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 // Map v4i64 to QQ registers but do not make the type legal for any
270 // operations. v4i64 is only used for REG_SEQUENCE to load / store quad
272 addRegisterClass(MVT::v4i64, ARM::QQPRRegisterClass);
274 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
275 // neither Neon nor VFP support any arithmetic operations on it.
276 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
277 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
278 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
279 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
280 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
281 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
282 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
283 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
284 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
285 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
286 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
287 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
288 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
289 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
290 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
291 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
292 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
293 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
294 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
295 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
296 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
297 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
298 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
299 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
301 // Neon does not support some operations on v1i64 and v2i64 types.
302 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
303 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
304 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
305 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
307 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
308 setTargetDAGCombine(ISD::SHL);
309 setTargetDAGCombine(ISD::SRL);
310 setTargetDAGCombine(ISD::SRA);
311 setTargetDAGCombine(ISD::SIGN_EXTEND);
312 setTargetDAGCombine(ISD::ZERO_EXTEND);
313 setTargetDAGCombine(ISD::ANY_EXTEND);
314 setTargetDAGCombine(ISD::SELECT_CC);
317 computeRegisterProperties();
319 // ARM does not have f32 extending load.
320 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
322 // ARM does not have i1 sign extending load.
323 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
325 // ARM supports all 4 flavors of integer indexed load / store.
326 if (!Subtarget->isThumb1Only()) {
327 for (unsigned im = (unsigned)ISD::PRE_INC;
328 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
329 setIndexedLoadAction(im, MVT::i1, Legal);
330 setIndexedLoadAction(im, MVT::i8, Legal);
331 setIndexedLoadAction(im, MVT::i16, Legal);
332 setIndexedLoadAction(im, MVT::i32, Legal);
333 setIndexedStoreAction(im, MVT::i1, Legal);
334 setIndexedStoreAction(im, MVT::i8, Legal);
335 setIndexedStoreAction(im, MVT::i16, Legal);
336 setIndexedStoreAction(im, MVT::i32, Legal);
340 // i64 operation support.
341 if (Subtarget->isThumb1Only()) {
342 setOperationAction(ISD::MUL, MVT::i64, Expand);
343 setOperationAction(ISD::MULHU, MVT::i32, Expand);
344 setOperationAction(ISD::MULHS, MVT::i32, Expand);
345 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
346 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
348 setOperationAction(ISD::MUL, MVT::i64, Expand);
349 setOperationAction(ISD::MULHU, MVT::i32, Expand);
350 if (!Subtarget->hasV6Ops())
351 setOperationAction(ISD::MULHS, MVT::i32, Expand);
353 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
354 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
355 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
356 setOperationAction(ISD::SRL, MVT::i64, Custom);
357 setOperationAction(ISD::SRA, MVT::i64, Custom);
359 // ARM does not have ROTL.
360 setOperationAction(ISD::ROTL, MVT::i32, Expand);
361 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
362 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
363 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
364 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
366 // Only ARMv6 has BSWAP.
367 if (!Subtarget->hasV6Ops())
368 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
370 // These are expanded into libcalls.
371 if (!Subtarget->hasDivide()) {
372 // v7M has a hardware divider
373 setOperationAction(ISD::SDIV, MVT::i32, Expand);
374 setOperationAction(ISD::UDIV, MVT::i32, Expand);
376 setOperationAction(ISD::SREM, MVT::i32, Expand);
377 setOperationAction(ISD::UREM, MVT::i32, Expand);
378 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
379 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
381 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
382 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
383 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
384 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
385 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
387 setOperationAction(ISD::TRAP, MVT::Other, Legal);
389 // Use the default implementation.
390 setOperationAction(ISD::VASTART, MVT::Other, Custom);
391 setOperationAction(ISD::VAARG, MVT::Other, Expand);
392 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
393 setOperationAction(ISD::VAEND, MVT::Other, Expand);
394 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
395 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
396 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
397 // FIXME: Shouldn't need this, since no register is used, but the legalizer
398 // doesn't yet know how to not do that for SjLj.
399 setExceptionSelectorRegister(ARM::R0);
400 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
401 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
403 // If the subtarget does not have extract instructions, sign_extend_inreg
404 // needs to be expanded. Extract is available in ARM mode on v6 and up,
405 // and on most Thumb2 implementations.
406 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
407 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
408 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
409 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
411 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
413 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
414 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
415 // iff target supports vfp2.
416 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
418 // We want to custom lower some of our intrinsics.
419 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
421 setOperationAction(ISD::SETCC, MVT::i32, Expand);
422 setOperationAction(ISD::SETCC, MVT::f32, Expand);
423 setOperationAction(ISD::SETCC, MVT::f64, Expand);
424 setOperationAction(ISD::SELECT, MVT::i32, Expand);
425 setOperationAction(ISD::SELECT, MVT::f32, Expand);
426 setOperationAction(ISD::SELECT, MVT::f64, Expand);
427 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
428 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
429 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
431 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
432 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
433 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
434 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
435 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
437 // We don't support sin/cos/fmod/copysign/pow
438 setOperationAction(ISD::FSIN, MVT::f64, Expand);
439 setOperationAction(ISD::FSIN, MVT::f32, Expand);
440 setOperationAction(ISD::FCOS, MVT::f32, Expand);
441 setOperationAction(ISD::FCOS, MVT::f64, Expand);
442 setOperationAction(ISD::FREM, MVT::f64, Expand);
443 setOperationAction(ISD::FREM, MVT::f32, Expand);
444 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
445 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
446 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
448 setOperationAction(ISD::FPOW, MVT::f64, Expand);
449 setOperationAction(ISD::FPOW, MVT::f32, Expand);
451 // Various VFP goodness
452 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
453 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
454 if (Subtarget->hasVFP2()) {
455 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
456 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
457 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
458 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
460 // Special handling for half-precision FP.
461 if (!Subtarget->hasFP16()) {
462 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
463 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
467 // We have target-specific dag combine patterns for the following nodes:
468 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
469 setTargetDAGCombine(ISD::ADD);
470 setTargetDAGCombine(ISD::SUB);
472 setStackPointerRegisterToSaveRestore(ARM::SP);
473 setSchedulingPreference(SchedulingForRegPressure);
475 // FIXME: If-converter should use instruction latency to determine
476 // profitability rather than relying on fixed limits.
477 if (Subtarget->getCPUString() == "generic") {
478 // Generic (and overly aggressive) if-conversion limits.
479 setIfCvtBlockSizeLimit(10);
480 setIfCvtDupBlockSizeLimit(2);
481 } else if (Subtarget->hasV7Ops()) {
482 setIfCvtBlockSizeLimit(3);
483 setIfCvtDupBlockSizeLimit(1);
484 } else if (Subtarget->hasV6Ops()) {
485 setIfCvtBlockSizeLimit(2);
486 setIfCvtDupBlockSizeLimit(1);
488 setIfCvtBlockSizeLimit(3);
489 setIfCvtDupBlockSizeLimit(2);
492 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
493 // Do not enable CodePlacementOpt for now: it currently runs after the
494 // ARMConstantIslandPass and messes up branch relaxation and placement
495 // of constant islands.
496 // benefitFromCodePlacementOpt = true;
499 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
502 case ARMISD::Wrapper: return "ARMISD::Wrapper";
503 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
504 case ARMISD::CALL: return "ARMISD::CALL";
505 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
506 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
507 case ARMISD::tCALL: return "ARMISD::tCALL";
508 case ARMISD::BRCOND: return "ARMISD::BRCOND";
509 case ARMISD::BR_JT: return "ARMISD::BR_JT";
510 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
511 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
512 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
513 case ARMISD::CMP: return "ARMISD::CMP";
514 case ARMISD::CMPZ: return "ARMISD::CMPZ";
515 case ARMISD::CMPFP: return "ARMISD::CMPFP";
516 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
517 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
518 case ARMISD::CMOV: return "ARMISD::CMOV";
519 case ARMISD::CNEG: return "ARMISD::CNEG";
521 case ARMISD::RBIT: return "ARMISD::RBIT";
523 case ARMISD::FTOSI: return "ARMISD::FTOSI";
524 case ARMISD::FTOUI: return "ARMISD::FTOUI";
525 case ARMISD::SITOF: return "ARMISD::SITOF";
526 case ARMISD::UITOF: return "ARMISD::UITOF";
528 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
529 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
530 case ARMISD::RRX: return "ARMISD::RRX";
532 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
533 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
535 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
536 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
538 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
540 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
542 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
543 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
545 case ARMISD::VCEQ: return "ARMISD::VCEQ";
546 case ARMISD::VCGE: return "ARMISD::VCGE";
547 case ARMISD::VCGEU: return "ARMISD::VCGEU";
548 case ARMISD::VCGT: return "ARMISD::VCGT";
549 case ARMISD::VCGTU: return "ARMISD::VCGTU";
550 case ARMISD::VTST: return "ARMISD::VTST";
552 case ARMISD::VSHL: return "ARMISD::VSHL";
553 case ARMISD::VSHRs: return "ARMISD::VSHRs";
554 case ARMISD::VSHRu: return "ARMISD::VSHRu";
555 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
556 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
557 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
558 case ARMISD::VSHRN: return "ARMISD::VSHRN";
559 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
560 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
561 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
562 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
563 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
564 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
565 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
566 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
567 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
568 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
569 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
570 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
571 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
572 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
573 case ARMISD::VDUP: return "ARMISD::VDUP";
574 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
575 case ARMISD::VEXT: return "ARMISD::VEXT";
576 case ARMISD::VREV64: return "ARMISD::VREV64";
577 case ARMISD::VREV32: return "ARMISD::VREV32";
578 case ARMISD::VREV16: return "ARMISD::VREV16";
579 case ARMISD::VZIP: return "ARMISD::VZIP";
580 case ARMISD::VUZP: return "ARMISD::VUZP";
581 case ARMISD::VTRN: return "ARMISD::VTRN";
582 case ARMISD::FMAX: return "ARMISD::FMAX";
583 case ARMISD::FMIN: return "ARMISD::FMIN";
587 /// getFunctionAlignment - Return the Log2 alignment of this function.
588 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
589 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
592 //===----------------------------------------------------------------------===//
594 //===----------------------------------------------------------------------===//
596 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
597 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
599 default: llvm_unreachable("Unknown condition code!");
600 case ISD::SETNE: return ARMCC::NE;
601 case ISD::SETEQ: return ARMCC::EQ;
602 case ISD::SETGT: return ARMCC::GT;
603 case ISD::SETGE: return ARMCC::GE;
604 case ISD::SETLT: return ARMCC::LT;
605 case ISD::SETLE: return ARMCC::LE;
606 case ISD::SETUGT: return ARMCC::HI;
607 case ISD::SETUGE: return ARMCC::HS;
608 case ISD::SETULT: return ARMCC::LO;
609 case ISD::SETULE: return ARMCC::LS;
613 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
614 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
615 ARMCC::CondCodes &CondCode2) {
616 CondCode2 = ARMCC::AL;
618 default: llvm_unreachable("Unknown FP condition!");
620 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
622 case ISD::SETOGT: CondCode = ARMCC::GT; break;
624 case ISD::SETOGE: CondCode = ARMCC::GE; break;
625 case ISD::SETOLT: CondCode = ARMCC::MI; break;
626 case ISD::SETOLE: CondCode = ARMCC::LS; break;
627 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
628 case ISD::SETO: CondCode = ARMCC::VC; break;
629 case ISD::SETUO: CondCode = ARMCC::VS; break;
630 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
631 case ISD::SETUGT: CondCode = ARMCC::HI; break;
632 case ISD::SETUGE: CondCode = ARMCC::PL; break;
634 case ISD::SETULT: CondCode = ARMCC::LT; break;
636 case ISD::SETULE: CondCode = ARMCC::LE; break;
638 case ISD::SETUNE: CondCode = ARMCC::NE; break;
642 //===----------------------------------------------------------------------===//
643 // Calling Convention Implementation
644 //===----------------------------------------------------------------------===//
646 #include "ARMGenCallingConv.inc"
648 // APCS f64 is in register pairs, possibly split to stack
649 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
650 CCValAssign::LocInfo &LocInfo,
651 CCState &State, bool CanFail) {
652 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
654 // Try to get the first register.
655 if (unsigned Reg = State.AllocateReg(RegList, 4))
656 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
658 // For the 2nd half of a v2f64, do not fail.
662 // Put the whole thing on the stack.
663 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
664 State.AllocateStack(8, 4),
669 // Try to get the second register.
670 if (unsigned Reg = State.AllocateReg(RegList, 4))
671 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
673 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
674 State.AllocateStack(4, 4),
679 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
680 CCValAssign::LocInfo &LocInfo,
681 ISD::ArgFlagsTy &ArgFlags,
683 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
685 if (LocVT == MVT::v2f64 &&
686 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
688 return true; // we handled it
691 // AAPCS f64 is in aligned register pairs
692 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
693 CCValAssign::LocInfo &LocInfo,
694 CCState &State, bool CanFail) {
695 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
696 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
698 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
700 // For the 2nd half of a v2f64, do not just fail.
704 // Put the whole thing on the stack.
705 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
706 State.AllocateStack(8, 8),
712 for (i = 0; i < 2; ++i)
713 if (HiRegList[i] == Reg)
716 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
717 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
722 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
723 CCValAssign::LocInfo &LocInfo,
724 ISD::ArgFlagsTy &ArgFlags,
726 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
728 if (LocVT == MVT::v2f64 &&
729 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
731 return true; // we handled it
734 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
735 CCValAssign::LocInfo &LocInfo, CCState &State) {
736 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
737 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
739 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
741 return false; // we didn't handle it
744 for (i = 0; i < 2; ++i)
745 if (HiRegList[i] == Reg)
748 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
749 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
754 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
755 CCValAssign::LocInfo &LocInfo,
756 ISD::ArgFlagsTy &ArgFlags,
758 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
760 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
762 return true; // we handled it
765 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
766 CCValAssign::LocInfo &LocInfo,
767 ISD::ArgFlagsTy &ArgFlags,
769 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
773 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
774 /// given CallingConvention value.
775 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
777 bool isVarArg) const {
780 llvm_unreachable("Unsupported calling convention");
782 case CallingConv::Fast:
783 // Use target triple & subtarget features to do actual dispatch.
784 if (Subtarget->isAAPCS_ABI()) {
785 if (Subtarget->hasVFP2() &&
786 FloatABIType == FloatABI::Hard && !isVarArg)
787 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
789 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
791 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
792 case CallingConv::ARM_AAPCS_VFP:
793 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
794 case CallingConv::ARM_AAPCS:
795 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
796 case CallingConv::ARM_APCS:
797 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
801 /// LowerCallResult - Lower the result values of a call into the
802 /// appropriate copies out of appropriate physical registers.
804 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
805 CallingConv::ID CallConv, bool isVarArg,
806 const SmallVectorImpl<ISD::InputArg> &Ins,
807 DebugLoc dl, SelectionDAG &DAG,
808 SmallVectorImpl<SDValue> &InVals) const {
810 // Assign locations to each value returned by this call.
811 SmallVector<CCValAssign, 16> RVLocs;
812 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
813 RVLocs, *DAG.getContext());
814 CCInfo.AnalyzeCallResult(Ins,
815 CCAssignFnForNode(CallConv, /* Return*/ true,
818 // Copy all of the result registers out of their specified physreg.
819 for (unsigned i = 0; i != RVLocs.size(); ++i) {
820 CCValAssign VA = RVLocs[i];
823 if (VA.needsCustom()) {
824 // Handle f64 or half of a v2f64.
825 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
827 Chain = Lo.getValue(1);
828 InFlag = Lo.getValue(2);
829 VA = RVLocs[++i]; // skip ahead to next loc
830 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
832 Chain = Hi.getValue(1);
833 InFlag = Hi.getValue(2);
834 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
836 if (VA.getLocVT() == MVT::v2f64) {
837 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
838 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
839 DAG.getConstant(0, MVT::i32));
841 VA = RVLocs[++i]; // skip ahead to next loc
842 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
843 Chain = Lo.getValue(1);
844 InFlag = Lo.getValue(2);
845 VA = RVLocs[++i]; // skip ahead to next loc
846 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
847 Chain = Hi.getValue(1);
848 InFlag = Hi.getValue(2);
849 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
850 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
851 DAG.getConstant(1, MVT::i32));
854 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
856 Chain = Val.getValue(1);
857 InFlag = Val.getValue(2);
860 switch (VA.getLocInfo()) {
861 default: llvm_unreachable("Unknown loc info!");
862 case CCValAssign::Full: break;
863 case CCValAssign::BCvt:
864 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
868 InVals.push_back(Val);
874 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
875 /// by "Src" to address "Dst" of size "Size". Alignment information is
876 /// specified by the specific parameter attribute. The copy will be passed as
877 /// a byval function parameter.
878 /// Sometimes what we are copying is the end of a larger object, the part that
879 /// does not fit in registers.
881 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
882 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
884 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
885 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
886 /*isVolatile=*/false, /*AlwaysInline=*/false,
890 /// LowerMemOpCallTo - Store the argument to the stack.
892 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
893 SDValue StackPtr, SDValue Arg,
894 DebugLoc dl, SelectionDAG &DAG,
895 const CCValAssign &VA,
896 ISD::ArgFlagsTy Flags) const {
897 unsigned LocMemOffset = VA.getLocMemOffset();
898 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
899 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
900 if (Flags.isByVal()) {
901 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
903 return DAG.getStore(Chain, dl, Arg, PtrOff,
904 PseudoSourceValue::getStack(), LocMemOffset,
908 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
909 SDValue Chain, SDValue &Arg,
910 RegsToPassVector &RegsToPass,
911 CCValAssign &VA, CCValAssign &NextVA,
913 SmallVector<SDValue, 8> &MemOpChains,
914 ISD::ArgFlagsTy Flags) const {
916 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
917 DAG.getVTList(MVT::i32, MVT::i32), Arg);
918 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
920 if (NextVA.isRegLoc())
921 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
923 assert(NextVA.isMemLoc());
924 if (StackPtr.getNode() == 0)
925 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
927 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
933 /// LowerCall - Lowering a call into a callseq_start <-
934 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
937 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
938 CallingConv::ID CallConv, bool isVarArg,
940 const SmallVectorImpl<ISD::OutputArg> &Outs,
941 const SmallVectorImpl<ISD::InputArg> &Ins,
942 DebugLoc dl, SelectionDAG &DAG,
943 SmallVectorImpl<SDValue> &InVals) const {
944 // ARM target does not yet support tail call optimization.
947 // Analyze operands of the call, assigning locations to each operand.
948 SmallVector<CCValAssign, 16> ArgLocs;
949 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
951 CCInfo.AnalyzeCallOperands(Outs,
952 CCAssignFnForNode(CallConv, /* Return*/ false,
955 // Get a count of how many bytes are to be pushed on the stack.
956 unsigned NumBytes = CCInfo.getNextStackOffset();
958 // Adjust the stack pointer for the new arguments...
959 // These operations are automatically eliminated by the prolog/epilog pass
960 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
962 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
964 RegsToPassVector RegsToPass;
965 SmallVector<SDValue, 8> MemOpChains;
967 // Walk the register/memloc assignments, inserting copies/loads. In the case
968 // of tail call optimization, arguments are handled later.
969 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
972 CCValAssign &VA = ArgLocs[i];
973 SDValue Arg = Outs[realArgIdx].Val;
974 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
976 // Promote the value if needed.
977 switch (VA.getLocInfo()) {
978 default: llvm_unreachable("Unknown loc info!");
979 case CCValAssign::Full: break;
980 case CCValAssign::SExt:
981 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
983 case CCValAssign::ZExt:
984 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
986 case CCValAssign::AExt:
987 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
989 case CCValAssign::BCvt:
990 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
994 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
995 if (VA.needsCustom()) {
996 if (VA.getLocVT() == MVT::v2f64) {
997 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
998 DAG.getConstant(0, MVT::i32));
999 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1000 DAG.getConstant(1, MVT::i32));
1002 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1003 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1005 VA = ArgLocs[++i]; // skip ahead to next loc
1006 if (VA.isRegLoc()) {
1007 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1008 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1010 assert(VA.isMemLoc());
1012 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1013 dl, DAG, VA, Flags));
1016 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1017 StackPtr, MemOpChains, Flags);
1019 } else if (VA.isRegLoc()) {
1020 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1022 assert(VA.isMemLoc());
1024 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1025 dl, DAG, VA, Flags));
1029 if (!MemOpChains.empty())
1030 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1031 &MemOpChains[0], MemOpChains.size());
1033 // Build a sequence of copy-to-reg nodes chained together with token chain
1034 // and flag operands which copy the outgoing args into the appropriate regs.
1036 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1037 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1038 RegsToPass[i].second, InFlag);
1039 InFlag = Chain.getValue(1);
1042 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1043 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1044 // node so that legalize doesn't hack it.
1045 bool isDirect = false;
1046 bool isARMFunc = false;
1047 bool isLocalARMFunc = false;
1048 MachineFunction &MF = DAG.getMachineFunction();
1049 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1051 if (EnableARMLongCalls) {
1052 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1053 && "long-calls with non-static relocation model!");
1054 // Handle a global address or an external symbol. If it's not one of
1055 // those, the target's already in a register, so we don't need to do
1057 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1058 const GlobalValue *GV = G->getGlobal();
1059 // Create a constant pool entry for the callee address
1060 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1061 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1064 // Get the address of the callee into a register
1065 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1066 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1067 Callee = DAG.getLoad(getPointerTy(), dl,
1068 DAG.getEntryNode(), CPAddr,
1069 PseudoSourceValue::getConstantPool(), 0,
1071 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1072 const char *Sym = S->getSymbol();
1074 // Create a constant pool entry for the callee address
1075 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1076 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1077 Sym, ARMPCLabelIndex, 0);
1078 // Get the address of the callee into a register
1079 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1080 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1081 Callee = DAG.getLoad(getPointerTy(), dl,
1082 DAG.getEntryNode(), CPAddr,
1083 PseudoSourceValue::getConstantPool(), 0,
1086 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1087 const GlobalValue *GV = G->getGlobal();
1089 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1090 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1091 getTargetMachine().getRelocationModel() != Reloc::Static;
1092 isARMFunc = !Subtarget->isThumb() || isStub;
1093 // ARM call to a local ARM function is predicable.
1094 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1095 // tBX takes a register source operand.
1096 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1097 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1098 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1101 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1102 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1103 Callee = DAG.getLoad(getPointerTy(), dl,
1104 DAG.getEntryNode(), CPAddr,
1105 PseudoSourceValue::getConstantPool(), 0,
1107 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1108 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1109 getPointerTy(), Callee, PICLabel);
1111 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1112 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1114 bool isStub = Subtarget->isTargetDarwin() &&
1115 getTargetMachine().getRelocationModel() != Reloc::Static;
1116 isARMFunc = !Subtarget->isThumb() || isStub;
1117 // tBX takes a register source operand.
1118 const char *Sym = S->getSymbol();
1119 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1120 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1121 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1122 Sym, ARMPCLabelIndex, 4);
1123 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1124 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1125 Callee = DAG.getLoad(getPointerTy(), dl,
1126 DAG.getEntryNode(), CPAddr,
1127 PseudoSourceValue::getConstantPool(), 0,
1129 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1130 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1131 getPointerTy(), Callee, PICLabel);
1133 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1136 // FIXME: handle tail calls differently.
1138 if (Subtarget->isThumb()) {
1139 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1140 CallOpc = ARMISD::CALL_NOLINK;
1142 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1144 CallOpc = (isDirect || Subtarget->hasV5TOps())
1145 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1146 : ARMISD::CALL_NOLINK;
1148 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1149 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1150 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1151 InFlag = Chain.getValue(1);
1154 std::vector<SDValue> Ops;
1155 Ops.push_back(Chain);
1156 Ops.push_back(Callee);
1158 // Add argument registers to the end of the list so that they are known live
1160 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1161 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1162 RegsToPass[i].second.getValueType()));
1164 if (InFlag.getNode())
1165 Ops.push_back(InFlag);
1166 // Returns a chain and a flag for retval copy to use.
1167 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1168 &Ops[0], Ops.size());
1169 InFlag = Chain.getValue(1);
1171 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1172 DAG.getIntPtrConstant(0, true), InFlag);
1174 InFlag = Chain.getValue(1);
1176 // Handle result values, copying them out of physregs into vregs that we
1178 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1183 ARMTargetLowering::LowerReturn(SDValue Chain,
1184 CallingConv::ID CallConv, bool isVarArg,
1185 const SmallVectorImpl<ISD::OutputArg> &Outs,
1186 DebugLoc dl, SelectionDAG &DAG) const {
1188 // CCValAssign - represent the assignment of the return value to a location.
1189 SmallVector<CCValAssign, 16> RVLocs;
1191 // CCState - Info about the registers and stack slots.
1192 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1195 // Analyze outgoing return values.
1196 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1199 // If this is the first return lowered for this function, add
1200 // the regs to the liveout set for the function.
1201 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1202 for (unsigned i = 0; i != RVLocs.size(); ++i)
1203 if (RVLocs[i].isRegLoc())
1204 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1209 // Copy the result values into the output registers.
1210 for (unsigned i = 0, realRVLocIdx = 0;
1212 ++i, ++realRVLocIdx) {
1213 CCValAssign &VA = RVLocs[i];
1214 assert(VA.isRegLoc() && "Can only return in registers!");
1216 SDValue Arg = Outs[realRVLocIdx].Val;
1218 switch (VA.getLocInfo()) {
1219 default: llvm_unreachable("Unknown loc info!");
1220 case CCValAssign::Full: break;
1221 case CCValAssign::BCvt:
1222 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1226 if (VA.needsCustom()) {
1227 if (VA.getLocVT() == MVT::v2f64) {
1228 // Extract the first half and return it in two registers.
1229 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1230 DAG.getConstant(0, MVT::i32));
1231 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1232 DAG.getVTList(MVT::i32, MVT::i32), Half);
1234 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1235 Flag = Chain.getValue(1);
1236 VA = RVLocs[++i]; // skip ahead to next loc
1237 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1238 HalfGPRs.getValue(1), Flag);
1239 Flag = Chain.getValue(1);
1240 VA = RVLocs[++i]; // skip ahead to next loc
1242 // Extract the 2nd half and fall through to handle it as an f64 value.
1243 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1244 DAG.getConstant(1, MVT::i32));
1246 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1248 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1249 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1250 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1251 Flag = Chain.getValue(1);
1252 VA = RVLocs[++i]; // skip ahead to next loc
1253 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1256 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1258 // Guarantee that all emitted copies are
1259 // stuck together, avoiding something bad.
1260 Flag = Chain.getValue(1);
1265 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1267 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1272 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1273 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1274 // one of the above mentioned nodes. It has to be wrapped because otherwise
1275 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1276 // be used to form addressing mode. These wrapped nodes will be selected
1278 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1279 EVT PtrVT = Op.getValueType();
1280 // FIXME there is no actual debug info here
1281 DebugLoc dl = Op.getDebugLoc();
1282 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1284 if (CP->isMachineConstantPoolEntry())
1285 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1286 CP->getAlignment());
1288 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1289 CP->getAlignment());
1290 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1293 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1294 SelectionDAG &DAG) const {
1295 MachineFunction &MF = DAG.getMachineFunction();
1296 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1297 unsigned ARMPCLabelIndex = 0;
1298 DebugLoc DL = Op.getDebugLoc();
1299 EVT PtrVT = getPointerTy();
1300 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1301 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1303 if (RelocM == Reloc::Static) {
1304 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1306 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1307 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1308 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1309 ARMCP::CPBlockAddress,
1311 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1313 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1314 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1315 PseudoSourceValue::getConstantPool(), 0,
1317 if (RelocM == Reloc::Static)
1319 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1320 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1323 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1325 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1326 SelectionDAG &DAG) const {
1327 DebugLoc dl = GA->getDebugLoc();
1328 EVT PtrVT = getPointerTy();
1329 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1330 MachineFunction &MF = DAG.getMachineFunction();
1331 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1332 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1333 ARMConstantPoolValue *CPV =
1334 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1335 ARMCP::CPValue, PCAdj, "tlsgd", true);
1336 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1337 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1338 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1339 PseudoSourceValue::getConstantPool(), 0,
1341 SDValue Chain = Argument.getValue(1);
1343 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1344 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1346 // call __tls_get_addr.
1349 Entry.Node = Argument;
1350 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1351 Args.push_back(Entry);
1352 // FIXME: is there useful debug info available here?
1353 std::pair<SDValue, SDValue> CallResult =
1354 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1355 false, false, false, false,
1356 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1357 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1358 return CallResult.first;
1361 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1362 // "local exec" model.
1364 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1365 SelectionDAG &DAG) const {
1366 const GlobalValue *GV = GA->getGlobal();
1367 DebugLoc dl = GA->getDebugLoc();
1369 SDValue Chain = DAG.getEntryNode();
1370 EVT PtrVT = getPointerTy();
1371 // Get the Thread Pointer
1372 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1374 if (GV->isDeclaration()) {
1375 MachineFunction &MF = DAG.getMachineFunction();
1376 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1377 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1378 // Initial exec model.
1379 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1380 ARMConstantPoolValue *CPV =
1381 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1382 ARMCP::CPValue, PCAdj, "gottpoff", true);
1383 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1384 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1385 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1386 PseudoSourceValue::getConstantPool(), 0,
1388 Chain = Offset.getValue(1);
1390 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1391 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1393 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1394 PseudoSourceValue::getConstantPool(), 0,
1398 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1399 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1400 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1401 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1402 PseudoSourceValue::getConstantPool(), 0,
1406 // The address of the thread local variable is the add of the thread
1407 // pointer with the offset of the variable.
1408 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1412 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1413 // TODO: implement the "local dynamic" model
1414 assert(Subtarget->isTargetELF() &&
1415 "TLS not implemented for non-ELF targets");
1416 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1417 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1418 // otherwise use the "Local Exec" TLS Model
1419 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1420 return LowerToTLSGeneralDynamicModel(GA, DAG);
1422 return LowerToTLSExecModels(GA, DAG);
1425 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1426 SelectionDAG &DAG) const {
1427 EVT PtrVT = getPointerTy();
1428 DebugLoc dl = Op.getDebugLoc();
1429 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1430 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1431 if (RelocM == Reloc::PIC_) {
1432 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1433 ARMConstantPoolValue *CPV =
1434 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1435 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1436 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1437 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1439 PseudoSourceValue::getConstantPool(), 0,
1441 SDValue Chain = Result.getValue(1);
1442 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1443 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1445 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1446 PseudoSourceValue::getGOT(), 0,
1450 // If we have T2 ops, we can materialize the address directly via movt/movw
1451 // pair. This is always cheaper.
1452 if (Subtarget->useMovt()) {
1453 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1454 DAG.getTargetGlobalAddress(GV, PtrVT));
1456 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1457 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1458 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1459 PseudoSourceValue::getConstantPool(), 0,
1465 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1466 SelectionDAG &DAG) const {
1467 MachineFunction &MF = DAG.getMachineFunction();
1468 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1469 unsigned ARMPCLabelIndex = 0;
1470 EVT PtrVT = getPointerTy();
1471 DebugLoc dl = Op.getDebugLoc();
1472 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1473 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1475 if (RelocM == Reloc::Static)
1476 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1478 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1479 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1480 ARMConstantPoolValue *CPV =
1481 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1482 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1484 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1486 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1487 PseudoSourceValue::getConstantPool(), 0,
1489 SDValue Chain = Result.getValue(1);
1491 if (RelocM == Reloc::PIC_) {
1492 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1493 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1496 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1497 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1498 PseudoSourceValue::getGOT(), 0,
1504 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1505 SelectionDAG &DAG) const {
1506 assert(Subtarget->isTargetELF() &&
1507 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1508 MachineFunction &MF = DAG.getMachineFunction();
1509 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1510 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1511 EVT PtrVT = getPointerTy();
1512 DebugLoc dl = Op.getDebugLoc();
1513 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1514 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1515 "_GLOBAL_OFFSET_TABLE_",
1516 ARMPCLabelIndex, PCAdj);
1517 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1518 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1519 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1520 PseudoSourceValue::getConstantPool(), 0,
1522 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1523 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1527 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1528 const ARMSubtarget *Subtarget)
1530 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1531 DebugLoc dl = Op.getDebugLoc();
1533 default: return SDValue(); // Don't custom lower most intrinsics.
1534 case Intrinsic::arm_thread_pointer: {
1535 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1536 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1538 case Intrinsic::eh_sjlj_lsda: {
1539 MachineFunction &MF = DAG.getMachineFunction();
1540 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1541 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1542 EVT PtrVT = getPointerTy();
1543 DebugLoc dl = Op.getDebugLoc();
1544 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1546 unsigned PCAdj = (RelocM != Reloc::PIC_)
1547 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1548 ARMConstantPoolValue *CPV =
1549 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1550 ARMCP::CPLSDA, PCAdj);
1551 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1552 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1554 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1555 PseudoSourceValue::getConstantPool(), 0,
1557 SDValue Chain = Result.getValue(1);
1559 if (RelocM == Reloc::PIC_) {
1560 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1561 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1565 case Intrinsic::eh_sjlj_setjmp:
1566 SDValue Val = Subtarget->isThumb() ?
1567 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1568 DAG.getConstant(0, MVT::i32);
1569 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1574 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1575 const ARMSubtarget *Subtarget) {
1576 DebugLoc dl = Op.getDebugLoc();
1577 SDValue Op5 = Op.getOperand(5);
1579 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1580 if (isDeviceBarrier) {
1581 if (Subtarget->hasV7Ops())
1582 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1584 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1585 DAG.getConstant(0, MVT::i32));
1587 if (Subtarget->hasV7Ops())
1588 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1590 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1591 DAG.getConstant(0, MVT::i32));
1596 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1597 MachineFunction &MF = DAG.getMachineFunction();
1598 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1600 // vastart just stores the address of the VarArgsFrameIndex slot into the
1601 // memory location argument.
1602 DebugLoc dl = Op.getDebugLoc();
1603 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1604 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1605 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1606 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1611 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1612 SelectionDAG &DAG) const {
1613 SDNode *Node = Op.getNode();
1614 DebugLoc dl = Node->getDebugLoc();
1615 EVT VT = Node->getValueType(0);
1616 SDValue Chain = Op.getOperand(0);
1617 SDValue Size = Op.getOperand(1);
1618 SDValue Align = Op.getOperand(2);
1620 // Chain the dynamic stack allocation so that it doesn't modify the stack
1621 // pointer when other instructions are using the stack.
1622 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1624 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1625 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1626 if (AlignVal > StackAlign)
1627 // Do this now since selection pass cannot introduce new target
1628 // independent node.
1629 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1631 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1632 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1633 // do even more horrible hack later.
1634 MachineFunction &MF = DAG.getMachineFunction();
1635 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1636 if (AFI->isThumb1OnlyFunction()) {
1638 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1640 uint32_t Val = C->getZExtValue();
1641 if (Val <= 508 && ((Val & 3) == 0))
1645 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1648 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1649 SDValue Ops1[] = { Chain, Size, Align };
1650 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1651 Chain = Res.getValue(1);
1652 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1653 DAG.getIntPtrConstant(0, true), SDValue());
1654 SDValue Ops2[] = { Res, Chain };
1655 return DAG.getMergeValues(Ops2, 2, dl);
1659 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1660 SDValue &Root, SelectionDAG &DAG,
1661 DebugLoc dl) const {
1662 MachineFunction &MF = DAG.getMachineFunction();
1663 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1665 TargetRegisterClass *RC;
1666 if (AFI->isThumb1OnlyFunction())
1667 RC = ARM::tGPRRegisterClass;
1669 RC = ARM::GPRRegisterClass;
1671 // Transform the arguments stored in physical registers into virtual ones.
1672 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1673 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1676 if (NextVA.isMemLoc()) {
1677 MachineFrameInfo *MFI = MF.getFrameInfo();
1678 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1680 // Create load node to retrieve arguments from the stack.
1681 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1682 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1683 PseudoSourceValue::getFixedStack(FI), 0,
1686 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1687 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1690 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1694 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1695 CallingConv::ID CallConv, bool isVarArg,
1696 const SmallVectorImpl<ISD::InputArg>
1698 DebugLoc dl, SelectionDAG &DAG,
1699 SmallVectorImpl<SDValue> &InVals)
1702 MachineFunction &MF = DAG.getMachineFunction();
1703 MachineFrameInfo *MFI = MF.getFrameInfo();
1705 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1707 // Assign locations to all of the incoming arguments.
1708 SmallVector<CCValAssign, 16> ArgLocs;
1709 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1711 CCInfo.AnalyzeFormalArguments(Ins,
1712 CCAssignFnForNode(CallConv, /* Return*/ false,
1715 SmallVector<SDValue, 16> ArgValues;
1717 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1718 CCValAssign &VA = ArgLocs[i];
1720 // Arguments stored in registers.
1721 if (VA.isRegLoc()) {
1722 EVT RegVT = VA.getLocVT();
1725 if (VA.needsCustom()) {
1726 // f64 and vector types are split up into multiple registers or
1727 // combinations of registers and stack slots.
1728 if (VA.getLocVT() == MVT::v2f64) {
1729 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1731 VA = ArgLocs[++i]; // skip ahead to next loc
1733 if (VA.isMemLoc()) {
1734 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
1736 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1737 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1738 PseudoSourceValue::getFixedStack(FI), 0,
1741 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1744 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1745 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1746 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1747 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1748 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1750 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1753 TargetRegisterClass *RC;
1755 if (RegVT == MVT::f32)
1756 RC = ARM::SPRRegisterClass;
1757 else if (RegVT == MVT::f64)
1758 RC = ARM::DPRRegisterClass;
1759 else if (RegVT == MVT::v2f64)
1760 RC = ARM::QPRRegisterClass;
1761 else if (RegVT == MVT::i32)
1762 RC = (AFI->isThumb1OnlyFunction() ?
1763 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1765 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1767 // Transform the arguments in physical registers into virtual ones.
1768 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1769 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1772 // If this is an 8 or 16-bit value, it is really passed promoted
1773 // to 32 bits. Insert an assert[sz]ext to capture this, then
1774 // truncate to the right size.
1775 switch (VA.getLocInfo()) {
1776 default: llvm_unreachable("Unknown loc info!");
1777 case CCValAssign::Full: break;
1778 case CCValAssign::BCvt:
1779 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1781 case CCValAssign::SExt:
1782 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1783 DAG.getValueType(VA.getValVT()));
1784 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1786 case CCValAssign::ZExt:
1787 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1788 DAG.getValueType(VA.getValVT()));
1789 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1793 InVals.push_back(ArgValue);
1795 } else { // VA.isRegLoc()
1798 assert(VA.isMemLoc());
1799 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1801 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1802 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1805 // Create load nodes to retrieve arguments from the stack.
1806 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1807 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1808 PseudoSourceValue::getFixedStack(FI), 0,
1815 static const unsigned GPRArgRegs[] = {
1816 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1819 unsigned NumGPRs = CCInfo.getFirstUnallocated
1820 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1822 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1823 unsigned VARegSize = (4 - NumGPRs) * 4;
1824 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1825 unsigned ArgOffset = CCInfo.getNextStackOffset();
1826 if (VARegSaveSize) {
1827 // If this function is vararg, store any remaining integer argument regs
1828 // to their spots on the stack so that they may be loaded by deferencing
1829 // the result of va_next.
1830 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1831 AFI->setVarArgsFrameIndex(
1832 MFI->CreateFixedObject(VARegSaveSize,
1833 ArgOffset + VARegSaveSize - VARegSize,
1835 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
1838 SmallVector<SDValue, 4> MemOps;
1839 for (; NumGPRs < 4; ++NumGPRs) {
1840 TargetRegisterClass *RC;
1841 if (AFI->isThumb1OnlyFunction())
1842 RC = ARM::tGPRRegisterClass;
1844 RC = ARM::GPRRegisterClass;
1846 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1847 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1849 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1850 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()), 0,
1852 MemOps.push_back(Store);
1853 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1854 DAG.getConstant(4, getPointerTy()));
1856 if (!MemOps.empty())
1857 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1858 &MemOps[0], MemOps.size());
1860 // This will point to the next argument passed via stack.
1861 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
1868 /// isFloatingPointZero - Return true if this is +0.0.
1869 static bool isFloatingPointZero(SDValue Op) {
1870 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1871 return CFP->getValueAPF().isPosZero();
1872 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1873 // Maybe this has already been legalized into the constant pool?
1874 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1875 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1876 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1877 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1878 return CFP->getValueAPF().isPosZero();
1884 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1885 /// the given operands.
1887 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1888 SDValue &ARMCC, SelectionDAG &DAG,
1889 DebugLoc dl) const {
1890 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1891 unsigned C = RHSC->getZExtValue();
1892 if (!isLegalICmpImmediate(C)) {
1893 // Constant does not fit, try adjusting it by one?
1898 if (isLegalICmpImmediate(C-1)) {
1899 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1900 RHS = DAG.getConstant(C-1, MVT::i32);
1905 if (C > 0 && isLegalICmpImmediate(C-1)) {
1906 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1907 RHS = DAG.getConstant(C-1, MVT::i32);
1912 if (isLegalICmpImmediate(C+1)) {
1913 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1914 RHS = DAG.getConstant(C+1, MVT::i32);
1919 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1920 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1921 RHS = DAG.getConstant(C+1, MVT::i32);
1928 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1929 ARMISD::NodeType CompareType;
1932 CompareType = ARMISD::CMP;
1937 CompareType = ARMISD::CMPZ;
1940 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1941 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1944 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1945 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1948 if (!isFloatingPointZero(RHS))
1949 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1951 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1952 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1955 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
1956 EVT VT = Op.getValueType();
1957 SDValue LHS = Op.getOperand(0);
1958 SDValue RHS = Op.getOperand(1);
1959 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1960 SDValue TrueVal = Op.getOperand(2);
1961 SDValue FalseVal = Op.getOperand(3);
1962 DebugLoc dl = Op.getDebugLoc();
1964 if (LHS.getValueType() == MVT::i32) {
1966 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1967 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1968 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1971 ARMCC::CondCodes CondCode, CondCode2;
1972 FPCCToARMCC(CC, CondCode, CondCode2);
1974 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1975 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1976 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1977 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1979 if (CondCode2 != ARMCC::AL) {
1980 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1981 // FIXME: Needs another CMP because flag can have but one use.
1982 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1983 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1984 Result, TrueVal, ARMCC2, CCR, Cmp2);
1989 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1990 SDValue Chain = Op.getOperand(0);
1991 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1992 SDValue LHS = Op.getOperand(2);
1993 SDValue RHS = Op.getOperand(3);
1994 SDValue Dest = Op.getOperand(4);
1995 DebugLoc dl = Op.getDebugLoc();
1997 if (LHS.getValueType() == MVT::i32) {
1999 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2000 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2001 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2002 Chain, Dest, ARMCC, CCR,Cmp);
2005 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2006 ARMCC::CondCodes CondCode, CondCode2;
2007 FPCCToARMCC(CC, CondCode, CondCode2);
2009 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2010 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2011 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2012 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2013 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2014 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2015 if (CondCode2 != ARMCC::AL) {
2016 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2017 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2018 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2023 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2024 SDValue Chain = Op.getOperand(0);
2025 SDValue Table = Op.getOperand(1);
2026 SDValue Index = Op.getOperand(2);
2027 DebugLoc dl = Op.getDebugLoc();
2029 EVT PTy = getPointerTy();
2030 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2031 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2032 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2033 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2034 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2035 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2036 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2037 if (Subtarget->isThumb2()) {
2038 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2039 // which does another jump to the destination. This also makes it easier
2040 // to translate it to TBB / TBH later.
2041 // FIXME: This might not work if the function is extremely large.
2042 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2043 Addr, Op.getOperand(2), JTI, UId);
2045 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2046 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2047 PseudoSourceValue::getJumpTable(), 0,
2049 Chain = Addr.getValue(1);
2050 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2051 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2053 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2054 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2055 Chain = Addr.getValue(1);
2056 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2060 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2061 DebugLoc dl = Op.getDebugLoc();
2064 switch (Op.getOpcode()) {
2066 assert(0 && "Invalid opcode!");
2067 case ISD::FP_TO_SINT:
2068 Opc = ARMISD::FTOSI;
2070 case ISD::FP_TO_UINT:
2071 Opc = ARMISD::FTOUI;
2074 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2075 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2078 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2079 EVT VT = Op.getValueType();
2080 DebugLoc dl = Op.getDebugLoc();
2083 switch (Op.getOpcode()) {
2085 assert(0 && "Invalid opcode!");
2086 case ISD::SINT_TO_FP:
2087 Opc = ARMISD::SITOF;
2089 case ISD::UINT_TO_FP:
2090 Opc = ARMISD::UITOF;
2094 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2095 return DAG.getNode(Opc, dl, VT, Op);
2098 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2099 // Implement fcopysign with a fabs and a conditional fneg.
2100 SDValue Tmp0 = Op.getOperand(0);
2101 SDValue Tmp1 = Op.getOperand(1);
2102 DebugLoc dl = Op.getDebugLoc();
2103 EVT VT = Op.getValueType();
2104 EVT SrcVT = Tmp1.getValueType();
2105 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2106 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2107 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2108 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2109 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2112 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2113 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2114 MFI->setFrameAddressIsTaken(true);
2115 EVT VT = Op.getValueType();
2116 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2117 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2118 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2119 ? ARM::R7 : ARM::R11;
2120 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2122 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2128 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2130 SDValue Dst, SDValue Src,
2131 SDValue Size, unsigned Align,
2132 bool isVolatile, bool AlwaysInline,
2136 uint64_t SrcSVOff) const {
2137 // Do repeated 4-byte loads and stores. To be improved.
2138 // This requires 4-byte alignment.
2139 if ((Align & 3) != 0)
2141 // This requires the copy size to be a constant, preferrably
2142 // within a subtarget-specific limit.
2143 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2146 uint64_t SizeVal = ConstantSize->getZExtValue();
2147 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2150 unsigned BytesLeft = SizeVal & 3;
2151 unsigned NumMemOps = SizeVal >> 2;
2152 unsigned EmittedNumMemOps = 0;
2154 unsigned VTSize = 4;
2156 const unsigned MAX_LOADS_IN_LDM = 6;
2157 SDValue TFOps[MAX_LOADS_IN_LDM];
2158 SDValue Loads[MAX_LOADS_IN_LDM];
2159 uint64_t SrcOff = 0, DstOff = 0;
2161 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2162 // same number of stores. The loads and stores will get combined into
2163 // ldm/stm later on.
2164 while (EmittedNumMemOps < NumMemOps) {
2166 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2167 Loads[i] = DAG.getLoad(VT, dl, Chain,
2168 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2169 DAG.getConstant(SrcOff, MVT::i32)),
2170 SrcSV, SrcSVOff + SrcOff, isVolatile, false, 0);
2171 TFOps[i] = Loads[i].getValue(1);
2174 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2177 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2178 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2179 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2180 DAG.getConstant(DstOff, MVT::i32)),
2181 DstSV, DstSVOff + DstOff, isVolatile, false, 0);
2184 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2186 EmittedNumMemOps += i;
2192 // Issue loads / stores for the trailing (1 - 3) bytes.
2193 unsigned BytesLeftSave = BytesLeft;
2196 if (BytesLeft >= 2) {
2204 Loads[i] = DAG.getLoad(VT, dl, Chain,
2205 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2206 DAG.getConstant(SrcOff, MVT::i32)),
2207 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2208 TFOps[i] = Loads[i].getValue(1);
2211 BytesLeft -= VTSize;
2213 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2216 BytesLeft = BytesLeftSave;
2218 if (BytesLeft >= 2) {
2226 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2227 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2228 DAG.getConstant(DstOff, MVT::i32)),
2229 DstSV, DstSVOff + DstOff, false, false, 0);
2232 BytesLeft -= VTSize;
2234 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2237 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2238 /// expand a bit convert where either the source or destination type is i64 to
2239 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2240 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2241 /// vectors), since the legalizer won't know what to do with that.
2242 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2243 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2244 DebugLoc dl = N->getDebugLoc();
2245 SDValue Op = N->getOperand(0);
2247 // This function is only supposed to be called for i64 types, either as the
2248 // source or destination of the bit convert.
2249 EVT SrcVT = Op.getValueType();
2250 EVT DstVT = N->getValueType(0);
2251 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2252 "ExpandBIT_CONVERT called for non-i64 type");
2254 // Turn i64->f64 into VMOVDRR.
2255 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2256 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2257 DAG.getConstant(0, MVT::i32));
2258 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2259 DAG.getConstant(1, MVT::i32));
2260 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2263 // Turn f64->i64 into VMOVRRD.
2264 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2265 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2266 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2267 // Merge the pieces into a single i64 value.
2268 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2274 /// getZeroVector - Returns a vector of specified type with all zero elements.
2276 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2277 assert(VT.isVector() && "Expected a vector type");
2279 // Zero vectors are used to represent vector negation and in those cases
2280 // will be implemented with the NEON VNEG instruction. However, VNEG does
2281 // not support i64 elements, so sometimes the zero vectors will need to be
2282 // explicitly constructed. For those cases, and potentially other uses in
2283 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2284 // to their dest type. This ensures they get CSE'd.
2286 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2287 SmallVector<SDValue, 8> Ops;
2290 if (VT.getSizeInBits() == 64) {
2291 Ops.assign(8, Cst); TVT = MVT::v8i8;
2293 Ops.assign(16, Cst); TVT = MVT::v16i8;
2295 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2297 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2300 /// getOnesVector - Returns a vector of specified type with all bits set.
2302 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2303 assert(VT.isVector() && "Expected a vector type");
2305 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2306 // dest type. This ensures they get CSE'd.
2308 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2309 SmallVector<SDValue, 8> Ops;
2312 if (VT.getSizeInBits() == 64) {
2313 Ops.assign(8, Cst); TVT = MVT::v8i8;
2315 Ops.assign(16, Cst); TVT = MVT::v16i8;
2317 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2319 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2322 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2323 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2324 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2325 SelectionDAG &DAG) const {
2326 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2327 EVT VT = Op.getValueType();
2328 unsigned VTBits = VT.getSizeInBits();
2329 DebugLoc dl = Op.getDebugLoc();
2330 SDValue ShOpLo = Op.getOperand(0);
2331 SDValue ShOpHi = Op.getOperand(1);
2332 SDValue ShAmt = Op.getOperand(2);
2334 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2336 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2338 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2339 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2340 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2341 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2342 DAG.getConstant(VTBits, MVT::i32));
2343 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2344 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2345 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2347 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2348 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2350 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2351 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2354 SDValue Ops[2] = { Lo, Hi };
2355 return DAG.getMergeValues(Ops, 2, dl);
2358 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2359 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2360 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2361 SelectionDAG &DAG) const {
2362 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2363 EVT VT = Op.getValueType();
2364 unsigned VTBits = VT.getSizeInBits();
2365 DebugLoc dl = Op.getDebugLoc();
2366 SDValue ShOpLo = Op.getOperand(0);
2367 SDValue ShOpHi = Op.getOperand(1);
2368 SDValue ShAmt = Op.getOperand(2);
2371 assert(Op.getOpcode() == ISD::SHL_PARTS);
2372 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2373 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2374 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2375 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2376 DAG.getConstant(VTBits, MVT::i32));
2377 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2378 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2380 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2381 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2382 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2384 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2385 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2388 SDValue Ops[2] = { Lo, Hi };
2389 return DAG.getMergeValues(Ops, 2, dl);
2392 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2393 const ARMSubtarget *ST) {
2394 EVT VT = N->getValueType(0);
2395 DebugLoc dl = N->getDebugLoc();
2397 if (!ST->hasV6T2Ops())
2400 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2401 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2404 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2405 const ARMSubtarget *ST) {
2406 EVT VT = N->getValueType(0);
2407 DebugLoc dl = N->getDebugLoc();
2409 // Lower vector shifts on NEON to use VSHL.
2410 if (VT.isVector()) {
2411 assert(ST->hasNEON() && "unexpected vector shift");
2413 // Left shifts translate directly to the vshiftu intrinsic.
2414 if (N->getOpcode() == ISD::SHL)
2415 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2416 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2417 N->getOperand(0), N->getOperand(1));
2419 assert((N->getOpcode() == ISD::SRA ||
2420 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2422 // NEON uses the same intrinsics for both left and right shifts. For
2423 // right shifts, the shift amounts are negative, so negate the vector of
2425 EVT ShiftVT = N->getOperand(1).getValueType();
2426 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2427 getZeroVector(ShiftVT, DAG, dl),
2429 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2430 Intrinsic::arm_neon_vshifts :
2431 Intrinsic::arm_neon_vshiftu);
2432 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2433 DAG.getConstant(vshiftInt, MVT::i32),
2434 N->getOperand(0), NegatedCount);
2437 // We can get here for a node like i32 = ISD::SHL i32, i64
2441 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2442 "Unknown shift to lower!");
2444 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2445 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2446 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2449 // If we are in thumb mode, we don't have RRX.
2450 if (ST->isThumb1Only()) return SDValue();
2452 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2453 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2454 DAG.getConstant(0, MVT::i32));
2455 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2456 DAG.getConstant(1, MVT::i32));
2458 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2459 // captures the result into a carry flag.
2460 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2461 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2463 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2464 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2466 // Merge the pieces into a single i64 value.
2467 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2470 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2471 SDValue TmpOp0, TmpOp1;
2472 bool Invert = false;
2476 SDValue Op0 = Op.getOperand(0);
2477 SDValue Op1 = Op.getOperand(1);
2478 SDValue CC = Op.getOperand(2);
2479 EVT VT = Op.getValueType();
2480 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2481 DebugLoc dl = Op.getDebugLoc();
2483 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2484 switch (SetCCOpcode) {
2485 default: llvm_unreachable("Illegal FP comparison"); break;
2487 case ISD::SETNE: Invert = true; // Fallthrough
2489 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2491 case ISD::SETLT: Swap = true; // Fallthrough
2493 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2495 case ISD::SETLE: Swap = true; // Fallthrough
2497 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2498 case ISD::SETUGE: Swap = true; // Fallthrough
2499 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2500 case ISD::SETUGT: Swap = true; // Fallthrough
2501 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2502 case ISD::SETUEQ: Invert = true; // Fallthrough
2504 // Expand this to (OLT | OGT).
2508 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2509 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2511 case ISD::SETUO: Invert = true; // Fallthrough
2513 // Expand this to (OLT | OGE).
2517 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2518 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2522 // Integer comparisons.
2523 switch (SetCCOpcode) {
2524 default: llvm_unreachable("Illegal integer comparison"); break;
2525 case ISD::SETNE: Invert = true;
2526 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2527 case ISD::SETLT: Swap = true;
2528 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2529 case ISD::SETLE: Swap = true;
2530 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2531 case ISD::SETULT: Swap = true;
2532 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2533 case ISD::SETULE: Swap = true;
2534 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2537 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2538 if (Opc == ARMISD::VCEQ) {
2541 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2543 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2546 // Ignore bitconvert.
2547 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2548 AndOp = AndOp.getOperand(0);
2550 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2552 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2553 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2560 std::swap(Op0, Op1);
2562 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2565 Result = DAG.getNOT(dl, Result, VT);
2570 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2571 /// VMOV instruction, and if so, return the constant being splatted.
2572 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2573 unsigned SplatBitSize, SelectionDAG &DAG) {
2574 switch (SplatBitSize) {
2576 // Any 1-byte value is OK.
2577 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2578 return DAG.getTargetConstant(SplatBits, MVT::i8);
2581 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2582 if ((SplatBits & ~0xff) == 0 ||
2583 (SplatBits & ~0xff00) == 0)
2584 return DAG.getTargetConstant(SplatBits, MVT::i16);
2588 // NEON's 32-bit VMOV supports splat values where:
2589 // * only one byte is nonzero, or
2590 // * the least significant byte is 0xff and the second byte is nonzero, or
2591 // * the least significant 2 bytes are 0xff and the third is nonzero.
2592 if ((SplatBits & ~0xff) == 0 ||
2593 (SplatBits & ~0xff00) == 0 ||
2594 (SplatBits & ~0xff0000) == 0 ||
2595 (SplatBits & ~0xff000000) == 0)
2596 return DAG.getTargetConstant(SplatBits, MVT::i32);
2598 if ((SplatBits & ~0xffff) == 0 &&
2599 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2600 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2602 if ((SplatBits & ~0xffffff) == 0 &&
2603 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2604 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2606 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2607 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2608 // VMOV.I32. A (very) minor optimization would be to replicate the value
2609 // and fall through here to test for a valid 64-bit splat. But, then the
2610 // caller would also need to check and handle the change in size.
2614 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2615 uint64_t BitMask = 0xff;
2617 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2618 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2620 else if ((SplatBits & BitMask) != 0)
2624 return DAG.getTargetConstant(Val, MVT::i64);
2628 llvm_unreachable("unexpected size for isVMOVSplat");
2635 /// getVMOVImm - If this is a build_vector of constants which can be
2636 /// formed by using a VMOV instruction of the specified element size,
2637 /// return the constant being splatted. The ByteSize field indicates the
2638 /// number of bytes of each element [1248].
2639 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2640 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2641 APInt SplatBits, SplatUndef;
2642 unsigned SplatBitSize;
2644 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2645 HasAnyUndefs, ByteSize * 8))
2648 if (SplatBitSize > ByteSize * 8)
2651 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2655 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2656 bool &ReverseVEXT, unsigned &Imm) {
2657 unsigned NumElts = VT.getVectorNumElements();
2658 ReverseVEXT = false;
2661 // If this is a VEXT shuffle, the immediate value is the index of the first
2662 // element. The other shuffle indices must be the successive elements after
2664 unsigned ExpectedElt = Imm;
2665 for (unsigned i = 1; i < NumElts; ++i) {
2666 // Increment the expected index. If it wraps around, it may still be
2667 // a VEXT but the source vectors must be swapped.
2669 if (ExpectedElt == NumElts * 2) {
2674 if (ExpectedElt != static_cast<unsigned>(M[i]))
2678 // Adjust the index value if the source operands will be swapped.
2685 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2686 /// instruction with the specified blocksize. (The order of the elements
2687 /// within each block of the vector is reversed.)
2688 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2689 unsigned BlockSize) {
2690 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2691 "Only possible block sizes for VREV are: 16, 32, 64");
2693 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2697 unsigned NumElts = VT.getVectorNumElements();
2698 unsigned BlockElts = M[0] + 1;
2700 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2703 for (unsigned i = 0; i < NumElts; ++i) {
2704 if ((unsigned) M[i] !=
2705 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2712 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2713 unsigned &WhichResult) {
2714 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2718 unsigned NumElts = VT.getVectorNumElements();
2719 WhichResult = (M[0] == 0 ? 0 : 1);
2720 for (unsigned i = 0; i < NumElts; i += 2) {
2721 if ((unsigned) M[i] != i + WhichResult ||
2722 (unsigned) M[i+1] != i + NumElts + WhichResult)
2728 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2729 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2730 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2731 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2732 unsigned &WhichResult) {
2733 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2737 unsigned NumElts = VT.getVectorNumElements();
2738 WhichResult = (M[0] == 0 ? 0 : 1);
2739 for (unsigned i = 0; i < NumElts; i += 2) {
2740 if ((unsigned) M[i] != i + WhichResult ||
2741 (unsigned) M[i+1] != i + WhichResult)
2747 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2748 unsigned &WhichResult) {
2749 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2753 unsigned NumElts = VT.getVectorNumElements();
2754 WhichResult = (M[0] == 0 ? 0 : 1);
2755 for (unsigned i = 0; i != NumElts; ++i) {
2756 if ((unsigned) M[i] != 2 * i + WhichResult)
2760 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2761 if (VT.is64BitVector() && EltSz == 32)
2767 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2768 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2769 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2770 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2771 unsigned &WhichResult) {
2772 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2776 unsigned Half = VT.getVectorNumElements() / 2;
2777 WhichResult = (M[0] == 0 ? 0 : 1);
2778 for (unsigned j = 0; j != 2; ++j) {
2779 unsigned Idx = WhichResult;
2780 for (unsigned i = 0; i != Half; ++i) {
2781 if ((unsigned) M[i + j * Half] != Idx)
2787 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2788 if (VT.is64BitVector() && EltSz == 32)
2794 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2795 unsigned &WhichResult) {
2796 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2800 unsigned NumElts = VT.getVectorNumElements();
2801 WhichResult = (M[0] == 0 ? 0 : 1);
2802 unsigned Idx = WhichResult * NumElts / 2;
2803 for (unsigned i = 0; i != NumElts; i += 2) {
2804 if ((unsigned) M[i] != Idx ||
2805 (unsigned) M[i+1] != Idx + NumElts)
2810 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2811 if (VT.is64BitVector() && EltSz == 32)
2817 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2818 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2819 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2820 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2821 unsigned &WhichResult) {
2822 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2826 unsigned NumElts = VT.getVectorNumElements();
2827 WhichResult = (M[0] == 0 ? 0 : 1);
2828 unsigned Idx = WhichResult * NumElts / 2;
2829 for (unsigned i = 0; i != NumElts; i += 2) {
2830 if ((unsigned) M[i] != Idx ||
2831 (unsigned) M[i+1] != Idx)
2836 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2837 if (VT.is64BitVector() && EltSz == 32)
2844 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2845 // Canonicalize all-zeros and all-ones vectors.
2846 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2847 if (ConstVal->isNullValue())
2848 return getZeroVector(VT, DAG, dl);
2849 if (ConstVal->isAllOnesValue())
2850 return getOnesVector(VT, DAG, dl);
2853 if (VT.is64BitVector()) {
2854 switch (Val.getValueType().getSizeInBits()) {
2855 case 8: CanonicalVT = MVT::v8i8; break;
2856 case 16: CanonicalVT = MVT::v4i16; break;
2857 case 32: CanonicalVT = MVT::v2i32; break;
2858 case 64: CanonicalVT = MVT::v1i64; break;
2859 default: llvm_unreachable("unexpected splat element type"); break;
2862 assert(VT.is128BitVector() && "unknown splat vector size");
2863 switch (Val.getValueType().getSizeInBits()) {
2864 case 8: CanonicalVT = MVT::v16i8; break;
2865 case 16: CanonicalVT = MVT::v8i16; break;
2866 case 32: CanonicalVT = MVT::v4i32; break;
2867 case 64: CanonicalVT = MVT::v2i64; break;
2868 default: llvm_unreachable("unexpected splat element type"); break;
2872 // Build a canonical splat for this value.
2873 SmallVector<SDValue, 8> Ops;
2874 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2875 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2877 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2880 // If this is a case we can't handle, return null and let the default
2881 // expansion code take care of it.
2882 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2883 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2884 DebugLoc dl = Op.getDebugLoc();
2885 EVT VT = Op.getValueType();
2887 APInt SplatBits, SplatUndef;
2888 unsigned SplatBitSize;
2890 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2891 if (SplatBitSize <= 64) {
2892 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2893 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2895 return BuildSplat(Val, VT, DAG, dl);
2899 // If there are only 2 elements in a 128-bit vector, insert them into an
2900 // undef vector. This handles the common case for 128-bit vector argument
2901 // passing, where the insertions should be translated to subreg accesses
2902 // with no real instructions.
2903 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2904 SDValue Val = DAG.getUNDEF(VT);
2905 SDValue Op0 = Op.getOperand(0);
2906 SDValue Op1 = Op.getOperand(1);
2907 if (Op0.getOpcode() != ISD::UNDEF)
2908 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2909 DAG.getIntPtrConstant(0));
2910 if (Op1.getOpcode() != ISD::UNDEF)
2911 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2912 DAG.getIntPtrConstant(1));
2919 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2920 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2921 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2922 /// are assumed to be legal.
2924 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2926 if (VT.getVectorNumElements() == 4 &&
2927 (VT.is128BitVector() || VT.is64BitVector())) {
2928 unsigned PFIndexes[4];
2929 for (unsigned i = 0; i != 4; ++i) {
2933 PFIndexes[i] = M[i];
2936 // Compute the index in the perfect shuffle table.
2937 unsigned PFTableIndex =
2938 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2939 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2940 unsigned Cost = (PFEntry >> 30);
2947 unsigned Imm, WhichResult;
2949 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2950 isVREVMask(M, VT, 64) ||
2951 isVREVMask(M, VT, 32) ||
2952 isVREVMask(M, VT, 16) ||
2953 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2954 isVTRNMask(M, VT, WhichResult) ||
2955 isVUZPMask(M, VT, WhichResult) ||
2956 isVZIPMask(M, VT, WhichResult) ||
2957 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2958 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2959 isVZIP_v_undef_Mask(M, VT, WhichResult));
2962 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2963 /// the specified operations to build the shuffle.
2964 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2965 SDValue RHS, SelectionDAG &DAG,
2967 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2968 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2969 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2972 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2981 OP_VUZPL, // VUZP, left result
2982 OP_VUZPR, // VUZP, right result
2983 OP_VZIPL, // VZIP, left result
2984 OP_VZIPR, // VZIP, right result
2985 OP_VTRNL, // VTRN, left result
2986 OP_VTRNR // VTRN, right result
2989 if (OpNum == OP_COPY) {
2990 if (LHSID == (1*9+2)*9+3) return LHS;
2991 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2995 SDValue OpLHS, OpRHS;
2996 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2997 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2998 EVT VT = OpLHS.getValueType();
3001 default: llvm_unreachable("Unknown shuffle opcode!");
3003 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3008 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3009 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3013 return DAG.getNode(ARMISD::VEXT, dl, VT,
3015 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3018 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3019 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3022 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3023 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3026 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3027 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3031 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3032 SDValue V1 = Op.getOperand(0);
3033 SDValue V2 = Op.getOperand(1);
3034 DebugLoc dl = Op.getDebugLoc();
3035 EVT VT = Op.getValueType();
3036 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3037 SmallVector<int, 8> ShuffleMask;
3039 // Convert shuffles that are directly supported on NEON to target-specific
3040 // DAG nodes, instead of keeping them as shuffles and matching them again
3041 // during code selection. This is more efficient and avoids the possibility
3042 // of inconsistencies between legalization and selection.
3043 // FIXME: floating-point vectors should be canonicalized to integer vectors
3044 // of the same time so that they get CSEd properly.
3045 SVN->getMask(ShuffleMask);
3047 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3048 int Lane = SVN->getSplatIndex();
3049 // If this is undef splat, generate it via "just" vdup, if possible.
3050 if (Lane == -1) Lane = 0;
3052 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3053 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3055 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3056 DAG.getConstant(Lane, MVT::i32));
3061 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3064 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3065 DAG.getConstant(Imm, MVT::i32));
3068 if (isVREVMask(ShuffleMask, VT, 64))
3069 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3070 if (isVREVMask(ShuffleMask, VT, 32))
3071 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3072 if (isVREVMask(ShuffleMask, VT, 16))
3073 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3075 // Check for Neon shuffles that modify both input vectors in place.
3076 // If both results are used, i.e., if there are two shuffles with the same
3077 // source operands and with masks corresponding to both results of one of
3078 // these operations, DAG memoization will ensure that a single node is
3079 // used for both shuffles.
3080 unsigned WhichResult;
3081 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3082 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3083 V1, V2).getValue(WhichResult);
3084 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3085 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3086 V1, V2).getValue(WhichResult);
3087 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3088 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3089 V1, V2).getValue(WhichResult);
3091 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3092 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3093 V1, V1).getValue(WhichResult);
3094 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3095 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3096 V1, V1).getValue(WhichResult);
3097 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3098 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3099 V1, V1).getValue(WhichResult);
3101 // If the shuffle is not directly supported and it has 4 elements, use
3102 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3103 if (VT.getVectorNumElements() == 4 &&
3104 (VT.is128BitVector() || VT.is64BitVector())) {
3105 unsigned PFIndexes[4];
3106 for (unsigned i = 0; i != 4; ++i) {
3107 if (ShuffleMask[i] < 0)
3110 PFIndexes[i] = ShuffleMask[i];
3113 // Compute the index in the perfect shuffle table.
3114 unsigned PFTableIndex =
3115 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3117 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3118 unsigned Cost = (PFEntry >> 30);
3121 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3127 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3128 EVT VT = Op.getValueType();
3129 DebugLoc dl = Op.getDebugLoc();
3130 SDValue Vec = Op.getOperand(0);
3131 SDValue Lane = Op.getOperand(1);
3132 assert(VT == MVT::i32 &&
3133 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3134 "unexpected type for custom-lowering vector extract");
3135 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3138 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3139 // The only time a CONCAT_VECTORS operation can have legal types is when
3140 // two 64-bit vectors are concatenated to a 128-bit vector.
3141 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3142 "unexpected CONCAT_VECTORS");
3143 DebugLoc dl = Op.getDebugLoc();
3144 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3145 SDValue Op0 = Op.getOperand(0);
3146 SDValue Op1 = Op.getOperand(1);
3147 if (Op0.getOpcode() != ISD::UNDEF)
3148 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3149 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3150 DAG.getIntPtrConstant(0));
3151 if (Op1.getOpcode() != ISD::UNDEF)
3152 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3153 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3154 DAG.getIntPtrConstant(1));
3155 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3158 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3159 switch (Op.getOpcode()) {
3160 default: llvm_unreachable("Don't know how to custom lower this!");
3161 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3162 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3163 case ISD::GlobalAddress:
3164 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3165 LowerGlobalAddressELF(Op, DAG);
3166 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3167 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3168 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3169 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3170 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3171 case ISD::VASTART: return LowerVASTART(Op, DAG);
3172 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3173 case ISD::SINT_TO_FP:
3174 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3175 case ISD::FP_TO_SINT:
3176 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3177 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3178 case ISD::RETURNADDR: break;
3179 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3180 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3181 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3183 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3186 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3187 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3188 case ISD::SRL_PARTS:
3189 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3190 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3191 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3192 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3193 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3194 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3195 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3200 /// ReplaceNodeResults - Replace the results of node with an illegal result
3201 /// type with new values built out of custom code.
3202 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3203 SmallVectorImpl<SDValue>&Results,
3204 SelectionDAG &DAG) const {
3206 switch (N->getOpcode()) {
3208 llvm_unreachable("Don't know how to custom expand this!");
3210 case ISD::BIT_CONVERT:
3211 Res = ExpandBIT_CONVERT(N, DAG);
3215 Res = LowerShift(N, DAG, Subtarget);
3219 Results.push_back(Res);
3222 //===----------------------------------------------------------------------===//
3223 // ARM Scheduler Hooks
3224 //===----------------------------------------------------------------------===//
3227 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3228 MachineBasicBlock *BB,
3229 unsigned Size) const {
3230 unsigned dest = MI->getOperand(0).getReg();
3231 unsigned ptr = MI->getOperand(1).getReg();
3232 unsigned oldval = MI->getOperand(2).getReg();
3233 unsigned newval = MI->getOperand(3).getReg();
3234 unsigned scratch = BB->getParent()->getRegInfo()
3235 .createVirtualRegister(ARM::GPRRegisterClass);
3236 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3237 DebugLoc dl = MI->getDebugLoc();
3238 bool isThumb2 = Subtarget->isThumb2();
3240 unsigned ldrOpc, strOpc;
3242 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3244 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3245 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3248 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3249 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3252 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3253 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3257 MachineFunction *MF = BB->getParent();
3258 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3259 MachineFunction::iterator It = BB;
3260 ++It; // insert the new blocks after the current block
3262 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3263 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3264 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3265 MF->insert(It, loop1MBB);
3266 MF->insert(It, loop2MBB);
3267 MF->insert(It, exitMBB);
3268 exitMBB->transferSuccessors(BB);
3272 // fallthrough --> loop1MBB
3273 BB->addSuccessor(loop1MBB);
3276 // ldrex dest, [ptr]
3280 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3281 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3282 .addReg(dest).addReg(oldval));
3283 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3284 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3285 BB->addSuccessor(loop2MBB);
3286 BB->addSuccessor(exitMBB);
3289 // strex scratch, newval, [ptr]
3293 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3295 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3296 .addReg(scratch).addImm(0));
3297 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3298 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3299 BB->addSuccessor(loop1MBB);
3300 BB->addSuccessor(exitMBB);
3306 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3312 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3313 unsigned Size, unsigned BinOpcode) const {
3314 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3315 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3317 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3318 MachineFunction *MF = BB->getParent();
3319 MachineFunction::iterator It = BB;
3322 unsigned dest = MI->getOperand(0).getReg();
3323 unsigned ptr = MI->getOperand(1).getReg();
3324 unsigned incr = MI->getOperand(2).getReg();
3325 DebugLoc dl = MI->getDebugLoc();
3327 bool isThumb2 = Subtarget->isThumb2();
3328 unsigned ldrOpc, strOpc;
3330 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3332 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3333 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3336 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3337 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3340 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3341 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3345 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3346 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3347 MF->insert(It, loopMBB);
3348 MF->insert(It, exitMBB);
3349 exitMBB->transferSuccessors(BB);
3351 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3352 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3353 unsigned scratch2 = (!BinOpcode) ? incr :
3354 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3358 // fallthrough --> loopMBB
3359 BB->addSuccessor(loopMBB);
3363 // <binop> scratch2, dest, incr
3364 // strex scratch, scratch2, ptr
3367 // fallthrough --> exitMBB
3369 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3371 // operand order needs to go the other way for NAND
3372 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3373 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3374 addReg(incr).addReg(dest)).addReg(0);
3376 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3377 addReg(dest).addReg(incr)).addReg(0);
3380 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3382 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3383 .addReg(scratch).addImm(0));
3384 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3385 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3387 BB->addSuccessor(loopMBB);
3388 BB->addSuccessor(exitMBB);
3394 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3400 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3401 MachineBasicBlock *BB) const {
3402 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3403 DebugLoc dl = MI->getDebugLoc();
3404 bool isThumb2 = Subtarget->isThumb2();
3405 switch (MI->getOpcode()) {
3408 llvm_unreachable("Unexpected instr type to insert");
3410 case ARM::ATOMIC_LOAD_ADD_I8:
3411 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3412 case ARM::ATOMIC_LOAD_ADD_I16:
3413 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3414 case ARM::ATOMIC_LOAD_ADD_I32:
3415 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3417 case ARM::ATOMIC_LOAD_AND_I8:
3418 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3419 case ARM::ATOMIC_LOAD_AND_I16:
3420 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3421 case ARM::ATOMIC_LOAD_AND_I32:
3422 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3424 case ARM::ATOMIC_LOAD_OR_I8:
3425 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3426 case ARM::ATOMIC_LOAD_OR_I16:
3427 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3428 case ARM::ATOMIC_LOAD_OR_I32:
3429 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3431 case ARM::ATOMIC_LOAD_XOR_I8:
3432 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3433 case ARM::ATOMIC_LOAD_XOR_I16:
3434 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3435 case ARM::ATOMIC_LOAD_XOR_I32:
3436 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3438 case ARM::ATOMIC_LOAD_NAND_I8:
3439 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3440 case ARM::ATOMIC_LOAD_NAND_I16:
3441 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3442 case ARM::ATOMIC_LOAD_NAND_I32:
3443 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3445 case ARM::ATOMIC_LOAD_SUB_I8:
3446 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3447 case ARM::ATOMIC_LOAD_SUB_I16:
3448 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3449 case ARM::ATOMIC_LOAD_SUB_I32:
3450 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3452 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3453 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3454 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3456 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3457 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3458 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3460 case ARM::tMOVCCr_pseudo: {
3461 // To "insert" a SELECT_CC instruction, we actually have to insert the
3462 // diamond control-flow pattern. The incoming instruction knows the
3463 // destination vreg to set, the condition code register to branch on, the
3464 // true/false values to select between, and a branch opcode to use.
3465 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3466 MachineFunction::iterator It = BB;
3472 // cmpTY ccX, r1, r2
3474 // fallthrough --> copy0MBB
3475 MachineBasicBlock *thisMBB = BB;
3476 MachineFunction *F = BB->getParent();
3477 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3478 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3479 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3480 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3481 F->insert(It, copy0MBB);
3482 F->insert(It, sinkMBB);
3483 // Update machine-CFG edges by first adding all successors of the current
3484 // block to the new block which will contain the Phi node for the select.
3485 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3486 E = BB->succ_end(); I != E; ++I)
3487 sinkMBB->addSuccessor(*I);
3488 // Next, remove all successors of the current block, and add the true
3489 // and fallthrough blocks as its successors.
3490 while (!BB->succ_empty())
3491 BB->removeSuccessor(BB->succ_begin());
3492 BB->addSuccessor(copy0MBB);
3493 BB->addSuccessor(sinkMBB);
3496 // %FalseValue = ...
3497 // # fallthrough to sinkMBB
3500 // Update machine-CFG edges
3501 BB->addSuccessor(sinkMBB);
3504 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3507 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3508 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3509 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3511 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3518 case ARM::t2SUBrSPi_:
3519 case ARM::t2SUBrSPi12_:
3520 case ARM::t2SUBrSPs_: {
3521 MachineFunction *MF = BB->getParent();
3522 unsigned DstReg = MI->getOperand(0).getReg();
3523 unsigned SrcReg = MI->getOperand(1).getReg();
3524 bool DstIsDead = MI->getOperand(0).isDead();
3525 bool SrcIsKill = MI->getOperand(1).isKill();
3527 if (SrcReg != ARM::SP) {
3528 // Copy the source to SP from virtual register.
3529 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3530 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3531 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3532 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3533 .addReg(SrcReg, getKillRegState(SrcIsKill));
3537 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3538 switch (MI->getOpcode()) {
3540 llvm_unreachable("Unexpected pseudo instruction!");
3546 OpOpc = ARM::tADDspr;
3549 OpOpc = ARM::tSUBspi;
3551 case ARM::t2SUBrSPi_:
3552 OpOpc = ARM::t2SUBrSPi;
3553 NeedPred = true; NeedCC = true;
3555 case ARM::t2SUBrSPi12_:
3556 OpOpc = ARM::t2SUBrSPi12;
3559 case ARM::t2SUBrSPs_:
3560 OpOpc = ARM::t2SUBrSPs;
3561 NeedPred = true; NeedCC = true; NeedOp3 = true;
3564 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3565 if (OpOpc == ARM::tAND)
3566 AddDefaultT1CC(MIB);
3567 MIB.addReg(ARM::SP);
3568 MIB.addOperand(MI->getOperand(2));
3570 MIB.addOperand(MI->getOperand(3));
3572 AddDefaultPred(MIB);
3576 // Copy the result from SP to virtual register.
3577 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3578 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3579 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3580 BuildMI(BB, dl, TII->get(CopyOpc))
3581 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3583 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3589 //===----------------------------------------------------------------------===//
3590 // ARM Optimization Hooks
3591 //===----------------------------------------------------------------------===//
3594 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3595 TargetLowering::DAGCombinerInfo &DCI) {
3596 SelectionDAG &DAG = DCI.DAG;
3597 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3598 EVT VT = N->getValueType(0);
3599 unsigned Opc = N->getOpcode();
3600 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3601 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3602 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3603 ISD::CondCode CC = ISD::SETCC_INVALID;
3606 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3608 SDValue CCOp = Slct.getOperand(0);
3609 if (CCOp.getOpcode() == ISD::SETCC)
3610 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3613 bool DoXform = false;
3615 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3618 if (LHS.getOpcode() == ISD::Constant &&
3619 cast<ConstantSDNode>(LHS)->isNullValue()) {
3621 } else if (CC != ISD::SETCC_INVALID &&
3622 RHS.getOpcode() == ISD::Constant &&
3623 cast<ConstantSDNode>(RHS)->isNullValue()) {
3624 std::swap(LHS, RHS);
3625 SDValue Op0 = Slct.getOperand(0);
3626 EVT OpVT = isSlctCC ? Op0.getValueType() :
3627 Op0.getOperand(0).getValueType();
3628 bool isInt = OpVT.isInteger();
3629 CC = ISD::getSetCCInverse(CC, isInt);
3631 if (!TLI.isCondCodeLegal(CC, OpVT))
3632 return SDValue(); // Inverse operator isn't legal.
3639 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3641 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3642 Slct.getOperand(0), Slct.getOperand(1), CC);
3643 SDValue CCOp = Slct.getOperand(0);
3645 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3646 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3647 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3648 CCOp, OtherOp, Result);
3653 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3654 static SDValue PerformADDCombine(SDNode *N,
3655 TargetLowering::DAGCombinerInfo &DCI) {
3656 // added by evan in r37685 with no testcase.
3657 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3659 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3660 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3661 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3662 if (Result.getNode()) return Result;
3664 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3665 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3666 if (Result.getNode()) return Result;
3672 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3673 static SDValue PerformSUBCombine(SDNode *N,
3674 TargetLowering::DAGCombinerInfo &DCI) {
3675 // added by evan in r37685 with no testcase.
3676 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3678 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3679 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3680 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3681 if (Result.getNode()) return Result;
3687 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3688 /// ARMISD::VMOVRRD.
3689 static SDValue PerformVMOVRRDCombine(SDNode *N,
3690 TargetLowering::DAGCombinerInfo &DCI) {
3691 // fmrrd(fmdrr x, y) -> x,y
3692 SDValue InDouble = N->getOperand(0);
3693 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3694 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3698 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3699 /// operand of a vector shift operation, where all the elements of the
3700 /// build_vector must have the same constant integer value.
3701 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3702 // Ignore bit_converts.
3703 while (Op.getOpcode() == ISD::BIT_CONVERT)
3704 Op = Op.getOperand(0);
3705 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3706 APInt SplatBits, SplatUndef;
3707 unsigned SplatBitSize;
3709 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3710 HasAnyUndefs, ElementBits) ||
3711 SplatBitSize > ElementBits)
3713 Cnt = SplatBits.getSExtValue();
3717 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3718 /// operand of a vector shift left operation. That value must be in the range:
3719 /// 0 <= Value < ElementBits for a left shift; or
3720 /// 0 <= Value <= ElementBits for a long left shift.
3721 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3722 assert(VT.isVector() && "vector shift count is not a vector type");
3723 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3724 if (! getVShiftImm(Op, ElementBits, Cnt))
3726 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3729 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3730 /// operand of a vector shift right operation. For a shift opcode, the value
3731 /// is positive, but for an intrinsic the value count must be negative. The
3732 /// absolute value must be in the range:
3733 /// 1 <= |Value| <= ElementBits for a right shift; or
3734 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3735 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3737 assert(VT.isVector() && "vector shift count is not a vector type");
3738 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3739 if (! getVShiftImm(Op, ElementBits, Cnt))
3743 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3746 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3747 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3748 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3751 // Don't do anything for most intrinsics.
3754 // Vector shifts: check for immediate versions and lower them.
3755 // Note: This is done during DAG combining instead of DAG legalizing because
3756 // the build_vectors for 64-bit vector element shift counts are generally
3757 // not legal, and it is hard to see their values after they get legalized to
3758 // loads from a constant pool.
3759 case Intrinsic::arm_neon_vshifts:
3760 case Intrinsic::arm_neon_vshiftu:
3761 case Intrinsic::arm_neon_vshiftls:
3762 case Intrinsic::arm_neon_vshiftlu:
3763 case Intrinsic::arm_neon_vshiftn:
3764 case Intrinsic::arm_neon_vrshifts:
3765 case Intrinsic::arm_neon_vrshiftu:
3766 case Intrinsic::arm_neon_vrshiftn:
3767 case Intrinsic::arm_neon_vqshifts:
3768 case Intrinsic::arm_neon_vqshiftu:
3769 case Intrinsic::arm_neon_vqshiftsu:
3770 case Intrinsic::arm_neon_vqshiftns:
3771 case Intrinsic::arm_neon_vqshiftnu:
3772 case Intrinsic::arm_neon_vqshiftnsu:
3773 case Intrinsic::arm_neon_vqrshiftns:
3774 case Intrinsic::arm_neon_vqrshiftnu:
3775 case Intrinsic::arm_neon_vqrshiftnsu: {
3776 EVT VT = N->getOperand(1).getValueType();
3778 unsigned VShiftOpc = 0;
3781 case Intrinsic::arm_neon_vshifts:
3782 case Intrinsic::arm_neon_vshiftu:
3783 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3784 VShiftOpc = ARMISD::VSHL;
3787 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3788 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3789 ARMISD::VSHRs : ARMISD::VSHRu);
3794 case Intrinsic::arm_neon_vshiftls:
3795 case Intrinsic::arm_neon_vshiftlu:
3796 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3798 llvm_unreachable("invalid shift count for vshll intrinsic");
3800 case Intrinsic::arm_neon_vrshifts:
3801 case Intrinsic::arm_neon_vrshiftu:
3802 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3806 case Intrinsic::arm_neon_vqshifts:
3807 case Intrinsic::arm_neon_vqshiftu:
3808 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3812 case Intrinsic::arm_neon_vqshiftsu:
3813 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3815 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3817 case Intrinsic::arm_neon_vshiftn:
3818 case Intrinsic::arm_neon_vrshiftn:
3819 case Intrinsic::arm_neon_vqshiftns:
3820 case Intrinsic::arm_neon_vqshiftnu:
3821 case Intrinsic::arm_neon_vqshiftnsu:
3822 case Intrinsic::arm_neon_vqrshiftns:
3823 case Intrinsic::arm_neon_vqrshiftnu:
3824 case Intrinsic::arm_neon_vqrshiftnsu:
3825 // Narrowing shifts require an immediate right shift.
3826 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3828 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3831 llvm_unreachable("unhandled vector shift");
3835 case Intrinsic::arm_neon_vshifts:
3836 case Intrinsic::arm_neon_vshiftu:
3837 // Opcode already set above.
3839 case Intrinsic::arm_neon_vshiftls:
3840 case Intrinsic::arm_neon_vshiftlu:
3841 if (Cnt == VT.getVectorElementType().getSizeInBits())
3842 VShiftOpc = ARMISD::VSHLLi;
3844 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3845 ARMISD::VSHLLs : ARMISD::VSHLLu);
3847 case Intrinsic::arm_neon_vshiftn:
3848 VShiftOpc = ARMISD::VSHRN; break;
3849 case Intrinsic::arm_neon_vrshifts:
3850 VShiftOpc = ARMISD::VRSHRs; break;
3851 case Intrinsic::arm_neon_vrshiftu:
3852 VShiftOpc = ARMISD::VRSHRu; break;
3853 case Intrinsic::arm_neon_vrshiftn:
3854 VShiftOpc = ARMISD::VRSHRN; break;
3855 case Intrinsic::arm_neon_vqshifts:
3856 VShiftOpc = ARMISD::VQSHLs; break;
3857 case Intrinsic::arm_neon_vqshiftu:
3858 VShiftOpc = ARMISD::VQSHLu; break;
3859 case Intrinsic::arm_neon_vqshiftsu:
3860 VShiftOpc = ARMISD::VQSHLsu; break;
3861 case Intrinsic::arm_neon_vqshiftns:
3862 VShiftOpc = ARMISD::VQSHRNs; break;
3863 case Intrinsic::arm_neon_vqshiftnu:
3864 VShiftOpc = ARMISD::VQSHRNu; break;
3865 case Intrinsic::arm_neon_vqshiftnsu:
3866 VShiftOpc = ARMISD::VQSHRNsu; break;
3867 case Intrinsic::arm_neon_vqrshiftns:
3868 VShiftOpc = ARMISD::VQRSHRNs; break;
3869 case Intrinsic::arm_neon_vqrshiftnu:
3870 VShiftOpc = ARMISD::VQRSHRNu; break;
3871 case Intrinsic::arm_neon_vqrshiftnsu:
3872 VShiftOpc = ARMISD::VQRSHRNsu; break;
3875 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3876 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3879 case Intrinsic::arm_neon_vshiftins: {
3880 EVT VT = N->getOperand(1).getValueType();
3882 unsigned VShiftOpc = 0;
3884 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3885 VShiftOpc = ARMISD::VSLI;
3886 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3887 VShiftOpc = ARMISD::VSRI;
3889 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3892 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3893 N->getOperand(1), N->getOperand(2),
3894 DAG.getConstant(Cnt, MVT::i32));
3897 case Intrinsic::arm_neon_vqrshifts:
3898 case Intrinsic::arm_neon_vqrshiftu:
3899 // No immediate versions of these to check for.
3906 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3907 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3908 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3909 /// vector element shift counts are generally not legal, and it is hard to see
3910 /// their values after they get legalized to loads from a constant pool.
3911 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3912 const ARMSubtarget *ST) {
3913 EVT VT = N->getValueType(0);
3915 // Nothing to be done for scalar shifts.
3916 if (! VT.isVector())
3919 assert(ST->hasNEON() && "unexpected vector shift");
3922 switch (N->getOpcode()) {
3923 default: llvm_unreachable("unexpected shift opcode");
3926 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3927 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3928 DAG.getConstant(Cnt, MVT::i32));
3933 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3934 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3935 ARMISD::VSHRs : ARMISD::VSHRu);
3936 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3937 DAG.getConstant(Cnt, MVT::i32));
3943 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3944 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3945 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3946 const ARMSubtarget *ST) {
3947 SDValue N0 = N->getOperand(0);
3949 // Check for sign- and zero-extensions of vector extract operations of 8-
3950 // and 16-bit vector elements. NEON supports these directly. They are
3951 // handled during DAG combining because type legalization will promote them
3952 // to 32-bit types and it is messy to recognize the operations after that.
3953 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3954 SDValue Vec = N0.getOperand(0);
3955 SDValue Lane = N0.getOperand(1);
3956 EVT VT = N->getValueType(0);
3957 EVT EltVT = N0.getValueType();
3958 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3960 if (VT == MVT::i32 &&
3961 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3962 TLI.isTypeLegal(Vec.getValueType())) {
3965 switch (N->getOpcode()) {
3966 default: llvm_unreachable("unexpected opcode");
3967 case ISD::SIGN_EXTEND:
3968 Opc = ARMISD::VGETLANEs;
3970 case ISD::ZERO_EXTEND:
3971 case ISD::ANY_EXTEND:
3972 Opc = ARMISD::VGETLANEu;
3975 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3982 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3983 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3984 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3985 const ARMSubtarget *ST) {
3986 // If the target supports NEON, try to use vmax/vmin instructions for f32
3987 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3988 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3989 // a NaN; only do the transformation when it matches that behavior.
3991 // For now only do this when using NEON for FP operations; if using VFP, it
3992 // is not obvious that the benefit outweighs the cost of switching to the
3994 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3995 N->getValueType(0) != MVT::f32)
3998 SDValue CondLHS = N->getOperand(0);
3999 SDValue CondRHS = N->getOperand(1);
4000 SDValue LHS = N->getOperand(2);
4001 SDValue RHS = N->getOperand(3);
4002 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4004 unsigned Opcode = 0;
4006 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4007 IsReversed = false; // x CC y ? x : y
4008 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4009 IsReversed = true ; // x CC y ? y : x
4023 // If LHS is NaN, an ordered comparison will be false and the result will
4024 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4025 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4026 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4027 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4029 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4030 // will return -0, so vmin can only be used for unsafe math or if one of
4031 // the operands is known to be nonzero.
4032 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4034 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4036 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4045 // If LHS is NaN, an ordered comparison will be false and the result will
4046 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4047 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4048 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4049 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4051 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4052 // will return +0, so vmax can only be used for unsafe math or if one of
4053 // the operands is known to be nonzero.
4054 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4056 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4058 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4064 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4067 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4068 DAGCombinerInfo &DCI) const {
4069 switch (N->getOpcode()) {
4071 case ISD::ADD: return PerformADDCombine(N, DCI);
4072 case ISD::SUB: return PerformSUBCombine(N, DCI);
4073 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4074 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4077 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4078 case ISD::SIGN_EXTEND:
4079 case ISD::ZERO_EXTEND:
4080 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4081 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4086 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4087 if (!Subtarget->hasV6Ops())
4088 // Pre-v6 does not support unaligned mem access.
4091 // v6+ may or may not support unaligned mem access depending on the system
4093 // FIXME: This is pretty conservative. Should we provide cmdline option to
4094 // control the behaviour?
4095 if (!Subtarget->isTargetDarwin())
4099 switch (VT.getSimpleVT().SimpleTy) {
4106 // FIXME: VLD1 etc with standard alignment is legal.
4110 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4115 switch (VT.getSimpleVT().SimpleTy) {
4116 default: return false;
4131 if ((V & (Scale - 1)) != 0)
4134 return V == (V & ((1LL << 5) - 1));
4137 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4138 const ARMSubtarget *Subtarget) {
4145 switch (VT.getSimpleVT().SimpleTy) {
4146 default: return false;
4151 // + imm12 or - imm8
4153 return V == (V & ((1LL << 8) - 1));
4154 return V == (V & ((1LL << 12) - 1));
4157 // Same as ARM mode. FIXME: NEON?
4158 if (!Subtarget->hasVFP2())
4163 return V == (V & ((1LL << 8) - 1));
4167 /// isLegalAddressImmediate - Return true if the integer value can be used
4168 /// as the offset of the target addressing mode for load / store of the
4170 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4171 const ARMSubtarget *Subtarget) {
4178 if (Subtarget->isThumb1Only())
4179 return isLegalT1AddressImmediate(V, VT);
4180 else if (Subtarget->isThumb2())
4181 return isLegalT2AddressImmediate(V, VT, Subtarget);
4186 switch (VT.getSimpleVT().SimpleTy) {
4187 default: return false;
4192 return V == (V & ((1LL << 12) - 1));
4195 return V == (V & ((1LL << 8) - 1));
4198 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4203 return V == (V & ((1LL << 8) - 1));
4207 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4209 int Scale = AM.Scale;
4213 switch (VT.getSimpleVT().SimpleTy) {
4214 default: return false;
4223 return Scale == 2 || Scale == 4 || Scale == 8;
4226 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4230 // Note, we allow "void" uses (basically, uses that aren't loads or
4231 // stores), because arm allows folding a scale into many arithmetic
4232 // operations. This should be made more precise and revisited later.
4234 // Allow r << imm, but the imm has to be a multiple of two.
4235 if (Scale & 1) return false;
4236 return isPowerOf2_32(Scale);
4240 /// isLegalAddressingMode - Return true if the addressing mode represented
4241 /// by AM is legal for this target, for a load/store of the specified type.
4242 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4243 const Type *Ty) const {
4244 EVT VT = getValueType(Ty, true);
4245 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4248 // Can never fold addr of global into load/store.
4253 case 0: // no scale reg, must be "r+i" or "r", or "i".
4256 if (Subtarget->isThumb1Only())
4260 // ARM doesn't support any R+R*scale+imm addr modes.
4267 if (Subtarget->isThumb2())
4268 return isLegalT2ScaledAddressingMode(AM, VT);
4270 int Scale = AM.Scale;
4271 switch (VT.getSimpleVT().SimpleTy) {
4272 default: return false;
4276 if (Scale < 0) Scale = -Scale;
4280 return isPowerOf2_32(Scale & ~1);
4284 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4289 // Note, we allow "void" uses (basically, uses that aren't loads or
4290 // stores), because arm allows folding a scale into many arithmetic
4291 // operations. This should be made more precise and revisited later.
4293 // Allow r << imm, but the imm has to be a multiple of two.
4294 if (Scale & 1) return false;
4295 return isPowerOf2_32(Scale);
4302 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4303 /// icmp immediate, that is the target has icmp instructions which can compare
4304 /// a register against the immediate without having to materialize the
4305 /// immediate into a register.
4306 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4307 if (!Subtarget->isThumb())
4308 return ARM_AM::getSOImmVal(Imm) != -1;
4309 if (Subtarget->isThumb2())
4310 return ARM_AM::getT2SOImmVal(Imm) != -1;
4311 return Imm >= 0 && Imm <= 255;
4314 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4315 bool isSEXTLoad, SDValue &Base,
4316 SDValue &Offset, bool &isInc,
4317 SelectionDAG &DAG) {
4318 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4321 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4323 Base = Ptr->getOperand(0);
4324 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4325 int RHSC = (int)RHS->getZExtValue();
4326 if (RHSC < 0 && RHSC > -256) {
4327 assert(Ptr->getOpcode() == ISD::ADD);
4329 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4333 isInc = (Ptr->getOpcode() == ISD::ADD);
4334 Offset = Ptr->getOperand(1);
4336 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4338 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4339 int RHSC = (int)RHS->getZExtValue();
4340 if (RHSC < 0 && RHSC > -0x1000) {
4341 assert(Ptr->getOpcode() == ISD::ADD);
4343 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4344 Base = Ptr->getOperand(0);
4349 if (Ptr->getOpcode() == ISD::ADD) {
4351 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4352 if (ShOpcVal != ARM_AM::no_shift) {
4353 Base = Ptr->getOperand(1);
4354 Offset = Ptr->getOperand(0);
4356 Base = Ptr->getOperand(0);
4357 Offset = Ptr->getOperand(1);
4362 isInc = (Ptr->getOpcode() == ISD::ADD);
4363 Base = Ptr->getOperand(0);
4364 Offset = Ptr->getOperand(1);
4368 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4372 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4373 bool isSEXTLoad, SDValue &Base,
4374 SDValue &Offset, bool &isInc,
4375 SelectionDAG &DAG) {
4376 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4379 Base = Ptr->getOperand(0);
4380 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4381 int RHSC = (int)RHS->getZExtValue();
4382 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4383 assert(Ptr->getOpcode() == ISD::ADD);
4385 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4387 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4388 isInc = Ptr->getOpcode() == ISD::ADD;
4389 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4397 /// getPreIndexedAddressParts - returns true by value, base pointer and
4398 /// offset pointer and addressing mode by reference if the node's address
4399 /// can be legally represented as pre-indexed load / store address.
4401 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4403 ISD::MemIndexedMode &AM,
4404 SelectionDAG &DAG) const {
4405 if (Subtarget->isThumb1Only())
4410 bool isSEXTLoad = false;
4411 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4412 Ptr = LD->getBasePtr();
4413 VT = LD->getMemoryVT();
4414 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4415 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4416 Ptr = ST->getBasePtr();
4417 VT = ST->getMemoryVT();
4422 bool isLegal = false;
4423 if (Subtarget->isThumb2())
4424 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4425 Offset, isInc, DAG);
4427 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4428 Offset, isInc, DAG);
4432 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4436 /// getPostIndexedAddressParts - returns true by value, base pointer and
4437 /// offset pointer and addressing mode by reference if this node can be
4438 /// combined with a load / store to form a post-indexed load / store.
4439 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4442 ISD::MemIndexedMode &AM,
4443 SelectionDAG &DAG) const {
4444 if (Subtarget->isThumb1Only())
4449 bool isSEXTLoad = false;
4450 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4451 VT = LD->getMemoryVT();
4452 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4453 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4454 VT = ST->getMemoryVT();
4459 bool isLegal = false;
4460 if (Subtarget->isThumb2())
4461 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4464 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4469 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4473 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4477 const SelectionDAG &DAG,
4478 unsigned Depth) const {
4479 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4480 switch (Op.getOpcode()) {
4482 case ARMISD::CMOV: {
4483 // Bits are known zero/one if known on the LHS and RHS.
4484 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4485 if (KnownZero == 0 && KnownOne == 0) return;
4487 APInt KnownZeroRHS, KnownOneRHS;
4488 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4489 KnownZeroRHS, KnownOneRHS, Depth+1);
4490 KnownZero &= KnownZeroRHS;
4491 KnownOne &= KnownOneRHS;
4497 //===----------------------------------------------------------------------===//
4498 // ARM Inline Assembly Support
4499 //===----------------------------------------------------------------------===//
4501 /// getConstraintType - Given a constraint letter, return the type of
4502 /// constraint it is for this target.
4503 ARMTargetLowering::ConstraintType
4504 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4505 if (Constraint.size() == 1) {
4506 switch (Constraint[0]) {
4508 case 'l': return C_RegisterClass;
4509 case 'w': return C_RegisterClass;
4512 return TargetLowering::getConstraintType(Constraint);
4515 std::pair<unsigned, const TargetRegisterClass*>
4516 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4518 if (Constraint.size() == 1) {
4519 // GCC ARM Constraint Letters
4520 switch (Constraint[0]) {
4522 if (Subtarget->isThumb())
4523 return std::make_pair(0U, ARM::tGPRRegisterClass);
4525 return std::make_pair(0U, ARM::GPRRegisterClass);
4527 return std::make_pair(0U, ARM::GPRRegisterClass);
4530 return std::make_pair(0U, ARM::SPRRegisterClass);
4531 if (VT.getSizeInBits() == 64)
4532 return std::make_pair(0U, ARM::DPRRegisterClass);
4533 if (VT.getSizeInBits() == 128)
4534 return std::make_pair(0U, ARM::QPRRegisterClass);
4538 if (StringRef("{cc}").equals_lower(Constraint))
4539 return std::make_pair(0U, ARM::CCRRegisterClass);
4541 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4544 std::vector<unsigned> ARMTargetLowering::
4545 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4547 if (Constraint.size() != 1)
4548 return std::vector<unsigned>();
4550 switch (Constraint[0]) { // GCC ARM Constraint Letters
4553 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4554 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4557 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4558 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4559 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4560 ARM::R12, ARM::LR, 0);
4563 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4564 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4565 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4566 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4567 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4568 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4569 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4570 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4571 if (VT.getSizeInBits() == 64)
4572 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4573 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4574 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4575 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4576 if (VT.getSizeInBits() == 128)
4577 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4578 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4582 return std::vector<unsigned>();
4585 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4586 /// vector. If it is invalid, don't add anything to Ops.
4587 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4590 std::vector<SDValue>&Ops,
4591 SelectionDAG &DAG) const {
4592 SDValue Result(0, 0);
4594 switch (Constraint) {
4596 case 'I': case 'J': case 'K': case 'L':
4597 case 'M': case 'N': case 'O':
4598 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4602 int64_t CVal64 = C->getSExtValue();
4603 int CVal = (int) CVal64;
4604 // None of these constraints allow values larger than 32 bits. Check
4605 // that the value fits in an int.
4609 switch (Constraint) {
4611 if (Subtarget->isThumb1Only()) {
4612 // This must be a constant between 0 and 255, for ADD
4614 if (CVal >= 0 && CVal <= 255)
4616 } else if (Subtarget->isThumb2()) {
4617 // A constant that can be used as an immediate value in a
4618 // data-processing instruction.
4619 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4622 // A constant that can be used as an immediate value in a
4623 // data-processing instruction.
4624 if (ARM_AM::getSOImmVal(CVal) != -1)
4630 if (Subtarget->isThumb()) { // FIXME thumb2
4631 // This must be a constant between -255 and -1, for negated ADD
4632 // immediates. This can be used in GCC with an "n" modifier that
4633 // prints the negated value, for use with SUB instructions. It is
4634 // not useful otherwise but is implemented for compatibility.
4635 if (CVal >= -255 && CVal <= -1)
4638 // This must be a constant between -4095 and 4095. It is not clear
4639 // what this constraint is intended for. Implemented for
4640 // compatibility with GCC.
4641 if (CVal >= -4095 && CVal <= 4095)
4647 if (Subtarget->isThumb1Only()) {
4648 // A 32-bit value where only one byte has a nonzero value. Exclude
4649 // zero to match GCC. This constraint is used by GCC internally for
4650 // constants that can be loaded with a move/shift combination.
4651 // It is not useful otherwise but is implemented for compatibility.
4652 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4654 } else if (Subtarget->isThumb2()) {
4655 // A constant whose bitwise inverse can be used as an immediate
4656 // value in a data-processing instruction. This can be used in GCC
4657 // with a "B" modifier that prints the inverted value, for use with
4658 // BIC and MVN instructions. It is not useful otherwise but is
4659 // implemented for compatibility.
4660 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4663 // A constant whose bitwise inverse can be used as an immediate
4664 // value in a data-processing instruction. This can be used in GCC
4665 // with a "B" modifier that prints the inverted value, for use with
4666 // BIC and MVN instructions. It is not useful otherwise but is
4667 // implemented for compatibility.
4668 if (ARM_AM::getSOImmVal(~CVal) != -1)
4674 if (Subtarget->isThumb1Only()) {
4675 // This must be a constant between -7 and 7,
4676 // for 3-operand ADD/SUB immediate instructions.
4677 if (CVal >= -7 && CVal < 7)
4679 } else if (Subtarget->isThumb2()) {
4680 // A constant whose negation can be used as an immediate value in a
4681 // data-processing instruction. This can be used in GCC with an "n"
4682 // modifier that prints the negated value, for use with SUB
4683 // instructions. It is not useful otherwise but is implemented for
4685 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4688 // A constant whose negation can be used as an immediate value in a
4689 // data-processing instruction. This can be used in GCC with an "n"
4690 // modifier that prints the negated value, for use with SUB
4691 // instructions. It is not useful otherwise but is implemented for
4693 if (ARM_AM::getSOImmVal(-CVal) != -1)
4699 if (Subtarget->isThumb()) { // FIXME thumb2
4700 // This must be a multiple of 4 between 0 and 1020, for
4701 // ADD sp + immediate.
4702 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4705 // A power of two or a constant between 0 and 32. This is used in
4706 // GCC for the shift amount on shifted register operands, but it is
4707 // useful in general for any shift amounts.
4708 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4714 if (Subtarget->isThumb()) { // FIXME thumb2
4715 // This must be a constant between 0 and 31, for shift amounts.
4716 if (CVal >= 0 && CVal <= 31)
4722 if (Subtarget->isThumb()) { // FIXME thumb2
4723 // This must be a multiple of 4 between -508 and 508, for
4724 // ADD/SUB sp = sp + immediate.
4725 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4730 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4734 if (Result.getNode()) {
4735 Ops.push_back(Result);
4738 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4743 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4744 // The ARM target isn't yet aware of offsets.
4748 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4749 APInt Imm = FPImm.bitcastToAPInt();
4750 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4751 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4752 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4754 // We can handle 4 bits of mantissa.
4755 // mantissa = (16+UInt(e:f:g:h))/16.
4756 if (Mantissa & 0x7ffff)
4759 if ((Mantissa & 0xf) != Mantissa)
4762 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4763 if (Exp < -3 || Exp > 4)
4765 Exp = ((Exp+3) & 0x7) ^ 4;
4767 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4770 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4771 APInt Imm = FPImm.bitcastToAPInt();
4772 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4773 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4774 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4776 // We can handle 4 bits of mantissa.
4777 // mantissa = (16+UInt(e:f:g:h))/16.
4778 if (Mantissa & 0xffffffffffffLL)
4781 if ((Mantissa & 0xf) != Mantissa)
4784 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4785 if (Exp < -3 || Exp > 4)
4787 Exp = ((Exp+3) & 0x7) ^ 4;
4789 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4792 /// isFPImmLegal - Returns true if the target can instruction select the
4793 /// specified FP immediate natively. If false, the legalizer will
4794 /// materialize the FP immediate as a load from a constant pool.
4795 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4796 if (!Subtarget->hasVFP3())
4799 return ARM::getVFPf32Imm(Imm) != -1;
4801 return ARM::getVFPf64Imm(Imm) != -1;