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 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/Instruction.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Type.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/MC/MCSectionMachO.h"
41 #include "llvm/Target/TargetOptions.h"
42 #include "llvm/ADT/VectorExtras.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/MathExtras.h"
46 #include "llvm/Support/raw_ostream.h"
51 EnableARMLongCalls("arm-long-calls", cl::Hidden,
52 cl::desc("Generate calls via indirect call instructions."),
55 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
56 CCValAssign::LocInfo &LocInfo,
57 ISD::ArgFlagsTy &ArgFlags,
59 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
60 CCValAssign::LocInfo &LocInfo,
61 ISD::ArgFlagsTy &ArgFlags,
63 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
64 CCValAssign::LocInfo &LocInfo,
65 ISD::ArgFlagsTy &ArgFlags,
67 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
68 CCValAssign::LocInfo &LocInfo,
69 ISD::ArgFlagsTy &ArgFlags,
72 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
73 EVT PromotedBitwiseVT) {
74 if (VT != PromotedLdStVT) {
75 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
76 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
77 PromotedLdStVT.getSimpleVT());
79 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
80 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
81 PromotedLdStVT.getSimpleVT());
84 EVT ElemTy = VT.getVectorElementType();
85 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
86 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
87 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
88 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
89 if (ElemTy != MVT::i32) {
90 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
91 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
95 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
97 if (llvm::ModelWithRegSequence())
98 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
100 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
101 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
102 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
103 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
104 if (VT.isInteger()) {
105 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
106 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
107 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
110 // Promote all bit-wise operations.
111 if (VT.isInteger() && VT != PromotedBitwiseVT) {
112 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
113 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
114 PromotedBitwiseVT.getSimpleVT());
115 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
116 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
117 PromotedBitwiseVT.getSimpleVT());
118 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
119 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
120 PromotedBitwiseVT.getSimpleVT());
123 // Neon does not support vector divide/remainder operations.
124 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
125 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
128 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
129 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
132 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
133 addRegisterClass(VT, ARM::DPRRegisterClass);
134 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
137 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
138 addRegisterClass(VT, ARM::QPRRegisterClass);
139 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
142 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
143 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
144 return new TargetLoweringObjectFileMachO();
146 return new ARMElfTargetObjectFile();
149 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
150 : TargetLowering(TM, createTLOF(TM)) {
151 Subtarget = &TM.getSubtarget<ARMSubtarget>();
153 if (Subtarget->isTargetDarwin()) {
154 // Uses VFP for Thumb libfuncs if available.
155 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
156 // Single-precision floating-point arithmetic.
157 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
158 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
159 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
160 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
162 // Double-precision floating-point arithmetic.
163 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
164 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
165 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
166 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
168 // Single-precision comparisons.
169 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
170 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
171 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
172 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
173 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
174 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
175 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
176 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
178 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
179 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
180 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
181 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
182 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
183 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
187 // Double-precision comparisons.
188 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
189 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
190 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
191 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
192 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
193 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
194 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
195 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
197 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
206 // Floating-point to integer conversions.
207 // i64 conversions are done via library routines even when generating VFP
208 // instructions, so use the same ones.
209 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
210 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
211 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
212 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
214 // Conversions between floating types.
215 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
216 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
218 // Integer to floating-point conversions.
219 // i64 conversions are done via library routines even when generating VFP
220 // instructions, so use the same ones.
221 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
222 // e.g., __floatunsidf vs. __floatunssidfvfp.
223 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
224 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
225 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
226 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
230 // These libcalls are not available in 32-bit.
231 setLibcallName(RTLIB::SHL_I128, 0);
232 setLibcallName(RTLIB::SRL_I128, 0);
233 setLibcallName(RTLIB::SRA_I128, 0);
235 // Libcalls should use the AAPCS base standard ABI, even if hard float
236 // is in effect, as per the ARM RTABI specification, section 4.1.2.
237 if (Subtarget->isAAPCS_ABI()) {
238 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
239 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
240 CallingConv::ARM_AAPCS);
244 if (Subtarget->isThumb1Only())
245 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
247 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
248 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
249 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
250 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
252 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
255 if (Subtarget->hasNEON()) {
256 addDRTypeForNEON(MVT::v2f32);
257 addDRTypeForNEON(MVT::v8i8);
258 addDRTypeForNEON(MVT::v4i16);
259 addDRTypeForNEON(MVT::v2i32);
260 addDRTypeForNEON(MVT::v1i64);
262 addQRTypeForNEON(MVT::v4f32);
263 addQRTypeForNEON(MVT::v2f64);
264 addQRTypeForNEON(MVT::v16i8);
265 addQRTypeForNEON(MVT::v8i16);
266 addQRTypeForNEON(MVT::v4i32);
267 addQRTypeForNEON(MVT::v2i64);
269 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
270 // neither Neon nor VFP support any arithmetic operations on it.
271 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
272 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
273 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
274 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
275 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
276 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
277 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
278 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
279 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
280 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
281 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
282 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
283 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
284 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
285 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
286 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
287 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
288 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
289 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
290 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
291 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
292 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
293 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
294 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
296 // Neon does not support some operations on v1i64 and v2i64 types.
297 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
298 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
299 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
300 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
302 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
303 setTargetDAGCombine(ISD::SHL);
304 setTargetDAGCombine(ISD::SRL);
305 setTargetDAGCombine(ISD::SRA);
306 setTargetDAGCombine(ISD::SIGN_EXTEND);
307 setTargetDAGCombine(ISD::ZERO_EXTEND);
308 setTargetDAGCombine(ISD::ANY_EXTEND);
309 setTargetDAGCombine(ISD::SELECT_CC);
312 computeRegisterProperties();
314 // ARM does not have f32 extending load.
315 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
317 // ARM does not have i1 sign extending load.
318 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
320 // ARM supports all 4 flavors of integer indexed load / store.
321 if (!Subtarget->isThumb1Only()) {
322 for (unsigned im = (unsigned)ISD::PRE_INC;
323 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
324 setIndexedLoadAction(im, MVT::i1, Legal);
325 setIndexedLoadAction(im, MVT::i8, Legal);
326 setIndexedLoadAction(im, MVT::i16, Legal);
327 setIndexedLoadAction(im, MVT::i32, Legal);
328 setIndexedStoreAction(im, MVT::i1, Legal);
329 setIndexedStoreAction(im, MVT::i8, Legal);
330 setIndexedStoreAction(im, MVT::i16, Legal);
331 setIndexedStoreAction(im, MVT::i32, Legal);
335 // i64 operation support.
336 if (Subtarget->isThumb1Only()) {
337 setOperationAction(ISD::MUL, MVT::i64, Expand);
338 setOperationAction(ISD::MULHU, MVT::i32, Expand);
339 setOperationAction(ISD::MULHS, MVT::i32, Expand);
340 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
341 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
343 setOperationAction(ISD::MUL, MVT::i64, Expand);
344 setOperationAction(ISD::MULHU, MVT::i32, Expand);
345 if (!Subtarget->hasV6Ops())
346 setOperationAction(ISD::MULHS, MVT::i32, Expand);
348 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
349 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
350 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
351 setOperationAction(ISD::SRL, MVT::i64, Custom);
352 setOperationAction(ISD::SRA, MVT::i64, Custom);
354 // ARM does not have ROTL.
355 setOperationAction(ISD::ROTL, MVT::i32, Expand);
356 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
357 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
358 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
359 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
361 // Only ARMv6 has BSWAP.
362 if (!Subtarget->hasV6Ops())
363 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
365 // These are expanded into libcalls.
366 if (!Subtarget->hasDivide()) {
367 // v7M has a hardware divider
368 setOperationAction(ISD::SDIV, MVT::i32, Expand);
369 setOperationAction(ISD::UDIV, MVT::i32, Expand);
371 setOperationAction(ISD::SREM, MVT::i32, Expand);
372 setOperationAction(ISD::UREM, MVT::i32, Expand);
373 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
374 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
376 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
377 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
378 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
379 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
380 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
382 setOperationAction(ISD::TRAP, MVT::Other, Legal);
384 // Use the default implementation.
385 setOperationAction(ISD::VASTART, MVT::Other, Custom);
386 setOperationAction(ISD::VAARG, MVT::Other, Expand);
387 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
388 setOperationAction(ISD::VAEND, MVT::Other, Expand);
389 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
390 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
391 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
392 // FIXME: Shouldn't need this, since no register is used, but the legalizer
393 // doesn't yet know how to not do that for SjLj.
394 setExceptionSelectorRegister(ARM::R0);
395 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
396 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
398 // If the subtarget does not have extract instructions, sign_extend_inreg
399 // needs to be expanded. Extract is available in ARM mode on v6 and up,
400 // and on most Thumb2 implementations.
401 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
402 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
403 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
404 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
406 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
408 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
409 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
410 // iff target supports vfp2.
411 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
413 // We want to custom lower some of our intrinsics.
414 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
415 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
417 setOperationAction(ISD::SETCC, MVT::i32, Expand);
418 setOperationAction(ISD::SETCC, MVT::f32, Expand);
419 setOperationAction(ISD::SETCC, MVT::f64, Expand);
420 setOperationAction(ISD::SELECT, MVT::i32, Expand);
421 setOperationAction(ISD::SELECT, MVT::f32, Expand);
422 setOperationAction(ISD::SELECT, MVT::f64, Expand);
423 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
424 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
425 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
427 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
428 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
429 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
430 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
431 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
433 // We don't support sin/cos/fmod/copysign/pow
434 setOperationAction(ISD::FSIN, MVT::f64, Expand);
435 setOperationAction(ISD::FSIN, MVT::f32, Expand);
436 setOperationAction(ISD::FCOS, MVT::f32, Expand);
437 setOperationAction(ISD::FCOS, MVT::f64, Expand);
438 setOperationAction(ISD::FREM, MVT::f64, Expand);
439 setOperationAction(ISD::FREM, MVT::f32, Expand);
440 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
441 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
442 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
444 setOperationAction(ISD::FPOW, MVT::f64, Expand);
445 setOperationAction(ISD::FPOW, MVT::f32, Expand);
447 // Various VFP goodness
448 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
449 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
450 if (Subtarget->hasVFP2()) {
451 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
452 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
453 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
454 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
456 // Special handling for half-precision FP.
457 if (!Subtarget->hasFP16()) {
458 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
459 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
463 // We have target-specific dag combine patterns for the following nodes:
464 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
465 setTargetDAGCombine(ISD::ADD);
466 setTargetDAGCombine(ISD::SUB);
467 setTargetDAGCombine(ISD::MUL);
469 setStackPointerRegisterToSaveRestore(ARM::SP);
471 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
472 setSchedulingPreference(Sched::RegPressure);
474 setSchedulingPreference(Sched::Hybrid);
476 // FIXME: If-converter should use instruction latency to determine
477 // profitability rather than relying on fixed limits.
478 if (Subtarget->getCPUString() == "generic") {
479 // Generic (and overly aggressive) if-conversion limits.
480 setIfCvtBlockSizeLimit(10);
481 setIfCvtDupBlockSizeLimit(2);
482 } else if (Subtarget->hasV7Ops()) {
483 setIfCvtBlockSizeLimit(3);
484 setIfCvtDupBlockSizeLimit(1);
485 } else if (Subtarget->hasV6Ops()) {
486 setIfCvtBlockSizeLimit(2);
487 setIfCvtDupBlockSizeLimit(1);
489 setIfCvtBlockSizeLimit(3);
490 setIfCvtDupBlockSizeLimit(2);
493 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
494 // Do not enable CodePlacementOpt for now: it currently runs after the
495 // ARMConstantIslandPass and messes up branch relaxation and placement
496 // of constant islands.
497 // benefitFromCodePlacementOpt = true;
500 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
503 case ARMISD::Wrapper: return "ARMISD::Wrapper";
504 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
505 case ARMISD::CALL: return "ARMISD::CALL";
506 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
507 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
508 case ARMISD::tCALL: return "ARMISD::tCALL";
509 case ARMISD::BRCOND: return "ARMISD::BRCOND";
510 case ARMISD::BR_JT: return "ARMISD::BR_JT";
511 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
512 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
513 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
514 case ARMISD::CMP: return "ARMISD::CMP";
515 case ARMISD::CMPZ: return "ARMISD::CMPZ";
516 case ARMISD::CMPFP: return "ARMISD::CMPFP";
517 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
518 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
519 case ARMISD::CMOV: return "ARMISD::CMOV";
520 case ARMISD::CNEG: return "ARMISD::CNEG";
522 case ARMISD::RBIT: return "ARMISD::RBIT";
524 case ARMISD::FTOSI: return "ARMISD::FTOSI";
525 case ARMISD::FTOUI: return "ARMISD::FTOUI";
526 case ARMISD::SITOF: return "ARMISD::SITOF";
527 case ARMISD::UITOF: return "ARMISD::UITOF";
529 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
530 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
531 case ARMISD::RRX: return "ARMISD::RRX";
533 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
534 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
536 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
537 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
539 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
541 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
543 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
544 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
546 case ARMISD::VCEQ: return "ARMISD::VCEQ";
547 case ARMISD::VCGE: return "ARMISD::VCGE";
548 case ARMISD::VCGEU: return "ARMISD::VCGEU";
549 case ARMISD::VCGT: return "ARMISD::VCGT";
550 case ARMISD::VCGTU: return "ARMISD::VCGTU";
551 case ARMISD::VTST: return "ARMISD::VTST";
553 case ARMISD::VSHL: return "ARMISD::VSHL";
554 case ARMISD::VSHRs: return "ARMISD::VSHRs";
555 case ARMISD::VSHRu: return "ARMISD::VSHRu";
556 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
557 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
558 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
559 case ARMISD::VSHRN: return "ARMISD::VSHRN";
560 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
561 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
562 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
563 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
564 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
565 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
566 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
567 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
568 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
569 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
570 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
571 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
572 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
573 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
574 case ARMISD::VDUP: return "ARMISD::VDUP";
575 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
576 case ARMISD::VEXT: return "ARMISD::VEXT";
577 case ARMISD::VREV64: return "ARMISD::VREV64";
578 case ARMISD::VREV32: return "ARMISD::VREV32";
579 case ARMISD::VREV16: return "ARMISD::VREV16";
580 case ARMISD::VZIP: return "ARMISD::VZIP";
581 case ARMISD::VUZP: return "ARMISD::VUZP";
582 case ARMISD::VTRN: return "ARMISD::VTRN";
583 case ARMISD::FMAX: return "ARMISD::FMAX";
584 case ARMISD::FMIN: return "ARMISD::FMIN";
588 /// getRegClassFor - Return the register class that should be used for the
589 /// specified value type.
590 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
591 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
592 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
593 // load / store 4 to 8 consecutive D registers.
594 if (Subtarget->hasNEON()) {
595 if (VT == MVT::v4i64)
596 return ARM::QQPRRegisterClass;
597 else if (VT == MVT::v8i64)
598 return ARM::QQQQPRRegisterClass;
600 return TargetLowering::getRegClassFor(VT);
603 /// getFunctionAlignment - Return the Log2 alignment of this function.
604 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
605 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
608 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
609 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
610 EVT VT = N->getValueType(i);
611 if (VT.isFloatingPoint() || VT.isVector())
612 return Sched::Latency;
614 return Sched::RegPressure;
617 //===----------------------------------------------------------------------===//
619 //===----------------------------------------------------------------------===//
621 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
622 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
624 default: llvm_unreachable("Unknown condition code!");
625 case ISD::SETNE: return ARMCC::NE;
626 case ISD::SETEQ: return ARMCC::EQ;
627 case ISD::SETGT: return ARMCC::GT;
628 case ISD::SETGE: return ARMCC::GE;
629 case ISD::SETLT: return ARMCC::LT;
630 case ISD::SETLE: return ARMCC::LE;
631 case ISD::SETUGT: return ARMCC::HI;
632 case ISD::SETUGE: return ARMCC::HS;
633 case ISD::SETULT: return ARMCC::LO;
634 case ISD::SETULE: return ARMCC::LS;
638 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
639 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
640 ARMCC::CondCodes &CondCode2) {
641 CondCode2 = ARMCC::AL;
643 default: llvm_unreachable("Unknown FP condition!");
645 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
647 case ISD::SETOGT: CondCode = ARMCC::GT; break;
649 case ISD::SETOGE: CondCode = ARMCC::GE; break;
650 case ISD::SETOLT: CondCode = ARMCC::MI; break;
651 case ISD::SETOLE: CondCode = ARMCC::LS; break;
652 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
653 case ISD::SETO: CondCode = ARMCC::VC; break;
654 case ISD::SETUO: CondCode = ARMCC::VS; break;
655 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
656 case ISD::SETUGT: CondCode = ARMCC::HI; break;
657 case ISD::SETUGE: CondCode = ARMCC::PL; break;
659 case ISD::SETULT: CondCode = ARMCC::LT; break;
661 case ISD::SETULE: CondCode = ARMCC::LE; break;
663 case ISD::SETUNE: CondCode = ARMCC::NE; break;
667 //===----------------------------------------------------------------------===//
668 // Calling Convention Implementation
669 //===----------------------------------------------------------------------===//
671 #include "ARMGenCallingConv.inc"
673 // APCS f64 is in register pairs, possibly split to stack
674 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
675 CCValAssign::LocInfo &LocInfo,
676 CCState &State, bool CanFail) {
677 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
679 // Try to get the first register.
680 if (unsigned Reg = State.AllocateReg(RegList, 4))
681 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
683 // For the 2nd half of a v2f64, do not fail.
687 // Put the whole thing on the stack.
688 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
689 State.AllocateStack(8, 4),
694 // Try to get the second register.
695 if (unsigned Reg = State.AllocateReg(RegList, 4))
696 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
698 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
699 State.AllocateStack(4, 4),
704 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
705 CCValAssign::LocInfo &LocInfo,
706 ISD::ArgFlagsTy &ArgFlags,
708 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
710 if (LocVT == MVT::v2f64 &&
711 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
713 return true; // we handled it
716 // AAPCS f64 is in aligned register pairs
717 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
718 CCValAssign::LocInfo &LocInfo,
719 CCState &State, bool CanFail) {
720 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
721 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
723 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
725 // For the 2nd half of a v2f64, do not just fail.
729 // Put the whole thing on the stack.
730 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
731 State.AllocateStack(8, 8),
737 for (i = 0; i < 2; ++i)
738 if (HiRegList[i] == Reg)
741 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
742 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
747 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
748 CCValAssign::LocInfo &LocInfo,
749 ISD::ArgFlagsTy &ArgFlags,
751 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
753 if (LocVT == MVT::v2f64 &&
754 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
756 return true; // we handled it
759 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
760 CCValAssign::LocInfo &LocInfo, CCState &State) {
761 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
762 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
764 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
766 return false; // we didn't handle it
769 for (i = 0; i < 2; ++i)
770 if (HiRegList[i] == Reg)
773 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
774 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
779 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
780 CCValAssign::LocInfo &LocInfo,
781 ISD::ArgFlagsTy &ArgFlags,
783 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
785 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
787 return true; // we handled it
790 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
791 CCValAssign::LocInfo &LocInfo,
792 ISD::ArgFlagsTy &ArgFlags,
794 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
798 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
799 /// given CallingConvention value.
800 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
802 bool isVarArg) const {
805 llvm_unreachable("Unsupported calling convention");
807 case CallingConv::Fast:
808 // Use target triple & subtarget features to do actual dispatch.
809 if (Subtarget->isAAPCS_ABI()) {
810 if (Subtarget->hasVFP2() &&
811 FloatABIType == FloatABI::Hard && !isVarArg)
812 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
814 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
816 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
817 case CallingConv::ARM_AAPCS_VFP:
818 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
819 case CallingConv::ARM_AAPCS:
820 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
821 case CallingConv::ARM_APCS:
822 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
826 /// LowerCallResult - Lower the result values of a call into the
827 /// appropriate copies out of appropriate physical registers.
829 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
830 CallingConv::ID CallConv, bool isVarArg,
831 const SmallVectorImpl<ISD::InputArg> &Ins,
832 DebugLoc dl, SelectionDAG &DAG,
833 SmallVectorImpl<SDValue> &InVals) const {
835 // Assign locations to each value returned by this call.
836 SmallVector<CCValAssign, 16> RVLocs;
837 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
838 RVLocs, *DAG.getContext());
839 CCInfo.AnalyzeCallResult(Ins,
840 CCAssignFnForNode(CallConv, /* Return*/ true,
843 // Copy all of the result registers out of their specified physreg.
844 for (unsigned i = 0; i != RVLocs.size(); ++i) {
845 CCValAssign VA = RVLocs[i];
848 if (VA.needsCustom()) {
849 // Handle f64 or half of a v2f64.
850 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
852 Chain = Lo.getValue(1);
853 InFlag = Lo.getValue(2);
854 VA = RVLocs[++i]; // skip ahead to next loc
855 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
857 Chain = Hi.getValue(1);
858 InFlag = Hi.getValue(2);
859 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
861 if (VA.getLocVT() == MVT::v2f64) {
862 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
863 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
864 DAG.getConstant(0, MVT::i32));
866 VA = RVLocs[++i]; // skip ahead to next loc
867 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
868 Chain = Lo.getValue(1);
869 InFlag = Lo.getValue(2);
870 VA = RVLocs[++i]; // skip ahead to next loc
871 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
872 Chain = Hi.getValue(1);
873 InFlag = Hi.getValue(2);
874 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
875 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
876 DAG.getConstant(1, MVT::i32));
879 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
881 Chain = Val.getValue(1);
882 InFlag = Val.getValue(2);
885 switch (VA.getLocInfo()) {
886 default: llvm_unreachable("Unknown loc info!");
887 case CCValAssign::Full: break;
888 case CCValAssign::BCvt:
889 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
893 InVals.push_back(Val);
899 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
900 /// by "Src" to address "Dst" of size "Size". Alignment information is
901 /// specified by the specific parameter attribute. The copy will be passed as
902 /// a byval function parameter.
903 /// Sometimes what we are copying is the end of a larger object, the part that
904 /// does not fit in registers.
906 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
907 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
909 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
910 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
911 /*isVolatile=*/false, /*AlwaysInline=*/false,
915 /// LowerMemOpCallTo - Store the argument to the stack.
917 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
918 SDValue StackPtr, SDValue Arg,
919 DebugLoc dl, SelectionDAG &DAG,
920 const CCValAssign &VA,
921 ISD::ArgFlagsTy Flags) const {
922 unsigned LocMemOffset = VA.getLocMemOffset();
923 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
924 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
925 if (Flags.isByVal()) {
926 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
928 return DAG.getStore(Chain, dl, Arg, PtrOff,
929 PseudoSourceValue::getStack(), LocMemOffset,
933 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
934 SDValue Chain, SDValue &Arg,
935 RegsToPassVector &RegsToPass,
936 CCValAssign &VA, CCValAssign &NextVA,
938 SmallVector<SDValue, 8> &MemOpChains,
939 ISD::ArgFlagsTy Flags) const {
941 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
942 DAG.getVTList(MVT::i32, MVT::i32), Arg);
943 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
945 if (NextVA.isRegLoc())
946 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
948 assert(NextVA.isMemLoc());
949 if (StackPtr.getNode() == 0)
950 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
952 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
958 /// LowerCall - Lowering a call into a callseq_start <-
959 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
962 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
963 CallingConv::ID CallConv, bool isVarArg,
965 const SmallVectorImpl<ISD::OutputArg> &Outs,
966 const SmallVectorImpl<ISD::InputArg> &Ins,
967 DebugLoc dl, SelectionDAG &DAG,
968 SmallVectorImpl<SDValue> &InVals) const {
969 // ARM target does not yet support tail call optimization.
972 // Analyze operands of the call, assigning locations to each operand.
973 SmallVector<CCValAssign, 16> ArgLocs;
974 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
976 CCInfo.AnalyzeCallOperands(Outs,
977 CCAssignFnForNode(CallConv, /* Return*/ false,
980 // Get a count of how many bytes are to be pushed on the stack.
981 unsigned NumBytes = CCInfo.getNextStackOffset();
983 // Adjust the stack pointer for the new arguments...
984 // These operations are automatically eliminated by the prolog/epilog pass
985 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
987 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
989 RegsToPassVector RegsToPass;
990 SmallVector<SDValue, 8> MemOpChains;
992 // Walk the register/memloc assignments, inserting copies/loads. In the case
993 // of tail call optimization, arguments are handled later.
994 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
997 CCValAssign &VA = ArgLocs[i];
998 SDValue Arg = Outs[realArgIdx].Val;
999 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1001 // Promote the value if needed.
1002 switch (VA.getLocInfo()) {
1003 default: llvm_unreachable("Unknown loc info!");
1004 case CCValAssign::Full: break;
1005 case CCValAssign::SExt:
1006 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1008 case CCValAssign::ZExt:
1009 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1011 case CCValAssign::AExt:
1012 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1014 case CCValAssign::BCvt:
1015 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1019 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1020 if (VA.needsCustom()) {
1021 if (VA.getLocVT() == MVT::v2f64) {
1022 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1023 DAG.getConstant(0, MVT::i32));
1024 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1025 DAG.getConstant(1, MVT::i32));
1027 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1028 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1030 VA = ArgLocs[++i]; // skip ahead to next loc
1031 if (VA.isRegLoc()) {
1032 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1033 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1035 assert(VA.isMemLoc());
1037 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1038 dl, DAG, VA, Flags));
1041 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1042 StackPtr, MemOpChains, Flags);
1044 } else if (VA.isRegLoc()) {
1045 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1047 assert(VA.isMemLoc());
1049 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1050 dl, DAG, VA, Flags));
1054 if (!MemOpChains.empty())
1055 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1056 &MemOpChains[0], MemOpChains.size());
1058 // Build a sequence of copy-to-reg nodes chained together with token chain
1059 // and flag operands which copy the outgoing args into the appropriate regs.
1061 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1062 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1063 RegsToPass[i].second, InFlag);
1064 InFlag = Chain.getValue(1);
1067 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1068 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1069 // node so that legalize doesn't hack it.
1070 bool isDirect = false;
1071 bool isARMFunc = false;
1072 bool isLocalARMFunc = false;
1073 MachineFunction &MF = DAG.getMachineFunction();
1074 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1076 if (EnableARMLongCalls) {
1077 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1078 && "long-calls with non-static relocation model!");
1079 // Handle a global address or an external symbol. If it's not one of
1080 // those, the target's already in a register, so we don't need to do
1082 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1083 const GlobalValue *GV = G->getGlobal();
1084 // Create a constant pool entry for the callee address
1085 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1086 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1089 // Get the address of the callee into a register
1090 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1091 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1092 Callee = DAG.getLoad(getPointerTy(), dl,
1093 DAG.getEntryNode(), CPAddr,
1094 PseudoSourceValue::getConstantPool(), 0,
1096 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1097 const char *Sym = S->getSymbol();
1099 // Create a constant pool entry for the callee address
1100 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1101 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1102 Sym, ARMPCLabelIndex, 0);
1103 // Get the address of the callee into a register
1104 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1105 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1106 Callee = DAG.getLoad(getPointerTy(), dl,
1107 DAG.getEntryNode(), CPAddr,
1108 PseudoSourceValue::getConstantPool(), 0,
1111 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1112 const GlobalValue *GV = G->getGlobal();
1114 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1115 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1116 getTargetMachine().getRelocationModel() != Reloc::Static;
1117 isARMFunc = !Subtarget->isThumb() || isStub;
1118 // ARM call to a local ARM function is predicable.
1119 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1120 // tBX takes a register source operand.
1121 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1122 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1123 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1126 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1127 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1128 Callee = DAG.getLoad(getPointerTy(), dl,
1129 DAG.getEntryNode(), CPAddr,
1130 PseudoSourceValue::getConstantPool(), 0,
1132 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1133 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1134 getPointerTy(), Callee, PICLabel);
1136 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1137 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1139 bool isStub = Subtarget->isTargetDarwin() &&
1140 getTargetMachine().getRelocationModel() != Reloc::Static;
1141 isARMFunc = !Subtarget->isThumb() || isStub;
1142 // tBX takes a register source operand.
1143 const char *Sym = S->getSymbol();
1144 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1145 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1146 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1147 Sym, ARMPCLabelIndex, 4);
1148 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1149 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1150 Callee = DAG.getLoad(getPointerTy(), dl,
1151 DAG.getEntryNode(), CPAddr,
1152 PseudoSourceValue::getConstantPool(), 0,
1154 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1155 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1156 getPointerTy(), Callee, PICLabel);
1158 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1161 // FIXME: handle tail calls differently.
1163 if (Subtarget->isThumb()) {
1164 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1165 CallOpc = ARMISD::CALL_NOLINK;
1167 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1169 CallOpc = (isDirect || Subtarget->hasV5TOps())
1170 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1171 : ARMISD::CALL_NOLINK;
1173 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1174 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1175 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1176 InFlag = Chain.getValue(1);
1179 std::vector<SDValue> Ops;
1180 Ops.push_back(Chain);
1181 Ops.push_back(Callee);
1183 // Add argument registers to the end of the list so that they are known live
1185 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1186 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1187 RegsToPass[i].second.getValueType()));
1189 if (InFlag.getNode())
1190 Ops.push_back(InFlag);
1191 // Returns a chain and a flag for retval copy to use.
1192 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1193 &Ops[0], Ops.size());
1194 InFlag = Chain.getValue(1);
1196 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1197 DAG.getIntPtrConstant(0, true), InFlag);
1199 InFlag = Chain.getValue(1);
1201 // Handle result values, copying them out of physregs into vregs that we
1203 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1208 ARMTargetLowering::LowerReturn(SDValue Chain,
1209 CallingConv::ID CallConv, bool isVarArg,
1210 const SmallVectorImpl<ISD::OutputArg> &Outs,
1211 DebugLoc dl, SelectionDAG &DAG) const {
1213 // CCValAssign - represent the assignment of the return value to a location.
1214 SmallVector<CCValAssign, 16> RVLocs;
1216 // CCState - Info about the registers and stack slots.
1217 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1220 // Analyze outgoing return values.
1221 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1224 // If this is the first return lowered for this function, add
1225 // the regs to the liveout set for the function.
1226 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1227 for (unsigned i = 0; i != RVLocs.size(); ++i)
1228 if (RVLocs[i].isRegLoc())
1229 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1234 // Copy the result values into the output registers.
1235 for (unsigned i = 0, realRVLocIdx = 0;
1237 ++i, ++realRVLocIdx) {
1238 CCValAssign &VA = RVLocs[i];
1239 assert(VA.isRegLoc() && "Can only return in registers!");
1241 SDValue Arg = Outs[realRVLocIdx].Val;
1243 switch (VA.getLocInfo()) {
1244 default: llvm_unreachable("Unknown loc info!");
1245 case CCValAssign::Full: break;
1246 case CCValAssign::BCvt:
1247 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1251 if (VA.needsCustom()) {
1252 if (VA.getLocVT() == MVT::v2f64) {
1253 // Extract the first half and return it in two registers.
1254 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1255 DAG.getConstant(0, MVT::i32));
1256 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1257 DAG.getVTList(MVT::i32, MVT::i32), Half);
1259 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1260 Flag = Chain.getValue(1);
1261 VA = RVLocs[++i]; // skip ahead to next loc
1262 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1263 HalfGPRs.getValue(1), Flag);
1264 Flag = Chain.getValue(1);
1265 VA = RVLocs[++i]; // skip ahead to next loc
1267 // Extract the 2nd half and fall through to handle it as an f64 value.
1268 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1269 DAG.getConstant(1, MVT::i32));
1271 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1273 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1274 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1275 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1276 Flag = Chain.getValue(1);
1277 VA = RVLocs[++i]; // skip ahead to next loc
1278 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1281 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1283 // Guarantee that all emitted copies are
1284 // stuck together, avoiding something bad.
1285 Flag = Chain.getValue(1);
1290 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1292 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1297 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1298 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1299 // one of the above mentioned nodes. It has to be wrapped because otherwise
1300 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1301 // be used to form addressing mode. These wrapped nodes will be selected
1303 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1304 EVT PtrVT = Op.getValueType();
1305 // FIXME there is no actual debug info here
1306 DebugLoc dl = Op.getDebugLoc();
1307 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1309 if (CP->isMachineConstantPoolEntry())
1310 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1311 CP->getAlignment());
1313 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1314 CP->getAlignment());
1315 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1318 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1319 SelectionDAG &DAG) const {
1320 MachineFunction &MF = DAG.getMachineFunction();
1321 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1322 unsigned ARMPCLabelIndex = 0;
1323 DebugLoc DL = Op.getDebugLoc();
1324 EVT PtrVT = getPointerTy();
1325 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1326 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1328 if (RelocM == Reloc::Static) {
1329 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1331 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1332 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1333 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1334 ARMCP::CPBlockAddress,
1336 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1338 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1339 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1340 PseudoSourceValue::getConstantPool(), 0,
1342 if (RelocM == Reloc::Static)
1344 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1345 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1348 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1350 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1351 SelectionDAG &DAG) const {
1352 DebugLoc dl = GA->getDebugLoc();
1353 EVT PtrVT = getPointerTy();
1354 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1355 MachineFunction &MF = DAG.getMachineFunction();
1356 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1357 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1358 ARMConstantPoolValue *CPV =
1359 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1360 ARMCP::CPValue, PCAdj, "tlsgd", true);
1361 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1362 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1363 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1364 PseudoSourceValue::getConstantPool(), 0,
1366 SDValue Chain = Argument.getValue(1);
1368 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1369 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1371 // call __tls_get_addr.
1374 Entry.Node = Argument;
1375 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1376 Args.push_back(Entry);
1377 // FIXME: is there useful debug info available here?
1378 std::pair<SDValue, SDValue> CallResult =
1379 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1380 false, false, false, false,
1381 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1382 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1383 return CallResult.first;
1386 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1387 // "local exec" model.
1389 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1390 SelectionDAG &DAG) const {
1391 const GlobalValue *GV = GA->getGlobal();
1392 DebugLoc dl = GA->getDebugLoc();
1394 SDValue Chain = DAG.getEntryNode();
1395 EVT PtrVT = getPointerTy();
1396 // Get the Thread Pointer
1397 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1399 if (GV->isDeclaration()) {
1400 MachineFunction &MF = DAG.getMachineFunction();
1401 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1402 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1403 // Initial exec model.
1404 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1405 ARMConstantPoolValue *CPV =
1406 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1407 ARMCP::CPValue, PCAdj, "gottpoff", true);
1408 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1409 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1410 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1411 PseudoSourceValue::getConstantPool(), 0,
1413 Chain = Offset.getValue(1);
1415 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1416 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1418 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1419 PseudoSourceValue::getConstantPool(), 0,
1423 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1424 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1425 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1426 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1427 PseudoSourceValue::getConstantPool(), 0,
1431 // The address of the thread local variable is the add of the thread
1432 // pointer with the offset of the variable.
1433 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1437 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1438 // TODO: implement the "local dynamic" model
1439 assert(Subtarget->isTargetELF() &&
1440 "TLS not implemented for non-ELF targets");
1441 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1442 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1443 // otherwise use the "Local Exec" TLS Model
1444 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1445 return LowerToTLSGeneralDynamicModel(GA, DAG);
1447 return LowerToTLSExecModels(GA, DAG);
1450 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1451 SelectionDAG &DAG) const {
1452 EVT PtrVT = getPointerTy();
1453 DebugLoc dl = Op.getDebugLoc();
1454 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1455 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1456 if (RelocM == Reloc::PIC_) {
1457 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1458 ARMConstantPoolValue *CPV =
1459 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1460 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1461 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1462 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1464 PseudoSourceValue::getConstantPool(), 0,
1466 SDValue Chain = Result.getValue(1);
1467 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1468 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1470 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1471 PseudoSourceValue::getGOT(), 0,
1475 // If we have T2 ops, we can materialize the address directly via movt/movw
1476 // pair. This is always cheaper.
1477 if (Subtarget->useMovt()) {
1478 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1479 DAG.getTargetGlobalAddress(GV, PtrVT));
1481 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1482 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1483 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1484 PseudoSourceValue::getConstantPool(), 0,
1490 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1491 SelectionDAG &DAG) const {
1492 MachineFunction &MF = DAG.getMachineFunction();
1493 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1494 unsigned ARMPCLabelIndex = 0;
1495 EVT PtrVT = getPointerTy();
1496 DebugLoc dl = Op.getDebugLoc();
1497 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1498 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1500 if (RelocM == Reloc::Static)
1501 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1503 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1504 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1505 ARMConstantPoolValue *CPV =
1506 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1507 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1509 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1511 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1512 PseudoSourceValue::getConstantPool(), 0,
1514 SDValue Chain = Result.getValue(1);
1516 if (RelocM == Reloc::PIC_) {
1517 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1518 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1521 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1522 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1523 PseudoSourceValue::getGOT(), 0,
1529 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1530 SelectionDAG &DAG) const {
1531 assert(Subtarget->isTargetELF() &&
1532 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1533 MachineFunction &MF = DAG.getMachineFunction();
1534 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1535 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1536 EVT PtrVT = getPointerTy();
1537 DebugLoc dl = Op.getDebugLoc();
1538 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1539 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1540 "_GLOBAL_OFFSET_TABLE_",
1541 ARMPCLabelIndex, PCAdj);
1542 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1543 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1544 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1545 PseudoSourceValue::getConstantPool(), 0,
1547 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1548 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1552 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1553 DebugLoc dl = Op.getDebugLoc();
1554 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1555 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1559 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1560 const ARMSubtarget *Subtarget)
1562 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1563 DebugLoc dl = Op.getDebugLoc();
1565 default: return SDValue(); // Don't custom lower most intrinsics.
1566 case Intrinsic::arm_thread_pointer: {
1567 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1568 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1570 case Intrinsic::eh_sjlj_lsda: {
1571 MachineFunction &MF = DAG.getMachineFunction();
1572 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1573 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1574 EVT PtrVT = getPointerTy();
1575 DebugLoc dl = Op.getDebugLoc();
1576 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1578 unsigned PCAdj = (RelocM != Reloc::PIC_)
1579 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1580 ARMConstantPoolValue *CPV =
1581 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1582 ARMCP::CPLSDA, PCAdj);
1583 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1584 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1586 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1587 PseudoSourceValue::getConstantPool(), 0,
1589 SDValue Chain = Result.getValue(1);
1591 if (RelocM == Reloc::PIC_) {
1592 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1593 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1597 case Intrinsic::eh_sjlj_setjmp:
1598 SDValue Val = Subtarget->isThumb() ?
1599 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1600 DAG.getConstant(0, MVT::i32);
1601 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1606 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1607 const ARMSubtarget *Subtarget) {
1608 DebugLoc dl = Op.getDebugLoc();
1609 SDValue Op5 = Op.getOperand(5);
1611 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1612 if (isDeviceBarrier) {
1613 if (Subtarget->hasV7Ops())
1614 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1616 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1617 DAG.getConstant(0, MVT::i32));
1619 if (Subtarget->hasV7Ops())
1620 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1622 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1623 DAG.getConstant(0, MVT::i32));
1628 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1629 MachineFunction &MF = DAG.getMachineFunction();
1630 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1632 // vastart just stores the address of the VarArgsFrameIndex slot into the
1633 // memory location argument.
1634 DebugLoc dl = Op.getDebugLoc();
1635 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1636 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1637 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1638 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1643 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1644 SelectionDAG &DAG) const {
1645 SDNode *Node = Op.getNode();
1646 DebugLoc dl = Node->getDebugLoc();
1647 EVT VT = Node->getValueType(0);
1648 SDValue Chain = Op.getOperand(0);
1649 SDValue Size = Op.getOperand(1);
1650 SDValue Align = Op.getOperand(2);
1652 // Chain the dynamic stack allocation so that it doesn't modify the stack
1653 // pointer when other instructions are using the stack.
1654 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1656 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1657 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1658 if (AlignVal > StackAlign)
1659 // Do this now since selection pass cannot introduce new target
1660 // independent node.
1661 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1663 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1664 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1665 // do even more horrible hack later.
1666 MachineFunction &MF = DAG.getMachineFunction();
1667 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1668 if (AFI->isThumb1OnlyFunction()) {
1670 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1672 uint32_t Val = C->getZExtValue();
1673 if (Val <= 508 && ((Val & 3) == 0))
1677 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1680 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1681 SDValue Ops1[] = { Chain, Size, Align };
1682 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1683 Chain = Res.getValue(1);
1684 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1685 DAG.getIntPtrConstant(0, true), SDValue());
1686 SDValue Ops2[] = { Res, Chain };
1687 return DAG.getMergeValues(Ops2, 2, dl);
1691 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1692 SDValue &Root, SelectionDAG &DAG,
1693 DebugLoc dl) const {
1694 MachineFunction &MF = DAG.getMachineFunction();
1695 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1697 TargetRegisterClass *RC;
1698 if (AFI->isThumb1OnlyFunction())
1699 RC = ARM::tGPRRegisterClass;
1701 RC = ARM::GPRRegisterClass;
1703 // Transform the arguments stored in physical registers into virtual ones.
1704 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1705 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1708 if (NextVA.isMemLoc()) {
1709 MachineFrameInfo *MFI = MF.getFrameInfo();
1710 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1712 // Create load node to retrieve arguments from the stack.
1713 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1714 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1715 PseudoSourceValue::getFixedStack(FI), 0,
1718 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1719 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1722 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1726 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1727 CallingConv::ID CallConv, bool isVarArg,
1728 const SmallVectorImpl<ISD::InputArg>
1730 DebugLoc dl, SelectionDAG &DAG,
1731 SmallVectorImpl<SDValue> &InVals)
1734 MachineFunction &MF = DAG.getMachineFunction();
1735 MachineFrameInfo *MFI = MF.getFrameInfo();
1737 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1739 // Assign locations to all of the incoming arguments.
1740 SmallVector<CCValAssign, 16> ArgLocs;
1741 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1743 CCInfo.AnalyzeFormalArguments(Ins,
1744 CCAssignFnForNode(CallConv, /* Return*/ false,
1747 SmallVector<SDValue, 16> ArgValues;
1749 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1750 CCValAssign &VA = ArgLocs[i];
1752 // Arguments stored in registers.
1753 if (VA.isRegLoc()) {
1754 EVT RegVT = VA.getLocVT();
1757 if (VA.needsCustom()) {
1758 // f64 and vector types are split up into multiple registers or
1759 // combinations of registers and stack slots.
1760 if (VA.getLocVT() == MVT::v2f64) {
1761 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1763 VA = ArgLocs[++i]; // skip ahead to next loc
1765 if (VA.isMemLoc()) {
1766 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
1768 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1769 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1770 PseudoSourceValue::getFixedStack(FI), 0,
1773 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1776 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1777 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1778 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1779 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1780 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1782 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1785 TargetRegisterClass *RC;
1787 if (RegVT == MVT::f32)
1788 RC = ARM::SPRRegisterClass;
1789 else if (RegVT == MVT::f64)
1790 RC = ARM::DPRRegisterClass;
1791 else if (RegVT == MVT::v2f64)
1792 RC = ARM::QPRRegisterClass;
1793 else if (RegVT == MVT::i32)
1794 RC = (AFI->isThumb1OnlyFunction() ?
1795 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1797 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1799 // Transform the arguments in physical registers into virtual ones.
1800 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1801 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1804 // If this is an 8 or 16-bit value, it is really passed promoted
1805 // to 32 bits. Insert an assert[sz]ext to capture this, then
1806 // truncate to the right size.
1807 switch (VA.getLocInfo()) {
1808 default: llvm_unreachable("Unknown loc info!");
1809 case CCValAssign::Full: break;
1810 case CCValAssign::BCvt:
1811 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1813 case CCValAssign::SExt:
1814 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1815 DAG.getValueType(VA.getValVT()));
1816 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1818 case CCValAssign::ZExt:
1819 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1820 DAG.getValueType(VA.getValVT()));
1821 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1825 InVals.push_back(ArgValue);
1827 } else { // VA.isRegLoc()
1830 assert(VA.isMemLoc());
1831 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1833 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1834 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1837 // Create load nodes to retrieve arguments from the stack.
1838 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1839 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1840 PseudoSourceValue::getFixedStack(FI), 0,
1847 static const unsigned GPRArgRegs[] = {
1848 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1851 unsigned NumGPRs = CCInfo.getFirstUnallocated
1852 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1854 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1855 unsigned VARegSize = (4 - NumGPRs) * 4;
1856 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1857 unsigned ArgOffset = CCInfo.getNextStackOffset();
1858 if (VARegSaveSize) {
1859 // If this function is vararg, store any remaining integer argument regs
1860 // to their spots on the stack so that they may be loaded by deferencing
1861 // the result of va_next.
1862 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1863 AFI->setVarArgsFrameIndex(
1864 MFI->CreateFixedObject(VARegSaveSize,
1865 ArgOffset + VARegSaveSize - VARegSize,
1867 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
1870 SmallVector<SDValue, 4> MemOps;
1871 for (; NumGPRs < 4; ++NumGPRs) {
1872 TargetRegisterClass *RC;
1873 if (AFI->isThumb1OnlyFunction())
1874 RC = ARM::tGPRRegisterClass;
1876 RC = ARM::GPRRegisterClass;
1878 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1879 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1881 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1882 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()), 0,
1884 MemOps.push_back(Store);
1885 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1886 DAG.getConstant(4, getPointerTy()));
1888 if (!MemOps.empty())
1889 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1890 &MemOps[0], MemOps.size());
1892 // This will point to the next argument passed via stack.
1893 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
1900 /// isFloatingPointZero - Return true if this is +0.0.
1901 static bool isFloatingPointZero(SDValue Op) {
1902 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1903 return CFP->getValueAPF().isPosZero();
1904 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1905 // Maybe this has already been legalized into the constant pool?
1906 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1907 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1908 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1909 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1910 return CFP->getValueAPF().isPosZero();
1916 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1917 /// the given operands.
1919 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1920 SDValue &ARMCC, SelectionDAG &DAG,
1921 DebugLoc dl) const {
1922 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1923 unsigned C = RHSC->getZExtValue();
1924 if (!isLegalICmpImmediate(C)) {
1925 // Constant does not fit, try adjusting it by one?
1930 if (isLegalICmpImmediate(C-1)) {
1931 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1932 RHS = DAG.getConstant(C-1, MVT::i32);
1937 if (C > 0 && isLegalICmpImmediate(C-1)) {
1938 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1939 RHS = DAG.getConstant(C-1, MVT::i32);
1944 if (isLegalICmpImmediate(C+1)) {
1945 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1946 RHS = DAG.getConstant(C+1, MVT::i32);
1951 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1952 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1953 RHS = DAG.getConstant(C+1, MVT::i32);
1960 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1961 ARMISD::NodeType CompareType;
1964 CompareType = ARMISD::CMP;
1969 CompareType = ARMISD::CMPZ;
1972 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1973 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1976 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1977 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1980 if (!isFloatingPointZero(RHS))
1981 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1983 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1984 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1987 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
1988 EVT VT = Op.getValueType();
1989 SDValue LHS = Op.getOperand(0);
1990 SDValue RHS = Op.getOperand(1);
1991 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1992 SDValue TrueVal = Op.getOperand(2);
1993 SDValue FalseVal = Op.getOperand(3);
1994 DebugLoc dl = Op.getDebugLoc();
1996 if (LHS.getValueType() == MVT::i32) {
1998 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1999 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2000 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2003 ARMCC::CondCodes CondCode, CondCode2;
2004 FPCCToARMCC(CC, CondCode, CondCode2);
2006 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2007 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2008 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2009 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2011 if (CondCode2 != ARMCC::AL) {
2012 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2013 // FIXME: Needs another CMP because flag can have but one use.
2014 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2015 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2016 Result, TrueVal, ARMCC2, CCR, Cmp2);
2021 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2022 SDValue Chain = Op.getOperand(0);
2023 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2024 SDValue LHS = Op.getOperand(2);
2025 SDValue RHS = Op.getOperand(3);
2026 SDValue Dest = Op.getOperand(4);
2027 DebugLoc dl = Op.getDebugLoc();
2029 if (LHS.getValueType() == MVT::i32) {
2031 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2032 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2033 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2034 Chain, Dest, ARMCC, CCR,Cmp);
2037 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2038 ARMCC::CondCodes CondCode, CondCode2;
2039 FPCCToARMCC(CC, CondCode, CondCode2);
2041 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2042 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2043 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2044 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2045 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2046 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2047 if (CondCode2 != ARMCC::AL) {
2048 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2049 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2050 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2055 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2056 SDValue Chain = Op.getOperand(0);
2057 SDValue Table = Op.getOperand(1);
2058 SDValue Index = Op.getOperand(2);
2059 DebugLoc dl = Op.getDebugLoc();
2061 EVT PTy = getPointerTy();
2062 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2063 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2064 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2065 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2066 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2067 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2068 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2069 if (Subtarget->isThumb2()) {
2070 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2071 // which does another jump to the destination. This also makes it easier
2072 // to translate it to TBB / TBH later.
2073 // FIXME: This might not work if the function is extremely large.
2074 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2075 Addr, Op.getOperand(2), JTI, UId);
2077 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2078 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2079 PseudoSourceValue::getJumpTable(), 0,
2081 Chain = Addr.getValue(1);
2082 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2083 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2085 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2086 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2087 Chain = Addr.getValue(1);
2088 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2092 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2093 DebugLoc dl = Op.getDebugLoc();
2096 switch (Op.getOpcode()) {
2098 assert(0 && "Invalid opcode!");
2099 case ISD::FP_TO_SINT:
2100 Opc = ARMISD::FTOSI;
2102 case ISD::FP_TO_UINT:
2103 Opc = ARMISD::FTOUI;
2106 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2107 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2110 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2111 EVT VT = Op.getValueType();
2112 DebugLoc dl = Op.getDebugLoc();
2115 switch (Op.getOpcode()) {
2117 assert(0 && "Invalid opcode!");
2118 case ISD::SINT_TO_FP:
2119 Opc = ARMISD::SITOF;
2121 case ISD::UINT_TO_FP:
2122 Opc = ARMISD::UITOF;
2126 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2127 return DAG.getNode(Opc, dl, VT, Op);
2130 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2131 // Implement fcopysign with a fabs and a conditional fneg.
2132 SDValue Tmp0 = Op.getOperand(0);
2133 SDValue Tmp1 = Op.getOperand(1);
2134 DebugLoc dl = Op.getDebugLoc();
2135 EVT VT = Op.getValueType();
2136 EVT SrcVT = Tmp1.getValueType();
2137 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2138 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2139 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2140 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2141 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2144 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2145 MachineFunction &MF = DAG.getMachineFunction();
2146 MachineFrameInfo *MFI = MF.getFrameInfo();
2147 MFI->setReturnAddressIsTaken(true);
2149 EVT VT = Op.getValueType();
2150 DebugLoc dl = Op.getDebugLoc();
2151 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2153 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2154 SDValue Offset = DAG.getConstant(4, MVT::i32);
2155 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2156 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2157 NULL, 0, false, false, 0);
2160 // Return LR, which contains the return address. Mark it an implicit live-in.
2161 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2162 TargetRegisterClass *RC = AFI->isThumb1OnlyFunction()
2163 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
2164 unsigned Reg = MF.addLiveIn(ARM::LR, RC);
2165 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2168 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2169 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2170 MFI->setFrameAddressIsTaken(true);
2172 EVT VT = Op.getValueType();
2173 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2174 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2175 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2176 ? ARM::R7 : ARM::R11;
2177 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2179 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2184 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2185 /// expand a bit convert where either the source or destination type is i64 to
2186 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2187 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2188 /// vectors), since the legalizer won't know what to do with that.
2189 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2190 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2191 DebugLoc dl = N->getDebugLoc();
2192 SDValue Op = N->getOperand(0);
2194 // This function is only supposed to be called for i64 types, either as the
2195 // source or destination of the bit convert.
2196 EVT SrcVT = Op.getValueType();
2197 EVT DstVT = N->getValueType(0);
2198 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2199 "ExpandBIT_CONVERT called for non-i64 type");
2201 // Turn i64->f64 into VMOVDRR.
2202 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2203 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2204 DAG.getConstant(0, MVT::i32));
2205 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2206 DAG.getConstant(1, MVT::i32));
2207 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2210 // Turn f64->i64 into VMOVRRD.
2211 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2212 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2213 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2214 // Merge the pieces into a single i64 value.
2215 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2221 /// getZeroVector - Returns a vector of specified type with all zero elements.
2223 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2224 assert(VT.isVector() && "Expected a vector type");
2226 // Zero vectors are used to represent vector negation and in those cases
2227 // will be implemented with the NEON VNEG instruction. However, VNEG does
2228 // not support i64 elements, so sometimes the zero vectors will need to be
2229 // explicitly constructed. For those cases, and potentially other uses in
2230 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2231 // to their dest type. This ensures they get CSE'd.
2233 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2234 SmallVector<SDValue, 8> Ops;
2237 if (VT.getSizeInBits() == 64) {
2238 Ops.assign(8, Cst); TVT = MVT::v8i8;
2240 Ops.assign(16, Cst); TVT = MVT::v16i8;
2242 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2244 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2247 /// getOnesVector - Returns a vector of specified type with all bits set.
2249 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2250 assert(VT.isVector() && "Expected a vector type");
2252 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2253 // dest type. This ensures they get CSE'd.
2255 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2256 SmallVector<SDValue, 8> Ops;
2259 if (VT.getSizeInBits() == 64) {
2260 Ops.assign(8, Cst); TVT = MVT::v8i8;
2262 Ops.assign(16, Cst); TVT = MVT::v16i8;
2264 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2266 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2269 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2270 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2271 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2272 SelectionDAG &DAG) const {
2273 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2274 EVT VT = Op.getValueType();
2275 unsigned VTBits = VT.getSizeInBits();
2276 DebugLoc dl = Op.getDebugLoc();
2277 SDValue ShOpLo = Op.getOperand(0);
2278 SDValue ShOpHi = Op.getOperand(1);
2279 SDValue ShAmt = Op.getOperand(2);
2281 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2283 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2285 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2286 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2287 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2288 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2289 DAG.getConstant(VTBits, MVT::i32));
2290 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2291 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2292 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2294 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2295 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2297 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2298 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2301 SDValue Ops[2] = { Lo, Hi };
2302 return DAG.getMergeValues(Ops, 2, dl);
2305 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2306 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2307 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2308 SelectionDAG &DAG) const {
2309 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2310 EVT VT = Op.getValueType();
2311 unsigned VTBits = VT.getSizeInBits();
2312 DebugLoc dl = Op.getDebugLoc();
2313 SDValue ShOpLo = Op.getOperand(0);
2314 SDValue ShOpHi = Op.getOperand(1);
2315 SDValue ShAmt = Op.getOperand(2);
2318 assert(Op.getOpcode() == ISD::SHL_PARTS);
2319 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2320 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2321 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2322 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2323 DAG.getConstant(VTBits, MVT::i32));
2324 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2325 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2327 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2328 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2329 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2331 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2332 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2335 SDValue Ops[2] = { Lo, Hi };
2336 return DAG.getMergeValues(Ops, 2, dl);
2339 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2340 const ARMSubtarget *ST) {
2341 EVT VT = N->getValueType(0);
2342 DebugLoc dl = N->getDebugLoc();
2344 if (!ST->hasV6T2Ops())
2347 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2348 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2351 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2352 const ARMSubtarget *ST) {
2353 EVT VT = N->getValueType(0);
2354 DebugLoc dl = N->getDebugLoc();
2356 // Lower vector shifts on NEON to use VSHL.
2357 if (VT.isVector()) {
2358 assert(ST->hasNEON() && "unexpected vector shift");
2360 // Left shifts translate directly to the vshiftu intrinsic.
2361 if (N->getOpcode() == ISD::SHL)
2362 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2363 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2364 N->getOperand(0), N->getOperand(1));
2366 assert((N->getOpcode() == ISD::SRA ||
2367 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2369 // NEON uses the same intrinsics for both left and right shifts. For
2370 // right shifts, the shift amounts are negative, so negate the vector of
2372 EVT ShiftVT = N->getOperand(1).getValueType();
2373 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2374 getZeroVector(ShiftVT, DAG, dl),
2376 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2377 Intrinsic::arm_neon_vshifts :
2378 Intrinsic::arm_neon_vshiftu);
2379 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2380 DAG.getConstant(vshiftInt, MVT::i32),
2381 N->getOperand(0), NegatedCount);
2384 // We can get here for a node like i32 = ISD::SHL i32, i64
2388 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2389 "Unknown shift to lower!");
2391 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2392 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2393 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2396 // If we are in thumb mode, we don't have RRX.
2397 if (ST->isThumb1Only()) return SDValue();
2399 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2400 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2401 DAG.getConstant(0, MVT::i32));
2402 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2403 DAG.getConstant(1, MVT::i32));
2405 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2406 // captures the result into a carry flag.
2407 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2408 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2410 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2411 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2413 // Merge the pieces into a single i64 value.
2414 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2417 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2418 SDValue TmpOp0, TmpOp1;
2419 bool Invert = false;
2423 SDValue Op0 = Op.getOperand(0);
2424 SDValue Op1 = Op.getOperand(1);
2425 SDValue CC = Op.getOperand(2);
2426 EVT VT = Op.getValueType();
2427 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2428 DebugLoc dl = Op.getDebugLoc();
2430 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2431 switch (SetCCOpcode) {
2432 default: llvm_unreachable("Illegal FP comparison"); break;
2434 case ISD::SETNE: Invert = true; // Fallthrough
2436 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2438 case ISD::SETLT: Swap = true; // Fallthrough
2440 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2442 case ISD::SETLE: Swap = true; // Fallthrough
2444 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2445 case ISD::SETUGE: Swap = true; // Fallthrough
2446 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2447 case ISD::SETUGT: Swap = true; // Fallthrough
2448 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2449 case ISD::SETUEQ: Invert = true; // Fallthrough
2451 // Expand this to (OLT | OGT).
2455 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2456 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2458 case ISD::SETUO: Invert = true; // Fallthrough
2460 // Expand this to (OLT | OGE).
2464 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2465 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2469 // Integer comparisons.
2470 switch (SetCCOpcode) {
2471 default: llvm_unreachable("Illegal integer comparison"); break;
2472 case ISD::SETNE: Invert = true;
2473 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2474 case ISD::SETLT: Swap = true;
2475 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2476 case ISD::SETLE: Swap = true;
2477 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2478 case ISD::SETULT: Swap = true;
2479 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2480 case ISD::SETULE: Swap = true;
2481 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2484 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2485 if (Opc == ARMISD::VCEQ) {
2488 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2490 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2493 // Ignore bitconvert.
2494 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2495 AndOp = AndOp.getOperand(0);
2497 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2499 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2500 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2507 std::swap(Op0, Op1);
2509 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2512 Result = DAG.getNOT(dl, Result, VT);
2517 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2518 /// VMOV instruction, and if so, return the constant being splatted.
2519 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2520 unsigned SplatBitSize, SelectionDAG &DAG) {
2521 switch (SplatBitSize) {
2523 // Any 1-byte value is OK.
2524 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2525 return DAG.getTargetConstant(SplatBits, MVT::i8);
2528 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2529 if ((SplatBits & ~0xff) == 0 ||
2530 (SplatBits & ~0xff00) == 0)
2531 return DAG.getTargetConstant(SplatBits, MVT::i16);
2535 // NEON's 32-bit VMOV supports splat values where:
2536 // * only one byte is nonzero, or
2537 // * the least significant byte is 0xff and the second byte is nonzero, or
2538 // * the least significant 2 bytes are 0xff and the third is nonzero.
2539 if ((SplatBits & ~0xff) == 0 ||
2540 (SplatBits & ~0xff00) == 0 ||
2541 (SplatBits & ~0xff0000) == 0 ||
2542 (SplatBits & ~0xff000000) == 0)
2543 return DAG.getTargetConstant(SplatBits, MVT::i32);
2545 if ((SplatBits & ~0xffff) == 0 &&
2546 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2547 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2549 if ((SplatBits & ~0xffffff) == 0 &&
2550 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2551 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2553 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2554 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2555 // VMOV.I32. A (very) minor optimization would be to replicate the value
2556 // and fall through here to test for a valid 64-bit splat. But, then the
2557 // caller would also need to check and handle the change in size.
2561 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2562 uint64_t BitMask = 0xff;
2564 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2565 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2567 else if ((SplatBits & BitMask) != 0)
2571 return DAG.getTargetConstant(Val, MVT::i64);
2575 llvm_unreachable("unexpected size for isVMOVSplat");
2582 /// getVMOVImm - If this is a build_vector of constants which can be
2583 /// formed by using a VMOV instruction of the specified element size,
2584 /// return the constant being splatted. The ByteSize field indicates the
2585 /// number of bytes of each element [1248].
2586 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2587 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2588 APInt SplatBits, SplatUndef;
2589 unsigned SplatBitSize;
2591 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2592 HasAnyUndefs, ByteSize * 8))
2595 if (SplatBitSize > ByteSize * 8)
2598 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2602 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2603 bool &ReverseVEXT, unsigned &Imm) {
2604 unsigned NumElts = VT.getVectorNumElements();
2605 ReverseVEXT = false;
2608 // If this is a VEXT shuffle, the immediate value is the index of the first
2609 // element. The other shuffle indices must be the successive elements after
2611 unsigned ExpectedElt = Imm;
2612 for (unsigned i = 1; i < NumElts; ++i) {
2613 // Increment the expected index. If it wraps around, it may still be
2614 // a VEXT but the source vectors must be swapped.
2616 if (ExpectedElt == NumElts * 2) {
2621 if (ExpectedElt != static_cast<unsigned>(M[i]))
2625 // Adjust the index value if the source operands will be swapped.
2632 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2633 /// instruction with the specified blocksize. (The order of the elements
2634 /// within each block of the vector is reversed.)
2635 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2636 unsigned BlockSize) {
2637 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2638 "Only possible block sizes for VREV are: 16, 32, 64");
2640 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2644 unsigned NumElts = VT.getVectorNumElements();
2645 unsigned BlockElts = M[0] + 1;
2647 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2650 for (unsigned i = 0; i < NumElts; ++i) {
2651 if ((unsigned) M[i] !=
2652 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2659 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2660 unsigned &WhichResult) {
2661 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2665 unsigned NumElts = VT.getVectorNumElements();
2666 WhichResult = (M[0] == 0 ? 0 : 1);
2667 for (unsigned i = 0; i < NumElts; i += 2) {
2668 if ((unsigned) M[i] != i + WhichResult ||
2669 (unsigned) M[i+1] != i + NumElts + WhichResult)
2675 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2676 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2677 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2678 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2679 unsigned &WhichResult) {
2680 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2684 unsigned NumElts = VT.getVectorNumElements();
2685 WhichResult = (M[0] == 0 ? 0 : 1);
2686 for (unsigned i = 0; i < NumElts; i += 2) {
2687 if ((unsigned) M[i] != i + WhichResult ||
2688 (unsigned) M[i+1] != i + WhichResult)
2694 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2695 unsigned &WhichResult) {
2696 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2700 unsigned NumElts = VT.getVectorNumElements();
2701 WhichResult = (M[0] == 0 ? 0 : 1);
2702 for (unsigned i = 0; i != NumElts; ++i) {
2703 if ((unsigned) M[i] != 2 * i + WhichResult)
2707 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2708 if (VT.is64BitVector() && EltSz == 32)
2714 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2715 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2716 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2717 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2718 unsigned &WhichResult) {
2719 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2723 unsigned Half = VT.getVectorNumElements() / 2;
2724 WhichResult = (M[0] == 0 ? 0 : 1);
2725 for (unsigned j = 0; j != 2; ++j) {
2726 unsigned Idx = WhichResult;
2727 for (unsigned i = 0; i != Half; ++i) {
2728 if ((unsigned) M[i + j * Half] != Idx)
2734 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2735 if (VT.is64BitVector() && EltSz == 32)
2741 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2742 unsigned &WhichResult) {
2743 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2747 unsigned NumElts = VT.getVectorNumElements();
2748 WhichResult = (M[0] == 0 ? 0 : 1);
2749 unsigned Idx = WhichResult * NumElts / 2;
2750 for (unsigned i = 0; i != NumElts; i += 2) {
2751 if ((unsigned) M[i] != Idx ||
2752 (unsigned) M[i+1] != Idx + NumElts)
2757 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2758 if (VT.is64BitVector() && EltSz == 32)
2764 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2765 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2766 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2767 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2768 unsigned &WhichResult) {
2769 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2773 unsigned NumElts = VT.getVectorNumElements();
2774 WhichResult = (M[0] == 0 ? 0 : 1);
2775 unsigned Idx = WhichResult * NumElts / 2;
2776 for (unsigned i = 0; i != NumElts; i += 2) {
2777 if ((unsigned) M[i] != Idx ||
2778 (unsigned) M[i+1] != Idx)
2783 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2784 if (VT.is64BitVector() && EltSz == 32)
2791 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2792 // Canonicalize all-zeros and all-ones vectors.
2793 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2794 if (ConstVal->isNullValue())
2795 return getZeroVector(VT, DAG, dl);
2796 if (ConstVal->isAllOnesValue())
2797 return getOnesVector(VT, DAG, dl);
2800 if (VT.is64BitVector()) {
2801 switch (Val.getValueType().getSizeInBits()) {
2802 case 8: CanonicalVT = MVT::v8i8; break;
2803 case 16: CanonicalVT = MVT::v4i16; break;
2804 case 32: CanonicalVT = MVT::v2i32; break;
2805 case 64: CanonicalVT = MVT::v1i64; break;
2806 default: llvm_unreachable("unexpected splat element type"); break;
2809 assert(VT.is128BitVector() && "unknown splat vector size");
2810 switch (Val.getValueType().getSizeInBits()) {
2811 case 8: CanonicalVT = MVT::v16i8; break;
2812 case 16: CanonicalVT = MVT::v8i16; break;
2813 case 32: CanonicalVT = MVT::v4i32; break;
2814 case 64: CanonicalVT = MVT::v2i64; break;
2815 default: llvm_unreachable("unexpected splat element type"); break;
2819 // Build a canonical splat for this value.
2820 SmallVector<SDValue, 8> Ops;
2821 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2822 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2824 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2827 // If this is a case we can't handle, return null and let the default
2828 // expansion code take care of it.
2829 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2830 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2831 DebugLoc dl = Op.getDebugLoc();
2832 EVT VT = Op.getValueType();
2834 APInt SplatBits, SplatUndef;
2835 unsigned SplatBitSize;
2837 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2838 if (SplatBitSize <= 64) {
2839 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2840 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2842 return BuildSplat(Val, VT, DAG, dl);
2846 // Scan through the operands to see if only one value is used.
2847 unsigned NumElts = VT.getVectorNumElements();
2848 bool isOnlyLowElement = true;
2849 bool usesOnlyOneValue = true;
2850 bool isConstant = true;
2852 for (unsigned i = 0; i < NumElts; ++i) {
2853 SDValue V = Op.getOperand(i);
2854 if (V.getOpcode() == ISD::UNDEF)
2857 isOnlyLowElement = false;
2858 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
2861 if (!Value.getNode())
2863 else if (V != Value)
2864 usesOnlyOneValue = false;
2867 if (!Value.getNode())
2868 return DAG.getUNDEF(VT);
2870 if (isOnlyLowElement)
2871 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
2873 // If all elements are constants, fall back to the default expansion, which
2874 // will generate a load from the constant pool.
2878 // Use VDUP for non-constant splats.
2879 if (usesOnlyOneValue)
2880 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
2882 // Vectors with 32- or 64-bit elements can be built by directly assigning
2883 // the subregisters.
2884 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
2885 if (EltSize >= 32) {
2886 // Do the expansion with floating-point types, since that is what the VFP
2887 // registers are defined to use, and since i64 is not legal.
2888 EVT EltVT = EVT::getFloatingPointVT(EltSize);
2889 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
2890 SDValue Val = DAG.getUNDEF(VecVT);
2891 for (unsigned i = 0; i < NumElts; ++i) {
2892 SDValue Elt = Op.getOperand(i);
2893 if (Elt.getOpcode() == ISD::UNDEF)
2895 Elt = DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Elt);
2896 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Val, Elt,
2897 DAG.getConstant(i, MVT::i32));
2899 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
2905 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2906 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2907 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2908 /// are assumed to be legal.
2910 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2912 if (VT.getVectorNumElements() == 4 &&
2913 (VT.is128BitVector() || VT.is64BitVector())) {
2914 unsigned PFIndexes[4];
2915 for (unsigned i = 0; i != 4; ++i) {
2919 PFIndexes[i] = M[i];
2922 // Compute the index in the perfect shuffle table.
2923 unsigned PFTableIndex =
2924 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2925 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2926 unsigned Cost = (PFEntry >> 30);
2933 unsigned Imm, WhichResult;
2935 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2936 isVREVMask(M, VT, 64) ||
2937 isVREVMask(M, VT, 32) ||
2938 isVREVMask(M, VT, 16) ||
2939 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2940 isVTRNMask(M, VT, WhichResult) ||
2941 isVUZPMask(M, VT, WhichResult) ||
2942 isVZIPMask(M, VT, WhichResult) ||
2943 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2944 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2945 isVZIP_v_undef_Mask(M, VT, WhichResult));
2948 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2949 /// the specified operations to build the shuffle.
2950 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2951 SDValue RHS, SelectionDAG &DAG,
2953 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2954 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2955 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2958 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2967 OP_VUZPL, // VUZP, left result
2968 OP_VUZPR, // VUZP, right result
2969 OP_VZIPL, // VZIP, left result
2970 OP_VZIPR, // VZIP, right result
2971 OP_VTRNL, // VTRN, left result
2972 OP_VTRNR // VTRN, right result
2975 if (OpNum == OP_COPY) {
2976 if (LHSID == (1*9+2)*9+3) return LHS;
2977 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2981 SDValue OpLHS, OpRHS;
2982 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2983 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2984 EVT VT = OpLHS.getValueType();
2987 default: llvm_unreachable("Unknown shuffle opcode!");
2989 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2994 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2995 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2999 return DAG.getNode(ARMISD::VEXT, dl, VT,
3001 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3004 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3005 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3008 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3009 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3012 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3013 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3017 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3018 SDValue V1 = Op.getOperand(0);
3019 SDValue V2 = Op.getOperand(1);
3020 DebugLoc dl = Op.getDebugLoc();
3021 EVT VT = Op.getValueType();
3022 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3023 SmallVector<int, 8> ShuffleMask;
3025 // Convert shuffles that are directly supported on NEON to target-specific
3026 // DAG nodes, instead of keeping them as shuffles and matching them again
3027 // during code selection. This is more efficient and avoids the possibility
3028 // of inconsistencies between legalization and selection.
3029 // FIXME: floating-point vectors should be canonicalized to integer vectors
3030 // of the same time so that they get CSEd properly.
3031 SVN->getMask(ShuffleMask);
3033 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3034 int Lane = SVN->getSplatIndex();
3035 // If this is undef splat, generate it via "just" vdup, if possible.
3036 if (Lane == -1) Lane = 0;
3038 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3039 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3041 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3042 DAG.getConstant(Lane, MVT::i32));
3047 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3050 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3051 DAG.getConstant(Imm, MVT::i32));
3054 if (isVREVMask(ShuffleMask, VT, 64))
3055 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3056 if (isVREVMask(ShuffleMask, VT, 32))
3057 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3058 if (isVREVMask(ShuffleMask, VT, 16))
3059 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3061 // Check for Neon shuffles that modify both input vectors in place.
3062 // If both results are used, i.e., if there are two shuffles with the same
3063 // source operands and with masks corresponding to both results of one of
3064 // these operations, DAG memoization will ensure that a single node is
3065 // used for both shuffles.
3066 unsigned WhichResult;
3067 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3068 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3069 V1, V2).getValue(WhichResult);
3070 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3071 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3072 V1, V2).getValue(WhichResult);
3073 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3074 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3075 V1, V2).getValue(WhichResult);
3077 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3078 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3079 V1, V1).getValue(WhichResult);
3080 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3081 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3082 V1, V1).getValue(WhichResult);
3083 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3084 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3085 V1, V1).getValue(WhichResult);
3087 // If the shuffle is not directly supported and it has 4 elements, use
3088 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3089 unsigned NumElts = VT.getVectorNumElements();
3091 unsigned PFIndexes[4];
3092 for (unsigned i = 0; i != 4; ++i) {
3093 if (ShuffleMask[i] < 0)
3096 PFIndexes[i] = ShuffleMask[i];
3099 // Compute the index in the perfect shuffle table.
3100 unsigned PFTableIndex =
3101 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3102 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3103 unsigned Cost = (PFEntry >> 30);
3106 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3109 // Implement shuffles with 32- or 64-bit elements as subreg copies.
3110 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3111 if (EltSize >= 32) {
3112 // Do the expansion with floating-point types, since that is what the VFP
3113 // registers are defined to use, and since i64 is not legal.
3114 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3115 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3116 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3117 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3118 SDValue Val = DAG.getUNDEF(VecVT);
3119 for (unsigned i = 0; i < NumElts; ++i) {
3120 if (ShuffleMask[i] < 0)
3122 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3123 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3124 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3126 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Val,
3127 Elt, DAG.getConstant(i, MVT::i32));
3129 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3135 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3136 EVT VT = Op.getValueType();
3137 DebugLoc dl = Op.getDebugLoc();
3138 SDValue Vec = Op.getOperand(0);
3139 SDValue Lane = Op.getOperand(1);
3140 assert(VT == MVT::i32 &&
3141 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3142 "unexpected type for custom-lowering vector extract");
3143 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3146 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3147 // The only time a CONCAT_VECTORS operation can have legal types is when
3148 // two 64-bit vectors are concatenated to a 128-bit vector.
3149 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3150 "unexpected CONCAT_VECTORS");
3151 DebugLoc dl = Op.getDebugLoc();
3152 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3153 SDValue Op0 = Op.getOperand(0);
3154 SDValue Op1 = Op.getOperand(1);
3155 if (Op0.getOpcode() != ISD::UNDEF)
3156 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3157 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3158 DAG.getIntPtrConstant(0));
3159 if (Op1.getOpcode() != ISD::UNDEF)
3160 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3161 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3162 DAG.getIntPtrConstant(1));
3163 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3166 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3167 switch (Op.getOpcode()) {
3168 default: llvm_unreachable("Don't know how to custom lower this!");
3169 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3170 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3171 case ISD::GlobalAddress:
3172 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3173 LowerGlobalAddressELF(Op, DAG);
3174 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3175 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3176 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3177 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3178 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3179 case ISD::VASTART: return LowerVASTART(Op, DAG);
3180 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3181 case ISD::SINT_TO_FP:
3182 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3183 case ISD::FP_TO_SINT:
3184 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3185 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3186 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3187 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3188 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3189 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3190 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3192 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3195 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3196 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3197 case ISD::SRL_PARTS:
3198 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3199 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3200 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3201 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3202 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3203 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3204 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3209 /// ReplaceNodeResults - Replace the results of node with an illegal result
3210 /// type with new values built out of custom code.
3211 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3212 SmallVectorImpl<SDValue>&Results,
3213 SelectionDAG &DAG) const {
3215 switch (N->getOpcode()) {
3217 llvm_unreachable("Don't know how to custom expand this!");
3219 case ISD::BIT_CONVERT:
3220 Res = ExpandBIT_CONVERT(N, DAG);
3224 Res = LowerShift(N, DAG, Subtarget);
3228 Results.push_back(Res);
3231 //===----------------------------------------------------------------------===//
3232 // ARM Scheduler Hooks
3233 //===----------------------------------------------------------------------===//
3236 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3237 MachineBasicBlock *BB,
3238 unsigned Size) const {
3239 unsigned dest = MI->getOperand(0).getReg();
3240 unsigned ptr = MI->getOperand(1).getReg();
3241 unsigned oldval = MI->getOperand(2).getReg();
3242 unsigned newval = MI->getOperand(3).getReg();
3243 unsigned scratch = BB->getParent()->getRegInfo()
3244 .createVirtualRegister(ARM::GPRRegisterClass);
3245 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3246 DebugLoc dl = MI->getDebugLoc();
3247 bool isThumb2 = Subtarget->isThumb2();
3249 unsigned ldrOpc, strOpc;
3251 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3253 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3254 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3257 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3258 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3261 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3262 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3266 MachineFunction *MF = BB->getParent();
3267 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3268 MachineFunction::iterator It = BB;
3269 ++It; // insert the new blocks after the current block
3271 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3272 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3273 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3274 MF->insert(It, loop1MBB);
3275 MF->insert(It, loop2MBB);
3276 MF->insert(It, exitMBB);
3277 exitMBB->transferSuccessors(BB);
3281 // fallthrough --> loop1MBB
3282 BB->addSuccessor(loop1MBB);
3285 // ldrex dest, [ptr]
3289 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3290 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3291 .addReg(dest).addReg(oldval));
3292 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3293 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3294 BB->addSuccessor(loop2MBB);
3295 BB->addSuccessor(exitMBB);
3298 // strex scratch, newval, [ptr]
3302 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3304 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3305 .addReg(scratch).addImm(0));
3306 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3307 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3308 BB->addSuccessor(loop1MBB);
3309 BB->addSuccessor(exitMBB);
3315 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3321 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3322 unsigned Size, unsigned BinOpcode) const {
3323 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3324 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3326 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3327 MachineFunction *MF = BB->getParent();
3328 MachineFunction::iterator It = BB;
3331 unsigned dest = MI->getOperand(0).getReg();
3332 unsigned ptr = MI->getOperand(1).getReg();
3333 unsigned incr = MI->getOperand(2).getReg();
3334 DebugLoc dl = MI->getDebugLoc();
3336 bool isThumb2 = Subtarget->isThumb2();
3337 unsigned ldrOpc, strOpc;
3339 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3341 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3342 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3345 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3346 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3349 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3350 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3354 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3355 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3356 MF->insert(It, loopMBB);
3357 MF->insert(It, exitMBB);
3358 exitMBB->transferSuccessors(BB);
3360 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3361 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3362 unsigned scratch2 = (!BinOpcode) ? incr :
3363 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3367 // fallthrough --> loopMBB
3368 BB->addSuccessor(loopMBB);
3372 // <binop> scratch2, dest, incr
3373 // strex scratch, scratch2, ptr
3376 // fallthrough --> exitMBB
3378 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3380 // operand order needs to go the other way for NAND
3381 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3382 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3383 addReg(incr).addReg(dest)).addReg(0);
3385 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3386 addReg(dest).addReg(incr)).addReg(0);
3389 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3391 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3392 .addReg(scratch).addImm(0));
3393 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3394 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3396 BB->addSuccessor(loopMBB);
3397 BB->addSuccessor(exitMBB);
3403 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3409 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3410 MachineBasicBlock *BB) const {
3411 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3412 DebugLoc dl = MI->getDebugLoc();
3413 bool isThumb2 = Subtarget->isThumb2();
3414 switch (MI->getOpcode()) {
3417 llvm_unreachable("Unexpected instr type to insert");
3419 case ARM::ATOMIC_LOAD_ADD_I8:
3420 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3421 case ARM::ATOMIC_LOAD_ADD_I16:
3422 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3423 case ARM::ATOMIC_LOAD_ADD_I32:
3424 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3426 case ARM::ATOMIC_LOAD_AND_I8:
3427 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3428 case ARM::ATOMIC_LOAD_AND_I16:
3429 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3430 case ARM::ATOMIC_LOAD_AND_I32:
3431 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3433 case ARM::ATOMIC_LOAD_OR_I8:
3434 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3435 case ARM::ATOMIC_LOAD_OR_I16:
3436 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3437 case ARM::ATOMIC_LOAD_OR_I32:
3438 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3440 case ARM::ATOMIC_LOAD_XOR_I8:
3441 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3442 case ARM::ATOMIC_LOAD_XOR_I16:
3443 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3444 case ARM::ATOMIC_LOAD_XOR_I32:
3445 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3447 case ARM::ATOMIC_LOAD_NAND_I8:
3448 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3449 case ARM::ATOMIC_LOAD_NAND_I16:
3450 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3451 case ARM::ATOMIC_LOAD_NAND_I32:
3452 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3454 case ARM::ATOMIC_LOAD_SUB_I8:
3455 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3456 case ARM::ATOMIC_LOAD_SUB_I16:
3457 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3458 case ARM::ATOMIC_LOAD_SUB_I32:
3459 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3461 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3462 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3463 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3465 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3466 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3467 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3469 case ARM::tMOVCCr_pseudo: {
3470 // To "insert" a SELECT_CC instruction, we actually have to insert the
3471 // diamond control-flow pattern. The incoming instruction knows the
3472 // destination vreg to set, the condition code register to branch on, the
3473 // true/false values to select between, and a branch opcode to use.
3474 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3475 MachineFunction::iterator It = BB;
3481 // cmpTY ccX, r1, r2
3483 // fallthrough --> copy0MBB
3484 MachineBasicBlock *thisMBB = BB;
3485 MachineFunction *F = BB->getParent();
3486 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3487 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3488 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3489 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3490 F->insert(It, copy0MBB);
3491 F->insert(It, sinkMBB);
3492 // Update machine-CFG edges by first adding all successors of the current
3493 // block to the new block which will contain the Phi node for the select.
3494 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3495 E = BB->succ_end(); I != E; ++I)
3496 sinkMBB->addSuccessor(*I);
3497 // Next, remove all successors of the current block, and add the true
3498 // and fallthrough blocks as its successors.
3499 while (!BB->succ_empty())
3500 BB->removeSuccessor(BB->succ_begin());
3501 BB->addSuccessor(copy0MBB);
3502 BB->addSuccessor(sinkMBB);
3505 // %FalseValue = ...
3506 // # fallthrough to sinkMBB
3509 // Update machine-CFG edges
3510 BB->addSuccessor(sinkMBB);
3513 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3516 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3517 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3518 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3520 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3527 case ARM::t2SUBrSPi_:
3528 case ARM::t2SUBrSPi12_:
3529 case ARM::t2SUBrSPs_: {
3530 MachineFunction *MF = BB->getParent();
3531 unsigned DstReg = MI->getOperand(0).getReg();
3532 unsigned SrcReg = MI->getOperand(1).getReg();
3533 bool DstIsDead = MI->getOperand(0).isDead();
3534 bool SrcIsKill = MI->getOperand(1).isKill();
3536 if (SrcReg != ARM::SP) {
3537 // Copy the source to SP from virtual register.
3538 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3539 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3540 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3541 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3542 .addReg(SrcReg, getKillRegState(SrcIsKill));
3546 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3547 switch (MI->getOpcode()) {
3549 llvm_unreachable("Unexpected pseudo instruction!");
3555 OpOpc = ARM::tADDspr;
3558 OpOpc = ARM::tSUBspi;
3560 case ARM::t2SUBrSPi_:
3561 OpOpc = ARM::t2SUBrSPi;
3562 NeedPred = true; NeedCC = true;
3564 case ARM::t2SUBrSPi12_:
3565 OpOpc = ARM::t2SUBrSPi12;
3568 case ARM::t2SUBrSPs_:
3569 OpOpc = ARM::t2SUBrSPs;
3570 NeedPred = true; NeedCC = true; NeedOp3 = true;
3573 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3574 if (OpOpc == ARM::tAND)
3575 AddDefaultT1CC(MIB);
3576 MIB.addReg(ARM::SP);
3577 MIB.addOperand(MI->getOperand(2));
3579 MIB.addOperand(MI->getOperand(3));
3581 AddDefaultPred(MIB);
3585 // Copy the result from SP to virtual register.
3586 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3587 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3588 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3589 BuildMI(BB, dl, TII->get(CopyOpc))
3590 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3592 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3598 //===----------------------------------------------------------------------===//
3599 // ARM Optimization Hooks
3600 //===----------------------------------------------------------------------===//
3603 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3604 TargetLowering::DAGCombinerInfo &DCI) {
3605 SelectionDAG &DAG = DCI.DAG;
3606 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3607 EVT VT = N->getValueType(0);
3608 unsigned Opc = N->getOpcode();
3609 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3610 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3611 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3612 ISD::CondCode CC = ISD::SETCC_INVALID;
3615 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3617 SDValue CCOp = Slct.getOperand(0);
3618 if (CCOp.getOpcode() == ISD::SETCC)
3619 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3622 bool DoXform = false;
3624 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3627 if (LHS.getOpcode() == ISD::Constant &&
3628 cast<ConstantSDNode>(LHS)->isNullValue()) {
3630 } else if (CC != ISD::SETCC_INVALID &&
3631 RHS.getOpcode() == ISD::Constant &&
3632 cast<ConstantSDNode>(RHS)->isNullValue()) {
3633 std::swap(LHS, RHS);
3634 SDValue Op0 = Slct.getOperand(0);
3635 EVT OpVT = isSlctCC ? Op0.getValueType() :
3636 Op0.getOperand(0).getValueType();
3637 bool isInt = OpVT.isInteger();
3638 CC = ISD::getSetCCInverse(CC, isInt);
3640 if (!TLI.isCondCodeLegal(CC, OpVT))
3641 return SDValue(); // Inverse operator isn't legal.
3648 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3650 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3651 Slct.getOperand(0), Slct.getOperand(1), CC);
3652 SDValue CCOp = Slct.getOperand(0);
3654 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3655 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3656 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3657 CCOp, OtherOp, Result);
3662 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3663 static SDValue PerformADDCombine(SDNode *N,
3664 TargetLowering::DAGCombinerInfo &DCI) {
3665 // added by evan in r37685 with no testcase.
3666 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3668 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3669 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3670 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3671 if (Result.getNode()) return Result;
3673 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3674 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3675 if (Result.getNode()) return Result;
3681 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3682 static SDValue PerformSUBCombine(SDNode *N,
3683 TargetLowering::DAGCombinerInfo &DCI) {
3684 // added by evan in r37685 with no testcase.
3685 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3687 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3688 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3689 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3690 if (Result.getNode()) return Result;
3696 static SDValue PerformMULCombine(SDNode *N,
3697 TargetLowering::DAGCombinerInfo &DCI,
3698 const ARMSubtarget *Subtarget) {
3699 SelectionDAG &DAG = DCI.DAG;
3701 if (Subtarget->isThumb1Only())
3704 if (DAG.getMachineFunction().
3705 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
3708 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
3711 EVT VT = N->getValueType(0);
3715 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
3719 uint64_t MulAmt = C->getZExtValue();
3720 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
3721 ShiftAmt = ShiftAmt & (32 - 1);
3722 SDValue V = N->getOperand(0);
3723 DebugLoc DL = N->getDebugLoc();
3726 MulAmt >>= ShiftAmt;
3727 if (isPowerOf2_32(MulAmt - 1)) {
3728 // (mul x, 2^N + 1) => (add (shl x, N), x)
3729 Res = DAG.getNode(ISD::ADD, DL, VT,
3730 V, DAG.getNode(ISD::SHL, DL, VT,
3731 V, DAG.getConstant(Log2_32(MulAmt-1),
3733 } else if (isPowerOf2_32(MulAmt + 1)) {
3734 // (mul x, 2^N - 1) => (sub (shl x, N), x)
3735 Res = DAG.getNode(ISD::SUB, DL, VT,
3736 DAG.getNode(ISD::SHL, DL, VT,
3737 V, DAG.getConstant(Log2_32(MulAmt+1),
3744 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
3745 DAG.getConstant(ShiftAmt, MVT::i32));
3747 // Do not add new nodes to DAG combiner worklist.
3748 DCI.CombineTo(N, Res, false);
3752 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3753 /// ARMISD::VMOVRRD.
3754 static SDValue PerformVMOVRRDCombine(SDNode *N,
3755 TargetLowering::DAGCombinerInfo &DCI) {
3756 // fmrrd(fmdrr x, y) -> x,y
3757 SDValue InDouble = N->getOperand(0);
3758 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3759 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3763 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3764 /// operand of a vector shift operation, where all the elements of the
3765 /// build_vector must have the same constant integer value.
3766 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3767 // Ignore bit_converts.
3768 while (Op.getOpcode() == ISD::BIT_CONVERT)
3769 Op = Op.getOperand(0);
3770 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3771 APInt SplatBits, SplatUndef;
3772 unsigned SplatBitSize;
3774 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3775 HasAnyUndefs, ElementBits) ||
3776 SplatBitSize > ElementBits)
3778 Cnt = SplatBits.getSExtValue();
3782 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3783 /// operand of a vector shift left operation. That value must be in the range:
3784 /// 0 <= Value < ElementBits for a left shift; or
3785 /// 0 <= Value <= ElementBits for a long left shift.
3786 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3787 assert(VT.isVector() && "vector shift count is not a vector type");
3788 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3789 if (! getVShiftImm(Op, ElementBits, Cnt))
3791 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3794 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3795 /// operand of a vector shift right operation. For a shift opcode, the value
3796 /// is positive, but for an intrinsic the value count must be negative. The
3797 /// absolute value must be in the range:
3798 /// 1 <= |Value| <= ElementBits for a right shift; or
3799 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3800 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3802 assert(VT.isVector() && "vector shift count is not a vector type");
3803 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3804 if (! getVShiftImm(Op, ElementBits, Cnt))
3808 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3811 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3812 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3813 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3816 // Don't do anything for most intrinsics.
3819 // Vector shifts: check for immediate versions and lower them.
3820 // Note: This is done during DAG combining instead of DAG legalizing because
3821 // the build_vectors for 64-bit vector element shift counts are generally
3822 // not legal, and it is hard to see their values after they get legalized to
3823 // loads from a constant pool.
3824 case Intrinsic::arm_neon_vshifts:
3825 case Intrinsic::arm_neon_vshiftu:
3826 case Intrinsic::arm_neon_vshiftls:
3827 case Intrinsic::arm_neon_vshiftlu:
3828 case Intrinsic::arm_neon_vshiftn:
3829 case Intrinsic::arm_neon_vrshifts:
3830 case Intrinsic::arm_neon_vrshiftu:
3831 case Intrinsic::arm_neon_vrshiftn:
3832 case Intrinsic::arm_neon_vqshifts:
3833 case Intrinsic::arm_neon_vqshiftu:
3834 case Intrinsic::arm_neon_vqshiftsu:
3835 case Intrinsic::arm_neon_vqshiftns:
3836 case Intrinsic::arm_neon_vqshiftnu:
3837 case Intrinsic::arm_neon_vqshiftnsu:
3838 case Intrinsic::arm_neon_vqrshiftns:
3839 case Intrinsic::arm_neon_vqrshiftnu:
3840 case Intrinsic::arm_neon_vqrshiftnsu: {
3841 EVT VT = N->getOperand(1).getValueType();
3843 unsigned VShiftOpc = 0;
3846 case Intrinsic::arm_neon_vshifts:
3847 case Intrinsic::arm_neon_vshiftu:
3848 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3849 VShiftOpc = ARMISD::VSHL;
3852 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3853 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3854 ARMISD::VSHRs : ARMISD::VSHRu);
3859 case Intrinsic::arm_neon_vshiftls:
3860 case Intrinsic::arm_neon_vshiftlu:
3861 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3863 llvm_unreachable("invalid shift count for vshll intrinsic");
3865 case Intrinsic::arm_neon_vrshifts:
3866 case Intrinsic::arm_neon_vrshiftu:
3867 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3871 case Intrinsic::arm_neon_vqshifts:
3872 case Intrinsic::arm_neon_vqshiftu:
3873 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3877 case Intrinsic::arm_neon_vqshiftsu:
3878 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3880 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3882 case Intrinsic::arm_neon_vshiftn:
3883 case Intrinsic::arm_neon_vrshiftn:
3884 case Intrinsic::arm_neon_vqshiftns:
3885 case Intrinsic::arm_neon_vqshiftnu:
3886 case Intrinsic::arm_neon_vqshiftnsu:
3887 case Intrinsic::arm_neon_vqrshiftns:
3888 case Intrinsic::arm_neon_vqrshiftnu:
3889 case Intrinsic::arm_neon_vqrshiftnsu:
3890 // Narrowing shifts require an immediate right shift.
3891 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3893 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3896 llvm_unreachable("unhandled vector shift");
3900 case Intrinsic::arm_neon_vshifts:
3901 case Intrinsic::arm_neon_vshiftu:
3902 // Opcode already set above.
3904 case Intrinsic::arm_neon_vshiftls:
3905 case Intrinsic::arm_neon_vshiftlu:
3906 if (Cnt == VT.getVectorElementType().getSizeInBits())
3907 VShiftOpc = ARMISD::VSHLLi;
3909 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3910 ARMISD::VSHLLs : ARMISD::VSHLLu);
3912 case Intrinsic::arm_neon_vshiftn:
3913 VShiftOpc = ARMISD::VSHRN; break;
3914 case Intrinsic::arm_neon_vrshifts:
3915 VShiftOpc = ARMISD::VRSHRs; break;
3916 case Intrinsic::arm_neon_vrshiftu:
3917 VShiftOpc = ARMISD::VRSHRu; break;
3918 case Intrinsic::arm_neon_vrshiftn:
3919 VShiftOpc = ARMISD::VRSHRN; break;
3920 case Intrinsic::arm_neon_vqshifts:
3921 VShiftOpc = ARMISD::VQSHLs; break;
3922 case Intrinsic::arm_neon_vqshiftu:
3923 VShiftOpc = ARMISD::VQSHLu; break;
3924 case Intrinsic::arm_neon_vqshiftsu:
3925 VShiftOpc = ARMISD::VQSHLsu; break;
3926 case Intrinsic::arm_neon_vqshiftns:
3927 VShiftOpc = ARMISD::VQSHRNs; break;
3928 case Intrinsic::arm_neon_vqshiftnu:
3929 VShiftOpc = ARMISD::VQSHRNu; break;
3930 case Intrinsic::arm_neon_vqshiftnsu:
3931 VShiftOpc = ARMISD::VQSHRNsu; break;
3932 case Intrinsic::arm_neon_vqrshiftns:
3933 VShiftOpc = ARMISD::VQRSHRNs; break;
3934 case Intrinsic::arm_neon_vqrshiftnu:
3935 VShiftOpc = ARMISD::VQRSHRNu; break;
3936 case Intrinsic::arm_neon_vqrshiftnsu:
3937 VShiftOpc = ARMISD::VQRSHRNsu; break;
3940 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3941 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3944 case Intrinsic::arm_neon_vshiftins: {
3945 EVT VT = N->getOperand(1).getValueType();
3947 unsigned VShiftOpc = 0;
3949 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3950 VShiftOpc = ARMISD::VSLI;
3951 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3952 VShiftOpc = ARMISD::VSRI;
3954 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3957 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3958 N->getOperand(1), N->getOperand(2),
3959 DAG.getConstant(Cnt, MVT::i32));
3962 case Intrinsic::arm_neon_vqrshifts:
3963 case Intrinsic::arm_neon_vqrshiftu:
3964 // No immediate versions of these to check for.
3971 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3972 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3973 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3974 /// vector element shift counts are generally not legal, and it is hard to see
3975 /// their values after they get legalized to loads from a constant pool.
3976 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3977 const ARMSubtarget *ST) {
3978 EVT VT = N->getValueType(0);
3980 // Nothing to be done for scalar shifts.
3981 if (! VT.isVector())
3984 assert(ST->hasNEON() && "unexpected vector shift");
3987 switch (N->getOpcode()) {
3988 default: llvm_unreachable("unexpected shift opcode");
3991 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3992 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3993 DAG.getConstant(Cnt, MVT::i32));
3998 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3999 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4000 ARMISD::VSHRs : ARMISD::VSHRu);
4001 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4002 DAG.getConstant(Cnt, MVT::i32));
4008 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4009 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4010 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4011 const ARMSubtarget *ST) {
4012 SDValue N0 = N->getOperand(0);
4014 // Check for sign- and zero-extensions of vector extract operations of 8-
4015 // and 16-bit vector elements. NEON supports these directly. They are
4016 // handled during DAG combining because type legalization will promote them
4017 // to 32-bit types and it is messy to recognize the operations after that.
4018 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4019 SDValue Vec = N0.getOperand(0);
4020 SDValue Lane = N0.getOperand(1);
4021 EVT VT = N->getValueType(0);
4022 EVT EltVT = N0.getValueType();
4023 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4025 if (VT == MVT::i32 &&
4026 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4027 TLI.isTypeLegal(Vec.getValueType())) {
4030 switch (N->getOpcode()) {
4031 default: llvm_unreachable("unexpected opcode");
4032 case ISD::SIGN_EXTEND:
4033 Opc = ARMISD::VGETLANEs;
4035 case ISD::ZERO_EXTEND:
4036 case ISD::ANY_EXTEND:
4037 Opc = ARMISD::VGETLANEu;
4040 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4047 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4048 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4049 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4050 const ARMSubtarget *ST) {
4051 // If the target supports NEON, try to use vmax/vmin instructions for f32
4052 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4053 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4054 // a NaN; only do the transformation when it matches that behavior.
4056 // For now only do this when using NEON for FP operations; if using VFP, it
4057 // is not obvious that the benefit outweighs the cost of switching to the
4059 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4060 N->getValueType(0) != MVT::f32)
4063 SDValue CondLHS = N->getOperand(0);
4064 SDValue CondRHS = N->getOperand(1);
4065 SDValue LHS = N->getOperand(2);
4066 SDValue RHS = N->getOperand(3);
4067 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4069 unsigned Opcode = 0;
4071 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4072 IsReversed = false; // x CC y ? x : y
4073 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4074 IsReversed = true ; // x CC y ? y : x
4088 // If LHS is NaN, an ordered comparison will be false and the result will
4089 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4090 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4091 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4092 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4094 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4095 // will return -0, so vmin can only be used for unsafe math or if one of
4096 // the operands is known to be nonzero.
4097 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4099 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4101 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4110 // If LHS is NaN, an ordered comparison will be false and the result will
4111 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4112 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4113 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4114 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4116 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4117 // will return +0, so vmax can only be used for unsafe math or if one of
4118 // the operands is known to be nonzero.
4119 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4121 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4123 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4129 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4132 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4133 DAGCombinerInfo &DCI) const {
4134 switch (N->getOpcode()) {
4136 case ISD::ADD: return PerformADDCombine(N, DCI);
4137 case ISD::SUB: return PerformSUBCombine(N, DCI);
4138 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4139 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4140 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4143 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4144 case ISD::SIGN_EXTEND:
4145 case ISD::ZERO_EXTEND:
4146 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4147 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4152 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4153 if (!Subtarget->hasV6Ops())
4154 // Pre-v6 does not support unaligned mem access.
4157 // v6+ may or may not support unaligned mem access depending on the system
4159 // FIXME: This is pretty conservative. Should we provide cmdline option to
4160 // control the behaviour?
4161 if (!Subtarget->isTargetDarwin())
4165 switch (VT.getSimpleVT().SimpleTy) {
4172 // FIXME: VLD1 etc with standard alignment is legal.
4176 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4181 switch (VT.getSimpleVT().SimpleTy) {
4182 default: return false;
4197 if ((V & (Scale - 1)) != 0)
4200 return V == (V & ((1LL << 5) - 1));
4203 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4204 const ARMSubtarget *Subtarget) {
4211 switch (VT.getSimpleVT().SimpleTy) {
4212 default: return false;
4217 // + imm12 or - imm8
4219 return V == (V & ((1LL << 8) - 1));
4220 return V == (V & ((1LL << 12) - 1));
4223 // Same as ARM mode. FIXME: NEON?
4224 if (!Subtarget->hasVFP2())
4229 return V == (V & ((1LL << 8) - 1));
4233 /// isLegalAddressImmediate - Return true if the integer value can be used
4234 /// as the offset of the target addressing mode for load / store of the
4236 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4237 const ARMSubtarget *Subtarget) {
4244 if (Subtarget->isThumb1Only())
4245 return isLegalT1AddressImmediate(V, VT);
4246 else if (Subtarget->isThumb2())
4247 return isLegalT2AddressImmediate(V, VT, Subtarget);
4252 switch (VT.getSimpleVT().SimpleTy) {
4253 default: return false;
4258 return V == (V & ((1LL << 12) - 1));
4261 return V == (V & ((1LL << 8) - 1));
4264 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4269 return V == (V & ((1LL << 8) - 1));
4273 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4275 int Scale = AM.Scale;
4279 switch (VT.getSimpleVT().SimpleTy) {
4280 default: return false;
4289 return Scale == 2 || Scale == 4 || Scale == 8;
4292 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4296 // Note, we allow "void" uses (basically, uses that aren't loads or
4297 // stores), because arm allows folding a scale into many arithmetic
4298 // operations. This should be made more precise and revisited later.
4300 // Allow r << imm, but the imm has to be a multiple of two.
4301 if (Scale & 1) return false;
4302 return isPowerOf2_32(Scale);
4306 /// isLegalAddressingMode - Return true if the addressing mode represented
4307 /// by AM is legal for this target, for a load/store of the specified type.
4308 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4309 const Type *Ty) const {
4310 EVT VT = getValueType(Ty, true);
4311 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4314 // Can never fold addr of global into load/store.
4319 case 0: // no scale reg, must be "r+i" or "r", or "i".
4322 if (Subtarget->isThumb1Only())
4326 // ARM doesn't support any R+R*scale+imm addr modes.
4333 if (Subtarget->isThumb2())
4334 return isLegalT2ScaledAddressingMode(AM, VT);
4336 int Scale = AM.Scale;
4337 switch (VT.getSimpleVT().SimpleTy) {
4338 default: return false;
4342 if (Scale < 0) Scale = -Scale;
4346 return isPowerOf2_32(Scale & ~1);
4350 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4355 // Note, we allow "void" uses (basically, uses that aren't loads or
4356 // stores), because arm allows folding a scale into many arithmetic
4357 // operations. This should be made more precise and revisited later.
4359 // Allow r << imm, but the imm has to be a multiple of two.
4360 if (Scale & 1) return false;
4361 return isPowerOf2_32(Scale);
4368 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4369 /// icmp immediate, that is the target has icmp instructions which can compare
4370 /// a register against the immediate without having to materialize the
4371 /// immediate into a register.
4372 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4373 if (!Subtarget->isThumb())
4374 return ARM_AM::getSOImmVal(Imm) != -1;
4375 if (Subtarget->isThumb2())
4376 return ARM_AM::getT2SOImmVal(Imm) != -1;
4377 return Imm >= 0 && Imm <= 255;
4380 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4381 bool isSEXTLoad, SDValue &Base,
4382 SDValue &Offset, bool &isInc,
4383 SelectionDAG &DAG) {
4384 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4387 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4389 Base = Ptr->getOperand(0);
4390 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4391 int RHSC = (int)RHS->getZExtValue();
4392 if (RHSC < 0 && RHSC > -256) {
4393 assert(Ptr->getOpcode() == ISD::ADD);
4395 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4399 isInc = (Ptr->getOpcode() == ISD::ADD);
4400 Offset = Ptr->getOperand(1);
4402 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4404 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4405 int RHSC = (int)RHS->getZExtValue();
4406 if (RHSC < 0 && RHSC > -0x1000) {
4407 assert(Ptr->getOpcode() == ISD::ADD);
4409 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4410 Base = Ptr->getOperand(0);
4415 if (Ptr->getOpcode() == ISD::ADD) {
4417 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4418 if (ShOpcVal != ARM_AM::no_shift) {
4419 Base = Ptr->getOperand(1);
4420 Offset = Ptr->getOperand(0);
4422 Base = Ptr->getOperand(0);
4423 Offset = Ptr->getOperand(1);
4428 isInc = (Ptr->getOpcode() == ISD::ADD);
4429 Base = Ptr->getOperand(0);
4430 Offset = Ptr->getOperand(1);
4434 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4438 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4439 bool isSEXTLoad, SDValue &Base,
4440 SDValue &Offset, bool &isInc,
4441 SelectionDAG &DAG) {
4442 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4445 Base = Ptr->getOperand(0);
4446 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4447 int RHSC = (int)RHS->getZExtValue();
4448 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4449 assert(Ptr->getOpcode() == ISD::ADD);
4451 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4453 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4454 isInc = Ptr->getOpcode() == ISD::ADD;
4455 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4463 /// getPreIndexedAddressParts - returns true by value, base pointer and
4464 /// offset pointer and addressing mode by reference if the node's address
4465 /// can be legally represented as pre-indexed load / store address.
4467 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4469 ISD::MemIndexedMode &AM,
4470 SelectionDAG &DAG) const {
4471 if (Subtarget->isThumb1Only())
4476 bool isSEXTLoad = false;
4477 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4478 Ptr = LD->getBasePtr();
4479 VT = LD->getMemoryVT();
4480 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4481 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4482 Ptr = ST->getBasePtr();
4483 VT = ST->getMemoryVT();
4488 bool isLegal = false;
4489 if (Subtarget->isThumb2())
4490 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4491 Offset, isInc, DAG);
4493 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4494 Offset, isInc, DAG);
4498 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4502 /// getPostIndexedAddressParts - returns true by value, base pointer and
4503 /// offset pointer and addressing mode by reference if this node can be
4504 /// combined with a load / store to form a post-indexed load / store.
4505 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4508 ISD::MemIndexedMode &AM,
4509 SelectionDAG &DAG) const {
4510 if (Subtarget->isThumb1Only())
4515 bool isSEXTLoad = false;
4516 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4517 VT = LD->getMemoryVT();
4518 Ptr = LD->getBasePtr();
4519 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4520 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4521 VT = ST->getMemoryVT();
4522 Ptr = ST->getBasePtr();
4527 bool isLegal = false;
4528 if (Subtarget->isThumb2())
4529 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4532 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4538 // Swap base ptr and offset to catch more post-index load / store when
4539 // it's legal. In Thumb2 mode, offset must be an immediate.
4540 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4541 !Subtarget->isThumb2())
4542 std::swap(Base, Offset);
4544 // Post-indexed load / store update the base pointer.
4549 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4553 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4557 const SelectionDAG &DAG,
4558 unsigned Depth) const {
4559 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4560 switch (Op.getOpcode()) {
4562 case ARMISD::CMOV: {
4563 // Bits are known zero/one if known on the LHS and RHS.
4564 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4565 if (KnownZero == 0 && KnownOne == 0) return;
4567 APInt KnownZeroRHS, KnownOneRHS;
4568 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4569 KnownZeroRHS, KnownOneRHS, Depth+1);
4570 KnownZero &= KnownZeroRHS;
4571 KnownOne &= KnownOneRHS;
4577 //===----------------------------------------------------------------------===//
4578 // ARM Inline Assembly Support
4579 //===----------------------------------------------------------------------===//
4581 /// getConstraintType - Given a constraint letter, return the type of
4582 /// constraint it is for this target.
4583 ARMTargetLowering::ConstraintType
4584 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4585 if (Constraint.size() == 1) {
4586 switch (Constraint[0]) {
4588 case 'l': return C_RegisterClass;
4589 case 'w': return C_RegisterClass;
4592 return TargetLowering::getConstraintType(Constraint);
4595 std::pair<unsigned, const TargetRegisterClass*>
4596 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4598 if (Constraint.size() == 1) {
4599 // GCC ARM Constraint Letters
4600 switch (Constraint[0]) {
4602 if (Subtarget->isThumb())
4603 return std::make_pair(0U, ARM::tGPRRegisterClass);
4605 return std::make_pair(0U, ARM::GPRRegisterClass);
4607 return std::make_pair(0U, ARM::GPRRegisterClass);
4610 return std::make_pair(0U, ARM::SPRRegisterClass);
4611 if (VT.getSizeInBits() == 64)
4612 return std::make_pair(0U, ARM::DPRRegisterClass);
4613 if (VT.getSizeInBits() == 128)
4614 return std::make_pair(0U, ARM::QPRRegisterClass);
4618 if (StringRef("{cc}").equals_lower(Constraint))
4619 return std::make_pair(0U, ARM::CCRRegisterClass);
4621 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4624 std::vector<unsigned> ARMTargetLowering::
4625 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4627 if (Constraint.size() != 1)
4628 return std::vector<unsigned>();
4630 switch (Constraint[0]) { // GCC ARM Constraint Letters
4633 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4634 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4637 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4638 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4639 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4640 ARM::R12, ARM::LR, 0);
4643 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4644 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4645 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4646 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4647 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4648 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4649 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4650 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4651 if (VT.getSizeInBits() == 64)
4652 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4653 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4654 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4655 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4656 if (VT.getSizeInBits() == 128)
4657 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4658 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4662 return std::vector<unsigned>();
4665 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4666 /// vector. If it is invalid, don't add anything to Ops.
4667 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4670 std::vector<SDValue>&Ops,
4671 SelectionDAG &DAG) const {
4672 SDValue Result(0, 0);
4674 switch (Constraint) {
4676 case 'I': case 'J': case 'K': case 'L':
4677 case 'M': case 'N': case 'O':
4678 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4682 int64_t CVal64 = C->getSExtValue();
4683 int CVal = (int) CVal64;
4684 // None of these constraints allow values larger than 32 bits. Check
4685 // that the value fits in an int.
4689 switch (Constraint) {
4691 if (Subtarget->isThumb1Only()) {
4692 // This must be a constant between 0 and 255, for ADD
4694 if (CVal >= 0 && CVal <= 255)
4696 } else if (Subtarget->isThumb2()) {
4697 // A constant that can be used as an immediate value in a
4698 // data-processing instruction.
4699 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4702 // A constant that can be used as an immediate value in a
4703 // data-processing instruction.
4704 if (ARM_AM::getSOImmVal(CVal) != -1)
4710 if (Subtarget->isThumb()) { // FIXME thumb2
4711 // This must be a constant between -255 and -1, for negated ADD
4712 // immediates. This can be used in GCC with an "n" modifier that
4713 // prints the negated value, for use with SUB instructions. It is
4714 // not useful otherwise but is implemented for compatibility.
4715 if (CVal >= -255 && CVal <= -1)
4718 // This must be a constant between -4095 and 4095. It is not clear
4719 // what this constraint is intended for. Implemented for
4720 // compatibility with GCC.
4721 if (CVal >= -4095 && CVal <= 4095)
4727 if (Subtarget->isThumb1Only()) {
4728 // A 32-bit value where only one byte has a nonzero value. Exclude
4729 // zero to match GCC. This constraint is used by GCC internally for
4730 // constants that can be loaded with a move/shift combination.
4731 // It is not useful otherwise but is implemented for compatibility.
4732 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4734 } else if (Subtarget->isThumb2()) {
4735 // A constant whose bitwise inverse can be used as an immediate
4736 // value in a data-processing instruction. This can be used in GCC
4737 // with a "B" modifier that prints the inverted value, for use with
4738 // BIC and MVN instructions. It is not useful otherwise but is
4739 // implemented for compatibility.
4740 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4743 // A constant whose bitwise inverse can be used as an immediate
4744 // value in a data-processing instruction. This can be used in GCC
4745 // with a "B" modifier that prints the inverted value, for use with
4746 // BIC and MVN instructions. It is not useful otherwise but is
4747 // implemented for compatibility.
4748 if (ARM_AM::getSOImmVal(~CVal) != -1)
4754 if (Subtarget->isThumb1Only()) {
4755 // This must be a constant between -7 and 7,
4756 // for 3-operand ADD/SUB immediate instructions.
4757 if (CVal >= -7 && CVal < 7)
4759 } else if (Subtarget->isThumb2()) {
4760 // A constant whose negation can be used as an immediate value in a
4761 // data-processing instruction. This can be used in GCC with an "n"
4762 // modifier that prints the negated value, for use with SUB
4763 // instructions. It is not useful otherwise but is implemented for
4765 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4768 // A constant whose negation can be used as an immediate value in a
4769 // data-processing instruction. This can be used in GCC with an "n"
4770 // modifier that prints the negated value, for use with SUB
4771 // instructions. It is not useful otherwise but is implemented for
4773 if (ARM_AM::getSOImmVal(-CVal) != -1)
4779 if (Subtarget->isThumb()) { // FIXME thumb2
4780 // This must be a multiple of 4 between 0 and 1020, for
4781 // ADD sp + immediate.
4782 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4785 // A power of two or a constant between 0 and 32. This is used in
4786 // GCC for the shift amount on shifted register operands, but it is
4787 // useful in general for any shift amounts.
4788 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4794 if (Subtarget->isThumb()) { // FIXME thumb2
4795 // This must be a constant between 0 and 31, for shift amounts.
4796 if (CVal >= 0 && CVal <= 31)
4802 if (Subtarget->isThumb()) { // FIXME thumb2
4803 // This must be a multiple of 4 between -508 and 508, for
4804 // ADD/SUB sp = sp + immediate.
4805 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4810 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4814 if (Result.getNode()) {
4815 Ops.push_back(Result);
4818 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4823 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4824 // The ARM target isn't yet aware of offsets.
4828 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4829 APInt Imm = FPImm.bitcastToAPInt();
4830 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4831 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4832 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4834 // We can handle 4 bits of mantissa.
4835 // mantissa = (16+UInt(e:f:g:h))/16.
4836 if (Mantissa & 0x7ffff)
4839 if ((Mantissa & 0xf) != Mantissa)
4842 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4843 if (Exp < -3 || Exp > 4)
4845 Exp = ((Exp+3) & 0x7) ^ 4;
4847 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4850 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4851 APInt Imm = FPImm.bitcastToAPInt();
4852 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4853 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4854 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4856 // We can handle 4 bits of mantissa.
4857 // mantissa = (16+UInt(e:f:g:h))/16.
4858 if (Mantissa & 0xffffffffffffLL)
4861 if ((Mantissa & 0xf) != Mantissa)
4864 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4865 if (Exp < -3 || Exp > 4)
4867 Exp = ((Exp+3) & 0x7) ^ 4;
4869 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4872 /// isFPImmLegal - Returns true if the target can instruction select the
4873 /// specified FP immediate natively. If false, the legalizer will
4874 /// materialize the FP immediate as a load from a constant pool.
4875 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4876 if (!Subtarget->hasVFP3())
4879 return ARM::getVFPf32Imm(Imm) != -1;
4881 return ARM::getVFPf64Imm(Imm) != -1;