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 // Use the default implementation.
388 setOperationAction(ISD::VASTART, MVT::Other, Custom);
389 setOperationAction(ISD::VAARG, MVT::Other, Expand);
390 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
391 setOperationAction(ISD::VAEND, MVT::Other, Expand);
392 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
393 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
394 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
395 // FIXME: Shouldn't need this, since no register is used, but the legalizer
396 // doesn't yet know how to not do that for SjLj.
397 setExceptionSelectorRegister(ARM::R0);
398 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
399 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
401 // If the subtarget does not have extract instructions, sign_extend_inreg
402 // needs to be expanded. Extract is available in ARM mode on v6 and up,
403 // and on most Thumb2 implementations.
404 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
405 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
406 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
407 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
409 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
411 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
412 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
413 // iff target supports vfp2.
414 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
416 // We want to custom lower some of our intrinsics.
417 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
419 setOperationAction(ISD::SETCC, MVT::i32, Expand);
420 setOperationAction(ISD::SETCC, MVT::f32, Expand);
421 setOperationAction(ISD::SETCC, MVT::f64, Expand);
422 setOperationAction(ISD::SELECT, MVT::i32, Expand);
423 setOperationAction(ISD::SELECT, MVT::f32, Expand);
424 setOperationAction(ISD::SELECT, MVT::f64, Expand);
425 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
426 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
427 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
429 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
430 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
431 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
432 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
433 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
435 // We don't support sin/cos/fmod/copysign/pow
436 setOperationAction(ISD::FSIN, MVT::f64, Expand);
437 setOperationAction(ISD::FSIN, MVT::f32, Expand);
438 setOperationAction(ISD::FCOS, MVT::f32, Expand);
439 setOperationAction(ISD::FCOS, MVT::f64, Expand);
440 setOperationAction(ISD::FREM, MVT::f64, Expand);
441 setOperationAction(ISD::FREM, MVT::f32, Expand);
442 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
443 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
444 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
446 setOperationAction(ISD::FPOW, MVT::f64, Expand);
447 setOperationAction(ISD::FPOW, MVT::f32, Expand);
449 // Various VFP goodness
450 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
451 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
452 if (Subtarget->hasVFP2()) {
453 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
454 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
455 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
456 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
458 // Special handling for half-precision FP.
459 if (!Subtarget->hasFP16()) {
460 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
461 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
465 // We have target-specific dag combine patterns for the following nodes:
466 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
467 setTargetDAGCombine(ISD::ADD);
468 setTargetDAGCombine(ISD::SUB);
470 setStackPointerRegisterToSaveRestore(ARM::SP);
471 setSchedulingPreference(SchedulingForRegPressure);
473 // FIXME: If-converter should use instruction latency to determine
474 // profitability rather than relying on fixed limits.
475 if (Subtarget->getCPUString() == "generic") {
476 // Generic (and overly aggressive) if-conversion limits.
477 setIfCvtBlockSizeLimit(10);
478 setIfCvtDupBlockSizeLimit(2);
479 } else if (Subtarget->hasV7Ops()) {
480 setIfCvtBlockSizeLimit(3);
481 setIfCvtDupBlockSizeLimit(1);
482 } else if (Subtarget->hasV6Ops()) {
483 setIfCvtBlockSizeLimit(2);
484 setIfCvtDupBlockSizeLimit(1);
486 setIfCvtBlockSizeLimit(3);
487 setIfCvtDupBlockSizeLimit(2);
490 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
491 // Do not enable CodePlacementOpt for now: it currently runs after the
492 // ARMConstantIslandPass and messes up branch relaxation and placement
493 // of constant islands.
494 // benefitFromCodePlacementOpt = true;
497 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
500 case ARMISD::Wrapper: return "ARMISD::Wrapper";
501 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
502 case ARMISD::CALL: return "ARMISD::CALL";
503 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
504 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
505 case ARMISD::tCALL: return "ARMISD::tCALL";
506 case ARMISD::BRCOND: return "ARMISD::BRCOND";
507 case ARMISD::BR_JT: return "ARMISD::BR_JT";
508 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
509 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
510 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
511 case ARMISD::CMP: return "ARMISD::CMP";
512 case ARMISD::CMPZ: return "ARMISD::CMPZ";
513 case ARMISD::CMPFP: return "ARMISD::CMPFP";
514 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
515 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
516 case ARMISD::CMOV: return "ARMISD::CMOV";
517 case ARMISD::CNEG: return "ARMISD::CNEG";
519 case ARMISD::RBIT: return "ARMISD::RBIT";
521 case ARMISD::FTOSI: return "ARMISD::FTOSI";
522 case ARMISD::FTOUI: return "ARMISD::FTOUI";
523 case ARMISD::SITOF: return "ARMISD::SITOF";
524 case ARMISD::UITOF: return "ARMISD::UITOF";
526 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
527 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
528 case ARMISD::RRX: return "ARMISD::RRX";
530 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
531 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
533 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
534 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
536 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
538 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
540 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
541 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
543 case ARMISD::VCEQ: return "ARMISD::VCEQ";
544 case ARMISD::VCGE: return "ARMISD::VCGE";
545 case ARMISD::VCGEU: return "ARMISD::VCGEU";
546 case ARMISD::VCGT: return "ARMISD::VCGT";
547 case ARMISD::VCGTU: return "ARMISD::VCGTU";
548 case ARMISD::VTST: return "ARMISD::VTST";
550 case ARMISD::VSHL: return "ARMISD::VSHL";
551 case ARMISD::VSHRs: return "ARMISD::VSHRs";
552 case ARMISD::VSHRu: return "ARMISD::VSHRu";
553 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
554 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
555 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
556 case ARMISD::VSHRN: return "ARMISD::VSHRN";
557 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
558 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
559 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
560 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
561 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
562 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
563 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
564 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
565 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
566 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
567 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
568 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
569 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
570 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
571 case ARMISD::VDUP: return "ARMISD::VDUP";
572 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
573 case ARMISD::VEXT: return "ARMISD::VEXT";
574 case ARMISD::VREV64: return "ARMISD::VREV64";
575 case ARMISD::VREV32: return "ARMISD::VREV32";
576 case ARMISD::VREV16: return "ARMISD::VREV16";
577 case ARMISD::VZIP: return "ARMISD::VZIP";
578 case ARMISD::VUZP: return "ARMISD::VUZP";
579 case ARMISD::VTRN: return "ARMISD::VTRN";
580 case ARMISD::FMAX: return "ARMISD::FMAX";
581 case ARMISD::FMIN: return "ARMISD::FMIN";
585 /// getFunctionAlignment - Return the Log2 alignment of this function.
586 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
587 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
590 //===----------------------------------------------------------------------===//
592 //===----------------------------------------------------------------------===//
594 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
595 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
597 default: llvm_unreachable("Unknown condition code!");
598 case ISD::SETNE: return ARMCC::NE;
599 case ISD::SETEQ: return ARMCC::EQ;
600 case ISD::SETGT: return ARMCC::GT;
601 case ISD::SETGE: return ARMCC::GE;
602 case ISD::SETLT: return ARMCC::LT;
603 case ISD::SETLE: return ARMCC::LE;
604 case ISD::SETUGT: return ARMCC::HI;
605 case ISD::SETUGE: return ARMCC::HS;
606 case ISD::SETULT: return ARMCC::LO;
607 case ISD::SETULE: return ARMCC::LS;
611 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
612 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
613 ARMCC::CondCodes &CondCode2) {
614 CondCode2 = ARMCC::AL;
616 default: llvm_unreachable("Unknown FP condition!");
618 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
620 case ISD::SETOGT: CondCode = ARMCC::GT; break;
622 case ISD::SETOGE: CondCode = ARMCC::GE; break;
623 case ISD::SETOLT: CondCode = ARMCC::MI; break;
624 case ISD::SETOLE: CondCode = ARMCC::LS; break;
625 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
626 case ISD::SETO: CondCode = ARMCC::VC; break;
627 case ISD::SETUO: CondCode = ARMCC::VS; break;
628 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
629 case ISD::SETUGT: CondCode = ARMCC::HI; break;
630 case ISD::SETUGE: CondCode = ARMCC::PL; break;
632 case ISD::SETULT: CondCode = ARMCC::LT; break;
634 case ISD::SETULE: CondCode = ARMCC::LE; break;
636 case ISD::SETUNE: CondCode = ARMCC::NE; break;
640 //===----------------------------------------------------------------------===//
641 // Calling Convention Implementation
642 //===----------------------------------------------------------------------===//
644 #include "ARMGenCallingConv.inc"
646 // APCS f64 is in register pairs, possibly split to stack
647 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
648 CCValAssign::LocInfo &LocInfo,
649 CCState &State, bool CanFail) {
650 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
652 // Try to get the first register.
653 if (unsigned Reg = State.AllocateReg(RegList, 4))
654 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
656 // For the 2nd half of a v2f64, do not fail.
660 // Put the whole thing on the stack.
661 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
662 State.AllocateStack(8, 4),
667 // Try to get the second register.
668 if (unsigned Reg = State.AllocateReg(RegList, 4))
669 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
671 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
672 State.AllocateStack(4, 4),
677 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
678 CCValAssign::LocInfo &LocInfo,
679 ISD::ArgFlagsTy &ArgFlags,
681 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
683 if (LocVT == MVT::v2f64 &&
684 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
686 return true; // we handled it
689 // AAPCS f64 is in aligned register pairs
690 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
691 CCValAssign::LocInfo &LocInfo,
692 CCState &State, bool CanFail) {
693 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
694 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
696 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
698 // For the 2nd half of a v2f64, do not just fail.
702 // Put the whole thing on the stack.
703 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
704 State.AllocateStack(8, 8),
710 for (i = 0; i < 2; ++i)
711 if (HiRegList[i] == Reg)
714 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
715 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
720 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
721 CCValAssign::LocInfo &LocInfo,
722 ISD::ArgFlagsTy &ArgFlags,
724 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
726 if (LocVT == MVT::v2f64 &&
727 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
729 return true; // we handled it
732 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
733 CCValAssign::LocInfo &LocInfo, CCState &State) {
734 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
735 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
737 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
739 return false; // we didn't handle it
742 for (i = 0; i < 2; ++i)
743 if (HiRegList[i] == Reg)
746 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
747 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
752 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
753 CCValAssign::LocInfo &LocInfo,
754 ISD::ArgFlagsTy &ArgFlags,
756 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
758 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
760 return true; // we handled it
763 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
764 CCValAssign::LocInfo &LocInfo,
765 ISD::ArgFlagsTy &ArgFlags,
767 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
771 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
772 /// given CallingConvention value.
773 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
775 bool isVarArg) const {
778 llvm_unreachable("Unsupported calling convention");
780 case CallingConv::Fast:
781 // Use target triple & subtarget features to do actual dispatch.
782 if (Subtarget->isAAPCS_ABI()) {
783 if (Subtarget->hasVFP2() &&
784 FloatABIType == FloatABI::Hard && !isVarArg)
785 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
787 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
789 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
790 case CallingConv::ARM_AAPCS_VFP:
791 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
792 case CallingConv::ARM_AAPCS:
793 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
794 case CallingConv::ARM_APCS:
795 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
799 /// LowerCallResult - Lower the result values of a call into the
800 /// appropriate copies out of appropriate physical registers.
802 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
803 CallingConv::ID CallConv, bool isVarArg,
804 const SmallVectorImpl<ISD::InputArg> &Ins,
805 DebugLoc dl, SelectionDAG &DAG,
806 SmallVectorImpl<SDValue> &InVals) const {
808 // Assign locations to each value returned by this call.
809 SmallVector<CCValAssign, 16> RVLocs;
810 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
811 RVLocs, *DAG.getContext());
812 CCInfo.AnalyzeCallResult(Ins,
813 CCAssignFnForNode(CallConv, /* Return*/ true,
816 // Copy all of the result registers out of their specified physreg.
817 for (unsigned i = 0; i != RVLocs.size(); ++i) {
818 CCValAssign VA = RVLocs[i];
821 if (VA.needsCustom()) {
822 // Handle f64 or half of a v2f64.
823 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
825 Chain = Lo.getValue(1);
826 InFlag = Lo.getValue(2);
827 VA = RVLocs[++i]; // skip ahead to next loc
828 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
830 Chain = Hi.getValue(1);
831 InFlag = Hi.getValue(2);
832 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
834 if (VA.getLocVT() == MVT::v2f64) {
835 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
836 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
837 DAG.getConstant(0, MVT::i32));
839 VA = RVLocs[++i]; // skip ahead to next loc
840 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
841 Chain = Lo.getValue(1);
842 InFlag = Lo.getValue(2);
843 VA = RVLocs[++i]; // skip ahead to next loc
844 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
845 Chain = Hi.getValue(1);
846 InFlag = Hi.getValue(2);
847 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
848 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
849 DAG.getConstant(1, MVT::i32));
852 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
854 Chain = Val.getValue(1);
855 InFlag = Val.getValue(2);
858 switch (VA.getLocInfo()) {
859 default: llvm_unreachable("Unknown loc info!");
860 case CCValAssign::Full: break;
861 case CCValAssign::BCvt:
862 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
866 InVals.push_back(Val);
872 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
873 /// by "Src" to address "Dst" of size "Size". Alignment information is
874 /// specified by the specific parameter attribute. The copy will be passed as
875 /// a byval function parameter.
876 /// Sometimes what we are copying is the end of a larger object, the part that
877 /// does not fit in registers.
879 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
880 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
882 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
883 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
884 /*isVolatile=*/false, /*AlwaysInline=*/false,
888 /// LowerMemOpCallTo - Store the argument to the stack.
890 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
891 SDValue StackPtr, SDValue Arg,
892 DebugLoc dl, SelectionDAG &DAG,
893 const CCValAssign &VA,
894 ISD::ArgFlagsTy Flags) const {
895 unsigned LocMemOffset = VA.getLocMemOffset();
896 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
897 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
898 if (Flags.isByVal()) {
899 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
901 return DAG.getStore(Chain, dl, Arg, PtrOff,
902 PseudoSourceValue::getStack(), LocMemOffset,
906 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
907 SDValue Chain, SDValue &Arg,
908 RegsToPassVector &RegsToPass,
909 CCValAssign &VA, CCValAssign &NextVA,
911 SmallVector<SDValue, 8> &MemOpChains,
912 ISD::ArgFlagsTy Flags) const {
914 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
915 DAG.getVTList(MVT::i32, MVT::i32), Arg);
916 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
918 if (NextVA.isRegLoc())
919 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
921 assert(NextVA.isMemLoc());
922 if (StackPtr.getNode() == 0)
923 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
925 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
931 /// LowerCall - Lowering a call into a callseq_start <-
932 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
935 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
936 CallingConv::ID CallConv, bool isVarArg,
938 const SmallVectorImpl<ISD::OutputArg> &Outs,
939 const SmallVectorImpl<ISD::InputArg> &Ins,
940 DebugLoc dl, SelectionDAG &DAG,
941 SmallVectorImpl<SDValue> &InVals) const {
942 // ARM target does not yet support tail call optimization.
945 // Analyze operands of the call, assigning locations to each operand.
946 SmallVector<CCValAssign, 16> ArgLocs;
947 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
949 CCInfo.AnalyzeCallOperands(Outs,
950 CCAssignFnForNode(CallConv, /* Return*/ false,
953 // Get a count of how many bytes are to be pushed on the stack.
954 unsigned NumBytes = CCInfo.getNextStackOffset();
956 // Adjust the stack pointer for the new arguments...
957 // These operations are automatically eliminated by the prolog/epilog pass
958 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
960 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
962 RegsToPassVector RegsToPass;
963 SmallVector<SDValue, 8> MemOpChains;
965 // Walk the register/memloc assignments, inserting copies/loads. In the case
966 // of tail call optimization, arguments are handled later.
967 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
970 CCValAssign &VA = ArgLocs[i];
971 SDValue Arg = Outs[realArgIdx].Val;
972 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
974 // Promote the value if needed.
975 switch (VA.getLocInfo()) {
976 default: llvm_unreachable("Unknown loc info!");
977 case CCValAssign::Full: break;
978 case CCValAssign::SExt:
979 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
981 case CCValAssign::ZExt:
982 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
984 case CCValAssign::AExt:
985 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
987 case CCValAssign::BCvt:
988 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
992 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
993 if (VA.needsCustom()) {
994 if (VA.getLocVT() == MVT::v2f64) {
995 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
996 DAG.getConstant(0, MVT::i32));
997 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
998 DAG.getConstant(1, MVT::i32));
1000 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1001 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1003 VA = ArgLocs[++i]; // skip ahead to next loc
1004 if (VA.isRegLoc()) {
1005 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1006 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1008 assert(VA.isMemLoc());
1010 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1011 dl, DAG, VA, Flags));
1014 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1015 StackPtr, MemOpChains, Flags);
1017 } else if (VA.isRegLoc()) {
1018 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1020 assert(VA.isMemLoc());
1022 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1023 dl, DAG, VA, Flags));
1027 if (!MemOpChains.empty())
1028 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1029 &MemOpChains[0], MemOpChains.size());
1031 // Build a sequence of copy-to-reg nodes chained together with token chain
1032 // and flag operands which copy the outgoing args into the appropriate regs.
1034 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1035 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1036 RegsToPass[i].second, InFlag);
1037 InFlag = Chain.getValue(1);
1040 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1041 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1042 // node so that legalize doesn't hack it.
1043 bool isDirect = false;
1044 bool isARMFunc = false;
1045 bool isLocalARMFunc = false;
1046 MachineFunction &MF = DAG.getMachineFunction();
1047 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1049 if (EnableARMLongCalls) {
1050 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1051 && "long-calls with non-static relocation model!");
1052 // Handle a global address or an external symbol. If it's not one of
1053 // those, the target's already in a register, so we don't need to do
1055 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1056 const GlobalValue *GV = G->getGlobal();
1057 // Create a constant pool entry for the callee address
1058 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1059 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1062 // Get the address of the callee into a register
1063 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1064 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1065 Callee = DAG.getLoad(getPointerTy(), dl,
1066 DAG.getEntryNode(), CPAddr,
1067 PseudoSourceValue::getConstantPool(), 0,
1069 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1070 const char *Sym = S->getSymbol();
1072 // Create a constant pool entry for the callee address
1073 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1074 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1075 Sym, ARMPCLabelIndex, 0);
1076 // Get the address of the callee into a register
1077 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1078 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1079 Callee = DAG.getLoad(getPointerTy(), dl,
1080 DAG.getEntryNode(), CPAddr,
1081 PseudoSourceValue::getConstantPool(), 0,
1084 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1085 const GlobalValue *GV = G->getGlobal();
1087 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1088 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1089 getTargetMachine().getRelocationModel() != Reloc::Static;
1090 isARMFunc = !Subtarget->isThumb() || isStub;
1091 // ARM call to a local ARM function is predicable.
1092 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1093 // tBX takes a register source operand.
1094 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1095 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1096 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1099 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1100 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1101 Callee = DAG.getLoad(getPointerTy(), dl,
1102 DAG.getEntryNode(), CPAddr,
1103 PseudoSourceValue::getConstantPool(), 0,
1105 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1106 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1107 getPointerTy(), Callee, PICLabel);
1109 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1110 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1112 bool isStub = Subtarget->isTargetDarwin() &&
1113 getTargetMachine().getRelocationModel() != Reloc::Static;
1114 isARMFunc = !Subtarget->isThumb() || isStub;
1115 // tBX takes a register source operand.
1116 const char *Sym = S->getSymbol();
1117 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1118 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1119 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1120 Sym, ARMPCLabelIndex, 4);
1121 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1122 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1123 Callee = DAG.getLoad(getPointerTy(), dl,
1124 DAG.getEntryNode(), CPAddr,
1125 PseudoSourceValue::getConstantPool(), 0,
1127 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1128 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1129 getPointerTy(), Callee, PICLabel);
1131 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1134 // FIXME: handle tail calls differently.
1136 if (Subtarget->isThumb()) {
1137 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1138 CallOpc = ARMISD::CALL_NOLINK;
1140 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1142 CallOpc = (isDirect || Subtarget->hasV5TOps())
1143 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1144 : ARMISD::CALL_NOLINK;
1146 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1147 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1148 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1149 InFlag = Chain.getValue(1);
1152 std::vector<SDValue> Ops;
1153 Ops.push_back(Chain);
1154 Ops.push_back(Callee);
1156 // Add argument registers to the end of the list so that they are known live
1158 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1159 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1160 RegsToPass[i].second.getValueType()));
1162 if (InFlag.getNode())
1163 Ops.push_back(InFlag);
1164 // Returns a chain and a flag for retval copy to use.
1165 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1166 &Ops[0], Ops.size());
1167 InFlag = Chain.getValue(1);
1169 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1170 DAG.getIntPtrConstant(0, true), InFlag);
1172 InFlag = Chain.getValue(1);
1174 // Handle result values, copying them out of physregs into vregs that we
1176 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1181 ARMTargetLowering::LowerReturn(SDValue Chain,
1182 CallingConv::ID CallConv, bool isVarArg,
1183 const SmallVectorImpl<ISD::OutputArg> &Outs,
1184 DebugLoc dl, SelectionDAG &DAG) const {
1186 // CCValAssign - represent the assignment of the return value to a location.
1187 SmallVector<CCValAssign, 16> RVLocs;
1189 // CCState - Info about the registers and stack slots.
1190 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1193 // Analyze outgoing return values.
1194 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1197 // If this is the first return lowered for this function, add
1198 // the regs to the liveout set for the function.
1199 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1200 for (unsigned i = 0; i != RVLocs.size(); ++i)
1201 if (RVLocs[i].isRegLoc())
1202 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1207 // Copy the result values into the output registers.
1208 for (unsigned i = 0, realRVLocIdx = 0;
1210 ++i, ++realRVLocIdx) {
1211 CCValAssign &VA = RVLocs[i];
1212 assert(VA.isRegLoc() && "Can only return in registers!");
1214 SDValue Arg = Outs[realRVLocIdx].Val;
1216 switch (VA.getLocInfo()) {
1217 default: llvm_unreachable("Unknown loc info!");
1218 case CCValAssign::Full: break;
1219 case CCValAssign::BCvt:
1220 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1224 if (VA.needsCustom()) {
1225 if (VA.getLocVT() == MVT::v2f64) {
1226 // Extract the first half and return it in two registers.
1227 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1228 DAG.getConstant(0, MVT::i32));
1229 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1230 DAG.getVTList(MVT::i32, MVT::i32), Half);
1232 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1233 Flag = Chain.getValue(1);
1234 VA = RVLocs[++i]; // skip ahead to next loc
1235 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1236 HalfGPRs.getValue(1), Flag);
1237 Flag = Chain.getValue(1);
1238 VA = RVLocs[++i]; // skip ahead to next loc
1240 // Extract the 2nd half and fall through to handle it as an f64 value.
1241 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1242 DAG.getConstant(1, MVT::i32));
1244 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1246 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1247 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1248 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1249 Flag = Chain.getValue(1);
1250 VA = RVLocs[++i]; // skip ahead to next loc
1251 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1254 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1256 // Guarantee that all emitted copies are
1257 // stuck together, avoiding something bad.
1258 Flag = Chain.getValue(1);
1263 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1265 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1270 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1271 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1272 // one of the above mentioned nodes. It has to be wrapped because otherwise
1273 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1274 // be used to form addressing mode. These wrapped nodes will be selected
1276 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1277 EVT PtrVT = Op.getValueType();
1278 // FIXME there is no actual debug info here
1279 DebugLoc dl = Op.getDebugLoc();
1280 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1282 if (CP->isMachineConstantPoolEntry())
1283 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1284 CP->getAlignment());
1286 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1287 CP->getAlignment());
1288 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1291 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1292 SelectionDAG &DAG) const {
1293 MachineFunction &MF = DAG.getMachineFunction();
1294 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1295 unsigned ARMPCLabelIndex = 0;
1296 DebugLoc DL = Op.getDebugLoc();
1297 EVT PtrVT = getPointerTy();
1298 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1299 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1301 if (RelocM == Reloc::Static) {
1302 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1304 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1305 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1306 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1307 ARMCP::CPBlockAddress,
1309 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1311 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1312 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1313 PseudoSourceValue::getConstantPool(), 0,
1315 if (RelocM == Reloc::Static)
1317 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1318 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1321 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1323 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1324 SelectionDAG &DAG) const {
1325 DebugLoc dl = GA->getDebugLoc();
1326 EVT PtrVT = getPointerTy();
1327 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1328 MachineFunction &MF = DAG.getMachineFunction();
1329 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1330 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1331 ARMConstantPoolValue *CPV =
1332 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1333 ARMCP::CPValue, PCAdj, "tlsgd", true);
1334 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1335 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1336 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1337 PseudoSourceValue::getConstantPool(), 0,
1339 SDValue Chain = Argument.getValue(1);
1341 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1342 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1344 // call __tls_get_addr.
1347 Entry.Node = Argument;
1348 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1349 Args.push_back(Entry);
1350 // FIXME: is there useful debug info available here?
1351 std::pair<SDValue, SDValue> CallResult =
1352 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1353 false, false, false, false,
1354 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1355 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1356 return CallResult.first;
1359 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1360 // "local exec" model.
1362 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1363 SelectionDAG &DAG) const {
1364 const GlobalValue *GV = GA->getGlobal();
1365 DebugLoc dl = GA->getDebugLoc();
1367 SDValue Chain = DAG.getEntryNode();
1368 EVT PtrVT = getPointerTy();
1369 // Get the Thread Pointer
1370 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1372 if (GV->isDeclaration()) {
1373 MachineFunction &MF = DAG.getMachineFunction();
1374 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1375 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1376 // Initial exec model.
1377 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1378 ARMConstantPoolValue *CPV =
1379 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1380 ARMCP::CPValue, PCAdj, "gottpoff", true);
1381 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1382 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1383 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1384 PseudoSourceValue::getConstantPool(), 0,
1386 Chain = Offset.getValue(1);
1388 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1389 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1391 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1392 PseudoSourceValue::getConstantPool(), 0,
1396 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1397 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1398 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1399 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1400 PseudoSourceValue::getConstantPool(), 0,
1404 // The address of the thread local variable is the add of the thread
1405 // pointer with the offset of the variable.
1406 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1410 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1411 // TODO: implement the "local dynamic" model
1412 assert(Subtarget->isTargetELF() &&
1413 "TLS not implemented for non-ELF targets");
1414 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1415 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1416 // otherwise use the "Local Exec" TLS Model
1417 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1418 return LowerToTLSGeneralDynamicModel(GA, DAG);
1420 return LowerToTLSExecModels(GA, DAG);
1423 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1424 SelectionDAG &DAG) const {
1425 EVT PtrVT = getPointerTy();
1426 DebugLoc dl = Op.getDebugLoc();
1427 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1428 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1429 if (RelocM == Reloc::PIC_) {
1430 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1431 ARMConstantPoolValue *CPV =
1432 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1433 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1434 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1435 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1437 PseudoSourceValue::getConstantPool(), 0,
1439 SDValue Chain = Result.getValue(1);
1440 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1441 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1443 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1444 PseudoSourceValue::getGOT(), 0,
1448 // If we have T2 ops, we can materialize the address directly via movt/movw
1449 // pair. This is always cheaper.
1450 if (Subtarget->useMovt()) {
1451 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1452 DAG.getTargetGlobalAddress(GV, PtrVT));
1454 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1455 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1456 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1457 PseudoSourceValue::getConstantPool(), 0,
1463 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1464 SelectionDAG &DAG) const {
1465 MachineFunction &MF = DAG.getMachineFunction();
1466 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1467 unsigned ARMPCLabelIndex = 0;
1468 EVT PtrVT = getPointerTy();
1469 DebugLoc dl = Op.getDebugLoc();
1470 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1471 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1473 if (RelocM == Reloc::Static)
1474 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1476 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1477 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1478 ARMConstantPoolValue *CPV =
1479 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1480 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1482 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1484 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1485 PseudoSourceValue::getConstantPool(), 0,
1487 SDValue Chain = Result.getValue(1);
1489 if (RelocM == Reloc::PIC_) {
1490 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1491 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1494 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1495 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1496 PseudoSourceValue::getGOT(), 0,
1502 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1503 SelectionDAG &DAG) const {
1504 assert(Subtarget->isTargetELF() &&
1505 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1506 MachineFunction &MF = DAG.getMachineFunction();
1507 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1508 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1509 EVT PtrVT = getPointerTy();
1510 DebugLoc dl = Op.getDebugLoc();
1511 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1512 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1513 "_GLOBAL_OFFSET_TABLE_",
1514 ARMPCLabelIndex, PCAdj);
1515 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1516 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1517 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1518 PseudoSourceValue::getConstantPool(), 0,
1520 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1521 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1525 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1526 const ARMSubtarget *Subtarget)
1528 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1529 DebugLoc dl = Op.getDebugLoc();
1531 default: return SDValue(); // Don't custom lower most intrinsics.
1532 case Intrinsic::arm_thread_pointer: {
1533 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1534 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1536 case Intrinsic::eh_sjlj_lsda: {
1537 MachineFunction &MF = DAG.getMachineFunction();
1538 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1539 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1540 EVT PtrVT = getPointerTy();
1541 DebugLoc dl = Op.getDebugLoc();
1542 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1544 unsigned PCAdj = (RelocM != Reloc::PIC_)
1545 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1546 ARMConstantPoolValue *CPV =
1547 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1548 ARMCP::CPLSDA, PCAdj);
1549 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1550 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1552 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1553 PseudoSourceValue::getConstantPool(), 0,
1555 SDValue Chain = Result.getValue(1);
1557 if (RelocM == Reloc::PIC_) {
1558 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1559 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1563 case Intrinsic::eh_sjlj_setjmp:
1564 SDValue Val = Subtarget->isThumb() ?
1565 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1566 DAG.getConstant(0, MVT::i32);
1567 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1572 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1573 const ARMSubtarget *Subtarget) {
1574 DebugLoc dl = Op.getDebugLoc();
1575 SDValue Op5 = Op.getOperand(5);
1577 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1578 if (isDeviceBarrier) {
1579 if (Subtarget->hasV7Ops())
1580 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1582 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1583 DAG.getConstant(0, MVT::i32));
1585 if (Subtarget->hasV7Ops())
1586 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1588 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1589 DAG.getConstant(0, MVT::i32));
1594 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1595 MachineFunction &MF = DAG.getMachineFunction();
1596 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1598 // vastart just stores the address of the VarArgsFrameIndex slot into the
1599 // memory location argument.
1600 DebugLoc dl = Op.getDebugLoc();
1601 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1602 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1603 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1604 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1609 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1610 SelectionDAG &DAG) const {
1611 SDNode *Node = Op.getNode();
1612 DebugLoc dl = Node->getDebugLoc();
1613 EVT VT = Node->getValueType(0);
1614 SDValue Chain = Op.getOperand(0);
1615 SDValue Size = Op.getOperand(1);
1616 SDValue Align = Op.getOperand(2);
1618 // Chain the dynamic stack allocation so that it doesn't modify the stack
1619 // pointer when other instructions are using the stack.
1620 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1622 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1623 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1624 if (AlignVal > StackAlign)
1625 // Do this now since selection pass cannot introduce new target
1626 // independent node.
1627 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1629 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1630 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1631 // do even more horrible hack later.
1632 MachineFunction &MF = DAG.getMachineFunction();
1633 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1634 if (AFI->isThumb1OnlyFunction()) {
1636 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1638 uint32_t Val = C->getZExtValue();
1639 if (Val <= 508 && ((Val & 3) == 0))
1643 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1646 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1647 SDValue Ops1[] = { Chain, Size, Align };
1648 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1649 Chain = Res.getValue(1);
1650 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1651 DAG.getIntPtrConstant(0, true), SDValue());
1652 SDValue Ops2[] = { Res, Chain };
1653 return DAG.getMergeValues(Ops2, 2, dl);
1657 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1658 SDValue &Root, SelectionDAG &DAG,
1659 DebugLoc dl) const {
1660 MachineFunction &MF = DAG.getMachineFunction();
1661 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1663 TargetRegisterClass *RC;
1664 if (AFI->isThumb1OnlyFunction())
1665 RC = ARM::tGPRRegisterClass;
1667 RC = ARM::GPRRegisterClass;
1669 // Transform the arguments stored in physical registers into virtual ones.
1670 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1671 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1674 if (NextVA.isMemLoc()) {
1675 MachineFrameInfo *MFI = MF.getFrameInfo();
1676 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1678 // Create load node to retrieve arguments from the stack.
1679 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1680 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1681 PseudoSourceValue::getFixedStack(FI), 0,
1684 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1685 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1688 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1692 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1693 CallingConv::ID CallConv, bool isVarArg,
1694 const SmallVectorImpl<ISD::InputArg>
1696 DebugLoc dl, SelectionDAG &DAG,
1697 SmallVectorImpl<SDValue> &InVals)
1700 MachineFunction &MF = DAG.getMachineFunction();
1701 MachineFrameInfo *MFI = MF.getFrameInfo();
1703 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1705 // Assign locations to all of the incoming arguments.
1706 SmallVector<CCValAssign, 16> ArgLocs;
1707 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1709 CCInfo.AnalyzeFormalArguments(Ins,
1710 CCAssignFnForNode(CallConv, /* Return*/ false,
1713 SmallVector<SDValue, 16> ArgValues;
1715 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1716 CCValAssign &VA = ArgLocs[i];
1718 // Arguments stored in registers.
1719 if (VA.isRegLoc()) {
1720 EVT RegVT = VA.getLocVT();
1723 if (VA.needsCustom()) {
1724 // f64 and vector types are split up into multiple registers or
1725 // combinations of registers and stack slots.
1726 if (VA.getLocVT() == MVT::v2f64) {
1727 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1729 VA = ArgLocs[++i]; // skip ahead to next loc
1731 if (VA.isMemLoc()) {
1732 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
1734 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1735 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1736 PseudoSourceValue::getFixedStack(FI), 0,
1739 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1742 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1743 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1744 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1745 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1746 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1748 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1751 TargetRegisterClass *RC;
1753 if (RegVT == MVT::f32)
1754 RC = ARM::SPRRegisterClass;
1755 else if (RegVT == MVT::f64)
1756 RC = ARM::DPRRegisterClass;
1757 else if (RegVT == MVT::v2f64)
1758 RC = ARM::QPRRegisterClass;
1759 else if (RegVT == MVT::i32)
1760 RC = (AFI->isThumb1OnlyFunction() ?
1761 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1763 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1765 // Transform the arguments in physical registers into virtual ones.
1766 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1767 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1770 // If this is an 8 or 16-bit value, it is really passed promoted
1771 // to 32 bits. Insert an assert[sz]ext to capture this, then
1772 // truncate to the right size.
1773 switch (VA.getLocInfo()) {
1774 default: llvm_unreachable("Unknown loc info!");
1775 case CCValAssign::Full: break;
1776 case CCValAssign::BCvt:
1777 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1779 case CCValAssign::SExt:
1780 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1781 DAG.getValueType(VA.getValVT()));
1782 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1784 case CCValAssign::ZExt:
1785 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1786 DAG.getValueType(VA.getValVT()));
1787 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1791 InVals.push_back(ArgValue);
1793 } else { // VA.isRegLoc()
1796 assert(VA.isMemLoc());
1797 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1799 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1800 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1803 // Create load nodes to retrieve arguments from the stack.
1804 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1805 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1806 PseudoSourceValue::getFixedStack(FI), 0,
1813 static const unsigned GPRArgRegs[] = {
1814 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1817 unsigned NumGPRs = CCInfo.getFirstUnallocated
1818 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1820 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1821 unsigned VARegSize = (4 - NumGPRs) * 4;
1822 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1823 unsigned ArgOffset = CCInfo.getNextStackOffset();
1824 if (VARegSaveSize) {
1825 // If this function is vararg, store any remaining integer argument regs
1826 // to their spots on the stack so that they may be loaded by deferencing
1827 // the result of va_next.
1828 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1829 AFI->setVarArgsFrameIndex(
1830 MFI->CreateFixedObject(VARegSaveSize,
1831 ArgOffset + VARegSaveSize - VARegSize,
1833 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
1836 SmallVector<SDValue, 4> MemOps;
1837 for (; NumGPRs < 4; ++NumGPRs) {
1838 TargetRegisterClass *RC;
1839 if (AFI->isThumb1OnlyFunction())
1840 RC = ARM::tGPRRegisterClass;
1842 RC = ARM::GPRRegisterClass;
1844 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1845 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1847 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1848 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()), 0,
1850 MemOps.push_back(Store);
1851 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1852 DAG.getConstant(4, getPointerTy()));
1854 if (!MemOps.empty())
1855 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1856 &MemOps[0], MemOps.size());
1858 // This will point to the next argument passed via stack.
1859 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
1866 /// isFloatingPointZero - Return true if this is +0.0.
1867 static bool isFloatingPointZero(SDValue Op) {
1868 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1869 return CFP->getValueAPF().isPosZero();
1870 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1871 // Maybe this has already been legalized into the constant pool?
1872 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1873 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1874 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1875 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1876 return CFP->getValueAPF().isPosZero();
1882 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1883 /// the given operands.
1885 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1886 SDValue &ARMCC, SelectionDAG &DAG,
1887 DebugLoc dl) const {
1888 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1889 unsigned C = RHSC->getZExtValue();
1890 if (!isLegalICmpImmediate(C)) {
1891 // Constant does not fit, try adjusting it by one?
1896 if (isLegalICmpImmediate(C-1)) {
1897 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1898 RHS = DAG.getConstant(C-1, MVT::i32);
1903 if (C > 0 && isLegalICmpImmediate(C-1)) {
1904 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1905 RHS = DAG.getConstant(C-1, MVT::i32);
1910 if (isLegalICmpImmediate(C+1)) {
1911 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1912 RHS = DAG.getConstant(C+1, MVT::i32);
1917 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1918 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1919 RHS = DAG.getConstant(C+1, MVT::i32);
1926 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1927 ARMISD::NodeType CompareType;
1930 CompareType = ARMISD::CMP;
1935 CompareType = ARMISD::CMPZ;
1938 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1939 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1942 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1943 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1946 if (!isFloatingPointZero(RHS))
1947 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1949 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1950 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1953 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
1954 EVT VT = Op.getValueType();
1955 SDValue LHS = Op.getOperand(0);
1956 SDValue RHS = Op.getOperand(1);
1957 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1958 SDValue TrueVal = Op.getOperand(2);
1959 SDValue FalseVal = Op.getOperand(3);
1960 DebugLoc dl = Op.getDebugLoc();
1962 if (LHS.getValueType() == MVT::i32) {
1964 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1965 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1966 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1969 ARMCC::CondCodes CondCode, CondCode2;
1970 FPCCToARMCC(CC, CondCode, CondCode2);
1972 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1973 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1974 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1975 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1977 if (CondCode2 != ARMCC::AL) {
1978 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1979 // FIXME: Needs another CMP because flag can have but one use.
1980 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1981 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1982 Result, TrueVal, ARMCC2, CCR, Cmp2);
1987 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1988 SDValue Chain = Op.getOperand(0);
1989 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1990 SDValue LHS = Op.getOperand(2);
1991 SDValue RHS = Op.getOperand(3);
1992 SDValue Dest = Op.getOperand(4);
1993 DebugLoc dl = Op.getDebugLoc();
1995 if (LHS.getValueType() == MVT::i32) {
1997 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1998 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1999 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2000 Chain, Dest, ARMCC, CCR,Cmp);
2003 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2004 ARMCC::CondCodes CondCode, CondCode2;
2005 FPCCToARMCC(CC, CondCode, CondCode2);
2007 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2008 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2009 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2010 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2011 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2012 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2013 if (CondCode2 != ARMCC::AL) {
2014 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2015 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2016 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2021 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2022 SDValue Chain = Op.getOperand(0);
2023 SDValue Table = Op.getOperand(1);
2024 SDValue Index = Op.getOperand(2);
2025 DebugLoc dl = Op.getDebugLoc();
2027 EVT PTy = getPointerTy();
2028 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2029 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2030 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2031 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2032 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2033 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2034 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2035 if (Subtarget->isThumb2()) {
2036 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2037 // which does another jump to the destination. This also makes it easier
2038 // to translate it to TBB / TBH later.
2039 // FIXME: This might not work if the function is extremely large.
2040 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2041 Addr, Op.getOperand(2), JTI, UId);
2043 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2044 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2045 PseudoSourceValue::getJumpTable(), 0,
2047 Chain = Addr.getValue(1);
2048 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2049 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2051 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2052 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2053 Chain = Addr.getValue(1);
2054 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2058 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2059 DebugLoc dl = Op.getDebugLoc();
2062 switch (Op.getOpcode()) {
2064 assert(0 && "Invalid opcode!");
2065 case ISD::FP_TO_SINT:
2066 Opc = ARMISD::FTOSI;
2068 case ISD::FP_TO_UINT:
2069 Opc = ARMISD::FTOUI;
2072 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2073 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2076 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2077 EVT VT = Op.getValueType();
2078 DebugLoc dl = Op.getDebugLoc();
2081 switch (Op.getOpcode()) {
2083 assert(0 && "Invalid opcode!");
2084 case ISD::SINT_TO_FP:
2085 Opc = ARMISD::SITOF;
2087 case ISD::UINT_TO_FP:
2088 Opc = ARMISD::UITOF;
2092 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2093 return DAG.getNode(Opc, dl, VT, Op);
2096 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2097 // Implement fcopysign with a fabs and a conditional fneg.
2098 SDValue Tmp0 = Op.getOperand(0);
2099 SDValue Tmp1 = Op.getOperand(1);
2100 DebugLoc dl = Op.getDebugLoc();
2101 EVT VT = Op.getValueType();
2102 EVT SrcVT = Tmp1.getValueType();
2103 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2104 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2105 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2106 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2107 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2110 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2111 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2112 MFI->setFrameAddressIsTaken(true);
2113 EVT VT = Op.getValueType();
2114 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2115 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2116 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2117 ? ARM::R7 : ARM::R11;
2118 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2120 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2126 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2128 SDValue Dst, SDValue Src,
2129 SDValue Size, unsigned Align,
2130 bool isVolatile, bool AlwaysInline,
2134 uint64_t SrcSVOff) const {
2135 // Do repeated 4-byte loads and stores. To be improved.
2136 // This requires 4-byte alignment.
2137 if ((Align & 3) != 0)
2139 // This requires the copy size to be a constant, preferrably
2140 // within a subtarget-specific limit.
2141 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2144 uint64_t SizeVal = ConstantSize->getZExtValue();
2145 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2148 unsigned BytesLeft = SizeVal & 3;
2149 unsigned NumMemOps = SizeVal >> 2;
2150 unsigned EmittedNumMemOps = 0;
2152 unsigned VTSize = 4;
2154 const unsigned MAX_LOADS_IN_LDM = 6;
2155 SDValue TFOps[MAX_LOADS_IN_LDM];
2156 SDValue Loads[MAX_LOADS_IN_LDM];
2157 uint64_t SrcOff = 0, DstOff = 0;
2159 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2160 // same number of stores. The loads and stores will get combined into
2161 // ldm/stm later on.
2162 while (EmittedNumMemOps < NumMemOps) {
2164 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2165 Loads[i] = DAG.getLoad(VT, dl, Chain,
2166 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2167 DAG.getConstant(SrcOff, MVT::i32)),
2168 SrcSV, SrcSVOff + SrcOff, isVolatile, false, 0);
2169 TFOps[i] = Loads[i].getValue(1);
2172 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2175 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2176 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2177 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2178 DAG.getConstant(DstOff, MVT::i32)),
2179 DstSV, DstSVOff + DstOff, isVolatile, false, 0);
2182 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2184 EmittedNumMemOps += i;
2190 // Issue loads / stores for the trailing (1 - 3) bytes.
2191 unsigned BytesLeftSave = BytesLeft;
2194 if (BytesLeft >= 2) {
2202 Loads[i] = DAG.getLoad(VT, dl, Chain,
2203 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2204 DAG.getConstant(SrcOff, MVT::i32)),
2205 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2206 TFOps[i] = Loads[i].getValue(1);
2209 BytesLeft -= VTSize;
2211 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2214 BytesLeft = BytesLeftSave;
2216 if (BytesLeft >= 2) {
2224 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2225 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2226 DAG.getConstant(DstOff, MVT::i32)),
2227 DstSV, DstSVOff + DstOff, false, false, 0);
2230 BytesLeft -= VTSize;
2232 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2235 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2236 /// expand a bit convert where either the source or destination type is i64 to
2237 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2238 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2239 /// vectors), since the legalizer won't know what to do with that.
2240 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2241 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2242 DebugLoc dl = N->getDebugLoc();
2243 SDValue Op = N->getOperand(0);
2245 // This function is only supposed to be called for i64 types, either as the
2246 // source or destination of the bit convert.
2247 EVT SrcVT = Op.getValueType();
2248 EVT DstVT = N->getValueType(0);
2249 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2250 "ExpandBIT_CONVERT called for non-i64 type");
2252 // Turn i64->f64 into VMOVDRR.
2253 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2254 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2255 DAG.getConstant(0, MVT::i32));
2256 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2257 DAG.getConstant(1, MVT::i32));
2258 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2261 // Turn f64->i64 into VMOVRRD.
2262 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2263 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2264 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2265 // Merge the pieces into a single i64 value.
2266 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2272 /// getZeroVector - Returns a vector of specified type with all zero elements.
2274 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2275 assert(VT.isVector() && "Expected a vector type");
2277 // Zero vectors are used to represent vector negation and in those cases
2278 // will be implemented with the NEON VNEG instruction. However, VNEG does
2279 // not support i64 elements, so sometimes the zero vectors will need to be
2280 // explicitly constructed. For those cases, and potentially other uses in
2281 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2282 // to their dest type. This ensures they get CSE'd.
2284 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2285 SmallVector<SDValue, 8> Ops;
2288 if (VT.getSizeInBits() == 64) {
2289 Ops.assign(8, Cst); TVT = MVT::v8i8;
2291 Ops.assign(16, Cst); TVT = MVT::v16i8;
2293 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2295 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2298 /// getOnesVector - Returns a vector of specified type with all bits set.
2300 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2301 assert(VT.isVector() && "Expected a vector type");
2303 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2304 // dest type. This ensures they get CSE'd.
2306 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2307 SmallVector<SDValue, 8> Ops;
2310 if (VT.getSizeInBits() == 64) {
2311 Ops.assign(8, Cst); TVT = MVT::v8i8;
2313 Ops.assign(16, Cst); TVT = MVT::v16i8;
2315 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2317 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2320 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2321 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2322 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2323 SelectionDAG &DAG) const {
2324 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2325 EVT VT = Op.getValueType();
2326 unsigned VTBits = VT.getSizeInBits();
2327 DebugLoc dl = Op.getDebugLoc();
2328 SDValue ShOpLo = Op.getOperand(0);
2329 SDValue ShOpHi = Op.getOperand(1);
2330 SDValue ShAmt = Op.getOperand(2);
2332 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2334 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2336 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2337 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2338 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2339 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2340 DAG.getConstant(VTBits, MVT::i32));
2341 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2342 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2343 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2345 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2346 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2348 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2349 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2352 SDValue Ops[2] = { Lo, Hi };
2353 return DAG.getMergeValues(Ops, 2, dl);
2356 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2357 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2358 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2359 SelectionDAG &DAG) const {
2360 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2361 EVT VT = Op.getValueType();
2362 unsigned VTBits = VT.getSizeInBits();
2363 DebugLoc dl = Op.getDebugLoc();
2364 SDValue ShOpLo = Op.getOperand(0);
2365 SDValue ShOpHi = Op.getOperand(1);
2366 SDValue ShAmt = Op.getOperand(2);
2369 assert(Op.getOpcode() == ISD::SHL_PARTS);
2370 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2371 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2372 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2373 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2374 DAG.getConstant(VTBits, MVT::i32));
2375 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2376 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2378 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2379 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2380 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2382 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2383 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2386 SDValue Ops[2] = { Lo, Hi };
2387 return DAG.getMergeValues(Ops, 2, dl);
2390 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2391 const ARMSubtarget *ST) {
2392 EVT VT = N->getValueType(0);
2393 DebugLoc dl = N->getDebugLoc();
2395 if (!ST->hasV6T2Ops())
2398 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2399 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2402 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2403 const ARMSubtarget *ST) {
2404 EVT VT = N->getValueType(0);
2405 DebugLoc dl = N->getDebugLoc();
2407 // Lower vector shifts on NEON to use VSHL.
2408 if (VT.isVector()) {
2409 assert(ST->hasNEON() && "unexpected vector shift");
2411 // Left shifts translate directly to the vshiftu intrinsic.
2412 if (N->getOpcode() == ISD::SHL)
2413 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2414 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2415 N->getOperand(0), N->getOperand(1));
2417 assert((N->getOpcode() == ISD::SRA ||
2418 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2420 // NEON uses the same intrinsics for both left and right shifts. For
2421 // right shifts, the shift amounts are negative, so negate the vector of
2423 EVT ShiftVT = N->getOperand(1).getValueType();
2424 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2425 getZeroVector(ShiftVT, DAG, dl),
2427 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2428 Intrinsic::arm_neon_vshifts :
2429 Intrinsic::arm_neon_vshiftu);
2430 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2431 DAG.getConstant(vshiftInt, MVT::i32),
2432 N->getOperand(0), NegatedCount);
2435 // We can get here for a node like i32 = ISD::SHL i32, i64
2439 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2440 "Unknown shift to lower!");
2442 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2443 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2444 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2447 // If we are in thumb mode, we don't have RRX.
2448 if (ST->isThumb1Only()) return SDValue();
2450 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2451 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2452 DAG.getConstant(0, MVT::i32));
2453 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2454 DAG.getConstant(1, MVT::i32));
2456 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2457 // captures the result into a carry flag.
2458 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2459 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2461 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2462 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2464 // Merge the pieces into a single i64 value.
2465 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2468 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2469 SDValue TmpOp0, TmpOp1;
2470 bool Invert = false;
2474 SDValue Op0 = Op.getOperand(0);
2475 SDValue Op1 = Op.getOperand(1);
2476 SDValue CC = Op.getOperand(2);
2477 EVT VT = Op.getValueType();
2478 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2479 DebugLoc dl = Op.getDebugLoc();
2481 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2482 switch (SetCCOpcode) {
2483 default: llvm_unreachable("Illegal FP comparison"); break;
2485 case ISD::SETNE: Invert = true; // Fallthrough
2487 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2489 case ISD::SETLT: Swap = true; // Fallthrough
2491 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2493 case ISD::SETLE: Swap = true; // Fallthrough
2495 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2496 case ISD::SETUGE: Swap = true; // Fallthrough
2497 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2498 case ISD::SETUGT: Swap = true; // Fallthrough
2499 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2500 case ISD::SETUEQ: Invert = true; // Fallthrough
2502 // Expand this to (OLT | OGT).
2506 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2507 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2509 case ISD::SETUO: Invert = true; // Fallthrough
2511 // Expand this to (OLT | OGE).
2515 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2516 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2520 // Integer comparisons.
2521 switch (SetCCOpcode) {
2522 default: llvm_unreachable("Illegal integer comparison"); break;
2523 case ISD::SETNE: Invert = true;
2524 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2525 case ISD::SETLT: Swap = true;
2526 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2527 case ISD::SETLE: Swap = true;
2528 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2529 case ISD::SETULT: Swap = true;
2530 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2531 case ISD::SETULE: Swap = true;
2532 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2535 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2536 if (Opc == ARMISD::VCEQ) {
2539 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2541 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2544 // Ignore bitconvert.
2545 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2546 AndOp = AndOp.getOperand(0);
2548 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2550 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2551 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2558 std::swap(Op0, Op1);
2560 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2563 Result = DAG.getNOT(dl, Result, VT);
2568 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2569 /// VMOV instruction, and if so, return the constant being splatted.
2570 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2571 unsigned SplatBitSize, SelectionDAG &DAG) {
2572 switch (SplatBitSize) {
2574 // Any 1-byte value is OK.
2575 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2576 return DAG.getTargetConstant(SplatBits, MVT::i8);
2579 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2580 if ((SplatBits & ~0xff) == 0 ||
2581 (SplatBits & ~0xff00) == 0)
2582 return DAG.getTargetConstant(SplatBits, MVT::i16);
2586 // NEON's 32-bit VMOV supports splat values where:
2587 // * only one byte is nonzero, or
2588 // * the least significant byte is 0xff and the second byte is nonzero, or
2589 // * the least significant 2 bytes are 0xff and the third is nonzero.
2590 if ((SplatBits & ~0xff) == 0 ||
2591 (SplatBits & ~0xff00) == 0 ||
2592 (SplatBits & ~0xff0000) == 0 ||
2593 (SplatBits & ~0xff000000) == 0)
2594 return DAG.getTargetConstant(SplatBits, MVT::i32);
2596 if ((SplatBits & ~0xffff) == 0 &&
2597 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2598 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2600 if ((SplatBits & ~0xffffff) == 0 &&
2601 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2602 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2604 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2605 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2606 // VMOV.I32. A (very) minor optimization would be to replicate the value
2607 // and fall through here to test for a valid 64-bit splat. But, then the
2608 // caller would also need to check and handle the change in size.
2612 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2613 uint64_t BitMask = 0xff;
2615 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2616 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2618 else if ((SplatBits & BitMask) != 0)
2622 return DAG.getTargetConstant(Val, MVT::i64);
2626 llvm_unreachable("unexpected size for isVMOVSplat");
2633 /// getVMOVImm - If this is a build_vector of constants which can be
2634 /// formed by using a VMOV instruction of the specified element size,
2635 /// return the constant being splatted. The ByteSize field indicates the
2636 /// number of bytes of each element [1248].
2637 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2638 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2639 APInt SplatBits, SplatUndef;
2640 unsigned SplatBitSize;
2642 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2643 HasAnyUndefs, ByteSize * 8))
2646 if (SplatBitSize > ByteSize * 8)
2649 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2653 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2654 bool &ReverseVEXT, unsigned &Imm) {
2655 unsigned NumElts = VT.getVectorNumElements();
2656 ReverseVEXT = false;
2659 // If this is a VEXT shuffle, the immediate value is the index of the first
2660 // element. The other shuffle indices must be the successive elements after
2662 unsigned ExpectedElt = Imm;
2663 for (unsigned i = 1; i < NumElts; ++i) {
2664 // Increment the expected index. If it wraps around, it may still be
2665 // a VEXT but the source vectors must be swapped.
2667 if (ExpectedElt == NumElts * 2) {
2672 if (ExpectedElt != static_cast<unsigned>(M[i]))
2676 // Adjust the index value if the source operands will be swapped.
2683 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2684 /// instruction with the specified blocksize. (The order of the elements
2685 /// within each block of the vector is reversed.)
2686 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2687 unsigned BlockSize) {
2688 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2689 "Only possible block sizes for VREV are: 16, 32, 64");
2691 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2695 unsigned NumElts = VT.getVectorNumElements();
2696 unsigned BlockElts = M[0] + 1;
2698 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2701 for (unsigned i = 0; i < NumElts; ++i) {
2702 if ((unsigned) M[i] !=
2703 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2710 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2711 unsigned &WhichResult) {
2712 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2716 unsigned NumElts = VT.getVectorNumElements();
2717 WhichResult = (M[0] == 0 ? 0 : 1);
2718 for (unsigned i = 0; i < NumElts; i += 2) {
2719 if ((unsigned) M[i] != i + WhichResult ||
2720 (unsigned) M[i+1] != i + NumElts + WhichResult)
2726 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2727 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2728 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2729 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2730 unsigned &WhichResult) {
2731 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2735 unsigned NumElts = VT.getVectorNumElements();
2736 WhichResult = (M[0] == 0 ? 0 : 1);
2737 for (unsigned i = 0; i < NumElts; i += 2) {
2738 if ((unsigned) M[i] != i + WhichResult ||
2739 (unsigned) M[i+1] != i + WhichResult)
2745 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2746 unsigned &WhichResult) {
2747 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2751 unsigned NumElts = VT.getVectorNumElements();
2752 WhichResult = (M[0] == 0 ? 0 : 1);
2753 for (unsigned i = 0; i != NumElts; ++i) {
2754 if ((unsigned) M[i] != 2 * i + WhichResult)
2758 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2759 if (VT.is64BitVector() && EltSz == 32)
2765 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2766 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2767 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2768 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2769 unsigned &WhichResult) {
2770 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2774 unsigned Half = VT.getVectorNumElements() / 2;
2775 WhichResult = (M[0] == 0 ? 0 : 1);
2776 for (unsigned j = 0; j != 2; ++j) {
2777 unsigned Idx = WhichResult;
2778 for (unsigned i = 0; i != Half; ++i) {
2779 if ((unsigned) M[i + j * Half] != Idx)
2785 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2786 if (VT.is64BitVector() && EltSz == 32)
2792 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2793 unsigned &WhichResult) {
2794 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2798 unsigned NumElts = VT.getVectorNumElements();
2799 WhichResult = (M[0] == 0 ? 0 : 1);
2800 unsigned Idx = WhichResult * NumElts / 2;
2801 for (unsigned i = 0; i != NumElts; i += 2) {
2802 if ((unsigned) M[i] != Idx ||
2803 (unsigned) M[i+1] != Idx + NumElts)
2808 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2809 if (VT.is64BitVector() && EltSz == 32)
2815 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2816 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2817 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2818 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2819 unsigned &WhichResult) {
2820 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2824 unsigned NumElts = VT.getVectorNumElements();
2825 WhichResult = (M[0] == 0 ? 0 : 1);
2826 unsigned Idx = WhichResult * NumElts / 2;
2827 for (unsigned i = 0; i != NumElts; i += 2) {
2828 if ((unsigned) M[i] != Idx ||
2829 (unsigned) M[i+1] != Idx)
2834 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2835 if (VT.is64BitVector() && EltSz == 32)
2842 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2843 // Canonicalize all-zeros and all-ones vectors.
2844 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2845 if (ConstVal->isNullValue())
2846 return getZeroVector(VT, DAG, dl);
2847 if (ConstVal->isAllOnesValue())
2848 return getOnesVector(VT, DAG, dl);
2851 if (VT.is64BitVector()) {
2852 switch (Val.getValueType().getSizeInBits()) {
2853 case 8: CanonicalVT = MVT::v8i8; break;
2854 case 16: CanonicalVT = MVT::v4i16; break;
2855 case 32: CanonicalVT = MVT::v2i32; break;
2856 case 64: CanonicalVT = MVT::v1i64; break;
2857 default: llvm_unreachable("unexpected splat element type"); break;
2860 assert(VT.is128BitVector() && "unknown splat vector size");
2861 switch (Val.getValueType().getSizeInBits()) {
2862 case 8: CanonicalVT = MVT::v16i8; break;
2863 case 16: CanonicalVT = MVT::v8i16; break;
2864 case 32: CanonicalVT = MVT::v4i32; break;
2865 case 64: CanonicalVT = MVT::v2i64; break;
2866 default: llvm_unreachable("unexpected splat element type"); break;
2870 // Build a canonical splat for this value.
2871 SmallVector<SDValue, 8> Ops;
2872 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2873 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2875 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2878 // If this is a case we can't handle, return null and let the default
2879 // expansion code take care of it.
2880 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2881 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2882 DebugLoc dl = Op.getDebugLoc();
2883 EVT VT = Op.getValueType();
2885 APInt SplatBits, SplatUndef;
2886 unsigned SplatBitSize;
2888 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2889 if (SplatBitSize <= 64) {
2890 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2891 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2893 return BuildSplat(Val, VT, DAG, dl);
2897 // If there are only 2 elements in a 128-bit vector, insert them into an
2898 // undef vector. This handles the common case for 128-bit vector argument
2899 // passing, where the insertions should be translated to subreg accesses
2900 // with no real instructions.
2901 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2902 SDValue Val = DAG.getUNDEF(VT);
2903 SDValue Op0 = Op.getOperand(0);
2904 SDValue Op1 = Op.getOperand(1);
2905 if (Op0.getOpcode() != ISD::UNDEF)
2906 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2907 DAG.getIntPtrConstant(0));
2908 if (Op1.getOpcode() != ISD::UNDEF)
2909 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2910 DAG.getIntPtrConstant(1));
2917 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2918 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2919 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2920 /// are assumed to be legal.
2922 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2924 if (VT.getVectorNumElements() == 4 &&
2925 (VT.is128BitVector() || VT.is64BitVector())) {
2926 unsigned PFIndexes[4];
2927 for (unsigned i = 0; i != 4; ++i) {
2931 PFIndexes[i] = M[i];
2934 // Compute the index in the perfect shuffle table.
2935 unsigned PFTableIndex =
2936 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2937 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2938 unsigned Cost = (PFEntry >> 30);
2945 unsigned Imm, WhichResult;
2947 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2948 isVREVMask(M, VT, 64) ||
2949 isVREVMask(M, VT, 32) ||
2950 isVREVMask(M, VT, 16) ||
2951 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2952 isVTRNMask(M, VT, WhichResult) ||
2953 isVUZPMask(M, VT, WhichResult) ||
2954 isVZIPMask(M, VT, WhichResult) ||
2955 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2956 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2957 isVZIP_v_undef_Mask(M, VT, WhichResult));
2960 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2961 /// the specified operations to build the shuffle.
2962 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2963 SDValue RHS, SelectionDAG &DAG,
2965 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2966 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2967 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2970 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2979 OP_VUZPL, // VUZP, left result
2980 OP_VUZPR, // VUZP, right result
2981 OP_VZIPL, // VZIP, left result
2982 OP_VZIPR, // VZIP, right result
2983 OP_VTRNL, // VTRN, left result
2984 OP_VTRNR // VTRN, right result
2987 if (OpNum == OP_COPY) {
2988 if (LHSID == (1*9+2)*9+3) return LHS;
2989 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2993 SDValue OpLHS, OpRHS;
2994 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2995 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2996 EVT VT = OpLHS.getValueType();
2999 default: llvm_unreachable("Unknown shuffle opcode!");
3001 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3006 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3007 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3011 return DAG.getNode(ARMISD::VEXT, dl, VT,
3013 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3016 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3017 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3020 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3021 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3024 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3025 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3029 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3030 SDValue V1 = Op.getOperand(0);
3031 SDValue V2 = Op.getOperand(1);
3032 DebugLoc dl = Op.getDebugLoc();
3033 EVT VT = Op.getValueType();
3034 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3035 SmallVector<int, 8> ShuffleMask;
3037 // Convert shuffles that are directly supported on NEON to target-specific
3038 // DAG nodes, instead of keeping them as shuffles and matching them again
3039 // during code selection. This is more efficient and avoids the possibility
3040 // of inconsistencies between legalization and selection.
3041 // FIXME: floating-point vectors should be canonicalized to integer vectors
3042 // of the same time so that they get CSEd properly.
3043 SVN->getMask(ShuffleMask);
3045 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3046 int Lane = SVN->getSplatIndex();
3047 // If this is undef splat, generate it via "just" vdup, if possible.
3048 if (Lane == -1) Lane = 0;
3050 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3051 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3053 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3054 DAG.getConstant(Lane, MVT::i32));
3059 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3062 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3063 DAG.getConstant(Imm, MVT::i32));
3066 if (isVREVMask(ShuffleMask, VT, 64))
3067 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3068 if (isVREVMask(ShuffleMask, VT, 32))
3069 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3070 if (isVREVMask(ShuffleMask, VT, 16))
3071 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3073 // Check for Neon shuffles that modify both input vectors in place.
3074 // If both results are used, i.e., if there are two shuffles with the same
3075 // source operands and with masks corresponding to both results of one of
3076 // these operations, DAG memoization will ensure that a single node is
3077 // used for both shuffles.
3078 unsigned WhichResult;
3079 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3080 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3081 V1, V2).getValue(WhichResult);
3082 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3083 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3084 V1, V2).getValue(WhichResult);
3085 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3086 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3087 V1, V2).getValue(WhichResult);
3089 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3090 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3091 V1, V1).getValue(WhichResult);
3092 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3093 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3094 V1, V1).getValue(WhichResult);
3095 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3096 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3097 V1, V1).getValue(WhichResult);
3099 // If the shuffle is not directly supported and it has 4 elements, use
3100 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3101 if (VT.getVectorNumElements() == 4 &&
3102 (VT.is128BitVector() || VT.is64BitVector())) {
3103 unsigned PFIndexes[4];
3104 for (unsigned i = 0; i != 4; ++i) {
3105 if (ShuffleMask[i] < 0)
3108 PFIndexes[i] = ShuffleMask[i];
3111 // Compute the index in the perfect shuffle table.
3112 unsigned PFTableIndex =
3113 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3115 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3116 unsigned Cost = (PFEntry >> 30);
3119 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3125 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3126 EVT VT = Op.getValueType();
3127 DebugLoc dl = Op.getDebugLoc();
3128 SDValue Vec = Op.getOperand(0);
3129 SDValue Lane = Op.getOperand(1);
3130 assert(VT == MVT::i32 &&
3131 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3132 "unexpected type for custom-lowering vector extract");
3133 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3136 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3137 // The only time a CONCAT_VECTORS operation can have legal types is when
3138 // two 64-bit vectors are concatenated to a 128-bit vector.
3139 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3140 "unexpected CONCAT_VECTORS");
3141 DebugLoc dl = Op.getDebugLoc();
3142 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3143 SDValue Op0 = Op.getOperand(0);
3144 SDValue Op1 = Op.getOperand(1);
3145 if (Op0.getOpcode() != ISD::UNDEF)
3146 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3147 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3148 DAG.getIntPtrConstant(0));
3149 if (Op1.getOpcode() != ISD::UNDEF)
3150 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3151 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3152 DAG.getIntPtrConstant(1));
3153 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3156 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3157 switch (Op.getOpcode()) {
3158 default: llvm_unreachable("Don't know how to custom lower this!");
3159 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3160 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3161 case ISD::GlobalAddress:
3162 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3163 LowerGlobalAddressELF(Op, DAG);
3164 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3165 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3166 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3167 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3168 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3169 case ISD::VASTART: return LowerVASTART(Op, DAG);
3170 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3171 case ISD::SINT_TO_FP:
3172 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3173 case ISD::FP_TO_SINT:
3174 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3175 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3176 case ISD::RETURNADDR: break;
3177 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3178 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3179 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3181 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3184 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3185 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3186 case ISD::SRL_PARTS:
3187 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3188 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3189 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3190 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3191 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3192 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3193 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3198 /// ReplaceNodeResults - Replace the results of node with an illegal result
3199 /// type with new values built out of custom code.
3200 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3201 SmallVectorImpl<SDValue>&Results,
3202 SelectionDAG &DAG) const {
3204 switch (N->getOpcode()) {
3206 llvm_unreachable("Don't know how to custom expand this!");
3208 case ISD::BIT_CONVERT:
3209 Res = ExpandBIT_CONVERT(N, DAG);
3213 Res = LowerShift(N, DAG, Subtarget);
3217 Results.push_back(Res);
3220 //===----------------------------------------------------------------------===//
3221 // ARM Scheduler Hooks
3222 //===----------------------------------------------------------------------===//
3225 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3226 MachineBasicBlock *BB,
3227 unsigned Size) const {
3228 unsigned dest = MI->getOperand(0).getReg();
3229 unsigned ptr = MI->getOperand(1).getReg();
3230 unsigned oldval = MI->getOperand(2).getReg();
3231 unsigned newval = MI->getOperand(3).getReg();
3232 unsigned scratch = BB->getParent()->getRegInfo()
3233 .createVirtualRegister(ARM::GPRRegisterClass);
3234 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3235 DebugLoc dl = MI->getDebugLoc();
3236 bool isThumb2 = Subtarget->isThumb2();
3238 unsigned ldrOpc, strOpc;
3240 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3242 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3243 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3246 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3247 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3250 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3251 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3255 MachineFunction *MF = BB->getParent();
3256 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3257 MachineFunction::iterator It = BB;
3258 ++It; // insert the new blocks after the current block
3260 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3261 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3262 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3263 MF->insert(It, loop1MBB);
3264 MF->insert(It, loop2MBB);
3265 MF->insert(It, exitMBB);
3266 exitMBB->transferSuccessors(BB);
3270 // fallthrough --> loop1MBB
3271 BB->addSuccessor(loop1MBB);
3274 // ldrex dest, [ptr]
3278 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3279 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3280 .addReg(dest).addReg(oldval));
3281 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3282 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3283 BB->addSuccessor(loop2MBB);
3284 BB->addSuccessor(exitMBB);
3287 // strex scratch, newval, [ptr]
3291 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3293 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3294 .addReg(scratch).addImm(0));
3295 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3296 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3297 BB->addSuccessor(loop1MBB);
3298 BB->addSuccessor(exitMBB);
3304 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3310 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3311 unsigned Size, unsigned BinOpcode) const {
3312 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3313 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3315 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3316 MachineFunction *MF = BB->getParent();
3317 MachineFunction::iterator It = BB;
3320 unsigned dest = MI->getOperand(0).getReg();
3321 unsigned ptr = MI->getOperand(1).getReg();
3322 unsigned incr = MI->getOperand(2).getReg();
3323 DebugLoc dl = MI->getDebugLoc();
3325 bool isThumb2 = Subtarget->isThumb2();
3326 unsigned ldrOpc, strOpc;
3328 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3330 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3331 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3334 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3335 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3338 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3339 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3343 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3344 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3345 MF->insert(It, loopMBB);
3346 MF->insert(It, exitMBB);
3347 exitMBB->transferSuccessors(BB);
3349 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3350 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3351 unsigned scratch2 = (!BinOpcode) ? incr :
3352 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3356 // fallthrough --> loopMBB
3357 BB->addSuccessor(loopMBB);
3361 // <binop> scratch2, dest, incr
3362 // strex scratch, scratch2, ptr
3365 // fallthrough --> exitMBB
3367 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3369 // operand order needs to go the other way for NAND
3370 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3371 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3372 addReg(incr).addReg(dest)).addReg(0);
3374 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3375 addReg(dest).addReg(incr)).addReg(0);
3378 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3380 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3381 .addReg(scratch).addImm(0));
3382 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3383 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3385 BB->addSuccessor(loopMBB);
3386 BB->addSuccessor(exitMBB);
3392 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3398 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3399 MachineBasicBlock *BB) const {
3400 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3401 DebugLoc dl = MI->getDebugLoc();
3402 bool isThumb2 = Subtarget->isThumb2();
3403 switch (MI->getOpcode()) {
3406 llvm_unreachable("Unexpected instr type to insert");
3408 case ARM::ATOMIC_LOAD_ADD_I8:
3409 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3410 case ARM::ATOMIC_LOAD_ADD_I16:
3411 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3412 case ARM::ATOMIC_LOAD_ADD_I32:
3413 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3415 case ARM::ATOMIC_LOAD_AND_I8:
3416 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3417 case ARM::ATOMIC_LOAD_AND_I16:
3418 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3419 case ARM::ATOMIC_LOAD_AND_I32:
3420 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3422 case ARM::ATOMIC_LOAD_OR_I8:
3423 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3424 case ARM::ATOMIC_LOAD_OR_I16:
3425 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3426 case ARM::ATOMIC_LOAD_OR_I32:
3427 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3429 case ARM::ATOMIC_LOAD_XOR_I8:
3430 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3431 case ARM::ATOMIC_LOAD_XOR_I16:
3432 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3433 case ARM::ATOMIC_LOAD_XOR_I32:
3434 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3436 case ARM::ATOMIC_LOAD_NAND_I8:
3437 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3438 case ARM::ATOMIC_LOAD_NAND_I16:
3439 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3440 case ARM::ATOMIC_LOAD_NAND_I32:
3441 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3443 case ARM::ATOMIC_LOAD_SUB_I8:
3444 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3445 case ARM::ATOMIC_LOAD_SUB_I16:
3446 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3447 case ARM::ATOMIC_LOAD_SUB_I32:
3448 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3450 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3451 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3452 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3454 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3455 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3456 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3458 case ARM::tMOVCCr_pseudo: {
3459 // To "insert" a SELECT_CC instruction, we actually have to insert the
3460 // diamond control-flow pattern. The incoming instruction knows the
3461 // destination vreg to set, the condition code register to branch on, the
3462 // true/false values to select between, and a branch opcode to use.
3463 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3464 MachineFunction::iterator It = BB;
3470 // cmpTY ccX, r1, r2
3472 // fallthrough --> copy0MBB
3473 MachineBasicBlock *thisMBB = BB;
3474 MachineFunction *F = BB->getParent();
3475 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3476 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3477 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3478 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3479 F->insert(It, copy0MBB);
3480 F->insert(It, sinkMBB);
3481 // Update machine-CFG edges by first adding all successors of the current
3482 // block to the new block which will contain the Phi node for the select.
3483 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3484 E = BB->succ_end(); I != E; ++I)
3485 sinkMBB->addSuccessor(*I);
3486 // Next, remove all successors of the current block, and add the true
3487 // and fallthrough blocks as its successors.
3488 while (!BB->succ_empty())
3489 BB->removeSuccessor(BB->succ_begin());
3490 BB->addSuccessor(copy0MBB);
3491 BB->addSuccessor(sinkMBB);
3494 // %FalseValue = ...
3495 // # fallthrough to sinkMBB
3498 // Update machine-CFG edges
3499 BB->addSuccessor(sinkMBB);
3502 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3505 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3506 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3507 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3509 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3516 case ARM::t2SUBrSPi_:
3517 case ARM::t2SUBrSPi12_:
3518 case ARM::t2SUBrSPs_: {
3519 MachineFunction *MF = BB->getParent();
3520 unsigned DstReg = MI->getOperand(0).getReg();
3521 unsigned SrcReg = MI->getOperand(1).getReg();
3522 bool DstIsDead = MI->getOperand(0).isDead();
3523 bool SrcIsKill = MI->getOperand(1).isKill();
3525 if (SrcReg != ARM::SP) {
3526 // Copy the source to SP from virtual register.
3527 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3528 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3529 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3530 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3531 .addReg(SrcReg, getKillRegState(SrcIsKill));
3535 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3536 switch (MI->getOpcode()) {
3538 llvm_unreachable("Unexpected pseudo instruction!");
3544 OpOpc = ARM::tADDspr;
3547 OpOpc = ARM::tSUBspi;
3549 case ARM::t2SUBrSPi_:
3550 OpOpc = ARM::t2SUBrSPi;
3551 NeedPred = true; NeedCC = true;
3553 case ARM::t2SUBrSPi12_:
3554 OpOpc = ARM::t2SUBrSPi12;
3557 case ARM::t2SUBrSPs_:
3558 OpOpc = ARM::t2SUBrSPs;
3559 NeedPred = true; NeedCC = true; NeedOp3 = true;
3562 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3563 if (OpOpc == ARM::tAND)
3564 AddDefaultT1CC(MIB);
3565 MIB.addReg(ARM::SP);
3566 MIB.addOperand(MI->getOperand(2));
3568 MIB.addOperand(MI->getOperand(3));
3570 AddDefaultPred(MIB);
3574 // Copy the result from SP to virtual register.
3575 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3576 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3577 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3578 BuildMI(BB, dl, TII->get(CopyOpc))
3579 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3581 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3587 //===----------------------------------------------------------------------===//
3588 // ARM Optimization Hooks
3589 //===----------------------------------------------------------------------===//
3592 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3593 TargetLowering::DAGCombinerInfo &DCI) {
3594 SelectionDAG &DAG = DCI.DAG;
3595 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3596 EVT VT = N->getValueType(0);
3597 unsigned Opc = N->getOpcode();
3598 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3599 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3600 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3601 ISD::CondCode CC = ISD::SETCC_INVALID;
3604 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3606 SDValue CCOp = Slct.getOperand(0);
3607 if (CCOp.getOpcode() == ISD::SETCC)
3608 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3611 bool DoXform = false;
3613 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3616 if (LHS.getOpcode() == ISD::Constant &&
3617 cast<ConstantSDNode>(LHS)->isNullValue()) {
3619 } else if (CC != ISD::SETCC_INVALID &&
3620 RHS.getOpcode() == ISD::Constant &&
3621 cast<ConstantSDNode>(RHS)->isNullValue()) {
3622 std::swap(LHS, RHS);
3623 SDValue Op0 = Slct.getOperand(0);
3624 EVT OpVT = isSlctCC ? Op0.getValueType() :
3625 Op0.getOperand(0).getValueType();
3626 bool isInt = OpVT.isInteger();
3627 CC = ISD::getSetCCInverse(CC, isInt);
3629 if (!TLI.isCondCodeLegal(CC, OpVT))
3630 return SDValue(); // Inverse operator isn't legal.
3637 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3639 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3640 Slct.getOperand(0), Slct.getOperand(1), CC);
3641 SDValue CCOp = Slct.getOperand(0);
3643 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3644 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3645 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3646 CCOp, OtherOp, Result);
3651 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3652 static SDValue PerformADDCombine(SDNode *N,
3653 TargetLowering::DAGCombinerInfo &DCI) {
3654 // added by evan in r37685 with no testcase.
3655 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3657 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3658 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3659 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3660 if (Result.getNode()) return Result;
3662 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3663 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3664 if (Result.getNode()) return Result;
3670 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3671 static SDValue PerformSUBCombine(SDNode *N,
3672 TargetLowering::DAGCombinerInfo &DCI) {
3673 // added by evan in r37685 with no testcase.
3674 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3676 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3677 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3678 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3679 if (Result.getNode()) return Result;
3685 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3686 /// ARMISD::VMOVRRD.
3687 static SDValue PerformVMOVRRDCombine(SDNode *N,
3688 TargetLowering::DAGCombinerInfo &DCI) {
3689 // fmrrd(fmdrr x, y) -> x,y
3690 SDValue InDouble = N->getOperand(0);
3691 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3692 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3696 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3697 /// operand of a vector shift operation, where all the elements of the
3698 /// build_vector must have the same constant integer value.
3699 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3700 // Ignore bit_converts.
3701 while (Op.getOpcode() == ISD::BIT_CONVERT)
3702 Op = Op.getOperand(0);
3703 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3704 APInt SplatBits, SplatUndef;
3705 unsigned SplatBitSize;
3707 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3708 HasAnyUndefs, ElementBits) ||
3709 SplatBitSize > ElementBits)
3711 Cnt = SplatBits.getSExtValue();
3715 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3716 /// operand of a vector shift left operation. That value must be in the range:
3717 /// 0 <= Value < ElementBits for a left shift; or
3718 /// 0 <= Value <= ElementBits for a long left shift.
3719 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3720 assert(VT.isVector() && "vector shift count is not a vector type");
3721 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3722 if (! getVShiftImm(Op, ElementBits, Cnt))
3724 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3727 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3728 /// operand of a vector shift right operation. For a shift opcode, the value
3729 /// is positive, but for an intrinsic the value count must be negative. The
3730 /// absolute value must be in the range:
3731 /// 1 <= |Value| <= ElementBits for a right shift; or
3732 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3733 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3735 assert(VT.isVector() && "vector shift count is not a vector type");
3736 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3737 if (! getVShiftImm(Op, ElementBits, Cnt))
3741 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3744 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3745 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3746 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3749 // Don't do anything for most intrinsics.
3752 // Vector shifts: check for immediate versions and lower them.
3753 // Note: This is done during DAG combining instead of DAG legalizing because
3754 // the build_vectors for 64-bit vector element shift counts are generally
3755 // not legal, and it is hard to see their values after they get legalized to
3756 // loads from a constant pool.
3757 case Intrinsic::arm_neon_vshifts:
3758 case Intrinsic::arm_neon_vshiftu:
3759 case Intrinsic::arm_neon_vshiftls:
3760 case Intrinsic::arm_neon_vshiftlu:
3761 case Intrinsic::arm_neon_vshiftn:
3762 case Intrinsic::arm_neon_vrshifts:
3763 case Intrinsic::arm_neon_vrshiftu:
3764 case Intrinsic::arm_neon_vrshiftn:
3765 case Intrinsic::arm_neon_vqshifts:
3766 case Intrinsic::arm_neon_vqshiftu:
3767 case Intrinsic::arm_neon_vqshiftsu:
3768 case Intrinsic::arm_neon_vqshiftns:
3769 case Intrinsic::arm_neon_vqshiftnu:
3770 case Intrinsic::arm_neon_vqshiftnsu:
3771 case Intrinsic::arm_neon_vqrshiftns:
3772 case Intrinsic::arm_neon_vqrshiftnu:
3773 case Intrinsic::arm_neon_vqrshiftnsu: {
3774 EVT VT = N->getOperand(1).getValueType();
3776 unsigned VShiftOpc = 0;
3779 case Intrinsic::arm_neon_vshifts:
3780 case Intrinsic::arm_neon_vshiftu:
3781 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3782 VShiftOpc = ARMISD::VSHL;
3785 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3786 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3787 ARMISD::VSHRs : ARMISD::VSHRu);
3792 case Intrinsic::arm_neon_vshiftls:
3793 case Intrinsic::arm_neon_vshiftlu:
3794 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3796 llvm_unreachable("invalid shift count for vshll intrinsic");
3798 case Intrinsic::arm_neon_vrshifts:
3799 case Intrinsic::arm_neon_vrshiftu:
3800 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3804 case Intrinsic::arm_neon_vqshifts:
3805 case Intrinsic::arm_neon_vqshiftu:
3806 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3810 case Intrinsic::arm_neon_vqshiftsu:
3811 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3813 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3815 case Intrinsic::arm_neon_vshiftn:
3816 case Intrinsic::arm_neon_vrshiftn:
3817 case Intrinsic::arm_neon_vqshiftns:
3818 case Intrinsic::arm_neon_vqshiftnu:
3819 case Intrinsic::arm_neon_vqshiftnsu:
3820 case Intrinsic::arm_neon_vqrshiftns:
3821 case Intrinsic::arm_neon_vqrshiftnu:
3822 case Intrinsic::arm_neon_vqrshiftnsu:
3823 // Narrowing shifts require an immediate right shift.
3824 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3826 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3829 llvm_unreachable("unhandled vector shift");
3833 case Intrinsic::arm_neon_vshifts:
3834 case Intrinsic::arm_neon_vshiftu:
3835 // Opcode already set above.
3837 case Intrinsic::arm_neon_vshiftls:
3838 case Intrinsic::arm_neon_vshiftlu:
3839 if (Cnt == VT.getVectorElementType().getSizeInBits())
3840 VShiftOpc = ARMISD::VSHLLi;
3842 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3843 ARMISD::VSHLLs : ARMISD::VSHLLu);
3845 case Intrinsic::arm_neon_vshiftn:
3846 VShiftOpc = ARMISD::VSHRN; break;
3847 case Intrinsic::arm_neon_vrshifts:
3848 VShiftOpc = ARMISD::VRSHRs; break;
3849 case Intrinsic::arm_neon_vrshiftu:
3850 VShiftOpc = ARMISD::VRSHRu; break;
3851 case Intrinsic::arm_neon_vrshiftn:
3852 VShiftOpc = ARMISD::VRSHRN; break;
3853 case Intrinsic::arm_neon_vqshifts:
3854 VShiftOpc = ARMISD::VQSHLs; break;
3855 case Intrinsic::arm_neon_vqshiftu:
3856 VShiftOpc = ARMISD::VQSHLu; break;
3857 case Intrinsic::arm_neon_vqshiftsu:
3858 VShiftOpc = ARMISD::VQSHLsu; break;
3859 case Intrinsic::arm_neon_vqshiftns:
3860 VShiftOpc = ARMISD::VQSHRNs; break;
3861 case Intrinsic::arm_neon_vqshiftnu:
3862 VShiftOpc = ARMISD::VQSHRNu; break;
3863 case Intrinsic::arm_neon_vqshiftnsu:
3864 VShiftOpc = ARMISD::VQSHRNsu; break;
3865 case Intrinsic::arm_neon_vqrshiftns:
3866 VShiftOpc = ARMISD::VQRSHRNs; break;
3867 case Intrinsic::arm_neon_vqrshiftnu:
3868 VShiftOpc = ARMISD::VQRSHRNu; break;
3869 case Intrinsic::arm_neon_vqrshiftnsu:
3870 VShiftOpc = ARMISD::VQRSHRNsu; break;
3873 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3874 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3877 case Intrinsic::arm_neon_vshiftins: {
3878 EVT VT = N->getOperand(1).getValueType();
3880 unsigned VShiftOpc = 0;
3882 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3883 VShiftOpc = ARMISD::VSLI;
3884 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3885 VShiftOpc = ARMISD::VSRI;
3887 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3890 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3891 N->getOperand(1), N->getOperand(2),
3892 DAG.getConstant(Cnt, MVT::i32));
3895 case Intrinsic::arm_neon_vqrshifts:
3896 case Intrinsic::arm_neon_vqrshiftu:
3897 // No immediate versions of these to check for.
3904 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3905 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3906 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3907 /// vector element shift counts are generally not legal, and it is hard to see
3908 /// their values after they get legalized to loads from a constant pool.
3909 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3910 const ARMSubtarget *ST) {
3911 EVT VT = N->getValueType(0);
3913 // Nothing to be done for scalar shifts.
3914 if (! VT.isVector())
3917 assert(ST->hasNEON() && "unexpected vector shift");
3920 switch (N->getOpcode()) {
3921 default: llvm_unreachable("unexpected shift opcode");
3924 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3925 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3926 DAG.getConstant(Cnt, MVT::i32));
3931 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3932 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3933 ARMISD::VSHRs : ARMISD::VSHRu);
3934 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3935 DAG.getConstant(Cnt, MVT::i32));
3941 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3942 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3943 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3944 const ARMSubtarget *ST) {
3945 SDValue N0 = N->getOperand(0);
3947 // Check for sign- and zero-extensions of vector extract operations of 8-
3948 // and 16-bit vector elements. NEON supports these directly. They are
3949 // handled during DAG combining because type legalization will promote them
3950 // to 32-bit types and it is messy to recognize the operations after that.
3951 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3952 SDValue Vec = N0.getOperand(0);
3953 SDValue Lane = N0.getOperand(1);
3954 EVT VT = N->getValueType(0);
3955 EVT EltVT = N0.getValueType();
3956 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3958 if (VT == MVT::i32 &&
3959 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3960 TLI.isTypeLegal(Vec.getValueType())) {
3963 switch (N->getOpcode()) {
3964 default: llvm_unreachable("unexpected opcode");
3965 case ISD::SIGN_EXTEND:
3966 Opc = ARMISD::VGETLANEs;
3968 case ISD::ZERO_EXTEND:
3969 case ISD::ANY_EXTEND:
3970 Opc = ARMISD::VGETLANEu;
3973 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3980 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3981 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3982 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3983 const ARMSubtarget *ST) {
3984 // If the target supports NEON, try to use vmax/vmin instructions for f32
3985 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3986 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3987 // a NaN; only do the transformation when it matches that behavior.
3989 // For now only do this when using NEON for FP operations; if using VFP, it
3990 // is not obvious that the benefit outweighs the cost of switching to the
3992 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3993 N->getValueType(0) != MVT::f32)
3996 SDValue CondLHS = N->getOperand(0);
3997 SDValue CondRHS = N->getOperand(1);
3998 SDValue LHS = N->getOperand(2);
3999 SDValue RHS = N->getOperand(3);
4000 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4002 unsigned Opcode = 0;
4004 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4005 IsReversed = false; // x CC y ? x : y
4006 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4007 IsReversed = true ; // x CC y ? y : x
4021 // If LHS is NaN, an ordered comparison will be false and the result will
4022 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4023 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4024 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4025 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4027 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4028 // will return -0, so vmin can only be used for unsafe math or if one of
4029 // the operands is known to be nonzero.
4030 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4032 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4034 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4043 // If LHS is NaN, an ordered comparison will be false and the result will
4044 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4045 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4046 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4047 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4049 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4050 // will return +0, so vmax can only be used for unsafe math or if one of
4051 // the operands is known to be nonzero.
4052 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4054 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4056 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4062 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4065 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4066 DAGCombinerInfo &DCI) const {
4067 switch (N->getOpcode()) {
4069 case ISD::ADD: return PerformADDCombine(N, DCI);
4070 case ISD::SUB: return PerformSUBCombine(N, DCI);
4071 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4072 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4075 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4076 case ISD::SIGN_EXTEND:
4077 case ISD::ZERO_EXTEND:
4078 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4079 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4084 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4085 if (!Subtarget->hasV6Ops())
4086 // Pre-v6 does not support unaligned mem access.
4089 // v6+ may or may not support unaligned mem access depending on the system
4091 // FIXME: This is pretty conservative. Should we provide cmdline option to
4092 // control the behaviour?
4093 if (!Subtarget->isTargetDarwin())
4097 switch (VT.getSimpleVT().SimpleTy) {
4104 // FIXME: VLD1 etc with standard alignment is legal.
4108 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4113 switch (VT.getSimpleVT().SimpleTy) {
4114 default: return false;
4129 if ((V & (Scale - 1)) != 0)
4132 return V == (V & ((1LL << 5) - 1));
4135 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4136 const ARMSubtarget *Subtarget) {
4143 switch (VT.getSimpleVT().SimpleTy) {
4144 default: return false;
4149 // + imm12 or - imm8
4151 return V == (V & ((1LL << 8) - 1));
4152 return V == (V & ((1LL << 12) - 1));
4155 // Same as ARM mode. FIXME: NEON?
4156 if (!Subtarget->hasVFP2())
4161 return V == (V & ((1LL << 8) - 1));
4165 /// isLegalAddressImmediate - Return true if the integer value can be used
4166 /// as the offset of the target addressing mode for load / store of the
4168 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4169 const ARMSubtarget *Subtarget) {
4176 if (Subtarget->isThumb1Only())
4177 return isLegalT1AddressImmediate(V, VT);
4178 else if (Subtarget->isThumb2())
4179 return isLegalT2AddressImmediate(V, VT, Subtarget);
4184 switch (VT.getSimpleVT().SimpleTy) {
4185 default: return false;
4190 return V == (V & ((1LL << 12) - 1));
4193 return V == (V & ((1LL << 8) - 1));
4196 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4201 return V == (V & ((1LL << 8) - 1));
4205 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4207 int Scale = AM.Scale;
4211 switch (VT.getSimpleVT().SimpleTy) {
4212 default: return false;
4221 return Scale == 2 || Scale == 4 || Scale == 8;
4224 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4228 // Note, we allow "void" uses (basically, uses that aren't loads or
4229 // stores), because arm allows folding a scale into many arithmetic
4230 // operations. This should be made more precise and revisited later.
4232 // Allow r << imm, but the imm has to be a multiple of two.
4233 if (Scale & 1) return false;
4234 return isPowerOf2_32(Scale);
4238 /// isLegalAddressingMode - Return true if the addressing mode represented
4239 /// by AM is legal for this target, for a load/store of the specified type.
4240 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4241 const Type *Ty) const {
4242 EVT VT = getValueType(Ty, true);
4243 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4246 // Can never fold addr of global into load/store.
4251 case 0: // no scale reg, must be "r+i" or "r", or "i".
4254 if (Subtarget->isThumb1Only())
4258 // ARM doesn't support any R+R*scale+imm addr modes.
4265 if (Subtarget->isThumb2())
4266 return isLegalT2ScaledAddressingMode(AM, VT);
4268 int Scale = AM.Scale;
4269 switch (VT.getSimpleVT().SimpleTy) {
4270 default: return false;
4274 if (Scale < 0) Scale = -Scale;
4278 return isPowerOf2_32(Scale & ~1);
4282 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4287 // Note, we allow "void" uses (basically, uses that aren't loads or
4288 // stores), because arm allows folding a scale into many arithmetic
4289 // operations. This should be made more precise and revisited later.
4291 // Allow r << imm, but the imm has to be a multiple of two.
4292 if (Scale & 1) return false;
4293 return isPowerOf2_32(Scale);
4300 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4301 /// icmp immediate, that is the target has icmp instructions which can compare
4302 /// a register against the immediate without having to materialize the
4303 /// immediate into a register.
4304 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4305 if (!Subtarget->isThumb())
4306 return ARM_AM::getSOImmVal(Imm) != -1;
4307 if (Subtarget->isThumb2())
4308 return ARM_AM::getT2SOImmVal(Imm) != -1;
4309 return Imm >= 0 && Imm <= 255;
4312 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4313 bool isSEXTLoad, SDValue &Base,
4314 SDValue &Offset, bool &isInc,
4315 SelectionDAG &DAG) {
4316 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4319 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4321 Base = Ptr->getOperand(0);
4322 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4323 int RHSC = (int)RHS->getZExtValue();
4324 if (RHSC < 0 && RHSC > -256) {
4325 assert(Ptr->getOpcode() == ISD::ADD);
4327 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4331 isInc = (Ptr->getOpcode() == ISD::ADD);
4332 Offset = Ptr->getOperand(1);
4334 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4336 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4337 int RHSC = (int)RHS->getZExtValue();
4338 if (RHSC < 0 && RHSC > -0x1000) {
4339 assert(Ptr->getOpcode() == ISD::ADD);
4341 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4342 Base = Ptr->getOperand(0);
4347 if (Ptr->getOpcode() == ISD::ADD) {
4349 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4350 if (ShOpcVal != ARM_AM::no_shift) {
4351 Base = Ptr->getOperand(1);
4352 Offset = Ptr->getOperand(0);
4354 Base = Ptr->getOperand(0);
4355 Offset = Ptr->getOperand(1);
4360 isInc = (Ptr->getOpcode() == ISD::ADD);
4361 Base = Ptr->getOperand(0);
4362 Offset = Ptr->getOperand(1);
4366 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4370 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4371 bool isSEXTLoad, SDValue &Base,
4372 SDValue &Offset, bool &isInc,
4373 SelectionDAG &DAG) {
4374 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4377 Base = Ptr->getOperand(0);
4378 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4379 int RHSC = (int)RHS->getZExtValue();
4380 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4381 assert(Ptr->getOpcode() == ISD::ADD);
4383 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4385 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4386 isInc = Ptr->getOpcode() == ISD::ADD;
4387 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4395 /// getPreIndexedAddressParts - returns true by value, base pointer and
4396 /// offset pointer and addressing mode by reference if the node's address
4397 /// can be legally represented as pre-indexed load / store address.
4399 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4401 ISD::MemIndexedMode &AM,
4402 SelectionDAG &DAG) const {
4403 if (Subtarget->isThumb1Only())
4408 bool isSEXTLoad = false;
4409 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4410 Ptr = LD->getBasePtr();
4411 VT = LD->getMemoryVT();
4412 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4413 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4414 Ptr = ST->getBasePtr();
4415 VT = ST->getMemoryVT();
4420 bool isLegal = false;
4421 if (Subtarget->isThumb2())
4422 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4423 Offset, isInc, DAG);
4425 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4426 Offset, isInc, DAG);
4430 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4434 /// getPostIndexedAddressParts - returns true by value, base pointer and
4435 /// offset pointer and addressing mode by reference if this node can be
4436 /// combined with a load / store to form a post-indexed load / store.
4437 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4440 ISD::MemIndexedMode &AM,
4441 SelectionDAG &DAG) const {
4442 if (Subtarget->isThumb1Only())
4447 bool isSEXTLoad = false;
4448 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4449 VT = LD->getMemoryVT();
4450 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4451 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4452 VT = ST->getMemoryVT();
4457 bool isLegal = false;
4458 if (Subtarget->isThumb2())
4459 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4462 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4467 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4471 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4475 const SelectionDAG &DAG,
4476 unsigned Depth) const {
4477 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4478 switch (Op.getOpcode()) {
4480 case ARMISD::CMOV: {
4481 // Bits are known zero/one if known on the LHS and RHS.
4482 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4483 if (KnownZero == 0 && KnownOne == 0) return;
4485 APInt KnownZeroRHS, KnownOneRHS;
4486 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4487 KnownZeroRHS, KnownOneRHS, Depth+1);
4488 KnownZero &= KnownZeroRHS;
4489 KnownOne &= KnownOneRHS;
4495 //===----------------------------------------------------------------------===//
4496 // ARM Inline Assembly Support
4497 //===----------------------------------------------------------------------===//
4499 /// getConstraintType - Given a constraint letter, return the type of
4500 /// constraint it is for this target.
4501 ARMTargetLowering::ConstraintType
4502 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4503 if (Constraint.size() == 1) {
4504 switch (Constraint[0]) {
4506 case 'l': return C_RegisterClass;
4507 case 'w': return C_RegisterClass;
4510 return TargetLowering::getConstraintType(Constraint);
4513 std::pair<unsigned, const TargetRegisterClass*>
4514 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4516 if (Constraint.size() == 1) {
4517 // GCC ARM Constraint Letters
4518 switch (Constraint[0]) {
4520 if (Subtarget->isThumb())
4521 return std::make_pair(0U, ARM::tGPRRegisterClass);
4523 return std::make_pair(0U, ARM::GPRRegisterClass);
4525 return std::make_pair(0U, ARM::GPRRegisterClass);
4528 return std::make_pair(0U, ARM::SPRRegisterClass);
4529 if (VT.getSizeInBits() == 64)
4530 return std::make_pair(0U, ARM::DPRRegisterClass);
4531 if (VT.getSizeInBits() == 128)
4532 return std::make_pair(0U, ARM::QPRRegisterClass);
4536 if (StringRef("{cc}").equals_lower(Constraint))
4537 return std::make_pair(0U, ARM::CCRRegisterClass);
4539 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4542 std::vector<unsigned> ARMTargetLowering::
4543 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4545 if (Constraint.size() != 1)
4546 return std::vector<unsigned>();
4548 switch (Constraint[0]) { // GCC ARM Constraint Letters
4551 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4552 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4555 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4556 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4557 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4558 ARM::R12, ARM::LR, 0);
4561 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4562 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4563 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4564 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4565 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4566 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4567 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4568 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4569 if (VT.getSizeInBits() == 64)
4570 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4571 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4572 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4573 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4574 if (VT.getSizeInBits() == 128)
4575 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4576 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4580 return std::vector<unsigned>();
4583 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4584 /// vector. If it is invalid, don't add anything to Ops.
4585 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4588 std::vector<SDValue>&Ops,
4589 SelectionDAG &DAG) const {
4590 SDValue Result(0, 0);
4592 switch (Constraint) {
4594 case 'I': case 'J': case 'K': case 'L':
4595 case 'M': case 'N': case 'O':
4596 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4600 int64_t CVal64 = C->getSExtValue();
4601 int CVal = (int) CVal64;
4602 // None of these constraints allow values larger than 32 bits. Check
4603 // that the value fits in an int.
4607 switch (Constraint) {
4609 if (Subtarget->isThumb1Only()) {
4610 // This must be a constant between 0 and 255, for ADD
4612 if (CVal >= 0 && CVal <= 255)
4614 } else if (Subtarget->isThumb2()) {
4615 // A constant that can be used as an immediate value in a
4616 // data-processing instruction.
4617 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4620 // A constant that can be used as an immediate value in a
4621 // data-processing instruction.
4622 if (ARM_AM::getSOImmVal(CVal) != -1)
4628 if (Subtarget->isThumb()) { // FIXME thumb2
4629 // This must be a constant between -255 and -1, for negated ADD
4630 // immediates. This can be used in GCC with an "n" modifier that
4631 // prints the negated value, for use with SUB instructions. It is
4632 // not useful otherwise but is implemented for compatibility.
4633 if (CVal >= -255 && CVal <= -1)
4636 // This must be a constant between -4095 and 4095. It is not clear
4637 // what this constraint is intended for. Implemented for
4638 // compatibility with GCC.
4639 if (CVal >= -4095 && CVal <= 4095)
4645 if (Subtarget->isThumb1Only()) {
4646 // A 32-bit value where only one byte has a nonzero value. Exclude
4647 // zero to match GCC. This constraint is used by GCC internally for
4648 // constants that can be loaded with a move/shift combination.
4649 // It is not useful otherwise but is implemented for compatibility.
4650 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4652 } else if (Subtarget->isThumb2()) {
4653 // A constant whose bitwise inverse can be used as an immediate
4654 // value in a data-processing instruction. This can be used in GCC
4655 // with a "B" modifier that prints the inverted value, for use with
4656 // BIC and MVN instructions. It is not useful otherwise but is
4657 // implemented for compatibility.
4658 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4661 // A constant whose bitwise inverse can be used as an immediate
4662 // value in a data-processing instruction. This can be used in GCC
4663 // with a "B" modifier that prints the inverted value, for use with
4664 // BIC and MVN instructions. It is not useful otherwise but is
4665 // implemented for compatibility.
4666 if (ARM_AM::getSOImmVal(~CVal) != -1)
4672 if (Subtarget->isThumb1Only()) {
4673 // This must be a constant between -7 and 7,
4674 // for 3-operand ADD/SUB immediate instructions.
4675 if (CVal >= -7 && CVal < 7)
4677 } else if (Subtarget->isThumb2()) {
4678 // A constant whose negation can be used as an immediate value in a
4679 // data-processing instruction. This can be used in GCC with an "n"
4680 // modifier that prints the negated value, for use with SUB
4681 // instructions. It is not useful otherwise but is implemented for
4683 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4686 // A constant whose negation can be used as an immediate value in a
4687 // data-processing instruction. This can be used in GCC with an "n"
4688 // modifier that prints the negated value, for use with SUB
4689 // instructions. It is not useful otherwise but is implemented for
4691 if (ARM_AM::getSOImmVal(-CVal) != -1)
4697 if (Subtarget->isThumb()) { // FIXME thumb2
4698 // This must be a multiple of 4 between 0 and 1020, for
4699 // ADD sp + immediate.
4700 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4703 // A power of two or a constant between 0 and 32. This is used in
4704 // GCC for the shift amount on shifted register operands, but it is
4705 // useful in general for any shift amounts.
4706 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4712 if (Subtarget->isThumb()) { // FIXME thumb2
4713 // This must be a constant between 0 and 31, for shift amounts.
4714 if (CVal >= 0 && CVal <= 31)
4720 if (Subtarget->isThumb()) { // FIXME thumb2
4721 // This must be a multiple of 4 between -508 and 508, for
4722 // ADD/SUB sp = sp + immediate.
4723 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4728 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4732 if (Result.getNode()) {
4733 Ops.push_back(Result);
4736 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4741 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4742 // The ARM target isn't yet aware of offsets.
4746 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4747 APInt Imm = FPImm.bitcastToAPInt();
4748 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4749 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4750 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4752 // We can handle 4 bits of mantissa.
4753 // mantissa = (16+UInt(e:f:g:h))/16.
4754 if (Mantissa & 0x7ffff)
4757 if ((Mantissa & 0xf) != Mantissa)
4760 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4761 if (Exp < -3 || Exp > 4)
4763 Exp = ((Exp+3) & 0x7) ^ 4;
4765 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4768 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4769 APInt Imm = FPImm.bitcastToAPInt();
4770 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4771 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4772 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4774 // We can handle 4 bits of mantissa.
4775 // mantissa = (16+UInt(e:f:g:h))/16.
4776 if (Mantissa & 0xffffffffffffLL)
4779 if ((Mantissa & 0xf) != Mantissa)
4782 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4783 if (Exp < -3 || Exp > 4)
4785 Exp = ((Exp+3) & 0x7) ^ 4;
4787 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4790 /// isFPImmLegal - Returns true if the target can instruction select the
4791 /// specified FP immediate natively. If false, the legalizer will
4792 /// materialize the FP immediate as a load from a constant pool.
4793 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4794 if (!Subtarget->hasVFP3())
4797 return ARM::getVFPf32Imm(Imm) != -1;
4799 return ARM::getVFPf64Imm(Imm) != -1;