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 (Subtarget->hasNEON()) {
589 if (VT == MVT::v4i64)
590 return ARM::QQPRRegisterClass;
591 else if (VT == MVT::v8i64)
592 return ARM::QQQQPRRegisterClass;
594 return TargetLowering::getRegClassFor(VT);
597 /// getFunctionAlignment - Return the Log2 alignment of this function.
598 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
599 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
602 //===----------------------------------------------------------------------===//
604 //===----------------------------------------------------------------------===//
606 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
607 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
609 default: llvm_unreachable("Unknown condition code!");
610 case ISD::SETNE: return ARMCC::NE;
611 case ISD::SETEQ: return ARMCC::EQ;
612 case ISD::SETGT: return ARMCC::GT;
613 case ISD::SETGE: return ARMCC::GE;
614 case ISD::SETLT: return ARMCC::LT;
615 case ISD::SETLE: return ARMCC::LE;
616 case ISD::SETUGT: return ARMCC::HI;
617 case ISD::SETUGE: return ARMCC::HS;
618 case ISD::SETULT: return ARMCC::LO;
619 case ISD::SETULE: return ARMCC::LS;
623 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
624 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
625 ARMCC::CondCodes &CondCode2) {
626 CondCode2 = ARMCC::AL;
628 default: llvm_unreachable("Unknown FP condition!");
630 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
632 case ISD::SETOGT: CondCode = ARMCC::GT; break;
634 case ISD::SETOGE: CondCode = ARMCC::GE; break;
635 case ISD::SETOLT: CondCode = ARMCC::MI; break;
636 case ISD::SETOLE: CondCode = ARMCC::LS; break;
637 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
638 case ISD::SETO: CondCode = ARMCC::VC; break;
639 case ISD::SETUO: CondCode = ARMCC::VS; break;
640 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
641 case ISD::SETUGT: CondCode = ARMCC::HI; break;
642 case ISD::SETUGE: CondCode = ARMCC::PL; break;
644 case ISD::SETULT: CondCode = ARMCC::LT; break;
646 case ISD::SETULE: CondCode = ARMCC::LE; break;
648 case ISD::SETUNE: CondCode = ARMCC::NE; break;
652 //===----------------------------------------------------------------------===//
653 // Calling Convention Implementation
654 //===----------------------------------------------------------------------===//
656 #include "ARMGenCallingConv.inc"
658 // APCS f64 is in register pairs, possibly split to stack
659 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
660 CCValAssign::LocInfo &LocInfo,
661 CCState &State, bool CanFail) {
662 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
664 // Try to get the first register.
665 if (unsigned Reg = State.AllocateReg(RegList, 4))
666 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
668 // For the 2nd half of a v2f64, do not fail.
672 // Put the whole thing on the stack.
673 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
674 State.AllocateStack(8, 4),
679 // Try to get the second register.
680 if (unsigned Reg = State.AllocateReg(RegList, 4))
681 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
683 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
684 State.AllocateStack(4, 4),
689 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
690 CCValAssign::LocInfo &LocInfo,
691 ISD::ArgFlagsTy &ArgFlags,
693 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
695 if (LocVT == MVT::v2f64 &&
696 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
698 return true; // we handled it
701 // AAPCS f64 is in aligned register pairs
702 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
703 CCValAssign::LocInfo &LocInfo,
704 CCState &State, bool CanFail) {
705 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
706 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
708 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
710 // For the 2nd half of a v2f64, do not just fail.
714 // Put the whole thing on the stack.
715 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
716 State.AllocateStack(8, 8),
722 for (i = 0; i < 2; ++i)
723 if (HiRegList[i] == Reg)
726 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
727 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
732 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
733 CCValAssign::LocInfo &LocInfo,
734 ISD::ArgFlagsTy &ArgFlags,
736 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
738 if (LocVT == MVT::v2f64 &&
739 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
741 return true; // we handled it
744 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
745 CCValAssign::LocInfo &LocInfo, CCState &State) {
746 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
747 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
749 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
751 return false; // we didn't handle it
754 for (i = 0; i < 2; ++i)
755 if (HiRegList[i] == Reg)
758 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
759 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
764 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
765 CCValAssign::LocInfo &LocInfo,
766 ISD::ArgFlagsTy &ArgFlags,
768 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
770 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
772 return true; // we handled it
775 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
776 CCValAssign::LocInfo &LocInfo,
777 ISD::ArgFlagsTy &ArgFlags,
779 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
783 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
784 /// given CallingConvention value.
785 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
787 bool isVarArg) const {
790 llvm_unreachable("Unsupported calling convention");
792 case CallingConv::Fast:
793 // Use target triple & subtarget features to do actual dispatch.
794 if (Subtarget->isAAPCS_ABI()) {
795 if (Subtarget->hasVFP2() &&
796 FloatABIType == FloatABI::Hard && !isVarArg)
797 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
799 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
801 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
802 case CallingConv::ARM_AAPCS_VFP:
803 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
804 case CallingConv::ARM_AAPCS:
805 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
806 case CallingConv::ARM_APCS:
807 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
811 /// LowerCallResult - Lower the result values of a call into the
812 /// appropriate copies out of appropriate physical registers.
814 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
815 CallingConv::ID CallConv, bool isVarArg,
816 const SmallVectorImpl<ISD::InputArg> &Ins,
817 DebugLoc dl, SelectionDAG &DAG,
818 SmallVectorImpl<SDValue> &InVals) const {
820 // Assign locations to each value returned by this call.
821 SmallVector<CCValAssign, 16> RVLocs;
822 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
823 RVLocs, *DAG.getContext());
824 CCInfo.AnalyzeCallResult(Ins,
825 CCAssignFnForNode(CallConv, /* Return*/ true,
828 // Copy all of the result registers out of their specified physreg.
829 for (unsigned i = 0; i != RVLocs.size(); ++i) {
830 CCValAssign VA = RVLocs[i];
833 if (VA.needsCustom()) {
834 // Handle f64 or half of a v2f64.
835 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
837 Chain = Lo.getValue(1);
838 InFlag = Lo.getValue(2);
839 VA = RVLocs[++i]; // skip ahead to next loc
840 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
842 Chain = Hi.getValue(1);
843 InFlag = Hi.getValue(2);
844 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
846 if (VA.getLocVT() == MVT::v2f64) {
847 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
848 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
849 DAG.getConstant(0, MVT::i32));
851 VA = RVLocs[++i]; // skip ahead to next loc
852 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
853 Chain = Lo.getValue(1);
854 InFlag = Lo.getValue(2);
855 VA = RVLocs[++i]; // skip ahead to next loc
856 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
857 Chain = Hi.getValue(1);
858 InFlag = Hi.getValue(2);
859 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
860 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
861 DAG.getConstant(1, MVT::i32));
864 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
866 Chain = Val.getValue(1);
867 InFlag = Val.getValue(2);
870 switch (VA.getLocInfo()) {
871 default: llvm_unreachable("Unknown loc info!");
872 case CCValAssign::Full: break;
873 case CCValAssign::BCvt:
874 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
878 InVals.push_back(Val);
884 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
885 /// by "Src" to address "Dst" of size "Size". Alignment information is
886 /// specified by the specific parameter attribute. The copy will be passed as
887 /// a byval function parameter.
888 /// Sometimes what we are copying is the end of a larger object, the part that
889 /// does not fit in registers.
891 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
892 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
894 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
895 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
896 /*isVolatile=*/false, /*AlwaysInline=*/false,
900 /// LowerMemOpCallTo - Store the argument to the stack.
902 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
903 SDValue StackPtr, SDValue Arg,
904 DebugLoc dl, SelectionDAG &DAG,
905 const CCValAssign &VA,
906 ISD::ArgFlagsTy Flags) const {
907 unsigned LocMemOffset = VA.getLocMemOffset();
908 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
909 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
910 if (Flags.isByVal()) {
911 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
913 return DAG.getStore(Chain, dl, Arg, PtrOff,
914 PseudoSourceValue::getStack(), LocMemOffset,
918 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
919 SDValue Chain, SDValue &Arg,
920 RegsToPassVector &RegsToPass,
921 CCValAssign &VA, CCValAssign &NextVA,
923 SmallVector<SDValue, 8> &MemOpChains,
924 ISD::ArgFlagsTy Flags) const {
926 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
927 DAG.getVTList(MVT::i32, MVT::i32), Arg);
928 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
930 if (NextVA.isRegLoc())
931 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
933 assert(NextVA.isMemLoc());
934 if (StackPtr.getNode() == 0)
935 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
937 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
943 /// LowerCall - Lowering a call into a callseq_start <-
944 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
947 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
948 CallingConv::ID CallConv, bool isVarArg,
950 const SmallVectorImpl<ISD::OutputArg> &Outs,
951 const SmallVectorImpl<ISD::InputArg> &Ins,
952 DebugLoc dl, SelectionDAG &DAG,
953 SmallVectorImpl<SDValue> &InVals) const {
954 // ARM target does not yet support tail call optimization.
957 // Analyze operands of the call, assigning locations to each operand.
958 SmallVector<CCValAssign, 16> ArgLocs;
959 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
961 CCInfo.AnalyzeCallOperands(Outs,
962 CCAssignFnForNode(CallConv, /* Return*/ false,
965 // Get a count of how many bytes are to be pushed on the stack.
966 unsigned NumBytes = CCInfo.getNextStackOffset();
968 // Adjust the stack pointer for the new arguments...
969 // These operations are automatically eliminated by the prolog/epilog pass
970 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
972 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
974 RegsToPassVector RegsToPass;
975 SmallVector<SDValue, 8> MemOpChains;
977 // Walk the register/memloc assignments, inserting copies/loads. In the case
978 // of tail call optimization, arguments are handled later.
979 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
982 CCValAssign &VA = ArgLocs[i];
983 SDValue Arg = Outs[realArgIdx].Val;
984 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
986 // Promote the value if needed.
987 switch (VA.getLocInfo()) {
988 default: llvm_unreachable("Unknown loc info!");
989 case CCValAssign::Full: break;
990 case CCValAssign::SExt:
991 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
993 case CCValAssign::ZExt:
994 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
996 case CCValAssign::AExt:
997 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
999 case CCValAssign::BCvt:
1000 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1004 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1005 if (VA.needsCustom()) {
1006 if (VA.getLocVT() == MVT::v2f64) {
1007 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1008 DAG.getConstant(0, MVT::i32));
1009 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1010 DAG.getConstant(1, MVT::i32));
1012 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1013 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1015 VA = ArgLocs[++i]; // skip ahead to next loc
1016 if (VA.isRegLoc()) {
1017 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1018 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1020 assert(VA.isMemLoc());
1022 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1023 dl, DAG, VA, Flags));
1026 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1027 StackPtr, MemOpChains, Flags);
1029 } else if (VA.isRegLoc()) {
1030 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1032 assert(VA.isMemLoc());
1034 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1035 dl, DAG, VA, Flags));
1039 if (!MemOpChains.empty())
1040 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1041 &MemOpChains[0], MemOpChains.size());
1043 // Build a sequence of copy-to-reg nodes chained together with token chain
1044 // and flag operands which copy the outgoing args into the appropriate regs.
1046 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1047 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1048 RegsToPass[i].second, InFlag);
1049 InFlag = Chain.getValue(1);
1052 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1053 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1054 // node so that legalize doesn't hack it.
1055 bool isDirect = false;
1056 bool isARMFunc = false;
1057 bool isLocalARMFunc = false;
1058 MachineFunction &MF = DAG.getMachineFunction();
1059 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1061 if (EnableARMLongCalls) {
1062 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1063 && "long-calls with non-static relocation model!");
1064 // Handle a global address or an external symbol. If it's not one of
1065 // those, the target's already in a register, so we don't need to do
1067 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1068 const GlobalValue *GV = G->getGlobal();
1069 // Create a constant pool entry for the callee address
1070 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1071 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1074 // Get the address of the callee into a register
1075 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1076 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1077 Callee = DAG.getLoad(getPointerTy(), dl,
1078 DAG.getEntryNode(), CPAddr,
1079 PseudoSourceValue::getConstantPool(), 0,
1081 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1082 const char *Sym = S->getSymbol();
1084 // Create a constant pool entry for the callee address
1085 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1086 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1087 Sym, ARMPCLabelIndex, 0);
1088 // Get the address of the callee into a register
1089 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1090 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1091 Callee = DAG.getLoad(getPointerTy(), dl,
1092 DAG.getEntryNode(), CPAddr,
1093 PseudoSourceValue::getConstantPool(), 0,
1096 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1097 const GlobalValue *GV = G->getGlobal();
1099 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1100 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1101 getTargetMachine().getRelocationModel() != Reloc::Static;
1102 isARMFunc = !Subtarget->isThumb() || isStub;
1103 // ARM call to a local ARM function is predicable.
1104 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1105 // tBX takes a register source operand.
1106 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1107 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1108 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1111 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1112 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1113 Callee = DAG.getLoad(getPointerTy(), dl,
1114 DAG.getEntryNode(), CPAddr,
1115 PseudoSourceValue::getConstantPool(), 0,
1117 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1118 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1119 getPointerTy(), Callee, PICLabel);
1121 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1122 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1124 bool isStub = Subtarget->isTargetDarwin() &&
1125 getTargetMachine().getRelocationModel() != Reloc::Static;
1126 isARMFunc = !Subtarget->isThumb() || isStub;
1127 // tBX takes a register source operand.
1128 const char *Sym = S->getSymbol();
1129 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1130 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1131 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1132 Sym, ARMPCLabelIndex, 4);
1133 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1134 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1135 Callee = DAG.getLoad(getPointerTy(), dl,
1136 DAG.getEntryNode(), CPAddr,
1137 PseudoSourceValue::getConstantPool(), 0,
1139 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1140 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1141 getPointerTy(), Callee, PICLabel);
1143 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1146 // FIXME: handle tail calls differently.
1148 if (Subtarget->isThumb()) {
1149 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1150 CallOpc = ARMISD::CALL_NOLINK;
1152 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1154 CallOpc = (isDirect || Subtarget->hasV5TOps())
1155 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1156 : ARMISD::CALL_NOLINK;
1158 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1159 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1160 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1161 InFlag = Chain.getValue(1);
1164 std::vector<SDValue> Ops;
1165 Ops.push_back(Chain);
1166 Ops.push_back(Callee);
1168 // Add argument registers to the end of the list so that they are known live
1170 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1171 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1172 RegsToPass[i].second.getValueType()));
1174 if (InFlag.getNode())
1175 Ops.push_back(InFlag);
1176 // Returns a chain and a flag for retval copy to use.
1177 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1178 &Ops[0], Ops.size());
1179 InFlag = Chain.getValue(1);
1181 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1182 DAG.getIntPtrConstant(0, true), InFlag);
1184 InFlag = Chain.getValue(1);
1186 // Handle result values, copying them out of physregs into vregs that we
1188 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1193 ARMTargetLowering::LowerReturn(SDValue Chain,
1194 CallingConv::ID CallConv, bool isVarArg,
1195 const SmallVectorImpl<ISD::OutputArg> &Outs,
1196 DebugLoc dl, SelectionDAG &DAG) const {
1198 // CCValAssign - represent the assignment of the return value to a location.
1199 SmallVector<CCValAssign, 16> RVLocs;
1201 // CCState - Info about the registers and stack slots.
1202 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1205 // Analyze outgoing return values.
1206 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1209 // If this is the first return lowered for this function, add
1210 // the regs to the liveout set for the function.
1211 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1212 for (unsigned i = 0; i != RVLocs.size(); ++i)
1213 if (RVLocs[i].isRegLoc())
1214 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1219 // Copy the result values into the output registers.
1220 for (unsigned i = 0, realRVLocIdx = 0;
1222 ++i, ++realRVLocIdx) {
1223 CCValAssign &VA = RVLocs[i];
1224 assert(VA.isRegLoc() && "Can only return in registers!");
1226 SDValue Arg = Outs[realRVLocIdx].Val;
1228 switch (VA.getLocInfo()) {
1229 default: llvm_unreachable("Unknown loc info!");
1230 case CCValAssign::Full: break;
1231 case CCValAssign::BCvt:
1232 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1236 if (VA.needsCustom()) {
1237 if (VA.getLocVT() == MVT::v2f64) {
1238 // Extract the first half and return it in two registers.
1239 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1240 DAG.getConstant(0, MVT::i32));
1241 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1242 DAG.getVTList(MVT::i32, MVT::i32), Half);
1244 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1245 Flag = Chain.getValue(1);
1246 VA = RVLocs[++i]; // skip ahead to next loc
1247 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1248 HalfGPRs.getValue(1), Flag);
1249 Flag = Chain.getValue(1);
1250 VA = RVLocs[++i]; // skip ahead to next loc
1252 // Extract the 2nd half and fall through to handle it as an f64 value.
1253 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1254 DAG.getConstant(1, MVT::i32));
1256 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1258 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1259 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1260 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1261 Flag = Chain.getValue(1);
1262 VA = RVLocs[++i]; // skip ahead to next loc
1263 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1266 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1268 // Guarantee that all emitted copies are
1269 // stuck together, avoiding something bad.
1270 Flag = Chain.getValue(1);
1275 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1277 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1282 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1283 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1284 // one of the above mentioned nodes. It has to be wrapped because otherwise
1285 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1286 // be used to form addressing mode. These wrapped nodes will be selected
1288 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1289 EVT PtrVT = Op.getValueType();
1290 // FIXME there is no actual debug info here
1291 DebugLoc dl = Op.getDebugLoc();
1292 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1294 if (CP->isMachineConstantPoolEntry())
1295 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1296 CP->getAlignment());
1298 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1299 CP->getAlignment());
1300 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1303 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1304 SelectionDAG &DAG) const {
1305 MachineFunction &MF = DAG.getMachineFunction();
1306 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1307 unsigned ARMPCLabelIndex = 0;
1308 DebugLoc DL = Op.getDebugLoc();
1309 EVT PtrVT = getPointerTy();
1310 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1311 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1313 if (RelocM == Reloc::Static) {
1314 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1316 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1317 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1318 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1319 ARMCP::CPBlockAddress,
1321 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1323 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1324 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1325 PseudoSourceValue::getConstantPool(), 0,
1327 if (RelocM == Reloc::Static)
1329 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1330 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1333 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1335 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1336 SelectionDAG &DAG) const {
1337 DebugLoc dl = GA->getDebugLoc();
1338 EVT PtrVT = getPointerTy();
1339 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1340 MachineFunction &MF = DAG.getMachineFunction();
1341 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1342 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1343 ARMConstantPoolValue *CPV =
1344 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1345 ARMCP::CPValue, PCAdj, "tlsgd", true);
1346 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1347 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1348 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1349 PseudoSourceValue::getConstantPool(), 0,
1351 SDValue Chain = Argument.getValue(1);
1353 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1354 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1356 // call __tls_get_addr.
1359 Entry.Node = Argument;
1360 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1361 Args.push_back(Entry);
1362 // FIXME: is there useful debug info available here?
1363 std::pair<SDValue, SDValue> CallResult =
1364 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1365 false, false, false, false,
1366 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1367 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1368 return CallResult.first;
1371 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1372 // "local exec" model.
1374 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1375 SelectionDAG &DAG) const {
1376 const GlobalValue *GV = GA->getGlobal();
1377 DebugLoc dl = GA->getDebugLoc();
1379 SDValue Chain = DAG.getEntryNode();
1380 EVT PtrVT = getPointerTy();
1381 // Get the Thread Pointer
1382 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1384 if (GV->isDeclaration()) {
1385 MachineFunction &MF = DAG.getMachineFunction();
1386 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1387 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1388 // Initial exec model.
1389 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1390 ARMConstantPoolValue *CPV =
1391 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1392 ARMCP::CPValue, PCAdj, "gottpoff", true);
1393 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1394 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1395 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1396 PseudoSourceValue::getConstantPool(), 0,
1398 Chain = Offset.getValue(1);
1400 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1401 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1403 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1404 PseudoSourceValue::getConstantPool(), 0,
1408 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1409 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1410 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1411 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1412 PseudoSourceValue::getConstantPool(), 0,
1416 // The address of the thread local variable is the add of the thread
1417 // pointer with the offset of the variable.
1418 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1422 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1423 // TODO: implement the "local dynamic" model
1424 assert(Subtarget->isTargetELF() &&
1425 "TLS not implemented for non-ELF targets");
1426 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1427 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1428 // otherwise use the "Local Exec" TLS Model
1429 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1430 return LowerToTLSGeneralDynamicModel(GA, DAG);
1432 return LowerToTLSExecModels(GA, DAG);
1435 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1436 SelectionDAG &DAG) const {
1437 EVT PtrVT = getPointerTy();
1438 DebugLoc dl = Op.getDebugLoc();
1439 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1440 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1441 if (RelocM == Reloc::PIC_) {
1442 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1443 ARMConstantPoolValue *CPV =
1444 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1445 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1446 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1447 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1449 PseudoSourceValue::getConstantPool(), 0,
1451 SDValue Chain = Result.getValue(1);
1452 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1453 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1455 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1456 PseudoSourceValue::getGOT(), 0,
1460 // If we have T2 ops, we can materialize the address directly via movt/movw
1461 // pair. This is always cheaper.
1462 if (Subtarget->useMovt()) {
1463 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1464 DAG.getTargetGlobalAddress(GV, PtrVT));
1466 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1467 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1468 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1469 PseudoSourceValue::getConstantPool(), 0,
1475 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1476 SelectionDAG &DAG) const {
1477 MachineFunction &MF = DAG.getMachineFunction();
1478 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1479 unsigned ARMPCLabelIndex = 0;
1480 EVT PtrVT = getPointerTy();
1481 DebugLoc dl = Op.getDebugLoc();
1482 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1483 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1485 if (RelocM == Reloc::Static)
1486 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1488 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1489 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1490 ARMConstantPoolValue *CPV =
1491 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1492 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1494 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1496 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1497 PseudoSourceValue::getConstantPool(), 0,
1499 SDValue Chain = Result.getValue(1);
1501 if (RelocM == Reloc::PIC_) {
1502 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1503 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1506 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1507 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1508 PseudoSourceValue::getGOT(), 0,
1514 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1515 SelectionDAG &DAG) const {
1516 assert(Subtarget->isTargetELF() &&
1517 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1518 MachineFunction &MF = DAG.getMachineFunction();
1519 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1520 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1521 EVT PtrVT = getPointerTy();
1522 DebugLoc dl = Op.getDebugLoc();
1523 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1524 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1525 "_GLOBAL_OFFSET_TABLE_",
1526 ARMPCLabelIndex, PCAdj);
1527 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1528 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1529 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1530 PseudoSourceValue::getConstantPool(), 0,
1532 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1533 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1537 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1538 const ARMSubtarget *Subtarget)
1540 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1541 DebugLoc dl = Op.getDebugLoc();
1543 default: return SDValue(); // Don't custom lower most intrinsics.
1544 case Intrinsic::arm_thread_pointer: {
1545 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1546 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1548 case Intrinsic::eh_sjlj_lsda: {
1549 MachineFunction &MF = DAG.getMachineFunction();
1550 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1551 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1552 EVT PtrVT = getPointerTy();
1553 DebugLoc dl = Op.getDebugLoc();
1554 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1556 unsigned PCAdj = (RelocM != Reloc::PIC_)
1557 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1558 ARMConstantPoolValue *CPV =
1559 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1560 ARMCP::CPLSDA, PCAdj);
1561 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1562 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1564 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1565 PseudoSourceValue::getConstantPool(), 0,
1567 SDValue Chain = Result.getValue(1);
1569 if (RelocM == Reloc::PIC_) {
1570 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1571 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1575 case Intrinsic::eh_sjlj_setjmp:
1576 SDValue Val = Subtarget->isThumb() ?
1577 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1578 DAG.getConstant(0, MVT::i32);
1579 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1584 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1585 const ARMSubtarget *Subtarget) {
1586 DebugLoc dl = Op.getDebugLoc();
1587 SDValue Op5 = Op.getOperand(5);
1589 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1590 if (isDeviceBarrier) {
1591 if (Subtarget->hasV7Ops())
1592 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1594 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1595 DAG.getConstant(0, MVT::i32));
1597 if (Subtarget->hasV7Ops())
1598 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1600 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1601 DAG.getConstant(0, MVT::i32));
1606 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1607 MachineFunction &MF = DAG.getMachineFunction();
1608 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1610 // vastart just stores the address of the VarArgsFrameIndex slot into the
1611 // memory location argument.
1612 DebugLoc dl = Op.getDebugLoc();
1613 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1614 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1615 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1616 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1621 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1622 SelectionDAG &DAG) const {
1623 SDNode *Node = Op.getNode();
1624 DebugLoc dl = Node->getDebugLoc();
1625 EVT VT = Node->getValueType(0);
1626 SDValue Chain = Op.getOperand(0);
1627 SDValue Size = Op.getOperand(1);
1628 SDValue Align = Op.getOperand(2);
1630 // Chain the dynamic stack allocation so that it doesn't modify the stack
1631 // pointer when other instructions are using the stack.
1632 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1634 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1635 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1636 if (AlignVal > StackAlign)
1637 // Do this now since selection pass cannot introduce new target
1638 // independent node.
1639 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1641 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1642 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1643 // do even more horrible hack later.
1644 MachineFunction &MF = DAG.getMachineFunction();
1645 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1646 if (AFI->isThumb1OnlyFunction()) {
1648 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1650 uint32_t Val = C->getZExtValue();
1651 if (Val <= 508 && ((Val & 3) == 0))
1655 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1658 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1659 SDValue Ops1[] = { Chain, Size, Align };
1660 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1661 Chain = Res.getValue(1);
1662 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1663 DAG.getIntPtrConstant(0, true), SDValue());
1664 SDValue Ops2[] = { Res, Chain };
1665 return DAG.getMergeValues(Ops2, 2, dl);
1669 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1670 SDValue &Root, SelectionDAG &DAG,
1671 DebugLoc dl) const {
1672 MachineFunction &MF = DAG.getMachineFunction();
1673 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1675 TargetRegisterClass *RC;
1676 if (AFI->isThumb1OnlyFunction())
1677 RC = ARM::tGPRRegisterClass;
1679 RC = ARM::GPRRegisterClass;
1681 // Transform the arguments stored in physical registers into virtual ones.
1682 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1683 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1686 if (NextVA.isMemLoc()) {
1687 MachineFrameInfo *MFI = MF.getFrameInfo();
1688 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1690 // Create load node to retrieve arguments from the stack.
1691 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1692 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1693 PseudoSourceValue::getFixedStack(FI), 0,
1696 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1697 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1700 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1704 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1705 CallingConv::ID CallConv, bool isVarArg,
1706 const SmallVectorImpl<ISD::InputArg>
1708 DebugLoc dl, SelectionDAG &DAG,
1709 SmallVectorImpl<SDValue> &InVals)
1712 MachineFunction &MF = DAG.getMachineFunction();
1713 MachineFrameInfo *MFI = MF.getFrameInfo();
1715 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1717 // Assign locations to all of the incoming arguments.
1718 SmallVector<CCValAssign, 16> ArgLocs;
1719 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1721 CCInfo.AnalyzeFormalArguments(Ins,
1722 CCAssignFnForNode(CallConv, /* Return*/ false,
1725 SmallVector<SDValue, 16> ArgValues;
1727 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1728 CCValAssign &VA = ArgLocs[i];
1730 // Arguments stored in registers.
1731 if (VA.isRegLoc()) {
1732 EVT RegVT = VA.getLocVT();
1735 if (VA.needsCustom()) {
1736 // f64 and vector types are split up into multiple registers or
1737 // combinations of registers and stack slots.
1738 if (VA.getLocVT() == MVT::v2f64) {
1739 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1741 VA = ArgLocs[++i]; // skip ahead to next loc
1743 if (VA.isMemLoc()) {
1744 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
1746 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1747 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1748 PseudoSourceValue::getFixedStack(FI), 0,
1751 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1754 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1755 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1756 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1757 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1758 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1760 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1763 TargetRegisterClass *RC;
1765 if (RegVT == MVT::f32)
1766 RC = ARM::SPRRegisterClass;
1767 else if (RegVT == MVT::f64)
1768 RC = ARM::DPRRegisterClass;
1769 else if (RegVT == MVT::v2f64)
1770 RC = ARM::QPRRegisterClass;
1771 else if (RegVT == MVT::i32)
1772 RC = (AFI->isThumb1OnlyFunction() ?
1773 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1775 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1777 // Transform the arguments in physical registers into virtual ones.
1778 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1779 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1782 // If this is an 8 or 16-bit value, it is really passed promoted
1783 // to 32 bits. Insert an assert[sz]ext to capture this, then
1784 // truncate to the right size.
1785 switch (VA.getLocInfo()) {
1786 default: llvm_unreachable("Unknown loc info!");
1787 case CCValAssign::Full: break;
1788 case CCValAssign::BCvt:
1789 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1791 case CCValAssign::SExt:
1792 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1793 DAG.getValueType(VA.getValVT()));
1794 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1796 case CCValAssign::ZExt:
1797 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1798 DAG.getValueType(VA.getValVT()));
1799 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1803 InVals.push_back(ArgValue);
1805 } else { // VA.isRegLoc()
1808 assert(VA.isMemLoc());
1809 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1811 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1812 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1815 // Create load nodes to retrieve arguments from the stack.
1816 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1817 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1818 PseudoSourceValue::getFixedStack(FI), 0,
1825 static const unsigned GPRArgRegs[] = {
1826 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1829 unsigned NumGPRs = CCInfo.getFirstUnallocated
1830 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1832 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1833 unsigned VARegSize = (4 - NumGPRs) * 4;
1834 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1835 unsigned ArgOffset = CCInfo.getNextStackOffset();
1836 if (VARegSaveSize) {
1837 // If this function is vararg, store any remaining integer argument regs
1838 // to their spots on the stack so that they may be loaded by deferencing
1839 // the result of va_next.
1840 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1841 AFI->setVarArgsFrameIndex(
1842 MFI->CreateFixedObject(VARegSaveSize,
1843 ArgOffset + VARegSaveSize - VARegSize,
1845 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
1848 SmallVector<SDValue, 4> MemOps;
1849 for (; NumGPRs < 4; ++NumGPRs) {
1850 TargetRegisterClass *RC;
1851 if (AFI->isThumb1OnlyFunction())
1852 RC = ARM::tGPRRegisterClass;
1854 RC = ARM::GPRRegisterClass;
1856 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1857 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1859 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1860 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()), 0,
1862 MemOps.push_back(Store);
1863 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1864 DAG.getConstant(4, getPointerTy()));
1866 if (!MemOps.empty())
1867 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1868 &MemOps[0], MemOps.size());
1870 // This will point to the next argument passed via stack.
1871 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
1878 /// isFloatingPointZero - Return true if this is +0.0.
1879 static bool isFloatingPointZero(SDValue Op) {
1880 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1881 return CFP->getValueAPF().isPosZero();
1882 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1883 // Maybe this has already been legalized into the constant pool?
1884 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1885 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1886 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1887 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1888 return CFP->getValueAPF().isPosZero();
1894 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1895 /// the given operands.
1897 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1898 SDValue &ARMCC, SelectionDAG &DAG,
1899 DebugLoc dl) const {
1900 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1901 unsigned C = RHSC->getZExtValue();
1902 if (!isLegalICmpImmediate(C)) {
1903 // Constant does not fit, try adjusting it by one?
1908 if (isLegalICmpImmediate(C-1)) {
1909 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1910 RHS = DAG.getConstant(C-1, MVT::i32);
1915 if (C > 0 && isLegalICmpImmediate(C-1)) {
1916 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1917 RHS = DAG.getConstant(C-1, MVT::i32);
1922 if (isLegalICmpImmediate(C+1)) {
1923 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1924 RHS = DAG.getConstant(C+1, MVT::i32);
1929 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1930 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1931 RHS = DAG.getConstant(C+1, MVT::i32);
1938 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1939 ARMISD::NodeType CompareType;
1942 CompareType = ARMISD::CMP;
1947 CompareType = ARMISD::CMPZ;
1950 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1951 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1954 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1955 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1958 if (!isFloatingPointZero(RHS))
1959 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1961 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1962 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1965 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
1966 EVT VT = Op.getValueType();
1967 SDValue LHS = Op.getOperand(0);
1968 SDValue RHS = Op.getOperand(1);
1969 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1970 SDValue TrueVal = Op.getOperand(2);
1971 SDValue FalseVal = Op.getOperand(3);
1972 DebugLoc dl = Op.getDebugLoc();
1974 if (LHS.getValueType() == MVT::i32) {
1976 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1977 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1978 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1981 ARMCC::CondCodes CondCode, CondCode2;
1982 FPCCToARMCC(CC, CondCode, CondCode2);
1984 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1985 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1986 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1987 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1989 if (CondCode2 != ARMCC::AL) {
1990 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1991 // FIXME: Needs another CMP because flag can have but one use.
1992 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1993 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1994 Result, TrueVal, ARMCC2, CCR, Cmp2);
1999 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2000 SDValue Chain = Op.getOperand(0);
2001 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2002 SDValue LHS = Op.getOperand(2);
2003 SDValue RHS = Op.getOperand(3);
2004 SDValue Dest = Op.getOperand(4);
2005 DebugLoc dl = Op.getDebugLoc();
2007 if (LHS.getValueType() == MVT::i32) {
2009 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2010 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2011 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2012 Chain, Dest, ARMCC, CCR,Cmp);
2015 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2016 ARMCC::CondCodes CondCode, CondCode2;
2017 FPCCToARMCC(CC, CondCode, CondCode2);
2019 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2020 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2021 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2022 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2023 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2024 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2025 if (CondCode2 != ARMCC::AL) {
2026 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2027 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2028 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2033 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2034 SDValue Chain = Op.getOperand(0);
2035 SDValue Table = Op.getOperand(1);
2036 SDValue Index = Op.getOperand(2);
2037 DebugLoc dl = Op.getDebugLoc();
2039 EVT PTy = getPointerTy();
2040 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2041 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2042 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2043 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2044 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2045 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2046 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2047 if (Subtarget->isThumb2()) {
2048 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2049 // which does another jump to the destination. This also makes it easier
2050 // to translate it to TBB / TBH later.
2051 // FIXME: This might not work if the function is extremely large.
2052 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2053 Addr, Op.getOperand(2), JTI, UId);
2055 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2056 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2057 PseudoSourceValue::getJumpTable(), 0,
2059 Chain = Addr.getValue(1);
2060 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2061 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2063 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2064 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2065 Chain = Addr.getValue(1);
2066 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2070 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2071 DebugLoc dl = Op.getDebugLoc();
2074 switch (Op.getOpcode()) {
2076 assert(0 && "Invalid opcode!");
2077 case ISD::FP_TO_SINT:
2078 Opc = ARMISD::FTOSI;
2080 case ISD::FP_TO_UINT:
2081 Opc = ARMISD::FTOUI;
2084 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2085 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2088 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2089 EVT VT = Op.getValueType();
2090 DebugLoc dl = Op.getDebugLoc();
2093 switch (Op.getOpcode()) {
2095 assert(0 && "Invalid opcode!");
2096 case ISD::SINT_TO_FP:
2097 Opc = ARMISD::SITOF;
2099 case ISD::UINT_TO_FP:
2100 Opc = ARMISD::UITOF;
2104 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2105 return DAG.getNode(Opc, dl, VT, Op);
2108 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2109 // Implement fcopysign with a fabs and a conditional fneg.
2110 SDValue Tmp0 = Op.getOperand(0);
2111 SDValue Tmp1 = Op.getOperand(1);
2112 DebugLoc dl = Op.getDebugLoc();
2113 EVT VT = Op.getValueType();
2114 EVT SrcVT = Tmp1.getValueType();
2115 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2116 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2117 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2118 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2119 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2122 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2123 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2124 MFI->setFrameAddressIsTaken(true);
2125 EVT VT = Op.getValueType();
2126 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2127 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2128 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2129 ? ARM::R7 : ARM::R11;
2130 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2132 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2137 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2138 /// expand a bit convert where either the source or destination type is i64 to
2139 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2140 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2141 /// vectors), since the legalizer won't know what to do with that.
2142 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2143 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2144 DebugLoc dl = N->getDebugLoc();
2145 SDValue Op = N->getOperand(0);
2147 // This function is only supposed to be called for i64 types, either as the
2148 // source or destination of the bit convert.
2149 EVT SrcVT = Op.getValueType();
2150 EVT DstVT = N->getValueType(0);
2151 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2152 "ExpandBIT_CONVERT called for non-i64 type");
2154 // Turn i64->f64 into VMOVDRR.
2155 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2156 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2157 DAG.getConstant(0, MVT::i32));
2158 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2159 DAG.getConstant(1, MVT::i32));
2160 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2163 // Turn f64->i64 into VMOVRRD.
2164 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2165 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2166 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2167 // Merge the pieces into a single i64 value.
2168 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2174 /// getZeroVector - Returns a vector of specified type with all zero elements.
2176 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2177 assert(VT.isVector() && "Expected a vector type");
2179 // Zero vectors are used to represent vector negation and in those cases
2180 // will be implemented with the NEON VNEG instruction. However, VNEG does
2181 // not support i64 elements, so sometimes the zero vectors will need to be
2182 // explicitly constructed. For those cases, and potentially other uses in
2183 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2184 // to their dest type. This ensures they get CSE'd.
2186 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2187 SmallVector<SDValue, 8> Ops;
2190 if (VT.getSizeInBits() == 64) {
2191 Ops.assign(8, Cst); TVT = MVT::v8i8;
2193 Ops.assign(16, Cst); TVT = MVT::v16i8;
2195 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2197 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2200 /// getOnesVector - Returns a vector of specified type with all bits set.
2202 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2203 assert(VT.isVector() && "Expected a vector type");
2205 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2206 // dest type. This ensures they get CSE'd.
2208 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2209 SmallVector<SDValue, 8> Ops;
2212 if (VT.getSizeInBits() == 64) {
2213 Ops.assign(8, Cst); TVT = MVT::v8i8;
2215 Ops.assign(16, Cst); TVT = MVT::v16i8;
2217 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2219 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2222 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2223 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2224 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2225 SelectionDAG &DAG) const {
2226 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2227 EVT VT = Op.getValueType();
2228 unsigned VTBits = VT.getSizeInBits();
2229 DebugLoc dl = Op.getDebugLoc();
2230 SDValue ShOpLo = Op.getOperand(0);
2231 SDValue ShOpHi = Op.getOperand(1);
2232 SDValue ShAmt = Op.getOperand(2);
2234 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2236 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2238 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2239 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2240 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2241 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2242 DAG.getConstant(VTBits, MVT::i32));
2243 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2244 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2245 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2247 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2248 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2250 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2251 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2254 SDValue Ops[2] = { Lo, Hi };
2255 return DAG.getMergeValues(Ops, 2, dl);
2258 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2259 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2260 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2261 SelectionDAG &DAG) const {
2262 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2263 EVT VT = Op.getValueType();
2264 unsigned VTBits = VT.getSizeInBits();
2265 DebugLoc dl = Op.getDebugLoc();
2266 SDValue ShOpLo = Op.getOperand(0);
2267 SDValue ShOpHi = Op.getOperand(1);
2268 SDValue ShAmt = Op.getOperand(2);
2271 assert(Op.getOpcode() == ISD::SHL_PARTS);
2272 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2273 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2274 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2275 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2276 DAG.getConstant(VTBits, MVT::i32));
2277 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2278 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2280 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2281 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2282 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2284 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2285 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2288 SDValue Ops[2] = { Lo, Hi };
2289 return DAG.getMergeValues(Ops, 2, dl);
2292 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2293 const ARMSubtarget *ST) {
2294 EVT VT = N->getValueType(0);
2295 DebugLoc dl = N->getDebugLoc();
2297 if (!ST->hasV6T2Ops())
2300 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2301 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2304 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2305 const ARMSubtarget *ST) {
2306 EVT VT = N->getValueType(0);
2307 DebugLoc dl = N->getDebugLoc();
2309 // Lower vector shifts on NEON to use VSHL.
2310 if (VT.isVector()) {
2311 assert(ST->hasNEON() && "unexpected vector shift");
2313 // Left shifts translate directly to the vshiftu intrinsic.
2314 if (N->getOpcode() == ISD::SHL)
2315 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2316 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2317 N->getOperand(0), N->getOperand(1));
2319 assert((N->getOpcode() == ISD::SRA ||
2320 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2322 // NEON uses the same intrinsics for both left and right shifts. For
2323 // right shifts, the shift amounts are negative, so negate the vector of
2325 EVT ShiftVT = N->getOperand(1).getValueType();
2326 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2327 getZeroVector(ShiftVT, DAG, dl),
2329 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2330 Intrinsic::arm_neon_vshifts :
2331 Intrinsic::arm_neon_vshiftu);
2332 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2333 DAG.getConstant(vshiftInt, MVT::i32),
2334 N->getOperand(0), NegatedCount);
2337 // We can get here for a node like i32 = ISD::SHL i32, i64
2341 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2342 "Unknown shift to lower!");
2344 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2345 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2346 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2349 // If we are in thumb mode, we don't have RRX.
2350 if (ST->isThumb1Only()) return SDValue();
2352 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2353 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2354 DAG.getConstant(0, MVT::i32));
2355 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2356 DAG.getConstant(1, MVT::i32));
2358 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2359 // captures the result into a carry flag.
2360 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2361 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2363 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2364 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2366 // Merge the pieces into a single i64 value.
2367 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2370 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2371 SDValue TmpOp0, TmpOp1;
2372 bool Invert = false;
2376 SDValue Op0 = Op.getOperand(0);
2377 SDValue Op1 = Op.getOperand(1);
2378 SDValue CC = Op.getOperand(2);
2379 EVT VT = Op.getValueType();
2380 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2381 DebugLoc dl = Op.getDebugLoc();
2383 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2384 switch (SetCCOpcode) {
2385 default: llvm_unreachable("Illegal FP comparison"); break;
2387 case ISD::SETNE: Invert = true; // Fallthrough
2389 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2391 case ISD::SETLT: Swap = true; // Fallthrough
2393 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2395 case ISD::SETLE: Swap = true; // Fallthrough
2397 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2398 case ISD::SETUGE: Swap = true; // Fallthrough
2399 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2400 case ISD::SETUGT: Swap = true; // Fallthrough
2401 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2402 case ISD::SETUEQ: Invert = true; // Fallthrough
2404 // Expand this to (OLT | OGT).
2408 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2409 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2411 case ISD::SETUO: Invert = true; // Fallthrough
2413 // Expand this to (OLT | OGE).
2417 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2418 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2422 // Integer comparisons.
2423 switch (SetCCOpcode) {
2424 default: llvm_unreachable("Illegal integer comparison"); break;
2425 case ISD::SETNE: Invert = true;
2426 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2427 case ISD::SETLT: Swap = true;
2428 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2429 case ISD::SETLE: Swap = true;
2430 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2431 case ISD::SETULT: Swap = true;
2432 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2433 case ISD::SETULE: Swap = true;
2434 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2437 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2438 if (Opc == ARMISD::VCEQ) {
2441 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2443 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2446 // Ignore bitconvert.
2447 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2448 AndOp = AndOp.getOperand(0);
2450 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2452 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2453 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2460 std::swap(Op0, Op1);
2462 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2465 Result = DAG.getNOT(dl, Result, VT);
2470 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2471 /// VMOV instruction, and if so, return the constant being splatted.
2472 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2473 unsigned SplatBitSize, SelectionDAG &DAG) {
2474 switch (SplatBitSize) {
2476 // Any 1-byte value is OK.
2477 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2478 return DAG.getTargetConstant(SplatBits, MVT::i8);
2481 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2482 if ((SplatBits & ~0xff) == 0 ||
2483 (SplatBits & ~0xff00) == 0)
2484 return DAG.getTargetConstant(SplatBits, MVT::i16);
2488 // NEON's 32-bit VMOV supports splat values where:
2489 // * only one byte is nonzero, or
2490 // * the least significant byte is 0xff and the second byte is nonzero, or
2491 // * the least significant 2 bytes are 0xff and the third is nonzero.
2492 if ((SplatBits & ~0xff) == 0 ||
2493 (SplatBits & ~0xff00) == 0 ||
2494 (SplatBits & ~0xff0000) == 0 ||
2495 (SplatBits & ~0xff000000) == 0)
2496 return DAG.getTargetConstant(SplatBits, MVT::i32);
2498 if ((SplatBits & ~0xffff) == 0 &&
2499 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2500 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2502 if ((SplatBits & ~0xffffff) == 0 &&
2503 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2504 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2506 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2507 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2508 // VMOV.I32. A (very) minor optimization would be to replicate the value
2509 // and fall through here to test for a valid 64-bit splat. But, then the
2510 // caller would also need to check and handle the change in size.
2514 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2515 uint64_t BitMask = 0xff;
2517 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2518 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2520 else if ((SplatBits & BitMask) != 0)
2524 return DAG.getTargetConstant(Val, MVT::i64);
2528 llvm_unreachable("unexpected size for isVMOVSplat");
2535 /// getVMOVImm - If this is a build_vector of constants which can be
2536 /// formed by using a VMOV instruction of the specified element size,
2537 /// return the constant being splatted. The ByteSize field indicates the
2538 /// number of bytes of each element [1248].
2539 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2540 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2541 APInt SplatBits, SplatUndef;
2542 unsigned SplatBitSize;
2544 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2545 HasAnyUndefs, ByteSize * 8))
2548 if (SplatBitSize > ByteSize * 8)
2551 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2555 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2556 bool &ReverseVEXT, unsigned &Imm) {
2557 unsigned NumElts = VT.getVectorNumElements();
2558 ReverseVEXT = false;
2561 // If this is a VEXT shuffle, the immediate value is the index of the first
2562 // element. The other shuffle indices must be the successive elements after
2564 unsigned ExpectedElt = Imm;
2565 for (unsigned i = 1; i < NumElts; ++i) {
2566 // Increment the expected index. If it wraps around, it may still be
2567 // a VEXT but the source vectors must be swapped.
2569 if (ExpectedElt == NumElts * 2) {
2574 if (ExpectedElt != static_cast<unsigned>(M[i]))
2578 // Adjust the index value if the source operands will be swapped.
2585 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2586 /// instruction with the specified blocksize. (The order of the elements
2587 /// within each block of the vector is reversed.)
2588 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2589 unsigned BlockSize) {
2590 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2591 "Only possible block sizes for VREV are: 16, 32, 64");
2593 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2597 unsigned NumElts = VT.getVectorNumElements();
2598 unsigned BlockElts = M[0] + 1;
2600 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2603 for (unsigned i = 0; i < NumElts; ++i) {
2604 if ((unsigned) M[i] !=
2605 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2612 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2613 unsigned &WhichResult) {
2614 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2618 unsigned NumElts = VT.getVectorNumElements();
2619 WhichResult = (M[0] == 0 ? 0 : 1);
2620 for (unsigned i = 0; i < NumElts; i += 2) {
2621 if ((unsigned) M[i] != i + WhichResult ||
2622 (unsigned) M[i+1] != i + NumElts + WhichResult)
2628 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2629 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2630 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2631 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2632 unsigned &WhichResult) {
2633 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2637 unsigned NumElts = VT.getVectorNumElements();
2638 WhichResult = (M[0] == 0 ? 0 : 1);
2639 for (unsigned i = 0; i < NumElts; i += 2) {
2640 if ((unsigned) M[i] != i + WhichResult ||
2641 (unsigned) M[i+1] != i + WhichResult)
2647 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2648 unsigned &WhichResult) {
2649 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2653 unsigned NumElts = VT.getVectorNumElements();
2654 WhichResult = (M[0] == 0 ? 0 : 1);
2655 for (unsigned i = 0; i != NumElts; ++i) {
2656 if ((unsigned) M[i] != 2 * i + WhichResult)
2660 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2661 if (VT.is64BitVector() && EltSz == 32)
2667 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2668 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2669 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2670 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2671 unsigned &WhichResult) {
2672 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2676 unsigned Half = VT.getVectorNumElements() / 2;
2677 WhichResult = (M[0] == 0 ? 0 : 1);
2678 for (unsigned j = 0; j != 2; ++j) {
2679 unsigned Idx = WhichResult;
2680 for (unsigned i = 0; i != Half; ++i) {
2681 if ((unsigned) M[i + j * Half] != Idx)
2687 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2688 if (VT.is64BitVector() && EltSz == 32)
2694 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2695 unsigned &WhichResult) {
2696 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2700 unsigned NumElts = VT.getVectorNumElements();
2701 WhichResult = (M[0] == 0 ? 0 : 1);
2702 unsigned Idx = WhichResult * NumElts / 2;
2703 for (unsigned i = 0; i != NumElts; i += 2) {
2704 if ((unsigned) M[i] != Idx ||
2705 (unsigned) M[i+1] != Idx + NumElts)
2710 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2711 if (VT.is64BitVector() && EltSz == 32)
2717 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2718 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2719 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2720 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2721 unsigned &WhichResult) {
2722 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2726 unsigned NumElts = VT.getVectorNumElements();
2727 WhichResult = (M[0] == 0 ? 0 : 1);
2728 unsigned Idx = WhichResult * NumElts / 2;
2729 for (unsigned i = 0; i != NumElts; i += 2) {
2730 if ((unsigned) M[i] != Idx ||
2731 (unsigned) M[i+1] != Idx)
2736 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2737 if (VT.is64BitVector() && EltSz == 32)
2744 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2745 // Canonicalize all-zeros and all-ones vectors.
2746 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2747 if (ConstVal->isNullValue())
2748 return getZeroVector(VT, DAG, dl);
2749 if (ConstVal->isAllOnesValue())
2750 return getOnesVector(VT, DAG, dl);
2753 if (VT.is64BitVector()) {
2754 switch (Val.getValueType().getSizeInBits()) {
2755 case 8: CanonicalVT = MVT::v8i8; break;
2756 case 16: CanonicalVT = MVT::v4i16; break;
2757 case 32: CanonicalVT = MVT::v2i32; break;
2758 case 64: CanonicalVT = MVT::v1i64; break;
2759 default: llvm_unreachable("unexpected splat element type"); break;
2762 assert(VT.is128BitVector() && "unknown splat vector size");
2763 switch (Val.getValueType().getSizeInBits()) {
2764 case 8: CanonicalVT = MVT::v16i8; break;
2765 case 16: CanonicalVT = MVT::v8i16; break;
2766 case 32: CanonicalVT = MVT::v4i32; break;
2767 case 64: CanonicalVT = MVT::v2i64; break;
2768 default: llvm_unreachable("unexpected splat element type"); break;
2772 // Build a canonical splat for this value.
2773 SmallVector<SDValue, 8> Ops;
2774 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2775 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2777 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2780 // If this is a case we can't handle, return null and let the default
2781 // expansion code take care of it.
2782 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2783 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2784 DebugLoc dl = Op.getDebugLoc();
2785 EVT VT = Op.getValueType();
2787 APInt SplatBits, SplatUndef;
2788 unsigned SplatBitSize;
2790 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2791 if (SplatBitSize <= 64) {
2792 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2793 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2795 return BuildSplat(Val, VT, DAG, dl);
2799 // If there are only 2 elements in a 128-bit vector, insert them into an
2800 // undef vector. This handles the common case for 128-bit vector argument
2801 // passing, where the insertions should be translated to subreg accesses
2802 // with no real instructions.
2803 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2804 SDValue Val = DAG.getUNDEF(VT);
2805 SDValue Op0 = Op.getOperand(0);
2806 SDValue Op1 = Op.getOperand(1);
2807 if (Op0.getOpcode() != ISD::UNDEF)
2808 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2809 DAG.getIntPtrConstant(0));
2810 if (Op1.getOpcode() != ISD::UNDEF)
2811 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2812 DAG.getIntPtrConstant(1));
2819 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2820 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2821 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2822 /// are assumed to be legal.
2824 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2826 if (VT.getVectorNumElements() == 4 &&
2827 (VT.is128BitVector() || VT.is64BitVector())) {
2828 unsigned PFIndexes[4];
2829 for (unsigned i = 0; i != 4; ++i) {
2833 PFIndexes[i] = M[i];
2836 // Compute the index in the perfect shuffle table.
2837 unsigned PFTableIndex =
2838 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2839 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2840 unsigned Cost = (PFEntry >> 30);
2847 unsigned Imm, WhichResult;
2849 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2850 isVREVMask(M, VT, 64) ||
2851 isVREVMask(M, VT, 32) ||
2852 isVREVMask(M, VT, 16) ||
2853 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2854 isVTRNMask(M, VT, WhichResult) ||
2855 isVUZPMask(M, VT, WhichResult) ||
2856 isVZIPMask(M, VT, WhichResult) ||
2857 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2858 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2859 isVZIP_v_undef_Mask(M, VT, WhichResult));
2862 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2863 /// the specified operations to build the shuffle.
2864 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2865 SDValue RHS, SelectionDAG &DAG,
2867 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2868 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2869 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2872 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2881 OP_VUZPL, // VUZP, left result
2882 OP_VUZPR, // VUZP, right result
2883 OP_VZIPL, // VZIP, left result
2884 OP_VZIPR, // VZIP, right result
2885 OP_VTRNL, // VTRN, left result
2886 OP_VTRNR // VTRN, right result
2889 if (OpNum == OP_COPY) {
2890 if (LHSID == (1*9+2)*9+3) return LHS;
2891 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2895 SDValue OpLHS, OpRHS;
2896 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2897 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2898 EVT VT = OpLHS.getValueType();
2901 default: llvm_unreachable("Unknown shuffle opcode!");
2903 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2908 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2909 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2913 return DAG.getNode(ARMISD::VEXT, dl, VT,
2915 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2918 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2919 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2922 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2923 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2926 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2927 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2931 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2932 SDValue V1 = Op.getOperand(0);
2933 SDValue V2 = Op.getOperand(1);
2934 DebugLoc dl = Op.getDebugLoc();
2935 EVT VT = Op.getValueType();
2936 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2937 SmallVector<int, 8> ShuffleMask;
2939 // Convert shuffles that are directly supported on NEON to target-specific
2940 // DAG nodes, instead of keeping them as shuffles and matching them again
2941 // during code selection. This is more efficient and avoids the possibility
2942 // of inconsistencies between legalization and selection.
2943 // FIXME: floating-point vectors should be canonicalized to integer vectors
2944 // of the same time so that they get CSEd properly.
2945 SVN->getMask(ShuffleMask);
2947 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2948 int Lane = SVN->getSplatIndex();
2949 // If this is undef splat, generate it via "just" vdup, if possible.
2950 if (Lane == -1) Lane = 0;
2952 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2953 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2955 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2956 DAG.getConstant(Lane, MVT::i32));
2961 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2964 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2965 DAG.getConstant(Imm, MVT::i32));
2968 if (isVREVMask(ShuffleMask, VT, 64))
2969 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2970 if (isVREVMask(ShuffleMask, VT, 32))
2971 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2972 if (isVREVMask(ShuffleMask, VT, 16))
2973 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2975 // Check for Neon shuffles that modify both input vectors in place.
2976 // If both results are used, i.e., if there are two shuffles with the same
2977 // source operands and with masks corresponding to both results of one of
2978 // these operations, DAG memoization will ensure that a single node is
2979 // used for both shuffles.
2980 unsigned WhichResult;
2981 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2982 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2983 V1, V2).getValue(WhichResult);
2984 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2985 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2986 V1, V2).getValue(WhichResult);
2987 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2988 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2989 V1, V2).getValue(WhichResult);
2991 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
2992 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2993 V1, V1).getValue(WhichResult);
2994 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2995 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2996 V1, V1).getValue(WhichResult);
2997 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2998 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2999 V1, V1).getValue(WhichResult);
3001 // If the shuffle is not directly supported and it has 4 elements, use
3002 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3003 if (VT.getVectorNumElements() == 4 &&
3004 (VT.is128BitVector() || VT.is64BitVector())) {
3005 unsigned PFIndexes[4];
3006 for (unsigned i = 0; i != 4; ++i) {
3007 if (ShuffleMask[i] < 0)
3010 PFIndexes[i] = ShuffleMask[i];
3013 // Compute the index in the perfect shuffle table.
3014 unsigned PFTableIndex =
3015 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3017 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3018 unsigned Cost = (PFEntry >> 30);
3021 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3027 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3028 EVT VT = Op.getValueType();
3029 DebugLoc dl = Op.getDebugLoc();
3030 SDValue Vec = Op.getOperand(0);
3031 SDValue Lane = Op.getOperand(1);
3032 assert(VT == MVT::i32 &&
3033 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3034 "unexpected type for custom-lowering vector extract");
3035 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3038 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3039 // The only time a CONCAT_VECTORS operation can have legal types is when
3040 // two 64-bit vectors are concatenated to a 128-bit vector.
3041 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3042 "unexpected CONCAT_VECTORS");
3043 DebugLoc dl = Op.getDebugLoc();
3044 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3045 SDValue Op0 = Op.getOperand(0);
3046 SDValue Op1 = Op.getOperand(1);
3047 if (Op0.getOpcode() != ISD::UNDEF)
3048 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3049 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3050 DAG.getIntPtrConstant(0));
3051 if (Op1.getOpcode() != ISD::UNDEF)
3052 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3053 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3054 DAG.getIntPtrConstant(1));
3055 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3058 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3059 switch (Op.getOpcode()) {
3060 default: llvm_unreachable("Don't know how to custom lower this!");
3061 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3062 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3063 case ISD::GlobalAddress:
3064 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3065 LowerGlobalAddressELF(Op, DAG);
3066 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3067 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3068 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3069 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3070 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3071 case ISD::VASTART: return LowerVASTART(Op, DAG);
3072 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3073 case ISD::SINT_TO_FP:
3074 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3075 case ISD::FP_TO_SINT:
3076 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3077 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3078 case ISD::RETURNADDR: break;
3079 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3080 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3081 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3083 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3086 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3087 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3088 case ISD::SRL_PARTS:
3089 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3090 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3091 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3092 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3093 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3094 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3095 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3100 /// ReplaceNodeResults - Replace the results of node with an illegal result
3101 /// type with new values built out of custom code.
3102 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3103 SmallVectorImpl<SDValue>&Results,
3104 SelectionDAG &DAG) const {
3106 switch (N->getOpcode()) {
3108 llvm_unreachable("Don't know how to custom expand this!");
3110 case ISD::BIT_CONVERT:
3111 Res = ExpandBIT_CONVERT(N, DAG);
3115 Res = LowerShift(N, DAG, Subtarget);
3119 Results.push_back(Res);
3122 //===----------------------------------------------------------------------===//
3123 // ARM Scheduler Hooks
3124 //===----------------------------------------------------------------------===//
3127 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3128 MachineBasicBlock *BB,
3129 unsigned Size) const {
3130 unsigned dest = MI->getOperand(0).getReg();
3131 unsigned ptr = MI->getOperand(1).getReg();
3132 unsigned oldval = MI->getOperand(2).getReg();
3133 unsigned newval = MI->getOperand(3).getReg();
3134 unsigned scratch = BB->getParent()->getRegInfo()
3135 .createVirtualRegister(ARM::GPRRegisterClass);
3136 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3137 DebugLoc dl = MI->getDebugLoc();
3138 bool isThumb2 = Subtarget->isThumb2();
3140 unsigned ldrOpc, strOpc;
3142 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3144 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3145 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3148 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3149 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3152 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3153 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3157 MachineFunction *MF = BB->getParent();
3158 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3159 MachineFunction::iterator It = BB;
3160 ++It; // insert the new blocks after the current block
3162 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3163 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3164 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3165 MF->insert(It, loop1MBB);
3166 MF->insert(It, loop2MBB);
3167 MF->insert(It, exitMBB);
3168 exitMBB->transferSuccessors(BB);
3172 // fallthrough --> loop1MBB
3173 BB->addSuccessor(loop1MBB);
3176 // ldrex dest, [ptr]
3180 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3181 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3182 .addReg(dest).addReg(oldval));
3183 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3184 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3185 BB->addSuccessor(loop2MBB);
3186 BB->addSuccessor(exitMBB);
3189 // strex scratch, newval, [ptr]
3193 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3195 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3196 .addReg(scratch).addImm(0));
3197 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3198 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3199 BB->addSuccessor(loop1MBB);
3200 BB->addSuccessor(exitMBB);
3206 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3212 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3213 unsigned Size, unsigned BinOpcode) const {
3214 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3215 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3217 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3218 MachineFunction *MF = BB->getParent();
3219 MachineFunction::iterator It = BB;
3222 unsigned dest = MI->getOperand(0).getReg();
3223 unsigned ptr = MI->getOperand(1).getReg();
3224 unsigned incr = MI->getOperand(2).getReg();
3225 DebugLoc dl = MI->getDebugLoc();
3227 bool isThumb2 = Subtarget->isThumb2();
3228 unsigned ldrOpc, strOpc;
3230 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3232 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3233 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3236 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3237 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3240 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3241 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3245 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3246 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3247 MF->insert(It, loopMBB);
3248 MF->insert(It, exitMBB);
3249 exitMBB->transferSuccessors(BB);
3251 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3252 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3253 unsigned scratch2 = (!BinOpcode) ? incr :
3254 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3258 // fallthrough --> loopMBB
3259 BB->addSuccessor(loopMBB);
3263 // <binop> scratch2, dest, incr
3264 // strex scratch, scratch2, ptr
3267 // fallthrough --> exitMBB
3269 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3271 // operand order needs to go the other way for NAND
3272 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3273 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3274 addReg(incr).addReg(dest)).addReg(0);
3276 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3277 addReg(dest).addReg(incr)).addReg(0);
3280 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3282 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3283 .addReg(scratch).addImm(0));
3284 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3285 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3287 BB->addSuccessor(loopMBB);
3288 BB->addSuccessor(exitMBB);
3294 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3300 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3301 MachineBasicBlock *BB) const {
3302 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3303 DebugLoc dl = MI->getDebugLoc();
3304 bool isThumb2 = Subtarget->isThumb2();
3305 switch (MI->getOpcode()) {
3308 llvm_unreachable("Unexpected instr type to insert");
3310 case ARM::ATOMIC_LOAD_ADD_I8:
3311 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3312 case ARM::ATOMIC_LOAD_ADD_I16:
3313 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3314 case ARM::ATOMIC_LOAD_ADD_I32:
3315 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3317 case ARM::ATOMIC_LOAD_AND_I8:
3318 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3319 case ARM::ATOMIC_LOAD_AND_I16:
3320 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3321 case ARM::ATOMIC_LOAD_AND_I32:
3322 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3324 case ARM::ATOMIC_LOAD_OR_I8:
3325 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3326 case ARM::ATOMIC_LOAD_OR_I16:
3327 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3328 case ARM::ATOMIC_LOAD_OR_I32:
3329 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3331 case ARM::ATOMIC_LOAD_XOR_I8:
3332 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3333 case ARM::ATOMIC_LOAD_XOR_I16:
3334 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3335 case ARM::ATOMIC_LOAD_XOR_I32:
3336 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3338 case ARM::ATOMIC_LOAD_NAND_I8:
3339 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3340 case ARM::ATOMIC_LOAD_NAND_I16:
3341 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3342 case ARM::ATOMIC_LOAD_NAND_I32:
3343 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3345 case ARM::ATOMIC_LOAD_SUB_I8:
3346 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3347 case ARM::ATOMIC_LOAD_SUB_I16:
3348 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3349 case ARM::ATOMIC_LOAD_SUB_I32:
3350 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3352 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3353 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3354 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3356 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3357 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3358 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3360 case ARM::tMOVCCr_pseudo: {
3361 // To "insert" a SELECT_CC instruction, we actually have to insert the
3362 // diamond control-flow pattern. The incoming instruction knows the
3363 // destination vreg to set, the condition code register to branch on, the
3364 // true/false values to select between, and a branch opcode to use.
3365 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3366 MachineFunction::iterator It = BB;
3372 // cmpTY ccX, r1, r2
3374 // fallthrough --> copy0MBB
3375 MachineBasicBlock *thisMBB = BB;
3376 MachineFunction *F = BB->getParent();
3377 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3378 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3379 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3380 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3381 F->insert(It, copy0MBB);
3382 F->insert(It, sinkMBB);
3383 // Update machine-CFG edges by first adding all successors of the current
3384 // block to the new block which will contain the Phi node for the select.
3385 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3386 E = BB->succ_end(); I != E; ++I)
3387 sinkMBB->addSuccessor(*I);
3388 // Next, remove all successors of the current block, and add the true
3389 // and fallthrough blocks as its successors.
3390 while (!BB->succ_empty())
3391 BB->removeSuccessor(BB->succ_begin());
3392 BB->addSuccessor(copy0MBB);
3393 BB->addSuccessor(sinkMBB);
3396 // %FalseValue = ...
3397 // # fallthrough to sinkMBB
3400 // Update machine-CFG edges
3401 BB->addSuccessor(sinkMBB);
3404 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3407 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3408 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3409 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3411 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3418 case ARM::t2SUBrSPi_:
3419 case ARM::t2SUBrSPi12_:
3420 case ARM::t2SUBrSPs_: {
3421 MachineFunction *MF = BB->getParent();
3422 unsigned DstReg = MI->getOperand(0).getReg();
3423 unsigned SrcReg = MI->getOperand(1).getReg();
3424 bool DstIsDead = MI->getOperand(0).isDead();
3425 bool SrcIsKill = MI->getOperand(1).isKill();
3427 if (SrcReg != ARM::SP) {
3428 // Copy the source to SP from virtual register.
3429 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3430 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3431 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3432 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3433 .addReg(SrcReg, getKillRegState(SrcIsKill));
3437 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3438 switch (MI->getOpcode()) {
3440 llvm_unreachable("Unexpected pseudo instruction!");
3446 OpOpc = ARM::tADDspr;
3449 OpOpc = ARM::tSUBspi;
3451 case ARM::t2SUBrSPi_:
3452 OpOpc = ARM::t2SUBrSPi;
3453 NeedPred = true; NeedCC = true;
3455 case ARM::t2SUBrSPi12_:
3456 OpOpc = ARM::t2SUBrSPi12;
3459 case ARM::t2SUBrSPs_:
3460 OpOpc = ARM::t2SUBrSPs;
3461 NeedPred = true; NeedCC = true; NeedOp3 = true;
3464 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3465 if (OpOpc == ARM::tAND)
3466 AddDefaultT1CC(MIB);
3467 MIB.addReg(ARM::SP);
3468 MIB.addOperand(MI->getOperand(2));
3470 MIB.addOperand(MI->getOperand(3));
3472 AddDefaultPred(MIB);
3476 // Copy the result from SP to virtual register.
3477 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3478 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3479 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3480 BuildMI(BB, dl, TII->get(CopyOpc))
3481 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3483 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3489 //===----------------------------------------------------------------------===//
3490 // ARM Optimization Hooks
3491 //===----------------------------------------------------------------------===//
3494 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3495 TargetLowering::DAGCombinerInfo &DCI) {
3496 SelectionDAG &DAG = DCI.DAG;
3497 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3498 EVT VT = N->getValueType(0);
3499 unsigned Opc = N->getOpcode();
3500 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3501 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3502 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3503 ISD::CondCode CC = ISD::SETCC_INVALID;
3506 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3508 SDValue CCOp = Slct.getOperand(0);
3509 if (CCOp.getOpcode() == ISD::SETCC)
3510 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3513 bool DoXform = false;
3515 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3518 if (LHS.getOpcode() == ISD::Constant &&
3519 cast<ConstantSDNode>(LHS)->isNullValue()) {
3521 } else if (CC != ISD::SETCC_INVALID &&
3522 RHS.getOpcode() == ISD::Constant &&
3523 cast<ConstantSDNode>(RHS)->isNullValue()) {
3524 std::swap(LHS, RHS);
3525 SDValue Op0 = Slct.getOperand(0);
3526 EVT OpVT = isSlctCC ? Op0.getValueType() :
3527 Op0.getOperand(0).getValueType();
3528 bool isInt = OpVT.isInteger();
3529 CC = ISD::getSetCCInverse(CC, isInt);
3531 if (!TLI.isCondCodeLegal(CC, OpVT))
3532 return SDValue(); // Inverse operator isn't legal.
3539 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3541 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3542 Slct.getOperand(0), Slct.getOperand(1), CC);
3543 SDValue CCOp = Slct.getOperand(0);
3545 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3546 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3547 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3548 CCOp, OtherOp, Result);
3553 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3554 static SDValue PerformADDCombine(SDNode *N,
3555 TargetLowering::DAGCombinerInfo &DCI) {
3556 // added by evan in r37685 with no testcase.
3557 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3559 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3560 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3561 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3562 if (Result.getNode()) return Result;
3564 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3565 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3566 if (Result.getNode()) return Result;
3572 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3573 static SDValue PerformSUBCombine(SDNode *N,
3574 TargetLowering::DAGCombinerInfo &DCI) {
3575 // added by evan in r37685 with no testcase.
3576 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3578 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3579 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3580 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3581 if (Result.getNode()) return Result;
3587 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3588 /// ARMISD::VMOVRRD.
3589 static SDValue PerformVMOVRRDCombine(SDNode *N,
3590 TargetLowering::DAGCombinerInfo &DCI) {
3591 // fmrrd(fmdrr x, y) -> x,y
3592 SDValue InDouble = N->getOperand(0);
3593 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3594 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3598 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3599 /// operand of a vector shift operation, where all the elements of the
3600 /// build_vector must have the same constant integer value.
3601 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3602 // Ignore bit_converts.
3603 while (Op.getOpcode() == ISD::BIT_CONVERT)
3604 Op = Op.getOperand(0);
3605 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3606 APInt SplatBits, SplatUndef;
3607 unsigned SplatBitSize;
3609 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3610 HasAnyUndefs, ElementBits) ||
3611 SplatBitSize > ElementBits)
3613 Cnt = SplatBits.getSExtValue();
3617 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3618 /// operand of a vector shift left operation. That value must be in the range:
3619 /// 0 <= Value < ElementBits for a left shift; or
3620 /// 0 <= Value <= ElementBits for a long left shift.
3621 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3622 assert(VT.isVector() && "vector shift count is not a vector type");
3623 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3624 if (! getVShiftImm(Op, ElementBits, Cnt))
3626 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3629 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3630 /// operand of a vector shift right operation. For a shift opcode, the value
3631 /// is positive, but for an intrinsic the value count must be negative. The
3632 /// absolute value must be in the range:
3633 /// 1 <= |Value| <= ElementBits for a right shift; or
3634 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3635 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3637 assert(VT.isVector() && "vector shift count is not a vector type");
3638 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3639 if (! getVShiftImm(Op, ElementBits, Cnt))
3643 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3646 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3647 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3648 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3651 // Don't do anything for most intrinsics.
3654 // Vector shifts: check for immediate versions and lower them.
3655 // Note: This is done during DAG combining instead of DAG legalizing because
3656 // the build_vectors for 64-bit vector element shift counts are generally
3657 // not legal, and it is hard to see their values after they get legalized to
3658 // loads from a constant pool.
3659 case Intrinsic::arm_neon_vshifts:
3660 case Intrinsic::arm_neon_vshiftu:
3661 case Intrinsic::arm_neon_vshiftls:
3662 case Intrinsic::arm_neon_vshiftlu:
3663 case Intrinsic::arm_neon_vshiftn:
3664 case Intrinsic::arm_neon_vrshifts:
3665 case Intrinsic::arm_neon_vrshiftu:
3666 case Intrinsic::arm_neon_vrshiftn:
3667 case Intrinsic::arm_neon_vqshifts:
3668 case Intrinsic::arm_neon_vqshiftu:
3669 case Intrinsic::arm_neon_vqshiftsu:
3670 case Intrinsic::arm_neon_vqshiftns:
3671 case Intrinsic::arm_neon_vqshiftnu:
3672 case Intrinsic::arm_neon_vqshiftnsu:
3673 case Intrinsic::arm_neon_vqrshiftns:
3674 case Intrinsic::arm_neon_vqrshiftnu:
3675 case Intrinsic::arm_neon_vqrshiftnsu: {
3676 EVT VT = N->getOperand(1).getValueType();
3678 unsigned VShiftOpc = 0;
3681 case Intrinsic::arm_neon_vshifts:
3682 case Intrinsic::arm_neon_vshiftu:
3683 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3684 VShiftOpc = ARMISD::VSHL;
3687 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3688 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3689 ARMISD::VSHRs : ARMISD::VSHRu);
3694 case Intrinsic::arm_neon_vshiftls:
3695 case Intrinsic::arm_neon_vshiftlu:
3696 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3698 llvm_unreachable("invalid shift count for vshll intrinsic");
3700 case Intrinsic::arm_neon_vrshifts:
3701 case Intrinsic::arm_neon_vrshiftu:
3702 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3706 case Intrinsic::arm_neon_vqshifts:
3707 case Intrinsic::arm_neon_vqshiftu:
3708 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3712 case Intrinsic::arm_neon_vqshiftsu:
3713 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3715 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3717 case Intrinsic::arm_neon_vshiftn:
3718 case Intrinsic::arm_neon_vrshiftn:
3719 case Intrinsic::arm_neon_vqshiftns:
3720 case Intrinsic::arm_neon_vqshiftnu:
3721 case Intrinsic::arm_neon_vqshiftnsu:
3722 case Intrinsic::arm_neon_vqrshiftns:
3723 case Intrinsic::arm_neon_vqrshiftnu:
3724 case Intrinsic::arm_neon_vqrshiftnsu:
3725 // Narrowing shifts require an immediate right shift.
3726 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3728 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3731 llvm_unreachable("unhandled vector shift");
3735 case Intrinsic::arm_neon_vshifts:
3736 case Intrinsic::arm_neon_vshiftu:
3737 // Opcode already set above.
3739 case Intrinsic::arm_neon_vshiftls:
3740 case Intrinsic::arm_neon_vshiftlu:
3741 if (Cnt == VT.getVectorElementType().getSizeInBits())
3742 VShiftOpc = ARMISD::VSHLLi;
3744 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3745 ARMISD::VSHLLs : ARMISD::VSHLLu);
3747 case Intrinsic::arm_neon_vshiftn:
3748 VShiftOpc = ARMISD::VSHRN; break;
3749 case Intrinsic::arm_neon_vrshifts:
3750 VShiftOpc = ARMISD::VRSHRs; break;
3751 case Intrinsic::arm_neon_vrshiftu:
3752 VShiftOpc = ARMISD::VRSHRu; break;
3753 case Intrinsic::arm_neon_vrshiftn:
3754 VShiftOpc = ARMISD::VRSHRN; break;
3755 case Intrinsic::arm_neon_vqshifts:
3756 VShiftOpc = ARMISD::VQSHLs; break;
3757 case Intrinsic::arm_neon_vqshiftu:
3758 VShiftOpc = ARMISD::VQSHLu; break;
3759 case Intrinsic::arm_neon_vqshiftsu:
3760 VShiftOpc = ARMISD::VQSHLsu; break;
3761 case Intrinsic::arm_neon_vqshiftns:
3762 VShiftOpc = ARMISD::VQSHRNs; break;
3763 case Intrinsic::arm_neon_vqshiftnu:
3764 VShiftOpc = ARMISD::VQSHRNu; break;
3765 case Intrinsic::arm_neon_vqshiftnsu:
3766 VShiftOpc = ARMISD::VQSHRNsu; break;
3767 case Intrinsic::arm_neon_vqrshiftns:
3768 VShiftOpc = ARMISD::VQRSHRNs; break;
3769 case Intrinsic::arm_neon_vqrshiftnu:
3770 VShiftOpc = ARMISD::VQRSHRNu; break;
3771 case Intrinsic::arm_neon_vqrshiftnsu:
3772 VShiftOpc = ARMISD::VQRSHRNsu; break;
3775 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3776 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3779 case Intrinsic::arm_neon_vshiftins: {
3780 EVT VT = N->getOperand(1).getValueType();
3782 unsigned VShiftOpc = 0;
3784 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3785 VShiftOpc = ARMISD::VSLI;
3786 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3787 VShiftOpc = ARMISD::VSRI;
3789 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3792 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3793 N->getOperand(1), N->getOperand(2),
3794 DAG.getConstant(Cnt, MVT::i32));
3797 case Intrinsic::arm_neon_vqrshifts:
3798 case Intrinsic::arm_neon_vqrshiftu:
3799 // No immediate versions of these to check for.
3806 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3807 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3808 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3809 /// vector element shift counts are generally not legal, and it is hard to see
3810 /// their values after they get legalized to loads from a constant pool.
3811 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3812 const ARMSubtarget *ST) {
3813 EVT VT = N->getValueType(0);
3815 // Nothing to be done for scalar shifts.
3816 if (! VT.isVector())
3819 assert(ST->hasNEON() && "unexpected vector shift");
3822 switch (N->getOpcode()) {
3823 default: llvm_unreachable("unexpected shift opcode");
3826 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3827 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3828 DAG.getConstant(Cnt, MVT::i32));
3833 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3834 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3835 ARMISD::VSHRs : ARMISD::VSHRu);
3836 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3837 DAG.getConstant(Cnt, MVT::i32));
3843 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3844 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3845 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3846 const ARMSubtarget *ST) {
3847 SDValue N0 = N->getOperand(0);
3849 // Check for sign- and zero-extensions of vector extract operations of 8-
3850 // and 16-bit vector elements. NEON supports these directly. They are
3851 // handled during DAG combining because type legalization will promote them
3852 // to 32-bit types and it is messy to recognize the operations after that.
3853 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3854 SDValue Vec = N0.getOperand(0);
3855 SDValue Lane = N0.getOperand(1);
3856 EVT VT = N->getValueType(0);
3857 EVT EltVT = N0.getValueType();
3858 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3860 if (VT == MVT::i32 &&
3861 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3862 TLI.isTypeLegal(Vec.getValueType())) {
3865 switch (N->getOpcode()) {
3866 default: llvm_unreachable("unexpected opcode");
3867 case ISD::SIGN_EXTEND:
3868 Opc = ARMISD::VGETLANEs;
3870 case ISD::ZERO_EXTEND:
3871 case ISD::ANY_EXTEND:
3872 Opc = ARMISD::VGETLANEu;
3875 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3882 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3883 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3884 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3885 const ARMSubtarget *ST) {
3886 // If the target supports NEON, try to use vmax/vmin instructions for f32
3887 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3888 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3889 // a NaN; only do the transformation when it matches that behavior.
3891 // For now only do this when using NEON for FP operations; if using VFP, it
3892 // is not obvious that the benefit outweighs the cost of switching to the
3894 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3895 N->getValueType(0) != MVT::f32)
3898 SDValue CondLHS = N->getOperand(0);
3899 SDValue CondRHS = N->getOperand(1);
3900 SDValue LHS = N->getOperand(2);
3901 SDValue RHS = N->getOperand(3);
3902 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3904 unsigned Opcode = 0;
3906 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3907 IsReversed = false; // x CC y ? x : y
3908 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3909 IsReversed = true ; // x CC y ? y : x
3923 // If LHS is NaN, an ordered comparison will be false and the result will
3924 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3925 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3926 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3927 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3929 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3930 // will return -0, so vmin can only be used for unsafe math or if one of
3931 // the operands is known to be nonzero.
3932 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3934 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3936 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3945 // If LHS is NaN, an ordered comparison will be false and the result will
3946 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3947 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3948 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3949 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3951 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3952 // will return +0, so vmax can only be used for unsafe math or if one of
3953 // the operands is known to be nonzero.
3954 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3956 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3958 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3964 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3967 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3968 DAGCombinerInfo &DCI) const {
3969 switch (N->getOpcode()) {
3971 case ISD::ADD: return PerformADDCombine(N, DCI);
3972 case ISD::SUB: return PerformSUBCombine(N, DCI);
3973 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3974 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
3977 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
3978 case ISD::SIGN_EXTEND:
3979 case ISD::ZERO_EXTEND:
3980 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
3981 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
3986 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3987 if (!Subtarget->hasV6Ops())
3988 // Pre-v6 does not support unaligned mem access.
3991 // v6+ may or may not support unaligned mem access depending on the system
3993 // FIXME: This is pretty conservative. Should we provide cmdline option to
3994 // control the behaviour?
3995 if (!Subtarget->isTargetDarwin())
3999 switch (VT.getSimpleVT().SimpleTy) {
4006 // FIXME: VLD1 etc with standard alignment is legal.
4010 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4015 switch (VT.getSimpleVT().SimpleTy) {
4016 default: return false;
4031 if ((V & (Scale - 1)) != 0)
4034 return V == (V & ((1LL << 5) - 1));
4037 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4038 const ARMSubtarget *Subtarget) {
4045 switch (VT.getSimpleVT().SimpleTy) {
4046 default: return false;
4051 // + imm12 or - imm8
4053 return V == (V & ((1LL << 8) - 1));
4054 return V == (V & ((1LL << 12) - 1));
4057 // Same as ARM mode. FIXME: NEON?
4058 if (!Subtarget->hasVFP2())
4063 return V == (V & ((1LL << 8) - 1));
4067 /// isLegalAddressImmediate - Return true if the integer value can be used
4068 /// as the offset of the target addressing mode for load / store of the
4070 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4071 const ARMSubtarget *Subtarget) {
4078 if (Subtarget->isThumb1Only())
4079 return isLegalT1AddressImmediate(V, VT);
4080 else if (Subtarget->isThumb2())
4081 return isLegalT2AddressImmediate(V, VT, Subtarget);
4086 switch (VT.getSimpleVT().SimpleTy) {
4087 default: return false;
4092 return V == (V & ((1LL << 12) - 1));
4095 return V == (V & ((1LL << 8) - 1));
4098 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4103 return V == (V & ((1LL << 8) - 1));
4107 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4109 int Scale = AM.Scale;
4113 switch (VT.getSimpleVT().SimpleTy) {
4114 default: return false;
4123 return Scale == 2 || Scale == 4 || Scale == 8;
4126 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4130 // Note, we allow "void" uses (basically, uses that aren't loads or
4131 // stores), because arm allows folding a scale into many arithmetic
4132 // operations. This should be made more precise and revisited later.
4134 // Allow r << imm, but the imm has to be a multiple of two.
4135 if (Scale & 1) return false;
4136 return isPowerOf2_32(Scale);
4140 /// isLegalAddressingMode - Return true if the addressing mode represented
4141 /// by AM is legal for this target, for a load/store of the specified type.
4142 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4143 const Type *Ty) const {
4144 EVT VT = getValueType(Ty, true);
4145 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4148 // Can never fold addr of global into load/store.
4153 case 0: // no scale reg, must be "r+i" or "r", or "i".
4156 if (Subtarget->isThumb1Only())
4160 // ARM doesn't support any R+R*scale+imm addr modes.
4167 if (Subtarget->isThumb2())
4168 return isLegalT2ScaledAddressingMode(AM, VT);
4170 int Scale = AM.Scale;
4171 switch (VT.getSimpleVT().SimpleTy) {
4172 default: return false;
4176 if (Scale < 0) Scale = -Scale;
4180 return isPowerOf2_32(Scale & ~1);
4184 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4189 // Note, we allow "void" uses (basically, uses that aren't loads or
4190 // stores), because arm allows folding a scale into many arithmetic
4191 // operations. This should be made more precise and revisited later.
4193 // Allow r << imm, but the imm has to be a multiple of two.
4194 if (Scale & 1) return false;
4195 return isPowerOf2_32(Scale);
4202 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4203 /// icmp immediate, that is the target has icmp instructions which can compare
4204 /// a register against the immediate without having to materialize the
4205 /// immediate into a register.
4206 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4207 if (!Subtarget->isThumb())
4208 return ARM_AM::getSOImmVal(Imm) != -1;
4209 if (Subtarget->isThumb2())
4210 return ARM_AM::getT2SOImmVal(Imm) != -1;
4211 return Imm >= 0 && Imm <= 255;
4214 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4215 bool isSEXTLoad, SDValue &Base,
4216 SDValue &Offset, bool &isInc,
4217 SelectionDAG &DAG) {
4218 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4221 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4223 Base = Ptr->getOperand(0);
4224 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4225 int RHSC = (int)RHS->getZExtValue();
4226 if (RHSC < 0 && RHSC > -256) {
4227 assert(Ptr->getOpcode() == ISD::ADD);
4229 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4233 isInc = (Ptr->getOpcode() == ISD::ADD);
4234 Offset = Ptr->getOperand(1);
4236 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4238 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4239 int RHSC = (int)RHS->getZExtValue();
4240 if (RHSC < 0 && RHSC > -0x1000) {
4241 assert(Ptr->getOpcode() == ISD::ADD);
4243 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4244 Base = Ptr->getOperand(0);
4249 if (Ptr->getOpcode() == ISD::ADD) {
4251 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4252 if (ShOpcVal != ARM_AM::no_shift) {
4253 Base = Ptr->getOperand(1);
4254 Offset = Ptr->getOperand(0);
4256 Base = Ptr->getOperand(0);
4257 Offset = Ptr->getOperand(1);
4262 isInc = (Ptr->getOpcode() == ISD::ADD);
4263 Base = Ptr->getOperand(0);
4264 Offset = Ptr->getOperand(1);
4268 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4272 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4273 bool isSEXTLoad, SDValue &Base,
4274 SDValue &Offset, bool &isInc,
4275 SelectionDAG &DAG) {
4276 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4279 Base = Ptr->getOperand(0);
4280 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4281 int RHSC = (int)RHS->getZExtValue();
4282 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4283 assert(Ptr->getOpcode() == ISD::ADD);
4285 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4287 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4288 isInc = Ptr->getOpcode() == ISD::ADD;
4289 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4297 /// getPreIndexedAddressParts - returns true by value, base pointer and
4298 /// offset pointer and addressing mode by reference if the node's address
4299 /// can be legally represented as pre-indexed load / store address.
4301 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4303 ISD::MemIndexedMode &AM,
4304 SelectionDAG &DAG) const {
4305 if (Subtarget->isThumb1Only())
4310 bool isSEXTLoad = false;
4311 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4312 Ptr = LD->getBasePtr();
4313 VT = LD->getMemoryVT();
4314 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4315 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4316 Ptr = ST->getBasePtr();
4317 VT = ST->getMemoryVT();
4322 bool isLegal = false;
4323 if (Subtarget->isThumb2())
4324 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4325 Offset, isInc, DAG);
4327 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4328 Offset, isInc, DAG);
4332 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4336 /// getPostIndexedAddressParts - returns true by value, base pointer and
4337 /// offset pointer and addressing mode by reference if this node can be
4338 /// combined with a load / store to form a post-indexed load / store.
4339 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4342 ISD::MemIndexedMode &AM,
4343 SelectionDAG &DAG) const {
4344 if (Subtarget->isThumb1Only())
4349 bool isSEXTLoad = false;
4350 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4351 VT = LD->getMemoryVT();
4352 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4353 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4354 VT = ST->getMemoryVT();
4359 bool isLegal = false;
4360 if (Subtarget->isThumb2())
4361 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4364 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4369 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4373 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4377 const SelectionDAG &DAG,
4378 unsigned Depth) const {
4379 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4380 switch (Op.getOpcode()) {
4382 case ARMISD::CMOV: {
4383 // Bits are known zero/one if known on the LHS and RHS.
4384 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4385 if (KnownZero == 0 && KnownOne == 0) return;
4387 APInt KnownZeroRHS, KnownOneRHS;
4388 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4389 KnownZeroRHS, KnownOneRHS, Depth+1);
4390 KnownZero &= KnownZeroRHS;
4391 KnownOne &= KnownOneRHS;
4397 //===----------------------------------------------------------------------===//
4398 // ARM Inline Assembly Support
4399 //===----------------------------------------------------------------------===//
4401 /// getConstraintType - Given a constraint letter, return the type of
4402 /// constraint it is for this target.
4403 ARMTargetLowering::ConstraintType
4404 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4405 if (Constraint.size() == 1) {
4406 switch (Constraint[0]) {
4408 case 'l': return C_RegisterClass;
4409 case 'w': return C_RegisterClass;
4412 return TargetLowering::getConstraintType(Constraint);
4415 std::pair<unsigned, const TargetRegisterClass*>
4416 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4418 if (Constraint.size() == 1) {
4419 // GCC ARM Constraint Letters
4420 switch (Constraint[0]) {
4422 if (Subtarget->isThumb())
4423 return std::make_pair(0U, ARM::tGPRRegisterClass);
4425 return std::make_pair(0U, ARM::GPRRegisterClass);
4427 return std::make_pair(0U, ARM::GPRRegisterClass);
4430 return std::make_pair(0U, ARM::SPRRegisterClass);
4431 if (VT.getSizeInBits() == 64)
4432 return std::make_pair(0U, ARM::DPRRegisterClass);
4433 if (VT.getSizeInBits() == 128)
4434 return std::make_pair(0U, ARM::QPRRegisterClass);
4438 if (StringRef("{cc}").equals_lower(Constraint))
4439 return std::make_pair(0U, ARM::CCRRegisterClass);
4441 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4444 std::vector<unsigned> ARMTargetLowering::
4445 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4447 if (Constraint.size() != 1)
4448 return std::vector<unsigned>();
4450 switch (Constraint[0]) { // GCC ARM Constraint Letters
4453 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4454 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4457 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4458 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4459 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4460 ARM::R12, ARM::LR, 0);
4463 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4464 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4465 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4466 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4467 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4468 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4469 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4470 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4471 if (VT.getSizeInBits() == 64)
4472 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4473 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4474 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4475 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4476 if (VT.getSizeInBits() == 128)
4477 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4478 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4482 return std::vector<unsigned>();
4485 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4486 /// vector. If it is invalid, don't add anything to Ops.
4487 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4490 std::vector<SDValue>&Ops,
4491 SelectionDAG &DAG) const {
4492 SDValue Result(0, 0);
4494 switch (Constraint) {
4496 case 'I': case 'J': case 'K': case 'L':
4497 case 'M': case 'N': case 'O':
4498 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4502 int64_t CVal64 = C->getSExtValue();
4503 int CVal = (int) CVal64;
4504 // None of these constraints allow values larger than 32 bits. Check
4505 // that the value fits in an int.
4509 switch (Constraint) {
4511 if (Subtarget->isThumb1Only()) {
4512 // This must be a constant between 0 and 255, for ADD
4514 if (CVal >= 0 && CVal <= 255)
4516 } else if (Subtarget->isThumb2()) {
4517 // A constant that can be used as an immediate value in a
4518 // data-processing instruction.
4519 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4522 // A constant that can be used as an immediate value in a
4523 // data-processing instruction.
4524 if (ARM_AM::getSOImmVal(CVal) != -1)
4530 if (Subtarget->isThumb()) { // FIXME thumb2
4531 // This must be a constant between -255 and -1, for negated ADD
4532 // immediates. This can be used in GCC with an "n" modifier that
4533 // prints the negated value, for use with SUB instructions. It is
4534 // not useful otherwise but is implemented for compatibility.
4535 if (CVal >= -255 && CVal <= -1)
4538 // This must be a constant between -4095 and 4095. It is not clear
4539 // what this constraint is intended for. Implemented for
4540 // compatibility with GCC.
4541 if (CVal >= -4095 && CVal <= 4095)
4547 if (Subtarget->isThumb1Only()) {
4548 // A 32-bit value where only one byte has a nonzero value. Exclude
4549 // zero to match GCC. This constraint is used by GCC internally for
4550 // constants that can be loaded with a move/shift combination.
4551 // It is not useful otherwise but is implemented for compatibility.
4552 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4554 } else if (Subtarget->isThumb2()) {
4555 // A constant whose bitwise inverse can be used as an immediate
4556 // value in a data-processing instruction. This can be used in GCC
4557 // with a "B" modifier that prints the inverted value, for use with
4558 // BIC and MVN instructions. It is not useful otherwise but is
4559 // implemented for compatibility.
4560 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4563 // A constant whose bitwise inverse can be used as an immediate
4564 // value in a data-processing instruction. This can be used in GCC
4565 // with a "B" modifier that prints the inverted value, for use with
4566 // BIC and MVN instructions. It is not useful otherwise but is
4567 // implemented for compatibility.
4568 if (ARM_AM::getSOImmVal(~CVal) != -1)
4574 if (Subtarget->isThumb1Only()) {
4575 // This must be a constant between -7 and 7,
4576 // for 3-operand ADD/SUB immediate instructions.
4577 if (CVal >= -7 && CVal < 7)
4579 } else if (Subtarget->isThumb2()) {
4580 // A constant whose negation can be used as an immediate value in a
4581 // data-processing instruction. This can be used in GCC with an "n"
4582 // modifier that prints the negated value, for use with SUB
4583 // instructions. It is not useful otherwise but is implemented for
4585 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4588 // A constant whose negation can be used as an immediate value in a
4589 // data-processing instruction. This can be used in GCC with an "n"
4590 // modifier that prints the negated value, for use with SUB
4591 // instructions. It is not useful otherwise but is implemented for
4593 if (ARM_AM::getSOImmVal(-CVal) != -1)
4599 if (Subtarget->isThumb()) { // FIXME thumb2
4600 // This must be a multiple of 4 between 0 and 1020, for
4601 // ADD sp + immediate.
4602 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4605 // A power of two or a constant between 0 and 32. This is used in
4606 // GCC for the shift amount on shifted register operands, but it is
4607 // useful in general for any shift amounts.
4608 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4614 if (Subtarget->isThumb()) { // FIXME thumb2
4615 // This must be a constant between 0 and 31, for shift amounts.
4616 if (CVal >= 0 && CVal <= 31)
4622 if (Subtarget->isThumb()) { // FIXME thumb2
4623 // This must be a multiple of 4 between -508 and 508, for
4624 // ADD/SUB sp = sp + immediate.
4625 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4630 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4634 if (Result.getNode()) {
4635 Ops.push_back(Result);
4638 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4643 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4644 // The ARM target isn't yet aware of offsets.
4648 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4649 APInt Imm = FPImm.bitcastToAPInt();
4650 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4651 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4652 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4654 // We can handle 4 bits of mantissa.
4655 // mantissa = (16+UInt(e:f:g:h))/16.
4656 if (Mantissa & 0x7ffff)
4659 if ((Mantissa & 0xf) != Mantissa)
4662 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4663 if (Exp < -3 || Exp > 4)
4665 Exp = ((Exp+3) & 0x7) ^ 4;
4667 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4670 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4671 APInt Imm = FPImm.bitcastToAPInt();
4672 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4673 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4674 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4676 // We can handle 4 bits of mantissa.
4677 // mantissa = (16+UInt(e:f:g:h))/16.
4678 if (Mantissa & 0xffffffffffffLL)
4681 if ((Mantissa & 0xf) != Mantissa)
4684 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4685 if (Exp < -3 || Exp > 4)
4687 Exp = ((Exp+3) & 0x7) ^ 4;
4689 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4692 /// isFPImmLegal - Returns true if the target can instruction select the
4693 /// specified FP immediate natively. If false, the legalizer will
4694 /// materialize the FP immediate as a load from a constant pool.
4695 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4696 if (!Subtarget->hasVFP3())
4699 return ARM::getVFPf32Imm(Imm) != -1;
4701 return ARM::getVFPf64Imm(Imm) != -1;