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);
416 setOperationAction(ISD::SETCC, MVT::i32, Expand);
417 setOperationAction(ISD::SETCC, MVT::f32, Expand);
418 setOperationAction(ISD::SETCC, MVT::f64, Expand);
419 setOperationAction(ISD::SELECT, MVT::i32, Expand);
420 setOperationAction(ISD::SELECT, MVT::f32, Expand);
421 setOperationAction(ISD::SELECT, MVT::f64, Expand);
422 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
423 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
424 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
426 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
427 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
428 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
429 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
430 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
432 // We don't support sin/cos/fmod/copysign/pow
433 setOperationAction(ISD::FSIN, MVT::f64, Expand);
434 setOperationAction(ISD::FSIN, MVT::f32, Expand);
435 setOperationAction(ISD::FCOS, MVT::f32, Expand);
436 setOperationAction(ISD::FCOS, MVT::f64, Expand);
437 setOperationAction(ISD::FREM, MVT::f64, Expand);
438 setOperationAction(ISD::FREM, MVT::f32, Expand);
439 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
440 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
441 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
443 setOperationAction(ISD::FPOW, MVT::f64, Expand);
444 setOperationAction(ISD::FPOW, MVT::f32, Expand);
446 // Various VFP goodness
447 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
448 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
449 if (Subtarget->hasVFP2()) {
450 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
451 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
452 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
453 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
455 // Special handling for half-precision FP.
456 if (!Subtarget->hasFP16()) {
457 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
458 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
462 // We have target-specific dag combine patterns for the following nodes:
463 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
464 setTargetDAGCombine(ISD::ADD);
465 setTargetDAGCombine(ISD::SUB);
467 setStackPointerRegisterToSaveRestore(ARM::SP);
468 setSchedulingPreference(SchedulingForRegPressure);
470 // FIXME: If-converter should use instruction latency to determine
471 // profitability rather than relying on fixed limits.
472 if (Subtarget->getCPUString() == "generic") {
473 // Generic (and overly aggressive) if-conversion limits.
474 setIfCvtBlockSizeLimit(10);
475 setIfCvtDupBlockSizeLimit(2);
476 } else if (Subtarget->hasV7Ops()) {
477 setIfCvtBlockSizeLimit(3);
478 setIfCvtDupBlockSizeLimit(1);
479 } else if (Subtarget->hasV6Ops()) {
480 setIfCvtBlockSizeLimit(2);
481 setIfCvtDupBlockSizeLimit(1);
483 setIfCvtBlockSizeLimit(3);
484 setIfCvtDupBlockSizeLimit(2);
487 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
488 // Do not enable CodePlacementOpt for now: it currently runs after the
489 // ARMConstantIslandPass and messes up branch relaxation and placement
490 // of constant islands.
491 // benefitFromCodePlacementOpt = true;
494 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
497 case ARMISD::Wrapper: return "ARMISD::Wrapper";
498 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
499 case ARMISD::CALL: return "ARMISD::CALL";
500 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
501 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
502 case ARMISD::tCALL: return "ARMISD::tCALL";
503 case ARMISD::BRCOND: return "ARMISD::BRCOND";
504 case ARMISD::BR_JT: return "ARMISD::BR_JT";
505 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
506 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
507 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
508 case ARMISD::CMP: return "ARMISD::CMP";
509 case ARMISD::CMPZ: return "ARMISD::CMPZ";
510 case ARMISD::CMPFP: return "ARMISD::CMPFP";
511 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
512 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
513 case ARMISD::CMOV: return "ARMISD::CMOV";
514 case ARMISD::CNEG: return "ARMISD::CNEG";
516 case ARMISD::RBIT: return "ARMISD::RBIT";
518 case ARMISD::FTOSI: return "ARMISD::FTOSI";
519 case ARMISD::FTOUI: return "ARMISD::FTOUI";
520 case ARMISD::SITOF: return "ARMISD::SITOF";
521 case ARMISD::UITOF: return "ARMISD::UITOF";
523 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
524 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
525 case ARMISD::RRX: return "ARMISD::RRX";
527 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
528 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
530 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
531 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
533 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
535 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
537 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
538 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
540 case ARMISD::VCEQ: return "ARMISD::VCEQ";
541 case ARMISD::VCGE: return "ARMISD::VCGE";
542 case ARMISD::VCGEU: return "ARMISD::VCGEU";
543 case ARMISD::VCGT: return "ARMISD::VCGT";
544 case ARMISD::VCGTU: return "ARMISD::VCGTU";
545 case ARMISD::VTST: return "ARMISD::VTST";
547 case ARMISD::VSHL: return "ARMISD::VSHL";
548 case ARMISD::VSHRs: return "ARMISD::VSHRs";
549 case ARMISD::VSHRu: return "ARMISD::VSHRu";
550 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
551 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
552 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
553 case ARMISD::VSHRN: return "ARMISD::VSHRN";
554 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
555 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
556 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
557 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
558 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
559 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
560 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
561 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
562 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
563 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
564 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
565 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
566 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
567 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
568 case ARMISD::VDUP: return "ARMISD::VDUP";
569 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
570 case ARMISD::VEXT: return "ARMISD::VEXT";
571 case ARMISD::VREV64: return "ARMISD::VREV64";
572 case ARMISD::VREV32: return "ARMISD::VREV32";
573 case ARMISD::VREV16: return "ARMISD::VREV16";
574 case ARMISD::VZIP: return "ARMISD::VZIP";
575 case ARMISD::VUZP: return "ARMISD::VUZP";
576 case ARMISD::VTRN: return "ARMISD::VTRN";
577 case ARMISD::FMAX: return "ARMISD::FMAX";
578 case ARMISD::FMIN: return "ARMISD::FMIN";
582 /// getRegClassFor - Return the register class that should be used for the
583 /// specified value type.
584 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
585 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
586 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
587 // load / store 4 to 8 consecutive D registers.
588 if (VT == MVT::v4i64)
589 return ARM::QQPRRegisterClass;
590 else if (VT == MVT::v8i64)
591 return ARM::QQQQPRRegisterClass;
592 return TargetLowering::getRegClassFor(VT);
595 /// getFunctionAlignment - Return the Log2 alignment of this function.
596 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
597 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
600 //===----------------------------------------------------------------------===//
602 //===----------------------------------------------------------------------===//
604 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
605 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
607 default: llvm_unreachable("Unknown condition code!");
608 case ISD::SETNE: return ARMCC::NE;
609 case ISD::SETEQ: return ARMCC::EQ;
610 case ISD::SETGT: return ARMCC::GT;
611 case ISD::SETGE: return ARMCC::GE;
612 case ISD::SETLT: return ARMCC::LT;
613 case ISD::SETLE: return ARMCC::LE;
614 case ISD::SETUGT: return ARMCC::HI;
615 case ISD::SETUGE: return ARMCC::HS;
616 case ISD::SETULT: return ARMCC::LO;
617 case ISD::SETULE: return ARMCC::LS;
621 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
622 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
623 ARMCC::CondCodes &CondCode2) {
624 CondCode2 = ARMCC::AL;
626 default: llvm_unreachable("Unknown FP condition!");
628 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
630 case ISD::SETOGT: CondCode = ARMCC::GT; break;
632 case ISD::SETOGE: CondCode = ARMCC::GE; break;
633 case ISD::SETOLT: CondCode = ARMCC::MI; break;
634 case ISD::SETOLE: CondCode = ARMCC::LS; break;
635 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
636 case ISD::SETO: CondCode = ARMCC::VC; break;
637 case ISD::SETUO: CondCode = ARMCC::VS; break;
638 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
639 case ISD::SETUGT: CondCode = ARMCC::HI; break;
640 case ISD::SETUGE: CondCode = ARMCC::PL; break;
642 case ISD::SETULT: CondCode = ARMCC::LT; break;
644 case ISD::SETULE: CondCode = ARMCC::LE; break;
646 case ISD::SETUNE: CondCode = ARMCC::NE; break;
650 //===----------------------------------------------------------------------===//
651 // Calling Convention Implementation
652 //===----------------------------------------------------------------------===//
654 #include "ARMGenCallingConv.inc"
656 // APCS f64 is in register pairs, possibly split to stack
657 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
658 CCValAssign::LocInfo &LocInfo,
659 CCState &State, bool CanFail) {
660 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
662 // Try to get the first register.
663 if (unsigned Reg = State.AllocateReg(RegList, 4))
664 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
666 // For the 2nd half of a v2f64, do not fail.
670 // Put the whole thing on the stack.
671 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
672 State.AllocateStack(8, 4),
677 // Try to get the second register.
678 if (unsigned Reg = State.AllocateReg(RegList, 4))
679 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
681 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
682 State.AllocateStack(4, 4),
687 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
688 CCValAssign::LocInfo &LocInfo,
689 ISD::ArgFlagsTy &ArgFlags,
691 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
693 if (LocVT == MVT::v2f64 &&
694 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
696 return true; // we handled it
699 // AAPCS f64 is in aligned register pairs
700 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
701 CCValAssign::LocInfo &LocInfo,
702 CCState &State, bool CanFail) {
703 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
704 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
706 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
708 // For the 2nd half of a v2f64, do not just fail.
712 // Put the whole thing on the stack.
713 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
714 State.AllocateStack(8, 8),
720 for (i = 0; i < 2; ++i)
721 if (HiRegList[i] == Reg)
724 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
725 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
730 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
731 CCValAssign::LocInfo &LocInfo,
732 ISD::ArgFlagsTy &ArgFlags,
734 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
736 if (LocVT == MVT::v2f64 &&
737 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
739 return true; // we handled it
742 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
743 CCValAssign::LocInfo &LocInfo, CCState &State) {
744 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
745 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
747 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
749 return false; // we didn't handle it
752 for (i = 0; i < 2; ++i)
753 if (HiRegList[i] == Reg)
756 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
757 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
762 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
763 CCValAssign::LocInfo &LocInfo,
764 ISD::ArgFlagsTy &ArgFlags,
766 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
768 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
770 return true; // we handled it
773 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
774 CCValAssign::LocInfo &LocInfo,
775 ISD::ArgFlagsTy &ArgFlags,
777 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
781 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
782 /// given CallingConvention value.
783 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
785 bool isVarArg) const {
788 llvm_unreachable("Unsupported calling convention");
790 case CallingConv::Fast:
791 // Use target triple & subtarget features to do actual dispatch.
792 if (Subtarget->isAAPCS_ABI()) {
793 if (Subtarget->hasVFP2() &&
794 FloatABIType == FloatABI::Hard && !isVarArg)
795 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
797 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
799 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
800 case CallingConv::ARM_AAPCS_VFP:
801 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
802 case CallingConv::ARM_AAPCS:
803 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
804 case CallingConv::ARM_APCS:
805 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
809 /// LowerCallResult - Lower the result values of a call into the
810 /// appropriate copies out of appropriate physical registers.
812 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
813 CallingConv::ID CallConv, bool isVarArg,
814 const SmallVectorImpl<ISD::InputArg> &Ins,
815 DebugLoc dl, SelectionDAG &DAG,
816 SmallVectorImpl<SDValue> &InVals) const {
818 // Assign locations to each value returned by this call.
819 SmallVector<CCValAssign, 16> RVLocs;
820 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
821 RVLocs, *DAG.getContext());
822 CCInfo.AnalyzeCallResult(Ins,
823 CCAssignFnForNode(CallConv, /* Return*/ true,
826 // Copy all of the result registers out of their specified physreg.
827 for (unsigned i = 0; i != RVLocs.size(); ++i) {
828 CCValAssign VA = RVLocs[i];
831 if (VA.needsCustom()) {
832 // Handle f64 or half of a v2f64.
833 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
835 Chain = Lo.getValue(1);
836 InFlag = Lo.getValue(2);
837 VA = RVLocs[++i]; // skip ahead to next loc
838 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
840 Chain = Hi.getValue(1);
841 InFlag = Hi.getValue(2);
842 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
844 if (VA.getLocVT() == MVT::v2f64) {
845 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
846 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
847 DAG.getConstant(0, MVT::i32));
849 VA = RVLocs[++i]; // skip ahead to next loc
850 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
851 Chain = Lo.getValue(1);
852 InFlag = Lo.getValue(2);
853 VA = RVLocs[++i]; // skip ahead to next loc
854 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
855 Chain = Hi.getValue(1);
856 InFlag = Hi.getValue(2);
857 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
858 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
859 DAG.getConstant(1, MVT::i32));
862 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
864 Chain = Val.getValue(1);
865 InFlag = Val.getValue(2);
868 switch (VA.getLocInfo()) {
869 default: llvm_unreachable("Unknown loc info!");
870 case CCValAssign::Full: break;
871 case CCValAssign::BCvt:
872 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
876 InVals.push_back(Val);
882 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
883 /// by "Src" to address "Dst" of size "Size". Alignment information is
884 /// specified by the specific parameter attribute. The copy will be passed as
885 /// a byval function parameter.
886 /// Sometimes what we are copying is the end of a larger object, the part that
887 /// does not fit in registers.
889 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
890 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
892 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
893 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
894 /*isVolatile=*/false, /*AlwaysInline=*/false,
898 /// LowerMemOpCallTo - Store the argument to the stack.
900 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
901 SDValue StackPtr, SDValue Arg,
902 DebugLoc dl, SelectionDAG &DAG,
903 const CCValAssign &VA,
904 ISD::ArgFlagsTy Flags) const {
905 unsigned LocMemOffset = VA.getLocMemOffset();
906 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
907 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
908 if (Flags.isByVal()) {
909 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
911 return DAG.getStore(Chain, dl, Arg, PtrOff,
912 PseudoSourceValue::getStack(), LocMemOffset,
916 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
917 SDValue Chain, SDValue &Arg,
918 RegsToPassVector &RegsToPass,
919 CCValAssign &VA, CCValAssign &NextVA,
921 SmallVector<SDValue, 8> &MemOpChains,
922 ISD::ArgFlagsTy Flags) const {
924 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
925 DAG.getVTList(MVT::i32, MVT::i32), Arg);
926 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
928 if (NextVA.isRegLoc())
929 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
931 assert(NextVA.isMemLoc());
932 if (StackPtr.getNode() == 0)
933 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
935 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
941 /// LowerCall - Lowering a call into a callseq_start <-
942 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
945 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
946 CallingConv::ID CallConv, bool isVarArg,
948 const SmallVectorImpl<ISD::OutputArg> &Outs,
949 const SmallVectorImpl<ISD::InputArg> &Ins,
950 DebugLoc dl, SelectionDAG &DAG,
951 SmallVectorImpl<SDValue> &InVals) const {
952 // ARM target does not yet support tail call optimization.
955 // Analyze operands of the call, assigning locations to each operand.
956 SmallVector<CCValAssign, 16> ArgLocs;
957 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
959 CCInfo.AnalyzeCallOperands(Outs,
960 CCAssignFnForNode(CallConv, /* Return*/ false,
963 // Get a count of how many bytes are to be pushed on the stack.
964 unsigned NumBytes = CCInfo.getNextStackOffset();
966 // Adjust the stack pointer for the new arguments...
967 // These operations are automatically eliminated by the prolog/epilog pass
968 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
970 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
972 RegsToPassVector RegsToPass;
973 SmallVector<SDValue, 8> MemOpChains;
975 // Walk the register/memloc assignments, inserting copies/loads. In the case
976 // of tail call optimization, arguments are handled later.
977 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
980 CCValAssign &VA = ArgLocs[i];
981 SDValue Arg = Outs[realArgIdx].Val;
982 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
984 // Promote the value if needed.
985 switch (VA.getLocInfo()) {
986 default: llvm_unreachable("Unknown loc info!");
987 case CCValAssign::Full: break;
988 case CCValAssign::SExt:
989 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
991 case CCValAssign::ZExt:
992 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
994 case CCValAssign::AExt:
995 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
997 case CCValAssign::BCvt:
998 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1002 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1003 if (VA.needsCustom()) {
1004 if (VA.getLocVT() == MVT::v2f64) {
1005 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1006 DAG.getConstant(0, MVT::i32));
1007 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1008 DAG.getConstant(1, MVT::i32));
1010 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1011 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1013 VA = ArgLocs[++i]; // skip ahead to next loc
1014 if (VA.isRegLoc()) {
1015 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1016 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1018 assert(VA.isMemLoc());
1020 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1021 dl, DAG, VA, Flags));
1024 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1025 StackPtr, MemOpChains, Flags);
1027 } else if (VA.isRegLoc()) {
1028 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1030 assert(VA.isMemLoc());
1032 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1033 dl, DAG, VA, Flags));
1037 if (!MemOpChains.empty())
1038 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1039 &MemOpChains[0], MemOpChains.size());
1041 // Build a sequence of copy-to-reg nodes chained together with token chain
1042 // and flag operands which copy the outgoing args into the appropriate regs.
1044 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1045 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1046 RegsToPass[i].second, InFlag);
1047 InFlag = Chain.getValue(1);
1050 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1051 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1052 // node so that legalize doesn't hack it.
1053 bool isDirect = false;
1054 bool isARMFunc = false;
1055 bool isLocalARMFunc = false;
1056 MachineFunction &MF = DAG.getMachineFunction();
1057 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1059 if (EnableARMLongCalls) {
1060 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1061 && "long-calls with non-static relocation model!");
1062 // Handle a global address or an external symbol. If it's not one of
1063 // those, the target's already in a register, so we don't need to do
1065 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1066 const GlobalValue *GV = G->getGlobal();
1067 // Create a constant pool entry for the callee address
1068 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1069 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1072 // Get the address of the callee into a register
1073 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1074 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1075 Callee = DAG.getLoad(getPointerTy(), dl,
1076 DAG.getEntryNode(), CPAddr,
1077 PseudoSourceValue::getConstantPool(), 0,
1079 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1080 const char *Sym = S->getSymbol();
1082 // Create a constant pool entry for the callee address
1083 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1084 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1085 Sym, ARMPCLabelIndex, 0);
1086 // Get the address of the callee into a register
1087 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1088 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1089 Callee = DAG.getLoad(getPointerTy(), dl,
1090 DAG.getEntryNode(), CPAddr,
1091 PseudoSourceValue::getConstantPool(), 0,
1094 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1095 const GlobalValue *GV = G->getGlobal();
1097 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1098 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1099 getTargetMachine().getRelocationModel() != Reloc::Static;
1100 isARMFunc = !Subtarget->isThumb() || isStub;
1101 // ARM call to a local ARM function is predicable.
1102 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1103 // tBX takes a register source operand.
1104 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1105 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1106 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1109 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1110 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1111 Callee = DAG.getLoad(getPointerTy(), dl,
1112 DAG.getEntryNode(), CPAddr,
1113 PseudoSourceValue::getConstantPool(), 0,
1115 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1116 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1117 getPointerTy(), Callee, PICLabel);
1119 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1120 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1122 bool isStub = Subtarget->isTargetDarwin() &&
1123 getTargetMachine().getRelocationModel() != Reloc::Static;
1124 isARMFunc = !Subtarget->isThumb() || isStub;
1125 // tBX takes a register source operand.
1126 const char *Sym = S->getSymbol();
1127 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1128 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1129 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1130 Sym, ARMPCLabelIndex, 4);
1131 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1132 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1133 Callee = DAG.getLoad(getPointerTy(), dl,
1134 DAG.getEntryNode(), CPAddr,
1135 PseudoSourceValue::getConstantPool(), 0,
1137 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1138 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1139 getPointerTy(), Callee, PICLabel);
1141 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1144 // FIXME: handle tail calls differently.
1146 if (Subtarget->isThumb()) {
1147 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1148 CallOpc = ARMISD::CALL_NOLINK;
1150 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1152 CallOpc = (isDirect || Subtarget->hasV5TOps())
1153 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1154 : ARMISD::CALL_NOLINK;
1156 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1157 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1158 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1159 InFlag = Chain.getValue(1);
1162 std::vector<SDValue> Ops;
1163 Ops.push_back(Chain);
1164 Ops.push_back(Callee);
1166 // Add argument registers to the end of the list so that they are known live
1168 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1169 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1170 RegsToPass[i].second.getValueType()));
1172 if (InFlag.getNode())
1173 Ops.push_back(InFlag);
1174 // Returns a chain and a flag for retval copy to use.
1175 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1176 &Ops[0], Ops.size());
1177 InFlag = Chain.getValue(1);
1179 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1180 DAG.getIntPtrConstant(0, true), InFlag);
1182 InFlag = Chain.getValue(1);
1184 // Handle result values, copying them out of physregs into vregs that we
1186 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1191 ARMTargetLowering::LowerReturn(SDValue Chain,
1192 CallingConv::ID CallConv, bool isVarArg,
1193 const SmallVectorImpl<ISD::OutputArg> &Outs,
1194 DebugLoc dl, SelectionDAG &DAG) const {
1196 // CCValAssign - represent the assignment of the return value to a location.
1197 SmallVector<CCValAssign, 16> RVLocs;
1199 // CCState - Info about the registers and stack slots.
1200 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1203 // Analyze outgoing return values.
1204 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1207 // If this is the first return lowered for this function, add
1208 // the regs to the liveout set for the function.
1209 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1210 for (unsigned i = 0; i != RVLocs.size(); ++i)
1211 if (RVLocs[i].isRegLoc())
1212 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1217 // Copy the result values into the output registers.
1218 for (unsigned i = 0, realRVLocIdx = 0;
1220 ++i, ++realRVLocIdx) {
1221 CCValAssign &VA = RVLocs[i];
1222 assert(VA.isRegLoc() && "Can only return in registers!");
1224 SDValue Arg = Outs[realRVLocIdx].Val;
1226 switch (VA.getLocInfo()) {
1227 default: llvm_unreachable("Unknown loc info!");
1228 case CCValAssign::Full: break;
1229 case CCValAssign::BCvt:
1230 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1234 if (VA.needsCustom()) {
1235 if (VA.getLocVT() == MVT::v2f64) {
1236 // Extract the first half and return it in two registers.
1237 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1238 DAG.getConstant(0, MVT::i32));
1239 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1240 DAG.getVTList(MVT::i32, MVT::i32), Half);
1242 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1243 Flag = Chain.getValue(1);
1244 VA = RVLocs[++i]; // skip ahead to next loc
1245 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1246 HalfGPRs.getValue(1), Flag);
1247 Flag = Chain.getValue(1);
1248 VA = RVLocs[++i]; // skip ahead to next loc
1250 // Extract the 2nd half and fall through to handle it as an f64 value.
1251 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1252 DAG.getConstant(1, MVT::i32));
1254 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1256 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1257 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1258 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1259 Flag = Chain.getValue(1);
1260 VA = RVLocs[++i]; // skip ahead to next loc
1261 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1264 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1266 // Guarantee that all emitted copies are
1267 // stuck together, avoiding something bad.
1268 Flag = Chain.getValue(1);
1273 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1275 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1280 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1281 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1282 // one of the above mentioned nodes. It has to be wrapped because otherwise
1283 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1284 // be used to form addressing mode. These wrapped nodes will be selected
1286 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1287 EVT PtrVT = Op.getValueType();
1288 // FIXME there is no actual debug info here
1289 DebugLoc dl = Op.getDebugLoc();
1290 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1292 if (CP->isMachineConstantPoolEntry())
1293 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1294 CP->getAlignment());
1296 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1297 CP->getAlignment());
1298 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1301 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1302 SelectionDAG &DAG) const {
1303 MachineFunction &MF = DAG.getMachineFunction();
1304 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1305 unsigned ARMPCLabelIndex = 0;
1306 DebugLoc DL = Op.getDebugLoc();
1307 EVT PtrVT = getPointerTy();
1308 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1309 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1311 if (RelocM == Reloc::Static) {
1312 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1314 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1315 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1316 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1317 ARMCP::CPBlockAddress,
1319 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1321 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1322 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1323 PseudoSourceValue::getConstantPool(), 0,
1325 if (RelocM == Reloc::Static)
1327 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1328 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1331 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1333 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1334 SelectionDAG &DAG) const {
1335 DebugLoc dl = GA->getDebugLoc();
1336 EVT PtrVT = getPointerTy();
1337 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1338 MachineFunction &MF = DAG.getMachineFunction();
1339 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1340 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1341 ARMConstantPoolValue *CPV =
1342 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1343 ARMCP::CPValue, PCAdj, "tlsgd", true);
1344 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1345 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1346 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1347 PseudoSourceValue::getConstantPool(), 0,
1349 SDValue Chain = Argument.getValue(1);
1351 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1352 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1354 // call __tls_get_addr.
1357 Entry.Node = Argument;
1358 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1359 Args.push_back(Entry);
1360 // FIXME: is there useful debug info available here?
1361 std::pair<SDValue, SDValue> CallResult =
1362 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1363 false, false, false, false,
1364 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1365 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1366 return CallResult.first;
1369 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1370 // "local exec" model.
1372 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1373 SelectionDAG &DAG) const {
1374 const GlobalValue *GV = GA->getGlobal();
1375 DebugLoc dl = GA->getDebugLoc();
1377 SDValue Chain = DAG.getEntryNode();
1378 EVT PtrVT = getPointerTy();
1379 // Get the Thread Pointer
1380 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1382 if (GV->isDeclaration()) {
1383 MachineFunction &MF = DAG.getMachineFunction();
1384 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1385 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1386 // Initial exec model.
1387 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1388 ARMConstantPoolValue *CPV =
1389 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1390 ARMCP::CPValue, PCAdj, "gottpoff", true);
1391 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1392 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1393 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1394 PseudoSourceValue::getConstantPool(), 0,
1396 Chain = Offset.getValue(1);
1398 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1399 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1401 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1402 PseudoSourceValue::getConstantPool(), 0,
1406 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1407 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1408 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1409 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1410 PseudoSourceValue::getConstantPool(), 0,
1414 // The address of the thread local variable is the add of the thread
1415 // pointer with the offset of the variable.
1416 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1420 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1421 // TODO: implement the "local dynamic" model
1422 assert(Subtarget->isTargetELF() &&
1423 "TLS not implemented for non-ELF targets");
1424 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1425 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1426 // otherwise use the "Local Exec" TLS Model
1427 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1428 return LowerToTLSGeneralDynamicModel(GA, DAG);
1430 return LowerToTLSExecModels(GA, DAG);
1433 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1434 SelectionDAG &DAG) const {
1435 EVT PtrVT = getPointerTy();
1436 DebugLoc dl = Op.getDebugLoc();
1437 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1438 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1439 if (RelocM == Reloc::PIC_) {
1440 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1441 ARMConstantPoolValue *CPV =
1442 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1443 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1444 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1445 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1447 PseudoSourceValue::getConstantPool(), 0,
1449 SDValue Chain = Result.getValue(1);
1450 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1451 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1453 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1454 PseudoSourceValue::getGOT(), 0,
1458 // If we have T2 ops, we can materialize the address directly via movt/movw
1459 // pair. This is always cheaper.
1460 if (Subtarget->useMovt()) {
1461 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1462 DAG.getTargetGlobalAddress(GV, PtrVT));
1464 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1465 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1466 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1467 PseudoSourceValue::getConstantPool(), 0,
1473 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1474 SelectionDAG &DAG) const {
1475 MachineFunction &MF = DAG.getMachineFunction();
1476 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1477 unsigned ARMPCLabelIndex = 0;
1478 EVT PtrVT = getPointerTy();
1479 DebugLoc dl = Op.getDebugLoc();
1480 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1481 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1483 if (RelocM == Reloc::Static)
1484 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1486 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1487 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1488 ARMConstantPoolValue *CPV =
1489 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1490 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1492 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1494 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1495 PseudoSourceValue::getConstantPool(), 0,
1497 SDValue Chain = Result.getValue(1);
1499 if (RelocM == Reloc::PIC_) {
1500 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1501 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1504 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1505 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1506 PseudoSourceValue::getGOT(), 0,
1512 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1513 SelectionDAG &DAG) const {
1514 assert(Subtarget->isTargetELF() &&
1515 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1516 MachineFunction &MF = DAG.getMachineFunction();
1517 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1518 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1519 EVT PtrVT = getPointerTy();
1520 DebugLoc dl = Op.getDebugLoc();
1521 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1522 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1523 "_GLOBAL_OFFSET_TABLE_",
1524 ARMPCLabelIndex, PCAdj);
1525 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1526 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1527 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1528 PseudoSourceValue::getConstantPool(), 0,
1530 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1531 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1535 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1536 const ARMSubtarget *Subtarget)
1538 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1539 DebugLoc dl = Op.getDebugLoc();
1541 default: return SDValue(); // Don't custom lower most intrinsics.
1542 case Intrinsic::arm_thread_pointer: {
1543 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1544 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1546 case Intrinsic::eh_sjlj_lsda: {
1547 MachineFunction &MF = DAG.getMachineFunction();
1548 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1549 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1550 EVT PtrVT = getPointerTy();
1551 DebugLoc dl = Op.getDebugLoc();
1552 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1554 unsigned PCAdj = (RelocM != Reloc::PIC_)
1555 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1556 ARMConstantPoolValue *CPV =
1557 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1558 ARMCP::CPLSDA, PCAdj);
1559 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1560 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1562 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1563 PseudoSourceValue::getConstantPool(), 0,
1565 SDValue Chain = Result.getValue(1);
1567 if (RelocM == Reloc::PIC_) {
1568 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1569 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1573 case Intrinsic::eh_sjlj_setjmp:
1574 SDValue Val = Subtarget->isThumb() ?
1575 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1576 DAG.getConstant(0, MVT::i32);
1577 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1582 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1583 const ARMSubtarget *Subtarget) {
1584 DebugLoc dl = Op.getDebugLoc();
1585 SDValue Op5 = Op.getOperand(5);
1587 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1588 if (isDeviceBarrier) {
1589 if (Subtarget->hasV7Ops())
1590 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1592 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1593 DAG.getConstant(0, MVT::i32));
1595 if (Subtarget->hasV7Ops())
1596 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1598 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1599 DAG.getConstant(0, MVT::i32));
1604 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1605 MachineFunction &MF = DAG.getMachineFunction();
1606 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1608 // vastart just stores the address of the VarArgsFrameIndex slot into the
1609 // memory location argument.
1610 DebugLoc dl = Op.getDebugLoc();
1611 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1612 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1613 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1614 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1619 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1620 SelectionDAG &DAG) const {
1621 SDNode *Node = Op.getNode();
1622 DebugLoc dl = Node->getDebugLoc();
1623 EVT VT = Node->getValueType(0);
1624 SDValue Chain = Op.getOperand(0);
1625 SDValue Size = Op.getOperand(1);
1626 SDValue Align = Op.getOperand(2);
1628 // Chain the dynamic stack allocation so that it doesn't modify the stack
1629 // pointer when other instructions are using the stack.
1630 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1632 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1633 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1634 if (AlignVal > StackAlign)
1635 // Do this now since selection pass cannot introduce new target
1636 // independent node.
1637 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1639 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1640 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1641 // do even more horrible hack later.
1642 MachineFunction &MF = DAG.getMachineFunction();
1643 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1644 if (AFI->isThumb1OnlyFunction()) {
1646 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1648 uint32_t Val = C->getZExtValue();
1649 if (Val <= 508 && ((Val & 3) == 0))
1653 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1656 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1657 SDValue Ops1[] = { Chain, Size, Align };
1658 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1659 Chain = Res.getValue(1);
1660 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1661 DAG.getIntPtrConstant(0, true), SDValue());
1662 SDValue Ops2[] = { Res, Chain };
1663 return DAG.getMergeValues(Ops2, 2, dl);
1667 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1668 SDValue &Root, SelectionDAG &DAG,
1669 DebugLoc dl) const {
1670 MachineFunction &MF = DAG.getMachineFunction();
1671 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1673 TargetRegisterClass *RC;
1674 if (AFI->isThumb1OnlyFunction())
1675 RC = ARM::tGPRRegisterClass;
1677 RC = ARM::GPRRegisterClass;
1679 // Transform the arguments stored in physical registers into virtual ones.
1680 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1681 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1684 if (NextVA.isMemLoc()) {
1685 MachineFrameInfo *MFI = MF.getFrameInfo();
1686 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1688 // Create load node to retrieve arguments from the stack.
1689 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1690 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1691 PseudoSourceValue::getFixedStack(FI), 0,
1694 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1695 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1698 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1702 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1703 CallingConv::ID CallConv, bool isVarArg,
1704 const SmallVectorImpl<ISD::InputArg>
1706 DebugLoc dl, SelectionDAG &DAG,
1707 SmallVectorImpl<SDValue> &InVals)
1710 MachineFunction &MF = DAG.getMachineFunction();
1711 MachineFrameInfo *MFI = MF.getFrameInfo();
1713 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1715 // Assign locations to all of the incoming arguments.
1716 SmallVector<CCValAssign, 16> ArgLocs;
1717 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1719 CCInfo.AnalyzeFormalArguments(Ins,
1720 CCAssignFnForNode(CallConv, /* Return*/ false,
1723 SmallVector<SDValue, 16> ArgValues;
1725 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1726 CCValAssign &VA = ArgLocs[i];
1728 // Arguments stored in registers.
1729 if (VA.isRegLoc()) {
1730 EVT RegVT = VA.getLocVT();
1733 if (VA.needsCustom()) {
1734 // f64 and vector types are split up into multiple registers or
1735 // combinations of registers and stack slots.
1736 if (VA.getLocVT() == MVT::v2f64) {
1737 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1739 VA = ArgLocs[++i]; // skip ahead to next loc
1741 if (VA.isMemLoc()) {
1742 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
1744 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1745 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1746 PseudoSourceValue::getFixedStack(FI), 0,
1749 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1752 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1753 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1754 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1755 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1756 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1758 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1761 TargetRegisterClass *RC;
1763 if (RegVT == MVT::f32)
1764 RC = ARM::SPRRegisterClass;
1765 else if (RegVT == MVT::f64)
1766 RC = ARM::DPRRegisterClass;
1767 else if (RegVT == MVT::v2f64)
1768 RC = ARM::QPRRegisterClass;
1769 else if (RegVT == MVT::i32)
1770 RC = (AFI->isThumb1OnlyFunction() ?
1771 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1773 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1775 // Transform the arguments in physical registers into virtual ones.
1776 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1777 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1780 // If this is an 8 or 16-bit value, it is really passed promoted
1781 // to 32 bits. Insert an assert[sz]ext to capture this, then
1782 // truncate to the right size.
1783 switch (VA.getLocInfo()) {
1784 default: llvm_unreachable("Unknown loc info!");
1785 case CCValAssign::Full: break;
1786 case CCValAssign::BCvt:
1787 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1789 case CCValAssign::SExt:
1790 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1791 DAG.getValueType(VA.getValVT()));
1792 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1794 case CCValAssign::ZExt:
1795 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1796 DAG.getValueType(VA.getValVT()));
1797 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1801 InVals.push_back(ArgValue);
1803 } else { // VA.isRegLoc()
1806 assert(VA.isMemLoc());
1807 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1809 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1810 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1813 // Create load nodes to retrieve arguments from the stack.
1814 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1815 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1816 PseudoSourceValue::getFixedStack(FI), 0,
1823 static const unsigned GPRArgRegs[] = {
1824 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1827 unsigned NumGPRs = CCInfo.getFirstUnallocated
1828 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1830 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1831 unsigned VARegSize = (4 - NumGPRs) * 4;
1832 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1833 unsigned ArgOffset = CCInfo.getNextStackOffset();
1834 if (VARegSaveSize) {
1835 // If this function is vararg, store any remaining integer argument regs
1836 // to their spots on the stack so that they may be loaded by deferencing
1837 // the result of va_next.
1838 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1839 AFI->setVarArgsFrameIndex(
1840 MFI->CreateFixedObject(VARegSaveSize,
1841 ArgOffset + VARegSaveSize - VARegSize,
1843 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
1846 SmallVector<SDValue, 4> MemOps;
1847 for (; NumGPRs < 4; ++NumGPRs) {
1848 TargetRegisterClass *RC;
1849 if (AFI->isThumb1OnlyFunction())
1850 RC = ARM::tGPRRegisterClass;
1852 RC = ARM::GPRRegisterClass;
1854 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1855 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1857 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1858 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()), 0,
1860 MemOps.push_back(Store);
1861 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1862 DAG.getConstant(4, getPointerTy()));
1864 if (!MemOps.empty())
1865 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1866 &MemOps[0], MemOps.size());
1868 // This will point to the next argument passed via stack.
1869 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
1876 /// isFloatingPointZero - Return true if this is +0.0.
1877 static bool isFloatingPointZero(SDValue Op) {
1878 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1879 return CFP->getValueAPF().isPosZero();
1880 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1881 // Maybe this has already been legalized into the constant pool?
1882 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1883 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1884 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1885 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1886 return CFP->getValueAPF().isPosZero();
1892 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1893 /// the given operands.
1895 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1896 SDValue &ARMCC, SelectionDAG &DAG,
1897 DebugLoc dl) const {
1898 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1899 unsigned C = RHSC->getZExtValue();
1900 if (!isLegalICmpImmediate(C)) {
1901 // Constant does not fit, try adjusting it by one?
1906 if (isLegalICmpImmediate(C-1)) {
1907 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1908 RHS = DAG.getConstant(C-1, MVT::i32);
1913 if (C > 0 && isLegalICmpImmediate(C-1)) {
1914 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1915 RHS = DAG.getConstant(C-1, MVT::i32);
1920 if (isLegalICmpImmediate(C+1)) {
1921 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1922 RHS = DAG.getConstant(C+1, MVT::i32);
1927 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1928 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1929 RHS = DAG.getConstant(C+1, MVT::i32);
1936 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1937 ARMISD::NodeType CompareType;
1940 CompareType = ARMISD::CMP;
1945 CompareType = ARMISD::CMPZ;
1948 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1949 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1952 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1953 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1956 if (!isFloatingPointZero(RHS))
1957 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1959 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1960 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1963 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
1964 EVT VT = Op.getValueType();
1965 SDValue LHS = Op.getOperand(0);
1966 SDValue RHS = Op.getOperand(1);
1967 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1968 SDValue TrueVal = Op.getOperand(2);
1969 SDValue FalseVal = Op.getOperand(3);
1970 DebugLoc dl = Op.getDebugLoc();
1972 if (LHS.getValueType() == MVT::i32) {
1974 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1975 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1976 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1979 ARMCC::CondCodes CondCode, CondCode2;
1980 FPCCToARMCC(CC, CondCode, CondCode2);
1982 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1983 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1984 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1985 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1987 if (CondCode2 != ARMCC::AL) {
1988 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1989 // FIXME: Needs another CMP because flag can have but one use.
1990 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1991 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1992 Result, TrueVal, ARMCC2, CCR, Cmp2);
1997 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1998 SDValue Chain = Op.getOperand(0);
1999 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2000 SDValue LHS = Op.getOperand(2);
2001 SDValue RHS = Op.getOperand(3);
2002 SDValue Dest = Op.getOperand(4);
2003 DebugLoc dl = Op.getDebugLoc();
2005 if (LHS.getValueType() == MVT::i32) {
2007 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2008 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2009 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2010 Chain, Dest, ARMCC, CCR,Cmp);
2013 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2014 ARMCC::CondCodes CondCode, CondCode2;
2015 FPCCToARMCC(CC, CondCode, CondCode2);
2017 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2018 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2019 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2020 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2021 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2022 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2023 if (CondCode2 != ARMCC::AL) {
2024 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2025 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2026 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2031 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2032 SDValue Chain = Op.getOperand(0);
2033 SDValue Table = Op.getOperand(1);
2034 SDValue Index = Op.getOperand(2);
2035 DebugLoc dl = Op.getDebugLoc();
2037 EVT PTy = getPointerTy();
2038 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2039 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2040 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2041 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2042 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2043 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2044 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2045 if (Subtarget->isThumb2()) {
2046 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2047 // which does another jump to the destination. This also makes it easier
2048 // to translate it to TBB / TBH later.
2049 // FIXME: This might not work if the function is extremely large.
2050 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2051 Addr, Op.getOperand(2), JTI, UId);
2053 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2054 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2055 PseudoSourceValue::getJumpTable(), 0,
2057 Chain = Addr.getValue(1);
2058 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2059 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2061 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2062 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2063 Chain = Addr.getValue(1);
2064 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2068 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2069 DebugLoc dl = Op.getDebugLoc();
2072 switch (Op.getOpcode()) {
2074 assert(0 && "Invalid opcode!");
2075 case ISD::FP_TO_SINT:
2076 Opc = ARMISD::FTOSI;
2078 case ISD::FP_TO_UINT:
2079 Opc = ARMISD::FTOUI;
2082 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2083 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2086 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2087 EVT VT = Op.getValueType();
2088 DebugLoc dl = Op.getDebugLoc();
2091 switch (Op.getOpcode()) {
2093 assert(0 && "Invalid opcode!");
2094 case ISD::SINT_TO_FP:
2095 Opc = ARMISD::SITOF;
2097 case ISD::UINT_TO_FP:
2098 Opc = ARMISD::UITOF;
2102 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2103 return DAG.getNode(Opc, dl, VT, Op);
2106 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2107 // Implement fcopysign with a fabs and a conditional fneg.
2108 SDValue Tmp0 = Op.getOperand(0);
2109 SDValue Tmp1 = Op.getOperand(1);
2110 DebugLoc dl = Op.getDebugLoc();
2111 EVT VT = Op.getValueType();
2112 EVT SrcVT = Tmp1.getValueType();
2113 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2114 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2115 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2116 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2117 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2120 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2121 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2122 MFI->setFrameAddressIsTaken(true);
2123 EVT VT = Op.getValueType();
2124 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2125 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2126 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2127 ? ARM::R7 : ARM::R11;
2128 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2130 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2135 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2136 /// expand a bit convert where either the source or destination type is i64 to
2137 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2138 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2139 /// vectors), since the legalizer won't know what to do with that.
2140 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2141 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2142 DebugLoc dl = N->getDebugLoc();
2143 SDValue Op = N->getOperand(0);
2145 // This function is only supposed to be called for i64 types, either as the
2146 // source or destination of the bit convert.
2147 EVT SrcVT = Op.getValueType();
2148 EVT DstVT = N->getValueType(0);
2149 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2150 "ExpandBIT_CONVERT called for non-i64 type");
2152 // Turn i64->f64 into VMOVDRR.
2153 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2154 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2155 DAG.getConstant(0, MVT::i32));
2156 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2157 DAG.getConstant(1, MVT::i32));
2158 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2161 // Turn f64->i64 into VMOVRRD.
2162 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2163 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2164 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2165 // Merge the pieces into a single i64 value.
2166 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2172 /// getZeroVector - Returns a vector of specified type with all zero elements.
2174 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2175 assert(VT.isVector() && "Expected a vector type");
2177 // Zero vectors are used to represent vector negation and in those cases
2178 // will be implemented with the NEON VNEG instruction. However, VNEG does
2179 // not support i64 elements, so sometimes the zero vectors will need to be
2180 // explicitly constructed. For those cases, and potentially other uses in
2181 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2182 // to their dest type. This ensures they get CSE'd.
2184 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2185 SmallVector<SDValue, 8> Ops;
2188 if (VT.getSizeInBits() == 64) {
2189 Ops.assign(8, Cst); TVT = MVT::v8i8;
2191 Ops.assign(16, Cst); TVT = MVT::v16i8;
2193 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2195 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2198 /// getOnesVector - Returns a vector of specified type with all bits set.
2200 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2201 assert(VT.isVector() && "Expected a vector type");
2203 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2204 // dest type. This ensures they get CSE'd.
2206 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2207 SmallVector<SDValue, 8> Ops;
2210 if (VT.getSizeInBits() == 64) {
2211 Ops.assign(8, Cst); TVT = MVT::v8i8;
2213 Ops.assign(16, Cst); TVT = MVT::v16i8;
2215 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2217 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2220 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2221 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2222 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2223 SelectionDAG &DAG) const {
2224 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2225 EVT VT = Op.getValueType();
2226 unsigned VTBits = VT.getSizeInBits();
2227 DebugLoc dl = Op.getDebugLoc();
2228 SDValue ShOpLo = Op.getOperand(0);
2229 SDValue ShOpHi = Op.getOperand(1);
2230 SDValue ShAmt = Op.getOperand(2);
2232 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2234 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2236 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2237 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2238 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2239 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2240 DAG.getConstant(VTBits, MVT::i32));
2241 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2242 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2243 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2245 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2246 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2248 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2249 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2252 SDValue Ops[2] = { Lo, Hi };
2253 return DAG.getMergeValues(Ops, 2, dl);
2256 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2257 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2258 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2259 SelectionDAG &DAG) const {
2260 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2261 EVT VT = Op.getValueType();
2262 unsigned VTBits = VT.getSizeInBits();
2263 DebugLoc dl = Op.getDebugLoc();
2264 SDValue ShOpLo = Op.getOperand(0);
2265 SDValue ShOpHi = Op.getOperand(1);
2266 SDValue ShAmt = Op.getOperand(2);
2269 assert(Op.getOpcode() == ISD::SHL_PARTS);
2270 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2271 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2272 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2273 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2274 DAG.getConstant(VTBits, MVT::i32));
2275 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2276 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2278 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2279 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2280 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2282 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2283 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2286 SDValue Ops[2] = { Lo, Hi };
2287 return DAG.getMergeValues(Ops, 2, dl);
2290 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2291 const ARMSubtarget *ST) {
2292 EVT VT = N->getValueType(0);
2293 DebugLoc dl = N->getDebugLoc();
2295 if (!ST->hasV6T2Ops())
2298 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2299 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2302 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2303 const ARMSubtarget *ST) {
2304 EVT VT = N->getValueType(0);
2305 DebugLoc dl = N->getDebugLoc();
2307 // Lower vector shifts on NEON to use VSHL.
2308 if (VT.isVector()) {
2309 assert(ST->hasNEON() && "unexpected vector shift");
2311 // Left shifts translate directly to the vshiftu intrinsic.
2312 if (N->getOpcode() == ISD::SHL)
2313 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2314 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2315 N->getOperand(0), N->getOperand(1));
2317 assert((N->getOpcode() == ISD::SRA ||
2318 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2320 // NEON uses the same intrinsics for both left and right shifts. For
2321 // right shifts, the shift amounts are negative, so negate the vector of
2323 EVT ShiftVT = N->getOperand(1).getValueType();
2324 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2325 getZeroVector(ShiftVT, DAG, dl),
2327 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2328 Intrinsic::arm_neon_vshifts :
2329 Intrinsic::arm_neon_vshiftu);
2330 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2331 DAG.getConstant(vshiftInt, MVT::i32),
2332 N->getOperand(0), NegatedCount);
2335 // We can get here for a node like i32 = ISD::SHL i32, i64
2339 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2340 "Unknown shift to lower!");
2342 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2343 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2344 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2347 // If we are in thumb mode, we don't have RRX.
2348 if (ST->isThumb1Only()) return SDValue();
2350 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2351 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2352 DAG.getConstant(0, MVT::i32));
2353 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2354 DAG.getConstant(1, MVT::i32));
2356 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2357 // captures the result into a carry flag.
2358 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2359 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2361 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2362 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2364 // Merge the pieces into a single i64 value.
2365 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2368 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2369 SDValue TmpOp0, TmpOp1;
2370 bool Invert = false;
2374 SDValue Op0 = Op.getOperand(0);
2375 SDValue Op1 = Op.getOperand(1);
2376 SDValue CC = Op.getOperand(2);
2377 EVT VT = Op.getValueType();
2378 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2379 DebugLoc dl = Op.getDebugLoc();
2381 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2382 switch (SetCCOpcode) {
2383 default: llvm_unreachable("Illegal FP comparison"); break;
2385 case ISD::SETNE: Invert = true; // Fallthrough
2387 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2389 case ISD::SETLT: Swap = true; // Fallthrough
2391 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2393 case ISD::SETLE: Swap = true; // Fallthrough
2395 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2396 case ISD::SETUGE: Swap = true; // Fallthrough
2397 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2398 case ISD::SETUGT: Swap = true; // Fallthrough
2399 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2400 case ISD::SETUEQ: Invert = true; // Fallthrough
2402 // Expand this to (OLT | OGT).
2406 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2407 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2409 case ISD::SETUO: Invert = true; // Fallthrough
2411 // Expand this to (OLT | OGE).
2415 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2416 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2420 // Integer comparisons.
2421 switch (SetCCOpcode) {
2422 default: llvm_unreachable("Illegal integer comparison"); break;
2423 case ISD::SETNE: Invert = true;
2424 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2425 case ISD::SETLT: Swap = true;
2426 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2427 case ISD::SETLE: Swap = true;
2428 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2429 case ISD::SETULT: Swap = true;
2430 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2431 case ISD::SETULE: Swap = true;
2432 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2435 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2436 if (Opc == ARMISD::VCEQ) {
2439 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2441 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2444 // Ignore bitconvert.
2445 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2446 AndOp = AndOp.getOperand(0);
2448 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2450 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2451 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2458 std::swap(Op0, Op1);
2460 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2463 Result = DAG.getNOT(dl, Result, VT);
2468 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2469 /// VMOV instruction, and if so, return the constant being splatted.
2470 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2471 unsigned SplatBitSize, SelectionDAG &DAG) {
2472 switch (SplatBitSize) {
2474 // Any 1-byte value is OK.
2475 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2476 return DAG.getTargetConstant(SplatBits, MVT::i8);
2479 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2480 if ((SplatBits & ~0xff) == 0 ||
2481 (SplatBits & ~0xff00) == 0)
2482 return DAG.getTargetConstant(SplatBits, MVT::i16);
2486 // NEON's 32-bit VMOV supports splat values where:
2487 // * only one byte is nonzero, or
2488 // * the least significant byte is 0xff and the second byte is nonzero, or
2489 // * the least significant 2 bytes are 0xff and the third is nonzero.
2490 if ((SplatBits & ~0xff) == 0 ||
2491 (SplatBits & ~0xff00) == 0 ||
2492 (SplatBits & ~0xff0000) == 0 ||
2493 (SplatBits & ~0xff000000) == 0)
2494 return DAG.getTargetConstant(SplatBits, MVT::i32);
2496 if ((SplatBits & ~0xffff) == 0 &&
2497 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2498 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2500 if ((SplatBits & ~0xffffff) == 0 &&
2501 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2502 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2504 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2505 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2506 // VMOV.I32. A (very) minor optimization would be to replicate the value
2507 // and fall through here to test for a valid 64-bit splat. But, then the
2508 // caller would also need to check and handle the change in size.
2512 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2513 uint64_t BitMask = 0xff;
2515 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2516 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2518 else if ((SplatBits & BitMask) != 0)
2522 return DAG.getTargetConstant(Val, MVT::i64);
2526 llvm_unreachable("unexpected size for isVMOVSplat");
2533 /// getVMOVImm - If this is a build_vector of constants which can be
2534 /// formed by using a VMOV instruction of the specified element size,
2535 /// return the constant being splatted. The ByteSize field indicates the
2536 /// number of bytes of each element [1248].
2537 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2538 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2539 APInt SplatBits, SplatUndef;
2540 unsigned SplatBitSize;
2542 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2543 HasAnyUndefs, ByteSize * 8))
2546 if (SplatBitSize > ByteSize * 8)
2549 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2553 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2554 bool &ReverseVEXT, unsigned &Imm) {
2555 unsigned NumElts = VT.getVectorNumElements();
2556 ReverseVEXT = false;
2559 // If this is a VEXT shuffle, the immediate value is the index of the first
2560 // element. The other shuffle indices must be the successive elements after
2562 unsigned ExpectedElt = Imm;
2563 for (unsigned i = 1; i < NumElts; ++i) {
2564 // Increment the expected index. If it wraps around, it may still be
2565 // a VEXT but the source vectors must be swapped.
2567 if (ExpectedElt == NumElts * 2) {
2572 if (ExpectedElt != static_cast<unsigned>(M[i]))
2576 // Adjust the index value if the source operands will be swapped.
2583 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2584 /// instruction with the specified blocksize. (The order of the elements
2585 /// within each block of the vector is reversed.)
2586 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2587 unsigned BlockSize) {
2588 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2589 "Only possible block sizes for VREV are: 16, 32, 64");
2591 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2595 unsigned NumElts = VT.getVectorNumElements();
2596 unsigned BlockElts = M[0] + 1;
2598 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2601 for (unsigned i = 0; i < NumElts; ++i) {
2602 if ((unsigned) M[i] !=
2603 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2610 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2611 unsigned &WhichResult) {
2612 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2616 unsigned NumElts = VT.getVectorNumElements();
2617 WhichResult = (M[0] == 0 ? 0 : 1);
2618 for (unsigned i = 0; i < NumElts; i += 2) {
2619 if ((unsigned) M[i] != i + WhichResult ||
2620 (unsigned) M[i+1] != i + NumElts + WhichResult)
2626 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2627 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2628 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2629 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2630 unsigned &WhichResult) {
2631 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2635 unsigned NumElts = VT.getVectorNumElements();
2636 WhichResult = (M[0] == 0 ? 0 : 1);
2637 for (unsigned i = 0; i < NumElts; i += 2) {
2638 if ((unsigned) M[i] != i + WhichResult ||
2639 (unsigned) M[i+1] != i + WhichResult)
2645 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2646 unsigned &WhichResult) {
2647 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2651 unsigned NumElts = VT.getVectorNumElements();
2652 WhichResult = (M[0] == 0 ? 0 : 1);
2653 for (unsigned i = 0; i != NumElts; ++i) {
2654 if ((unsigned) M[i] != 2 * i + WhichResult)
2658 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2659 if (VT.is64BitVector() && EltSz == 32)
2665 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2666 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2667 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2668 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2669 unsigned &WhichResult) {
2670 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2674 unsigned Half = VT.getVectorNumElements() / 2;
2675 WhichResult = (M[0] == 0 ? 0 : 1);
2676 for (unsigned j = 0; j != 2; ++j) {
2677 unsigned Idx = WhichResult;
2678 for (unsigned i = 0; i != Half; ++i) {
2679 if ((unsigned) M[i + j * Half] != Idx)
2685 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2686 if (VT.is64BitVector() && EltSz == 32)
2692 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2693 unsigned &WhichResult) {
2694 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2698 unsigned NumElts = VT.getVectorNumElements();
2699 WhichResult = (M[0] == 0 ? 0 : 1);
2700 unsigned Idx = WhichResult * NumElts / 2;
2701 for (unsigned i = 0; i != NumElts; i += 2) {
2702 if ((unsigned) M[i] != Idx ||
2703 (unsigned) M[i+1] != Idx + NumElts)
2708 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2709 if (VT.is64BitVector() && EltSz == 32)
2715 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2716 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2717 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2718 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2719 unsigned &WhichResult) {
2720 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2724 unsigned NumElts = VT.getVectorNumElements();
2725 WhichResult = (M[0] == 0 ? 0 : 1);
2726 unsigned Idx = WhichResult * NumElts / 2;
2727 for (unsigned i = 0; i != NumElts; i += 2) {
2728 if ((unsigned) M[i] != Idx ||
2729 (unsigned) M[i+1] != Idx)
2734 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2735 if (VT.is64BitVector() && EltSz == 32)
2742 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2743 // Canonicalize all-zeros and all-ones vectors.
2744 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2745 if (ConstVal->isNullValue())
2746 return getZeroVector(VT, DAG, dl);
2747 if (ConstVal->isAllOnesValue())
2748 return getOnesVector(VT, DAG, dl);
2751 if (VT.is64BitVector()) {
2752 switch (Val.getValueType().getSizeInBits()) {
2753 case 8: CanonicalVT = MVT::v8i8; break;
2754 case 16: CanonicalVT = MVT::v4i16; break;
2755 case 32: CanonicalVT = MVT::v2i32; break;
2756 case 64: CanonicalVT = MVT::v1i64; break;
2757 default: llvm_unreachable("unexpected splat element type"); break;
2760 assert(VT.is128BitVector() && "unknown splat vector size");
2761 switch (Val.getValueType().getSizeInBits()) {
2762 case 8: CanonicalVT = MVT::v16i8; break;
2763 case 16: CanonicalVT = MVT::v8i16; break;
2764 case 32: CanonicalVT = MVT::v4i32; break;
2765 case 64: CanonicalVT = MVT::v2i64; break;
2766 default: llvm_unreachable("unexpected splat element type"); break;
2770 // Build a canonical splat for this value.
2771 SmallVector<SDValue, 8> Ops;
2772 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2773 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2775 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2778 // If this is a case we can't handle, return null and let the default
2779 // expansion code take care of it.
2780 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2781 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2782 DebugLoc dl = Op.getDebugLoc();
2783 EVT VT = Op.getValueType();
2785 APInt SplatBits, SplatUndef;
2786 unsigned SplatBitSize;
2788 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2789 if (SplatBitSize <= 64) {
2790 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2791 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2793 return BuildSplat(Val, VT, DAG, dl);
2797 // If there are only 2 elements in a 128-bit vector, insert them into an
2798 // undef vector. This handles the common case for 128-bit vector argument
2799 // passing, where the insertions should be translated to subreg accesses
2800 // with no real instructions.
2801 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2802 SDValue Val = DAG.getUNDEF(VT);
2803 SDValue Op0 = Op.getOperand(0);
2804 SDValue Op1 = Op.getOperand(1);
2805 if (Op0.getOpcode() != ISD::UNDEF)
2806 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2807 DAG.getIntPtrConstant(0));
2808 if (Op1.getOpcode() != ISD::UNDEF)
2809 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2810 DAG.getIntPtrConstant(1));
2817 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2818 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2819 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2820 /// are assumed to be legal.
2822 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2824 if (VT.getVectorNumElements() == 4 &&
2825 (VT.is128BitVector() || VT.is64BitVector())) {
2826 unsigned PFIndexes[4];
2827 for (unsigned i = 0; i != 4; ++i) {
2831 PFIndexes[i] = M[i];
2834 // Compute the index in the perfect shuffle table.
2835 unsigned PFTableIndex =
2836 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2837 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2838 unsigned Cost = (PFEntry >> 30);
2845 unsigned Imm, WhichResult;
2847 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2848 isVREVMask(M, VT, 64) ||
2849 isVREVMask(M, VT, 32) ||
2850 isVREVMask(M, VT, 16) ||
2851 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2852 isVTRNMask(M, VT, WhichResult) ||
2853 isVUZPMask(M, VT, WhichResult) ||
2854 isVZIPMask(M, VT, WhichResult) ||
2855 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2856 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2857 isVZIP_v_undef_Mask(M, VT, WhichResult));
2860 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2861 /// the specified operations to build the shuffle.
2862 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2863 SDValue RHS, SelectionDAG &DAG,
2865 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2866 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2867 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2870 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2879 OP_VUZPL, // VUZP, left result
2880 OP_VUZPR, // VUZP, right result
2881 OP_VZIPL, // VZIP, left result
2882 OP_VZIPR, // VZIP, right result
2883 OP_VTRNL, // VTRN, left result
2884 OP_VTRNR // VTRN, right result
2887 if (OpNum == OP_COPY) {
2888 if (LHSID == (1*9+2)*9+3) return LHS;
2889 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2893 SDValue OpLHS, OpRHS;
2894 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2895 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2896 EVT VT = OpLHS.getValueType();
2899 default: llvm_unreachable("Unknown shuffle opcode!");
2901 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2906 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2907 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2911 return DAG.getNode(ARMISD::VEXT, dl, VT,
2913 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2916 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2917 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2920 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2921 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2924 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2925 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2929 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2930 SDValue V1 = Op.getOperand(0);
2931 SDValue V2 = Op.getOperand(1);
2932 DebugLoc dl = Op.getDebugLoc();
2933 EVT VT = Op.getValueType();
2934 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2935 SmallVector<int, 8> ShuffleMask;
2937 // Convert shuffles that are directly supported on NEON to target-specific
2938 // DAG nodes, instead of keeping them as shuffles and matching them again
2939 // during code selection. This is more efficient and avoids the possibility
2940 // of inconsistencies between legalization and selection.
2941 // FIXME: floating-point vectors should be canonicalized to integer vectors
2942 // of the same time so that they get CSEd properly.
2943 SVN->getMask(ShuffleMask);
2945 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2946 int Lane = SVN->getSplatIndex();
2947 // If this is undef splat, generate it via "just" vdup, if possible.
2948 if (Lane == -1) Lane = 0;
2950 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2951 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2953 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2954 DAG.getConstant(Lane, MVT::i32));
2959 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2962 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2963 DAG.getConstant(Imm, MVT::i32));
2966 if (isVREVMask(ShuffleMask, VT, 64))
2967 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2968 if (isVREVMask(ShuffleMask, VT, 32))
2969 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2970 if (isVREVMask(ShuffleMask, VT, 16))
2971 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2973 // Check for Neon shuffles that modify both input vectors in place.
2974 // If both results are used, i.e., if there are two shuffles with the same
2975 // source operands and with masks corresponding to both results of one of
2976 // these operations, DAG memoization will ensure that a single node is
2977 // used for both shuffles.
2978 unsigned WhichResult;
2979 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2980 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2981 V1, V2).getValue(WhichResult);
2982 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2983 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2984 V1, V2).getValue(WhichResult);
2985 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2986 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2987 V1, V2).getValue(WhichResult);
2989 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
2990 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2991 V1, V1).getValue(WhichResult);
2992 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2993 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2994 V1, V1).getValue(WhichResult);
2995 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2996 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2997 V1, V1).getValue(WhichResult);
2999 // If the shuffle is not directly supported and it has 4 elements, use
3000 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3001 if (VT.getVectorNumElements() == 4 &&
3002 (VT.is128BitVector() || VT.is64BitVector())) {
3003 unsigned PFIndexes[4];
3004 for (unsigned i = 0; i != 4; ++i) {
3005 if (ShuffleMask[i] < 0)
3008 PFIndexes[i] = ShuffleMask[i];
3011 // Compute the index in the perfect shuffle table.
3012 unsigned PFTableIndex =
3013 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3015 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3016 unsigned Cost = (PFEntry >> 30);
3019 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3025 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3026 EVT VT = Op.getValueType();
3027 DebugLoc dl = Op.getDebugLoc();
3028 SDValue Vec = Op.getOperand(0);
3029 SDValue Lane = Op.getOperand(1);
3030 assert(VT == MVT::i32 &&
3031 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3032 "unexpected type for custom-lowering vector extract");
3033 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3036 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3037 // The only time a CONCAT_VECTORS operation can have legal types is when
3038 // two 64-bit vectors are concatenated to a 128-bit vector.
3039 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3040 "unexpected CONCAT_VECTORS");
3041 DebugLoc dl = Op.getDebugLoc();
3042 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3043 SDValue Op0 = Op.getOperand(0);
3044 SDValue Op1 = Op.getOperand(1);
3045 if (Op0.getOpcode() != ISD::UNDEF)
3046 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3047 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3048 DAG.getIntPtrConstant(0));
3049 if (Op1.getOpcode() != ISD::UNDEF)
3050 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3051 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3052 DAG.getIntPtrConstant(1));
3053 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3056 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3057 switch (Op.getOpcode()) {
3058 default: llvm_unreachable("Don't know how to custom lower this!");
3059 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3060 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3061 case ISD::GlobalAddress:
3062 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3063 LowerGlobalAddressELF(Op, DAG);
3064 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3065 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3066 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3067 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3068 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3069 case ISD::VASTART: return LowerVASTART(Op, DAG);
3070 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3071 case ISD::SINT_TO_FP:
3072 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3073 case ISD::FP_TO_SINT:
3074 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3075 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3076 case ISD::RETURNADDR: break;
3077 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3078 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3079 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3081 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3084 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3085 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3086 case ISD::SRL_PARTS:
3087 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3088 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3089 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3090 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3091 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3092 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3093 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3098 /// ReplaceNodeResults - Replace the results of node with an illegal result
3099 /// type with new values built out of custom code.
3100 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3101 SmallVectorImpl<SDValue>&Results,
3102 SelectionDAG &DAG) const {
3104 switch (N->getOpcode()) {
3106 llvm_unreachable("Don't know how to custom expand this!");
3108 case ISD::BIT_CONVERT:
3109 Res = ExpandBIT_CONVERT(N, DAG);
3113 Res = LowerShift(N, DAG, Subtarget);
3117 Results.push_back(Res);
3120 //===----------------------------------------------------------------------===//
3121 // ARM Scheduler Hooks
3122 //===----------------------------------------------------------------------===//
3125 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3126 MachineBasicBlock *BB,
3127 unsigned Size) const {
3128 unsigned dest = MI->getOperand(0).getReg();
3129 unsigned ptr = MI->getOperand(1).getReg();
3130 unsigned oldval = MI->getOperand(2).getReg();
3131 unsigned newval = MI->getOperand(3).getReg();
3132 unsigned scratch = BB->getParent()->getRegInfo()
3133 .createVirtualRegister(ARM::GPRRegisterClass);
3134 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3135 DebugLoc dl = MI->getDebugLoc();
3136 bool isThumb2 = Subtarget->isThumb2();
3138 unsigned ldrOpc, strOpc;
3140 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3142 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3143 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3146 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3147 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3150 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3151 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3155 MachineFunction *MF = BB->getParent();
3156 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3157 MachineFunction::iterator It = BB;
3158 ++It; // insert the new blocks after the current block
3160 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3161 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3162 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3163 MF->insert(It, loop1MBB);
3164 MF->insert(It, loop2MBB);
3165 MF->insert(It, exitMBB);
3166 exitMBB->transferSuccessors(BB);
3170 // fallthrough --> loop1MBB
3171 BB->addSuccessor(loop1MBB);
3174 // ldrex dest, [ptr]
3178 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3179 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3180 .addReg(dest).addReg(oldval));
3181 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3182 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3183 BB->addSuccessor(loop2MBB);
3184 BB->addSuccessor(exitMBB);
3187 // strex scratch, newval, [ptr]
3191 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3193 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3194 .addReg(scratch).addImm(0));
3195 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3196 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3197 BB->addSuccessor(loop1MBB);
3198 BB->addSuccessor(exitMBB);
3204 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3210 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3211 unsigned Size, unsigned BinOpcode) const {
3212 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3213 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3215 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3216 MachineFunction *MF = BB->getParent();
3217 MachineFunction::iterator It = BB;
3220 unsigned dest = MI->getOperand(0).getReg();
3221 unsigned ptr = MI->getOperand(1).getReg();
3222 unsigned incr = MI->getOperand(2).getReg();
3223 DebugLoc dl = MI->getDebugLoc();
3225 bool isThumb2 = Subtarget->isThumb2();
3226 unsigned ldrOpc, strOpc;
3228 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3230 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3231 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3234 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3235 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3238 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3239 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3243 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3244 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3245 MF->insert(It, loopMBB);
3246 MF->insert(It, exitMBB);
3247 exitMBB->transferSuccessors(BB);
3249 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3250 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3251 unsigned scratch2 = (!BinOpcode) ? incr :
3252 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3256 // fallthrough --> loopMBB
3257 BB->addSuccessor(loopMBB);
3261 // <binop> scratch2, dest, incr
3262 // strex scratch, scratch2, ptr
3265 // fallthrough --> exitMBB
3267 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3269 // operand order needs to go the other way for NAND
3270 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3271 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3272 addReg(incr).addReg(dest)).addReg(0);
3274 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3275 addReg(dest).addReg(incr)).addReg(0);
3278 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3280 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3281 .addReg(scratch).addImm(0));
3282 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3283 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3285 BB->addSuccessor(loopMBB);
3286 BB->addSuccessor(exitMBB);
3292 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3298 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3299 MachineBasicBlock *BB) const {
3300 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3301 DebugLoc dl = MI->getDebugLoc();
3302 bool isThumb2 = Subtarget->isThumb2();
3303 switch (MI->getOpcode()) {
3306 llvm_unreachable("Unexpected instr type to insert");
3308 case ARM::ATOMIC_LOAD_ADD_I8:
3309 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3310 case ARM::ATOMIC_LOAD_ADD_I16:
3311 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3312 case ARM::ATOMIC_LOAD_ADD_I32:
3313 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3315 case ARM::ATOMIC_LOAD_AND_I8:
3316 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3317 case ARM::ATOMIC_LOAD_AND_I16:
3318 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3319 case ARM::ATOMIC_LOAD_AND_I32:
3320 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3322 case ARM::ATOMIC_LOAD_OR_I8:
3323 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3324 case ARM::ATOMIC_LOAD_OR_I16:
3325 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3326 case ARM::ATOMIC_LOAD_OR_I32:
3327 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3329 case ARM::ATOMIC_LOAD_XOR_I8:
3330 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3331 case ARM::ATOMIC_LOAD_XOR_I16:
3332 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3333 case ARM::ATOMIC_LOAD_XOR_I32:
3334 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3336 case ARM::ATOMIC_LOAD_NAND_I8:
3337 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3338 case ARM::ATOMIC_LOAD_NAND_I16:
3339 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3340 case ARM::ATOMIC_LOAD_NAND_I32:
3341 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3343 case ARM::ATOMIC_LOAD_SUB_I8:
3344 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3345 case ARM::ATOMIC_LOAD_SUB_I16:
3346 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3347 case ARM::ATOMIC_LOAD_SUB_I32:
3348 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3350 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3351 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3352 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3354 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3355 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3356 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3358 case ARM::tMOVCCr_pseudo: {
3359 // To "insert" a SELECT_CC instruction, we actually have to insert the
3360 // diamond control-flow pattern. The incoming instruction knows the
3361 // destination vreg to set, the condition code register to branch on, the
3362 // true/false values to select between, and a branch opcode to use.
3363 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3364 MachineFunction::iterator It = BB;
3370 // cmpTY ccX, r1, r2
3372 // fallthrough --> copy0MBB
3373 MachineBasicBlock *thisMBB = BB;
3374 MachineFunction *F = BB->getParent();
3375 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3376 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3377 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3378 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3379 F->insert(It, copy0MBB);
3380 F->insert(It, sinkMBB);
3381 // Update machine-CFG edges by first adding all successors of the current
3382 // block to the new block which will contain the Phi node for the select.
3383 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3384 E = BB->succ_end(); I != E; ++I)
3385 sinkMBB->addSuccessor(*I);
3386 // Next, remove all successors of the current block, and add the true
3387 // and fallthrough blocks as its successors.
3388 while (!BB->succ_empty())
3389 BB->removeSuccessor(BB->succ_begin());
3390 BB->addSuccessor(copy0MBB);
3391 BB->addSuccessor(sinkMBB);
3394 // %FalseValue = ...
3395 // # fallthrough to sinkMBB
3398 // Update machine-CFG edges
3399 BB->addSuccessor(sinkMBB);
3402 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3405 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3406 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3407 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3409 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3416 case ARM::t2SUBrSPi_:
3417 case ARM::t2SUBrSPi12_:
3418 case ARM::t2SUBrSPs_: {
3419 MachineFunction *MF = BB->getParent();
3420 unsigned DstReg = MI->getOperand(0).getReg();
3421 unsigned SrcReg = MI->getOperand(1).getReg();
3422 bool DstIsDead = MI->getOperand(0).isDead();
3423 bool SrcIsKill = MI->getOperand(1).isKill();
3425 if (SrcReg != ARM::SP) {
3426 // Copy the source to SP from virtual register.
3427 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3428 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3429 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3430 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3431 .addReg(SrcReg, getKillRegState(SrcIsKill));
3435 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3436 switch (MI->getOpcode()) {
3438 llvm_unreachable("Unexpected pseudo instruction!");
3444 OpOpc = ARM::tADDspr;
3447 OpOpc = ARM::tSUBspi;
3449 case ARM::t2SUBrSPi_:
3450 OpOpc = ARM::t2SUBrSPi;
3451 NeedPred = true; NeedCC = true;
3453 case ARM::t2SUBrSPi12_:
3454 OpOpc = ARM::t2SUBrSPi12;
3457 case ARM::t2SUBrSPs_:
3458 OpOpc = ARM::t2SUBrSPs;
3459 NeedPred = true; NeedCC = true; NeedOp3 = true;
3462 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3463 if (OpOpc == ARM::tAND)
3464 AddDefaultT1CC(MIB);
3465 MIB.addReg(ARM::SP);
3466 MIB.addOperand(MI->getOperand(2));
3468 MIB.addOperand(MI->getOperand(3));
3470 AddDefaultPred(MIB);
3474 // Copy the result from SP to virtual register.
3475 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3476 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3477 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3478 BuildMI(BB, dl, TII->get(CopyOpc))
3479 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3481 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3487 //===----------------------------------------------------------------------===//
3488 // ARM Optimization Hooks
3489 //===----------------------------------------------------------------------===//
3492 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3493 TargetLowering::DAGCombinerInfo &DCI) {
3494 SelectionDAG &DAG = DCI.DAG;
3495 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3496 EVT VT = N->getValueType(0);
3497 unsigned Opc = N->getOpcode();
3498 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3499 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3500 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3501 ISD::CondCode CC = ISD::SETCC_INVALID;
3504 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3506 SDValue CCOp = Slct.getOperand(0);
3507 if (CCOp.getOpcode() == ISD::SETCC)
3508 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3511 bool DoXform = false;
3513 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3516 if (LHS.getOpcode() == ISD::Constant &&
3517 cast<ConstantSDNode>(LHS)->isNullValue()) {
3519 } else if (CC != ISD::SETCC_INVALID &&
3520 RHS.getOpcode() == ISD::Constant &&
3521 cast<ConstantSDNode>(RHS)->isNullValue()) {
3522 std::swap(LHS, RHS);
3523 SDValue Op0 = Slct.getOperand(0);
3524 EVT OpVT = isSlctCC ? Op0.getValueType() :
3525 Op0.getOperand(0).getValueType();
3526 bool isInt = OpVT.isInteger();
3527 CC = ISD::getSetCCInverse(CC, isInt);
3529 if (!TLI.isCondCodeLegal(CC, OpVT))
3530 return SDValue(); // Inverse operator isn't legal.
3537 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3539 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3540 Slct.getOperand(0), Slct.getOperand(1), CC);
3541 SDValue CCOp = Slct.getOperand(0);
3543 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3544 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3545 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3546 CCOp, OtherOp, Result);
3551 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3552 static SDValue PerformADDCombine(SDNode *N,
3553 TargetLowering::DAGCombinerInfo &DCI) {
3554 // added by evan in r37685 with no testcase.
3555 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3557 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3558 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3559 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3560 if (Result.getNode()) return Result;
3562 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3563 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3564 if (Result.getNode()) return Result;
3570 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3571 static SDValue PerformSUBCombine(SDNode *N,
3572 TargetLowering::DAGCombinerInfo &DCI) {
3573 // added by evan in r37685 with no testcase.
3574 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3576 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3577 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3578 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3579 if (Result.getNode()) return Result;
3585 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3586 /// ARMISD::VMOVRRD.
3587 static SDValue PerformVMOVRRDCombine(SDNode *N,
3588 TargetLowering::DAGCombinerInfo &DCI) {
3589 // fmrrd(fmdrr x, y) -> x,y
3590 SDValue InDouble = N->getOperand(0);
3591 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3592 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3596 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3597 /// operand of a vector shift operation, where all the elements of the
3598 /// build_vector must have the same constant integer value.
3599 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3600 // Ignore bit_converts.
3601 while (Op.getOpcode() == ISD::BIT_CONVERT)
3602 Op = Op.getOperand(0);
3603 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3604 APInt SplatBits, SplatUndef;
3605 unsigned SplatBitSize;
3607 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3608 HasAnyUndefs, ElementBits) ||
3609 SplatBitSize > ElementBits)
3611 Cnt = SplatBits.getSExtValue();
3615 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3616 /// operand of a vector shift left operation. That value must be in the range:
3617 /// 0 <= Value < ElementBits for a left shift; or
3618 /// 0 <= Value <= ElementBits for a long left shift.
3619 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3620 assert(VT.isVector() && "vector shift count is not a vector type");
3621 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3622 if (! getVShiftImm(Op, ElementBits, Cnt))
3624 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3627 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3628 /// operand of a vector shift right operation. For a shift opcode, the value
3629 /// is positive, but for an intrinsic the value count must be negative. The
3630 /// absolute value must be in the range:
3631 /// 1 <= |Value| <= ElementBits for a right shift; or
3632 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3633 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3635 assert(VT.isVector() && "vector shift count is not a vector type");
3636 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3637 if (! getVShiftImm(Op, ElementBits, Cnt))
3641 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3644 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3645 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3646 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3649 // Don't do anything for most intrinsics.
3652 // Vector shifts: check for immediate versions and lower them.
3653 // Note: This is done during DAG combining instead of DAG legalizing because
3654 // the build_vectors for 64-bit vector element shift counts are generally
3655 // not legal, and it is hard to see their values after they get legalized to
3656 // loads from a constant pool.
3657 case Intrinsic::arm_neon_vshifts:
3658 case Intrinsic::arm_neon_vshiftu:
3659 case Intrinsic::arm_neon_vshiftls:
3660 case Intrinsic::arm_neon_vshiftlu:
3661 case Intrinsic::arm_neon_vshiftn:
3662 case Intrinsic::arm_neon_vrshifts:
3663 case Intrinsic::arm_neon_vrshiftu:
3664 case Intrinsic::arm_neon_vrshiftn:
3665 case Intrinsic::arm_neon_vqshifts:
3666 case Intrinsic::arm_neon_vqshiftu:
3667 case Intrinsic::arm_neon_vqshiftsu:
3668 case Intrinsic::arm_neon_vqshiftns:
3669 case Intrinsic::arm_neon_vqshiftnu:
3670 case Intrinsic::arm_neon_vqshiftnsu:
3671 case Intrinsic::arm_neon_vqrshiftns:
3672 case Intrinsic::arm_neon_vqrshiftnu:
3673 case Intrinsic::arm_neon_vqrshiftnsu: {
3674 EVT VT = N->getOperand(1).getValueType();
3676 unsigned VShiftOpc = 0;
3679 case Intrinsic::arm_neon_vshifts:
3680 case Intrinsic::arm_neon_vshiftu:
3681 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3682 VShiftOpc = ARMISD::VSHL;
3685 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3686 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3687 ARMISD::VSHRs : ARMISD::VSHRu);
3692 case Intrinsic::arm_neon_vshiftls:
3693 case Intrinsic::arm_neon_vshiftlu:
3694 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3696 llvm_unreachable("invalid shift count for vshll intrinsic");
3698 case Intrinsic::arm_neon_vrshifts:
3699 case Intrinsic::arm_neon_vrshiftu:
3700 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3704 case Intrinsic::arm_neon_vqshifts:
3705 case Intrinsic::arm_neon_vqshiftu:
3706 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3710 case Intrinsic::arm_neon_vqshiftsu:
3711 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3713 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3715 case Intrinsic::arm_neon_vshiftn:
3716 case Intrinsic::arm_neon_vrshiftn:
3717 case Intrinsic::arm_neon_vqshiftns:
3718 case Intrinsic::arm_neon_vqshiftnu:
3719 case Intrinsic::arm_neon_vqshiftnsu:
3720 case Intrinsic::arm_neon_vqrshiftns:
3721 case Intrinsic::arm_neon_vqrshiftnu:
3722 case Intrinsic::arm_neon_vqrshiftnsu:
3723 // Narrowing shifts require an immediate right shift.
3724 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3726 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3729 llvm_unreachable("unhandled vector shift");
3733 case Intrinsic::arm_neon_vshifts:
3734 case Intrinsic::arm_neon_vshiftu:
3735 // Opcode already set above.
3737 case Intrinsic::arm_neon_vshiftls:
3738 case Intrinsic::arm_neon_vshiftlu:
3739 if (Cnt == VT.getVectorElementType().getSizeInBits())
3740 VShiftOpc = ARMISD::VSHLLi;
3742 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3743 ARMISD::VSHLLs : ARMISD::VSHLLu);
3745 case Intrinsic::arm_neon_vshiftn:
3746 VShiftOpc = ARMISD::VSHRN; break;
3747 case Intrinsic::arm_neon_vrshifts:
3748 VShiftOpc = ARMISD::VRSHRs; break;
3749 case Intrinsic::arm_neon_vrshiftu:
3750 VShiftOpc = ARMISD::VRSHRu; break;
3751 case Intrinsic::arm_neon_vrshiftn:
3752 VShiftOpc = ARMISD::VRSHRN; break;
3753 case Intrinsic::arm_neon_vqshifts:
3754 VShiftOpc = ARMISD::VQSHLs; break;
3755 case Intrinsic::arm_neon_vqshiftu:
3756 VShiftOpc = ARMISD::VQSHLu; break;
3757 case Intrinsic::arm_neon_vqshiftsu:
3758 VShiftOpc = ARMISD::VQSHLsu; break;
3759 case Intrinsic::arm_neon_vqshiftns:
3760 VShiftOpc = ARMISD::VQSHRNs; break;
3761 case Intrinsic::arm_neon_vqshiftnu:
3762 VShiftOpc = ARMISD::VQSHRNu; break;
3763 case Intrinsic::arm_neon_vqshiftnsu:
3764 VShiftOpc = ARMISD::VQSHRNsu; break;
3765 case Intrinsic::arm_neon_vqrshiftns:
3766 VShiftOpc = ARMISD::VQRSHRNs; break;
3767 case Intrinsic::arm_neon_vqrshiftnu:
3768 VShiftOpc = ARMISD::VQRSHRNu; break;
3769 case Intrinsic::arm_neon_vqrshiftnsu:
3770 VShiftOpc = ARMISD::VQRSHRNsu; break;
3773 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3774 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3777 case Intrinsic::arm_neon_vshiftins: {
3778 EVT VT = N->getOperand(1).getValueType();
3780 unsigned VShiftOpc = 0;
3782 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3783 VShiftOpc = ARMISD::VSLI;
3784 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3785 VShiftOpc = ARMISD::VSRI;
3787 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3790 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3791 N->getOperand(1), N->getOperand(2),
3792 DAG.getConstant(Cnt, MVT::i32));
3795 case Intrinsic::arm_neon_vqrshifts:
3796 case Intrinsic::arm_neon_vqrshiftu:
3797 // No immediate versions of these to check for.
3804 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3805 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3806 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3807 /// vector element shift counts are generally not legal, and it is hard to see
3808 /// their values after they get legalized to loads from a constant pool.
3809 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3810 const ARMSubtarget *ST) {
3811 EVT VT = N->getValueType(0);
3813 // Nothing to be done for scalar shifts.
3814 if (! VT.isVector())
3817 assert(ST->hasNEON() && "unexpected vector shift");
3820 switch (N->getOpcode()) {
3821 default: llvm_unreachable("unexpected shift opcode");
3824 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3825 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3826 DAG.getConstant(Cnt, MVT::i32));
3831 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3832 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3833 ARMISD::VSHRs : ARMISD::VSHRu);
3834 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3835 DAG.getConstant(Cnt, MVT::i32));
3841 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3842 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3843 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3844 const ARMSubtarget *ST) {
3845 SDValue N0 = N->getOperand(0);
3847 // Check for sign- and zero-extensions of vector extract operations of 8-
3848 // and 16-bit vector elements. NEON supports these directly. They are
3849 // handled during DAG combining because type legalization will promote them
3850 // to 32-bit types and it is messy to recognize the operations after that.
3851 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3852 SDValue Vec = N0.getOperand(0);
3853 SDValue Lane = N0.getOperand(1);
3854 EVT VT = N->getValueType(0);
3855 EVT EltVT = N0.getValueType();
3856 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3858 if (VT == MVT::i32 &&
3859 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3860 TLI.isTypeLegal(Vec.getValueType())) {
3863 switch (N->getOpcode()) {
3864 default: llvm_unreachable("unexpected opcode");
3865 case ISD::SIGN_EXTEND:
3866 Opc = ARMISD::VGETLANEs;
3868 case ISD::ZERO_EXTEND:
3869 case ISD::ANY_EXTEND:
3870 Opc = ARMISD::VGETLANEu;
3873 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3880 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3881 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3882 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3883 const ARMSubtarget *ST) {
3884 // If the target supports NEON, try to use vmax/vmin instructions for f32
3885 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3886 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3887 // a NaN; only do the transformation when it matches that behavior.
3889 // For now only do this when using NEON for FP operations; if using VFP, it
3890 // is not obvious that the benefit outweighs the cost of switching to the
3892 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3893 N->getValueType(0) != MVT::f32)
3896 SDValue CondLHS = N->getOperand(0);
3897 SDValue CondRHS = N->getOperand(1);
3898 SDValue LHS = N->getOperand(2);
3899 SDValue RHS = N->getOperand(3);
3900 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3902 unsigned Opcode = 0;
3904 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3905 IsReversed = false; // x CC y ? x : y
3906 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3907 IsReversed = true ; // x CC y ? y : x
3921 // If LHS is NaN, an ordered comparison will be false and the result will
3922 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3923 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3924 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3925 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3927 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3928 // will return -0, so vmin can only be used for unsafe math or if one of
3929 // the operands is known to be nonzero.
3930 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3932 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3934 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3943 // If LHS is NaN, an ordered comparison will be false and the result will
3944 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3945 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3946 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3947 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3949 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3950 // will return +0, so vmax can only be used for unsafe math or if one of
3951 // the operands is known to be nonzero.
3952 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3954 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3956 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3962 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3965 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3966 DAGCombinerInfo &DCI) const {
3967 switch (N->getOpcode()) {
3969 case ISD::ADD: return PerformADDCombine(N, DCI);
3970 case ISD::SUB: return PerformSUBCombine(N, DCI);
3971 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3972 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
3975 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
3976 case ISD::SIGN_EXTEND:
3977 case ISD::ZERO_EXTEND:
3978 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
3979 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
3984 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3985 if (!Subtarget->hasV6Ops())
3986 // Pre-v6 does not support unaligned mem access.
3989 // v6+ may or may not support unaligned mem access depending on the system
3991 // FIXME: This is pretty conservative. Should we provide cmdline option to
3992 // control the behaviour?
3993 if (!Subtarget->isTargetDarwin())
3997 switch (VT.getSimpleVT().SimpleTy) {
4004 // FIXME: VLD1 etc with standard alignment is legal.
4008 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4013 switch (VT.getSimpleVT().SimpleTy) {
4014 default: return false;
4029 if ((V & (Scale - 1)) != 0)
4032 return V == (V & ((1LL << 5) - 1));
4035 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4036 const ARMSubtarget *Subtarget) {
4043 switch (VT.getSimpleVT().SimpleTy) {
4044 default: return false;
4049 // + imm12 or - imm8
4051 return V == (V & ((1LL << 8) - 1));
4052 return V == (V & ((1LL << 12) - 1));
4055 // Same as ARM mode. FIXME: NEON?
4056 if (!Subtarget->hasVFP2())
4061 return V == (V & ((1LL << 8) - 1));
4065 /// isLegalAddressImmediate - Return true if the integer value can be used
4066 /// as the offset of the target addressing mode for load / store of the
4068 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4069 const ARMSubtarget *Subtarget) {
4076 if (Subtarget->isThumb1Only())
4077 return isLegalT1AddressImmediate(V, VT);
4078 else if (Subtarget->isThumb2())
4079 return isLegalT2AddressImmediate(V, VT, Subtarget);
4084 switch (VT.getSimpleVT().SimpleTy) {
4085 default: return false;
4090 return V == (V & ((1LL << 12) - 1));
4093 return V == (V & ((1LL << 8) - 1));
4096 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4101 return V == (V & ((1LL << 8) - 1));
4105 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4107 int Scale = AM.Scale;
4111 switch (VT.getSimpleVT().SimpleTy) {
4112 default: return false;
4121 return Scale == 2 || Scale == 4 || Scale == 8;
4124 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4128 // Note, we allow "void" uses (basically, uses that aren't loads or
4129 // stores), because arm allows folding a scale into many arithmetic
4130 // operations. This should be made more precise and revisited later.
4132 // Allow r << imm, but the imm has to be a multiple of two.
4133 if (Scale & 1) return false;
4134 return isPowerOf2_32(Scale);
4138 /// isLegalAddressingMode - Return true if the addressing mode represented
4139 /// by AM is legal for this target, for a load/store of the specified type.
4140 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4141 const Type *Ty) const {
4142 EVT VT = getValueType(Ty, true);
4143 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4146 // Can never fold addr of global into load/store.
4151 case 0: // no scale reg, must be "r+i" or "r", or "i".
4154 if (Subtarget->isThumb1Only())
4158 // ARM doesn't support any R+R*scale+imm addr modes.
4165 if (Subtarget->isThumb2())
4166 return isLegalT2ScaledAddressingMode(AM, VT);
4168 int Scale = AM.Scale;
4169 switch (VT.getSimpleVT().SimpleTy) {
4170 default: return false;
4174 if (Scale < 0) Scale = -Scale;
4178 return isPowerOf2_32(Scale & ~1);
4182 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4187 // Note, we allow "void" uses (basically, uses that aren't loads or
4188 // stores), because arm allows folding a scale into many arithmetic
4189 // operations. This should be made more precise and revisited later.
4191 // Allow r << imm, but the imm has to be a multiple of two.
4192 if (Scale & 1) return false;
4193 return isPowerOf2_32(Scale);
4200 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4201 /// icmp immediate, that is the target has icmp instructions which can compare
4202 /// a register against the immediate without having to materialize the
4203 /// immediate into a register.
4204 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4205 if (!Subtarget->isThumb())
4206 return ARM_AM::getSOImmVal(Imm) != -1;
4207 if (Subtarget->isThumb2())
4208 return ARM_AM::getT2SOImmVal(Imm) != -1;
4209 return Imm >= 0 && Imm <= 255;
4212 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4213 bool isSEXTLoad, SDValue &Base,
4214 SDValue &Offset, bool &isInc,
4215 SelectionDAG &DAG) {
4216 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4219 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4221 Base = Ptr->getOperand(0);
4222 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4223 int RHSC = (int)RHS->getZExtValue();
4224 if (RHSC < 0 && RHSC > -256) {
4225 assert(Ptr->getOpcode() == ISD::ADD);
4227 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4231 isInc = (Ptr->getOpcode() == ISD::ADD);
4232 Offset = Ptr->getOperand(1);
4234 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4236 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4237 int RHSC = (int)RHS->getZExtValue();
4238 if (RHSC < 0 && RHSC > -0x1000) {
4239 assert(Ptr->getOpcode() == ISD::ADD);
4241 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4242 Base = Ptr->getOperand(0);
4247 if (Ptr->getOpcode() == ISD::ADD) {
4249 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4250 if (ShOpcVal != ARM_AM::no_shift) {
4251 Base = Ptr->getOperand(1);
4252 Offset = Ptr->getOperand(0);
4254 Base = Ptr->getOperand(0);
4255 Offset = Ptr->getOperand(1);
4260 isInc = (Ptr->getOpcode() == ISD::ADD);
4261 Base = Ptr->getOperand(0);
4262 Offset = Ptr->getOperand(1);
4266 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4270 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4271 bool isSEXTLoad, SDValue &Base,
4272 SDValue &Offset, bool &isInc,
4273 SelectionDAG &DAG) {
4274 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4277 Base = Ptr->getOperand(0);
4278 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4279 int RHSC = (int)RHS->getZExtValue();
4280 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4281 assert(Ptr->getOpcode() == ISD::ADD);
4283 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4285 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4286 isInc = Ptr->getOpcode() == ISD::ADD;
4287 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4295 /// getPreIndexedAddressParts - returns true by value, base pointer and
4296 /// offset pointer and addressing mode by reference if the node's address
4297 /// can be legally represented as pre-indexed load / store address.
4299 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4301 ISD::MemIndexedMode &AM,
4302 SelectionDAG &DAG) const {
4303 if (Subtarget->isThumb1Only())
4308 bool isSEXTLoad = false;
4309 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4310 Ptr = LD->getBasePtr();
4311 VT = LD->getMemoryVT();
4312 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4313 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4314 Ptr = ST->getBasePtr();
4315 VT = ST->getMemoryVT();
4320 bool isLegal = false;
4321 if (Subtarget->isThumb2())
4322 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4323 Offset, isInc, DAG);
4325 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4326 Offset, isInc, DAG);
4330 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4334 /// getPostIndexedAddressParts - returns true by value, base pointer and
4335 /// offset pointer and addressing mode by reference if this node can be
4336 /// combined with a load / store to form a post-indexed load / store.
4337 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4340 ISD::MemIndexedMode &AM,
4341 SelectionDAG &DAG) const {
4342 if (Subtarget->isThumb1Only())
4347 bool isSEXTLoad = false;
4348 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4349 VT = LD->getMemoryVT();
4350 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4351 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4352 VT = ST->getMemoryVT();
4357 bool isLegal = false;
4358 if (Subtarget->isThumb2())
4359 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4362 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4367 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4371 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4375 const SelectionDAG &DAG,
4376 unsigned Depth) const {
4377 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4378 switch (Op.getOpcode()) {
4380 case ARMISD::CMOV: {
4381 // Bits are known zero/one if known on the LHS and RHS.
4382 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4383 if (KnownZero == 0 && KnownOne == 0) return;
4385 APInt KnownZeroRHS, KnownOneRHS;
4386 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4387 KnownZeroRHS, KnownOneRHS, Depth+1);
4388 KnownZero &= KnownZeroRHS;
4389 KnownOne &= KnownOneRHS;
4395 //===----------------------------------------------------------------------===//
4396 // ARM Inline Assembly Support
4397 //===----------------------------------------------------------------------===//
4399 /// getConstraintType - Given a constraint letter, return the type of
4400 /// constraint it is for this target.
4401 ARMTargetLowering::ConstraintType
4402 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4403 if (Constraint.size() == 1) {
4404 switch (Constraint[0]) {
4406 case 'l': return C_RegisterClass;
4407 case 'w': return C_RegisterClass;
4410 return TargetLowering::getConstraintType(Constraint);
4413 std::pair<unsigned, const TargetRegisterClass*>
4414 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4416 if (Constraint.size() == 1) {
4417 // GCC ARM Constraint Letters
4418 switch (Constraint[0]) {
4420 if (Subtarget->isThumb())
4421 return std::make_pair(0U, ARM::tGPRRegisterClass);
4423 return std::make_pair(0U, ARM::GPRRegisterClass);
4425 return std::make_pair(0U, ARM::GPRRegisterClass);
4428 return std::make_pair(0U, ARM::SPRRegisterClass);
4429 if (VT.getSizeInBits() == 64)
4430 return std::make_pair(0U, ARM::DPRRegisterClass);
4431 if (VT.getSizeInBits() == 128)
4432 return std::make_pair(0U, ARM::QPRRegisterClass);
4436 if (StringRef("{cc}").equals_lower(Constraint))
4437 return std::make_pair(0U, ARM::CCRRegisterClass);
4439 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4442 std::vector<unsigned> ARMTargetLowering::
4443 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4445 if (Constraint.size() != 1)
4446 return std::vector<unsigned>();
4448 switch (Constraint[0]) { // GCC ARM Constraint Letters
4451 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4452 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4455 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4456 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4457 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4458 ARM::R12, ARM::LR, 0);
4461 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4462 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4463 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4464 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4465 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4466 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4467 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4468 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4469 if (VT.getSizeInBits() == 64)
4470 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4471 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4472 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4473 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4474 if (VT.getSizeInBits() == 128)
4475 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4476 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4480 return std::vector<unsigned>();
4483 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4484 /// vector. If it is invalid, don't add anything to Ops.
4485 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4488 std::vector<SDValue>&Ops,
4489 SelectionDAG &DAG) const {
4490 SDValue Result(0, 0);
4492 switch (Constraint) {
4494 case 'I': case 'J': case 'K': case 'L':
4495 case 'M': case 'N': case 'O':
4496 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4500 int64_t CVal64 = C->getSExtValue();
4501 int CVal = (int) CVal64;
4502 // None of these constraints allow values larger than 32 bits. Check
4503 // that the value fits in an int.
4507 switch (Constraint) {
4509 if (Subtarget->isThumb1Only()) {
4510 // This must be a constant between 0 and 255, for ADD
4512 if (CVal >= 0 && CVal <= 255)
4514 } else if (Subtarget->isThumb2()) {
4515 // A constant that can be used as an immediate value in a
4516 // data-processing instruction.
4517 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4520 // A constant that can be used as an immediate value in a
4521 // data-processing instruction.
4522 if (ARM_AM::getSOImmVal(CVal) != -1)
4528 if (Subtarget->isThumb()) { // FIXME thumb2
4529 // This must be a constant between -255 and -1, for negated ADD
4530 // immediates. This can be used in GCC with an "n" modifier that
4531 // prints the negated value, for use with SUB instructions. It is
4532 // not useful otherwise but is implemented for compatibility.
4533 if (CVal >= -255 && CVal <= -1)
4536 // This must be a constant between -4095 and 4095. It is not clear
4537 // what this constraint is intended for. Implemented for
4538 // compatibility with GCC.
4539 if (CVal >= -4095 && CVal <= 4095)
4545 if (Subtarget->isThumb1Only()) {
4546 // A 32-bit value where only one byte has a nonzero value. Exclude
4547 // zero to match GCC. This constraint is used by GCC internally for
4548 // constants that can be loaded with a move/shift combination.
4549 // It is not useful otherwise but is implemented for compatibility.
4550 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4552 } else if (Subtarget->isThumb2()) {
4553 // A constant whose bitwise inverse can be used as an immediate
4554 // value in a data-processing instruction. This can be used in GCC
4555 // with a "B" modifier that prints the inverted value, for use with
4556 // BIC and MVN instructions. It is not useful otherwise but is
4557 // implemented for compatibility.
4558 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4561 // A constant whose bitwise inverse can be used as an immediate
4562 // value in a data-processing instruction. This can be used in GCC
4563 // with a "B" modifier that prints the inverted value, for use with
4564 // BIC and MVN instructions. It is not useful otherwise but is
4565 // implemented for compatibility.
4566 if (ARM_AM::getSOImmVal(~CVal) != -1)
4572 if (Subtarget->isThumb1Only()) {
4573 // This must be a constant between -7 and 7,
4574 // for 3-operand ADD/SUB immediate instructions.
4575 if (CVal >= -7 && CVal < 7)
4577 } else if (Subtarget->isThumb2()) {
4578 // A constant whose negation can be used as an immediate value in a
4579 // data-processing instruction. This can be used in GCC with an "n"
4580 // modifier that prints the negated value, for use with SUB
4581 // instructions. It is not useful otherwise but is implemented for
4583 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4586 // A constant whose negation can be used as an immediate value in a
4587 // data-processing instruction. This can be used in GCC with an "n"
4588 // modifier that prints the negated value, for use with SUB
4589 // instructions. It is not useful otherwise but is implemented for
4591 if (ARM_AM::getSOImmVal(-CVal) != -1)
4597 if (Subtarget->isThumb()) { // FIXME thumb2
4598 // This must be a multiple of 4 between 0 and 1020, for
4599 // ADD sp + immediate.
4600 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4603 // A power of two or a constant between 0 and 32. This is used in
4604 // GCC for the shift amount on shifted register operands, but it is
4605 // useful in general for any shift amounts.
4606 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4612 if (Subtarget->isThumb()) { // FIXME thumb2
4613 // This must be a constant between 0 and 31, for shift amounts.
4614 if (CVal >= 0 && CVal <= 31)
4620 if (Subtarget->isThumb()) { // FIXME thumb2
4621 // This must be a multiple of 4 between -508 and 508, for
4622 // ADD/SUB sp = sp + immediate.
4623 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4628 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4632 if (Result.getNode()) {
4633 Ops.push_back(Result);
4636 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4641 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4642 // The ARM target isn't yet aware of offsets.
4646 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4647 APInt Imm = FPImm.bitcastToAPInt();
4648 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4649 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4650 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4652 // We can handle 4 bits of mantissa.
4653 // mantissa = (16+UInt(e:f:g:h))/16.
4654 if (Mantissa & 0x7ffff)
4657 if ((Mantissa & 0xf) != Mantissa)
4660 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4661 if (Exp < -3 || Exp > 4)
4663 Exp = ((Exp+3) & 0x7) ^ 4;
4665 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4668 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4669 APInt Imm = FPImm.bitcastToAPInt();
4670 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4671 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4672 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4674 // We can handle 4 bits of mantissa.
4675 // mantissa = (16+UInt(e:f:g:h))/16.
4676 if (Mantissa & 0xffffffffffffLL)
4679 if ((Mantissa & 0xf) != Mantissa)
4682 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4683 if (Exp < -3 || Exp > 4)
4685 Exp = ((Exp+3) & 0x7) ^ 4;
4687 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4690 /// isFPImmLegal - Returns true if the target can instruction select the
4691 /// specified FP immediate natively. If false, the legalizer will
4692 /// materialize the FP immediate as a load from a constant pool.
4693 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4694 if (!Subtarget->hasVFP3())
4697 return ARM::getVFPf32Imm(Imm) != -1;
4699 return ARM::getVFPf64Imm(Imm) != -1;