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 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMAddressingModes.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Type.h"
33 #include "llvm/CodeGen/CallingConvLower.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineInstrBuilder.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/CodeGen/PseudoSourceValue.h"
40 #include "llvm/CodeGen/SelectionDAG.h"
41 #include "llvm/MC/MCSectionMachO.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/ADT/VectorExtras.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
52 STATISTIC(NumTailCalls, "Number of tail calls");
54 // This option should go away when tail calls fully work.
56 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
57 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions."),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
70 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
71 CCValAssign::LocInfo &LocInfo,
72 ISD::ArgFlagsTy &ArgFlags,
74 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
75 CCValAssign::LocInfo &LocInfo,
76 ISD::ArgFlagsTy &ArgFlags,
78 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
79 CCValAssign::LocInfo &LocInfo,
80 ISD::ArgFlagsTy &ArgFlags,
82 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
83 CCValAssign::LocInfo &LocInfo,
84 ISD::ArgFlagsTy &ArgFlags,
87 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
88 EVT PromotedBitwiseVT) {
89 if (VT != PromotedLdStVT) {
90 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
91 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
92 PromotedLdStVT.getSimpleVT());
94 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
95 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
96 PromotedLdStVT.getSimpleVT());
99 EVT ElemTy = VT.getVectorElementType();
100 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
101 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
102 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
103 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
104 if (ElemTy != MVT::i32) {
105 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
106 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
108 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
110 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
112 if (llvm::ModelWithRegSequence())
113 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
115 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
116 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
119 if (VT.isInteger()) {
120 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
121 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
122 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
125 // Promote all bit-wise operations.
126 if (VT.isInteger() && VT != PromotedBitwiseVT) {
127 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
128 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
129 PromotedBitwiseVT.getSimpleVT());
130 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
131 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
132 PromotedBitwiseVT.getSimpleVT());
133 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
134 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
135 PromotedBitwiseVT.getSimpleVT());
138 // Neon does not support vector divide/remainder operations.
139 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
141 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
142 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
143 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
144 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
147 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
148 addRegisterClass(VT, ARM::DPRRegisterClass);
149 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
152 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
153 addRegisterClass(VT, ARM::QPRRegisterClass);
154 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
157 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
158 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
159 return new TargetLoweringObjectFileMachO();
161 return new ARMElfTargetObjectFile();
164 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
165 : TargetLowering(TM, createTLOF(TM)) {
166 Subtarget = &TM.getSubtarget<ARMSubtarget>();
168 if (Subtarget->isTargetDarwin()) {
169 // Uses VFP for Thumb libfuncs if available.
170 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
171 // Single-precision floating-point arithmetic.
172 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
173 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
174 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
175 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
177 // Double-precision floating-point arithmetic.
178 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
179 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
180 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
181 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
183 // Single-precision comparisons.
184 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
185 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
186 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
187 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
188 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
189 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
190 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
191 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
193 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
202 // Double-precision comparisons.
203 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
204 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
205 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
206 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
207 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
208 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
209 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
210 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
212 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
221 // Floating-point to integer conversions.
222 // i64 conversions are done via library routines even when generating VFP
223 // instructions, so use the same ones.
224 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
225 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
226 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
227 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
229 // Conversions between floating types.
230 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
231 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
233 // Integer to floating-point conversions.
234 // i64 conversions are done via library routines even when generating VFP
235 // instructions, so use the same ones.
236 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
237 // e.g., __floatunsidf vs. __floatunssidfvfp.
238 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
239 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
240 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
241 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
245 // These libcalls are not available in 32-bit.
246 setLibcallName(RTLIB::SHL_I128, 0);
247 setLibcallName(RTLIB::SRL_I128, 0);
248 setLibcallName(RTLIB::SRA_I128, 0);
250 // Libcalls should use the AAPCS base standard ABI, even if hard float
251 // is in effect, as per the ARM RTABI specification, section 4.1.2.
252 if (Subtarget->isAAPCS_ABI()) {
253 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
254 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
255 CallingConv::ARM_AAPCS);
259 if (Subtarget->isThumb1Only())
260 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
262 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
263 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
264 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
265 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
267 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
270 if (Subtarget->hasNEON()) {
271 addDRTypeForNEON(MVT::v2f32);
272 addDRTypeForNEON(MVT::v8i8);
273 addDRTypeForNEON(MVT::v4i16);
274 addDRTypeForNEON(MVT::v2i32);
275 addDRTypeForNEON(MVT::v1i64);
277 addQRTypeForNEON(MVT::v4f32);
278 addQRTypeForNEON(MVT::v2f64);
279 addQRTypeForNEON(MVT::v16i8);
280 addQRTypeForNEON(MVT::v8i16);
281 addQRTypeForNEON(MVT::v4i32);
282 addQRTypeForNEON(MVT::v2i64);
284 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
285 // neither Neon nor VFP support any arithmetic operations on it.
286 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
287 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
288 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
289 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
290 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
291 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
292 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
293 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
294 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
295 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
296 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
297 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
298 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
299 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
300 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
301 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
302 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
303 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
304 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
305 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
306 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
307 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
308 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
309 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
311 // Neon does not support some operations on v1i64 and v2i64 types.
312 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
313 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
314 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
315 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
317 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
318 setTargetDAGCombine(ISD::SHL);
319 setTargetDAGCombine(ISD::SRL);
320 setTargetDAGCombine(ISD::SRA);
321 setTargetDAGCombine(ISD::SIGN_EXTEND);
322 setTargetDAGCombine(ISD::ZERO_EXTEND);
323 setTargetDAGCombine(ISD::ANY_EXTEND);
324 setTargetDAGCombine(ISD::SELECT_CC);
327 computeRegisterProperties();
329 // ARM does not have f32 extending load.
330 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
332 // ARM does not have i1 sign extending load.
333 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
335 // ARM supports all 4 flavors of integer indexed load / store.
336 if (!Subtarget->isThumb1Only()) {
337 for (unsigned im = (unsigned)ISD::PRE_INC;
338 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
339 setIndexedLoadAction(im, MVT::i1, Legal);
340 setIndexedLoadAction(im, MVT::i8, Legal);
341 setIndexedLoadAction(im, MVT::i16, Legal);
342 setIndexedLoadAction(im, MVT::i32, Legal);
343 setIndexedStoreAction(im, MVT::i1, Legal);
344 setIndexedStoreAction(im, MVT::i8, Legal);
345 setIndexedStoreAction(im, MVT::i16, Legal);
346 setIndexedStoreAction(im, MVT::i32, Legal);
350 // i64 operation support.
351 if (Subtarget->isThumb1Only()) {
352 setOperationAction(ISD::MUL, MVT::i64, Expand);
353 setOperationAction(ISD::MULHU, MVT::i32, Expand);
354 setOperationAction(ISD::MULHS, MVT::i32, Expand);
355 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
356 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
358 setOperationAction(ISD::MUL, MVT::i64, Expand);
359 setOperationAction(ISD::MULHU, MVT::i32, Expand);
360 if (!Subtarget->hasV6Ops())
361 setOperationAction(ISD::MULHS, MVT::i32, Expand);
363 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
364 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
365 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
366 setOperationAction(ISD::SRL, MVT::i64, Custom);
367 setOperationAction(ISD::SRA, MVT::i64, Custom);
369 // ARM does not have ROTL.
370 setOperationAction(ISD::ROTL, MVT::i32, Expand);
371 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
372 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
373 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
374 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
376 // Only ARMv6 has BSWAP.
377 if (!Subtarget->hasV6Ops())
378 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
380 // These are expanded into libcalls.
381 if (!Subtarget->hasDivide()) {
382 // v7M has a hardware divider
383 setOperationAction(ISD::SDIV, MVT::i32, Expand);
384 setOperationAction(ISD::UDIV, MVT::i32, Expand);
386 setOperationAction(ISD::SREM, MVT::i32, Expand);
387 setOperationAction(ISD::UREM, MVT::i32, Expand);
388 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
389 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
391 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
392 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
393 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
394 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
395 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
397 setOperationAction(ISD::TRAP, MVT::Other, Legal);
399 // Use the default implementation.
400 setOperationAction(ISD::VASTART, MVT::Other, Custom);
401 setOperationAction(ISD::VAARG, MVT::Other, Expand);
402 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
403 setOperationAction(ISD::VAEND, MVT::Other, Expand);
404 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
405 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
406 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
407 // FIXME: Shouldn't need this, since no register is used, but the legalizer
408 // doesn't yet know how to not do that for SjLj.
409 setExceptionSelectorRegister(ARM::R0);
410 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
411 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
413 // If the subtarget does not have extract instructions, sign_extend_inreg
414 // needs to be expanded. Extract is available in ARM mode on v6 and up,
415 // and on most Thumb2 implementations.
416 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
417 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
418 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
419 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
421 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
423 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
424 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
425 // iff target supports vfp2.
426 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
428 // We want to custom lower some of our intrinsics.
429 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
431 setOperationAction(ISD::SETCC, MVT::i32, Expand);
432 setOperationAction(ISD::SETCC, MVT::f32, Expand);
433 setOperationAction(ISD::SETCC, MVT::f64, Expand);
434 setOperationAction(ISD::SELECT, MVT::i32, Expand);
435 setOperationAction(ISD::SELECT, MVT::f32, Expand);
436 setOperationAction(ISD::SELECT, MVT::f64, Expand);
437 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
438 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
439 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
441 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
442 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
443 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
444 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
445 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
447 // We don't support sin/cos/fmod/copysign/pow
448 setOperationAction(ISD::FSIN, MVT::f64, Expand);
449 setOperationAction(ISD::FSIN, MVT::f32, Expand);
450 setOperationAction(ISD::FCOS, MVT::f32, Expand);
451 setOperationAction(ISD::FCOS, MVT::f64, Expand);
452 setOperationAction(ISD::FREM, MVT::f64, Expand);
453 setOperationAction(ISD::FREM, MVT::f32, Expand);
454 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
455 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
456 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
458 setOperationAction(ISD::FPOW, MVT::f64, Expand);
459 setOperationAction(ISD::FPOW, MVT::f32, Expand);
461 // Various VFP goodness
462 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
463 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
464 if (Subtarget->hasVFP2()) {
465 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
466 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
467 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
468 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
470 // Special handling for half-precision FP.
471 if (!Subtarget->hasFP16()) {
472 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
473 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
477 // We have target-specific dag combine patterns for the following nodes:
478 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
479 setTargetDAGCombine(ISD::ADD);
480 setTargetDAGCombine(ISD::SUB);
481 setTargetDAGCombine(ISD::MUL);
483 setStackPointerRegisterToSaveRestore(ARM::SP);
485 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
486 setSchedulingPreference(Sched::RegPressure);
488 setSchedulingPreference(Sched::Hybrid);
490 // FIXME: If-converter should use instruction latency to determine
491 // profitability rather than relying on fixed limits.
492 if (Subtarget->getCPUString() == "generic") {
493 // Generic (and overly aggressive) if-conversion limits.
494 setIfCvtBlockSizeLimit(10);
495 setIfCvtDupBlockSizeLimit(2);
496 } else if (Subtarget->hasV7Ops()) {
497 setIfCvtBlockSizeLimit(3);
498 setIfCvtDupBlockSizeLimit(1);
499 } else if (Subtarget->hasV6Ops()) {
500 setIfCvtBlockSizeLimit(2);
501 setIfCvtDupBlockSizeLimit(1);
503 setIfCvtBlockSizeLimit(3);
504 setIfCvtDupBlockSizeLimit(2);
507 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
508 // Do not enable CodePlacementOpt for now: it currently runs after the
509 // ARMConstantIslandPass and messes up branch relaxation and placement
510 // of constant islands.
511 // benefitFromCodePlacementOpt = true;
514 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
517 case ARMISD::Wrapper: return "ARMISD::Wrapper";
518 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
519 case ARMISD::CALL: return "ARMISD::CALL";
520 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
521 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
522 case ARMISD::tCALL: return "ARMISD::tCALL";
523 case ARMISD::BRCOND: return "ARMISD::BRCOND";
524 case ARMISD::BR_JT: return "ARMISD::BR_JT";
525 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
526 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
527 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
528 case ARMISD::CMP: return "ARMISD::CMP";
529 case ARMISD::CMPZ: return "ARMISD::CMPZ";
530 case ARMISD::CMPFP: return "ARMISD::CMPFP";
531 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
532 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
533 case ARMISD::CMOV: return "ARMISD::CMOV";
534 case ARMISD::CNEG: return "ARMISD::CNEG";
536 case ARMISD::RBIT: return "ARMISD::RBIT";
538 case ARMISD::FTOSI: return "ARMISD::FTOSI";
539 case ARMISD::FTOUI: return "ARMISD::FTOUI";
540 case ARMISD::SITOF: return "ARMISD::SITOF";
541 case ARMISD::UITOF: return "ARMISD::UITOF";
543 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
544 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
545 case ARMISD::RRX: return "ARMISD::RRX";
547 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
548 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
550 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
551 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
553 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
555 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
557 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
559 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
560 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
562 case ARMISD::VCEQ: return "ARMISD::VCEQ";
563 case ARMISD::VCGE: return "ARMISD::VCGE";
564 case ARMISD::VCGEU: return "ARMISD::VCGEU";
565 case ARMISD::VCGT: return "ARMISD::VCGT";
566 case ARMISD::VCGTU: return "ARMISD::VCGTU";
567 case ARMISD::VTST: return "ARMISD::VTST";
569 case ARMISD::VSHL: return "ARMISD::VSHL";
570 case ARMISD::VSHRs: return "ARMISD::VSHRs";
571 case ARMISD::VSHRu: return "ARMISD::VSHRu";
572 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
573 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
574 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
575 case ARMISD::VSHRN: return "ARMISD::VSHRN";
576 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
577 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
578 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
579 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
580 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
581 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
582 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
583 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
584 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
585 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
586 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
587 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
588 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
589 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
590 case ARMISD::VDUP: return "ARMISD::VDUP";
591 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
592 case ARMISD::VEXT: return "ARMISD::VEXT";
593 case ARMISD::VREV64: return "ARMISD::VREV64";
594 case ARMISD::VREV32: return "ARMISD::VREV32";
595 case ARMISD::VREV16: return "ARMISD::VREV16";
596 case ARMISD::VZIP: return "ARMISD::VZIP";
597 case ARMISD::VUZP: return "ARMISD::VUZP";
598 case ARMISD::VTRN: return "ARMISD::VTRN";
599 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
600 case ARMISD::FMAX: return "ARMISD::FMAX";
601 case ARMISD::FMIN: return "ARMISD::FMIN";
605 /// getRegClassFor - Return the register class that should be used for the
606 /// specified value type.
607 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
608 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
609 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
610 // load / store 4 to 8 consecutive D registers.
611 if (Subtarget->hasNEON()) {
612 if (VT == MVT::v4i64)
613 return ARM::QQPRRegisterClass;
614 else if (VT == MVT::v8i64)
615 return ARM::QQQQPRRegisterClass;
617 return TargetLowering::getRegClassFor(VT);
620 /// getFunctionAlignment - Return the Log2 alignment of this function.
621 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
622 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
625 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
626 unsigned NumVals = N->getNumValues();
628 return Sched::RegPressure;
630 for (unsigned i = 0; i != NumVals; ++i) {
631 EVT VT = N->getValueType(i);
632 if (VT.isFloatingPoint() || VT.isVector())
633 return Sched::Latency;
636 if (!N->isMachineOpcode())
637 return Sched::RegPressure;
639 // Load are scheduled for latency even if there instruction itinerary
641 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
642 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
644 return Sched::Latency;
646 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
647 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
648 return Sched::Latency;
649 return Sched::RegPressure;
652 //===----------------------------------------------------------------------===//
654 //===----------------------------------------------------------------------===//
656 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
657 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
659 default: llvm_unreachable("Unknown condition code!");
660 case ISD::SETNE: return ARMCC::NE;
661 case ISD::SETEQ: return ARMCC::EQ;
662 case ISD::SETGT: return ARMCC::GT;
663 case ISD::SETGE: return ARMCC::GE;
664 case ISD::SETLT: return ARMCC::LT;
665 case ISD::SETLE: return ARMCC::LE;
666 case ISD::SETUGT: return ARMCC::HI;
667 case ISD::SETUGE: return ARMCC::HS;
668 case ISD::SETULT: return ARMCC::LO;
669 case ISD::SETULE: return ARMCC::LS;
673 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
674 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
675 ARMCC::CondCodes &CondCode2) {
676 CondCode2 = ARMCC::AL;
678 default: llvm_unreachable("Unknown FP condition!");
680 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
682 case ISD::SETOGT: CondCode = ARMCC::GT; break;
684 case ISD::SETOGE: CondCode = ARMCC::GE; break;
685 case ISD::SETOLT: CondCode = ARMCC::MI; break;
686 case ISD::SETOLE: CondCode = ARMCC::LS; break;
687 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
688 case ISD::SETO: CondCode = ARMCC::VC; break;
689 case ISD::SETUO: CondCode = ARMCC::VS; break;
690 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
691 case ISD::SETUGT: CondCode = ARMCC::HI; break;
692 case ISD::SETUGE: CondCode = ARMCC::PL; break;
694 case ISD::SETULT: CondCode = ARMCC::LT; break;
696 case ISD::SETULE: CondCode = ARMCC::LE; break;
698 case ISD::SETUNE: CondCode = ARMCC::NE; break;
702 //===----------------------------------------------------------------------===//
703 // Calling Convention Implementation
704 //===----------------------------------------------------------------------===//
706 #include "ARMGenCallingConv.inc"
708 // APCS f64 is in register pairs, possibly split to stack
709 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
710 CCValAssign::LocInfo &LocInfo,
711 CCState &State, bool CanFail) {
712 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
714 // Try to get the first register.
715 if (unsigned Reg = State.AllocateReg(RegList, 4))
716 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
718 // For the 2nd half of a v2f64, do not fail.
722 // Put the whole thing on the stack.
723 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
724 State.AllocateStack(8, 4),
729 // Try to get the second register.
730 if (unsigned Reg = State.AllocateReg(RegList, 4))
731 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
733 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
734 State.AllocateStack(4, 4),
739 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
740 CCValAssign::LocInfo &LocInfo,
741 ISD::ArgFlagsTy &ArgFlags,
743 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
745 if (LocVT == MVT::v2f64 &&
746 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
748 return true; // we handled it
751 // AAPCS f64 is in aligned register pairs
752 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
753 CCValAssign::LocInfo &LocInfo,
754 CCState &State, bool CanFail) {
755 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
756 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
758 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
760 // For the 2nd half of a v2f64, do not just fail.
764 // Put the whole thing on the stack.
765 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
766 State.AllocateStack(8, 8),
772 for (i = 0; i < 2; ++i)
773 if (HiRegList[i] == Reg)
776 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
777 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
782 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
783 CCValAssign::LocInfo &LocInfo,
784 ISD::ArgFlagsTy &ArgFlags,
786 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
788 if (LocVT == MVT::v2f64 &&
789 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
791 return true; // we handled it
794 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
795 CCValAssign::LocInfo &LocInfo, CCState &State) {
796 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
797 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
799 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
801 return false; // we didn't handle it
804 for (i = 0; i < 2; ++i)
805 if (HiRegList[i] == Reg)
808 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
809 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
814 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
815 CCValAssign::LocInfo &LocInfo,
816 ISD::ArgFlagsTy &ArgFlags,
818 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
820 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
822 return true; // we handled it
825 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
826 CCValAssign::LocInfo &LocInfo,
827 ISD::ArgFlagsTy &ArgFlags,
829 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
833 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
834 /// given CallingConvention value.
835 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
837 bool isVarArg) const {
840 llvm_unreachable("Unsupported calling convention");
842 case CallingConv::Fast:
843 // Use target triple & subtarget features to do actual dispatch.
844 if (Subtarget->isAAPCS_ABI()) {
845 if (Subtarget->hasVFP2() &&
846 FloatABIType == FloatABI::Hard && !isVarArg)
847 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
849 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
851 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
852 case CallingConv::ARM_AAPCS_VFP:
853 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
854 case CallingConv::ARM_AAPCS:
855 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
856 case CallingConv::ARM_APCS:
857 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
861 /// LowerCallResult - Lower the result values of a call into the
862 /// appropriate copies out of appropriate physical registers.
864 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
865 CallingConv::ID CallConv, bool isVarArg,
866 const SmallVectorImpl<ISD::InputArg> &Ins,
867 DebugLoc dl, SelectionDAG &DAG,
868 SmallVectorImpl<SDValue> &InVals) const {
870 // Assign locations to each value returned by this call.
871 SmallVector<CCValAssign, 16> RVLocs;
872 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
873 RVLocs, *DAG.getContext());
874 CCInfo.AnalyzeCallResult(Ins,
875 CCAssignFnForNode(CallConv, /* Return*/ true,
878 // Copy all of the result registers out of their specified physreg.
879 for (unsigned i = 0; i != RVLocs.size(); ++i) {
880 CCValAssign VA = RVLocs[i];
883 if (VA.needsCustom()) {
884 // Handle f64 or half of a v2f64.
885 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
887 Chain = Lo.getValue(1);
888 InFlag = Lo.getValue(2);
889 VA = RVLocs[++i]; // skip ahead to next loc
890 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
892 Chain = Hi.getValue(1);
893 InFlag = Hi.getValue(2);
894 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
896 if (VA.getLocVT() == MVT::v2f64) {
897 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
898 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
899 DAG.getConstant(0, MVT::i32));
901 VA = RVLocs[++i]; // skip ahead to next loc
902 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
903 Chain = Lo.getValue(1);
904 InFlag = Lo.getValue(2);
905 VA = RVLocs[++i]; // skip ahead to next loc
906 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
907 Chain = Hi.getValue(1);
908 InFlag = Hi.getValue(2);
909 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
910 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
911 DAG.getConstant(1, MVT::i32));
914 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
916 Chain = Val.getValue(1);
917 InFlag = Val.getValue(2);
920 switch (VA.getLocInfo()) {
921 default: llvm_unreachable("Unknown loc info!");
922 case CCValAssign::Full: break;
923 case CCValAssign::BCvt:
924 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
928 InVals.push_back(Val);
934 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
935 /// by "Src" to address "Dst" of size "Size". Alignment information is
936 /// specified by the specific parameter attribute. The copy will be passed as
937 /// a byval function parameter.
938 /// Sometimes what we are copying is the end of a larger object, the part that
939 /// does not fit in registers.
941 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
942 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
944 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
945 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
946 /*isVolatile=*/false, /*AlwaysInline=*/false,
950 /// LowerMemOpCallTo - Store the argument to the stack.
952 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
953 SDValue StackPtr, SDValue Arg,
954 DebugLoc dl, SelectionDAG &DAG,
955 const CCValAssign &VA,
956 ISD::ArgFlagsTy Flags) const {
957 unsigned LocMemOffset = VA.getLocMemOffset();
958 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
959 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
960 if (Flags.isByVal()) {
961 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
963 return DAG.getStore(Chain, dl, Arg, PtrOff,
964 PseudoSourceValue::getStack(), LocMemOffset,
968 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
969 SDValue Chain, SDValue &Arg,
970 RegsToPassVector &RegsToPass,
971 CCValAssign &VA, CCValAssign &NextVA,
973 SmallVector<SDValue, 8> &MemOpChains,
974 ISD::ArgFlagsTy Flags) const {
976 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
977 DAG.getVTList(MVT::i32, MVT::i32), Arg);
978 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
980 if (NextVA.isRegLoc())
981 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
983 assert(NextVA.isMemLoc());
984 if (StackPtr.getNode() == 0)
985 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
987 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
993 /// LowerCall - Lowering a call into a callseq_start <-
994 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
997 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
998 CallingConv::ID CallConv, bool isVarArg,
1000 const SmallVectorImpl<ISD::OutputArg> &Outs,
1001 const SmallVectorImpl<ISD::InputArg> &Ins,
1002 DebugLoc dl, SelectionDAG &DAG,
1003 SmallVectorImpl<SDValue> &InVals) const {
1004 MachineFunction &MF = DAG.getMachineFunction();
1005 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1006 bool IsSibCall = false;
1007 // Temporarily disable tail calls so things don't break.
1008 if (!EnableARMTailCalls)
1011 // Check if it's really possible to do a tail call.
1012 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1013 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1015 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1016 // detected sibcalls.
1023 // Analyze operands of the call, assigning locations to each operand.
1024 SmallVector<CCValAssign, 16> ArgLocs;
1025 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1027 CCInfo.AnalyzeCallOperands(Outs,
1028 CCAssignFnForNode(CallConv, /* Return*/ false,
1031 // Get a count of how many bytes are to be pushed on the stack.
1032 unsigned NumBytes = CCInfo.getNextStackOffset();
1034 // For tail calls, memory operands are available in our caller's stack.
1038 // Adjust the stack pointer for the new arguments...
1039 // These operations are automatically eliminated by the prolog/epilog pass
1041 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1043 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1045 RegsToPassVector RegsToPass;
1046 SmallVector<SDValue, 8> MemOpChains;
1048 // Walk the register/memloc assignments, inserting copies/loads. In the case
1049 // of tail call optimization, arguments are handled later.
1050 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1052 ++i, ++realArgIdx) {
1053 CCValAssign &VA = ArgLocs[i];
1054 SDValue Arg = Outs[realArgIdx].Val;
1055 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1057 // Promote the value if needed.
1058 switch (VA.getLocInfo()) {
1059 default: llvm_unreachable("Unknown loc info!");
1060 case CCValAssign::Full: break;
1061 case CCValAssign::SExt:
1062 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1064 case CCValAssign::ZExt:
1065 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1067 case CCValAssign::AExt:
1068 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1070 case CCValAssign::BCvt:
1071 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1075 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1076 if (VA.needsCustom()) {
1077 if (VA.getLocVT() == MVT::v2f64) {
1078 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1079 DAG.getConstant(0, MVT::i32));
1080 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1081 DAG.getConstant(1, MVT::i32));
1083 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1084 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1086 VA = ArgLocs[++i]; // skip ahead to next loc
1087 if (VA.isRegLoc()) {
1088 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1089 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1091 assert(VA.isMemLoc());
1093 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1094 dl, DAG, VA, Flags));
1097 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1098 StackPtr, MemOpChains, Flags);
1100 } else if (VA.isRegLoc()) {
1101 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1103 assert(VA.isMemLoc());
1105 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1106 dl, DAG, VA, Flags));
1110 if (!MemOpChains.empty())
1111 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1112 &MemOpChains[0], MemOpChains.size());
1114 // Build a sequence of copy-to-reg nodes chained together with token chain
1115 // and flag operands which copy the outgoing args into the appropriate regs.
1117 // Tail call byval lowering might overwrite argument registers so in case of
1118 // tail call optimization the copies to registers are lowered later.
1120 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1121 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1122 RegsToPass[i].second, InFlag);
1123 InFlag = Chain.getValue(1);
1126 // For tail calls lower the arguments to the 'real' stack slot.
1128 // Force all the incoming stack arguments to be loaded from the stack
1129 // before any new outgoing arguments are stored to the stack, because the
1130 // outgoing stack slots may alias the incoming argument stack slots, and
1131 // the alias isn't otherwise explicit. This is slightly more conservative
1132 // than necessary, because it means that each store effectively depends
1133 // on every argument instead of just those arguments it would clobber.
1135 // Do not flag preceeding copytoreg stuff together with the following stuff.
1137 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1138 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1139 RegsToPass[i].second, InFlag);
1140 InFlag = Chain.getValue(1);
1145 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1146 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1147 // node so that legalize doesn't hack it.
1148 bool isDirect = false;
1149 bool isARMFunc = false;
1150 bool isLocalARMFunc = false;
1151 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1153 if (EnableARMLongCalls) {
1154 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1155 && "long-calls with non-static relocation model!");
1156 // Handle a global address or an external symbol. If it's not one of
1157 // those, the target's already in a register, so we don't need to do
1159 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1160 const GlobalValue *GV = G->getGlobal();
1161 // Create a constant pool entry for the callee address
1162 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1163 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1166 // Get the address of the callee into a register
1167 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1168 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1169 Callee = DAG.getLoad(getPointerTy(), dl,
1170 DAG.getEntryNode(), CPAddr,
1171 PseudoSourceValue::getConstantPool(), 0,
1173 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1174 const char *Sym = S->getSymbol();
1176 // Create a constant pool entry for the callee address
1177 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1178 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1179 Sym, ARMPCLabelIndex, 0);
1180 // Get the address of the callee into a register
1181 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1182 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1183 Callee = DAG.getLoad(getPointerTy(), dl,
1184 DAG.getEntryNode(), CPAddr,
1185 PseudoSourceValue::getConstantPool(), 0,
1188 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1189 const GlobalValue *GV = G->getGlobal();
1191 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1192 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1193 getTargetMachine().getRelocationModel() != Reloc::Static;
1194 isARMFunc = !Subtarget->isThumb() || isStub;
1195 // ARM call to a local ARM function is predicable.
1196 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1197 // tBX takes a register source operand.
1198 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1199 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1200 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1203 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1204 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1205 Callee = DAG.getLoad(getPointerTy(), dl,
1206 DAG.getEntryNode(), CPAddr,
1207 PseudoSourceValue::getConstantPool(), 0,
1209 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1210 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1211 getPointerTy(), Callee, PICLabel);
1213 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1214 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1216 bool isStub = Subtarget->isTargetDarwin() &&
1217 getTargetMachine().getRelocationModel() != Reloc::Static;
1218 isARMFunc = !Subtarget->isThumb() || isStub;
1219 // tBX takes a register source operand.
1220 const char *Sym = S->getSymbol();
1221 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1222 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1223 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1224 Sym, ARMPCLabelIndex, 4);
1225 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1226 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1227 Callee = DAG.getLoad(getPointerTy(), dl,
1228 DAG.getEntryNode(), CPAddr,
1229 PseudoSourceValue::getConstantPool(), 0,
1231 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1232 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1233 getPointerTy(), Callee, PICLabel);
1235 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1238 // FIXME: handle tail calls differently.
1240 if (Subtarget->isThumb()) {
1241 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1242 CallOpc = ARMISD::CALL_NOLINK;
1244 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1246 CallOpc = (isDirect || Subtarget->hasV5TOps())
1247 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1248 : ARMISD::CALL_NOLINK;
1250 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1251 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1252 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1253 InFlag = Chain.getValue(1);
1256 std::vector<SDValue> Ops;
1257 Ops.push_back(Chain);
1258 Ops.push_back(Callee);
1260 // Add argument registers to the end of the list so that they are known live
1262 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1263 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1264 RegsToPass[i].second.getValueType()));
1266 if (InFlag.getNode())
1267 Ops.push_back(InFlag);
1269 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1271 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1273 // Returns a chain and a flag for retval copy to use.
1274 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1275 InFlag = Chain.getValue(1);
1277 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1278 DAG.getIntPtrConstant(0, true), InFlag);
1280 InFlag = Chain.getValue(1);
1282 // Handle result values, copying them out of physregs into vregs that we
1284 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1288 /// MatchingStackOffset - Return true if the given stack call argument is
1289 /// already available in the same position (relatively) of the caller's
1290 /// incoming argument stack.
1292 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1293 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1294 const ARMInstrInfo *TII) {
1295 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1297 if (Arg.getOpcode() == ISD::CopyFromReg) {
1298 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1299 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1301 MachineInstr *Def = MRI->getVRegDef(VR);
1304 if (!Flags.isByVal()) {
1305 if (!TII->isLoadFromStackSlot(Def, FI))
1308 // unsigned Opcode = Def->getOpcode();
1309 // if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
1310 // Def->getOperand(1).isFI()) {
1311 // FI = Def->getOperand(1).getIndex();
1312 // Bytes = Flags.getByValSize();
1316 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1317 if (Flags.isByVal())
1318 // ByVal argument is passed in as a pointer but it's now being
1319 // dereferenced. e.g.
1320 // define @foo(%struct.X* %A) {
1321 // tail call @bar(%struct.X* byval %A)
1324 SDValue Ptr = Ld->getBasePtr();
1325 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1328 FI = FINode->getIndex();
1332 assert(FI != INT_MAX);
1333 if (!MFI->isFixedObjectIndex(FI))
1335 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1338 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1339 /// for tail call optimization. Targets which want to do tail call
1340 /// optimization should implement this function.
1342 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1343 CallingConv::ID CalleeCC,
1345 bool isCalleeStructRet,
1346 bool isCallerStructRet,
1347 const SmallVectorImpl<ISD::OutputArg> &Outs,
1348 const SmallVectorImpl<ISD::InputArg> &Ins,
1349 SelectionDAG& DAG) const {
1351 const Function *CallerF = DAG.getMachineFunction().getFunction();
1352 CallingConv::ID CallerCC = CallerF->getCallingConv();
1353 bool CCMatch = CallerCC == CalleeCC;
1355 // Look for obvious safe cases to perform tail call optimization that do not
1356 // require ABI changes. This is what gcc calls sibcall.
1358 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
1359 // emit a special epilogue.
1360 // Not sure yet if this is true on ARM.
1361 //?? if (RegInfo->needsStackRealignment(MF))
1364 // Do not sibcall optimize vararg calls unless the call site is not passing any
1366 if (isVarArg && !Outs.empty())
1369 // Also avoid sibcall optimization if either caller or callee uses struct
1370 // return semantics.
1371 if (isCalleeStructRet || isCallerStructRet)
1374 // If the calling conventions do not match, then we'd better make sure the
1375 // results are returned in the same way as what the caller expects.
1377 SmallVector<CCValAssign, 16> RVLocs1;
1378 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1379 RVLocs1, *DAG.getContext());
1380 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1382 SmallVector<CCValAssign, 16> RVLocs2;
1383 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1384 RVLocs2, *DAG.getContext());
1385 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1387 if (RVLocs1.size() != RVLocs2.size())
1389 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1390 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1392 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1394 if (RVLocs1[i].isRegLoc()) {
1395 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1398 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1404 // If the callee takes no arguments then go on to check the results of the
1406 if (!Outs.empty()) {
1407 // Check if stack adjustment is needed. For now, do not do this if any
1408 // argument is passed on the stack.
1409 SmallVector<CCValAssign, 16> ArgLocs;
1410 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1411 ArgLocs, *DAG.getContext());
1412 CCInfo.AnalyzeCallOperands(Outs,
1413 CCAssignFnForNode(CalleeCC, false, isVarArg));
1414 if (CCInfo.getNextStackOffset()) {
1415 MachineFunction &MF = DAG.getMachineFunction();
1417 // Check if the arguments are already laid out in the right way as
1418 // the caller's fixed stack objects.
1419 MachineFrameInfo *MFI = MF.getFrameInfo();
1420 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1421 const ARMInstrInfo *TII =
1422 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1423 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1425 ++i, ++realArgIdx) {
1426 CCValAssign &VA = ArgLocs[i];
1427 EVT RegVT = VA.getLocVT();
1428 SDValue Arg = Outs[realArgIdx].Val;
1429 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1430 if (VA.getLocInfo() == CCValAssign::Indirect)
1432 if (VA.needsCustom()) {
1433 // f64 and vector types are split into multiple registers or
1434 // register/stack-slot combinations. The types will not match
1435 // the registers; give up on memory f64 refs until we figure
1436 // out what to do about this.
1439 if (!ArgLocs[++i].isRegLoc())
1441 if (RegVT == MVT::v2f64) {
1442 if (!ArgLocs[++i].isRegLoc())
1444 if (!ArgLocs[++i].isRegLoc())
1447 } else if (!VA.isRegLoc()) {
1448 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1460 ARMTargetLowering::LowerReturn(SDValue Chain,
1461 CallingConv::ID CallConv, bool isVarArg,
1462 const SmallVectorImpl<ISD::OutputArg> &Outs,
1463 DebugLoc dl, SelectionDAG &DAG) const {
1465 // CCValAssign - represent the assignment of the return value to a location.
1466 SmallVector<CCValAssign, 16> RVLocs;
1468 // CCState - Info about the registers and stack slots.
1469 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1472 // Analyze outgoing return values.
1473 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1476 // If this is the first return lowered for this function, add
1477 // the regs to the liveout set for the function.
1478 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1479 for (unsigned i = 0; i != RVLocs.size(); ++i)
1480 if (RVLocs[i].isRegLoc())
1481 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1486 // Copy the result values into the output registers.
1487 for (unsigned i = 0, realRVLocIdx = 0;
1489 ++i, ++realRVLocIdx) {
1490 CCValAssign &VA = RVLocs[i];
1491 assert(VA.isRegLoc() && "Can only return in registers!");
1493 SDValue Arg = Outs[realRVLocIdx].Val;
1495 switch (VA.getLocInfo()) {
1496 default: llvm_unreachable("Unknown loc info!");
1497 case CCValAssign::Full: break;
1498 case CCValAssign::BCvt:
1499 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1503 if (VA.needsCustom()) {
1504 if (VA.getLocVT() == MVT::v2f64) {
1505 // Extract the first half and return it in two registers.
1506 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1507 DAG.getConstant(0, MVT::i32));
1508 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1509 DAG.getVTList(MVT::i32, MVT::i32), Half);
1511 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1512 Flag = Chain.getValue(1);
1513 VA = RVLocs[++i]; // skip ahead to next loc
1514 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1515 HalfGPRs.getValue(1), Flag);
1516 Flag = Chain.getValue(1);
1517 VA = RVLocs[++i]; // skip ahead to next loc
1519 // Extract the 2nd half and fall through to handle it as an f64 value.
1520 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1521 DAG.getConstant(1, MVT::i32));
1523 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1525 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1526 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1527 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1528 Flag = Chain.getValue(1);
1529 VA = RVLocs[++i]; // skip ahead to next loc
1530 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1533 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1535 // Guarantee that all emitted copies are
1536 // stuck together, avoiding something bad.
1537 Flag = Chain.getValue(1);
1542 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1544 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1549 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1550 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1551 // one of the above mentioned nodes. It has to be wrapped because otherwise
1552 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1553 // be used to form addressing mode. These wrapped nodes will be selected
1555 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1556 EVT PtrVT = Op.getValueType();
1557 // FIXME there is no actual debug info here
1558 DebugLoc dl = Op.getDebugLoc();
1559 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1561 if (CP->isMachineConstantPoolEntry())
1562 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1563 CP->getAlignment());
1565 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1566 CP->getAlignment());
1567 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1570 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1571 SelectionDAG &DAG) const {
1572 MachineFunction &MF = DAG.getMachineFunction();
1573 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1574 unsigned ARMPCLabelIndex = 0;
1575 DebugLoc DL = Op.getDebugLoc();
1576 EVT PtrVT = getPointerTy();
1577 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1578 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1580 if (RelocM == Reloc::Static) {
1581 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1583 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1584 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1585 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1586 ARMCP::CPBlockAddress,
1588 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1590 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1591 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1592 PseudoSourceValue::getConstantPool(), 0,
1594 if (RelocM == Reloc::Static)
1596 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1597 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1600 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1602 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1603 SelectionDAG &DAG) const {
1604 DebugLoc dl = GA->getDebugLoc();
1605 EVT PtrVT = getPointerTy();
1606 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1607 MachineFunction &MF = DAG.getMachineFunction();
1608 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1609 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1610 ARMConstantPoolValue *CPV =
1611 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1612 ARMCP::CPValue, PCAdj, "tlsgd", true);
1613 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1614 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1615 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1616 PseudoSourceValue::getConstantPool(), 0,
1618 SDValue Chain = Argument.getValue(1);
1620 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1621 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1623 // call __tls_get_addr.
1626 Entry.Node = Argument;
1627 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1628 Args.push_back(Entry);
1629 // FIXME: is there useful debug info available here?
1630 std::pair<SDValue, SDValue> CallResult =
1631 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1632 false, false, false, false,
1633 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1634 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1635 return CallResult.first;
1638 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1639 // "local exec" model.
1641 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1642 SelectionDAG &DAG) const {
1643 const GlobalValue *GV = GA->getGlobal();
1644 DebugLoc dl = GA->getDebugLoc();
1646 SDValue Chain = DAG.getEntryNode();
1647 EVT PtrVT = getPointerTy();
1648 // Get the Thread Pointer
1649 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1651 if (GV->isDeclaration()) {
1652 MachineFunction &MF = DAG.getMachineFunction();
1653 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1654 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1655 // Initial exec model.
1656 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1657 ARMConstantPoolValue *CPV =
1658 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1659 ARMCP::CPValue, PCAdj, "gottpoff", true);
1660 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1661 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1662 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1663 PseudoSourceValue::getConstantPool(), 0,
1665 Chain = Offset.getValue(1);
1667 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1668 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1670 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1671 PseudoSourceValue::getConstantPool(), 0,
1675 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1676 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1677 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1678 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1679 PseudoSourceValue::getConstantPool(), 0,
1683 // The address of the thread local variable is the add of the thread
1684 // pointer with the offset of the variable.
1685 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1689 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1690 // TODO: implement the "local dynamic" model
1691 assert(Subtarget->isTargetELF() &&
1692 "TLS not implemented for non-ELF targets");
1693 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1694 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1695 // otherwise use the "Local Exec" TLS Model
1696 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1697 return LowerToTLSGeneralDynamicModel(GA, DAG);
1699 return LowerToTLSExecModels(GA, DAG);
1702 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1703 SelectionDAG &DAG) const {
1704 EVT PtrVT = getPointerTy();
1705 DebugLoc dl = Op.getDebugLoc();
1706 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1707 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1708 if (RelocM == Reloc::PIC_) {
1709 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1710 ARMConstantPoolValue *CPV =
1711 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1712 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1713 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1714 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1716 PseudoSourceValue::getConstantPool(), 0,
1718 SDValue Chain = Result.getValue(1);
1719 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1720 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1722 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1723 PseudoSourceValue::getGOT(), 0,
1727 // If we have T2 ops, we can materialize the address directly via movt/movw
1728 // pair. This is always cheaper.
1729 if (Subtarget->useMovt()) {
1730 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1731 DAG.getTargetGlobalAddress(GV, PtrVT));
1733 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1734 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1735 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1736 PseudoSourceValue::getConstantPool(), 0,
1742 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1743 SelectionDAG &DAG) const {
1744 MachineFunction &MF = DAG.getMachineFunction();
1745 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1746 unsigned ARMPCLabelIndex = 0;
1747 EVT PtrVT = getPointerTy();
1748 DebugLoc dl = Op.getDebugLoc();
1749 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1750 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1752 if (RelocM == Reloc::Static)
1753 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1755 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1756 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1757 ARMConstantPoolValue *CPV =
1758 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1759 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1761 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1763 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1764 PseudoSourceValue::getConstantPool(), 0,
1766 SDValue Chain = Result.getValue(1);
1768 if (RelocM == Reloc::PIC_) {
1769 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1770 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1773 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1774 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1775 PseudoSourceValue::getGOT(), 0,
1781 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1782 SelectionDAG &DAG) const {
1783 assert(Subtarget->isTargetELF() &&
1784 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1785 MachineFunction &MF = DAG.getMachineFunction();
1786 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1787 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1788 EVT PtrVT = getPointerTy();
1789 DebugLoc dl = Op.getDebugLoc();
1790 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1791 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1792 "_GLOBAL_OFFSET_TABLE_",
1793 ARMPCLabelIndex, PCAdj);
1794 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1795 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1796 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1797 PseudoSourceValue::getConstantPool(), 0,
1799 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1800 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1804 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1805 DebugLoc dl = Op.getDebugLoc();
1806 SDValue Val = DAG.getConstant(0, MVT::i32);
1807 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1808 Op.getOperand(1), Val);
1812 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1813 DebugLoc dl = Op.getDebugLoc();
1814 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1815 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1819 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1820 const ARMSubtarget *Subtarget)
1822 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1823 DebugLoc dl = Op.getDebugLoc();
1825 default: return SDValue(); // Don't custom lower most intrinsics.
1826 case Intrinsic::arm_thread_pointer: {
1827 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1828 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1830 case Intrinsic::eh_sjlj_lsda: {
1831 MachineFunction &MF = DAG.getMachineFunction();
1832 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1833 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1834 EVT PtrVT = getPointerTy();
1835 DebugLoc dl = Op.getDebugLoc();
1836 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1838 unsigned PCAdj = (RelocM != Reloc::PIC_)
1839 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1840 ARMConstantPoolValue *CPV =
1841 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1842 ARMCP::CPLSDA, PCAdj);
1843 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1844 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1846 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1847 PseudoSourceValue::getConstantPool(), 0,
1849 SDValue Chain = Result.getValue(1);
1851 if (RelocM == Reloc::PIC_) {
1852 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1853 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1860 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1861 const ARMSubtarget *Subtarget) {
1862 DebugLoc dl = Op.getDebugLoc();
1863 SDValue Op5 = Op.getOperand(5);
1865 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1866 if (isDeviceBarrier) {
1867 if (Subtarget->hasV7Ops())
1868 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1870 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1871 DAG.getConstant(0, MVT::i32));
1873 if (Subtarget->hasV7Ops())
1874 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1876 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1877 DAG.getConstant(0, MVT::i32));
1882 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1883 MachineFunction &MF = DAG.getMachineFunction();
1884 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1886 // vastart just stores the address of the VarArgsFrameIndex slot into the
1887 // memory location argument.
1888 DebugLoc dl = Op.getDebugLoc();
1889 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1890 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1891 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1892 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1897 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1898 SelectionDAG &DAG) const {
1899 SDNode *Node = Op.getNode();
1900 DebugLoc dl = Node->getDebugLoc();
1901 EVT VT = Node->getValueType(0);
1902 SDValue Chain = Op.getOperand(0);
1903 SDValue Size = Op.getOperand(1);
1904 SDValue Align = Op.getOperand(2);
1906 // Chain the dynamic stack allocation so that it doesn't modify the stack
1907 // pointer when other instructions are using the stack.
1908 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1910 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1911 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1912 if (AlignVal > StackAlign)
1913 // Do this now since selection pass cannot introduce new target
1914 // independent node.
1915 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1917 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1918 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1919 // do even more horrible hack later.
1920 MachineFunction &MF = DAG.getMachineFunction();
1921 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1922 if (AFI->isThumb1OnlyFunction()) {
1924 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1926 uint32_t Val = C->getZExtValue();
1927 if (Val <= 508 && ((Val & 3) == 0))
1931 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1934 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1935 SDValue Ops1[] = { Chain, Size, Align };
1936 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1937 Chain = Res.getValue(1);
1938 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1939 DAG.getIntPtrConstant(0, true), SDValue());
1940 SDValue Ops2[] = { Res, Chain };
1941 return DAG.getMergeValues(Ops2, 2, dl);
1945 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1946 SDValue &Root, SelectionDAG &DAG,
1947 DebugLoc dl) const {
1948 MachineFunction &MF = DAG.getMachineFunction();
1949 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1951 TargetRegisterClass *RC;
1952 if (AFI->isThumb1OnlyFunction())
1953 RC = ARM::tGPRRegisterClass;
1955 RC = ARM::GPRRegisterClass;
1957 // Transform the arguments stored in physical registers into virtual ones.
1958 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1959 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1962 if (NextVA.isMemLoc()) {
1963 MachineFrameInfo *MFI = MF.getFrameInfo();
1964 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1966 // Create load node to retrieve arguments from the stack.
1967 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1968 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1969 PseudoSourceValue::getFixedStack(FI), 0,
1972 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1973 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1976 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1980 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1981 CallingConv::ID CallConv, bool isVarArg,
1982 const SmallVectorImpl<ISD::InputArg>
1984 DebugLoc dl, SelectionDAG &DAG,
1985 SmallVectorImpl<SDValue> &InVals)
1988 MachineFunction &MF = DAG.getMachineFunction();
1989 MachineFrameInfo *MFI = MF.getFrameInfo();
1991 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1993 // Assign locations to all of the incoming arguments.
1994 SmallVector<CCValAssign, 16> ArgLocs;
1995 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1997 CCInfo.AnalyzeFormalArguments(Ins,
1998 CCAssignFnForNode(CallConv, /* Return*/ false,
2001 SmallVector<SDValue, 16> ArgValues;
2003 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2004 CCValAssign &VA = ArgLocs[i];
2006 // Arguments stored in registers.
2007 if (VA.isRegLoc()) {
2008 EVT RegVT = VA.getLocVT();
2011 if (VA.needsCustom()) {
2012 // f64 and vector types are split up into multiple registers or
2013 // combinations of registers and stack slots.
2014 if (VA.getLocVT() == MVT::v2f64) {
2015 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2017 VA = ArgLocs[++i]; // skip ahead to next loc
2019 if (VA.isMemLoc()) {
2020 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
2022 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2023 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2024 PseudoSourceValue::getFixedStack(FI), 0,
2027 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2030 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2031 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2032 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2033 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2034 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2036 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2039 TargetRegisterClass *RC;
2041 if (RegVT == MVT::f32)
2042 RC = ARM::SPRRegisterClass;
2043 else if (RegVT == MVT::f64)
2044 RC = ARM::DPRRegisterClass;
2045 else if (RegVT == MVT::v2f64)
2046 RC = ARM::QPRRegisterClass;
2047 else if (RegVT == MVT::i32)
2048 RC = (AFI->isThumb1OnlyFunction() ?
2049 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2051 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2053 // Transform the arguments in physical registers into virtual ones.
2054 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2055 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2058 // If this is an 8 or 16-bit value, it is really passed promoted
2059 // to 32 bits. Insert an assert[sz]ext to capture this, then
2060 // truncate to the right size.
2061 switch (VA.getLocInfo()) {
2062 default: llvm_unreachable("Unknown loc info!");
2063 case CCValAssign::Full: break;
2064 case CCValAssign::BCvt:
2065 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2067 case CCValAssign::SExt:
2068 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2069 DAG.getValueType(VA.getValVT()));
2070 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2072 case CCValAssign::ZExt:
2073 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2074 DAG.getValueType(VA.getValVT()));
2075 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2079 InVals.push_back(ArgValue);
2081 } else { // VA.isRegLoc()
2084 assert(VA.isMemLoc());
2085 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2087 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2088 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2091 // Create load nodes to retrieve arguments from the stack.
2092 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2093 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2094 PseudoSourceValue::getFixedStack(FI), 0,
2101 static const unsigned GPRArgRegs[] = {
2102 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2105 unsigned NumGPRs = CCInfo.getFirstUnallocated
2106 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2108 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2109 unsigned VARegSize = (4 - NumGPRs) * 4;
2110 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2111 unsigned ArgOffset = CCInfo.getNextStackOffset();
2112 if (VARegSaveSize) {
2113 // If this function is vararg, store any remaining integer argument regs
2114 // to their spots on the stack so that they may be loaded by deferencing
2115 // the result of va_next.
2116 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2117 AFI->setVarArgsFrameIndex(
2118 MFI->CreateFixedObject(VARegSaveSize,
2119 ArgOffset + VARegSaveSize - VARegSize,
2121 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2124 SmallVector<SDValue, 4> MemOps;
2125 for (; NumGPRs < 4; ++NumGPRs) {
2126 TargetRegisterClass *RC;
2127 if (AFI->isThumb1OnlyFunction())
2128 RC = ARM::tGPRRegisterClass;
2130 RC = ARM::GPRRegisterClass;
2132 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2133 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2135 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2136 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2137 0, false, false, 0);
2138 MemOps.push_back(Store);
2139 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2140 DAG.getConstant(4, getPointerTy()));
2142 if (!MemOps.empty())
2143 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2144 &MemOps[0], MemOps.size());
2146 // This will point to the next argument passed via stack.
2147 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
2154 /// isFloatingPointZero - Return true if this is +0.0.
2155 static bool isFloatingPointZero(SDValue Op) {
2156 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2157 return CFP->getValueAPF().isPosZero();
2158 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2159 // Maybe this has already been legalized into the constant pool?
2160 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2161 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2162 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2163 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2164 return CFP->getValueAPF().isPosZero();
2170 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2171 /// the given operands.
2173 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2174 SDValue &ARMCC, SelectionDAG &DAG,
2175 DebugLoc dl) const {
2176 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2177 unsigned C = RHSC->getZExtValue();
2178 if (!isLegalICmpImmediate(C)) {
2179 // Constant does not fit, try adjusting it by one?
2184 if (isLegalICmpImmediate(C-1)) {
2185 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2186 RHS = DAG.getConstant(C-1, MVT::i32);
2191 if (C > 0 && isLegalICmpImmediate(C-1)) {
2192 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2193 RHS = DAG.getConstant(C-1, MVT::i32);
2198 if (isLegalICmpImmediate(C+1)) {
2199 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2200 RHS = DAG.getConstant(C+1, MVT::i32);
2205 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2206 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2207 RHS = DAG.getConstant(C+1, MVT::i32);
2214 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2215 ARMISD::NodeType CompareType;
2218 CompareType = ARMISD::CMP;
2223 CompareType = ARMISD::CMPZ;
2226 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2227 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2230 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2231 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2234 if (!isFloatingPointZero(RHS))
2235 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2237 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2238 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2241 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2242 EVT VT = Op.getValueType();
2243 SDValue LHS = Op.getOperand(0);
2244 SDValue RHS = Op.getOperand(1);
2245 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2246 SDValue TrueVal = Op.getOperand(2);
2247 SDValue FalseVal = Op.getOperand(3);
2248 DebugLoc dl = Op.getDebugLoc();
2250 if (LHS.getValueType() == MVT::i32) {
2252 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2253 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2254 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2257 ARMCC::CondCodes CondCode, CondCode2;
2258 FPCCToARMCC(CC, CondCode, CondCode2);
2260 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2261 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2262 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2263 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2265 if (CondCode2 != ARMCC::AL) {
2266 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2267 // FIXME: Needs another CMP because flag can have but one use.
2268 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2269 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2270 Result, TrueVal, ARMCC2, CCR, Cmp2);
2275 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2276 SDValue Chain = Op.getOperand(0);
2277 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2278 SDValue LHS = Op.getOperand(2);
2279 SDValue RHS = Op.getOperand(3);
2280 SDValue Dest = Op.getOperand(4);
2281 DebugLoc dl = Op.getDebugLoc();
2283 if (LHS.getValueType() == MVT::i32) {
2285 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2286 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2287 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2288 Chain, Dest, ARMCC, CCR,Cmp);
2291 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2292 ARMCC::CondCodes CondCode, CondCode2;
2293 FPCCToARMCC(CC, CondCode, CondCode2);
2295 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2296 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2297 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2298 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2299 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2300 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2301 if (CondCode2 != ARMCC::AL) {
2302 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2303 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2304 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2309 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2310 SDValue Chain = Op.getOperand(0);
2311 SDValue Table = Op.getOperand(1);
2312 SDValue Index = Op.getOperand(2);
2313 DebugLoc dl = Op.getDebugLoc();
2315 EVT PTy = getPointerTy();
2316 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2317 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2318 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2319 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2320 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2321 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2322 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2323 if (Subtarget->isThumb2()) {
2324 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2325 // which does another jump to the destination. This also makes it easier
2326 // to translate it to TBB / TBH later.
2327 // FIXME: This might not work if the function is extremely large.
2328 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2329 Addr, Op.getOperand(2), JTI, UId);
2331 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2332 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2333 PseudoSourceValue::getJumpTable(), 0,
2335 Chain = Addr.getValue(1);
2336 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2337 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2339 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2340 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2341 Chain = Addr.getValue(1);
2342 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2346 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2347 DebugLoc dl = Op.getDebugLoc();
2350 switch (Op.getOpcode()) {
2352 assert(0 && "Invalid opcode!");
2353 case ISD::FP_TO_SINT:
2354 Opc = ARMISD::FTOSI;
2356 case ISD::FP_TO_UINT:
2357 Opc = ARMISD::FTOUI;
2360 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2361 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2364 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2365 EVT VT = Op.getValueType();
2366 DebugLoc dl = Op.getDebugLoc();
2369 switch (Op.getOpcode()) {
2371 assert(0 && "Invalid opcode!");
2372 case ISD::SINT_TO_FP:
2373 Opc = ARMISD::SITOF;
2375 case ISD::UINT_TO_FP:
2376 Opc = ARMISD::UITOF;
2380 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2381 return DAG.getNode(Opc, dl, VT, Op);
2384 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2385 // Implement fcopysign with a fabs and a conditional fneg.
2386 SDValue Tmp0 = Op.getOperand(0);
2387 SDValue Tmp1 = Op.getOperand(1);
2388 DebugLoc dl = Op.getDebugLoc();
2389 EVT VT = Op.getValueType();
2390 EVT SrcVT = Tmp1.getValueType();
2391 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2392 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2393 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2394 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2395 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2398 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2399 MachineFunction &MF = DAG.getMachineFunction();
2400 MachineFrameInfo *MFI = MF.getFrameInfo();
2401 MFI->setReturnAddressIsTaken(true);
2403 EVT VT = Op.getValueType();
2404 DebugLoc dl = Op.getDebugLoc();
2405 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2407 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2408 SDValue Offset = DAG.getConstant(4, MVT::i32);
2409 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2410 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2411 NULL, 0, false, false, 0);
2414 // Return LR, which contains the return address. Mark it an implicit live-in.
2415 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2416 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2419 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2420 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2421 MFI->setFrameAddressIsTaken(true);
2423 EVT VT = Op.getValueType();
2424 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2425 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2426 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2427 ? ARM::R7 : ARM::R11;
2428 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2430 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2435 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2436 /// expand a bit convert where either the source or destination type is i64 to
2437 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2438 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2439 /// vectors), since the legalizer won't know what to do with that.
2440 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2441 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2442 DebugLoc dl = N->getDebugLoc();
2443 SDValue Op = N->getOperand(0);
2445 // This function is only supposed to be called for i64 types, either as the
2446 // source or destination of the bit convert.
2447 EVT SrcVT = Op.getValueType();
2448 EVT DstVT = N->getValueType(0);
2449 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2450 "ExpandBIT_CONVERT called for non-i64 type");
2452 // Turn i64->f64 into VMOVDRR.
2453 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2454 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2455 DAG.getConstant(0, MVT::i32));
2456 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2457 DAG.getConstant(1, MVT::i32));
2458 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2459 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2462 // Turn f64->i64 into VMOVRRD.
2463 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2464 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2465 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2466 // Merge the pieces into a single i64 value.
2467 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2473 /// getZeroVector - Returns a vector of specified type with all zero elements.
2475 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2476 assert(VT.isVector() && "Expected a vector type");
2478 // Zero vectors are used to represent vector negation and in those cases
2479 // will be implemented with the NEON VNEG instruction. However, VNEG does
2480 // not support i64 elements, so sometimes the zero vectors will need to be
2481 // explicitly constructed. For those cases, and potentially other uses in
2482 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2483 // to their dest type. This ensures they get CSE'd.
2485 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2486 SmallVector<SDValue, 8> Ops;
2489 if (VT.getSizeInBits() == 64) {
2490 Ops.assign(8, Cst); TVT = MVT::v8i8;
2492 Ops.assign(16, Cst); TVT = MVT::v16i8;
2494 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2496 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2499 /// getOnesVector - Returns a vector of specified type with all bits set.
2501 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2502 assert(VT.isVector() && "Expected a vector type");
2504 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2505 // dest type. This ensures they get CSE'd.
2507 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2508 SmallVector<SDValue, 8> Ops;
2511 if (VT.getSizeInBits() == 64) {
2512 Ops.assign(8, Cst); TVT = MVT::v8i8;
2514 Ops.assign(16, Cst); TVT = MVT::v16i8;
2516 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2518 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2521 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2522 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2523 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2524 SelectionDAG &DAG) const {
2525 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2526 EVT VT = Op.getValueType();
2527 unsigned VTBits = VT.getSizeInBits();
2528 DebugLoc dl = Op.getDebugLoc();
2529 SDValue ShOpLo = Op.getOperand(0);
2530 SDValue ShOpHi = Op.getOperand(1);
2531 SDValue ShAmt = Op.getOperand(2);
2533 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2535 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2537 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2538 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2539 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2540 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2541 DAG.getConstant(VTBits, MVT::i32));
2542 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2543 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2544 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2546 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2547 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2549 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2550 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2553 SDValue Ops[2] = { Lo, Hi };
2554 return DAG.getMergeValues(Ops, 2, dl);
2557 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2558 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2559 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2560 SelectionDAG &DAG) const {
2561 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2562 EVT VT = Op.getValueType();
2563 unsigned VTBits = VT.getSizeInBits();
2564 DebugLoc dl = Op.getDebugLoc();
2565 SDValue ShOpLo = Op.getOperand(0);
2566 SDValue ShOpHi = Op.getOperand(1);
2567 SDValue ShAmt = Op.getOperand(2);
2570 assert(Op.getOpcode() == ISD::SHL_PARTS);
2571 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2572 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2573 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2574 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2575 DAG.getConstant(VTBits, MVT::i32));
2576 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2577 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2579 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2580 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2581 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2583 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2584 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2587 SDValue Ops[2] = { Lo, Hi };
2588 return DAG.getMergeValues(Ops, 2, dl);
2591 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2592 const ARMSubtarget *ST) {
2593 EVT VT = N->getValueType(0);
2594 DebugLoc dl = N->getDebugLoc();
2596 if (!ST->hasV6T2Ops())
2599 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2600 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2603 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2604 const ARMSubtarget *ST) {
2605 EVT VT = N->getValueType(0);
2606 DebugLoc dl = N->getDebugLoc();
2608 // Lower vector shifts on NEON to use VSHL.
2609 if (VT.isVector()) {
2610 assert(ST->hasNEON() && "unexpected vector shift");
2612 // Left shifts translate directly to the vshiftu intrinsic.
2613 if (N->getOpcode() == ISD::SHL)
2614 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2615 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2616 N->getOperand(0), N->getOperand(1));
2618 assert((N->getOpcode() == ISD::SRA ||
2619 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2621 // NEON uses the same intrinsics for both left and right shifts. For
2622 // right shifts, the shift amounts are negative, so negate the vector of
2624 EVT ShiftVT = N->getOperand(1).getValueType();
2625 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2626 getZeroVector(ShiftVT, DAG, dl),
2628 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2629 Intrinsic::arm_neon_vshifts :
2630 Intrinsic::arm_neon_vshiftu);
2631 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2632 DAG.getConstant(vshiftInt, MVT::i32),
2633 N->getOperand(0), NegatedCount);
2636 // We can get here for a node like i32 = ISD::SHL i32, i64
2640 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2641 "Unknown shift to lower!");
2643 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2644 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2645 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2648 // If we are in thumb mode, we don't have RRX.
2649 if (ST->isThumb1Only()) return SDValue();
2651 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2652 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2653 DAG.getConstant(0, MVT::i32));
2654 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2655 DAG.getConstant(1, MVT::i32));
2657 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2658 // captures the result into a carry flag.
2659 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2660 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2662 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2663 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2665 // Merge the pieces into a single i64 value.
2666 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2669 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2670 SDValue TmpOp0, TmpOp1;
2671 bool Invert = false;
2675 SDValue Op0 = Op.getOperand(0);
2676 SDValue Op1 = Op.getOperand(1);
2677 SDValue CC = Op.getOperand(2);
2678 EVT VT = Op.getValueType();
2679 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2680 DebugLoc dl = Op.getDebugLoc();
2682 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2683 switch (SetCCOpcode) {
2684 default: llvm_unreachable("Illegal FP comparison"); break;
2686 case ISD::SETNE: Invert = true; // Fallthrough
2688 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2690 case ISD::SETLT: Swap = true; // Fallthrough
2692 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2694 case ISD::SETLE: Swap = true; // Fallthrough
2696 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2697 case ISD::SETUGE: Swap = true; // Fallthrough
2698 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2699 case ISD::SETUGT: Swap = true; // Fallthrough
2700 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2701 case ISD::SETUEQ: Invert = true; // Fallthrough
2703 // Expand this to (OLT | OGT).
2707 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2708 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2710 case ISD::SETUO: Invert = true; // Fallthrough
2712 // Expand this to (OLT | OGE).
2716 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2717 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2721 // Integer comparisons.
2722 switch (SetCCOpcode) {
2723 default: llvm_unreachable("Illegal integer comparison"); break;
2724 case ISD::SETNE: Invert = true;
2725 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2726 case ISD::SETLT: Swap = true;
2727 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2728 case ISD::SETLE: Swap = true;
2729 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2730 case ISD::SETULT: Swap = true;
2731 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2732 case ISD::SETULE: Swap = true;
2733 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2736 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2737 if (Opc == ARMISD::VCEQ) {
2740 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2742 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2745 // Ignore bitconvert.
2746 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2747 AndOp = AndOp.getOperand(0);
2749 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2751 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2752 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2759 std::swap(Op0, Op1);
2761 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2764 Result = DAG.getNOT(dl, Result, VT);
2769 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2770 /// valid vector constant for a NEON instruction with a "modified immediate"
2771 /// operand (e.g., VMOV). If so, return either the constant being
2772 /// splatted or the encoded value, depending on the DoEncode parameter. The
2773 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2774 /// bits7-0=Immediate.
2775 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2776 unsigned SplatBitSize, SelectionDAG &DAG,
2777 bool isVMOV, bool DoEncode) {
2778 unsigned Op, Cmode, Imm;
2781 // SplatBitSize is set to the smallest size that splats the vector, so a
2782 // zero vector will always have SplatBitSize == 8. However, NEON modified
2783 // immediate instructions others than VMOV do not support the 8-bit encoding
2784 // of a zero vector, and the default encoding of zero is supposed to be the
2790 switch (SplatBitSize) {
2792 // Any 1-byte value is OK. Op=0, Cmode=1110.
2793 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2800 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2802 if ((SplatBits & ~0xff) == 0) {
2803 // Value = 0x00nn: Op=x, Cmode=100x.
2808 if ((SplatBits & ~0xff00) == 0) {
2809 // Value = 0xnn00: Op=x, Cmode=101x.
2811 Imm = SplatBits >> 8;
2817 // NEON's 32-bit VMOV supports splat values where:
2818 // * only one byte is nonzero, or
2819 // * the least significant byte is 0xff and the second byte is nonzero, or
2820 // * the least significant 2 bytes are 0xff and the third is nonzero.
2822 if ((SplatBits & ~0xff) == 0) {
2823 // Value = 0x000000nn: Op=x, Cmode=000x.
2828 if ((SplatBits & ~0xff00) == 0) {
2829 // Value = 0x0000nn00: Op=x, Cmode=001x.
2831 Imm = SplatBits >> 8;
2834 if ((SplatBits & ~0xff0000) == 0) {
2835 // Value = 0x00nn0000: Op=x, Cmode=010x.
2837 Imm = SplatBits >> 16;
2840 if ((SplatBits & ~0xff000000) == 0) {
2841 // Value = 0xnn000000: Op=x, Cmode=011x.
2843 Imm = SplatBits >> 24;
2847 if ((SplatBits & ~0xffff) == 0 &&
2848 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2849 // Value = 0x0000nnff: Op=x, Cmode=1100.
2851 Imm = SplatBits >> 8;
2856 if ((SplatBits & ~0xffffff) == 0 &&
2857 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2858 // Value = 0x00nnffff: Op=x, Cmode=1101.
2860 Imm = SplatBits >> 16;
2861 SplatBits |= 0xffff;
2865 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2866 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2867 // VMOV.I32. A (very) minor optimization would be to replicate the value
2868 // and fall through here to test for a valid 64-bit splat. But, then the
2869 // caller would also need to check and handle the change in size.
2873 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2876 uint64_t BitMask = 0xff;
2878 unsigned ImmMask = 1;
2880 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2881 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2884 } else if ((SplatBits & BitMask) != 0) {
2890 // Op=1, Cmode=1110.
2899 llvm_unreachable("unexpected size for EncodeNEONModImm");
2904 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2905 return DAG.getTargetConstant(SplatBits, VT);
2909 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2910 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2911 /// size, return the encoded value for that immediate. The ByteSize field
2912 /// indicates the number of bytes of each element [1248].
2913 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2914 SelectionDAG &DAG) {
2915 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2916 APInt SplatBits, SplatUndef;
2917 unsigned SplatBitSize;
2919 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2920 HasAnyUndefs, ByteSize * 8))
2923 if (SplatBitSize > ByteSize * 8)
2926 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2927 SplatBitSize, DAG, isVMOV, true);
2930 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2931 bool &ReverseVEXT, unsigned &Imm) {
2932 unsigned NumElts = VT.getVectorNumElements();
2933 ReverseVEXT = false;
2936 // If this is a VEXT shuffle, the immediate value is the index of the first
2937 // element. The other shuffle indices must be the successive elements after
2939 unsigned ExpectedElt = Imm;
2940 for (unsigned i = 1; i < NumElts; ++i) {
2941 // Increment the expected index. If it wraps around, it may still be
2942 // a VEXT but the source vectors must be swapped.
2944 if (ExpectedElt == NumElts * 2) {
2949 if (ExpectedElt != static_cast<unsigned>(M[i]))
2953 // Adjust the index value if the source operands will be swapped.
2960 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2961 /// instruction with the specified blocksize. (The order of the elements
2962 /// within each block of the vector is reversed.)
2963 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2964 unsigned BlockSize) {
2965 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2966 "Only possible block sizes for VREV are: 16, 32, 64");
2968 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2972 unsigned NumElts = VT.getVectorNumElements();
2973 unsigned BlockElts = M[0] + 1;
2975 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2978 for (unsigned i = 0; i < NumElts; ++i) {
2979 if ((unsigned) M[i] !=
2980 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2987 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2988 unsigned &WhichResult) {
2989 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2993 unsigned NumElts = VT.getVectorNumElements();
2994 WhichResult = (M[0] == 0 ? 0 : 1);
2995 for (unsigned i = 0; i < NumElts; i += 2) {
2996 if ((unsigned) M[i] != i + WhichResult ||
2997 (unsigned) M[i+1] != i + NumElts + WhichResult)
3003 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3004 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3005 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3006 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3007 unsigned &WhichResult) {
3008 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3012 unsigned NumElts = VT.getVectorNumElements();
3013 WhichResult = (M[0] == 0 ? 0 : 1);
3014 for (unsigned i = 0; i < NumElts; i += 2) {
3015 if ((unsigned) M[i] != i + WhichResult ||
3016 (unsigned) M[i+1] != i + WhichResult)
3022 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3023 unsigned &WhichResult) {
3024 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3028 unsigned NumElts = VT.getVectorNumElements();
3029 WhichResult = (M[0] == 0 ? 0 : 1);
3030 for (unsigned i = 0; i != NumElts; ++i) {
3031 if ((unsigned) M[i] != 2 * i + WhichResult)
3035 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3036 if (VT.is64BitVector() && EltSz == 32)
3042 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3043 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3044 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3045 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3046 unsigned &WhichResult) {
3047 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3051 unsigned Half = VT.getVectorNumElements() / 2;
3052 WhichResult = (M[0] == 0 ? 0 : 1);
3053 for (unsigned j = 0; j != 2; ++j) {
3054 unsigned Idx = WhichResult;
3055 for (unsigned i = 0; i != Half; ++i) {
3056 if ((unsigned) M[i + j * Half] != Idx)
3062 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3063 if (VT.is64BitVector() && EltSz == 32)
3069 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3070 unsigned &WhichResult) {
3071 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3075 unsigned NumElts = VT.getVectorNumElements();
3076 WhichResult = (M[0] == 0 ? 0 : 1);
3077 unsigned Idx = WhichResult * NumElts / 2;
3078 for (unsigned i = 0; i != NumElts; i += 2) {
3079 if ((unsigned) M[i] != Idx ||
3080 (unsigned) M[i+1] != Idx + NumElts)
3085 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3086 if (VT.is64BitVector() && EltSz == 32)
3092 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3093 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3094 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3095 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3096 unsigned &WhichResult) {
3097 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3101 unsigned NumElts = VT.getVectorNumElements();
3102 WhichResult = (M[0] == 0 ? 0 : 1);
3103 unsigned Idx = WhichResult * NumElts / 2;
3104 for (unsigned i = 0; i != NumElts; i += 2) {
3105 if ((unsigned) M[i] != Idx ||
3106 (unsigned) M[i+1] != Idx)
3111 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3112 if (VT.is64BitVector() && EltSz == 32)
3119 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3120 // Canonicalize all-zeros and all-ones vectors.
3121 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3122 if (ConstVal->isNullValue())
3123 return getZeroVector(VT, DAG, dl);
3124 if (ConstVal->isAllOnesValue())
3125 return getOnesVector(VT, DAG, dl);
3128 if (VT.is64BitVector()) {
3129 switch (Val.getValueType().getSizeInBits()) {
3130 case 8: CanonicalVT = MVT::v8i8; break;
3131 case 16: CanonicalVT = MVT::v4i16; break;
3132 case 32: CanonicalVT = MVT::v2i32; break;
3133 case 64: CanonicalVT = MVT::v1i64; break;
3134 default: llvm_unreachable("unexpected splat element type"); break;
3137 assert(VT.is128BitVector() && "unknown splat vector size");
3138 switch (Val.getValueType().getSizeInBits()) {
3139 case 8: CanonicalVT = MVT::v16i8; break;
3140 case 16: CanonicalVT = MVT::v8i16; break;
3141 case 32: CanonicalVT = MVT::v4i32; break;
3142 case 64: CanonicalVT = MVT::v2i64; break;
3143 default: llvm_unreachable("unexpected splat element type"); break;
3147 // Build a canonical splat for this value.
3148 SmallVector<SDValue, 8> Ops;
3149 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3150 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3152 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3155 // If this is a case we can't handle, return null and let the default
3156 // expansion code take care of it.
3157 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3158 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3159 DebugLoc dl = Op.getDebugLoc();
3160 EVT VT = Op.getValueType();
3162 APInt SplatBits, SplatUndef;
3163 unsigned SplatBitSize;
3165 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3166 if (SplatBitSize <= 64) {
3167 // Check if an immediate VMOV works.
3168 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3169 SplatUndef.getZExtValue(),
3170 SplatBitSize, DAG, true, false);
3172 return BuildSplat(Val, VT, DAG, dl);
3176 // Scan through the operands to see if only one value is used.
3177 unsigned NumElts = VT.getVectorNumElements();
3178 bool isOnlyLowElement = true;
3179 bool usesOnlyOneValue = true;
3180 bool isConstant = true;
3182 for (unsigned i = 0; i < NumElts; ++i) {
3183 SDValue V = Op.getOperand(i);
3184 if (V.getOpcode() == ISD::UNDEF)
3187 isOnlyLowElement = false;
3188 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3191 if (!Value.getNode())
3193 else if (V != Value)
3194 usesOnlyOneValue = false;
3197 if (!Value.getNode())
3198 return DAG.getUNDEF(VT);
3200 if (isOnlyLowElement)
3201 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3203 // If all elements are constants, fall back to the default expansion, which
3204 // will generate a load from the constant pool.
3208 // Use VDUP for non-constant splats.
3209 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3210 if (usesOnlyOneValue && EltSize <= 32)
3211 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3213 // Vectors with 32- or 64-bit elements can be built by directly assigning
3214 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3215 // will be legalized.
3216 if (EltSize >= 32) {
3217 // Do the expansion with floating-point types, since that is what the VFP
3218 // registers are defined to use, and since i64 is not legal.
3219 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3220 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3221 SmallVector<SDValue, 8> Ops;
3222 for (unsigned i = 0; i < NumElts; ++i)
3223 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3224 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3225 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3231 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3232 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3233 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3234 /// are assumed to be legal.
3236 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3238 if (VT.getVectorNumElements() == 4 &&
3239 (VT.is128BitVector() || VT.is64BitVector())) {
3240 unsigned PFIndexes[4];
3241 for (unsigned i = 0; i != 4; ++i) {
3245 PFIndexes[i] = M[i];
3248 // Compute the index in the perfect shuffle table.
3249 unsigned PFTableIndex =
3250 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3251 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3252 unsigned Cost = (PFEntry >> 30);
3259 unsigned Imm, WhichResult;
3261 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3262 return (EltSize >= 32 ||
3263 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3264 isVREVMask(M, VT, 64) ||
3265 isVREVMask(M, VT, 32) ||
3266 isVREVMask(M, VT, 16) ||
3267 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3268 isVTRNMask(M, VT, WhichResult) ||
3269 isVUZPMask(M, VT, WhichResult) ||
3270 isVZIPMask(M, VT, WhichResult) ||
3271 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3272 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3273 isVZIP_v_undef_Mask(M, VT, WhichResult));
3276 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3277 /// the specified operations to build the shuffle.
3278 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3279 SDValue RHS, SelectionDAG &DAG,
3281 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3282 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3283 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3286 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3295 OP_VUZPL, // VUZP, left result
3296 OP_VUZPR, // VUZP, right result
3297 OP_VZIPL, // VZIP, left result
3298 OP_VZIPR, // VZIP, right result
3299 OP_VTRNL, // VTRN, left result
3300 OP_VTRNR // VTRN, right result
3303 if (OpNum == OP_COPY) {
3304 if (LHSID == (1*9+2)*9+3) return LHS;
3305 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3309 SDValue OpLHS, OpRHS;
3310 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3311 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3312 EVT VT = OpLHS.getValueType();
3315 default: llvm_unreachable("Unknown shuffle opcode!");
3317 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3322 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3323 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3327 return DAG.getNode(ARMISD::VEXT, dl, VT,
3329 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3332 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3333 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3336 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3337 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3340 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3341 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3345 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3346 SDValue V1 = Op.getOperand(0);
3347 SDValue V2 = Op.getOperand(1);
3348 DebugLoc dl = Op.getDebugLoc();
3349 EVT VT = Op.getValueType();
3350 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3351 SmallVector<int, 8> ShuffleMask;
3353 // Convert shuffles that are directly supported on NEON to target-specific
3354 // DAG nodes, instead of keeping them as shuffles and matching them again
3355 // during code selection. This is more efficient and avoids the possibility
3356 // of inconsistencies between legalization and selection.
3357 // FIXME: floating-point vectors should be canonicalized to integer vectors
3358 // of the same time so that they get CSEd properly.
3359 SVN->getMask(ShuffleMask);
3361 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3362 if (EltSize <= 32) {
3363 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3364 int Lane = SVN->getSplatIndex();
3365 // If this is undef splat, generate it via "just" vdup, if possible.
3366 if (Lane == -1) Lane = 0;
3368 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3369 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3371 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3372 DAG.getConstant(Lane, MVT::i32));
3377 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3380 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3381 DAG.getConstant(Imm, MVT::i32));
3384 if (isVREVMask(ShuffleMask, VT, 64))
3385 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3386 if (isVREVMask(ShuffleMask, VT, 32))
3387 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3388 if (isVREVMask(ShuffleMask, VT, 16))
3389 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3391 // Check for Neon shuffles that modify both input vectors in place.
3392 // If both results are used, i.e., if there are two shuffles with the same
3393 // source operands and with masks corresponding to both results of one of
3394 // these operations, DAG memoization will ensure that a single node is
3395 // used for both shuffles.
3396 unsigned WhichResult;
3397 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3398 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3399 V1, V2).getValue(WhichResult);
3400 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3401 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3402 V1, V2).getValue(WhichResult);
3403 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3404 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3405 V1, V2).getValue(WhichResult);
3407 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3408 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3409 V1, V1).getValue(WhichResult);
3410 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3411 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3412 V1, V1).getValue(WhichResult);
3413 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3414 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3415 V1, V1).getValue(WhichResult);
3418 // If the shuffle is not directly supported and it has 4 elements, use
3419 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3420 unsigned NumElts = VT.getVectorNumElements();
3422 unsigned PFIndexes[4];
3423 for (unsigned i = 0; i != 4; ++i) {
3424 if (ShuffleMask[i] < 0)
3427 PFIndexes[i] = ShuffleMask[i];
3430 // Compute the index in the perfect shuffle table.
3431 unsigned PFTableIndex =
3432 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3433 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3434 unsigned Cost = (PFEntry >> 30);
3437 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3440 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3441 if (EltSize >= 32) {
3442 // Do the expansion with floating-point types, since that is what the VFP
3443 // registers are defined to use, and since i64 is not legal.
3444 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3445 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3446 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3447 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3448 SmallVector<SDValue, 8> Ops;
3449 for (unsigned i = 0; i < NumElts; ++i) {
3450 if (ShuffleMask[i] < 0)
3451 Ops.push_back(DAG.getUNDEF(EltVT));
3453 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3454 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3455 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3458 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3459 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3465 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3466 EVT VT = Op.getValueType();
3467 DebugLoc dl = Op.getDebugLoc();
3468 SDValue Vec = Op.getOperand(0);
3469 SDValue Lane = Op.getOperand(1);
3470 assert(VT == MVT::i32 &&
3471 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3472 "unexpected type for custom-lowering vector extract");
3473 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3476 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3477 // The only time a CONCAT_VECTORS operation can have legal types is when
3478 // two 64-bit vectors are concatenated to a 128-bit vector.
3479 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3480 "unexpected CONCAT_VECTORS");
3481 DebugLoc dl = Op.getDebugLoc();
3482 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3483 SDValue Op0 = Op.getOperand(0);
3484 SDValue Op1 = Op.getOperand(1);
3485 if (Op0.getOpcode() != ISD::UNDEF)
3486 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3487 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3488 DAG.getIntPtrConstant(0));
3489 if (Op1.getOpcode() != ISD::UNDEF)
3490 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3491 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3492 DAG.getIntPtrConstant(1));
3493 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3496 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3497 switch (Op.getOpcode()) {
3498 default: llvm_unreachable("Don't know how to custom lower this!");
3499 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3500 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3501 case ISD::GlobalAddress:
3502 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3503 LowerGlobalAddressELF(Op, DAG);
3504 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3505 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3506 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3507 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3508 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3509 case ISD::VASTART: return LowerVASTART(Op, DAG);
3510 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3511 case ISD::SINT_TO_FP:
3512 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3513 case ISD::FP_TO_SINT:
3514 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3515 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3516 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3517 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3518 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3519 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3520 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3521 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3523 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3526 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3527 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3528 case ISD::SRL_PARTS:
3529 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3530 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3531 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3532 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3533 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3534 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3535 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3540 /// ReplaceNodeResults - Replace the results of node with an illegal result
3541 /// type with new values built out of custom code.
3542 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3543 SmallVectorImpl<SDValue>&Results,
3544 SelectionDAG &DAG) const {
3546 switch (N->getOpcode()) {
3548 llvm_unreachable("Don't know how to custom expand this!");
3550 case ISD::BIT_CONVERT:
3551 Res = ExpandBIT_CONVERT(N, DAG);
3555 Res = LowerShift(N, DAG, Subtarget);
3559 Results.push_back(Res);
3562 //===----------------------------------------------------------------------===//
3563 // ARM Scheduler Hooks
3564 //===----------------------------------------------------------------------===//
3567 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3568 MachineBasicBlock *BB,
3569 unsigned Size) const {
3570 unsigned dest = MI->getOperand(0).getReg();
3571 unsigned ptr = MI->getOperand(1).getReg();
3572 unsigned oldval = MI->getOperand(2).getReg();
3573 unsigned newval = MI->getOperand(3).getReg();
3574 unsigned scratch = BB->getParent()->getRegInfo()
3575 .createVirtualRegister(ARM::GPRRegisterClass);
3576 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3577 DebugLoc dl = MI->getDebugLoc();
3578 bool isThumb2 = Subtarget->isThumb2();
3580 unsigned ldrOpc, strOpc;
3582 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3584 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3585 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3588 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3589 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3592 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3593 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3597 MachineFunction *MF = BB->getParent();
3598 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3599 MachineFunction::iterator It = BB;
3600 ++It; // insert the new blocks after the current block
3602 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3603 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3604 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3605 MF->insert(It, loop1MBB);
3606 MF->insert(It, loop2MBB);
3607 MF->insert(It, exitMBB);
3608 exitMBB->transferSuccessors(BB);
3612 // fallthrough --> loop1MBB
3613 BB->addSuccessor(loop1MBB);
3616 // ldrex dest, [ptr]
3620 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3621 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3622 .addReg(dest).addReg(oldval));
3623 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3624 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3625 BB->addSuccessor(loop2MBB);
3626 BB->addSuccessor(exitMBB);
3629 // strex scratch, newval, [ptr]
3633 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3635 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3636 .addReg(scratch).addImm(0));
3637 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3638 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3639 BB->addSuccessor(loop1MBB);
3640 BB->addSuccessor(exitMBB);
3646 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3652 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3653 unsigned Size, unsigned BinOpcode) const {
3654 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3655 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3657 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3658 MachineFunction *MF = BB->getParent();
3659 MachineFunction::iterator It = BB;
3662 unsigned dest = MI->getOperand(0).getReg();
3663 unsigned ptr = MI->getOperand(1).getReg();
3664 unsigned incr = MI->getOperand(2).getReg();
3665 DebugLoc dl = MI->getDebugLoc();
3667 bool isThumb2 = Subtarget->isThumb2();
3668 unsigned ldrOpc, strOpc;
3670 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3672 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3673 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3676 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3677 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3680 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3681 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3685 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3686 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3687 MF->insert(It, loopMBB);
3688 MF->insert(It, exitMBB);
3689 exitMBB->transferSuccessors(BB);
3691 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3692 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3693 unsigned scratch2 = (!BinOpcode) ? incr :
3694 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3698 // fallthrough --> loopMBB
3699 BB->addSuccessor(loopMBB);
3703 // <binop> scratch2, dest, incr
3704 // strex scratch, scratch2, ptr
3707 // fallthrough --> exitMBB
3709 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3711 // operand order needs to go the other way for NAND
3712 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3713 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3714 addReg(incr).addReg(dest)).addReg(0);
3716 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3717 addReg(dest).addReg(incr)).addReg(0);
3720 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3722 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3723 .addReg(scratch).addImm(0));
3724 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3725 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3727 BB->addSuccessor(loopMBB);
3728 BB->addSuccessor(exitMBB);
3734 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3740 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3741 MachineBasicBlock *BB) const {
3742 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3743 DebugLoc dl = MI->getDebugLoc();
3744 bool isThumb2 = Subtarget->isThumb2();
3745 switch (MI->getOpcode()) {
3748 llvm_unreachable("Unexpected instr type to insert");
3750 case ARM::ATOMIC_LOAD_ADD_I8:
3751 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3752 case ARM::ATOMIC_LOAD_ADD_I16:
3753 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3754 case ARM::ATOMIC_LOAD_ADD_I32:
3755 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3757 case ARM::ATOMIC_LOAD_AND_I8:
3758 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3759 case ARM::ATOMIC_LOAD_AND_I16:
3760 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3761 case ARM::ATOMIC_LOAD_AND_I32:
3762 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3764 case ARM::ATOMIC_LOAD_OR_I8:
3765 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3766 case ARM::ATOMIC_LOAD_OR_I16:
3767 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3768 case ARM::ATOMIC_LOAD_OR_I32:
3769 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3771 case ARM::ATOMIC_LOAD_XOR_I8:
3772 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3773 case ARM::ATOMIC_LOAD_XOR_I16:
3774 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3775 case ARM::ATOMIC_LOAD_XOR_I32:
3776 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3778 case ARM::ATOMIC_LOAD_NAND_I8:
3779 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3780 case ARM::ATOMIC_LOAD_NAND_I16:
3781 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3782 case ARM::ATOMIC_LOAD_NAND_I32:
3783 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3785 case ARM::ATOMIC_LOAD_SUB_I8:
3786 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3787 case ARM::ATOMIC_LOAD_SUB_I16:
3788 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3789 case ARM::ATOMIC_LOAD_SUB_I32:
3790 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3792 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3793 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3794 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3796 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3797 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3798 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3800 case ARM::tMOVCCr_pseudo: {
3801 // To "insert" a SELECT_CC instruction, we actually have to insert the
3802 // diamond control-flow pattern. The incoming instruction knows the
3803 // destination vreg to set, the condition code register to branch on, the
3804 // true/false values to select between, and a branch opcode to use.
3805 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3806 MachineFunction::iterator It = BB;
3812 // cmpTY ccX, r1, r2
3814 // fallthrough --> copy0MBB
3815 MachineBasicBlock *thisMBB = BB;
3816 MachineFunction *F = BB->getParent();
3817 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3818 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3819 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3820 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3821 F->insert(It, copy0MBB);
3822 F->insert(It, sinkMBB);
3823 // Update machine-CFG edges by first adding all successors of the current
3824 // block to the new block which will contain the Phi node for the select.
3825 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3826 E = BB->succ_end(); I != E; ++I)
3827 sinkMBB->addSuccessor(*I);
3828 // Next, remove all successors of the current block, and add the true
3829 // and fallthrough blocks as its successors.
3830 while (!BB->succ_empty())
3831 BB->removeSuccessor(BB->succ_begin());
3832 BB->addSuccessor(copy0MBB);
3833 BB->addSuccessor(sinkMBB);
3836 // %FalseValue = ...
3837 // # fallthrough to sinkMBB
3840 // Update machine-CFG edges
3841 BB->addSuccessor(sinkMBB);
3844 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3847 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3848 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3849 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3851 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3858 case ARM::t2SUBrSPi_:
3859 case ARM::t2SUBrSPi12_:
3860 case ARM::t2SUBrSPs_: {
3861 MachineFunction *MF = BB->getParent();
3862 unsigned DstReg = MI->getOperand(0).getReg();
3863 unsigned SrcReg = MI->getOperand(1).getReg();
3864 bool DstIsDead = MI->getOperand(0).isDead();
3865 bool SrcIsKill = MI->getOperand(1).isKill();
3867 if (SrcReg != ARM::SP) {
3868 // Copy the source to SP from virtual register.
3869 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3870 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3871 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3872 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3873 .addReg(SrcReg, getKillRegState(SrcIsKill));
3877 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3878 switch (MI->getOpcode()) {
3880 llvm_unreachable("Unexpected pseudo instruction!");
3886 OpOpc = ARM::tADDspr;
3889 OpOpc = ARM::tSUBspi;
3891 case ARM::t2SUBrSPi_:
3892 OpOpc = ARM::t2SUBrSPi;
3893 NeedPred = true; NeedCC = true;
3895 case ARM::t2SUBrSPi12_:
3896 OpOpc = ARM::t2SUBrSPi12;
3899 case ARM::t2SUBrSPs_:
3900 OpOpc = ARM::t2SUBrSPs;
3901 NeedPred = true; NeedCC = true; NeedOp3 = true;
3904 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3905 if (OpOpc == ARM::tAND)
3906 AddDefaultT1CC(MIB);
3907 MIB.addReg(ARM::SP);
3908 MIB.addOperand(MI->getOperand(2));
3910 MIB.addOperand(MI->getOperand(3));
3912 AddDefaultPred(MIB);
3916 // Copy the result from SP to virtual register.
3917 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3918 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3919 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3920 BuildMI(BB, dl, TII->get(CopyOpc))
3921 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3923 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3929 //===----------------------------------------------------------------------===//
3930 // ARM Optimization Hooks
3931 //===----------------------------------------------------------------------===//
3934 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3935 TargetLowering::DAGCombinerInfo &DCI) {
3936 SelectionDAG &DAG = DCI.DAG;
3937 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3938 EVT VT = N->getValueType(0);
3939 unsigned Opc = N->getOpcode();
3940 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3941 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3942 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3943 ISD::CondCode CC = ISD::SETCC_INVALID;
3946 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3948 SDValue CCOp = Slct.getOperand(0);
3949 if (CCOp.getOpcode() == ISD::SETCC)
3950 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3953 bool DoXform = false;
3955 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3958 if (LHS.getOpcode() == ISD::Constant &&
3959 cast<ConstantSDNode>(LHS)->isNullValue()) {
3961 } else if (CC != ISD::SETCC_INVALID &&
3962 RHS.getOpcode() == ISD::Constant &&
3963 cast<ConstantSDNode>(RHS)->isNullValue()) {
3964 std::swap(LHS, RHS);
3965 SDValue Op0 = Slct.getOperand(0);
3966 EVT OpVT = isSlctCC ? Op0.getValueType() :
3967 Op0.getOperand(0).getValueType();
3968 bool isInt = OpVT.isInteger();
3969 CC = ISD::getSetCCInverse(CC, isInt);
3971 if (!TLI.isCondCodeLegal(CC, OpVT))
3972 return SDValue(); // Inverse operator isn't legal.
3979 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3981 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3982 Slct.getOperand(0), Slct.getOperand(1), CC);
3983 SDValue CCOp = Slct.getOperand(0);
3985 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3986 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3987 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3988 CCOp, OtherOp, Result);
3993 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3994 static SDValue PerformADDCombine(SDNode *N,
3995 TargetLowering::DAGCombinerInfo &DCI) {
3996 // added by evan in r37685 with no testcase.
3997 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3999 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4000 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4001 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4002 if (Result.getNode()) return Result;
4004 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4005 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4006 if (Result.getNode()) return Result;
4012 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4013 static SDValue PerformSUBCombine(SDNode *N,
4014 TargetLowering::DAGCombinerInfo &DCI) {
4015 // added by evan in r37685 with no testcase.
4016 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4018 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4019 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4020 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4021 if (Result.getNode()) return Result;
4027 static SDValue PerformMULCombine(SDNode *N,
4028 TargetLowering::DAGCombinerInfo &DCI,
4029 const ARMSubtarget *Subtarget) {
4030 SelectionDAG &DAG = DCI.DAG;
4032 if (Subtarget->isThumb1Only())
4035 if (DAG.getMachineFunction().
4036 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4039 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4042 EVT VT = N->getValueType(0);
4046 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4050 uint64_t MulAmt = C->getZExtValue();
4051 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4052 ShiftAmt = ShiftAmt & (32 - 1);
4053 SDValue V = N->getOperand(0);
4054 DebugLoc DL = N->getDebugLoc();
4057 MulAmt >>= ShiftAmt;
4058 if (isPowerOf2_32(MulAmt - 1)) {
4059 // (mul x, 2^N + 1) => (add (shl x, N), x)
4060 Res = DAG.getNode(ISD::ADD, DL, VT,
4061 V, DAG.getNode(ISD::SHL, DL, VT,
4062 V, DAG.getConstant(Log2_32(MulAmt-1),
4064 } else if (isPowerOf2_32(MulAmt + 1)) {
4065 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4066 Res = DAG.getNode(ISD::SUB, DL, VT,
4067 DAG.getNode(ISD::SHL, DL, VT,
4068 V, DAG.getConstant(Log2_32(MulAmt+1),
4075 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4076 DAG.getConstant(ShiftAmt, MVT::i32));
4078 // Do not add new nodes to DAG combiner worklist.
4079 DCI.CombineTo(N, Res, false);
4083 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4084 /// ARMISD::VMOVRRD.
4085 static SDValue PerformVMOVRRDCombine(SDNode *N,
4086 TargetLowering::DAGCombinerInfo &DCI) {
4087 // fmrrd(fmdrr x, y) -> x,y
4088 SDValue InDouble = N->getOperand(0);
4089 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4090 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4094 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4095 /// operand of a vector shift operation, where all the elements of the
4096 /// build_vector must have the same constant integer value.
4097 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4098 // Ignore bit_converts.
4099 while (Op.getOpcode() == ISD::BIT_CONVERT)
4100 Op = Op.getOperand(0);
4101 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4102 APInt SplatBits, SplatUndef;
4103 unsigned SplatBitSize;
4105 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4106 HasAnyUndefs, ElementBits) ||
4107 SplatBitSize > ElementBits)
4109 Cnt = SplatBits.getSExtValue();
4113 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4114 /// operand of a vector shift left operation. That value must be in the range:
4115 /// 0 <= Value < ElementBits for a left shift; or
4116 /// 0 <= Value <= ElementBits for a long left shift.
4117 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4118 assert(VT.isVector() && "vector shift count is not a vector type");
4119 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4120 if (! getVShiftImm(Op, ElementBits, Cnt))
4122 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4125 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4126 /// operand of a vector shift right operation. For a shift opcode, the value
4127 /// is positive, but for an intrinsic the value count must be negative. The
4128 /// absolute value must be in the range:
4129 /// 1 <= |Value| <= ElementBits for a right shift; or
4130 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4131 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4133 assert(VT.isVector() && "vector shift count is not a vector type");
4134 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4135 if (! getVShiftImm(Op, ElementBits, Cnt))
4139 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4142 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4143 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4144 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4147 // Don't do anything for most intrinsics.
4150 // Vector shifts: check for immediate versions and lower them.
4151 // Note: This is done during DAG combining instead of DAG legalizing because
4152 // the build_vectors for 64-bit vector element shift counts are generally
4153 // not legal, and it is hard to see their values after they get legalized to
4154 // loads from a constant pool.
4155 case Intrinsic::arm_neon_vshifts:
4156 case Intrinsic::arm_neon_vshiftu:
4157 case Intrinsic::arm_neon_vshiftls:
4158 case Intrinsic::arm_neon_vshiftlu:
4159 case Intrinsic::arm_neon_vshiftn:
4160 case Intrinsic::arm_neon_vrshifts:
4161 case Intrinsic::arm_neon_vrshiftu:
4162 case Intrinsic::arm_neon_vrshiftn:
4163 case Intrinsic::arm_neon_vqshifts:
4164 case Intrinsic::arm_neon_vqshiftu:
4165 case Intrinsic::arm_neon_vqshiftsu:
4166 case Intrinsic::arm_neon_vqshiftns:
4167 case Intrinsic::arm_neon_vqshiftnu:
4168 case Intrinsic::arm_neon_vqshiftnsu:
4169 case Intrinsic::arm_neon_vqrshiftns:
4170 case Intrinsic::arm_neon_vqrshiftnu:
4171 case Intrinsic::arm_neon_vqrshiftnsu: {
4172 EVT VT = N->getOperand(1).getValueType();
4174 unsigned VShiftOpc = 0;
4177 case Intrinsic::arm_neon_vshifts:
4178 case Intrinsic::arm_neon_vshiftu:
4179 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4180 VShiftOpc = ARMISD::VSHL;
4183 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4184 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4185 ARMISD::VSHRs : ARMISD::VSHRu);
4190 case Intrinsic::arm_neon_vshiftls:
4191 case Intrinsic::arm_neon_vshiftlu:
4192 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4194 llvm_unreachable("invalid shift count for vshll intrinsic");
4196 case Intrinsic::arm_neon_vrshifts:
4197 case Intrinsic::arm_neon_vrshiftu:
4198 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4202 case Intrinsic::arm_neon_vqshifts:
4203 case Intrinsic::arm_neon_vqshiftu:
4204 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4208 case Intrinsic::arm_neon_vqshiftsu:
4209 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4211 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4213 case Intrinsic::arm_neon_vshiftn:
4214 case Intrinsic::arm_neon_vrshiftn:
4215 case Intrinsic::arm_neon_vqshiftns:
4216 case Intrinsic::arm_neon_vqshiftnu:
4217 case Intrinsic::arm_neon_vqshiftnsu:
4218 case Intrinsic::arm_neon_vqrshiftns:
4219 case Intrinsic::arm_neon_vqrshiftnu:
4220 case Intrinsic::arm_neon_vqrshiftnsu:
4221 // Narrowing shifts require an immediate right shift.
4222 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4224 llvm_unreachable("invalid shift count for narrowing vector shift "
4228 llvm_unreachable("unhandled vector shift");
4232 case Intrinsic::arm_neon_vshifts:
4233 case Intrinsic::arm_neon_vshiftu:
4234 // Opcode already set above.
4236 case Intrinsic::arm_neon_vshiftls:
4237 case Intrinsic::arm_neon_vshiftlu:
4238 if (Cnt == VT.getVectorElementType().getSizeInBits())
4239 VShiftOpc = ARMISD::VSHLLi;
4241 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4242 ARMISD::VSHLLs : ARMISD::VSHLLu);
4244 case Intrinsic::arm_neon_vshiftn:
4245 VShiftOpc = ARMISD::VSHRN; break;
4246 case Intrinsic::arm_neon_vrshifts:
4247 VShiftOpc = ARMISD::VRSHRs; break;
4248 case Intrinsic::arm_neon_vrshiftu:
4249 VShiftOpc = ARMISD::VRSHRu; break;
4250 case Intrinsic::arm_neon_vrshiftn:
4251 VShiftOpc = ARMISD::VRSHRN; break;
4252 case Intrinsic::arm_neon_vqshifts:
4253 VShiftOpc = ARMISD::VQSHLs; break;
4254 case Intrinsic::arm_neon_vqshiftu:
4255 VShiftOpc = ARMISD::VQSHLu; break;
4256 case Intrinsic::arm_neon_vqshiftsu:
4257 VShiftOpc = ARMISD::VQSHLsu; break;
4258 case Intrinsic::arm_neon_vqshiftns:
4259 VShiftOpc = ARMISD::VQSHRNs; break;
4260 case Intrinsic::arm_neon_vqshiftnu:
4261 VShiftOpc = ARMISD::VQSHRNu; break;
4262 case Intrinsic::arm_neon_vqshiftnsu:
4263 VShiftOpc = ARMISD::VQSHRNsu; break;
4264 case Intrinsic::arm_neon_vqrshiftns:
4265 VShiftOpc = ARMISD::VQRSHRNs; break;
4266 case Intrinsic::arm_neon_vqrshiftnu:
4267 VShiftOpc = ARMISD::VQRSHRNu; break;
4268 case Intrinsic::arm_neon_vqrshiftnsu:
4269 VShiftOpc = ARMISD::VQRSHRNsu; break;
4272 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4273 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4276 case Intrinsic::arm_neon_vshiftins: {
4277 EVT VT = N->getOperand(1).getValueType();
4279 unsigned VShiftOpc = 0;
4281 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4282 VShiftOpc = ARMISD::VSLI;
4283 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4284 VShiftOpc = ARMISD::VSRI;
4286 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4289 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4290 N->getOperand(1), N->getOperand(2),
4291 DAG.getConstant(Cnt, MVT::i32));
4294 case Intrinsic::arm_neon_vqrshifts:
4295 case Intrinsic::arm_neon_vqrshiftu:
4296 // No immediate versions of these to check for.
4303 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4304 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4305 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4306 /// vector element shift counts are generally not legal, and it is hard to see
4307 /// their values after they get legalized to loads from a constant pool.
4308 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4309 const ARMSubtarget *ST) {
4310 EVT VT = N->getValueType(0);
4312 // Nothing to be done for scalar shifts.
4313 if (! VT.isVector())
4316 assert(ST->hasNEON() && "unexpected vector shift");
4319 switch (N->getOpcode()) {
4320 default: llvm_unreachable("unexpected shift opcode");
4323 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4324 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4325 DAG.getConstant(Cnt, MVT::i32));
4330 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4331 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4332 ARMISD::VSHRs : ARMISD::VSHRu);
4333 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4334 DAG.getConstant(Cnt, MVT::i32));
4340 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4341 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4342 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4343 const ARMSubtarget *ST) {
4344 SDValue N0 = N->getOperand(0);
4346 // Check for sign- and zero-extensions of vector extract operations of 8-
4347 // and 16-bit vector elements. NEON supports these directly. They are
4348 // handled during DAG combining because type legalization will promote them
4349 // to 32-bit types and it is messy to recognize the operations after that.
4350 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4351 SDValue Vec = N0.getOperand(0);
4352 SDValue Lane = N0.getOperand(1);
4353 EVT VT = N->getValueType(0);
4354 EVT EltVT = N0.getValueType();
4355 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4357 if (VT == MVT::i32 &&
4358 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4359 TLI.isTypeLegal(Vec.getValueType())) {
4362 switch (N->getOpcode()) {
4363 default: llvm_unreachable("unexpected opcode");
4364 case ISD::SIGN_EXTEND:
4365 Opc = ARMISD::VGETLANEs;
4367 case ISD::ZERO_EXTEND:
4368 case ISD::ANY_EXTEND:
4369 Opc = ARMISD::VGETLANEu;
4372 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4379 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4380 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4381 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4382 const ARMSubtarget *ST) {
4383 // If the target supports NEON, try to use vmax/vmin instructions for f32
4384 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4385 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4386 // a NaN; only do the transformation when it matches that behavior.
4388 // For now only do this when using NEON for FP operations; if using VFP, it
4389 // is not obvious that the benefit outweighs the cost of switching to the
4391 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4392 N->getValueType(0) != MVT::f32)
4395 SDValue CondLHS = N->getOperand(0);
4396 SDValue CondRHS = N->getOperand(1);
4397 SDValue LHS = N->getOperand(2);
4398 SDValue RHS = N->getOperand(3);
4399 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4401 unsigned Opcode = 0;
4403 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4404 IsReversed = false; // x CC y ? x : y
4405 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4406 IsReversed = true ; // x CC y ? y : x
4420 // If LHS is NaN, an ordered comparison will be false and the result will
4421 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4422 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4423 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4424 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4426 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4427 // will return -0, so vmin can only be used for unsafe math or if one of
4428 // the operands is known to be nonzero.
4429 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4431 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4433 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4442 // If LHS is NaN, an ordered comparison will be false and the result will
4443 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4444 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4445 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4446 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4448 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4449 // will return +0, so vmax can only be used for unsafe math or if one of
4450 // the operands is known to be nonzero.
4451 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4453 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4455 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4461 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4464 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4465 DAGCombinerInfo &DCI) const {
4466 switch (N->getOpcode()) {
4468 case ISD::ADD: return PerformADDCombine(N, DCI);
4469 case ISD::SUB: return PerformSUBCombine(N, DCI);
4470 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4471 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4472 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4475 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4476 case ISD::SIGN_EXTEND:
4477 case ISD::ZERO_EXTEND:
4478 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4479 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4484 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4485 if (!Subtarget->hasV6Ops())
4486 // Pre-v6 does not support unaligned mem access.
4489 // v6+ may or may not support unaligned mem access depending on the system
4491 // FIXME: This is pretty conservative. Should we provide cmdline option to
4492 // control the behaviour?
4493 if (!Subtarget->isTargetDarwin())
4497 switch (VT.getSimpleVT().SimpleTy) {
4504 // FIXME: VLD1 etc with standard alignment is legal.
4508 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4513 switch (VT.getSimpleVT().SimpleTy) {
4514 default: return false;
4529 if ((V & (Scale - 1)) != 0)
4532 return V == (V & ((1LL << 5) - 1));
4535 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4536 const ARMSubtarget *Subtarget) {
4543 switch (VT.getSimpleVT().SimpleTy) {
4544 default: return false;
4549 // + imm12 or - imm8
4551 return V == (V & ((1LL << 8) - 1));
4552 return V == (V & ((1LL << 12) - 1));
4555 // Same as ARM mode. FIXME: NEON?
4556 if (!Subtarget->hasVFP2())
4561 return V == (V & ((1LL << 8) - 1));
4565 /// isLegalAddressImmediate - Return true if the integer value can be used
4566 /// as the offset of the target addressing mode for load / store of the
4568 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4569 const ARMSubtarget *Subtarget) {
4576 if (Subtarget->isThumb1Only())
4577 return isLegalT1AddressImmediate(V, VT);
4578 else if (Subtarget->isThumb2())
4579 return isLegalT2AddressImmediate(V, VT, Subtarget);
4584 switch (VT.getSimpleVT().SimpleTy) {
4585 default: return false;
4590 return V == (V & ((1LL << 12) - 1));
4593 return V == (V & ((1LL << 8) - 1));
4596 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4601 return V == (V & ((1LL << 8) - 1));
4605 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4607 int Scale = AM.Scale;
4611 switch (VT.getSimpleVT().SimpleTy) {
4612 default: return false;
4621 return Scale == 2 || Scale == 4 || Scale == 8;
4624 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4628 // Note, we allow "void" uses (basically, uses that aren't loads or
4629 // stores), because arm allows folding a scale into many arithmetic
4630 // operations. This should be made more precise and revisited later.
4632 // Allow r << imm, but the imm has to be a multiple of two.
4633 if (Scale & 1) return false;
4634 return isPowerOf2_32(Scale);
4638 /// isLegalAddressingMode - Return true if the addressing mode represented
4639 /// by AM is legal for this target, for a load/store of the specified type.
4640 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4641 const Type *Ty) const {
4642 EVT VT = getValueType(Ty, true);
4643 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4646 // Can never fold addr of global into load/store.
4651 case 0: // no scale reg, must be "r+i" or "r", or "i".
4654 if (Subtarget->isThumb1Only())
4658 // ARM doesn't support any R+R*scale+imm addr modes.
4665 if (Subtarget->isThumb2())
4666 return isLegalT2ScaledAddressingMode(AM, VT);
4668 int Scale = AM.Scale;
4669 switch (VT.getSimpleVT().SimpleTy) {
4670 default: return false;
4674 if (Scale < 0) Scale = -Scale;
4678 return isPowerOf2_32(Scale & ~1);
4682 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4687 // Note, we allow "void" uses (basically, uses that aren't loads or
4688 // stores), because arm allows folding a scale into many arithmetic
4689 // operations. This should be made more precise and revisited later.
4691 // Allow r << imm, but the imm has to be a multiple of two.
4692 if (Scale & 1) return false;
4693 return isPowerOf2_32(Scale);
4700 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4701 /// icmp immediate, that is the target has icmp instructions which can compare
4702 /// a register against the immediate without having to materialize the
4703 /// immediate into a register.
4704 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4705 if (!Subtarget->isThumb())
4706 return ARM_AM::getSOImmVal(Imm) != -1;
4707 if (Subtarget->isThumb2())
4708 return ARM_AM::getT2SOImmVal(Imm) != -1;
4709 return Imm >= 0 && Imm <= 255;
4712 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4713 bool isSEXTLoad, SDValue &Base,
4714 SDValue &Offset, bool &isInc,
4715 SelectionDAG &DAG) {
4716 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4719 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4721 Base = Ptr->getOperand(0);
4722 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4723 int RHSC = (int)RHS->getZExtValue();
4724 if (RHSC < 0 && RHSC > -256) {
4725 assert(Ptr->getOpcode() == ISD::ADD);
4727 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4731 isInc = (Ptr->getOpcode() == ISD::ADD);
4732 Offset = Ptr->getOperand(1);
4734 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4736 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4737 int RHSC = (int)RHS->getZExtValue();
4738 if (RHSC < 0 && RHSC > -0x1000) {
4739 assert(Ptr->getOpcode() == ISD::ADD);
4741 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4742 Base = Ptr->getOperand(0);
4747 if (Ptr->getOpcode() == ISD::ADD) {
4749 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4750 if (ShOpcVal != ARM_AM::no_shift) {
4751 Base = Ptr->getOperand(1);
4752 Offset = Ptr->getOperand(0);
4754 Base = Ptr->getOperand(0);
4755 Offset = Ptr->getOperand(1);
4760 isInc = (Ptr->getOpcode() == ISD::ADD);
4761 Base = Ptr->getOperand(0);
4762 Offset = Ptr->getOperand(1);
4766 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4770 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4771 bool isSEXTLoad, SDValue &Base,
4772 SDValue &Offset, bool &isInc,
4773 SelectionDAG &DAG) {
4774 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4777 Base = Ptr->getOperand(0);
4778 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4779 int RHSC = (int)RHS->getZExtValue();
4780 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4781 assert(Ptr->getOpcode() == ISD::ADD);
4783 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4785 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4786 isInc = Ptr->getOpcode() == ISD::ADD;
4787 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4795 /// getPreIndexedAddressParts - returns true by value, base pointer and
4796 /// offset pointer and addressing mode by reference if the node's address
4797 /// can be legally represented as pre-indexed load / store address.
4799 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4801 ISD::MemIndexedMode &AM,
4802 SelectionDAG &DAG) const {
4803 if (Subtarget->isThumb1Only())
4808 bool isSEXTLoad = false;
4809 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4810 Ptr = LD->getBasePtr();
4811 VT = LD->getMemoryVT();
4812 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4813 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4814 Ptr = ST->getBasePtr();
4815 VT = ST->getMemoryVT();
4820 bool isLegal = false;
4821 if (Subtarget->isThumb2())
4822 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4823 Offset, isInc, DAG);
4825 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4826 Offset, isInc, DAG);
4830 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4834 /// getPostIndexedAddressParts - returns true by value, base pointer and
4835 /// offset pointer and addressing mode by reference if this node can be
4836 /// combined with a load / store to form a post-indexed load / store.
4837 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4840 ISD::MemIndexedMode &AM,
4841 SelectionDAG &DAG) const {
4842 if (Subtarget->isThumb1Only())
4847 bool isSEXTLoad = false;
4848 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4849 VT = LD->getMemoryVT();
4850 Ptr = LD->getBasePtr();
4851 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4852 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4853 VT = ST->getMemoryVT();
4854 Ptr = ST->getBasePtr();
4859 bool isLegal = false;
4860 if (Subtarget->isThumb2())
4861 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4864 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4870 // Swap base ptr and offset to catch more post-index load / store when
4871 // it's legal. In Thumb2 mode, offset must be an immediate.
4872 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4873 !Subtarget->isThumb2())
4874 std::swap(Base, Offset);
4876 // Post-indexed load / store update the base pointer.
4881 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4885 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4889 const SelectionDAG &DAG,
4890 unsigned Depth) const {
4891 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4892 switch (Op.getOpcode()) {
4894 case ARMISD::CMOV: {
4895 // Bits are known zero/one if known on the LHS and RHS.
4896 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4897 if (KnownZero == 0 && KnownOne == 0) return;
4899 APInt KnownZeroRHS, KnownOneRHS;
4900 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4901 KnownZeroRHS, KnownOneRHS, Depth+1);
4902 KnownZero &= KnownZeroRHS;
4903 KnownOne &= KnownOneRHS;
4909 //===----------------------------------------------------------------------===//
4910 // ARM Inline Assembly Support
4911 //===----------------------------------------------------------------------===//
4913 /// getConstraintType - Given a constraint letter, return the type of
4914 /// constraint it is for this target.
4915 ARMTargetLowering::ConstraintType
4916 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4917 if (Constraint.size() == 1) {
4918 switch (Constraint[0]) {
4920 case 'l': return C_RegisterClass;
4921 case 'w': return C_RegisterClass;
4924 return TargetLowering::getConstraintType(Constraint);
4927 std::pair<unsigned, const TargetRegisterClass*>
4928 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4930 if (Constraint.size() == 1) {
4931 // GCC ARM Constraint Letters
4932 switch (Constraint[0]) {
4934 if (Subtarget->isThumb())
4935 return std::make_pair(0U, ARM::tGPRRegisterClass);
4937 return std::make_pair(0U, ARM::GPRRegisterClass);
4939 return std::make_pair(0U, ARM::GPRRegisterClass);
4942 return std::make_pair(0U, ARM::SPRRegisterClass);
4943 if (VT.getSizeInBits() == 64)
4944 return std::make_pair(0U, ARM::DPRRegisterClass);
4945 if (VT.getSizeInBits() == 128)
4946 return std::make_pair(0U, ARM::QPRRegisterClass);
4950 if (StringRef("{cc}").equals_lower(Constraint))
4951 return std::make_pair(0U, ARM::CCRRegisterClass);
4953 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4956 std::vector<unsigned> ARMTargetLowering::
4957 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4959 if (Constraint.size() != 1)
4960 return std::vector<unsigned>();
4962 switch (Constraint[0]) { // GCC ARM Constraint Letters
4965 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4966 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4969 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4970 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4971 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4972 ARM::R12, ARM::LR, 0);
4975 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4976 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4977 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4978 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4979 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4980 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4981 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4982 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4983 if (VT.getSizeInBits() == 64)
4984 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4985 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4986 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4987 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4988 if (VT.getSizeInBits() == 128)
4989 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4990 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4994 return std::vector<unsigned>();
4997 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4998 /// vector. If it is invalid, don't add anything to Ops.
4999 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5002 std::vector<SDValue>&Ops,
5003 SelectionDAG &DAG) const {
5004 SDValue Result(0, 0);
5006 switch (Constraint) {
5008 case 'I': case 'J': case 'K': case 'L':
5009 case 'M': case 'N': case 'O':
5010 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5014 int64_t CVal64 = C->getSExtValue();
5015 int CVal = (int) CVal64;
5016 // None of these constraints allow values larger than 32 bits. Check
5017 // that the value fits in an int.
5021 switch (Constraint) {
5023 if (Subtarget->isThumb1Only()) {
5024 // This must be a constant between 0 and 255, for ADD
5026 if (CVal >= 0 && CVal <= 255)
5028 } else if (Subtarget->isThumb2()) {
5029 // A constant that can be used as an immediate value in a
5030 // data-processing instruction.
5031 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5034 // A constant that can be used as an immediate value in a
5035 // data-processing instruction.
5036 if (ARM_AM::getSOImmVal(CVal) != -1)
5042 if (Subtarget->isThumb()) { // FIXME thumb2
5043 // This must be a constant between -255 and -1, for negated ADD
5044 // immediates. This can be used in GCC with an "n" modifier that
5045 // prints the negated value, for use with SUB instructions. It is
5046 // not useful otherwise but is implemented for compatibility.
5047 if (CVal >= -255 && CVal <= -1)
5050 // This must be a constant between -4095 and 4095. It is not clear
5051 // what this constraint is intended for. Implemented for
5052 // compatibility with GCC.
5053 if (CVal >= -4095 && CVal <= 4095)
5059 if (Subtarget->isThumb1Only()) {
5060 // A 32-bit value where only one byte has a nonzero value. Exclude
5061 // zero to match GCC. This constraint is used by GCC internally for
5062 // constants that can be loaded with a move/shift combination.
5063 // It is not useful otherwise but is implemented for compatibility.
5064 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5066 } else if (Subtarget->isThumb2()) {
5067 // A constant whose bitwise inverse can be used as an immediate
5068 // value in a data-processing instruction. This can be used in GCC
5069 // with a "B" modifier that prints the inverted value, for use with
5070 // BIC and MVN instructions. It is not useful otherwise but is
5071 // implemented for compatibility.
5072 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5075 // A constant whose bitwise inverse can be used as an immediate
5076 // value in a data-processing instruction. This can be used in GCC
5077 // with a "B" modifier that prints the inverted value, for use with
5078 // BIC and MVN instructions. It is not useful otherwise but is
5079 // implemented for compatibility.
5080 if (ARM_AM::getSOImmVal(~CVal) != -1)
5086 if (Subtarget->isThumb1Only()) {
5087 // This must be a constant between -7 and 7,
5088 // for 3-operand ADD/SUB immediate instructions.
5089 if (CVal >= -7 && CVal < 7)
5091 } else if (Subtarget->isThumb2()) {
5092 // A constant whose negation can be used as an immediate value in a
5093 // data-processing instruction. This can be used in GCC with an "n"
5094 // modifier that prints the negated value, for use with SUB
5095 // instructions. It is not useful otherwise but is implemented for
5097 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5100 // A constant whose negation can be used as an immediate value in a
5101 // data-processing instruction. This can be used in GCC with an "n"
5102 // modifier that prints the negated value, for use with SUB
5103 // instructions. It is not useful otherwise but is implemented for
5105 if (ARM_AM::getSOImmVal(-CVal) != -1)
5111 if (Subtarget->isThumb()) { // FIXME thumb2
5112 // This must be a multiple of 4 between 0 and 1020, for
5113 // ADD sp + immediate.
5114 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5117 // A power of two or a constant between 0 and 32. This is used in
5118 // GCC for the shift amount on shifted register operands, but it is
5119 // useful in general for any shift amounts.
5120 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5126 if (Subtarget->isThumb()) { // FIXME thumb2
5127 // This must be a constant between 0 and 31, for shift amounts.
5128 if (CVal >= 0 && CVal <= 31)
5134 if (Subtarget->isThumb()) { // FIXME thumb2
5135 // This must be a multiple of 4 between -508 and 508, for
5136 // ADD/SUB sp = sp + immediate.
5137 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5142 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5146 if (Result.getNode()) {
5147 Ops.push_back(Result);
5150 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
5155 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5156 // The ARM target isn't yet aware of offsets.
5160 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5161 APInt Imm = FPImm.bitcastToAPInt();
5162 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5163 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5164 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5166 // We can handle 4 bits of mantissa.
5167 // mantissa = (16+UInt(e:f:g:h))/16.
5168 if (Mantissa & 0x7ffff)
5171 if ((Mantissa & 0xf) != Mantissa)
5174 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5175 if (Exp < -3 || Exp > 4)
5177 Exp = ((Exp+3) & 0x7) ^ 4;
5179 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5182 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5183 APInt Imm = FPImm.bitcastToAPInt();
5184 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5185 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5186 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5188 // We can handle 4 bits of mantissa.
5189 // mantissa = (16+UInt(e:f:g:h))/16.
5190 if (Mantissa & 0xffffffffffffLL)
5193 if ((Mantissa & 0xf) != Mantissa)
5196 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5197 if (Exp < -3 || Exp > 4)
5199 Exp = ((Exp+3) & 0x7) ^ 4;
5201 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5204 /// isFPImmLegal - Returns true if the target can instruction select the
5205 /// specified FP immediate natively. If false, the legalizer will
5206 /// materialize the FP immediate as a load from a constant pool.
5207 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5208 if (!Subtarget->hasVFP3())
5211 return ARM::getVFPf32Imm(Imm) != -1;
5213 return ARM::getVFPf64Imm(Imm) != -1;