
3869 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3870 case ARM::ATOMIC_LOAD_AND_I32:
3871 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3873 case ARM::ATOMIC_LOAD_OR_I8:
3874 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3875 case ARM::ATOMIC_LOAD_OR_I16:
3876 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3877 case ARM::ATOMIC_LOAD_OR_I32:
3878 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3880 case ARM::ATOMIC_LOAD_XOR_I8:
3881 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3882 case ARM::ATOMIC_LOAD_XOR_I16:
3883 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3884 case ARM::ATOMIC_LOAD_XOR_I32:
3885 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3887 case ARM::ATOMIC_LOAD_NAND_I8:
3888 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3889 case ARM::ATOMIC_LOAD_NAND_I16:
3890 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3891 case ARM::ATOMIC_LOAD_NAND_I32:
3892 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3894 case ARM::ATOMIC_LOAD_SUB_I8:
3895 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3896 case ARM::ATOMIC_LOAD_SUB_I16:
3897 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3898 case ARM::ATOMIC_LOAD_SUB_I32:
3899 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3901 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3902 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3903 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3905 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3906 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3907 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3909 case ARM::tMOVCCr_pseudo: {
3910 // To "insert" a SELECT_CC instruction, we actually have to insert the
3911 // diamond control-flow pattern. The incoming instruction knows the
3912 // destination vreg to set, the condition code register to branch on, the
3913 // true/false values to select between, and a branch opcode to use.
3914 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3915 MachineFunction::iterator It = BB;
3921 // cmpTY ccX, r1, r2
3923 // fallthrough --> copy0MBB
3924 MachineBasicBlock *thisMBB = BB;
3925 MachineFunction *F = BB->getParent();
3926 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3927 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3928 F->insert(It, copy0MBB);
3929 F->insert(It, sinkMBB);
3931 // Transfer the remainder of BB and its successor edges to sinkMBB.
3932 sinkMBB->splice(sinkMBB->begin(), BB,
3933 llvm::next(MachineBasicBlock::iterator(MI)),
3935 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
3937 BB->addSuccessor(copy0MBB);
3938 BB->addSuccessor(sinkMBB);
3940 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3941 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3944 // %FalseValue = ...
3945 // # fallthrough to sinkMBB
3948 // Update machine-CFG edges
3949 BB->addSuccessor(sinkMBB);
3952 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3955 BuildMI(*BB, BB->begin(), dl,
3956 TII->get(ARM::PHI), MI->getOperand(0).getReg())
3957 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3958 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3960 MI->eraseFromParent(); // The pseudo instruction is gone now.
3965 case ARM::BCCZi64: {
3966 // Compare both parts that make up the double comparison separately for
3968 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
3970 unsigned LHS1 = MI->getOperand(1).getReg();
3971 unsigned LHS2 = MI->getOperand(2).getReg();
3973 AddDefaultPred(BuildMI(BB, dl,
3974 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3975 .addReg(LHS1).addImm(0));
3976 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3977 .addReg(LHS2).addImm(0)
3978 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
3980 unsigned RHS1 = MI->getOperand(3).getReg();
3981 unsigned RHS2 = MI->getOperand(4).getReg();
3982 AddDefaultPred(BuildMI(BB, dl,
3983 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3984 .addReg(LHS1).addReg(RHS1));
3985 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3986 .addReg(LHS2).addReg(RHS2)
3987 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
3990 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
3991 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
3992 if (MI->getOperand(0).getImm() == ARMCC::NE)
3993 std::swap(destMBB, exitMBB);
3995 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3996 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
3997 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
4000 MI->eraseFromParent(); // The pseudo instruction is gone now.
4007 case ARM::t2SUBrSPi_:
4008 case ARM::t2SUBrSPi12_:
4009 case ARM::t2SUBrSPs_: {
4010 MachineFunction *MF = BB->getParent();
4011 unsigned DstReg = MI->getOperand(0).getReg();
4012 unsigned SrcReg = MI->getOperand(1).getReg();
4013 bool DstIsDead = MI->getOperand(0).isDead();
4014 bool SrcIsKill = MI->getOperand(1).isKill();
4016 if (SrcReg != ARM::SP) {
4017 // Copy the source to SP from virtual register.
4018 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
4019 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4020 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
4021 BuildMI(*BB, MI, dl, TII->get(CopyOpc), ARM::SP)
4022 .addReg(SrcReg, getKillRegState(SrcIsKill));
4026 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
4027 switch (MI->getOpcode()) {
4029 llvm_unreachable("Unexpected pseudo instruction!");
4035 OpOpc = ARM::tADDspr;
4038 OpOpc = ARM::tSUBspi;
4040 case ARM::t2SUBrSPi_:
4041 OpOpc = ARM::t2SUBrSPi;
4042 NeedPred = true; NeedCC = true;
4044 case ARM::t2SUBrSPi12_:
4045 OpOpc = ARM::t2SUBrSPi12;
4048 case ARM::t2SUBrSPs_:
4049 OpOpc = ARM::t2SUBrSPs;
4050 NeedPred = true; NeedCC = true; NeedOp3 = true;
4053 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(OpOpc), ARM::SP);
4054 if (OpOpc == ARM::tAND)
4055 AddDefaultT1CC(MIB);
4056 MIB.addReg(ARM::SP);
4057 MIB.addOperand(MI->getOperand(2));
4059 MIB.addOperand(MI->getOperand(3));
4061 AddDefaultPred(MIB);
4065 // Copy the result from SP to virtual register.
4066 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
4067 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
4068 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
4069 BuildMI(*BB, MI, dl, TII->get(CopyOpc))
4070 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
4072 MI->eraseFromParent(); // The pseudo instruction is gone now.
4078 //===----------------------------------------------------------------------===//
4079 // ARM Optimization Hooks
4080 //===----------------------------------------------------------------------===//
4083 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
4084 TargetLowering::DAGCombinerInfo &DCI) {
4085 SelectionDAG &DAG = DCI.DAG;
4086 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4087 EVT VT = N->getValueType(0);
4088 unsigned Opc = N->getOpcode();
4089 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
4090 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
4091 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
4092 ISD::CondCode CC = ISD::SETCC_INVALID;
4095 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
4097 SDValue CCOp = Slct.getOperand(0);
4098 if (CCOp.getOpcode() == ISD::SETCC)
4099 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
4102 bool DoXform = false;
4104 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
4107 if (LHS.getOpcode() == ISD::Constant &&
4108 cast<ConstantSDNode>(LHS)->isNullValue()) {
4110 } else if (CC != ISD::SETCC_INVALID &&
4111 RHS.getOpcode() == ISD::Constant &&
4112 cast<ConstantSDNode>(RHS)->isNullValue()) {
4113 std::swap(LHS, RHS);
4114 SDValue Op0 = Slct.getOperand(0);
4115 EVT OpVT = isSlctCC ? Op0.getValueType() :
4116 Op0.getOperand(0).getValueType();
4117 bool isInt = OpVT.isInteger();
4118 CC = ISD::getSetCCInverse(CC, isInt);
4120 if (!TLI.isCondCodeLegal(CC, OpVT))
4121 return SDValue(); // Inverse operator isn't legal.
4128 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
4130 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
4131 Slct.getOperand(0), Slct.getOperand(1), CC);
4132 SDValue CCOp = Slct.getOperand(0);
4134 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
4135 CCOp.getOperand(0), CCOp.getOperand(1), CC);
4136 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4137 CCOp, OtherOp, Result);
4142 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
4143 static SDValue PerformADDCombine(SDNode *N,
4144 TargetLowering::DAGCombinerInfo &DCI) {
4145 // added by evan in r37685 with no testcase.
4146 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4148 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4149 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4150 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4151 if (Result.getNode()) return Result;
4153 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4154 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4155 if (Result.getNode()) return Result;
4161 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4162 static SDValue PerformSUBCombine(SDNode *N,
4163 TargetLowering::DAGCombinerInfo &DCI) {
4164 // added by evan in r37685 with no testcase.
4165 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4167 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4168 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4169 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4170 if (Result.getNode()) return Result;
4176 static SDValue PerformMULCombine(SDNode *N,
4177 TargetLowering::DAGCombinerInfo &DCI,
4178 const ARMSubtarget *Subtarget) {
4179 SelectionDAG &DAG = DCI.DAG;
4181 if (Subtarget->isThumb1Only())
4184 if (DAG.getMachineFunction().
4185 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4188 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4191 EVT VT = N->getValueType(0);
4195 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4199 uint64_t MulAmt = C->getZExtValue();
4200 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4201 ShiftAmt = ShiftAmt & (32 - 1);
4202 SDValue V = N->getOperand(0);
4203 DebugLoc DL = N->getDebugLoc();
4206 MulAmt >>= ShiftAmt;
4207 if (isPowerOf2_32(MulAmt - 1)) {
4208 // (mul x, 2^N + 1) => (add (shl x, N), x)
4209 Res = DAG.getNode(ISD::ADD, DL, VT,
4210 V, DAG.getNode(ISD::SHL, DL, VT,
4211 V, DAG.getConstant(Log2_32(MulAmt-1),
4213 } else if (isPowerOf2_32(MulAmt + 1)) {
4214 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4215 Res = DAG.getNode(ISD::SUB, DL, VT,
4216 DAG.getNode(ISD::SHL, DL, VT,
4217 V, DAG.getConstant(Log2_32(MulAmt+1),
4224 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4225 DAG.getConstant(ShiftAmt, MVT::i32));
4227 // Do not add new nodes to DAG combiner worklist.
4228 DCI.CombineTo(N, Res, false);
4232 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
4233 static SDValue PerformORCombine(SDNode *N,
4234 TargetLowering::DAGCombinerInfo &DCI,
4235 const ARMSubtarget *Subtarget) {
4236 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
4239 // BFI is only available on V6T2+
4240 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
4243 SelectionDAG &DAG = DCI.DAG;
4244 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4245 DebugLoc DL = N->getDebugLoc();
4246 // 1) or (and A, mask), val => ARMbfi A, val, mask
4247 // iff (val & mask) == val
4249 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4250 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
4251 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4252 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
4253 // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
4254 // (i.e., copy a bitfield value into another bitfield of the same width)
4255 if (N0.getOpcode() != ISD::AND)
4258 EVT VT = N->getValueType(0);
4263 // The value and the mask need to be constants so we can verify this is
4264 // actually a bitfield set. If the mask is 0xffff, we can do better
4265 // via a movt instruction, so don't use BFI in that case.
4266 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4269 unsigned Mask = C->getZExtValue();
4273 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
4274 if ((C = dyn_cast<ConstantSDNode>(N1))) {
4275 unsigned Val = C->getZExtValue();
4276 if (!ARM::isBitFieldInvertedMask(Mask) || (Val & ~Mask) != Val)
4278 Val >>= CountTrailingZeros_32(~Mask);
4280 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
4281 DAG.getConstant(Val, MVT::i32),
4282 DAG.getConstant(Mask, MVT::i32));
4284 // Do not add new nodes to DAG combiner worklist.
4285 DCI.CombineTo(N, Res, false);
4286 } else if (N1.getOpcode() == ISD::AND) {
4287 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
4288 C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
4291 unsigned Mask2 = C->getZExtValue();
4293 if (ARM::isBitFieldInvertedMask(Mask) &&
4294 ARM::isBitFieldInvertedMask(~Mask2) &&
4295 (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
4296 // The pack halfword instruction works better for masks that fit it,
4297 // so use that when it's available.
4298 if (Subtarget->hasT2ExtractPack() &&
4299 (Mask == 0xffff || Mask == 0xffff0000))
4302 unsigned lsb = CountTrailingZeros_32(Mask2);
4303 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
4304 DAG.getConstant(lsb, MVT::i32));
4305 Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0), Res,
4306 DAG.getConstant(Mask, MVT::i32));
4307 // Do not add new nodes to DAG combiner worklist.
4308 DCI.CombineTo(N, Res, false);
4309 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
4310 ARM::isBitFieldInvertedMask(Mask2) &&
4311 (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
4312 // The pack halfword instruction works better for masks that fit it,
4313 // so use that when it's available.
4314 if (Subtarget->hasT2ExtractPack() &&
4315 (Mask2 == 0xffff || Mask2 == 0xffff0000))
4318 unsigned lsb = CountTrailingZeros_32(Mask);
4319 Res = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4320 DAG.getConstant(lsb, MVT::i32));
4321 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
4322 DAG.getConstant(Mask2, MVT::i32));
4323 // Do not add new nodes to DAG combiner worklist.
4324 DCI.CombineTo(N, Res, false);
4331 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4332 /// ARMISD::VMOVRRD.
4333 static SDValue PerformVMOVRRDCombine(SDNode *N,
4334 TargetLowering::DAGCombinerInfo &DCI) {
4335 // fmrrd(fmdrr x, y) -> x,y
4336 SDValue InDouble = N->getOperand(0);
4337 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4338 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4342 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
4343 /// ARMISD::VDUPLANE.
4344 static SDValue PerformVDUPLANECombine(SDNode *N,
4345 TargetLowering::DAGCombinerInfo &DCI) {
4346 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
4348 SDValue Op = N->getOperand(0);
4349 EVT VT = N->getValueType(0);
4351 // Ignore bit_converts.
4352 while (Op.getOpcode() == ISD::BIT_CONVERT)
4353 Op = Op.getOperand(0);
4354 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
4357 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
4358 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
4359 // The canonical VMOV for a zero vector uses a 32-bit element size.
4360 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4362 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
4364 if (EltSize > VT.getVectorElementType().getSizeInBits())
4367 SDValue Res = DCI.DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
4368 return DCI.CombineTo(N, Res, false);
4371 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4372 /// operand of a vector shift operation, where all the elements of the
4373 /// build_vector must have the same constant integer value.
4374 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4375 // Ignore bit_converts.
4376 while (Op.getOpcode() == ISD::BIT_CONVERT)
4377 Op = Op.getOperand(0);
4378 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4379 APInt SplatBits, SplatUndef;
4380 unsigned SplatBitSize;
4382 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4383 HasAnyUndefs, ElementBits) ||
4384 SplatBitSize > ElementBits)
4386 Cnt = SplatBits.getSExtValue();
4390 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4391 /// operand of a vector shift left operation. That value must be in the range:
4392 /// 0 <= Value < ElementBits for a left shift; or
4393 /// 0 <= Value <= ElementBits for a long left shift.
4394 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4395 assert(VT.isVector() && "vector shift count is not a vector type");
4396 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4397 if (! getVShiftImm(Op, ElementBits, Cnt))
4399 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4402 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4403 /// operand of a vector shift right operation. For a shift opcode, the value
4404 /// is positive, but for an intrinsic the value count must be negative. The
4405 /// absolute value must be in the range:
4406 /// 1 <= |Value| <= ElementBits for a right shift; or
4407 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4408 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4410 assert(VT.isVector() && "vector shift count is not a vector type");
4411 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4412 if (! getVShiftImm(Op, ElementBits, Cnt))
4416 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4419 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4420 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4421 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4424 // Don't do anything for most intrinsics.
4427 // Vector shifts: check for immediate versions and lower them.
4428 // Note: This is done during DAG combining instead of DAG legalizing because
4429 // the build_vectors for 64-bit vector element shift counts are generally
4430 // not legal, and it is hard to see their values after they get legalized to
4431 // loads from a constant pool.
4432 case Intrinsic::arm_neon_vshifts:
4433 case Intrinsic::arm_neon_vshiftu:
4434 case Intrinsic::arm_neon_vshiftls:
4435 case Intrinsic::arm_neon_vshiftlu:
4436 case Intrinsic::arm_neon_vshiftn:
4437 case Intrinsic::arm_neon_vrshifts:
4438 case Intrinsic::arm_neon_vrshiftu:
4439 case Intrinsic::arm_neon_vrshiftn:
4440 case Intrinsic::arm_neon_vqshifts:
4441 case Intrinsic::arm_neon_vqshiftu:
4442 case Intrinsic::arm_neon_vqshiftsu:
4443 case Intrinsic::arm_neon_vqshiftns:
4444 case Intrinsic::arm_neon_vqshiftnu:
4445 case Intrinsic::arm_neon_vqshiftnsu:
4446 case Intrinsic::arm_neon_vqrshiftns:
4447 case Intrinsic::arm_neon_vqrshiftnu:
4448 case Intrinsic::arm_neon_vqrshiftnsu: {
4449 EVT VT = N->getOperand(1).getValueType();
4451 unsigned VShiftOpc = 0;
4454 case Intrinsic::arm_neon_vshifts:
4455 case Intrinsic::arm_neon_vshiftu:
4456 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4457 VShiftOpc = ARMISD::VSHL;
4460 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4461 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4462 ARMISD::VSHRs : ARMISD::VSHRu);
4467 case Intrinsic::arm_neon_vshiftls:
4468 case Intrinsic::arm_neon_vshiftlu:
4469 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4471 llvm_unreachable("invalid shift count for vshll intrinsic");
4473 case Intrinsic::arm_neon_vrshifts:
4474 case Intrinsic::arm_neon_vrshiftu:
4475 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4479 case Intrinsic::arm_neon_vqshifts:
4480 case Intrinsic::arm_neon_vqshiftu:
4481 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4485 case Intrinsic::arm_neon_vqshiftsu:
4486 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4488 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4490 case Intrinsic::arm_neon_vshiftn:
4491 case Intrinsic::arm_neon_vrshiftn:
4492 case Intrinsic::arm_neon_vqshiftns:
4493 case Intrinsic::arm_neon_vqshiftnu:
4494 case Intrinsic::arm_neon_vqshiftnsu:
4495 case Intrinsic::arm_neon_vqrshiftns:
4496 case Intrinsic::arm_neon_vqrshiftnu:
4497 case Intrinsic::arm_neon_vqrshiftnsu:
4498 // Narrowing shifts require an immediate right shift.
4499 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4501 llvm_unreachable("invalid shift count for narrowing vector shift "
4505 llvm_unreachable("unhandled vector shift");
4509 case Intrinsic::arm_neon_vshifts:
4510 case Intrinsic::arm_neon_vshiftu:
4511 // Opcode already set above.
4513 case Intrinsic::arm_neon_vshiftls:
4514 case Intrinsic::arm_neon_vshiftlu:
4515 if (Cnt == VT.getVectorElementType().getSizeInBits())
4516 VShiftOpc = ARMISD::VSHLLi;
4518 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4519 ARMISD::VSHLLs : ARMISD::VSHLLu);
4521 case Intrinsic::arm_neon_vshiftn:
4522 VShiftOpc = ARMISD::VSHRN; break;
4523 case Intrinsic::arm_neon_vrshifts:
4524 VShiftOpc = ARMISD::VRSHRs; break;
4525 case Intrinsic::arm_neon_vrshiftu:
4526 VShiftOpc = ARMISD::VRSHRu; break;
4527 case Intrinsic::arm_neon_vrshiftn:
4528 VShiftOpc = ARMISD::VRSHRN; break;
4529 case Intrinsic::arm_neon_vqshifts:
4530 VShiftOpc = ARMISD::VQSHLs; break;
4531 case Intrinsic::arm_neon_vqshiftu:
4532 VShiftOpc = ARMISD::VQSHLu; break;
4533 case Intrinsic::arm_neon_vqshiftsu:
4534 VShiftOpc = ARMISD::VQSHLsu; break;
4535 case Intrinsic::arm_neon_vqshiftns:
4536 VShiftOpc = ARMISD::VQSHRNs; break;
4537 case Intrinsic::arm_neon_vqshiftnu:
4538 VShiftOpc = ARMISD::VQSHRNu; break;
4539 case Intrinsic::arm_neon_vqshiftnsu:
4540 VShiftOpc = ARMISD::VQSHRNsu; break;
4541 case Intrinsic::arm_neon_vqrshiftns:
4542 VShiftOpc = ARMISD::VQRSHRNs; break;
4543 case Intrinsic::arm_neon_vqrshiftnu:
4544 VShiftOpc = ARMISD::VQRSHRNu; break;
4545 case Intrinsic::arm_neon_vqrshiftnsu:
4546 VShiftOpc = ARMISD::VQRSHRNsu; break;
4549 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4550 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4553 case Intrinsic::arm_neon_vshiftins: {
4554 EVT VT = N->getOperand(1).getValueType();
4556 unsigned VShiftOpc = 0;
4558 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4559 VShiftOpc = ARMISD::VSLI;
4560 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4561 VShiftOpc = ARMISD::VSRI;
4563 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4566 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4567 N->getOperand(1), N->getOperand(2),
4568 DAG.getConstant(Cnt, MVT::i32));
4571 case Intrinsic::arm_neon_vqrshifts:
4572 case Intrinsic::arm_neon_vqrshiftu:
4573 // No immediate versions of these to check for.
4580 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4581 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4582 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4583 /// vector element shift counts are generally not legal, and it is hard to see
4584 /// their values after they get legalized to loads from a constant pool.
4585 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4586 const ARMSubtarget *ST) {
4587 EVT VT = N->getValueType(0);
4589 // Nothing to be done for scalar shifts.
4590 if (! VT.isVector())
4593 assert(ST->hasNEON() && "unexpected vector shift");
4596 switch (N->getOpcode()) {
4597 default: llvm_unreachable("unexpected shift opcode");
4600 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4601 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4602 DAG.getConstant(Cnt, MVT::i32));
4607 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4608 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4609 ARMISD::VSHRs : ARMISD::VSHRu);
4610 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4611 DAG.getConstant(Cnt, MVT::i32));
4617 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4618 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4619 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4620 const ARMSubtarget *ST) {
4621 SDValue N0 = N->getOperand(0);
4623 // Check for sign- and zero-extensions of vector extract operations of 8-
4624 // and 16-bit vector elements. NEON supports these directly. They are
4625 // handled during DAG combining because type legalization will promote them
4626 // to 32-bit types and it is messy to recognize the operations after that.
4627 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4628 SDValue Vec = N0.getOperand(0);
4629 SDValue Lane = N0.getOperand(1);
4630 EVT VT = N->getValueType(0);
4631 EVT EltVT = N0.getValueType();
4632 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4634 if (VT == MVT::i32 &&
4635 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4636 TLI.isTypeLegal(Vec.getValueType())) {
4639 switch (N->getOpcode()) {
4640 default: llvm_unreachable("unexpected opcode");
4641 case ISD::SIGN_EXTEND:
4642 Opc = ARMISD::VGETLANEs;
4644 case ISD::ZERO_EXTEND:
4645 case ISD::ANY_EXTEND:
4646 Opc = ARMISD::VGETLANEu;
4649 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4656 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4657 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4658 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4659 const ARMSubtarget *ST) {
4660 // If the target supports NEON, try to use vmax/vmin instructions for f32
4661 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
4662 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4663 // a NaN; only do the transformation when it matches that behavior.
4665 // For now only do this when using NEON for FP operations; if using VFP, it
4666 // is not obvious that the benefit outweighs the cost of switching to the
4668 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4669 N->getValueType(0) != MVT::f32)
4672 SDValue CondLHS = N->getOperand(0);
4673 SDValue CondRHS = N->getOperand(1);
4674 SDValue LHS = N->getOperand(2);
4675 SDValue RHS = N->getOperand(3);
4676 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4678 unsigned Opcode = 0;
4680 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4681 IsReversed = false; // x CC y ? x : y
4682 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4683 IsReversed = true ; // x CC y ? y : x
4697 // If LHS is NaN, an ordered comparison will be false and the result will
4698 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4699 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4700 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4701 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4703 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4704 // will return -0, so vmin can only be used for unsafe math or if one of
4705 // the operands is known to be nonzero.
4706 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4708 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4710 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4719 // If LHS is NaN, an ordered comparison will be false and the result will
4720 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4721 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4722 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4723 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4725 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4726 // will return +0, so vmax can only be used for unsafe math or if one of
4727 // the operands is known to be nonzero.
4728 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4730 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4732 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4738 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4741 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4742 DAGCombinerInfo &DCI) const {
4743 switch (N->getOpcode()) {
4745 case ISD::ADD: return PerformADDCombine(N, DCI);
4746 case ISD::SUB: return PerformSUBCombine(N, DCI);
4747 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4748 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
4749 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4750 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
4751 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4754 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4755 case ISD::SIGN_EXTEND:
4756 case ISD::ZERO_EXTEND:
4757 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4758 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4763 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4764 if (!Subtarget->hasV6Ops())
4765 // Pre-v6 does not support unaligned mem access.
4768 // v6+ may or may not support unaligned mem access depending on the system
4770 // FIXME: This is pretty conservative. Should we provide cmdline option to
4771 // control the behaviour?
4772 if (!Subtarget->isTargetDarwin())
4775 switch (VT.getSimpleVT().SimpleTy) {
4782 // FIXME: VLD1 etc with standard alignment is legal.
4786 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4791 switch (VT.getSimpleVT().SimpleTy) {
4792 default: return false;
4807 if ((V & (Scale - 1)) != 0)
4810 return V == (V & ((1LL << 5) - 1));
4813 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4814 const ARMSubtarget *Subtarget) {
4821 switch (VT.getSimpleVT().SimpleTy) {
4822 default: return false;
4827 // + imm12 or - imm8
4829 return V == (V & ((1LL << 8) - 1));
4830 return V == (V & ((1LL << 12) - 1));
4833 // Same as ARM mode. FIXME: NEON?
4834 if (!Subtarget->hasVFP2())
4839 return V == (V & ((1LL << 8) - 1));
4843 /// isLegalAddressImmediate - Return true if the integer value can be used
4844 /// as the offset of the target addressing mode for load / store of the
4846 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4847 const ARMSubtarget *Subtarget) {
4854 if (Subtarget->isThumb1Only())
4855 return isLegalT1AddressImmediate(V, VT);
4856 else if (Subtarget->isThumb2())
4857 return isLegalT2AddressImmediate(V, VT, Subtarget);
4862 switch (VT.getSimpleVT().SimpleTy) {
4863 default: return false;
4868 return V == (V & ((1LL << 12) - 1));
4871 return V == (V & ((1LL << 8) - 1));
4874 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4879 return V == (V & ((1LL << 8) - 1));
4883 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4885 int Scale = AM.Scale;
4889 switch (VT.getSimpleVT().SimpleTy) {
4890 default: return false;
4899 return Scale == 2 || Scale == 4 || Scale == 8;
4902 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4906 // Note, we allow "void" uses (basically, uses that aren't loads or
4907 // stores), because arm allows folding a scale into many arithmetic
4908 // operations. This should be made more precise and revisited later.
4910 // Allow r << imm, but the imm has to be a multiple of two.
4911 if (Scale & 1) return false;
4912 return isPowerOf2_32(Scale);
4916 /// isLegalAddressingMode - Return true if the addressing mode represented
4917 /// by AM is legal for this target, for a load/store of the specified type.
4918 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4919 const Type *Ty) const {
4920 EVT VT = getValueType(Ty, true);
4921 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4924 // Can never fold addr of global into load/store.
4929 case 0: // no scale reg, must be "r+i" or "r", or "i".
4932 if (Subtarget->isThumb1Only())
4936 // ARM doesn't support any R+R*scale+imm addr modes.
4943 if (Subtarget->isThumb2())
4944 return isLegalT2ScaledAddressingMode(AM, VT);
4946 int Scale = AM.Scale;
4947 switch (VT.getSimpleVT().SimpleTy) {
4948 default: return false;
4952 if (Scale < 0) Scale = -Scale;
4956 return isPowerOf2_32(Scale & ~1);
4960 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4965 // Note, we allow "void" uses (basically, uses that aren't loads or
4966 // stores), because arm allows folding a scale into many arithmetic
4967 // operations. This should be made more precise and revisited later.
4969 // Allow r << imm, but the imm has to be a multiple of two.
4970 if (Scale & 1) return false;
4971 return isPowerOf2_32(Scale);
4978 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4979 /// icmp immediate, that is the target has icmp instructions which can compare
4980 /// a register against the immediate without having to materialize the
4981 /// immediate into a register.
4982 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4983 if (!Subtarget->isThumb())
4984 return ARM_AM::getSOImmVal(Imm) != -1;
4985 if (Subtarget->isThumb2())
4986 return ARM_AM::getT2SOImmVal(Imm) != -1;
4987 return Imm >= 0 && Imm <= 255;
4990 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4991 bool isSEXTLoad, SDValue &Base,
4992 SDValue &Offset, bool &isInc,
4993 SelectionDAG &DAG) {
4994 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4997 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4999 Base = Ptr->getOperand(0);
5000 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5001 int RHSC = (int)RHS->getZExtValue();
5002 if (RHSC < 0 && RHSC > -256) {
5003 assert(Ptr->getOpcode() == ISD::ADD);
5005 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5009 isInc = (Ptr->getOpcode() == ISD::ADD);
5010 Offset = Ptr->getOperand(1);
5012 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
5014 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5015 int RHSC = (int)RHS->getZExtValue();
5016 if (RHSC < 0 && RHSC > -0x1000) {
5017 assert(Ptr->getOpcode() == ISD::ADD);
5019 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5020 Base = Ptr->getOperand(0);
5025 if (Ptr->getOpcode() == ISD::ADD) {
5027 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
5028 if (ShOpcVal != ARM_AM::no_shift) {
5029 Base = Ptr->getOperand(1);
5030 Offset = Ptr->getOperand(0);
5032 Base = Ptr->getOperand(0);
5033 Offset = Ptr->getOperand(1);
5038 isInc = (Ptr->getOpcode() == ISD::ADD);
5039 Base = Ptr->getOperand(0);
5040 Offset = Ptr->getOperand(1);
5044 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
5048 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
5049 bool isSEXTLoad, SDValue &Base,
5050 SDValue &Offset, bool &isInc,
5051 SelectionDAG &DAG) {
5052 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
5055 Base = Ptr->getOperand(0);
5056 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
5057 int RHSC = (int)RHS->getZExtValue();
5058 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
5059 assert(Ptr->getOpcode() == ISD::ADD);
5061 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
5063 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
5064 isInc = Ptr->getOpcode() == ISD::ADD;
5065 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
5073 /// getPreIndexedAddressParts - returns true by value, base pointer and
5074 /// offset pointer and addressing mode by reference if the node's address
5075 /// can be legally represented as pre-indexed load / store address.
5077 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
5079 ISD::MemIndexedMode &AM,
5080 SelectionDAG &DAG) const {
5081 if (Subtarget->isThumb1Only())
5086 bool isSEXTLoad = false;
5087 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5088 Ptr = LD->getBasePtr();
5089 VT = LD->getMemoryVT();
5090 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5091 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5092 Ptr = ST->getBasePtr();
5093 VT = ST->getMemoryVT();
5098 bool isLegal = false;
5099 if (Subtarget->isThumb2())
5100 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5101 Offset, isInc, DAG);
5103 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
5104 Offset, isInc, DAG);
5108 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
5112 /// getPostIndexedAddressParts - returns true by value, base pointer and
5113 /// offset pointer and addressing mode by reference if this node can be
5114 /// combined with a load / store to form a post-indexed load / store.
5115 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
5118 ISD::MemIndexedMode &AM,
5119 SelectionDAG &DAG) const {
5120 if (Subtarget->isThumb1Only())
5125 bool isSEXTLoad = false;
5126 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5127 VT = LD->getMemoryVT();
5128 Ptr = LD->getBasePtr();
5129 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
5130 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5131 VT = ST->getMemoryVT();
5132 Ptr = ST->getBasePtr();
5137 bool isLegal = false;
5138 if (Subtarget->isThumb2())
5139 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5142 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
5148 // Swap base ptr and offset to catch more post-index load / store when
5149 // it's legal. In Thumb2 mode, offset must be an immediate.
5150 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
5151 !Subtarget->isThumb2())
5152 std::swap(Base, Offset);
5154 // Post-indexed load / store update the base pointer.
5159 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
5163 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5167 const SelectionDAG &DAG,
5168 unsigned Depth) const {
5169 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5170 switch (Op.getOpcode()) {
5172 case ARMISD::CMOV: {
5173 // Bits are known zero/one if known on the LHS and RHS.
5174 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
5175 if (KnownZero == 0 && KnownOne == 0) return;
5177 APInt KnownZeroRHS, KnownOneRHS;
5178 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
5179 KnownZeroRHS, KnownOneRHS, Depth+1);
5180 KnownZero &= KnownZeroRHS;
5181 KnownOne &= KnownOneRHS;
5187 //===----------------------------------------------------------------------===//
5188 // ARM Inline Assembly Support
5189 //===----------------------------------------------------------------------===//
5191 /// getConstraintType - Given a constraint letter, return the type of
5192 /// constraint it is for this target.
5193 ARMTargetLowering::ConstraintType
5194 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
5195 if (Constraint.size() == 1) {
5196 switch (Constraint[0]) {
5198 case 'l': return C_RegisterClass;
5199 case 'w': return C_RegisterClass;
5202 return TargetLowering::getConstraintType(Constraint);
5205 std::pair<unsigned, const TargetRegisterClass*>
5206 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5208 if (Constraint.size() == 1) {
5209 // GCC ARM Constraint Letters
5210 switch (Constraint[0]) {
5212 if (Subtarget->isThumb())
5213 return std::make_pair(0U, ARM::tGPRRegisterClass);
5215 return std::make_pair(0U, ARM::GPRRegisterClass);
5217 return std::make_pair(0U, ARM::GPRRegisterClass);
5220 return std::make_pair(0U, ARM::SPRRegisterClass);
5221 if (VT.getSizeInBits() == 64)
5222 return std::make_pair(0U, ARM::DPRRegisterClass);
5223 if (VT.getSizeInBits() == 128)
5224 return std::make_pair(0U, ARM::QPRRegisterClass);
5228 if (StringRef("{cc}").equals_lower(Constraint))
5229 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
5231 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5234 std::vector<unsigned> ARMTargetLowering::
5235 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5237 if (Constraint.size() != 1)
5238 return std::vector<unsigned>();
5240 switch (Constraint[0]) { // GCC ARM Constraint Letters
5243 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5244 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5247 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
5248 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
5249 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
5250 ARM::R12, ARM::LR, 0);
5253 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
5254 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
5255 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
5256 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
5257 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
5258 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
5259 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
5260 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
5261 if (VT.getSizeInBits() == 64)
5262 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
5263 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
5264 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
5265 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
5266 if (VT.getSizeInBits() == 128)
5267 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
5268 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
5272 return std::vector<unsigned>();
5275 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5276 /// vector. If it is invalid, don't add anything to Ops.
5277 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5279 std::vector<SDValue>&Ops,
5280 SelectionDAG &DAG) const {
5281 SDValue Result(0, 0);
5283 switch (Constraint) {
5285 case 'I': case 'J': case 'K': case 'L':
5286 case 'M': case 'N': case 'O':
5287 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5291 int64_t CVal64 = C->getSExtValue();
5292 int CVal = (int) CVal64;
5293 // None of these constraints allow values larger than 32 bits. Check
5294 // that the value fits in an int.
5298 switch (Constraint) {
5300 if (Subtarget->isThumb1Only()) {
5301 // This must be a constant between 0 and 255, for ADD
5303 if (CVal >= 0 && CVal <= 255)
5305 } else if (Subtarget->isThumb2()) {
5306 // A constant that can be used as an immediate value in a
5307 // data-processing instruction.
5308 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5311 // A constant that can be used as an immediate value in a
5312 // data-processing instruction.
5313 if (ARM_AM::getSOImmVal(CVal) != -1)
5319 if (Subtarget->isThumb()) { // FIXME thumb2
5320 // This must be a constant between -255 and -1, for negated ADD
5321 // immediates. This can be used in GCC with an "n" modifier that
5322 // prints the negated value, for use with SUB instructions. It is
5323 // not useful otherwise but is implemented for compatibility.
5324 if (CVal >= -255 && CVal <= -1)
5327 // This must be a constant between -4095 and 4095. It is not clear
5328 // what this constraint is intended for. Implemented for
5329 // compatibility with GCC.
5330 if (CVal >= -4095 && CVal <= 4095)
5336 if (Subtarget->isThumb1Only()) {
5337 // A 32-bit value where only one byte has a nonzero value. Exclude
5338 // zero to match GCC. This constraint is used by GCC internally for
5339 // constants that can be loaded with a move/shift combination.
5340 // It is not useful otherwise but is implemented for compatibility.
5341 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5343 } else if (Subtarget->isThumb2()) {
5344 // A constant whose bitwise inverse can be used as an immediate
5345 // value in a data-processing instruction. This can be used in GCC
5346 // with a "B" modifier that prints the inverted value, for use with
5347 // BIC and MVN instructions. It is not useful otherwise but is
5348 // implemented for compatibility.
5349 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5352 // A constant whose bitwise inverse can be used as an immediate
5353 // value in a data-processing instruction. This can be used in GCC
5354 // with a "B" modifier that prints the inverted value, for use with
5355 // BIC and MVN instructions. It is not useful otherwise but is
5356 // implemented for compatibility.
5357 if (ARM_AM::getSOImmVal(~CVal) != -1)
5363 if (Subtarget->isThumb1Only()) {
5364 // This must be a constant between -7 and 7,
5365 // for 3-operand ADD/SUB immediate instructions.
5366 if (CVal >= -7 && CVal < 7)
5368 } else if (Subtarget->isThumb2()) {
5369 // A constant whose negation can be used as an immediate value in a
5370 // data-processing instruction. This can be used in GCC with an "n"
5371 // modifier that prints the negated value, for use with SUB
5372 // instructions. It is not useful otherwise but is implemented for
5374 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5377 // A constant whose negation can be used as an immediate value in a
5378 // data-processing instruction. This can be used in GCC with an "n"
5379 // modifier that prints the negated value, for use with SUB
5380 // instructions. It is not useful otherwise but is implemented for
5382 if (ARM_AM::getSOImmVal(-CVal) != -1)
5388 if (Subtarget->isThumb()) { // FIXME thumb2
5389 // This must be a multiple of 4 between 0 and 1020, for
5390 // ADD sp + immediate.
5391 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5394 // A power of two or a constant between 0 and 32. This is used in
5395 // GCC for the shift amount on shifted register operands, but it is
5396 // useful in general for any shift amounts.
5397 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5403 if (Subtarget->isThumb()) { // FIXME thumb2
5404 // This must be a constant between 0 and 31, for shift amounts.
5405 if (CVal >= 0 && CVal <= 31)
5411 if (Subtarget->isThumb()) { // FIXME thumb2
5412 // This must be a multiple of 4 between -508 and 508, for
5413 // ADD/SUB sp = sp + immediate.
5414 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5419 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5423 if (Result.getNode()) {
5424 Ops.push_back(Result);
5427 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5431 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5432 // The ARM target isn't yet aware of offsets.
5436 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5437 APInt Imm = FPImm.bitcastToAPInt();
5438 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5439 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5440 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5442 // We can handle 4 bits of mantissa.
5443 // mantissa = (16+UInt(e:f:g:h))/16.
5444 if (Mantissa & 0x7ffff)
5447 if ((Mantissa & 0xf) != Mantissa)
5450 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5451 if (Exp < -3 || Exp > 4)
5453 Exp = ((Exp+3) & 0x7) ^ 4;
5455 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5458 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5459 APInt Imm = FPImm.bitcastToAPInt();
5460 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5461 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5462 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5464 // We can handle 4 bits of mantissa.
5465 // mantissa = (16+UInt(e:f:g:h))/16.
5466 if (Mantissa & 0xffffffffffffLL)
5469 if ((Mantissa & 0xf) != Mantissa)
5472 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5473 if (Exp < -3 || Exp > 4)
5475 Exp = ((Exp+3) & 0x7) ^ 4;
5477 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5480 bool ARM::isBitFieldInvertedMask(unsigned v) {
5481 if (v == 0xffffffff)
5483 // there can be 1's on either or both "outsides", all the "inside"
5485 unsigned int lsb = 0, msb = 31;
5486 while (v & (1 << msb)) --msb;
5487 while (v & (1 << lsb)) ++lsb;
5488 for (unsigned int i = lsb; i <= msb; ++i) {
5495 /// isFPImmLegal - Returns true if the target can instruction select the
5496 /// specified FP immediate natively. If false, the legalizer will
5497 /// materialize the FP immediate as a load from a constant pool.
5498 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5499 if (!Subtarget->hasVFP3())
5502 return ARM::getVFPf32Imm(Imm) != -1;
5504 return ARM::getVFPf64Imm(Imm) != -1;