1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/Instruction.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Type.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/MC/MCSectionMachO.h"
41 #include "llvm/Target/TargetOptions.h"
42 #include "llvm/ADT/VectorExtras.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/MathExtras.h"
46 #include "llvm/Support/raw_ostream.h"
51 EnableARMLongCalls("arm-long-calls", cl::Hidden,
52 cl::desc("Generate calls via indirect call instructions."),
55 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
56 CCValAssign::LocInfo &LocInfo,
57 ISD::ArgFlagsTy &ArgFlags,
59 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
60 CCValAssign::LocInfo &LocInfo,
61 ISD::ArgFlagsTy &ArgFlags,
63 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
64 CCValAssign::LocInfo &LocInfo,
65 ISD::ArgFlagsTy &ArgFlags,
67 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
68 CCValAssign::LocInfo &LocInfo,
69 ISD::ArgFlagsTy &ArgFlags,
72 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
73 EVT PromotedBitwiseVT) {
74 if (VT != PromotedLdStVT) {
75 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
76 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
77 PromotedLdStVT.getSimpleVT());
79 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
80 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
81 PromotedLdStVT.getSimpleVT());
84 EVT ElemTy = VT.getVectorElementType();
85 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
86 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
87 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
88 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
89 if (ElemTy != MVT::i32) {
90 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
91 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
95 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
97 if (llvm::ModelWithRegSequence())
98 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
100 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
101 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
102 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
103 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
104 if (VT.isInteger()) {
105 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
106 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
107 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
110 // Promote all bit-wise operations.
111 if (VT.isInteger() && VT != PromotedBitwiseVT) {
112 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
113 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
114 PromotedBitwiseVT.getSimpleVT());
115 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
116 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
117 PromotedBitwiseVT.getSimpleVT());
118 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
119 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
120 PromotedBitwiseVT.getSimpleVT());
123 // Neon does not support vector divide/remainder operations.
124 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
125 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
128 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
129 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
132 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
133 addRegisterClass(VT, ARM::DPRRegisterClass);
134 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
137 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
138 addRegisterClass(VT, ARM::QPRRegisterClass);
139 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
142 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
143 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
144 return new TargetLoweringObjectFileMachO();
146 return new ARMElfTargetObjectFile();
149 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
150 : TargetLowering(TM, createTLOF(TM)) {
151 Subtarget = &TM.getSubtarget<ARMSubtarget>();
153 if (Subtarget->isTargetDarwin()) {
154 // Uses VFP for Thumb libfuncs if available.
155 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
156 // Single-precision floating-point arithmetic.
157 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
158 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
159 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
160 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
162 // Double-precision floating-point arithmetic.
163 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
164 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
165 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
166 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
168 // Single-precision comparisons.
169 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
170 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
171 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
172 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
173 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
174 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
175 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
176 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
178 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
179 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
180 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
181 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
182 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
183 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
187 // Double-precision comparisons.
188 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
189 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
190 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
191 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
192 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
193 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
194 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
195 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
197 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
206 // Floating-point to integer conversions.
207 // i64 conversions are done via library routines even when generating VFP
208 // instructions, so use the same ones.
209 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
210 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
211 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
212 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
214 // Conversions between floating types.
215 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
216 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
218 // Integer to floating-point conversions.
219 // i64 conversions are done via library routines even when generating VFP
220 // instructions, so use the same ones.
221 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
222 // e.g., __floatunsidf vs. __floatunssidfvfp.
223 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
224 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
225 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
226 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
230 // These libcalls are not available in 32-bit.
231 setLibcallName(RTLIB::SHL_I128, 0);
232 setLibcallName(RTLIB::SRL_I128, 0);
233 setLibcallName(RTLIB::SRA_I128, 0);
235 // Libcalls should use the AAPCS base standard ABI, even if hard float
236 // is in effect, as per the ARM RTABI specification, section 4.1.2.
237 if (Subtarget->isAAPCS_ABI()) {
238 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
239 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
240 CallingConv::ARM_AAPCS);
244 if (Subtarget->isThumb1Only())
245 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
247 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
248 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
249 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
250 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
252 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
255 if (Subtarget->hasNEON()) {
256 addDRTypeForNEON(MVT::v2f32);
257 addDRTypeForNEON(MVT::v8i8);
258 addDRTypeForNEON(MVT::v4i16);
259 addDRTypeForNEON(MVT::v2i32);
260 addDRTypeForNEON(MVT::v1i64);
262 addQRTypeForNEON(MVT::v4f32);
263 addQRTypeForNEON(MVT::v2f64);
264 addQRTypeForNEON(MVT::v16i8);
265 addQRTypeForNEON(MVT::v8i16);
266 addQRTypeForNEON(MVT::v4i32);
267 addQRTypeForNEON(MVT::v2i64);
269 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
270 // neither Neon nor VFP support any arithmetic operations on it.
271 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
272 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
273 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
274 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
275 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
276 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
277 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
278 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
279 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
280 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
281 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
282 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
283 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
284 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
285 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
286 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
287 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
288 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
289 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
290 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
291 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
292 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
293 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
294 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
296 // Neon does not support some operations on v1i64 and v2i64 types.
297 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
298 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
299 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
300 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
302 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
303 setTargetDAGCombine(ISD::SHL);
304 setTargetDAGCombine(ISD::SRL);
305 setTargetDAGCombine(ISD::SRA);
306 setTargetDAGCombine(ISD::SIGN_EXTEND);
307 setTargetDAGCombine(ISD::ZERO_EXTEND);
308 setTargetDAGCombine(ISD::ANY_EXTEND);
309 setTargetDAGCombine(ISD::SELECT_CC);
312 computeRegisterProperties();
314 // ARM does not have f32 extending load.
315 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
317 // ARM does not have i1 sign extending load.
318 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
320 // ARM supports all 4 flavors of integer indexed load / store.
321 if (!Subtarget->isThumb1Only()) {
322 for (unsigned im = (unsigned)ISD::PRE_INC;
323 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
324 setIndexedLoadAction(im, MVT::i1, Legal);
325 setIndexedLoadAction(im, MVT::i8, Legal);
326 setIndexedLoadAction(im, MVT::i16, Legal);
327 setIndexedLoadAction(im, MVT::i32, Legal);
328 setIndexedStoreAction(im, MVT::i1, Legal);
329 setIndexedStoreAction(im, MVT::i8, Legal);
330 setIndexedStoreAction(im, MVT::i16, Legal);
331 setIndexedStoreAction(im, MVT::i32, Legal);
335 // i64 operation support.
336 if (Subtarget->isThumb1Only()) {
337 setOperationAction(ISD::MUL, MVT::i64, Expand);
338 setOperationAction(ISD::MULHU, MVT::i32, Expand);
339 setOperationAction(ISD::MULHS, MVT::i32, Expand);
340 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
341 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
343 setOperationAction(ISD::MUL, MVT::i64, Expand);
344 setOperationAction(ISD::MULHU, MVT::i32, Expand);
345 if (!Subtarget->hasV6Ops())
346 setOperationAction(ISD::MULHS, MVT::i32, Expand);
348 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
349 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
350 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
351 setOperationAction(ISD::SRL, MVT::i64, Custom);
352 setOperationAction(ISD::SRA, MVT::i64, Custom);
354 // ARM does not have ROTL.
355 setOperationAction(ISD::ROTL, MVT::i32, Expand);
356 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
357 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
358 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
359 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
361 // Only ARMv6 has BSWAP.
362 if (!Subtarget->hasV6Ops())
363 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
365 // These are expanded into libcalls.
366 if (!Subtarget->hasDivide()) {
367 // v7M has a hardware divider
368 setOperationAction(ISD::SDIV, MVT::i32, Expand);
369 setOperationAction(ISD::UDIV, MVT::i32, Expand);
371 setOperationAction(ISD::SREM, MVT::i32, Expand);
372 setOperationAction(ISD::UREM, MVT::i32, Expand);
373 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
374 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
376 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
377 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
378 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
379 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
380 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
382 // Use the default implementation.
383 setOperationAction(ISD::VASTART, MVT::Other, Custom);
384 setOperationAction(ISD::VAARG, MVT::Other, Expand);
385 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
386 setOperationAction(ISD::VAEND, MVT::Other, Expand);
387 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
388 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
389 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
390 // FIXME: Shouldn't need this, since no register is used, but the legalizer
391 // doesn't yet know how to not do that for SjLj.
392 setExceptionSelectorRegister(ARM::R0);
393 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
394 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
396 // If the subtarget does not have extract instructions, sign_extend_inreg
397 // needs to be expanded. Extract is available in ARM mode on v6 and up,
398 // and on most Thumb2 implementations.
399 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
400 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
401 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
402 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
404 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
406 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
407 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
408 // iff target supports vfp2.
409 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
411 // We want to custom lower some of our intrinsics.
412 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
414 setOperationAction(ISD::SETCC, MVT::i32, Expand);
415 setOperationAction(ISD::SETCC, MVT::f32, Expand);
416 setOperationAction(ISD::SETCC, MVT::f64, Expand);
417 setOperationAction(ISD::SELECT, MVT::i32, Expand);
418 setOperationAction(ISD::SELECT, MVT::f32, Expand);
419 setOperationAction(ISD::SELECT, MVT::f64, Expand);
420 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
421 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
422 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
424 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
425 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
426 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
427 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
428 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
430 // We don't support sin/cos/fmod/copysign/pow
431 setOperationAction(ISD::FSIN, MVT::f64, Expand);
432 setOperationAction(ISD::FSIN, MVT::f32, Expand);
433 setOperationAction(ISD::FCOS, MVT::f32, Expand);
434 setOperationAction(ISD::FCOS, MVT::f64, Expand);
435 setOperationAction(ISD::FREM, MVT::f64, Expand);
436 setOperationAction(ISD::FREM, MVT::f32, Expand);
437 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
438 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
439 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
441 setOperationAction(ISD::FPOW, MVT::f64, Expand);
442 setOperationAction(ISD::FPOW, MVT::f32, Expand);
444 // Various VFP goodness
445 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
446 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
447 if (Subtarget->hasVFP2()) {
448 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
449 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
450 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
451 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
453 // Special handling for half-precision FP.
454 if (!Subtarget->hasFP16()) {
455 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
456 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
460 // We have target-specific dag combine patterns for the following nodes:
461 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
462 setTargetDAGCombine(ISD::ADD);
463 setTargetDAGCombine(ISD::SUB);
465 setStackPointerRegisterToSaveRestore(ARM::SP);
466 setSchedulingPreference(SchedulingForRegPressure);
468 // FIXME: If-converter should use instruction latency to determine
469 // profitability rather than relying on fixed limits.
470 if (Subtarget->getCPUString() == "generic") {
471 // Generic (and overly aggressive) if-conversion limits.
472 setIfCvtBlockSizeLimit(10);
473 setIfCvtDupBlockSizeLimit(2);
474 } else if (Subtarget->hasV7Ops()) {
475 setIfCvtBlockSizeLimit(3);
476 setIfCvtDupBlockSizeLimit(1);
477 } else if (Subtarget->hasV6Ops()) {
478 setIfCvtBlockSizeLimit(2);
479 setIfCvtDupBlockSizeLimit(1);
481 setIfCvtBlockSizeLimit(3);
482 setIfCvtDupBlockSizeLimit(2);
485 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
486 // Do not enable CodePlacementOpt for now: it currently runs after the
487 // ARMConstantIslandPass and messes up branch relaxation and placement
488 // of constant islands.
489 // benefitFromCodePlacementOpt = true;
492 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
495 case ARMISD::Wrapper: return "ARMISD::Wrapper";
496 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
497 case ARMISD::CALL: return "ARMISD::CALL";
498 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
499 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
500 case ARMISD::tCALL: return "ARMISD::tCALL";
501 case ARMISD::BRCOND: return "ARMISD::BRCOND";
502 case ARMISD::BR_JT: return "ARMISD::BR_JT";
503 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
504 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
505 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
506 case ARMISD::CMP: return "ARMISD::CMP";
507 case ARMISD::CMPZ: return "ARMISD::CMPZ";
508 case ARMISD::CMPFP: return "ARMISD::CMPFP";
509 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
510 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
511 case ARMISD::CMOV: return "ARMISD::CMOV";
512 case ARMISD::CNEG: return "ARMISD::CNEG";
514 case ARMISD::RBIT: return "ARMISD::RBIT";
516 case ARMISD::FTOSI: return "ARMISD::FTOSI";
517 case ARMISD::FTOUI: return "ARMISD::FTOUI";
518 case ARMISD::SITOF: return "ARMISD::SITOF";
519 case ARMISD::UITOF: return "ARMISD::UITOF";
521 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
522 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
523 case ARMISD::RRX: return "ARMISD::RRX";
525 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
526 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
528 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
529 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
531 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
533 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
535 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
536 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
538 case ARMISD::VCEQ: return "ARMISD::VCEQ";
539 case ARMISD::VCGE: return "ARMISD::VCGE";
540 case ARMISD::VCGEU: return "ARMISD::VCGEU";
541 case ARMISD::VCGT: return "ARMISD::VCGT";
542 case ARMISD::VCGTU: return "ARMISD::VCGTU";
543 case ARMISD::VTST: return "ARMISD::VTST";
545 case ARMISD::VSHL: return "ARMISD::VSHL";
546 case ARMISD::VSHRs: return "ARMISD::VSHRs";
547 case ARMISD::VSHRu: return "ARMISD::VSHRu";
548 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
549 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
550 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
551 case ARMISD::VSHRN: return "ARMISD::VSHRN";
552 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
553 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
554 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
555 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
556 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
557 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
558 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
559 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
560 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
561 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
562 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
563 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
564 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
565 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
566 case ARMISD::VDUP: return "ARMISD::VDUP";
567 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
568 case ARMISD::VEXT: return "ARMISD::VEXT";
569 case ARMISD::VREV64: return "ARMISD::VREV64";
570 case ARMISD::VREV32: return "ARMISD::VREV32";
571 case ARMISD::VREV16: return "ARMISD::VREV16";
572 case ARMISD::VZIP: return "ARMISD::VZIP";
573 case ARMISD::VUZP: return "ARMISD::VUZP";
574 case ARMISD::VTRN: return "ARMISD::VTRN";
575 case ARMISD::FMAX: return "ARMISD::FMAX";
576 case ARMISD::FMIN: return "ARMISD::FMIN";
580 /// getFunctionAlignment - Return the Log2 alignment of this function.
581 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
582 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
585 //===----------------------------------------------------------------------===//
587 //===----------------------------------------------------------------------===//
589 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
590 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
592 default: llvm_unreachable("Unknown condition code!");
593 case ISD::SETNE: return ARMCC::NE;
594 case ISD::SETEQ: return ARMCC::EQ;
595 case ISD::SETGT: return ARMCC::GT;
596 case ISD::SETGE: return ARMCC::GE;
597 case ISD::SETLT: return ARMCC::LT;
598 case ISD::SETLE: return ARMCC::LE;
599 case ISD::SETUGT: return ARMCC::HI;
600 case ISD::SETUGE: return ARMCC::HS;
601 case ISD::SETULT: return ARMCC::LO;
602 case ISD::SETULE: return ARMCC::LS;
606 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
607 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
608 ARMCC::CondCodes &CondCode2) {
609 CondCode2 = ARMCC::AL;
611 default: llvm_unreachable("Unknown FP condition!");
613 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
615 case ISD::SETOGT: CondCode = ARMCC::GT; break;
617 case ISD::SETOGE: CondCode = ARMCC::GE; break;
618 case ISD::SETOLT: CondCode = ARMCC::MI; break;
619 case ISD::SETOLE: CondCode = ARMCC::LS; break;
620 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
621 case ISD::SETO: CondCode = ARMCC::VC; break;
622 case ISD::SETUO: CondCode = ARMCC::VS; break;
623 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
624 case ISD::SETUGT: CondCode = ARMCC::HI; break;
625 case ISD::SETUGE: CondCode = ARMCC::PL; break;
627 case ISD::SETULT: CondCode = ARMCC::LT; break;
629 case ISD::SETULE: CondCode = ARMCC::LE; break;
631 case ISD::SETUNE: CondCode = ARMCC::NE; break;
635 //===----------------------------------------------------------------------===//
636 // Calling Convention Implementation
637 //===----------------------------------------------------------------------===//
639 #include "ARMGenCallingConv.inc"
641 // APCS f64 is in register pairs, possibly split to stack
642 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
643 CCValAssign::LocInfo &LocInfo,
644 CCState &State, bool CanFail) {
645 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
647 // Try to get the first register.
648 if (unsigned Reg = State.AllocateReg(RegList, 4))
649 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
651 // For the 2nd half of a v2f64, do not fail.
655 // Put the whole thing on the stack.
656 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
657 State.AllocateStack(8, 4),
662 // Try to get the second register.
663 if (unsigned Reg = State.AllocateReg(RegList, 4))
664 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
666 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
667 State.AllocateStack(4, 4),
672 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
673 CCValAssign::LocInfo &LocInfo,
674 ISD::ArgFlagsTy &ArgFlags,
676 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
678 if (LocVT == MVT::v2f64 &&
679 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
681 return true; // we handled it
684 // AAPCS f64 is in aligned register pairs
685 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
686 CCValAssign::LocInfo &LocInfo,
687 CCState &State, bool CanFail) {
688 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
689 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
691 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
693 // For the 2nd half of a v2f64, do not just fail.
697 // Put the whole thing on the stack.
698 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
699 State.AllocateStack(8, 8),
705 for (i = 0; i < 2; ++i)
706 if (HiRegList[i] == Reg)
709 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
710 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
715 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
716 CCValAssign::LocInfo &LocInfo,
717 ISD::ArgFlagsTy &ArgFlags,
719 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
721 if (LocVT == MVT::v2f64 &&
722 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
724 return true; // we handled it
727 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
728 CCValAssign::LocInfo &LocInfo, CCState &State) {
729 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
730 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
732 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
734 return false; // we didn't handle it
737 for (i = 0; i < 2; ++i)
738 if (HiRegList[i] == Reg)
741 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
742 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
747 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
748 CCValAssign::LocInfo &LocInfo,
749 ISD::ArgFlagsTy &ArgFlags,
751 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
753 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
755 return true; // we handled it
758 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
759 CCValAssign::LocInfo &LocInfo,
760 ISD::ArgFlagsTy &ArgFlags,
762 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
766 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
767 /// given CallingConvention value.
768 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
770 bool isVarArg) const {
773 llvm_unreachable("Unsupported calling convention");
775 case CallingConv::Fast:
776 // Use target triple & subtarget features to do actual dispatch.
777 if (Subtarget->isAAPCS_ABI()) {
778 if (Subtarget->hasVFP2() &&
779 FloatABIType == FloatABI::Hard && !isVarArg)
780 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
782 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
784 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
785 case CallingConv::ARM_AAPCS_VFP:
786 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
787 case CallingConv::ARM_AAPCS:
788 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
789 case CallingConv::ARM_APCS:
790 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
794 /// LowerCallResult - Lower the result values of a call into the
795 /// appropriate copies out of appropriate physical registers.
797 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
798 CallingConv::ID CallConv, bool isVarArg,
799 const SmallVectorImpl<ISD::InputArg> &Ins,
800 DebugLoc dl, SelectionDAG &DAG,
801 SmallVectorImpl<SDValue> &InVals) const {
803 // Assign locations to each value returned by this call.
804 SmallVector<CCValAssign, 16> RVLocs;
805 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
806 RVLocs, *DAG.getContext());
807 CCInfo.AnalyzeCallResult(Ins,
808 CCAssignFnForNode(CallConv, /* Return*/ true,
811 // Copy all of the result registers out of their specified physreg.
812 for (unsigned i = 0; i != RVLocs.size(); ++i) {
813 CCValAssign VA = RVLocs[i];
816 if (VA.needsCustom()) {
817 // Handle f64 or half of a v2f64.
818 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
820 Chain = Lo.getValue(1);
821 InFlag = Lo.getValue(2);
822 VA = RVLocs[++i]; // skip ahead to next loc
823 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
825 Chain = Hi.getValue(1);
826 InFlag = Hi.getValue(2);
827 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
829 if (VA.getLocVT() == MVT::v2f64) {
830 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
831 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
832 DAG.getConstant(0, MVT::i32));
834 VA = RVLocs[++i]; // skip ahead to next loc
835 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
836 Chain = Lo.getValue(1);
837 InFlag = Lo.getValue(2);
838 VA = RVLocs[++i]; // skip ahead to next loc
839 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
840 Chain = Hi.getValue(1);
841 InFlag = Hi.getValue(2);
842 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
843 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
844 DAG.getConstant(1, MVT::i32));
847 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
849 Chain = Val.getValue(1);
850 InFlag = Val.getValue(2);
853 switch (VA.getLocInfo()) {
854 default: llvm_unreachable("Unknown loc info!");
855 case CCValAssign::Full: break;
856 case CCValAssign::BCvt:
857 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
861 InVals.push_back(Val);
867 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
868 /// by "Src" to address "Dst" of size "Size". Alignment information is
869 /// specified by the specific parameter attribute. The copy will be passed as
870 /// a byval function parameter.
871 /// Sometimes what we are copying is the end of a larger object, the part that
872 /// does not fit in registers.
874 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
875 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
877 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
878 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
879 /*isVolatile=*/false, /*AlwaysInline=*/false,
883 /// LowerMemOpCallTo - Store the argument to the stack.
885 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
886 SDValue StackPtr, SDValue Arg,
887 DebugLoc dl, SelectionDAG &DAG,
888 const CCValAssign &VA,
889 ISD::ArgFlagsTy Flags) const {
890 unsigned LocMemOffset = VA.getLocMemOffset();
891 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
892 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
893 if (Flags.isByVal()) {
894 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
896 return DAG.getStore(Chain, dl, Arg, PtrOff,
897 PseudoSourceValue::getStack(), LocMemOffset,
901 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
902 SDValue Chain, SDValue &Arg,
903 RegsToPassVector &RegsToPass,
904 CCValAssign &VA, CCValAssign &NextVA,
906 SmallVector<SDValue, 8> &MemOpChains,
907 ISD::ArgFlagsTy Flags) const {
909 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
910 DAG.getVTList(MVT::i32, MVT::i32), Arg);
911 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
913 if (NextVA.isRegLoc())
914 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
916 assert(NextVA.isMemLoc());
917 if (StackPtr.getNode() == 0)
918 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
920 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
926 /// LowerCall - Lowering a call into a callseq_start <-
927 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
930 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
931 CallingConv::ID CallConv, bool isVarArg,
933 const SmallVectorImpl<ISD::OutputArg> &Outs,
934 const SmallVectorImpl<ISD::InputArg> &Ins,
935 DebugLoc dl, SelectionDAG &DAG,
936 SmallVectorImpl<SDValue> &InVals) const {
937 // ARM target does not yet support tail call optimization.
940 // Analyze operands of the call, assigning locations to each operand.
941 SmallVector<CCValAssign, 16> ArgLocs;
942 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
944 CCInfo.AnalyzeCallOperands(Outs,
945 CCAssignFnForNode(CallConv, /* Return*/ false,
948 // Get a count of how many bytes are to be pushed on the stack.
949 unsigned NumBytes = CCInfo.getNextStackOffset();
951 // Adjust the stack pointer for the new arguments...
952 // These operations are automatically eliminated by the prolog/epilog pass
953 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
955 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
957 RegsToPassVector RegsToPass;
958 SmallVector<SDValue, 8> MemOpChains;
960 // Walk the register/memloc assignments, inserting copies/loads. In the case
961 // of tail call optimization, arguments are handled later.
962 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
965 CCValAssign &VA = ArgLocs[i];
966 SDValue Arg = Outs[realArgIdx].Val;
967 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
969 // Promote the value if needed.
970 switch (VA.getLocInfo()) {
971 default: llvm_unreachable("Unknown loc info!");
972 case CCValAssign::Full: break;
973 case CCValAssign::SExt:
974 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
976 case CCValAssign::ZExt:
977 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
979 case CCValAssign::AExt:
980 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
982 case CCValAssign::BCvt:
983 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
987 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
988 if (VA.needsCustom()) {
989 if (VA.getLocVT() == MVT::v2f64) {
990 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
991 DAG.getConstant(0, MVT::i32));
992 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
993 DAG.getConstant(1, MVT::i32));
995 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
996 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
998 VA = ArgLocs[++i]; // skip ahead to next loc
1000 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1001 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1003 assert(VA.isMemLoc());
1005 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1006 dl, DAG, VA, Flags));
1009 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1010 StackPtr, MemOpChains, Flags);
1012 } else if (VA.isRegLoc()) {
1013 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1015 assert(VA.isMemLoc());
1017 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1018 dl, DAG, VA, Flags));
1022 if (!MemOpChains.empty())
1023 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1024 &MemOpChains[0], MemOpChains.size());
1026 // Build a sequence of copy-to-reg nodes chained together with token chain
1027 // and flag operands which copy the outgoing args into the appropriate regs.
1029 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1030 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1031 RegsToPass[i].second, InFlag);
1032 InFlag = Chain.getValue(1);
1035 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1036 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1037 // node so that legalize doesn't hack it.
1038 bool isDirect = false;
1039 bool isARMFunc = false;
1040 bool isLocalARMFunc = false;
1041 MachineFunction &MF = DAG.getMachineFunction();
1042 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1044 if (EnableARMLongCalls) {
1045 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1046 && "long-calls with non-static relocation model!");
1047 // Handle a global address or an external symbol. If it's not one of
1048 // those, the target's already in a register, so we don't need to do
1050 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1051 const GlobalValue *GV = G->getGlobal();
1052 // Create a constant pool entry for the callee address
1053 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1054 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1057 // Get the address of the callee into a register
1058 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1059 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1060 Callee = DAG.getLoad(getPointerTy(), dl,
1061 DAG.getEntryNode(), CPAddr,
1062 PseudoSourceValue::getConstantPool(), 0,
1064 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1065 const char *Sym = S->getSymbol();
1067 // Create a constant pool entry for the callee address
1068 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1069 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1070 Sym, ARMPCLabelIndex, 0);
1071 // Get the address of the callee into a register
1072 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1073 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1074 Callee = DAG.getLoad(getPointerTy(), dl,
1075 DAG.getEntryNode(), CPAddr,
1076 PseudoSourceValue::getConstantPool(), 0,
1079 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1080 const GlobalValue *GV = G->getGlobal();
1082 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1083 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1084 getTargetMachine().getRelocationModel() != Reloc::Static;
1085 isARMFunc = !Subtarget->isThumb() || isStub;
1086 // ARM call to a local ARM function is predicable.
1087 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1088 // tBX takes a register source operand.
1089 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1090 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1091 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1094 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1095 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1096 Callee = DAG.getLoad(getPointerTy(), dl,
1097 DAG.getEntryNode(), CPAddr,
1098 PseudoSourceValue::getConstantPool(), 0,
1100 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1101 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1102 getPointerTy(), Callee, PICLabel);
1104 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1105 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1107 bool isStub = Subtarget->isTargetDarwin() &&
1108 getTargetMachine().getRelocationModel() != Reloc::Static;
1109 isARMFunc = !Subtarget->isThumb() || isStub;
1110 // tBX takes a register source operand.
1111 const char *Sym = S->getSymbol();
1112 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1113 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1114 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1115 Sym, ARMPCLabelIndex, 4);
1116 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1117 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1118 Callee = DAG.getLoad(getPointerTy(), dl,
1119 DAG.getEntryNode(), CPAddr,
1120 PseudoSourceValue::getConstantPool(), 0,
1122 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1123 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1124 getPointerTy(), Callee, PICLabel);
1126 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1129 // FIXME: handle tail calls differently.
1131 if (Subtarget->isThumb()) {
1132 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1133 CallOpc = ARMISD::CALL_NOLINK;
1135 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1137 CallOpc = (isDirect || Subtarget->hasV5TOps())
1138 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1139 : ARMISD::CALL_NOLINK;
1141 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1142 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1143 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1144 InFlag = Chain.getValue(1);
1147 std::vector<SDValue> Ops;
1148 Ops.push_back(Chain);
1149 Ops.push_back(Callee);
1151 // Add argument registers to the end of the list so that they are known live
1153 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1154 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1155 RegsToPass[i].second.getValueType()));
1157 if (InFlag.getNode())
1158 Ops.push_back(InFlag);
1159 // Returns a chain and a flag for retval copy to use.
1160 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1161 &Ops[0], Ops.size());
1162 InFlag = Chain.getValue(1);
1164 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1165 DAG.getIntPtrConstant(0, true), InFlag);
1167 InFlag = Chain.getValue(1);
1169 // Handle result values, copying them out of physregs into vregs that we
1171 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1176 ARMTargetLowering::LowerReturn(SDValue Chain,
1177 CallingConv::ID CallConv, bool isVarArg,
1178 const SmallVectorImpl<ISD::OutputArg> &Outs,
1179 DebugLoc dl, SelectionDAG &DAG) const {
1181 // CCValAssign - represent the assignment of the return value to a location.
1182 SmallVector<CCValAssign, 16> RVLocs;
1184 // CCState - Info about the registers and stack slots.
1185 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1188 // Analyze outgoing return values.
1189 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1192 // If this is the first return lowered for this function, add
1193 // the regs to the liveout set for the function.
1194 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1195 for (unsigned i = 0; i != RVLocs.size(); ++i)
1196 if (RVLocs[i].isRegLoc())
1197 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1202 // Copy the result values into the output registers.
1203 for (unsigned i = 0, realRVLocIdx = 0;
1205 ++i, ++realRVLocIdx) {
1206 CCValAssign &VA = RVLocs[i];
1207 assert(VA.isRegLoc() && "Can only return in registers!");
1209 SDValue Arg = Outs[realRVLocIdx].Val;
1211 switch (VA.getLocInfo()) {
1212 default: llvm_unreachable("Unknown loc info!");
1213 case CCValAssign::Full: break;
1214 case CCValAssign::BCvt:
1215 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1219 if (VA.needsCustom()) {
1220 if (VA.getLocVT() == MVT::v2f64) {
1221 // Extract the first half and return it in two registers.
1222 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1223 DAG.getConstant(0, MVT::i32));
1224 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1225 DAG.getVTList(MVT::i32, MVT::i32), Half);
1227 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1228 Flag = Chain.getValue(1);
1229 VA = RVLocs[++i]; // skip ahead to next loc
1230 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1231 HalfGPRs.getValue(1), Flag);
1232 Flag = Chain.getValue(1);
1233 VA = RVLocs[++i]; // skip ahead to next loc
1235 // Extract the 2nd half and fall through to handle it as an f64 value.
1236 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1237 DAG.getConstant(1, MVT::i32));
1239 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1241 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1242 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1243 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1244 Flag = Chain.getValue(1);
1245 VA = RVLocs[++i]; // skip ahead to next loc
1246 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1249 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1251 // Guarantee that all emitted copies are
1252 // stuck together, avoiding something bad.
1253 Flag = Chain.getValue(1);
1258 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1260 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1265 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1266 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1267 // one of the above mentioned nodes. It has to be wrapped because otherwise
1268 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1269 // be used to form addressing mode. These wrapped nodes will be selected
1271 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1272 EVT PtrVT = Op.getValueType();
1273 // FIXME there is no actual debug info here
1274 DebugLoc dl = Op.getDebugLoc();
1275 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1277 if (CP->isMachineConstantPoolEntry())
1278 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1279 CP->getAlignment());
1281 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1282 CP->getAlignment());
1283 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1286 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1287 SelectionDAG &DAG) const {
1288 MachineFunction &MF = DAG.getMachineFunction();
1289 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1290 unsigned ARMPCLabelIndex = 0;
1291 DebugLoc DL = Op.getDebugLoc();
1292 EVT PtrVT = getPointerTy();
1293 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1294 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1296 if (RelocM == Reloc::Static) {
1297 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1299 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1300 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1301 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1302 ARMCP::CPBlockAddress,
1304 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1306 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1307 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1308 PseudoSourceValue::getConstantPool(), 0,
1310 if (RelocM == Reloc::Static)
1312 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1313 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1316 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1318 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1319 SelectionDAG &DAG) const {
1320 DebugLoc dl = GA->getDebugLoc();
1321 EVT PtrVT = getPointerTy();
1322 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1323 MachineFunction &MF = DAG.getMachineFunction();
1324 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1325 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1326 ARMConstantPoolValue *CPV =
1327 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1328 ARMCP::CPValue, PCAdj, "tlsgd", true);
1329 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1330 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1331 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1332 PseudoSourceValue::getConstantPool(), 0,
1334 SDValue Chain = Argument.getValue(1);
1336 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1337 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1339 // call __tls_get_addr.
1342 Entry.Node = Argument;
1343 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1344 Args.push_back(Entry);
1345 // FIXME: is there useful debug info available here?
1346 std::pair<SDValue, SDValue> CallResult =
1347 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1348 false, false, false, false,
1349 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1350 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1351 return CallResult.first;
1354 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1355 // "local exec" model.
1357 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1358 SelectionDAG &DAG) const {
1359 const GlobalValue *GV = GA->getGlobal();
1360 DebugLoc dl = GA->getDebugLoc();
1362 SDValue Chain = DAG.getEntryNode();
1363 EVT PtrVT = getPointerTy();
1364 // Get the Thread Pointer
1365 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1367 if (GV->isDeclaration()) {
1368 MachineFunction &MF = DAG.getMachineFunction();
1369 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1370 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1371 // Initial exec model.
1372 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1373 ARMConstantPoolValue *CPV =
1374 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1375 ARMCP::CPValue, PCAdj, "gottpoff", true);
1376 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1377 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1378 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1379 PseudoSourceValue::getConstantPool(), 0,
1381 Chain = Offset.getValue(1);
1383 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1384 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1386 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1387 PseudoSourceValue::getConstantPool(), 0,
1391 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1392 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1393 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1394 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1395 PseudoSourceValue::getConstantPool(), 0,
1399 // The address of the thread local variable is the add of the thread
1400 // pointer with the offset of the variable.
1401 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1405 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1406 // TODO: implement the "local dynamic" model
1407 assert(Subtarget->isTargetELF() &&
1408 "TLS not implemented for non-ELF targets");
1409 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1410 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1411 // otherwise use the "Local Exec" TLS Model
1412 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1413 return LowerToTLSGeneralDynamicModel(GA, DAG);
1415 return LowerToTLSExecModels(GA, DAG);
1418 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1419 SelectionDAG &DAG) const {
1420 EVT PtrVT = getPointerTy();
1421 DebugLoc dl = Op.getDebugLoc();
1422 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1423 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1424 if (RelocM == Reloc::PIC_) {
1425 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1426 ARMConstantPoolValue *CPV =
1427 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1428 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1429 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1430 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1432 PseudoSourceValue::getConstantPool(), 0,
1434 SDValue Chain = Result.getValue(1);
1435 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1436 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1438 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1439 PseudoSourceValue::getGOT(), 0,
1443 // If we have T2 ops, we can materialize the address directly via movt/movw
1444 // pair. This is always cheaper.
1445 if (Subtarget->useMovt()) {
1446 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1447 DAG.getTargetGlobalAddress(GV, PtrVT));
1449 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1450 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1451 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1452 PseudoSourceValue::getConstantPool(), 0,
1458 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1459 SelectionDAG &DAG) const {
1460 MachineFunction &MF = DAG.getMachineFunction();
1461 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1462 unsigned ARMPCLabelIndex = 0;
1463 EVT PtrVT = getPointerTy();
1464 DebugLoc dl = Op.getDebugLoc();
1465 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1466 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1468 if (RelocM == Reloc::Static)
1469 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1471 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1472 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1473 ARMConstantPoolValue *CPV =
1474 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1475 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1477 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1479 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1480 PseudoSourceValue::getConstantPool(), 0,
1482 SDValue Chain = Result.getValue(1);
1484 if (RelocM == Reloc::PIC_) {
1485 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1486 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1489 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1490 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1491 PseudoSourceValue::getGOT(), 0,
1497 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1498 SelectionDAG &DAG) const {
1499 assert(Subtarget->isTargetELF() &&
1500 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1501 MachineFunction &MF = DAG.getMachineFunction();
1502 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1503 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1504 EVT PtrVT = getPointerTy();
1505 DebugLoc dl = Op.getDebugLoc();
1506 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1507 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1508 "_GLOBAL_OFFSET_TABLE_",
1509 ARMPCLabelIndex, PCAdj);
1510 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1511 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1512 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1513 PseudoSourceValue::getConstantPool(), 0,
1515 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1516 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1520 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1521 const ARMSubtarget *Subtarget)
1523 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1524 DebugLoc dl = Op.getDebugLoc();
1526 default: return SDValue(); // Don't custom lower most intrinsics.
1527 case Intrinsic::arm_thread_pointer: {
1528 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1529 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1531 case Intrinsic::eh_sjlj_lsda: {
1532 MachineFunction &MF = DAG.getMachineFunction();
1533 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1534 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1535 EVT PtrVT = getPointerTy();
1536 DebugLoc dl = Op.getDebugLoc();
1537 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1539 unsigned PCAdj = (RelocM != Reloc::PIC_)
1540 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1541 ARMConstantPoolValue *CPV =
1542 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1543 ARMCP::CPLSDA, PCAdj);
1544 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1545 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1547 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1548 PseudoSourceValue::getConstantPool(), 0,
1550 SDValue Chain = Result.getValue(1);
1552 if (RelocM == Reloc::PIC_) {
1553 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1554 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1558 case Intrinsic::eh_sjlj_setjmp:
1559 SDValue Val = Subtarget->isThumb() ?
1560 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1561 DAG.getConstant(0, MVT::i32);
1562 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1567 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1568 const ARMSubtarget *Subtarget) {
1569 DebugLoc dl = Op.getDebugLoc();
1570 SDValue Op5 = Op.getOperand(5);
1572 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1573 if (isDeviceBarrier) {
1574 if (Subtarget->hasV7Ops())
1575 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1577 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1578 DAG.getConstant(0, MVT::i32));
1580 if (Subtarget->hasV7Ops())
1581 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1583 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1584 DAG.getConstant(0, MVT::i32));
1589 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1590 MachineFunction &MF = DAG.getMachineFunction();
1591 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1593 // vastart just stores the address of the VarArgsFrameIndex slot into the
1594 // memory location argument.
1595 DebugLoc dl = Op.getDebugLoc();
1596 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1597 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1598 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1599 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1604 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1605 SelectionDAG &DAG) const {
1606 SDNode *Node = Op.getNode();
1607 DebugLoc dl = Node->getDebugLoc();
1608 EVT VT = Node->getValueType(0);
1609 SDValue Chain = Op.getOperand(0);
1610 SDValue Size = Op.getOperand(1);
1611 SDValue Align = Op.getOperand(2);
1613 // Chain the dynamic stack allocation so that it doesn't modify the stack
1614 // pointer when other instructions are using the stack.
1615 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1617 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1618 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1619 if (AlignVal > StackAlign)
1620 // Do this now since selection pass cannot introduce new target
1621 // independent node.
1622 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1624 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1625 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1626 // do even more horrible hack later.
1627 MachineFunction &MF = DAG.getMachineFunction();
1628 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1629 if (AFI->isThumb1OnlyFunction()) {
1631 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1633 uint32_t Val = C->getZExtValue();
1634 if (Val <= 508 && ((Val & 3) == 0))
1638 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1641 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1642 SDValue Ops1[] = { Chain, Size, Align };
1643 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1644 Chain = Res.getValue(1);
1645 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1646 DAG.getIntPtrConstant(0, true), SDValue());
1647 SDValue Ops2[] = { Res, Chain };
1648 return DAG.getMergeValues(Ops2, 2, dl);
1652 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1653 SDValue &Root, SelectionDAG &DAG,
1654 DebugLoc dl) const {
1655 MachineFunction &MF = DAG.getMachineFunction();
1656 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1658 TargetRegisterClass *RC;
1659 if (AFI->isThumb1OnlyFunction())
1660 RC = ARM::tGPRRegisterClass;
1662 RC = ARM::GPRRegisterClass;
1664 // Transform the arguments stored in physical registers into virtual ones.
1665 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1666 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1669 if (NextVA.isMemLoc()) {
1670 MachineFrameInfo *MFI = MF.getFrameInfo();
1671 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1673 // Create load node to retrieve arguments from the stack.
1674 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1675 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1676 PseudoSourceValue::getFixedStack(FI), 0,
1679 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1680 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1683 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1687 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1688 CallingConv::ID CallConv, bool isVarArg,
1689 const SmallVectorImpl<ISD::InputArg>
1691 DebugLoc dl, SelectionDAG &DAG,
1692 SmallVectorImpl<SDValue> &InVals)
1695 MachineFunction &MF = DAG.getMachineFunction();
1696 MachineFrameInfo *MFI = MF.getFrameInfo();
1698 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1700 // Assign locations to all of the incoming arguments.
1701 SmallVector<CCValAssign, 16> ArgLocs;
1702 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1704 CCInfo.AnalyzeFormalArguments(Ins,
1705 CCAssignFnForNode(CallConv, /* Return*/ false,
1708 SmallVector<SDValue, 16> ArgValues;
1710 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1711 CCValAssign &VA = ArgLocs[i];
1713 // Arguments stored in registers.
1714 if (VA.isRegLoc()) {
1715 EVT RegVT = VA.getLocVT();
1718 if (VA.needsCustom()) {
1719 // f64 and vector types are split up into multiple registers or
1720 // combinations of registers and stack slots.
1721 if (VA.getLocVT() == MVT::v2f64) {
1722 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1724 VA = ArgLocs[++i]; // skip ahead to next loc
1726 if (VA.isMemLoc()) {
1727 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
1729 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1730 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1731 PseudoSourceValue::getFixedStack(FI), 0,
1734 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1737 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1738 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1739 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1740 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1741 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1743 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1746 TargetRegisterClass *RC;
1748 if (RegVT == MVT::f32)
1749 RC = ARM::SPRRegisterClass;
1750 else if (RegVT == MVT::f64)
1751 RC = ARM::DPRRegisterClass;
1752 else if (RegVT == MVT::v2f64)
1753 RC = ARM::QPRRegisterClass;
1754 else if (RegVT == MVT::i32)
1755 RC = (AFI->isThumb1OnlyFunction() ?
1756 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1758 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1760 // Transform the arguments in physical registers into virtual ones.
1761 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1762 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1765 // If this is an 8 or 16-bit value, it is really passed promoted
1766 // to 32 bits. Insert an assert[sz]ext to capture this, then
1767 // truncate to the right size.
1768 switch (VA.getLocInfo()) {
1769 default: llvm_unreachable("Unknown loc info!");
1770 case CCValAssign::Full: break;
1771 case CCValAssign::BCvt:
1772 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1774 case CCValAssign::SExt:
1775 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1776 DAG.getValueType(VA.getValVT()));
1777 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1779 case CCValAssign::ZExt:
1780 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1781 DAG.getValueType(VA.getValVT()));
1782 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1786 InVals.push_back(ArgValue);
1788 } else { // VA.isRegLoc()
1791 assert(VA.isMemLoc());
1792 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1794 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1795 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1798 // Create load nodes to retrieve arguments from the stack.
1799 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1800 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1801 PseudoSourceValue::getFixedStack(FI), 0,
1808 static const unsigned GPRArgRegs[] = {
1809 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1812 unsigned NumGPRs = CCInfo.getFirstUnallocated
1813 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1815 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1816 unsigned VARegSize = (4 - NumGPRs) * 4;
1817 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1818 unsigned ArgOffset = CCInfo.getNextStackOffset();
1819 if (VARegSaveSize) {
1820 // If this function is vararg, store any remaining integer argument regs
1821 // to their spots on the stack so that they may be loaded by deferencing
1822 // the result of va_next.
1823 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1824 AFI->setVarArgsFrameIndex(
1825 MFI->CreateFixedObject(VARegSaveSize,
1826 ArgOffset + VARegSaveSize - VARegSize,
1828 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
1831 SmallVector<SDValue, 4> MemOps;
1832 for (; NumGPRs < 4; ++NumGPRs) {
1833 TargetRegisterClass *RC;
1834 if (AFI->isThumb1OnlyFunction())
1835 RC = ARM::tGPRRegisterClass;
1837 RC = ARM::GPRRegisterClass;
1839 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1840 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1842 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1843 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()), 0,
1845 MemOps.push_back(Store);
1846 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1847 DAG.getConstant(4, getPointerTy()));
1849 if (!MemOps.empty())
1850 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1851 &MemOps[0], MemOps.size());
1853 // This will point to the next argument passed via stack.
1854 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
1861 /// isFloatingPointZero - Return true if this is +0.0.
1862 static bool isFloatingPointZero(SDValue Op) {
1863 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1864 return CFP->getValueAPF().isPosZero();
1865 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1866 // Maybe this has already been legalized into the constant pool?
1867 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1868 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1869 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1870 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1871 return CFP->getValueAPF().isPosZero();
1877 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1878 /// the given operands.
1880 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1881 SDValue &ARMCC, SelectionDAG &DAG,
1882 DebugLoc dl) const {
1883 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1884 unsigned C = RHSC->getZExtValue();
1885 if (!isLegalICmpImmediate(C)) {
1886 // Constant does not fit, try adjusting it by one?
1891 if (isLegalICmpImmediate(C-1)) {
1892 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1893 RHS = DAG.getConstant(C-1, MVT::i32);
1898 if (C > 0 && isLegalICmpImmediate(C-1)) {
1899 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1900 RHS = DAG.getConstant(C-1, MVT::i32);
1905 if (isLegalICmpImmediate(C+1)) {
1906 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1907 RHS = DAG.getConstant(C+1, MVT::i32);
1912 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1913 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1914 RHS = DAG.getConstant(C+1, MVT::i32);
1921 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1922 ARMISD::NodeType CompareType;
1925 CompareType = ARMISD::CMP;
1930 CompareType = ARMISD::CMPZ;
1933 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1934 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1937 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1938 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1941 if (!isFloatingPointZero(RHS))
1942 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1944 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1945 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1948 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
1949 EVT VT = Op.getValueType();
1950 SDValue LHS = Op.getOperand(0);
1951 SDValue RHS = Op.getOperand(1);
1952 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1953 SDValue TrueVal = Op.getOperand(2);
1954 SDValue FalseVal = Op.getOperand(3);
1955 DebugLoc dl = Op.getDebugLoc();
1957 if (LHS.getValueType() == MVT::i32) {
1959 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1960 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1961 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1964 ARMCC::CondCodes CondCode, CondCode2;
1965 FPCCToARMCC(CC, CondCode, CondCode2);
1967 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1968 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1969 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1970 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1972 if (CondCode2 != ARMCC::AL) {
1973 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1974 // FIXME: Needs another CMP because flag can have but one use.
1975 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1976 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1977 Result, TrueVal, ARMCC2, CCR, Cmp2);
1982 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1983 SDValue Chain = Op.getOperand(0);
1984 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1985 SDValue LHS = Op.getOperand(2);
1986 SDValue RHS = Op.getOperand(3);
1987 SDValue Dest = Op.getOperand(4);
1988 DebugLoc dl = Op.getDebugLoc();
1990 if (LHS.getValueType() == MVT::i32) {
1992 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1993 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1994 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1995 Chain, Dest, ARMCC, CCR,Cmp);
1998 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1999 ARMCC::CondCodes CondCode, CondCode2;
2000 FPCCToARMCC(CC, CondCode, CondCode2);
2002 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2003 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2004 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2005 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2006 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2007 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2008 if (CondCode2 != ARMCC::AL) {
2009 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2010 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2011 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2016 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2017 SDValue Chain = Op.getOperand(0);
2018 SDValue Table = Op.getOperand(1);
2019 SDValue Index = Op.getOperand(2);
2020 DebugLoc dl = Op.getDebugLoc();
2022 EVT PTy = getPointerTy();
2023 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2024 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2025 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2026 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2027 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2028 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2029 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2030 if (Subtarget->isThumb2()) {
2031 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2032 // which does another jump to the destination. This also makes it easier
2033 // to translate it to TBB / TBH later.
2034 // FIXME: This might not work if the function is extremely large.
2035 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2036 Addr, Op.getOperand(2), JTI, UId);
2038 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2039 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2040 PseudoSourceValue::getJumpTable(), 0,
2042 Chain = Addr.getValue(1);
2043 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2044 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2046 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2047 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2048 Chain = Addr.getValue(1);
2049 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2053 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2054 DebugLoc dl = Op.getDebugLoc();
2057 switch (Op.getOpcode()) {
2059 assert(0 && "Invalid opcode!");
2060 case ISD::FP_TO_SINT:
2061 Opc = ARMISD::FTOSI;
2063 case ISD::FP_TO_UINT:
2064 Opc = ARMISD::FTOUI;
2067 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2068 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2071 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2072 EVT VT = Op.getValueType();
2073 DebugLoc dl = Op.getDebugLoc();
2076 switch (Op.getOpcode()) {
2078 assert(0 && "Invalid opcode!");
2079 case ISD::SINT_TO_FP:
2080 Opc = ARMISD::SITOF;
2082 case ISD::UINT_TO_FP:
2083 Opc = ARMISD::UITOF;
2087 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2088 return DAG.getNode(Opc, dl, VT, Op);
2091 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2092 // Implement fcopysign with a fabs and a conditional fneg.
2093 SDValue Tmp0 = Op.getOperand(0);
2094 SDValue Tmp1 = Op.getOperand(1);
2095 DebugLoc dl = Op.getDebugLoc();
2096 EVT VT = Op.getValueType();
2097 EVT SrcVT = Tmp1.getValueType();
2098 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2099 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2100 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2101 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2102 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2105 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2106 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2107 MFI->setFrameAddressIsTaken(true);
2108 EVT VT = Op.getValueType();
2109 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2110 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2111 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2112 ? ARM::R7 : ARM::R11;
2113 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2115 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2121 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2123 SDValue Dst, SDValue Src,
2124 SDValue Size, unsigned Align,
2125 bool isVolatile, bool AlwaysInline,
2129 uint64_t SrcSVOff) const {
2130 // Do repeated 4-byte loads and stores. To be improved.
2131 // This requires 4-byte alignment.
2132 if ((Align & 3) != 0)
2134 // This requires the copy size to be a constant, preferrably
2135 // within a subtarget-specific limit.
2136 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2139 uint64_t SizeVal = ConstantSize->getZExtValue();
2140 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2143 unsigned BytesLeft = SizeVal & 3;
2144 unsigned NumMemOps = SizeVal >> 2;
2145 unsigned EmittedNumMemOps = 0;
2147 unsigned VTSize = 4;
2149 const unsigned MAX_LOADS_IN_LDM = 6;
2150 SDValue TFOps[MAX_LOADS_IN_LDM];
2151 SDValue Loads[MAX_LOADS_IN_LDM];
2152 uint64_t SrcOff = 0, DstOff = 0;
2154 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2155 // same number of stores. The loads and stores will get combined into
2156 // ldm/stm later on.
2157 while (EmittedNumMemOps < NumMemOps) {
2159 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2160 Loads[i] = DAG.getLoad(VT, dl, Chain,
2161 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2162 DAG.getConstant(SrcOff, MVT::i32)),
2163 SrcSV, SrcSVOff + SrcOff, isVolatile, false, 0);
2164 TFOps[i] = Loads[i].getValue(1);
2167 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2170 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2171 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2172 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2173 DAG.getConstant(DstOff, MVT::i32)),
2174 DstSV, DstSVOff + DstOff, isVolatile, false, 0);
2177 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2179 EmittedNumMemOps += i;
2185 // Issue loads / stores for the trailing (1 - 3) bytes.
2186 unsigned BytesLeftSave = BytesLeft;
2189 if (BytesLeft >= 2) {
2197 Loads[i] = DAG.getLoad(VT, dl, Chain,
2198 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2199 DAG.getConstant(SrcOff, MVT::i32)),
2200 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2201 TFOps[i] = Loads[i].getValue(1);
2204 BytesLeft -= VTSize;
2206 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2209 BytesLeft = BytesLeftSave;
2211 if (BytesLeft >= 2) {
2219 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2220 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2221 DAG.getConstant(DstOff, MVT::i32)),
2222 DstSV, DstSVOff + DstOff, false, false, 0);
2225 BytesLeft -= VTSize;
2227 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2230 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2231 /// expand a bit convert where either the source or destination type is i64 to
2232 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2233 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2234 /// vectors), since the legalizer won't know what to do with that.
2235 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2236 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2237 DebugLoc dl = N->getDebugLoc();
2238 SDValue Op = N->getOperand(0);
2240 // This function is only supposed to be called for i64 types, either as the
2241 // source or destination of the bit convert.
2242 EVT SrcVT = Op.getValueType();
2243 EVT DstVT = N->getValueType(0);
2244 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2245 "ExpandBIT_CONVERT called for non-i64 type");
2247 // Turn i64->f64 into VMOVDRR.
2248 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2249 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2250 DAG.getConstant(0, MVT::i32));
2251 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2252 DAG.getConstant(1, MVT::i32));
2253 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2256 // Turn f64->i64 into VMOVRRD.
2257 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2258 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2259 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2260 // Merge the pieces into a single i64 value.
2261 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2267 /// getZeroVector - Returns a vector of specified type with all zero elements.
2269 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2270 assert(VT.isVector() && "Expected a vector type");
2272 // Zero vectors are used to represent vector negation and in those cases
2273 // will be implemented with the NEON VNEG instruction. However, VNEG does
2274 // not support i64 elements, so sometimes the zero vectors will need to be
2275 // explicitly constructed. For those cases, and potentially other uses in
2276 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2277 // to their dest type. This ensures they get CSE'd.
2279 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2280 SmallVector<SDValue, 8> Ops;
2283 if (VT.getSizeInBits() == 64) {
2284 Ops.assign(8, Cst); TVT = MVT::v8i8;
2286 Ops.assign(16, Cst); TVT = MVT::v16i8;
2288 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2290 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2293 /// getOnesVector - Returns a vector of specified type with all bits set.
2295 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2296 assert(VT.isVector() && "Expected a vector type");
2298 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2299 // dest type. This ensures they get CSE'd.
2301 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2302 SmallVector<SDValue, 8> Ops;
2305 if (VT.getSizeInBits() == 64) {
2306 Ops.assign(8, Cst); TVT = MVT::v8i8;
2308 Ops.assign(16, Cst); TVT = MVT::v16i8;
2310 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2312 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2315 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2316 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2317 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2318 SelectionDAG &DAG) const {
2319 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2320 EVT VT = Op.getValueType();
2321 unsigned VTBits = VT.getSizeInBits();
2322 DebugLoc dl = Op.getDebugLoc();
2323 SDValue ShOpLo = Op.getOperand(0);
2324 SDValue ShOpHi = Op.getOperand(1);
2325 SDValue ShAmt = Op.getOperand(2);
2327 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2329 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2331 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2332 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2333 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2334 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2335 DAG.getConstant(VTBits, MVT::i32));
2336 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2337 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2338 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2340 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2341 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2343 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2344 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2347 SDValue Ops[2] = { Lo, Hi };
2348 return DAG.getMergeValues(Ops, 2, dl);
2351 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2352 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2353 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2354 SelectionDAG &DAG) const {
2355 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2356 EVT VT = Op.getValueType();
2357 unsigned VTBits = VT.getSizeInBits();
2358 DebugLoc dl = Op.getDebugLoc();
2359 SDValue ShOpLo = Op.getOperand(0);
2360 SDValue ShOpHi = Op.getOperand(1);
2361 SDValue ShAmt = Op.getOperand(2);
2364 assert(Op.getOpcode() == ISD::SHL_PARTS);
2365 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2366 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2367 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2368 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2369 DAG.getConstant(VTBits, MVT::i32));
2370 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2371 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2373 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2374 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2375 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2377 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2378 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2381 SDValue Ops[2] = { Lo, Hi };
2382 return DAG.getMergeValues(Ops, 2, dl);
2385 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2386 const ARMSubtarget *ST) {
2387 EVT VT = N->getValueType(0);
2388 DebugLoc dl = N->getDebugLoc();
2390 if (!ST->hasV6T2Ops())
2393 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2394 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2397 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2398 const ARMSubtarget *ST) {
2399 EVT VT = N->getValueType(0);
2400 DebugLoc dl = N->getDebugLoc();
2402 // Lower vector shifts on NEON to use VSHL.
2403 if (VT.isVector()) {
2404 assert(ST->hasNEON() && "unexpected vector shift");
2406 // Left shifts translate directly to the vshiftu intrinsic.
2407 if (N->getOpcode() == ISD::SHL)
2408 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2409 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2410 N->getOperand(0), N->getOperand(1));
2412 assert((N->getOpcode() == ISD::SRA ||
2413 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2415 // NEON uses the same intrinsics for both left and right shifts. For
2416 // right shifts, the shift amounts are negative, so negate the vector of
2418 EVT ShiftVT = N->getOperand(1).getValueType();
2419 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2420 getZeroVector(ShiftVT, DAG, dl),
2422 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2423 Intrinsic::arm_neon_vshifts :
2424 Intrinsic::arm_neon_vshiftu);
2425 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2426 DAG.getConstant(vshiftInt, MVT::i32),
2427 N->getOperand(0), NegatedCount);
2430 // We can get here for a node like i32 = ISD::SHL i32, i64
2434 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2435 "Unknown shift to lower!");
2437 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2438 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2439 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2442 // If we are in thumb mode, we don't have RRX.
2443 if (ST->isThumb1Only()) return SDValue();
2445 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2446 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2447 DAG.getConstant(0, MVT::i32));
2448 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2449 DAG.getConstant(1, MVT::i32));
2451 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2452 // captures the result into a carry flag.
2453 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2454 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2456 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2457 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2459 // Merge the pieces into a single i64 value.
2460 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2463 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2464 SDValue TmpOp0, TmpOp1;
2465 bool Invert = false;
2469 SDValue Op0 = Op.getOperand(0);
2470 SDValue Op1 = Op.getOperand(1);
2471 SDValue CC = Op.getOperand(2);
2472 EVT VT = Op.getValueType();
2473 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2474 DebugLoc dl = Op.getDebugLoc();
2476 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2477 switch (SetCCOpcode) {
2478 default: llvm_unreachable("Illegal FP comparison"); break;
2480 case ISD::SETNE: Invert = true; // Fallthrough
2482 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2484 case ISD::SETLT: Swap = true; // Fallthrough
2486 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2488 case ISD::SETLE: Swap = true; // Fallthrough
2490 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2491 case ISD::SETUGE: Swap = true; // Fallthrough
2492 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2493 case ISD::SETUGT: Swap = true; // Fallthrough
2494 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2495 case ISD::SETUEQ: Invert = true; // Fallthrough
2497 // Expand this to (OLT | OGT).
2501 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2502 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2504 case ISD::SETUO: Invert = true; // Fallthrough
2506 // Expand this to (OLT | OGE).
2510 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2511 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2515 // Integer comparisons.
2516 switch (SetCCOpcode) {
2517 default: llvm_unreachable("Illegal integer comparison"); break;
2518 case ISD::SETNE: Invert = true;
2519 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2520 case ISD::SETLT: Swap = true;
2521 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2522 case ISD::SETLE: Swap = true;
2523 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2524 case ISD::SETULT: Swap = true;
2525 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2526 case ISD::SETULE: Swap = true;
2527 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2530 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2531 if (Opc == ARMISD::VCEQ) {
2534 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2536 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2539 // Ignore bitconvert.
2540 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2541 AndOp = AndOp.getOperand(0);
2543 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2545 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2546 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2553 std::swap(Op0, Op1);
2555 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2558 Result = DAG.getNOT(dl, Result, VT);
2563 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2564 /// VMOV instruction, and if so, return the constant being splatted.
2565 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2566 unsigned SplatBitSize, SelectionDAG &DAG) {
2567 switch (SplatBitSize) {
2569 // Any 1-byte value is OK.
2570 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2571 return DAG.getTargetConstant(SplatBits, MVT::i8);
2574 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2575 if ((SplatBits & ~0xff) == 0 ||
2576 (SplatBits & ~0xff00) == 0)
2577 return DAG.getTargetConstant(SplatBits, MVT::i16);
2581 // NEON's 32-bit VMOV supports splat values where:
2582 // * only one byte is nonzero, or
2583 // * the least significant byte is 0xff and the second byte is nonzero, or
2584 // * the least significant 2 bytes are 0xff and the third is nonzero.
2585 if ((SplatBits & ~0xff) == 0 ||
2586 (SplatBits & ~0xff00) == 0 ||
2587 (SplatBits & ~0xff0000) == 0 ||
2588 (SplatBits & ~0xff000000) == 0)
2589 return DAG.getTargetConstant(SplatBits, MVT::i32);
2591 if ((SplatBits & ~0xffff) == 0 &&
2592 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2593 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2595 if ((SplatBits & ~0xffffff) == 0 &&
2596 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2597 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2599 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2600 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2601 // VMOV.I32. A (very) minor optimization would be to replicate the value
2602 // and fall through here to test for a valid 64-bit splat. But, then the
2603 // caller would also need to check and handle the change in size.
2607 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2608 uint64_t BitMask = 0xff;
2610 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2611 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2613 else if ((SplatBits & BitMask) != 0)
2617 return DAG.getTargetConstant(Val, MVT::i64);
2621 llvm_unreachable("unexpected size for isVMOVSplat");
2628 /// getVMOVImm - If this is a build_vector of constants which can be
2629 /// formed by using a VMOV instruction of the specified element size,
2630 /// return the constant being splatted. The ByteSize field indicates the
2631 /// number of bytes of each element [1248].
2632 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2633 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2634 APInt SplatBits, SplatUndef;
2635 unsigned SplatBitSize;
2637 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2638 HasAnyUndefs, ByteSize * 8))
2641 if (SplatBitSize > ByteSize * 8)
2644 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2648 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2649 bool &ReverseVEXT, unsigned &Imm) {
2650 unsigned NumElts = VT.getVectorNumElements();
2651 ReverseVEXT = false;
2654 // If this is a VEXT shuffle, the immediate value is the index of the first
2655 // element. The other shuffle indices must be the successive elements after
2657 unsigned ExpectedElt = Imm;
2658 for (unsigned i = 1; i < NumElts; ++i) {
2659 // Increment the expected index. If it wraps around, it may still be
2660 // a VEXT but the source vectors must be swapped.
2662 if (ExpectedElt == NumElts * 2) {
2667 if (ExpectedElt != static_cast<unsigned>(M[i]))
2671 // Adjust the index value if the source operands will be swapped.
2678 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2679 /// instruction with the specified blocksize. (The order of the elements
2680 /// within each block of the vector is reversed.)
2681 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2682 unsigned BlockSize) {
2683 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2684 "Only possible block sizes for VREV are: 16, 32, 64");
2686 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2690 unsigned NumElts = VT.getVectorNumElements();
2691 unsigned BlockElts = M[0] + 1;
2693 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2696 for (unsigned i = 0; i < NumElts; ++i) {
2697 if ((unsigned) M[i] !=
2698 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2705 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2706 unsigned &WhichResult) {
2707 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2711 unsigned NumElts = VT.getVectorNumElements();
2712 WhichResult = (M[0] == 0 ? 0 : 1);
2713 for (unsigned i = 0; i < NumElts; i += 2) {
2714 if ((unsigned) M[i] != i + WhichResult ||
2715 (unsigned) M[i+1] != i + NumElts + WhichResult)
2721 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2722 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2723 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2724 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2725 unsigned &WhichResult) {
2726 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2730 unsigned NumElts = VT.getVectorNumElements();
2731 WhichResult = (M[0] == 0 ? 0 : 1);
2732 for (unsigned i = 0; i < NumElts; i += 2) {
2733 if ((unsigned) M[i] != i + WhichResult ||
2734 (unsigned) M[i+1] != i + WhichResult)
2740 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2741 unsigned &WhichResult) {
2742 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2746 unsigned NumElts = VT.getVectorNumElements();
2747 WhichResult = (M[0] == 0 ? 0 : 1);
2748 for (unsigned i = 0; i != NumElts; ++i) {
2749 if ((unsigned) M[i] != 2 * i + WhichResult)
2753 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2754 if (VT.is64BitVector() && EltSz == 32)
2760 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2761 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2762 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2763 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2764 unsigned &WhichResult) {
2765 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2769 unsigned Half = VT.getVectorNumElements() / 2;
2770 WhichResult = (M[0] == 0 ? 0 : 1);
2771 for (unsigned j = 0; j != 2; ++j) {
2772 unsigned Idx = WhichResult;
2773 for (unsigned i = 0; i != Half; ++i) {
2774 if ((unsigned) M[i + j * Half] != Idx)
2780 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2781 if (VT.is64BitVector() && EltSz == 32)
2787 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2788 unsigned &WhichResult) {
2789 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2793 unsigned NumElts = VT.getVectorNumElements();
2794 WhichResult = (M[0] == 0 ? 0 : 1);
2795 unsigned Idx = WhichResult * NumElts / 2;
2796 for (unsigned i = 0; i != NumElts; i += 2) {
2797 if ((unsigned) M[i] != Idx ||
2798 (unsigned) M[i+1] != Idx + NumElts)
2803 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2804 if (VT.is64BitVector() && EltSz == 32)
2810 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2811 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2812 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2813 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2814 unsigned &WhichResult) {
2815 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2819 unsigned NumElts = VT.getVectorNumElements();
2820 WhichResult = (M[0] == 0 ? 0 : 1);
2821 unsigned Idx = WhichResult * NumElts / 2;
2822 for (unsigned i = 0; i != NumElts; i += 2) {
2823 if ((unsigned) M[i] != Idx ||
2824 (unsigned) M[i+1] != Idx)
2829 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2830 if (VT.is64BitVector() && EltSz == 32)
2837 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2838 // Canonicalize all-zeros and all-ones vectors.
2839 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2840 if (ConstVal->isNullValue())
2841 return getZeroVector(VT, DAG, dl);
2842 if (ConstVal->isAllOnesValue())
2843 return getOnesVector(VT, DAG, dl);
2846 if (VT.is64BitVector()) {
2847 switch (Val.getValueType().getSizeInBits()) {
2848 case 8: CanonicalVT = MVT::v8i8; break;
2849 case 16: CanonicalVT = MVT::v4i16; break;
2850 case 32: CanonicalVT = MVT::v2i32; break;
2851 case 64: CanonicalVT = MVT::v1i64; break;
2852 default: llvm_unreachable("unexpected splat element type"); break;
2855 assert(VT.is128BitVector() && "unknown splat vector size");
2856 switch (Val.getValueType().getSizeInBits()) {
2857 case 8: CanonicalVT = MVT::v16i8; break;
2858 case 16: CanonicalVT = MVT::v8i16; break;
2859 case 32: CanonicalVT = MVT::v4i32; break;
2860 case 64: CanonicalVT = MVT::v2i64; break;
2861 default: llvm_unreachable("unexpected splat element type"); break;
2865 // Build a canonical splat for this value.
2866 SmallVector<SDValue, 8> Ops;
2867 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2868 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2870 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2873 // If this is a case we can't handle, return null and let the default
2874 // expansion code take care of it.
2875 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2876 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2877 DebugLoc dl = Op.getDebugLoc();
2878 EVT VT = Op.getValueType();
2880 APInt SplatBits, SplatUndef;
2881 unsigned SplatBitSize;
2883 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2884 if (SplatBitSize <= 64) {
2885 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2886 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2888 return BuildSplat(Val, VT, DAG, dl);
2892 // If there are only 2 elements in a 128-bit vector, insert them into an
2893 // undef vector. This handles the common case for 128-bit vector argument
2894 // passing, where the insertions should be translated to subreg accesses
2895 // with no real instructions.
2896 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2897 SDValue Val = DAG.getUNDEF(VT);
2898 SDValue Op0 = Op.getOperand(0);
2899 SDValue Op1 = Op.getOperand(1);
2900 if (Op0.getOpcode() != ISD::UNDEF)
2901 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2902 DAG.getIntPtrConstant(0));
2903 if (Op1.getOpcode() != ISD::UNDEF)
2904 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2905 DAG.getIntPtrConstant(1));
2912 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2913 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2914 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2915 /// are assumed to be legal.
2917 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2919 if (VT.getVectorNumElements() == 4 &&
2920 (VT.is128BitVector() || VT.is64BitVector())) {
2921 unsigned PFIndexes[4];
2922 for (unsigned i = 0; i != 4; ++i) {
2926 PFIndexes[i] = M[i];
2929 // Compute the index in the perfect shuffle table.
2930 unsigned PFTableIndex =
2931 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2932 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2933 unsigned Cost = (PFEntry >> 30);
2940 unsigned Imm, WhichResult;
2942 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2943 isVREVMask(M, VT, 64) ||
2944 isVREVMask(M, VT, 32) ||
2945 isVREVMask(M, VT, 16) ||
2946 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2947 isVTRNMask(M, VT, WhichResult) ||
2948 isVUZPMask(M, VT, WhichResult) ||
2949 isVZIPMask(M, VT, WhichResult) ||
2950 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2951 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2952 isVZIP_v_undef_Mask(M, VT, WhichResult));
2955 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2956 /// the specified operations to build the shuffle.
2957 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2958 SDValue RHS, SelectionDAG &DAG,
2960 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2961 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2962 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2965 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2974 OP_VUZPL, // VUZP, left result
2975 OP_VUZPR, // VUZP, right result
2976 OP_VZIPL, // VZIP, left result
2977 OP_VZIPR, // VZIP, right result
2978 OP_VTRNL, // VTRN, left result
2979 OP_VTRNR // VTRN, right result
2982 if (OpNum == OP_COPY) {
2983 if (LHSID == (1*9+2)*9+3) return LHS;
2984 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2988 SDValue OpLHS, OpRHS;
2989 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2990 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2991 EVT VT = OpLHS.getValueType();
2994 default: llvm_unreachable("Unknown shuffle opcode!");
2996 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3001 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3002 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3006 return DAG.getNode(ARMISD::VEXT, dl, VT,
3008 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3011 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3012 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3015 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3016 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3019 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3020 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3024 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3025 SDValue V1 = Op.getOperand(0);
3026 SDValue V2 = Op.getOperand(1);
3027 DebugLoc dl = Op.getDebugLoc();
3028 EVT VT = Op.getValueType();
3029 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3030 SmallVector<int, 8> ShuffleMask;
3032 // Convert shuffles that are directly supported on NEON to target-specific
3033 // DAG nodes, instead of keeping them as shuffles and matching them again
3034 // during code selection. This is more efficient and avoids the possibility
3035 // of inconsistencies between legalization and selection.
3036 // FIXME: floating-point vectors should be canonicalized to integer vectors
3037 // of the same time so that they get CSEd properly.
3038 SVN->getMask(ShuffleMask);
3040 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3041 int Lane = SVN->getSplatIndex();
3042 // If this is undef splat, generate it via "just" vdup, if possible.
3043 if (Lane == -1) Lane = 0;
3045 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3046 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3048 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3049 DAG.getConstant(Lane, MVT::i32));
3054 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3057 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3058 DAG.getConstant(Imm, MVT::i32));
3061 if (isVREVMask(ShuffleMask, VT, 64))
3062 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3063 if (isVREVMask(ShuffleMask, VT, 32))
3064 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3065 if (isVREVMask(ShuffleMask, VT, 16))
3066 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3068 // Check for Neon shuffles that modify both input vectors in place.
3069 // If both results are used, i.e., if there are two shuffles with the same
3070 // source operands and with masks corresponding to both results of one of
3071 // these operations, DAG memoization will ensure that a single node is
3072 // used for both shuffles.
3073 unsigned WhichResult;
3074 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3075 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3076 V1, V2).getValue(WhichResult);
3077 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3078 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3079 V1, V2).getValue(WhichResult);
3080 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3081 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3082 V1, V2).getValue(WhichResult);
3084 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3085 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3086 V1, V1).getValue(WhichResult);
3087 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3088 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3089 V1, V1).getValue(WhichResult);
3090 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3091 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3092 V1, V1).getValue(WhichResult);
3094 // If the shuffle is not directly supported and it has 4 elements, use
3095 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3096 if (VT.getVectorNumElements() == 4 &&
3097 (VT.is128BitVector() || VT.is64BitVector())) {
3098 unsigned PFIndexes[4];
3099 for (unsigned i = 0; i != 4; ++i) {
3100 if (ShuffleMask[i] < 0)
3103 PFIndexes[i] = ShuffleMask[i];
3106 // Compute the index in the perfect shuffle table.
3107 unsigned PFTableIndex =
3108 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3110 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3111 unsigned Cost = (PFEntry >> 30);
3114 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3120 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3121 EVT VT = Op.getValueType();
3122 DebugLoc dl = Op.getDebugLoc();
3123 SDValue Vec = Op.getOperand(0);
3124 SDValue Lane = Op.getOperand(1);
3125 assert(VT == MVT::i32 &&
3126 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3127 "unexpected type for custom-lowering vector extract");
3128 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3131 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3132 // The only time a CONCAT_VECTORS operation can have legal types is when
3133 // two 64-bit vectors are concatenated to a 128-bit vector.
3134 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3135 "unexpected CONCAT_VECTORS");
3136 DebugLoc dl = Op.getDebugLoc();
3137 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3138 SDValue Op0 = Op.getOperand(0);
3139 SDValue Op1 = Op.getOperand(1);
3140 if (Op0.getOpcode() != ISD::UNDEF)
3141 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3142 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3143 DAG.getIntPtrConstant(0));
3144 if (Op1.getOpcode() != ISD::UNDEF)
3145 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3146 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3147 DAG.getIntPtrConstant(1));
3148 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3151 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3152 switch (Op.getOpcode()) {
3153 default: llvm_unreachable("Don't know how to custom lower this!");
3154 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3155 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3156 case ISD::GlobalAddress:
3157 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3158 LowerGlobalAddressELF(Op, DAG);
3159 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3160 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3161 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3162 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3163 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3164 case ISD::VASTART: return LowerVASTART(Op, DAG);
3165 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3166 case ISD::SINT_TO_FP:
3167 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3168 case ISD::FP_TO_SINT:
3169 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3170 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3171 case ISD::RETURNADDR: break;
3172 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3173 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3174 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3176 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3179 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3180 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3181 case ISD::SRL_PARTS:
3182 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3183 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3184 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3185 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3186 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3187 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3188 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3193 /// ReplaceNodeResults - Replace the results of node with an illegal result
3194 /// type with new values built out of custom code.
3195 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3196 SmallVectorImpl<SDValue>&Results,
3197 SelectionDAG &DAG) const {
3199 switch (N->getOpcode()) {
3201 llvm_unreachable("Don't know how to custom expand this!");
3203 case ISD::BIT_CONVERT:
3204 Res = ExpandBIT_CONVERT(N, DAG);
3208 Res = LowerShift(N, DAG, Subtarget);
3212 Results.push_back(Res);
3215 //===----------------------------------------------------------------------===//
3216 // ARM Scheduler Hooks
3217 //===----------------------------------------------------------------------===//
3220 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3221 MachineBasicBlock *BB,
3222 unsigned Size) const {
3223 unsigned dest = MI->getOperand(0).getReg();
3224 unsigned ptr = MI->getOperand(1).getReg();
3225 unsigned oldval = MI->getOperand(2).getReg();
3226 unsigned newval = MI->getOperand(3).getReg();
3227 unsigned scratch = BB->getParent()->getRegInfo()
3228 .createVirtualRegister(ARM::GPRRegisterClass);
3229 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3230 DebugLoc dl = MI->getDebugLoc();
3231 bool isThumb2 = Subtarget->isThumb2();
3233 unsigned ldrOpc, strOpc;
3235 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3237 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3238 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3241 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3242 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3245 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3246 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3250 MachineFunction *MF = BB->getParent();
3251 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3252 MachineFunction::iterator It = BB;
3253 ++It; // insert the new blocks after the current block
3255 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3256 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3257 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3258 MF->insert(It, loop1MBB);
3259 MF->insert(It, loop2MBB);
3260 MF->insert(It, exitMBB);
3261 exitMBB->transferSuccessors(BB);
3265 // fallthrough --> loop1MBB
3266 BB->addSuccessor(loop1MBB);
3269 // ldrex dest, [ptr]
3273 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3274 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3275 .addReg(dest).addReg(oldval));
3276 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3277 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3278 BB->addSuccessor(loop2MBB);
3279 BB->addSuccessor(exitMBB);
3282 // strex scratch, newval, [ptr]
3286 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3288 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3289 .addReg(scratch).addImm(0));
3290 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3291 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3292 BB->addSuccessor(loop1MBB);
3293 BB->addSuccessor(exitMBB);
3299 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3305 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3306 unsigned Size, unsigned BinOpcode) const {
3307 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3308 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3310 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3311 MachineFunction *MF = BB->getParent();
3312 MachineFunction::iterator It = BB;
3315 unsigned dest = MI->getOperand(0).getReg();
3316 unsigned ptr = MI->getOperand(1).getReg();
3317 unsigned incr = MI->getOperand(2).getReg();
3318 DebugLoc dl = MI->getDebugLoc();
3320 bool isThumb2 = Subtarget->isThumb2();
3321 unsigned ldrOpc, strOpc;
3323 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3325 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3326 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3329 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3330 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3333 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3334 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3338 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3339 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3340 MF->insert(It, loopMBB);
3341 MF->insert(It, exitMBB);
3342 exitMBB->transferSuccessors(BB);
3344 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3345 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3346 unsigned scratch2 = (!BinOpcode) ? incr :
3347 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3351 // fallthrough --> loopMBB
3352 BB->addSuccessor(loopMBB);
3356 // <binop> scratch2, dest, incr
3357 // strex scratch, scratch2, ptr
3360 // fallthrough --> exitMBB
3362 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3364 // operand order needs to go the other way for NAND
3365 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3366 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3367 addReg(incr).addReg(dest)).addReg(0);
3369 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3370 addReg(dest).addReg(incr)).addReg(0);
3373 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3375 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3376 .addReg(scratch).addImm(0));
3377 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3378 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3380 BB->addSuccessor(loopMBB);
3381 BB->addSuccessor(exitMBB);
3387 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3393 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3394 MachineBasicBlock *BB) const {
3395 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3396 DebugLoc dl = MI->getDebugLoc();
3397 bool isThumb2 = Subtarget->isThumb2();
3398 switch (MI->getOpcode()) {
3401 llvm_unreachable("Unexpected instr type to insert");
3403 case ARM::ATOMIC_LOAD_ADD_I8:
3404 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3405 case ARM::ATOMIC_LOAD_ADD_I16:
3406 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3407 case ARM::ATOMIC_LOAD_ADD_I32:
3408 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3410 case ARM::ATOMIC_LOAD_AND_I8:
3411 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3412 case ARM::ATOMIC_LOAD_AND_I16:
3413 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3414 case ARM::ATOMIC_LOAD_AND_I32:
3415 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3417 case ARM::ATOMIC_LOAD_OR_I8:
3418 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3419 case ARM::ATOMIC_LOAD_OR_I16:
3420 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3421 case ARM::ATOMIC_LOAD_OR_I32:
3422 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3424 case ARM::ATOMIC_LOAD_XOR_I8:
3425 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3426 case ARM::ATOMIC_LOAD_XOR_I16:
3427 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3428 case ARM::ATOMIC_LOAD_XOR_I32:
3429 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3431 case ARM::ATOMIC_LOAD_NAND_I8:
3432 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3433 case ARM::ATOMIC_LOAD_NAND_I16:
3434 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3435 case ARM::ATOMIC_LOAD_NAND_I32:
3436 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3438 case ARM::ATOMIC_LOAD_SUB_I8:
3439 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3440 case ARM::ATOMIC_LOAD_SUB_I16:
3441 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3442 case ARM::ATOMIC_LOAD_SUB_I32:
3443 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3445 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3446 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3447 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3449 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3450 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3451 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3453 case ARM::tMOVCCr_pseudo: {
3454 // To "insert" a SELECT_CC instruction, we actually have to insert the
3455 // diamond control-flow pattern. The incoming instruction knows the
3456 // destination vreg to set, the condition code register to branch on, the
3457 // true/false values to select between, and a branch opcode to use.
3458 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3459 MachineFunction::iterator It = BB;
3465 // cmpTY ccX, r1, r2
3467 // fallthrough --> copy0MBB
3468 MachineBasicBlock *thisMBB = BB;
3469 MachineFunction *F = BB->getParent();
3470 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3471 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3472 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3473 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3474 F->insert(It, copy0MBB);
3475 F->insert(It, sinkMBB);
3476 // Update machine-CFG edges by first adding all successors of the current
3477 // block to the new block which will contain the Phi node for the select.
3478 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3479 E = BB->succ_end(); I != E; ++I)
3480 sinkMBB->addSuccessor(*I);
3481 // Next, remove all successors of the current block, and add the true
3482 // and fallthrough blocks as its successors.
3483 while (!BB->succ_empty())
3484 BB->removeSuccessor(BB->succ_begin());
3485 BB->addSuccessor(copy0MBB);
3486 BB->addSuccessor(sinkMBB);
3489 // %FalseValue = ...
3490 // # fallthrough to sinkMBB
3493 // Update machine-CFG edges
3494 BB->addSuccessor(sinkMBB);
3497 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3500 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3501 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3502 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3504 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3511 case ARM::t2SUBrSPi_:
3512 case ARM::t2SUBrSPi12_:
3513 case ARM::t2SUBrSPs_: {
3514 MachineFunction *MF = BB->getParent();
3515 unsigned DstReg = MI->getOperand(0).getReg();
3516 unsigned SrcReg = MI->getOperand(1).getReg();
3517 bool DstIsDead = MI->getOperand(0).isDead();
3518 bool SrcIsKill = MI->getOperand(1).isKill();
3520 if (SrcReg != ARM::SP) {
3521 // Copy the source to SP from virtual register.
3522 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3523 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3524 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3525 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3526 .addReg(SrcReg, getKillRegState(SrcIsKill));
3530 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3531 switch (MI->getOpcode()) {
3533 llvm_unreachable("Unexpected pseudo instruction!");
3539 OpOpc = ARM::tADDspr;
3542 OpOpc = ARM::tSUBspi;
3544 case ARM::t2SUBrSPi_:
3545 OpOpc = ARM::t2SUBrSPi;
3546 NeedPred = true; NeedCC = true;
3548 case ARM::t2SUBrSPi12_:
3549 OpOpc = ARM::t2SUBrSPi12;
3552 case ARM::t2SUBrSPs_:
3553 OpOpc = ARM::t2SUBrSPs;
3554 NeedPred = true; NeedCC = true; NeedOp3 = true;
3557 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3558 if (OpOpc == ARM::tAND)
3559 AddDefaultT1CC(MIB);
3560 MIB.addReg(ARM::SP);
3561 MIB.addOperand(MI->getOperand(2));
3563 MIB.addOperand(MI->getOperand(3));
3565 AddDefaultPred(MIB);
3569 // Copy the result from SP to virtual register.
3570 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3571 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3572 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3573 BuildMI(BB, dl, TII->get(CopyOpc))
3574 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3576 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3582 //===----------------------------------------------------------------------===//
3583 // ARM Optimization Hooks
3584 //===----------------------------------------------------------------------===//
3587 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3588 TargetLowering::DAGCombinerInfo &DCI) {
3589 SelectionDAG &DAG = DCI.DAG;
3590 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3591 EVT VT = N->getValueType(0);
3592 unsigned Opc = N->getOpcode();
3593 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3594 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3595 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3596 ISD::CondCode CC = ISD::SETCC_INVALID;
3599 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3601 SDValue CCOp = Slct.getOperand(0);
3602 if (CCOp.getOpcode() == ISD::SETCC)
3603 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3606 bool DoXform = false;
3608 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3611 if (LHS.getOpcode() == ISD::Constant &&
3612 cast<ConstantSDNode>(LHS)->isNullValue()) {
3614 } else if (CC != ISD::SETCC_INVALID &&
3615 RHS.getOpcode() == ISD::Constant &&
3616 cast<ConstantSDNode>(RHS)->isNullValue()) {
3617 std::swap(LHS, RHS);
3618 SDValue Op0 = Slct.getOperand(0);
3619 EVT OpVT = isSlctCC ? Op0.getValueType() :
3620 Op0.getOperand(0).getValueType();
3621 bool isInt = OpVT.isInteger();
3622 CC = ISD::getSetCCInverse(CC, isInt);
3624 if (!TLI.isCondCodeLegal(CC, OpVT))
3625 return SDValue(); // Inverse operator isn't legal.
3632 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3634 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3635 Slct.getOperand(0), Slct.getOperand(1), CC);
3636 SDValue CCOp = Slct.getOperand(0);
3638 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3639 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3640 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3641 CCOp, OtherOp, Result);
3646 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3647 static SDValue PerformADDCombine(SDNode *N,
3648 TargetLowering::DAGCombinerInfo &DCI) {
3649 // added by evan in r37685 with no testcase.
3650 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3652 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3653 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3654 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3655 if (Result.getNode()) return Result;
3657 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3658 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3659 if (Result.getNode()) return Result;
3665 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3666 static SDValue PerformSUBCombine(SDNode *N,
3667 TargetLowering::DAGCombinerInfo &DCI) {
3668 // added by evan in r37685 with no testcase.
3669 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3671 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3672 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3673 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3674 if (Result.getNode()) return Result;
3680 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3681 /// ARMISD::VMOVRRD.
3682 static SDValue PerformVMOVRRDCombine(SDNode *N,
3683 TargetLowering::DAGCombinerInfo &DCI) {
3684 // fmrrd(fmdrr x, y) -> x,y
3685 SDValue InDouble = N->getOperand(0);
3686 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3687 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3691 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3692 /// operand of a vector shift operation, where all the elements of the
3693 /// build_vector must have the same constant integer value.
3694 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3695 // Ignore bit_converts.
3696 while (Op.getOpcode() == ISD::BIT_CONVERT)
3697 Op = Op.getOperand(0);
3698 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3699 APInt SplatBits, SplatUndef;
3700 unsigned SplatBitSize;
3702 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3703 HasAnyUndefs, ElementBits) ||
3704 SplatBitSize > ElementBits)
3706 Cnt = SplatBits.getSExtValue();
3710 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3711 /// operand of a vector shift left operation. That value must be in the range:
3712 /// 0 <= Value < ElementBits for a left shift; or
3713 /// 0 <= Value <= ElementBits for a long left shift.
3714 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3715 assert(VT.isVector() && "vector shift count is not a vector type");
3716 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3717 if (! getVShiftImm(Op, ElementBits, Cnt))
3719 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3722 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3723 /// operand of a vector shift right operation. For a shift opcode, the value
3724 /// is positive, but for an intrinsic the value count must be negative. The
3725 /// absolute value must be in the range:
3726 /// 1 <= |Value| <= ElementBits for a right shift; or
3727 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3728 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3730 assert(VT.isVector() && "vector shift count is not a vector type");
3731 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3732 if (! getVShiftImm(Op, ElementBits, Cnt))
3736 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3739 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3740 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3741 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3744 // Don't do anything for most intrinsics.
3747 // Vector shifts: check for immediate versions and lower them.
3748 // Note: This is done during DAG combining instead of DAG legalizing because
3749 // the build_vectors for 64-bit vector element shift counts are generally
3750 // not legal, and it is hard to see their values after they get legalized to
3751 // loads from a constant pool.
3752 case Intrinsic::arm_neon_vshifts:
3753 case Intrinsic::arm_neon_vshiftu:
3754 case Intrinsic::arm_neon_vshiftls:
3755 case Intrinsic::arm_neon_vshiftlu:
3756 case Intrinsic::arm_neon_vshiftn:
3757 case Intrinsic::arm_neon_vrshifts:
3758 case Intrinsic::arm_neon_vrshiftu:
3759 case Intrinsic::arm_neon_vrshiftn:
3760 case Intrinsic::arm_neon_vqshifts:
3761 case Intrinsic::arm_neon_vqshiftu:
3762 case Intrinsic::arm_neon_vqshiftsu:
3763 case Intrinsic::arm_neon_vqshiftns:
3764 case Intrinsic::arm_neon_vqshiftnu:
3765 case Intrinsic::arm_neon_vqshiftnsu:
3766 case Intrinsic::arm_neon_vqrshiftns:
3767 case Intrinsic::arm_neon_vqrshiftnu:
3768 case Intrinsic::arm_neon_vqrshiftnsu: {
3769 EVT VT = N->getOperand(1).getValueType();
3771 unsigned VShiftOpc = 0;
3774 case Intrinsic::arm_neon_vshifts:
3775 case Intrinsic::arm_neon_vshiftu:
3776 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3777 VShiftOpc = ARMISD::VSHL;
3780 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3781 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3782 ARMISD::VSHRs : ARMISD::VSHRu);
3787 case Intrinsic::arm_neon_vshiftls:
3788 case Intrinsic::arm_neon_vshiftlu:
3789 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3791 llvm_unreachable("invalid shift count for vshll intrinsic");
3793 case Intrinsic::arm_neon_vrshifts:
3794 case Intrinsic::arm_neon_vrshiftu:
3795 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3799 case Intrinsic::arm_neon_vqshifts:
3800 case Intrinsic::arm_neon_vqshiftu:
3801 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3805 case Intrinsic::arm_neon_vqshiftsu:
3806 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3808 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3810 case Intrinsic::arm_neon_vshiftn:
3811 case Intrinsic::arm_neon_vrshiftn:
3812 case Intrinsic::arm_neon_vqshiftns:
3813 case Intrinsic::arm_neon_vqshiftnu:
3814 case Intrinsic::arm_neon_vqshiftnsu:
3815 case Intrinsic::arm_neon_vqrshiftns:
3816 case Intrinsic::arm_neon_vqrshiftnu:
3817 case Intrinsic::arm_neon_vqrshiftnsu:
3818 // Narrowing shifts require an immediate right shift.
3819 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3821 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3824 llvm_unreachable("unhandled vector shift");
3828 case Intrinsic::arm_neon_vshifts:
3829 case Intrinsic::arm_neon_vshiftu:
3830 // Opcode already set above.
3832 case Intrinsic::arm_neon_vshiftls:
3833 case Intrinsic::arm_neon_vshiftlu:
3834 if (Cnt == VT.getVectorElementType().getSizeInBits())
3835 VShiftOpc = ARMISD::VSHLLi;
3837 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3838 ARMISD::VSHLLs : ARMISD::VSHLLu);
3840 case Intrinsic::arm_neon_vshiftn:
3841 VShiftOpc = ARMISD::VSHRN; break;
3842 case Intrinsic::arm_neon_vrshifts:
3843 VShiftOpc = ARMISD::VRSHRs; break;
3844 case Intrinsic::arm_neon_vrshiftu:
3845 VShiftOpc = ARMISD::VRSHRu; break;
3846 case Intrinsic::arm_neon_vrshiftn:
3847 VShiftOpc = ARMISD::VRSHRN; break;
3848 case Intrinsic::arm_neon_vqshifts:
3849 VShiftOpc = ARMISD::VQSHLs; break;
3850 case Intrinsic::arm_neon_vqshiftu:
3851 VShiftOpc = ARMISD::VQSHLu; break;
3852 case Intrinsic::arm_neon_vqshiftsu:
3853 VShiftOpc = ARMISD::VQSHLsu; break;
3854 case Intrinsic::arm_neon_vqshiftns:
3855 VShiftOpc = ARMISD::VQSHRNs; break;
3856 case Intrinsic::arm_neon_vqshiftnu:
3857 VShiftOpc = ARMISD::VQSHRNu; break;
3858 case Intrinsic::arm_neon_vqshiftnsu:
3859 VShiftOpc = ARMISD::VQSHRNsu; break;
3860 case Intrinsic::arm_neon_vqrshiftns:
3861 VShiftOpc = ARMISD::VQRSHRNs; break;
3862 case Intrinsic::arm_neon_vqrshiftnu:
3863 VShiftOpc = ARMISD::VQRSHRNu; break;
3864 case Intrinsic::arm_neon_vqrshiftnsu:
3865 VShiftOpc = ARMISD::VQRSHRNsu; break;
3868 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3869 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3872 case Intrinsic::arm_neon_vshiftins: {
3873 EVT VT = N->getOperand(1).getValueType();
3875 unsigned VShiftOpc = 0;
3877 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3878 VShiftOpc = ARMISD::VSLI;
3879 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3880 VShiftOpc = ARMISD::VSRI;
3882 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3885 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3886 N->getOperand(1), N->getOperand(2),
3887 DAG.getConstant(Cnt, MVT::i32));
3890 case Intrinsic::arm_neon_vqrshifts:
3891 case Intrinsic::arm_neon_vqrshiftu:
3892 // No immediate versions of these to check for.
3899 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3900 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3901 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3902 /// vector element shift counts are generally not legal, and it is hard to see
3903 /// their values after they get legalized to loads from a constant pool.
3904 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3905 const ARMSubtarget *ST) {
3906 EVT VT = N->getValueType(0);
3908 // Nothing to be done for scalar shifts.
3909 if (! VT.isVector())
3912 assert(ST->hasNEON() && "unexpected vector shift");
3915 switch (N->getOpcode()) {
3916 default: llvm_unreachable("unexpected shift opcode");
3919 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3920 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3921 DAG.getConstant(Cnt, MVT::i32));
3926 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3927 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3928 ARMISD::VSHRs : ARMISD::VSHRu);
3929 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3930 DAG.getConstant(Cnt, MVT::i32));
3936 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3937 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3938 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3939 const ARMSubtarget *ST) {
3940 SDValue N0 = N->getOperand(0);
3942 // Check for sign- and zero-extensions of vector extract operations of 8-
3943 // and 16-bit vector elements. NEON supports these directly. They are
3944 // handled during DAG combining because type legalization will promote them
3945 // to 32-bit types and it is messy to recognize the operations after that.
3946 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3947 SDValue Vec = N0.getOperand(0);
3948 SDValue Lane = N0.getOperand(1);
3949 EVT VT = N->getValueType(0);
3950 EVT EltVT = N0.getValueType();
3951 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3953 if (VT == MVT::i32 &&
3954 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3955 TLI.isTypeLegal(Vec.getValueType())) {
3958 switch (N->getOpcode()) {
3959 default: llvm_unreachable("unexpected opcode");
3960 case ISD::SIGN_EXTEND:
3961 Opc = ARMISD::VGETLANEs;
3963 case ISD::ZERO_EXTEND:
3964 case ISD::ANY_EXTEND:
3965 Opc = ARMISD::VGETLANEu;
3968 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3975 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3976 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3977 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3978 const ARMSubtarget *ST) {
3979 // If the target supports NEON, try to use vmax/vmin instructions for f32
3980 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3981 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3982 // a NaN; only do the transformation when it matches that behavior.
3984 // For now only do this when using NEON for FP operations; if using VFP, it
3985 // is not obvious that the benefit outweighs the cost of switching to the
3987 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3988 N->getValueType(0) != MVT::f32)
3991 SDValue CondLHS = N->getOperand(0);
3992 SDValue CondRHS = N->getOperand(1);
3993 SDValue LHS = N->getOperand(2);
3994 SDValue RHS = N->getOperand(3);
3995 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3997 unsigned Opcode = 0;
3999 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4000 IsReversed = false; // x CC y ? x : y
4001 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4002 IsReversed = true ; // x CC y ? y : x
4016 // If LHS is NaN, an ordered comparison will be false and the result will
4017 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4018 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4019 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4020 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4022 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4023 // will return -0, so vmin can only be used for unsafe math or if one of
4024 // the operands is known to be nonzero.
4025 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4027 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4029 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4038 // If LHS is NaN, an ordered comparison will be false and the result will
4039 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4040 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4041 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4042 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4044 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4045 // will return +0, so vmax can only be used for unsafe math or if one of
4046 // the operands is known to be nonzero.
4047 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4049 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4051 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4057 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4060 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4061 DAGCombinerInfo &DCI) const {
4062 switch (N->getOpcode()) {
4064 case ISD::ADD: return PerformADDCombine(N, DCI);
4065 case ISD::SUB: return PerformSUBCombine(N, DCI);
4066 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4067 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4070 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4071 case ISD::SIGN_EXTEND:
4072 case ISD::ZERO_EXTEND:
4073 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4074 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4079 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4080 if (!Subtarget->hasV6Ops())
4081 // Pre-v6 does not support unaligned mem access.
4084 // v6+ may or may not support unaligned mem access depending on the system
4086 // FIXME: This is pretty conservative. Should we provide cmdline option to
4087 // control the behaviour?
4088 if (!Subtarget->isTargetDarwin())
4092 switch (VT.getSimpleVT().SimpleTy) {
4099 // FIXME: VLD1 etc with standard alignment is legal.
4103 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4108 switch (VT.getSimpleVT().SimpleTy) {
4109 default: return false;
4124 if ((V & (Scale - 1)) != 0)
4127 return V == (V & ((1LL << 5) - 1));
4130 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4131 const ARMSubtarget *Subtarget) {
4138 switch (VT.getSimpleVT().SimpleTy) {
4139 default: return false;
4144 // + imm12 or - imm8
4146 return V == (V & ((1LL << 8) - 1));
4147 return V == (V & ((1LL << 12) - 1));
4150 // Same as ARM mode. FIXME: NEON?
4151 if (!Subtarget->hasVFP2())
4156 return V == (V & ((1LL << 8) - 1));
4160 /// isLegalAddressImmediate - Return true if the integer value can be used
4161 /// as the offset of the target addressing mode for load / store of the
4163 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4164 const ARMSubtarget *Subtarget) {
4171 if (Subtarget->isThumb1Only())
4172 return isLegalT1AddressImmediate(V, VT);
4173 else if (Subtarget->isThumb2())
4174 return isLegalT2AddressImmediate(V, VT, Subtarget);
4179 switch (VT.getSimpleVT().SimpleTy) {
4180 default: return false;
4185 return V == (V & ((1LL << 12) - 1));
4188 return V == (V & ((1LL << 8) - 1));
4191 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4196 return V == (V & ((1LL << 8) - 1));
4200 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4202 int Scale = AM.Scale;
4206 switch (VT.getSimpleVT().SimpleTy) {
4207 default: return false;
4216 return Scale == 2 || Scale == 4 || Scale == 8;
4219 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4223 // Note, we allow "void" uses (basically, uses that aren't loads or
4224 // stores), because arm allows folding a scale into many arithmetic
4225 // operations. This should be made more precise and revisited later.
4227 // Allow r << imm, but the imm has to be a multiple of two.
4228 if (Scale & 1) return false;
4229 return isPowerOf2_32(Scale);
4233 /// isLegalAddressingMode - Return true if the addressing mode represented
4234 /// by AM is legal for this target, for a load/store of the specified type.
4235 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4236 const Type *Ty) const {
4237 EVT VT = getValueType(Ty, true);
4238 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4241 // Can never fold addr of global into load/store.
4246 case 0: // no scale reg, must be "r+i" or "r", or "i".
4249 if (Subtarget->isThumb1Only())
4253 // ARM doesn't support any R+R*scale+imm addr modes.
4260 if (Subtarget->isThumb2())
4261 return isLegalT2ScaledAddressingMode(AM, VT);
4263 int Scale = AM.Scale;
4264 switch (VT.getSimpleVT().SimpleTy) {
4265 default: return false;
4269 if (Scale < 0) Scale = -Scale;
4273 return isPowerOf2_32(Scale & ~1);
4277 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4282 // Note, we allow "void" uses (basically, uses that aren't loads or
4283 // stores), because arm allows folding a scale into many arithmetic
4284 // operations. This should be made more precise and revisited later.
4286 // Allow r << imm, but the imm has to be a multiple of two.
4287 if (Scale & 1) return false;
4288 return isPowerOf2_32(Scale);
4295 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4296 /// icmp immediate, that is the target has icmp instructions which can compare
4297 /// a register against the immediate without having to materialize the
4298 /// immediate into a register.
4299 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4300 if (!Subtarget->isThumb())
4301 return ARM_AM::getSOImmVal(Imm) != -1;
4302 if (Subtarget->isThumb2())
4303 return ARM_AM::getT2SOImmVal(Imm) != -1;
4304 return Imm >= 0 && Imm <= 255;
4307 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4308 bool isSEXTLoad, SDValue &Base,
4309 SDValue &Offset, bool &isInc,
4310 SelectionDAG &DAG) {
4311 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4314 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4316 Base = Ptr->getOperand(0);
4317 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4318 int RHSC = (int)RHS->getZExtValue();
4319 if (RHSC < 0 && RHSC > -256) {
4320 assert(Ptr->getOpcode() == ISD::ADD);
4322 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4326 isInc = (Ptr->getOpcode() == ISD::ADD);
4327 Offset = Ptr->getOperand(1);
4329 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4331 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4332 int RHSC = (int)RHS->getZExtValue();
4333 if (RHSC < 0 && RHSC > -0x1000) {
4334 assert(Ptr->getOpcode() == ISD::ADD);
4336 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4337 Base = Ptr->getOperand(0);
4342 if (Ptr->getOpcode() == ISD::ADD) {
4344 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4345 if (ShOpcVal != ARM_AM::no_shift) {
4346 Base = Ptr->getOperand(1);
4347 Offset = Ptr->getOperand(0);
4349 Base = Ptr->getOperand(0);
4350 Offset = Ptr->getOperand(1);
4355 isInc = (Ptr->getOpcode() == ISD::ADD);
4356 Base = Ptr->getOperand(0);
4357 Offset = Ptr->getOperand(1);
4361 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4365 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4366 bool isSEXTLoad, SDValue &Base,
4367 SDValue &Offset, bool &isInc,
4368 SelectionDAG &DAG) {
4369 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4372 Base = Ptr->getOperand(0);
4373 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4374 int RHSC = (int)RHS->getZExtValue();
4375 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4376 assert(Ptr->getOpcode() == ISD::ADD);
4378 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4380 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4381 isInc = Ptr->getOpcode() == ISD::ADD;
4382 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4390 /// getPreIndexedAddressParts - returns true by value, base pointer and
4391 /// offset pointer and addressing mode by reference if the node's address
4392 /// can be legally represented as pre-indexed load / store address.
4394 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4396 ISD::MemIndexedMode &AM,
4397 SelectionDAG &DAG) const {
4398 if (Subtarget->isThumb1Only())
4403 bool isSEXTLoad = false;
4404 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4405 Ptr = LD->getBasePtr();
4406 VT = LD->getMemoryVT();
4407 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4408 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4409 Ptr = ST->getBasePtr();
4410 VT = ST->getMemoryVT();
4415 bool isLegal = false;
4416 if (Subtarget->isThumb2())
4417 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4418 Offset, isInc, DAG);
4420 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4421 Offset, isInc, DAG);
4425 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4429 /// getPostIndexedAddressParts - returns true by value, base pointer and
4430 /// offset pointer and addressing mode by reference if this node can be
4431 /// combined with a load / store to form a post-indexed load / store.
4432 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4435 ISD::MemIndexedMode &AM,
4436 SelectionDAG &DAG) const {
4437 if (Subtarget->isThumb1Only())
4442 bool isSEXTLoad = false;
4443 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4444 VT = LD->getMemoryVT();
4445 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4446 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4447 VT = ST->getMemoryVT();
4452 bool isLegal = false;
4453 if (Subtarget->isThumb2())
4454 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4457 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4462 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4466 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4470 const SelectionDAG &DAG,
4471 unsigned Depth) const {
4472 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4473 switch (Op.getOpcode()) {
4475 case ARMISD::CMOV: {
4476 // Bits are known zero/one if known on the LHS and RHS.
4477 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4478 if (KnownZero == 0 && KnownOne == 0) return;
4480 APInt KnownZeroRHS, KnownOneRHS;
4481 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4482 KnownZeroRHS, KnownOneRHS, Depth+1);
4483 KnownZero &= KnownZeroRHS;
4484 KnownOne &= KnownOneRHS;
4490 //===----------------------------------------------------------------------===//
4491 // ARM Inline Assembly Support
4492 //===----------------------------------------------------------------------===//
4494 /// getConstraintType - Given a constraint letter, return the type of
4495 /// constraint it is for this target.
4496 ARMTargetLowering::ConstraintType
4497 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4498 if (Constraint.size() == 1) {
4499 switch (Constraint[0]) {
4501 case 'l': return C_RegisterClass;
4502 case 'w': return C_RegisterClass;
4505 return TargetLowering::getConstraintType(Constraint);
4508 std::pair<unsigned, const TargetRegisterClass*>
4509 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4511 if (Constraint.size() == 1) {
4512 // GCC ARM Constraint Letters
4513 switch (Constraint[0]) {
4515 if (Subtarget->isThumb())
4516 return std::make_pair(0U, ARM::tGPRRegisterClass);
4518 return std::make_pair(0U, ARM::GPRRegisterClass);
4520 return std::make_pair(0U, ARM::GPRRegisterClass);
4523 return std::make_pair(0U, ARM::SPRRegisterClass);
4524 if (VT.getSizeInBits() == 64)
4525 return std::make_pair(0U, ARM::DPRRegisterClass);
4526 if (VT.getSizeInBits() == 128)
4527 return std::make_pair(0U, ARM::QPRRegisterClass);
4531 if (StringRef("{cc}").equals_lower(Constraint))
4532 return std::make_pair(0U, ARM::CCRRegisterClass);
4534 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4537 std::vector<unsigned> ARMTargetLowering::
4538 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4540 if (Constraint.size() != 1)
4541 return std::vector<unsigned>();
4543 switch (Constraint[0]) { // GCC ARM Constraint Letters
4546 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4547 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4550 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4551 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4552 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4553 ARM::R12, ARM::LR, 0);
4556 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4557 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4558 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4559 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4560 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4561 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4562 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4563 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4564 if (VT.getSizeInBits() == 64)
4565 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4566 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4567 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4568 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4569 if (VT.getSizeInBits() == 128)
4570 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4571 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4575 return std::vector<unsigned>();
4578 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4579 /// vector. If it is invalid, don't add anything to Ops.
4580 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4583 std::vector<SDValue>&Ops,
4584 SelectionDAG &DAG) const {
4585 SDValue Result(0, 0);
4587 switch (Constraint) {
4589 case 'I': case 'J': case 'K': case 'L':
4590 case 'M': case 'N': case 'O':
4591 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4595 int64_t CVal64 = C->getSExtValue();
4596 int CVal = (int) CVal64;
4597 // None of these constraints allow values larger than 32 bits. Check
4598 // that the value fits in an int.
4602 switch (Constraint) {
4604 if (Subtarget->isThumb1Only()) {
4605 // This must be a constant between 0 and 255, for ADD
4607 if (CVal >= 0 && CVal <= 255)
4609 } else if (Subtarget->isThumb2()) {
4610 // A constant that can be used as an immediate value in a
4611 // data-processing instruction.
4612 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4615 // A constant that can be used as an immediate value in a
4616 // data-processing instruction.
4617 if (ARM_AM::getSOImmVal(CVal) != -1)
4623 if (Subtarget->isThumb()) { // FIXME thumb2
4624 // This must be a constant between -255 and -1, for negated ADD
4625 // immediates. This can be used in GCC with an "n" modifier that
4626 // prints the negated value, for use with SUB instructions. It is
4627 // not useful otherwise but is implemented for compatibility.
4628 if (CVal >= -255 && CVal <= -1)
4631 // This must be a constant between -4095 and 4095. It is not clear
4632 // what this constraint is intended for. Implemented for
4633 // compatibility with GCC.
4634 if (CVal >= -4095 && CVal <= 4095)
4640 if (Subtarget->isThumb1Only()) {
4641 // A 32-bit value where only one byte has a nonzero value. Exclude
4642 // zero to match GCC. This constraint is used by GCC internally for
4643 // constants that can be loaded with a move/shift combination.
4644 // It is not useful otherwise but is implemented for compatibility.
4645 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4647 } else if (Subtarget->isThumb2()) {
4648 // A constant whose bitwise inverse can be used as an immediate
4649 // value in a data-processing instruction. This can be used in GCC
4650 // with a "B" modifier that prints the inverted value, for use with
4651 // BIC and MVN instructions. It is not useful otherwise but is
4652 // implemented for compatibility.
4653 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4656 // A constant whose bitwise inverse can be used as an immediate
4657 // value in a data-processing instruction. This can be used in GCC
4658 // with a "B" modifier that prints the inverted value, for use with
4659 // BIC and MVN instructions. It is not useful otherwise but is
4660 // implemented for compatibility.
4661 if (ARM_AM::getSOImmVal(~CVal) != -1)
4667 if (Subtarget->isThumb1Only()) {
4668 // This must be a constant between -7 and 7,
4669 // for 3-operand ADD/SUB immediate instructions.
4670 if (CVal >= -7 && CVal < 7)
4672 } else if (Subtarget->isThumb2()) {
4673 // A constant whose negation can be used as an immediate value in a
4674 // data-processing instruction. This can be used in GCC with an "n"
4675 // modifier that prints the negated value, for use with SUB
4676 // instructions. It is not useful otherwise but is implemented for
4678 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4681 // A constant whose negation can be used as an immediate value in a
4682 // data-processing instruction. This can be used in GCC with an "n"
4683 // modifier that prints the negated value, for use with SUB
4684 // instructions. It is not useful otherwise but is implemented for
4686 if (ARM_AM::getSOImmVal(-CVal) != -1)
4692 if (Subtarget->isThumb()) { // FIXME thumb2
4693 // This must be a multiple of 4 between 0 and 1020, for
4694 // ADD sp + immediate.
4695 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4698 // A power of two or a constant between 0 and 32. This is used in
4699 // GCC for the shift amount on shifted register operands, but it is
4700 // useful in general for any shift amounts.
4701 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4707 if (Subtarget->isThumb()) { // FIXME thumb2
4708 // This must be a constant between 0 and 31, for shift amounts.
4709 if (CVal >= 0 && CVal <= 31)
4715 if (Subtarget->isThumb()) { // FIXME thumb2
4716 // This must be a multiple of 4 between -508 and 508, for
4717 // ADD/SUB sp = sp + immediate.
4718 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4723 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4727 if (Result.getNode()) {
4728 Ops.push_back(Result);
4731 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4736 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4737 // The ARM target isn't yet aware of offsets.
4741 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4742 APInt Imm = FPImm.bitcastToAPInt();
4743 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4744 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4745 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4747 // We can handle 4 bits of mantissa.
4748 // mantissa = (16+UInt(e:f:g:h))/16.
4749 if (Mantissa & 0x7ffff)
4752 if ((Mantissa & 0xf) != Mantissa)
4755 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4756 if (Exp < -3 || Exp > 4)
4758 Exp = ((Exp+3) & 0x7) ^ 4;
4760 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4763 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4764 APInt Imm = FPImm.bitcastToAPInt();
4765 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4766 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4767 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4769 // We can handle 4 bits of mantissa.
4770 // mantissa = (16+UInt(e:f:g:h))/16.
4771 if (Mantissa & 0xffffffffffffLL)
4774 if ((Mantissa & 0xf) != Mantissa)
4777 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4778 if (Exp < -3 || Exp > 4)
4780 Exp = ((Exp+3) & 0x7) ^ 4;
4782 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4785 /// isFPImmLegal - Returns true if the target can instruction select the
4786 /// specified FP immediate natively. If false, the legalizer will
4787 /// materialize the FP immediate as a load from a constant pool.
4788 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4789 if (!Subtarget->hasVFP3())
4792 return ARM::getVFPf32Imm(Imm) != -1;
4794 return ARM::getVFPf64Imm(Imm) != -1;