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 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
113 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
116 if (VT.isInteger()) {
117 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
119 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
122 // Promote all bit-wise operations.
123 if (VT.isInteger() && VT != PromotedBitwiseVT) {
124 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
125 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
126 PromotedBitwiseVT.getSimpleVT());
127 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
128 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
129 PromotedBitwiseVT.getSimpleVT());
130 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
131 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
132 PromotedBitwiseVT.getSimpleVT());
135 // Neon does not support vector divide/remainder operations.
136 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
137 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
138 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
139 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
141 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
144 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
145 addRegisterClass(VT, ARM::DPRRegisterClass);
146 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
149 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
150 addRegisterClass(VT, ARM::QPRRegisterClass);
151 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
154 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
155 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
156 return new TargetLoweringObjectFileMachO();
158 return new ARMElfTargetObjectFile();
161 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
162 : TargetLowering(TM, createTLOF(TM)) {
163 Subtarget = &TM.getSubtarget<ARMSubtarget>();
165 if (Subtarget->isTargetDarwin()) {
166 // Uses VFP for Thumb libfuncs if available.
167 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
168 // Single-precision floating-point arithmetic.
169 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
170 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
171 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
172 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
174 // Double-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
176 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
177 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
178 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
180 // Single-precision comparisons.
181 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
182 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
183 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
184 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
185 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
186 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
187 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
188 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
190 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
199 // Double-precision comparisons.
200 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
201 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
202 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
203 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
204 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
205 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
206 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
207 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
209 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
218 // Floating-point to integer conversions.
219 // i64 conversions are done via library routines even when generating VFP
220 // instructions, so use the same ones.
221 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
222 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
223 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
224 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
226 // Conversions between floating types.
227 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
228 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
230 // Integer to floating-point conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
234 // e.g., __floatunsidf vs. __floatunssidfvfp.
235 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
236 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
237 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
238 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
242 // These libcalls are not available in 32-bit.
243 setLibcallName(RTLIB::SHL_I128, 0);
244 setLibcallName(RTLIB::SRL_I128, 0);
245 setLibcallName(RTLIB::SRA_I128, 0);
247 // Libcalls should use the AAPCS base standard ABI, even if hard float
248 // is in effect, as per the ARM RTABI specification, section 4.1.2.
249 if (Subtarget->isAAPCS_ABI()) {
250 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
251 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
252 CallingConv::ARM_AAPCS);
256 if (Subtarget->isThumb1Only())
257 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
259 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
260 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
261 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
262 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
264 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
267 if (Subtarget->hasNEON()) {
268 addDRTypeForNEON(MVT::v2f32);
269 addDRTypeForNEON(MVT::v8i8);
270 addDRTypeForNEON(MVT::v4i16);
271 addDRTypeForNEON(MVT::v2i32);
272 addDRTypeForNEON(MVT::v1i64);
274 addQRTypeForNEON(MVT::v4f32);
275 addQRTypeForNEON(MVT::v2f64);
276 addQRTypeForNEON(MVT::v16i8);
277 addQRTypeForNEON(MVT::v8i16);
278 addQRTypeForNEON(MVT::v4i32);
279 addQRTypeForNEON(MVT::v2i64);
281 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
282 // neither Neon nor VFP support any arithmetic operations on it.
283 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
284 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
285 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
286 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
287 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
288 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
289 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
290 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
291 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
292 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
293 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
294 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
295 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
296 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
297 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
298 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
299 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
300 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
301 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
302 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
303 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
304 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
305 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
306 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
308 // Neon does not support some operations on v1i64 and v2i64 types.
309 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
310 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
311 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
312 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
314 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
315 setTargetDAGCombine(ISD::SHL);
316 setTargetDAGCombine(ISD::SRL);
317 setTargetDAGCombine(ISD::SRA);
318 setTargetDAGCombine(ISD::SIGN_EXTEND);
319 setTargetDAGCombine(ISD::ZERO_EXTEND);
320 setTargetDAGCombine(ISD::ANY_EXTEND);
321 setTargetDAGCombine(ISD::SELECT_CC);
324 computeRegisterProperties();
326 // ARM does not have f32 extending load.
327 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
329 // ARM does not have i1 sign extending load.
330 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
332 // ARM supports all 4 flavors of integer indexed load / store.
333 if (!Subtarget->isThumb1Only()) {
334 for (unsigned im = (unsigned)ISD::PRE_INC;
335 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
336 setIndexedLoadAction(im, MVT::i1, Legal);
337 setIndexedLoadAction(im, MVT::i8, Legal);
338 setIndexedLoadAction(im, MVT::i16, Legal);
339 setIndexedLoadAction(im, MVT::i32, Legal);
340 setIndexedStoreAction(im, MVT::i1, Legal);
341 setIndexedStoreAction(im, MVT::i8, Legal);
342 setIndexedStoreAction(im, MVT::i16, Legal);
343 setIndexedStoreAction(im, MVT::i32, Legal);
347 // i64 operation support.
348 if (Subtarget->isThumb1Only()) {
349 setOperationAction(ISD::MUL, MVT::i64, Expand);
350 setOperationAction(ISD::MULHU, MVT::i32, Expand);
351 setOperationAction(ISD::MULHS, MVT::i32, Expand);
352 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
353 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
355 setOperationAction(ISD::MUL, MVT::i64, Expand);
356 setOperationAction(ISD::MULHU, MVT::i32, Expand);
357 if (!Subtarget->hasV6Ops())
358 setOperationAction(ISD::MULHS, MVT::i32, Expand);
360 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
361 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
362 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
363 setOperationAction(ISD::SRL, MVT::i64, Custom);
364 setOperationAction(ISD::SRA, MVT::i64, Custom);
366 // ARM does not have ROTL.
367 setOperationAction(ISD::ROTL, MVT::i32, Expand);
368 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
369 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
370 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
371 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
373 // Only ARMv6 has BSWAP.
374 if (!Subtarget->hasV6Ops())
375 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
377 // These are expanded into libcalls.
378 if (!Subtarget->hasDivide()) {
379 // v7M has a hardware divider
380 setOperationAction(ISD::SDIV, MVT::i32, Expand);
381 setOperationAction(ISD::UDIV, MVT::i32, Expand);
383 setOperationAction(ISD::SREM, MVT::i32, Expand);
384 setOperationAction(ISD::UREM, MVT::i32, Expand);
385 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
386 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
388 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
389 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
390 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
391 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
392 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
394 setOperationAction(ISD::TRAP, MVT::Other, Legal);
396 // Use the default implementation.
397 setOperationAction(ISD::VASTART, MVT::Other, Custom);
398 setOperationAction(ISD::VAARG, MVT::Other, Expand);
399 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
400 setOperationAction(ISD::VAEND, MVT::Other, Expand);
401 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
402 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
403 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
404 // FIXME: Shouldn't need this, since no register is used, but the legalizer
405 // doesn't yet know how to not do that for SjLj.
406 setExceptionSelectorRegister(ARM::R0);
407 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
408 // Handle atomics directly for ARMv[67] (except for Thumb1), otherwise
409 // use the default expansion.
410 TargetLowering::LegalizeAction AtomicAction =
411 (Subtarget->hasV7Ops() ||
412 (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())) ? Custom : Expand;
413 setOperationAction(ISD::MEMBARRIER, MVT::Other, AtomicAction);
415 // If the subtarget does not have extract instructions, sign_extend_inreg
416 // needs to be expanded. Extract is available in ARM mode on v6 and up,
417 // and on most Thumb2 implementations.
418 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
419 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
420 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
421 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
423 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
425 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
426 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
427 // iff target supports vfp2.
428 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
430 // We want to custom lower some of our intrinsics.
431 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
433 setOperationAction(ISD::SETCC, MVT::i32, Expand);
434 setOperationAction(ISD::SETCC, MVT::f32, Expand);
435 setOperationAction(ISD::SETCC, MVT::f64, Expand);
436 setOperationAction(ISD::SELECT, MVT::i32, Expand);
437 setOperationAction(ISD::SELECT, MVT::f32, Expand);
438 setOperationAction(ISD::SELECT, MVT::f64, Expand);
439 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
440 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
441 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
443 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
444 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
445 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
446 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
447 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
449 // We don't support sin/cos/fmod/copysign/pow
450 setOperationAction(ISD::FSIN, MVT::f64, Expand);
451 setOperationAction(ISD::FSIN, MVT::f32, Expand);
452 setOperationAction(ISD::FCOS, MVT::f32, Expand);
453 setOperationAction(ISD::FCOS, MVT::f64, Expand);
454 setOperationAction(ISD::FREM, MVT::f64, Expand);
455 setOperationAction(ISD::FREM, MVT::f32, Expand);
456 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
457 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
458 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
460 setOperationAction(ISD::FPOW, MVT::f64, Expand);
461 setOperationAction(ISD::FPOW, MVT::f32, Expand);
463 // Various VFP goodness
464 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
465 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
466 if (Subtarget->hasVFP2()) {
467 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
468 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
469 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
470 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
472 // Special handling for half-precision FP.
473 if (!Subtarget->hasFP16()) {
474 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
475 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
479 // We have target-specific dag combine patterns for the following nodes:
480 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
481 setTargetDAGCombine(ISD::ADD);
482 setTargetDAGCombine(ISD::SUB);
483 setTargetDAGCombine(ISD::MUL);
485 setStackPointerRegisterToSaveRestore(ARM::SP);
487 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
488 setSchedulingPreference(Sched::RegPressure);
490 setSchedulingPreference(Sched::Hybrid);
492 // FIXME: If-converter should use instruction latency to determine
493 // profitability rather than relying on fixed limits.
494 if (Subtarget->getCPUString() == "generic") {
495 // Generic (and overly aggressive) if-conversion limits.
496 setIfCvtBlockSizeLimit(10);
497 setIfCvtDupBlockSizeLimit(2);
498 } else if (Subtarget->hasV7Ops()) {
499 setIfCvtBlockSizeLimit(3);
500 setIfCvtDupBlockSizeLimit(1);
501 } else if (Subtarget->hasV6Ops()) {
502 setIfCvtBlockSizeLimit(2);
503 setIfCvtDupBlockSizeLimit(1);
505 setIfCvtBlockSizeLimit(3);
506 setIfCvtDupBlockSizeLimit(2);
509 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
510 // Do not enable CodePlacementOpt for now: it currently runs after the
511 // ARMConstantIslandPass and messes up branch relaxation and placement
512 // of constant islands.
513 // benefitFromCodePlacementOpt = true;
516 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
519 case ARMISD::Wrapper: return "ARMISD::Wrapper";
520 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
521 case ARMISD::CALL: return "ARMISD::CALL";
522 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
523 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
524 case ARMISD::tCALL: return "ARMISD::tCALL";
525 case ARMISD::BRCOND: return "ARMISD::BRCOND";
526 case ARMISD::BR_JT: return "ARMISD::BR_JT";
527 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
528 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
529 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
530 case ARMISD::CMP: return "ARMISD::CMP";
531 case ARMISD::CMPZ: return "ARMISD::CMPZ";
532 case ARMISD::CMPFP: return "ARMISD::CMPFP";
533 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
534 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
535 case ARMISD::CMOV: return "ARMISD::CMOV";
536 case ARMISD::CNEG: return "ARMISD::CNEG";
538 case ARMISD::RBIT: return "ARMISD::RBIT";
540 case ARMISD::FTOSI: return "ARMISD::FTOSI";
541 case ARMISD::FTOUI: return "ARMISD::FTOUI";
542 case ARMISD::SITOF: return "ARMISD::SITOF";
543 case ARMISD::UITOF: return "ARMISD::UITOF";
545 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
546 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
547 case ARMISD::RRX: return "ARMISD::RRX";
549 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
550 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
552 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
553 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
555 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
557 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
559 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
561 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
562 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
564 case ARMISD::VCEQ: return "ARMISD::VCEQ";
565 case ARMISD::VCGE: return "ARMISD::VCGE";
566 case ARMISD::VCGEU: return "ARMISD::VCGEU";
567 case ARMISD::VCGT: return "ARMISD::VCGT";
568 case ARMISD::VCGTU: return "ARMISD::VCGTU";
569 case ARMISD::VTST: return "ARMISD::VTST";
571 case ARMISD::VSHL: return "ARMISD::VSHL";
572 case ARMISD::VSHRs: return "ARMISD::VSHRs";
573 case ARMISD::VSHRu: return "ARMISD::VSHRu";
574 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
575 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
576 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
577 case ARMISD::VSHRN: return "ARMISD::VSHRN";
578 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
579 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
580 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
581 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
582 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
583 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
584 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
585 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
586 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
587 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
588 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
589 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
590 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
591 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
592 case ARMISD::VDUP: return "ARMISD::VDUP";
593 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
594 case ARMISD::VEXT: return "ARMISD::VEXT";
595 case ARMISD::VREV64: return "ARMISD::VREV64";
596 case ARMISD::VREV32: return "ARMISD::VREV32";
597 case ARMISD::VREV16: return "ARMISD::VREV16";
598 case ARMISD::VZIP: return "ARMISD::VZIP";
599 case ARMISD::VUZP: return "ARMISD::VUZP";
600 case ARMISD::VTRN: return "ARMISD::VTRN";
601 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
602 case ARMISD::FMAX: return "ARMISD::FMAX";
603 case ARMISD::FMIN: return "ARMISD::FMIN";
607 /// getRegClassFor - Return the register class that should be used for the
608 /// specified value type.
609 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
610 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
611 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
612 // load / store 4 to 8 consecutive D registers.
613 if (Subtarget->hasNEON()) {
614 if (VT == MVT::v4i64)
615 return ARM::QQPRRegisterClass;
616 else if (VT == MVT::v8i64)
617 return ARM::QQQQPRRegisterClass;
619 return TargetLowering::getRegClassFor(VT);
622 /// getFunctionAlignment - Return the Log2 alignment of this function.
623 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
624 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
627 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
628 unsigned NumVals = N->getNumValues();
630 return Sched::RegPressure;
632 for (unsigned i = 0; i != NumVals; ++i) {
633 EVT VT = N->getValueType(i);
634 if (VT.isFloatingPoint() || VT.isVector())
635 return Sched::Latency;
638 if (!N->isMachineOpcode())
639 return Sched::RegPressure;
641 // Load are scheduled for latency even if there instruction itinerary
643 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
644 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
646 return Sched::Latency;
648 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
649 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
650 return Sched::Latency;
651 return Sched::RegPressure;
654 //===----------------------------------------------------------------------===//
656 //===----------------------------------------------------------------------===//
658 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
659 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
661 default: llvm_unreachable("Unknown condition code!");
662 case ISD::SETNE: return ARMCC::NE;
663 case ISD::SETEQ: return ARMCC::EQ;
664 case ISD::SETGT: return ARMCC::GT;
665 case ISD::SETGE: return ARMCC::GE;
666 case ISD::SETLT: return ARMCC::LT;
667 case ISD::SETLE: return ARMCC::LE;
668 case ISD::SETUGT: return ARMCC::HI;
669 case ISD::SETUGE: return ARMCC::HS;
670 case ISD::SETULT: return ARMCC::LO;
671 case ISD::SETULE: return ARMCC::LS;
675 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
676 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
677 ARMCC::CondCodes &CondCode2) {
678 CondCode2 = ARMCC::AL;
680 default: llvm_unreachable("Unknown FP condition!");
682 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
684 case ISD::SETOGT: CondCode = ARMCC::GT; break;
686 case ISD::SETOGE: CondCode = ARMCC::GE; break;
687 case ISD::SETOLT: CondCode = ARMCC::MI; break;
688 case ISD::SETOLE: CondCode = ARMCC::LS; break;
689 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
690 case ISD::SETO: CondCode = ARMCC::VC; break;
691 case ISD::SETUO: CondCode = ARMCC::VS; break;
692 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
693 case ISD::SETUGT: CondCode = ARMCC::HI; break;
694 case ISD::SETUGE: CondCode = ARMCC::PL; break;
696 case ISD::SETULT: CondCode = ARMCC::LT; break;
698 case ISD::SETULE: CondCode = ARMCC::LE; break;
700 case ISD::SETUNE: CondCode = ARMCC::NE; break;
704 //===----------------------------------------------------------------------===//
705 // Calling Convention Implementation
706 //===----------------------------------------------------------------------===//
708 #include "ARMGenCallingConv.inc"
710 // APCS f64 is in register pairs, possibly split to stack
711 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
712 CCValAssign::LocInfo &LocInfo,
713 CCState &State, bool CanFail) {
714 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
716 // Try to get the first register.
717 if (unsigned Reg = State.AllocateReg(RegList, 4))
718 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
720 // For the 2nd half of a v2f64, do not fail.
724 // Put the whole thing on the stack.
725 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
726 State.AllocateStack(8, 4),
731 // Try to get the second register.
732 if (unsigned Reg = State.AllocateReg(RegList, 4))
733 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
735 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
736 State.AllocateStack(4, 4),
741 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
742 CCValAssign::LocInfo &LocInfo,
743 ISD::ArgFlagsTy &ArgFlags,
745 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
747 if (LocVT == MVT::v2f64 &&
748 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
750 return true; // we handled it
753 // AAPCS f64 is in aligned register pairs
754 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
755 CCValAssign::LocInfo &LocInfo,
756 CCState &State, bool CanFail) {
757 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
758 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
760 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
762 // For the 2nd half of a v2f64, do not just fail.
766 // Put the whole thing on the stack.
767 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
768 State.AllocateStack(8, 8),
774 for (i = 0; i < 2; ++i)
775 if (HiRegList[i] == Reg)
778 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
779 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
784 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
785 CCValAssign::LocInfo &LocInfo,
786 ISD::ArgFlagsTy &ArgFlags,
788 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
790 if (LocVT == MVT::v2f64 &&
791 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
793 return true; // we handled it
796 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
797 CCValAssign::LocInfo &LocInfo, CCState &State) {
798 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
799 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
801 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
803 return false; // we didn't handle it
806 for (i = 0; i < 2; ++i)
807 if (HiRegList[i] == Reg)
810 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
811 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
816 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
817 CCValAssign::LocInfo &LocInfo,
818 ISD::ArgFlagsTy &ArgFlags,
820 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
822 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
824 return true; // we handled it
827 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
828 CCValAssign::LocInfo &LocInfo,
829 ISD::ArgFlagsTy &ArgFlags,
831 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
835 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
836 /// given CallingConvention value.
837 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
839 bool isVarArg) const {
842 llvm_unreachable("Unsupported calling convention");
844 case CallingConv::Fast:
845 // Use target triple & subtarget features to do actual dispatch.
846 if (Subtarget->isAAPCS_ABI()) {
847 if (Subtarget->hasVFP2() &&
848 FloatABIType == FloatABI::Hard && !isVarArg)
849 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
851 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
853 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
854 case CallingConv::ARM_AAPCS_VFP:
855 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
856 case CallingConv::ARM_AAPCS:
857 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
858 case CallingConv::ARM_APCS:
859 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
863 /// LowerCallResult - Lower the result values of a call into the
864 /// appropriate copies out of appropriate physical registers.
866 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
867 CallingConv::ID CallConv, bool isVarArg,
868 const SmallVectorImpl<ISD::InputArg> &Ins,
869 DebugLoc dl, SelectionDAG &DAG,
870 SmallVectorImpl<SDValue> &InVals) const {
872 // Assign locations to each value returned by this call.
873 SmallVector<CCValAssign, 16> RVLocs;
874 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
875 RVLocs, *DAG.getContext());
876 CCInfo.AnalyzeCallResult(Ins,
877 CCAssignFnForNode(CallConv, /* Return*/ true,
880 // Copy all of the result registers out of their specified physreg.
881 for (unsigned i = 0; i != RVLocs.size(); ++i) {
882 CCValAssign VA = RVLocs[i];
885 if (VA.needsCustom()) {
886 // Handle f64 or half of a v2f64.
887 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
889 Chain = Lo.getValue(1);
890 InFlag = Lo.getValue(2);
891 VA = RVLocs[++i]; // skip ahead to next loc
892 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
894 Chain = Hi.getValue(1);
895 InFlag = Hi.getValue(2);
896 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
898 if (VA.getLocVT() == MVT::v2f64) {
899 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
900 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
901 DAG.getConstant(0, MVT::i32));
903 VA = RVLocs[++i]; // skip ahead to next loc
904 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
905 Chain = Lo.getValue(1);
906 InFlag = Lo.getValue(2);
907 VA = RVLocs[++i]; // skip ahead to next loc
908 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
909 Chain = Hi.getValue(1);
910 InFlag = Hi.getValue(2);
911 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
912 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
913 DAG.getConstant(1, MVT::i32));
916 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
918 Chain = Val.getValue(1);
919 InFlag = Val.getValue(2);
922 switch (VA.getLocInfo()) {
923 default: llvm_unreachable("Unknown loc info!");
924 case CCValAssign::Full: break;
925 case CCValAssign::BCvt:
926 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
930 InVals.push_back(Val);
936 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
937 /// by "Src" to address "Dst" of size "Size". Alignment information is
938 /// specified by the specific parameter attribute. The copy will be passed as
939 /// a byval function parameter.
940 /// Sometimes what we are copying is the end of a larger object, the part that
941 /// does not fit in registers.
943 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
944 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
946 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
947 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
948 /*isVolatile=*/false, /*AlwaysInline=*/false,
952 /// LowerMemOpCallTo - Store the argument to the stack.
954 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
955 SDValue StackPtr, SDValue Arg,
956 DebugLoc dl, SelectionDAG &DAG,
957 const CCValAssign &VA,
958 ISD::ArgFlagsTy Flags) const {
959 unsigned LocMemOffset = VA.getLocMemOffset();
960 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
961 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
962 if (Flags.isByVal()) {
963 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
965 return DAG.getStore(Chain, dl, Arg, PtrOff,
966 PseudoSourceValue::getStack(), LocMemOffset,
970 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
971 SDValue Chain, SDValue &Arg,
972 RegsToPassVector &RegsToPass,
973 CCValAssign &VA, CCValAssign &NextVA,
975 SmallVector<SDValue, 8> &MemOpChains,
976 ISD::ArgFlagsTy Flags) const {
978 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
979 DAG.getVTList(MVT::i32, MVT::i32), Arg);
980 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
982 if (NextVA.isRegLoc())
983 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
985 assert(NextVA.isMemLoc());
986 if (StackPtr.getNode() == 0)
987 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
989 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
995 /// LowerCall - Lowering a call into a callseq_start <-
996 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
999 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1000 CallingConv::ID CallConv, bool isVarArg,
1002 const SmallVectorImpl<ISD::OutputArg> &Outs,
1003 const SmallVectorImpl<ISD::InputArg> &Ins,
1004 DebugLoc dl, SelectionDAG &DAG,
1005 SmallVectorImpl<SDValue> &InVals) const {
1006 MachineFunction &MF = DAG.getMachineFunction();
1007 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1008 bool IsSibCall = false;
1009 // Temporarily disable tail calls so things don't break.
1010 if (!EnableARMTailCalls)
1013 // Check if it's really possible to do a tail call.
1014 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1015 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1017 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1018 // detected sibcalls.
1025 // Analyze operands of the call, assigning locations to each operand.
1026 SmallVector<CCValAssign, 16> ArgLocs;
1027 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1029 CCInfo.AnalyzeCallOperands(Outs,
1030 CCAssignFnForNode(CallConv, /* Return*/ false,
1033 // Get a count of how many bytes are to be pushed on the stack.
1034 unsigned NumBytes = CCInfo.getNextStackOffset();
1036 // For tail calls, memory operands are available in our caller's stack.
1040 // Adjust the stack pointer for the new arguments...
1041 // These operations are automatically eliminated by the prolog/epilog pass
1043 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1045 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1047 RegsToPassVector RegsToPass;
1048 SmallVector<SDValue, 8> MemOpChains;
1050 // Walk the register/memloc assignments, inserting copies/loads. In the case
1051 // of tail call optimization, arguments are handled later.
1052 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1054 ++i, ++realArgIdx) {
1055 CCValAssign &VA = ArgLocs[i];
1056 SDValue Arg = Outs[realArgIdx].Val;
1057 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1059 // Promote the value if needed.
1060 switch (VA.getLocInfo()) {
1061 default: llvm_unreachable("Unknown loc info!");
1062 case CCValAssign::Full: break;
1063 case CCValAssign::SExt:
1064 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1066 case CCValAssign::ZExt:
1067 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1069 case CCValAssign::AExt:
1070 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1072 case CCValAssign::BCvt:
1073 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1077 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1078 if (VA.needsCustom()) {
1079 if (VA.getLocVT() == MVT::v2f64) {
1080 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1081 DAG.getConstant(0, MVT::i32));
1082 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1083 DAG.getConstant(1, MVT::i32));
1085 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1086 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1088 VA = ArgLocs[++i]; // skip ahead to next loc
1089 if (VA.isRegLoc()) {
1090 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1091 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1093 assert(VA.isMemLoc());
1095 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1096 dl, DAG, VA, Flags));
1099 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1100 StackPtr, MemOpChains, Flags);
1102 } else if (VA.isRegLoc()) {
1103 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1104 } else if (!IsSibCall) {
1105 assert(VA.isMemLoc());
1107 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1108 dl, DAG, VA, Flags));
1112 if (!MemOpChains.empty())
1113 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1114 &MemOpChains[0], MemOpChains.size());
1116 // Build a sequence of copy-to-reg nodes chained together with token chain
1117 // and flag operands which copy the outgoing args into the appropriate regs.
1119 // Tail call byval lowering might overwrite argument registers so in case of
1120 // tail call optimization the copies to registers are lowered later.
1122 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1123 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1124 RegsToPass[i].second, InFlag);
1125 InFlag = Chain.getValue(1);
1128 // For tail calls lower the arguments to the 'real' stack slot.
1130 // Force all the incoming stack arguments to be loaded from the stack
1131 // before any new outgoing arguments are stored to the stack, because the
1132 // outgoing stack slots may alias the incoming argument stack slots, and
1133 // the alias isn't otherwise explicit. This is slightly more conservative
1134 // than necessary, because it means that each store effectively depends
1135 // on every argument instead of just those arguments it would clobber.
1137 // Do not flag preceeding copytoreg stuff together with the following stuff.
1139 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1140 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1141 RegsToPass[i].second, InFlag);
1142 InFlag = Chain.getValue(1);
1147 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1148 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1149 // node so that legalize doesn't hack it.
1150 bool isDirect = false;
1151 bool isARMFunc = false;
1152 bool isLocalARMFunc = false;
1153 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1155 if (EnableARMLongCalls) {
1156 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1157 && "long-calls with non-static relocation model!");
1158 // Handle a global address or an external symbol. If it's not one of
1159 // those, the target's already in a register, so we don't need to do
1161 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1162 const GlobalValue *GV = G->getGlobal();
1163 // Create a constant pool entry for the callee address
1164 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1165 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1168 // Get the address of the callee into a register
1169 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1170 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1171 Callee = DAG.getLoad(getPointerTy(), dl,
1172 DAG.getEntryNode(), CPAddr,
1173 PseudoSourceValue::getConstantPool(), 0,
1175 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1176 const char *Sym = S->getSymbol();
1178 // Create a constant pool entry for the callee address
1179 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1180 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1181 Sym, ARMPCLabelIndex, 0);
1182 // Get the address of the callee into a register
1183 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1184 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1185 Callee = DAG.getLoad(getPointerTy(), dl,
1186 DAG.getEntryNode(), CPAddr,
1187 PseudoSourceValue::getConstantPool(), 0,
1190 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1191 const GlobalValue *GV = G->getGlobal();
1193 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1194 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1195 getTargetMachine().getRelocationModel() != Reloc::Static;
1196 isARMFunc = !Subtarget->isThumb() || isStub;
1197 // ARM call to a local ARM function is predicable.
1198 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1199 // tBX takes a register source operand.
1200 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1201 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1202 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1205 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1206 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1207 Callee = DAG.getLoad(getPointerTy(), dl,
1208 DAG.getEntryNode(), CPAddr,
1209 PseudoSourceValue::getConstantPool(), 0,
1211 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1212 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1213 getPointerTy(), Callee, PICLabel);
1215 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1216 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1218 bool isStub = Subtarget->isTargetDarwin() &&
1219 getTargetMachine().getRelocationModel() != Reloc::Static;
1220 isARMFunc = !Subtarget->isThumb() || isStub;
1221 // tBX takes a register source operand.
1222 const char *Sym = S->getSymbol();
1223 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1224 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1225 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1226 Sym, ARMPCLabelIndex, 4);
1227 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1228 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1229 Callee = DAG.getLoad(getPointerTy(), dl,
1230 DAG.getEntryNode(), CPAddr,
1231 PseudoSourceValue::getConstantPool(), 0,
1233 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1234 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1235 getPointerTy(), Callee, PICLabel);
1237 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1240 // FIXME: handle tail calls differently.
1242 if (Subtarget->isThumb()) {
1243 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1244 CallOpc = ARMISD::CALL_NOLINK;
1246 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1248 CallOpc = (isDirect || Subtarget->hasV5TOps())
1249 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1250 : ARMISD::CALL_NOLINK;
1252 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1253 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1254 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1255 InFlag = Chain.getValue(1);
1258 std::vector<SDValue> Ops;
1259 Ops.push_back(Chain);
1260 Ops.push_back(Callee);
1262 // Add argument registers to the end of the list so that they are known live
1264 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1265 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1266 RegsToPass[i].second.getValueType()));
1268 if (InFlag.getNode())
1269 Ops.push_back(InFlag);
1271 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1273 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1275 // Returns a chain and a flag for retval copy to use.
1276 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1277 InFlag = Chain.getValue(1);
1279 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1280 DAG.getIntPtrConstant(0, true), InFlag);
1282 InFlag = Chain.getValue(1);
1284 // Handle result values, copying them out of physregs into vregs that we
1286 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1290 /// MatchingStackOffset - Return true if the given stack call argument is
1291 /// already available in the same position (relatively) of the caller's
1292 /// incoming argument stack.
1294 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1295 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1296 const ARMInstrInfo *TII) {
1297 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1299 if (Arg.getOpcode() == ISD::CopyFromReg) {
1300 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1301 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1303 MachineInstr *Def = MRI->getVRegDef(VR);
1306 if (!Flags.isByVal()) {
1307 if (!TII->isLoadFromStackSlot(Def, FI))
1310 // unsigned Opcode = Def->getOpcode();
1311 // if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
1312 // Def->getOperand(1).isFI()) {
1313 // FI = Def->getOperand(1).getIndex();
1314 // Bytes = Flags.getByValSize();
1318 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1319 if (Flags.isByVal())
1320 // ByVal argument is passed in as a pointer but it's now being
1321 // dereferenced. e.g.
1322 // define @foo(%struct.X* %A) {
1323 // tail call @bar(%struct.X* byval %A)
1326 SDValue Ptr = Ld->getBasePtr();
1327 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1330 FI = FINode->getIndex();
1334 assert(FI != INT_MAX);
1335 if (!MFI->isFixedObjectIndex(FI))
1337 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1340 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1341 /// for tail call optimization. Targets which want to do tail call
1342 /// optimization should implement this function.
1344 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1345 CallingConv::ID CalleeCC,
1347 bool isCalleeStructRet,
1348 bool isCallerStructRet,
1349 const SmallVectorImpl<ISD::OutputArg> &Outs,
1350 const SmallVectorImpl<ISD::InputArg> &Ins,
1351 SelectionDAG& DAG) const {
1353 const Function *CallerF = DAG.getMachineFunction().getFunction();
1354 CallingConv::ID CallerCC = CallerF->getCallingConv();
1355 bool CCMatch = CallerCC == CalleeCC;
1357 // Look for obvious safe cases to perform tail call optimization that do not
1358 // require ABI changes. This is what gcc calls sibcall.
1360 // Do not sibcall optimize vararg calls unless the call site is not passing
1362 if (isVarArg && !Outs.empty())
1365 // Also avoid sibcall optimization if either caller or callee uses struct
1366 // return semantics.
1367 if (isCalleeStructRet || isCallerStructRet)
1370 // On Thumb, for the moment, we can only do this to functions defined in this
1371 // compilation, or to indirect calls. A Thumb B to an ARM function is not
1372 // easily fixed up in the linker, unlike BL.
1373 if (Subtarget->isThumb()) {
1374 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1375 const GlobalValue *GV = G->getGlobal();
1376 if (GV->isDeclaration() || GV->isWeakForLinker())
1378 } else if (isa<ExternalSymbolSDNode>(Callee)) {
1383 // If the calling conventions do not match, then we'd better make sure the
1384 // results are returned in the same way as what the caller expects.
1386 SmallVector<CCValAssign, 16> RVLocs1;
1387 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1388 RVLocs1, *DAG.getContext());
1389 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1391 SmallVector<CCValAssign, 16> RVLocs2;
1392 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1393 RVLocs2, *DAG.getContext());
1394 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1396 if (RVLocs1.size() != RVLocs2.size())
1398 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1399 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1401 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1403 if (RVLocs1[i].isRegLoc()) {
1404 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1407 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1413 // If the callee takes no arguments then go on to check the results of the
1415 if (!Outs.empty()) {
1416 // Check if stack adjustment is needed. For now, do not do this if any
1417 // argument is passed on the stack.
1418 SmallVector<CCValAssign, 16> ArgLocs;
1419 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1420 ArgLocs, *DAG.getContext());
1421 CCInfo.AnalyzeCallOperands(Outs,
1422 CCAssignFnForNode(CalleeCC, false, isVarArg));
1423 if (CCInfo.getNextStackOffset()) {
1424 MachineFunction &MF = DAG.getMachineFunction();
1426 // Check if the arguments are already laid out in the right way as
1427 // the caller's fixed stack objects.
1428 MachineFrameInfo *MFI = MF.getFrameInfo();
1429 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1430 const ARMInstrInfo *TII =
1431 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1432 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1434 ++i, ++realArgIdx) {
1435 CCValAssign &VA = ArgLocs[i];
1436 EVT RegVT = VA.getLocVT();
1437 SDValue Arg = Outs[realArgIdx].Val;
1438 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1439 if (VA.getLocInfo() == CCValAssign::Indirect)
1441 if (VA.needsCustom()) {
1442 // f64 and vector types are split into multiple registers or
1443 // register/stack-slot combinations. The types will not match
1444 // the registers; give up on memory f64 refs until we figure
1445 // out what to do about this.
1448 if (!ArgLocs[++i].isRegLoc())
1450 if (RegVT == MVT::v2f64) {
1451 if (!ArgLocs[++i].isRegLoc())
1453 if (!ArgLocs[++i].isRegLoc())
1456 } else if (!VA.isRegLoc()) {
1457 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1469 ARMTargetLowering::LowerReturn(SDValue Chain,
1470 CallingConv::ID CallConv, bool isVarArg,
1471 const SmallVectorImpl<ISD::OutputArg> &Outs,
1472 DebugLoc dl, SelectionDAG &DAG) const {
1474 // CCValAssign - represent the assignment of the return value to a location.
1475 SmallVector<CCValAssign, 16> RVLocs;
1477 // CCState - Info about the registers and stack slots.
1478 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1481 // Analyze outgoing return values.
1482 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1485 // If this is the first return lowered for this function, add
1486 // the regs to the liveout set for the function.
1487 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1488 for (unsigned i = 0; i != RVLocs.size(); ++i)
1489 if (RVLocs[i].isRegLoc())
1490 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1495 // Copy the result values into the output registers.
1496 for (unsigned i = 0, realRVLocIdx = 0;
1498 ++i, ++realRVLocIdx) {
1499 CCValAssign &VA = RVLocs[i];
1500 assert(VA.isRegLoc() && "Can only return in registers!");
1502 SDValue Arg = Outs[realRVLocIdx].Val;
1504 switch (VA.getLocInfo()) {
1505 default: llvm_unreachable("Unknown loc info!");
1506 case CCValAssign::Full: break;
1507 case CCValAssign::BCvt:
1508 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1512 if (VA.needsCustom()) {
1513 if (VA.getLocVT() == MVT::v2f64) {
1514 // Extract the first half and return it in two registers.
1515 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1516 DAG.getConstant(0, MVT::i32));
1517 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1518 DAG.getVTList(MVT::i32, MVT::i32), Half);
1520 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1521 Flag = Chain.getValue(1);
1522 VA = RVLocs[++i]; // skip ahead to next loc
1523 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1524 HalfGPRs.getValue(1), Flag);
1525 Flag = Chain.getValue(1);
1526 VA = RVLocs[++i]; // skip ahead to next loc
1528 // Extract the 2nd half and fall through to handle it as an f64 value.
1529 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1530 DAG.getConstant(1, MVT::i32));
1532 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1534 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1535 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1536 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1537 Flag = Chain.getValue(1);
1538 VA = RVLocs[++i]; // skip ahead to next loc
1539 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1542 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1544 // Guarantee that all emitted copies are
1545 // stuck together, avoiding something bad.
1546 Flag = Chain.getValue(1);
1551 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1553 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1558 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1559 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1560 // one of the above mentioned nodes. It has to be wrapped because otherwise
1561 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1562 // be used to form addressing mode. These wrapped nodes will be selected
1564 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1565 EVT PtrVT = Op.getValueType();
1566 // FIXME there is no actual debug info here
1567 DebugLoc dl = Op.getDebugLoc();
1568 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1570 if (CP->isMachineConstantPoolEntry())
1571 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1572 CP->getAlignment());
1574 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1575 CP->getAlignment());
1576 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1579 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1580 SelectionDAG &DAG) const {
1581 MachineFunction &MF = DAG.getMachineFunction();
1582 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1583 unsigned ARMPCLabelIndex = 0;
1584 DebugLoc DL = Op.getDebugLoc();
1585 EVT PtrVT = getPointerTy();
1586 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1587 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1589 if (RelocM == Reloc::Static) {
1590 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1592 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1593 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1594 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1595 ARMCP::CPBlockAddress,
1597 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1599 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1600 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1601 PseudoSourceValue::getConstantPool(), 0,
1603 if (RelocM == Reloc::Static)
1605 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1606 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1609 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1611 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1612 SelectionDAG &DAG) const {
1613 DebugLoc dl = GA->getDebugLoc();
1614 EVT PtrVT = getPointerTy();
1615 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1616 MachineFunction &MF = DAG.getMachineFunction();
1617 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1618 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1619 ARMConstantPoolValue *CPV =
1620 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1621 ARMCP::CPValue, PCAdj, "tlsgd", true);
1622 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1623 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1624 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1625 PseudoSourceValue::getConstantPool(), 0,
1627 SDValue Chain = Argument.getValue(1);
1629 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1630 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1632 // call __tls_get_addr.
1635 Entry.Node = Argument;
1636 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1637 Args.push_back(Entry);
1638 // FIXME: is there useful debug info available here?
1639 std::pair<SDValue, SDValue> CallResult =
1640 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1641 false, false, false, false,
1642 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1643 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1644 return CallResult.first;
1647 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1648 // "local exec" model.
1650 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1651 SelectionDAG &DAG) const {
1652 const GlobalValue *GV = GA->getGlobal();
1653 DebugLoc dl = GA->getDebugLoc();
1655 SDValue Chain = DAG.getEntryNode();
1656 EVT PtrVT = getPointerTy();
1657 // Get the Thread Pointer
1658 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1660 if (GV->isDeclaration()) {
1661 MachineFunction &MF = DAG.getMachineFunction();
1662 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1663 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1664 // Initial exec model.
1665 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1666 ARMConstantPoolValue *CPV =
1667 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1668 ARMCP::CPValue, PCAdj, "gottpoff", true);
1669 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1670 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1671 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1672 PseudoSourceValue::getConstantPool(), 0,
1674 Chain = Offset.getValue(1);
1676 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1677 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1679 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1680 PseudoSourceValue::getConstantPool(), 0,
1684 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1685 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1686 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1687 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1688 PseudoSourceValue::getConstantPool(), 0,
1692 // The address of the thread local variable is the add of the thread
1693 // pointer with the offset of the variable.
1694 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1698 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1699 // TODO: implement the "local dynamic" model
1700 assert(Subtarget->isTargetELF() &&
1701 "TLS not implemented for non-ELF targets");
1702 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1703 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1704 // otherwise use the "Local Exec" TLS Model
1705 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1706 return LowerToTLSGeneralDynamicModel(GA, DAG);
1708 return LowerToTLSExecModels(GA, DAG);
1711 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1712 SelectionDAG &DAG) const {
1713 EVT PtrVT = getPointerTy();
1714 DebugLoc dl = Op.getDebugLoc();
1715 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1716 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1717 if (RelocM == Reloc::PIC_) {
1718 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1719 ARMConstantPoolValue *CPV =
1720 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1721 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1722 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1723 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1725 PseudoSourceValue::getConstantPool(), 0,
1727 SDValue Chain = Result.getValue(1);
1728 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1729 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1731 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1732 PseudoSourceValue::getGOT(), 0,
1736 // If we have T2 ops, we can materialize the address directly via movt/movw
1737 // pair. This is always cheaper.
1738 if (Subtarget->useMovt()) {
1739 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1740 DAG.getTargetGlobalAddress(GV, PtrVT));
1742 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1743 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1744 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1745 PseudoSourceValue::getConstantPool(), 0,
1751 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1752 SelectionDAG &DAG) const {
1753 MachineFunction &MF = DAG.getMachineFunction();
1754 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1755 unsigned ARMPCLabelIndex = 0;
1756 EVT PtrVT = getPointerTy();
1757 DebugLoc dl = Op.getDebugLoc();
1758 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1759 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1761 if (RelocM == Reloc::Static)
1762 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1764 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1765 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1766 ARMConstantPoolValue *CPV =
1767 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1768 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1770 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1772 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1773 PseudoSourceValue::getConstantPool(), 0,
1775 SDValue Chain = Result.getValue(1);
1777 if (RelocM == Reloc::PIC_) {
1778 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1779 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1782 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1783 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1784 PseudoSourceValue::getGOT(), 0,
1790 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1791 SelectionDAG &DAG) const {
1792 assert(Subtarget->isTargetELF() &&
1793 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1794 MachineFunction &MF = DAG.getMachineFunction();
1795 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1796 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1797 EVT PtrVT = getPointerTy();
1798 DebugLoc dl = Op.getDebugLoc();
1799 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1800 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1801 "_GLOBAL_OFFSET_TABLE_",
1802 ARMPCLabelIndex, PCAdj);
1803 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1804 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1805 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1806 PseudoSourceValue::getConstantPool(), 0,
1808 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1809 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1813 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1814 DebugLoc dl = Op.getDebugLoc();
1815 SDValue Val = DAG.getConstant(0, MVT::i32);
1816 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1817 Op.getOperand(1), Val);
1821 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1822 DebugLoc dl = Op.getDebugLoc();
1823 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1824 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1828 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1829 const ARMSubtarget *Subtarget) const {
1830 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1831 DebugLoc dl = Op.getDebugLoc();
1833 default: return SDValue(); // Don't custom lower most intrinsics.
1834 case Intrinsic::arm_thread_pointer: {
1835 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1836 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1838 case Intrinsic::eh_sjlj_lsda: {
1839 MachineFunction &MF = DAG.getMachineFunction();
1840 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1841 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1842 EVT PtrVT = getPointerTy();
1843 DebugLoc dl = Op.getDebugLoc();
1844 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1846 unsigned PCAdj = (RelocM != Reloc::PIC_)
1847 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1848 ARMConstantPoolValue *CPV =
1849 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1850 ARMCP::CPLSDA, PCAdj);
1851 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1852 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1854 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1855 PseudoSourceValue::getConstantPool(), 0,
1857 SDValue Chain = Result.getValue(1);
1859 if (RelocM == Reloc::PIC_) {
1860 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1861 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1868 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1869 const ARMSubtarget *Subtarget) {
1870 DebugLoc dl = Op.getDebugLoc();
1871 SDValue Op5 = Op.getOperand(5);
1872 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1873 // v6 and v7 can both handle barriers directly, but need handled a bit
1874 // differently. Thumb1 and pre-v6 ARM mode use a libcall instead and should
1876 unsigned Opc = isDeviceBarrier ? ARMISD::SYNCBARRIER : ARMISD::MEMBARRIER;
1877 if (Subtarget->hasV7Ops())
1878 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0));
1879 else if (Subtarget->hasV6Ops() && !Subtarget->isThumb1Only())
1880 return DAG.getNode(Opc, dl, MVT::Other, Op.getOperand(0),
1881 DAG.getConstant(0, MVT::i32));
1882 assert(0 && "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
1886 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1887 MachineFunction &MF = DAG.getMachineFunction();
1888 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1890 // vastart just stores the address of the VarArgsFrameIndex slot into the
1891 // memory location argument.
1892 DebugLoc dl = Op.getDebugLoc();
1893 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1894 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1895 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1896 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1901 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1902 SelectionDAG &DAG) const {
1903 SDNode *Node = Op.getNode();
1904 DebugLoc dl = Node->getDebugLoc();
1905 EVT VT = Node->getValueType(0);
1906 SDValue Chain = Op.getOperand(0);
1907 SDValue Size = Op.getOperand(1);
1908 SDValue Align = Op.getOperand(2);
1910 // Chain the dynamic stack allocation so that it doesn't modify the stack
1911 // pointer when other instructions are using the stack.
1912 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1914 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1915 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1916 if (AlignVal > StackAlign)
1917 // Do this now since selection pass cannot introduce new target
1918 // independent node.
1919 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1921 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1922 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1923 // do even more horrible hack later.
1924 MachineFunction &MF = DAG.getMachineFunction();
1925 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1926 if (AFI->isThumb1OnlyFunction()) {
1928 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1930 uint32_t Val = C->getZExtValue();
1931 if (Val <= 508 && ((Val & 3) == 0))
1935 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1938 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1939 SDValue Ops1[] = { Chain, Size, Align };
1940 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1941 Chain = Res.getValue(1);
1942 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1943 DAG.getIntPtrConstant(0, true), SDValue());
1944 SDValue Ops2[] = { Res, Chain };
1945 return DAG.getMergeValues(Ops2, 2, dl);
1949 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1950 SDValue &Root, SelectionDAG &DAG,
1951 DebugLoc dl) const {
1952 MachineFunction &MF = DAG.getMachineFunction();
1953 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1955 TargetRegisterClass *RC;
1956 if (AFI->isThumb1OnlyFunction())
1957 RC = ARM::tGPRRegisterClass;
1959 RC = ARM::GPRRegisterClass;
1961 // Transform the arguments stored in physical registers into virtual ones.
1962 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1963 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1966 if (NextVA.isMemLoc()) {
1967 MachineFrameInfo *MFI = MF.getFrameInfo();
1968 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1970 // Create load node to retrieve arguments from the stack.
1971 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1972 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1973 PseudoSourceValue::getFixedStack(FI), 0,
1976 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1977 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1980 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1984 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1985 CallingConv::ID CallConv, bool isVarArg,
1986 const SmallVectorImpl<ISD::InputArg>
1988 DebugLoc dl, SelectionDAG &DAG,
1989 SmallVectorImpl<SDValue> &InVals)
1992 MachineFunction &MF = DAG.getMachineFunction();
1993 MachineFrameInfo *MFI = MF.getFrameInfo();
1995 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1997 // Assign locations to all of the incoming arguments.
1998 SmallVector<CCValAssign, 16> ArgLocs;
1999 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
2001 CCInfo.AnalyzeFormalArguments(Ins,
2002 CCAssignFnForNode(CallConv, /* Return*/ false,
2005 SmallVector<SDValue, 16> ArgValues;
2007 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2008 CCValAssign &VA = ArgLocs[i];
2010 // Arguments stored in registers.
2011 if (VA.isRegLoc()) {
2012 EVT RegVT = VA.getLocVT();
2015 if (VA.needsCustom()) {
2016 // f64 and vector types are split up into multiple registers or
2017 // combinations of registers and stack slots.
2018 if (VA.getLocVT() == MVT::v2f64) {
2019 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2021 VA = ArgLocs[++i]; // skip ahead to next loc
2023 if (VA.isMemLoc()) {
2024 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
2026 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2027 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2028 PseudoSourceValue::getFixedStack(FI), 0,
2031 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2034 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2035 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2036 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2037 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2038 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2040 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2043 TargetRegisterClass *RC;
2045 if (RegVT == MVT::f32)
2046 RC = ARM::SPRRegisterClass;
2047 else if (RegVT == MVT::f64)
2048 RC = ARM::DPRRegisterClass;
2049 else if (RegVT == MVT::v2f64)
2050 RC = ARM::QPRRegisterClass;
2051 else if (RegVT == MVT::i32)
2052 RC = (AFI->isThumb1OnlyFunction() ?
2053 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2055 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2057 // Transform the arguments in physical registers into virtual ones.
2058 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2059 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2062 // If this is an 8 or 16-bit value, it is really passed promoted
2063 // to 32 bits. Insert an assert[sz]ext to capture this, then
2064 // truncate to the right size.
2065 switch (VA.getLocInfo()) {
2066 default: llvm_unreachable("Unknown loc info!");
2067 case CCValAssign::Full: break;
2068 case CCValAssign::BCvt:
2069 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2071 case CCValAssign::SExt:
2072 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2073 DAG.getValueType(VA.getValVT()));
2074 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2076 case CCValAssign::ZExt:
2077 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2078 DAG.getValueType(VA.getValVT()));
2079 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2083 InVals.push_back(ArgValue);
2085 } else { // VA.isRegLoc()
2088 assert(VA.isMemLoc());
2089 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2091 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2092 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2095 // Create load nodes to retrieve arguments from the stack.
2096 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2097 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2098 PseudoSourceValue::getFixedStack(FI), 0,
2105 static const unsigned GPRArgRegs[] = {
2106 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2109 unsigned NumGPRs = CCInfo.getFirstUnallocated
2110 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2112 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2113 unsigned VARegSize = (4 - NumGPRs) * 4;
2114 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2115 unsigned ArgOffset = CCInfo.getNextStackOffset();
2116 if (VARegSaveSize) {
2117 // If this function is vararg, store any remaining integer argument regs
2118 // to their spots on the stack so that they may be loaded by deferencing
2119 // the result of va_next.
2120 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2121 AFI->setVarArgsFrameIndex(
2122 MFI->CreateFixedObject(VARegSaveSize,
2123 ArgOffset + VARegSaveSize - VARegSize,
2125 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2128 SmallVector<SDValue, 4> MemOps;
2129 for (; NumGPRs < 4; ++NumGPRs) {
2130 TargetRegisterClass *RC;
2131 if (AFI->isThumb1OnlyFunction())
2132 RC = ARM::tGPRRegisterClass;
2134 RC = ARM::GPRRegisterClass;
2136 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2137 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2139 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2140 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2141 0, false, false, 0);
2142 MemOps.push_back(Store);
2143 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2144 DAG.getConstant(4, getPointerTy()));
2146 if (!MemOps.empty())
2147 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2148 &MemOps[0], MemOps.size());
2150 // This will point to the next argument passed via stack.
2151 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
2158 /// isFloatingPointZero - Return true if this is +0.0.
2159 static bool isFloatingPointZero(SDValue Op) {
2160 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2161 return CFP->getValueAPF().isPosZero();
2162 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2163 // Maybe this has already been legalized into the constant pool?
2164 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2165 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2166 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2167 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2168 return CFP->getValueAPF().isPosZero();
2174 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2175 /// the given operands.
2177 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2178 SDValue &ARMCC, SelectionDAG &DAG,
2179 DebugLoc dl) const {
2180 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2181 unsigned C = RHSC->getZExtValue();
2182 if (!isLegalICmpImmediate(C)) {
2183 // Constant does not fit, try adjusting it by one?
2188 if (isLegalICmpImmediate(C-1)) {
2189 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2190 RHS = DAG.getConstant(C-1, MVT::i32);
2195 if (C > 0 && isLegalICmpImmediate(C-1)) {
2196 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2197 RHS = DAG.getConstant(C-1, MVT::i32);
2202 if (isLegalICmpImmediate(C+1)) {
2203 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2204 RHS = DAG.getConstant(C+1, MVT::i32);
2209 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2210 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2211 RHS = DAG.getConstant(C+1, MVT::i32);
2218 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2219 ARMISD::NodeType CompareType;
2222 CompareType = ARMISD::CMP;
2227 CompareType = ARMISD::CMPZ;
2230 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2231 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2234 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2235 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2238 if (!isFloatingPointZero(RHS))
2239 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2241 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2242 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2245 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2246 EVT VT = Op.getValueType();
2247 SDValue LHS = Op.getOperand(0);
2248 SDValue RHS = Op.getOperand(1);
2249 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2250 SDValue TrueVal = Op.getOperand(2);
2251 SDValue FalseVal = Op.getOperand(3);
2252 DebugLoc dl = Op.getDebugLoc();
2254 if (LHS.getValueType() == MVT::i32) {
2256 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2257 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2258 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2261 ARMCC::CondCodes CondCode, CondCode2;
2262 FPCCToARMCC(CC, CondCode, CondCode2);
2264 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2265 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2266 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2267 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2269 if (CondCode2 != ARMCC::AL) {
2270 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2271 // FIXME: Needs another CMP because flag can have but one use.
2272 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2273 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2274 Result, TrueVal, ARMCC2, CCR, Cmp2);
2279 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2280 SDValue Chain = Op.getOperand(0);
2281 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2282 SDValue LHS = Op.getOperand(2);
2283 SDValue RHS = Op.getOperand(3);
2284 SDValue Dest = Op.getOperand(4);
2285 DebugLoc dl = Op.getDebugLoc();
2287 if (LHS.getValueType() == MVT::i32) {
2289 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2290 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2291 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2292 Chain, Dest, ARMCC, CCR,Cmp);
2295 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2296 ARMCC::CondCodes CondCode, CondCode2;
2297 FPCCToARMCC(CC, CondCode, CondCode2);
2299 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2300 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2301 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2302 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2303 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2304 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2305 if (CondCode2 != ARMCC::AL) {
2306 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2307 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2308 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2313 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2314 SDValue Chain = Op.getOperand(0);
2315 SDValue Table = Op.getOperand(1);
2316 SDValue Index = Op.getOperand(2);
2317 DebugLoc dl = Op.getDebugLoc();
2319 EVT PTy = getPointerTy();
2320 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2321 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2322 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2323 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2324 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2325 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2326 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2327 if (Subtarget->isThumb2()) {
2328 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2329 // which does another jump to the destination. This also makes it easier
2330 // to translate it to TBB / TBH later.
2331 // FIXME: This might not work if the function is extremely large.
2332 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2333 Addr, Op.getOperand(2), JTI, UId);
2335 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2336 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2337 PseudoSourceValue::getJumpTable(), 0,
2339 Chain = Addr.getValue(1);
2340 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2341 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2343 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2344 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2345 Chain = Addr.getValue(1);
2346 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2350 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2351 DebugLoc dl = Op.getDebugLoc();
2354 switch (Op.getOpcode()) {
2356 assert(0 && "Invalid opcode!");
2357 case ISD::FP_TO_SINT:
2358 Opc = ARMISD::FTOSI;
2360 case ISD::FP_TO_UINT:
2361 Opc = ARMISD::FTOUI;
2364 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2365 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2368 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2369 EVT VT = Op.getValueType();
2370 DebugLoc dl = Op.getDebugLoc();
2373 switch (Op.getOpcode()) {
2375 assert(0 && "Invalid opcode!");
2376 case ISD::SINT_TO_FP:
2377 Opc = ARMISD::SITOF;
2379 case ISD::UINT_TO_FP:
2380 Opc = ARMISD::UITOF;
2384 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2385 return DAG.getNode(Opc, dl, VT, Op);
2388 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2389 // Implement fcopysign with a fabs and a conditional fneg.
2390 SDValue Tmp0 = Op.getOperand(0);
2391 SDValue Tmp1 = Op.getOperand(1);
2392 DebugLoc dl = Op.getDebugLoc();
2393 EVT VT = Op.getValueType();
2394 EVT SrcVT = Tmp1.getValueType();
2395 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2396 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2397 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2398 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2399 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2402 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2403 MachineFunction &MF = DAG.getMachineFunction();
2404 MachineFrameInfo *MFI = MF.getFrameInfo();
2405 MFI->setReturnAddressIsTaken(true);
2407 EVT VT = Op.getValueType();
2408 DebugLoc dl = Op.getDebugLoc();
2409 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2411 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2412 SDValue Offset = DAG.getConstant(4, MVT::i32);
2413 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2414 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2415 NULL, 0, false, false, 0);
2418 // Return LR, which contains the return address. Mark it an implicit live-in.
2419 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2420 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2423 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2424 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2425 MFI->setFrameAddressIsTaken(true);
2427 EVT VT = Op.getValueType();
2428 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2429 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2430 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2431 ? ARM::R7 : ARM::R11;
2432 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2434 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2439 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2440 /// expand a bit convert where either the source or destination type is i64 to
2441 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2442 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2443 /// vectors), since the legalizer won't know what to do with that.
2444 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2445 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2446 DebugLoc dl = N->getDebugLoc();
2447 SDValue Op = N->getOperand(0);
2449 // This function is only supposed to be called for i64 types, either as the
2450 // source or destination of the bit convert.
2451 EVT SrcVT = Op.getValueType();
2452 EVT DstVT = N->getValueType(0);
2453 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2454 "ExpandBIT_CONVERT called for non-i64 type");
2456 // Turn i64->f64 into VMOVDRR.
2457 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2458 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2459 DAG.getConstant(0, MVT::i32));
2460 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2461 DAG.getConstant(1, MVT::i32));
2462 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2463 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2466 // Turn f64->i64 into VMOVRRD.
2467 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2468 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2469 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2470 // Merge the pieces into a single i64 value.
2471 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2477 /// getZeroVector - Returns a vector of specified type with all zero elements.
2479 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2480 assert(VT.isVector() && "Expected a vector type");
2482 // Zero vectors are used to represent vector negation and in those cases
2483 // will be implemented with the NEON VNEG instruction. However, VNEG does
2484 // not support i64 elements, so sometimes the zero vectors will need to be
2485 // explicitly constructed. For those cases, and potentially other uses in
2486 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2487 // to their dest type. This ensures they get CSE'd.
2489 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2490 SmallVector<SDValue, 8> Ops;
2493 if (VT.getSizeInBits() == 64) {
2494 Ops.assign(8, Cst); TVT = MVT::v8i8;
2496 Ops.assign(16, Cst); TVT = MVT::v16i8;
2498 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2500 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2503 /// getOnesVector - Returns a vector of specified type with all bits set.
2505 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2506 assert(VT.isVector() && "Expected a vector type");
2508 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2509 // dest type. This ensures they get CSE'd.
2511 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2512 SmallVector<SDValue, 8> Ops;
2515 if (VT.getSizeInBits() == 64) {
2516 Ops.assign(8, Cst); TVT = MVT::v8i8;
2518 Ops.assign(16, Cst); TVT = MVT::v16i8;
2520 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2522 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2525 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2526 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2527 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2528 SelectionDAG &DAG) const {
2529 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2530 EVT VT = Op.getValueType();
2531 unsigned VTBits = VT.getSizeInBits();
2532 DebugLoc dl = Op.getDebugLoc();
2533 SDValue ShOpLo = Op.getOperand(0);
2534 SDValue ShOpHi = Op.getOperand(1);
2535 SDValue ShAmt = Op.getOperand(2);
2537 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2539 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2541 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2542 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2543 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2544 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2545 DAG.getConstant(VTBits, MVT::i32));
2546 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2547 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2548 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2550 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2551 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2553 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2554 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2557 SDValue Ops[2] = { Lo, Hi };
2558 return DAG.getMergeValues(Ops, 2, dl);
2561 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2562 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2563 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2564 SelectionDAG &DAG) const {
2565 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2566 EVT VT = Op.getValueType();
2567 unsigned VTBits = VT.getSizeInBits();
2568 DebugLoc dl = Op.getDebugLoc();
2569 SDValue ShOpLo = Op.getOperand(0);
2570 SDValue ShOpHi = Op.getOperand(1);
2571 SDValue ShAmt = Op.getOperand(2);
2574 assert(Op.getOpcode() == ISD::SHL_PARTS);
2575 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2576 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2577 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2578 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2579 DAG.getConstant(VTBits, MVT::i32));
2580 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2581 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2583 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2584 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2585 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2587 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2588 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2591 SDValue Ops[2] = { Lo, Hi };
2592 return DAG.getMergeValues(Ops, 2, dl);
2595 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2596 const ARMSubtarget *ST) {
2597 EVT VT = N->getValueType(0);
2598 DebugLoc dl = N->getDebugLoc();
2600 if (!ST->hasV6T2Ops())
2603 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2604 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2607 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2608 const ARMSubtarget *ST) {
2609 EVT VT = N->getValueType(0);
2610 DebugLoc dl = N->getDebugLoc();
2612 // Lower vector shifts on NEON to use VSHL.
2613 if (VT.isVector()) {
2614 assert(ST->hasNEON() && "unexpected vector shift");
2616 // Left shifts translate directly to the vshiftu intrinsic.
2617 if (N->getOpcode() == ISD::SHL)
2618 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2619 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2620 N->getOperand(0), N->getOperand(1));
2622 assert((N->getOpcode() == ISD::SRA ||
2623 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2625 // NEON uses the same intrinsics for both left and right shifts. For
2626 // right shifts, the shift amounts are negative, so negate the vector of
2628 EVT ShiftVT = N->getOperand(1).getValueType();
2629 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2630 getZeroVector(ShiftVT, DAG, dl),
2632 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2633 Intrinsic::arm_neon_vshifts :
2634 Intrinsic::arm_neon_vshiftu);
2635 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2636 DAG.getConstant(vshiftInt, MVT::i32),
2637 N->getOperand(0), NegatedCount);
2640 // We can get here for a node like i32 = ISD::SHL i32, i64
2644 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2645 "Unknown shift to lower!");
2647 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2648 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2649 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2652 // If we are in thumb mode, we don't have RRX.
2653 if (ST->isThumb1Only()) return SDValue();
2655 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2656 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2657 DAG.getConstant(0, MVT::i32));
2658 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2659 DAG.getConstant(1, MVT::i32));
2661 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2662 // captures the result into a carry flag.
2663 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2664 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2666 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2667 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2669 // Merge the pieces into a single i64 value.
2670 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2673 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2674 SDValue TmpOp0, TmpOp1;
2675 bool Invert = false;
2679 SDValue Op0 = Op.getOperand(0);
2680 SDValue Op1 = Op.getOperand(1);
2681 SDValue CC = Op.getOperand(2);
2682 EVT VT = Op.getValueType();
2683 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2684 DebugLoc dl = Op.getDebugLoc();
2686 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2687 switch (SetCCOpcode) {
2688 default: llvm_unreachable("Illegal FP comparison"); break;
2690 case ISD::SETNE: Invert = true; // Fallthrough
2692 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2694 case ISD::SETLT: Swap = true; // Fallthrough
2696 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2698 case ISD::SETLE: Swap = true; // Fallthrough
2700 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2701 case ISD::SETUGE: Swap = true; // Fallthrough
2702 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2703 case ISD::SETUGT: Swap = true; // Fallthrough
2704 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2705 case ISD::SETUEQ: Invert = true; // Fallthrough
2707 // Expand this to (OLT | OGT).
2711 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2712 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2714 case ISD::SETUO: Invert = true; // Fallthrough
2716 // Expand this to (OLT | OGE).
2720 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2721 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2725 // Integer comparisons.
2726 switch (SetCCOpcode) {
2727 default: llvm_unreachable("Illegal integer comparison"); break;
2728 case ISD::SETNE: Invert = true;
2729 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2730 case ISD::SETLT: Swap = true;
2731 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2732 case ISD::SETLE: Swap = true;
2733 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2734 case ISD::SETULT: Swap = true;
2735 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2736 case ISD::SETULE: Swap = true;
2737 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2740 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2741 if (Opc == ARMISD::VCEQ) {
2744 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2746 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2749 // Ignore bitconvert.
2750 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2751 AndOp = AndOp.getOperand(0);
2753 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2755 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2756 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2763 std::swap(Op0, Op1);
2765 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2768 Result = DAG.getNOT(dl, Result, VT);
2773 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2774 /// valid vector constant for a NEON instruction with a "modified immediate"
2775 /// operand (e.g., VMOV). If so, return either the constant being
2776 /// splatted or the encoded value, depending on the DoEncode parameter. The
2777 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2778 /// bits7-0=Immediate.
2779 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2780 unsigned SplatBitSize, SelectionDAG &DAG,
2781 bool isVMOV, bool DoEncode) {
2782 unsigned Op, Cmode, Imm;
2785 // SplatBitSize is set to the smallest size that splats the vector, so a
2786 // zero vector will always have SplatBitSize == 8. However, NEON modified
2787 // immediate instructions others than VMOV do not support the 8-bit encoding
2788 // of a zero vector, and the default encoding of zero is supposed to be the
2794 switch (SplatBitSize) {
2796 // Any 1-byte value is OK. Op=0, Cmode=1110.
2797 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2804 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2806 if ((SplatBits & ~0xff) == 0) {
2807 // Value = 0x00nn: Op=x, Cmode=100x.
2812 if ((SplatBits & ~0xff00) == 0) {
2813 // Value = 0xnn00: Op=x, Cmode=101x.
2815 Imm = SplatBits >> 8;
2821 // NEON's 32-bit VMOV supports splat values where:
2822 // * only one byte is nonzero, or
2823 // * the least significant byte is 0xff and the second byte is nonzero, or
2824 // * the least significant 2 bytes are 0xff and the third is nonzero.
2826 if ((SplatBits & ~0xff) == 0) {
2827 // Value = 0x000000nn: Op=x, Cmode=000x.
2832 if ((SplatBits & ~0xff00) == 0) {
2833 // Value = 0x0000nn00: Op=x, Cmode=001x.
2835 Imm = SplatBits >> 8;
2838 if ((SplatBits & ~0xff0000) == 0) {
2839 // Value = 0x00nn0000: Op=x, Cmode=010x.
2841 Imm = SplatBits >> 16;
2844 if ((SplatBits & ~0xff000000) == 0) {
2845 // Value = 0xnn000000: Op=x, Cmode=011x.
2847 Imm = SplatBits >> 24;
2851 if ((SplatBits & ~0xffff) == 0 &&
2852 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2853 // Value = 0x0000nnff: Op=x, Cmode=1100.
2855 Imm = SplatBits >> 8;
2860 if ((SplatBits & ~0xffffff) == 0 &&
2861 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2862 // Value = 0x00nnffff: Op=x, Cmode=1101.
2864 Imm = SplatBits >> 16;
2865 SplatBits |= 0xffff;
2869 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2870 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2871 // VMOV.I32. A (very) minor optimization would be to replicate the value
2872 // and fall through here to test for a valid 64-bit splat. But, then the
2873 // caller would also need to check and handle the change in size.
2877 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2880 uint64_t BitMask = 0xff;
2882 unsigned ImmMask = 1;
2884 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2885 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2888 } else if ((SplatBits & BitMask) != 0) {
2894 // Op=1, Cmode=1110.
2903 llvm_unreachable("unexpected size for EncodeNEONModImm");
2908 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2909 return DAG.getTargetConstant(SplatBits, VT);
2913 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2914 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2915 /// size, return the encoded value for that immediate. The ByteSize field
2916 /// indicates the number of bytes of each element [1248].
2917 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2918 SelectionDAG &DAG) {
2919 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2920 APInt SplatBits, SplatUndef;
2921 unsigned SplatBitSize;
2923 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2924 HasAnyUndefs, ByteSize * 8))
2927 if (SplatBitSize > ByteSize * 8)
2930 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2931 SplatBitSize, DAG, isVMOV, true);
2934 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2935 bool &ReverseVEXT, unsigned &Imm) {
2936 unsigned NumElts = VT.getVectorNumElements();
2937 ReverseVEXT = false;
2940 // If this is a VEXT shuffle, the immediate value is the index of the first
2941 // element. The other shuffle indices must be the successive elements after
2943 unsigned ExpectedElt = Imm;
2944 for (unsigned i = 1; i < NumElts; ++i) {
2945 // Increment the expected index. If it wraps around, it may still be
2946 // a VEXT but the source vectors must be swapped.
2948 if (ExpectedElt == NumElts * 2) {
2953 if (ExpectedElt != static_cast<unsigned>(M[i]))
2957 // Adjust the index value if the source operands will be swapped.
2964 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2965 /// instruction with the specified blocksize. (The order of the elements
2966 /// within each block of the vector is reversed.)
2967 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2968 unsigned BlockSize) {
2969 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2970 "Only possible block sizes for VREV are: 16, 32, 64");
2972 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2976 unsigned NumElts = VT.getVectorNumElements();
2977 unsigned BlockElts = M[0] + 1;
2979 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2982 for (unsigned i = 0; i < NumElts; ++i) {
2983 if ((unsigned) M[i] !=
2984 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2991 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2992 unsigned &WhichResult) {
2993 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2997 unsigned NumElts = VT.getVectorNumElements();
2998 WhichResult = (M[0] == 0 ? 0 : 1);
2999 for (unsigned i = 0; i < NumElts; i += 2) {
3000 if ((unsigned) M[i] != i + WhichResult ||
3001 (unsigned) M[i+1] != i + NumElts + WhichResult)
3007 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3008 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3009 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3010 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3011 unsigned &WhichResult) {
3012 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3016 unsigned NumElts = VT.getVectorNumElements();
3017 WhichResult = (M[0] == 0 ? 0 : 1);
3018 for (unsigned i = 0; i < NumElts; i += 2) {
3019 if ((unsigned) M[i] != i + WhichResult ||
3020 (unsigned) M[i+1] != i + WhichResult)
3026 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3027 unsigned &WhichResult) {
3028 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3032 unsigned NumElts = VT.getVectorNumElements();
3033 WhichResult = (M[0] == 0 ? 0 : 1);
3034 for (unsigned i = 0; i != NumElts; ++i) {
3035 if ((unsigned) M[i] != 2 * i + WhichResult)
3039 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3040 if (VT.is64BitVector() && EltSz == 32)
3046 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3047 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3048 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3049 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3050 unsigned &WhichResult) {
3051 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3055 unsigned Half = VT.getVectorNumElements() / 2;
3056 WhichResult = (M[0] == 0 ? 0 : 1);
3057 for (unsigned j = 0; j != 2; ++j) {
3058 unsigned Idx = WhichResult;
3059 for (unsigned i = 0; i != Half; ++i) {
3060 if ((unsigned) M[i + j * Half] != Idx)
3066 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3067 if (VT.is64BitVector() && EltSz == 32)
3073 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3074 unsigned &WhichResult) {
3075 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3079 unsigned NumElts = VT.getVectorNumElements();
3080 WhichResult = (M[0] == 0 ? 0 : 1);
3081 unsigned Idx = WhichResult * NumElts / 2;
3082 for (unsigned i = 0; i != NumElts; i += 2) {
3083 if ((unsigned) M[i] != Idx ||
3084 (unsigned) M[i+1] != Idx + NumElts)
3089 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3090 if (VT.is64BitVector() && EltSz == 32)
3096 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3097 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3098 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3099 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3100 unsigned &WhichResult) {
3101 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3105 unsigned NumElts = VT.getVectorNumElements();
3106 WhichResult = (M[0] == 0 ? 0 : 1);
3107 unsigned Idx = WhichResult * NumElts / 2;
3108 for (unsigned i = 0; i != NumElts; i += 2) {
3109 if ((unsigned) M[i] != Idx ||
3110 (unsigned) M[i+1] != Idx)
3115 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3116 if (VT.is64BitVector() && EltSz == 32)
3123 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3124 // Canonicalize all-zeros and all-ones vectors.
3125 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3126 if (ConstVal->isNullValue())
3127 return getZeroVector(VT, DAG, dl);
3128 if (ConstVal->isAllOnesValue())
3129 return getOnesVector(VT, DAG, dl);
3132 if (VT.is64BitVector()) {
3133 switch (Val.getValueType().getSizeInBits()) {
3134 case 8: CanonicalVT = MVT::v8i8; break;
3135 case 16: CanonicalVT = MVT::v4i16; break;
3136 case 32: CanonicalVT = MVT::v2i32; break;
3137 case 64: CanonicalVT = MVT::v1i64; break;
3138 default: llvm_unreachable("unexpected splat element type"); break;
3141 assert(VT.is128BitVector() && "unknown splat vector size");
3142 switch (Val.getValueType().getSizeInBits()) {
3143 case 8: CanonicalVT = MVT::v16i8; break;
3144 case 16: CanonicalVT = MVT::v8i16; break;
3145 case 32: CanonicalVT = MVT::v4i32; break;
3146 case 64: CanonicalVT = MVT::v2i64; break;
3147 default: llvm_unreachable("unexpected splat element type"); break;
3151 // Build a canonical splat for this value.
3152 SmallVector<SDValue, 8> Ops;
3153 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3154 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3156 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3159 // If this is a case we can't handle, return null and let the default
3160 // expansion code take care of it.
3161 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3162 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3163 DebugLoc dl = Op.getDebugLoc();
3164 EVT VT = Op.getValueType();
3166 APInt SplatBits, SplatUndef;
3167 unsigned SplatBitSize;
3169 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3170 if (SplatBitSize <= 64) {
3171 // Check if an immediate VMOV works.
3172 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3173 SplatUndef.getZExtValue(),
3174 SplatBitSize, DAG, true, false);
3176 return BuildSplat(Val, VT, DAG, dl);
3180 // Scan through the operands to see if only one value is used.
3181 unsigned NumElts = VT.getVectorNumElements();
3182 bool isOnlyLowElement = true;
3183 bool usesOnlyOneValue = true;
3184 bool isConstant = true;
3186 for (unsigned i = 0; i < NumElts; ++i) {
3187 SDValue V = Op.getOperand(i);
3188 if (V.getOpcode() == ISD::UNDEF)
3191 isOnlyLowElement = false;
3192 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3195 if (!Value.getNode())
3197 else if (V != Value)
3198 usesOnlyOneValue = false;
3201 if (!Value.getNode())
3202 return DAG.getUNDEF(VT);
3204 if (isOnlyLowElement)
3205 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3207 // If all elements are constants, fall back to the default expansion, which
3208 // will generate a load from the constant pool.
3212 // Use VDUP for non-constant splats.
3213 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3214 if (usesOnlyOneValue && EltSize <= 32)
3215 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3217 // Vectors with 32- or 64-bit elements can be built by directly assigning
3218 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3219 // will be legalized.
3220 if (EltSize >= 32) {
3221 // Do the expansion with floating-point types, since that is what the VFP
3222 // registers are defined to use, and since i64 is not legal.
3223 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3224 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3225 SmallVector<SDValue, 8> Ops;
3226 for (unsigned i = 0; i < NumElts; ++i)
3227 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3228 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3229 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3235 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3236 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3237 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3238 /// are assumed to be legal.
3240 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3242 if (VT.getVectorNumElements() == 4 &&
3243 (VT.is128BitVector() || VT.is64BitVector())) {
3244 unsigned PFIndexes[4];
3245 for (unsigned i = 0; i != 4; ++i) {
3249 PFIndexes[i] = M[i];
3252 // Compute the index in the perfect shuffle table.
3253 unsigned PFTableIndex =
3254 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3255 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3256 unsigned Cost = (PFEntry >> 30);
3263 unsigned Imm, WhichResult;
3265 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3266 return (EltSize >= 32 ||
3267 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3268 isVREVMask(M, VT, 64) ||
3269 isVREVMask(M, VT, 32) ||
3270 isVREVMask(M, VT, 16) ||
3271 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3272 isVTRNMask(M, VT, WhichResult) ||
3273 isVUZPMask(M, VT, WhichResult) ||
3274 isVZIPMask(M, VT, WhichResult) ||
3275 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3276 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3277 isVZIP_v_undef_Mask(M, VT, WhichResult));
3280 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3281 /// the specified operations to build the shuffle.
3282 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3283 SDValue RHS, SelectionDAG &DAG,
3285 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3286 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3287 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3290 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3299 OP_VUZPL, // VUZP, left result
3300 OP_VUZPR, // VUZP, right result
3301 OP_VZIPL, // VZIP, left result
3302 OP_VZIPR, // VZIP, right result
3303 OP_VTRNL, // VTRN, left result
3304 OP_VTRNR // VTRN, right result
3307 if (OpNum == OP_COPY) {
3308 if (LHSID == (1*9+2)*9+3) return LHS;
3309 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3313 SDValue OpLHS, OpRHS;
3314 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3315 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3316 EVT VT = OpLHS.getValueType();
3319 default: llvm_unreachable("Unknown shuffle opcode!");
3321 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3326 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3327 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3331 return DAG.getNode(ARMISD::VEXT, dl, VT,
3333 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3336 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3337 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3340 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3341 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3344 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3345 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3349 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3350 SDValue V1 = Op.getOperand(0);
3351 SDValue V2 = Op.getOperand(1);
3352 DebugLoc dl = Op.getDebugLoc();
3353 EVT VT = Op.getValueType();
3354 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3355 SmallVector<int, 8> ShuffleMask;
3357 // Convert shuffles that are directly supported on NEON to target-specific
3358 // DAG nodes, instead of keeping them as shuffles and matching them again
3359 // during code selection. This is more efficient and avoids the possibility
3360 // of inconsistencies between legalization and selection.
3361 // FIXME: floating-point vectors should be canonicalized to integer vectors
3362 // of the same time so that they get CSEd properly.
3363 SVN->getMask(ShuffleMask);
3365 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3366 if (EltSize <= 32) {
3367 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3368 int Lane = SVN->getSplatIndex();
3369 // If this is undef splat, generate it via "just" vdup, if possible.
3370 if (Lane == -1) Lane = 0;
3372 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3373 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3375 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3376 DAG.getConstant(Lane, MVT::i32));
3381 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3384 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3385 DAG.getConstant(Imm, MVT::i32));
3388 if (isVREVMask(ShuffleMask, VT, 64))
3389 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3390 if (isVREVMask(ShuffleMask, VT, 32))
3391 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3392 if (isVREVMask(ShuffleMask, VT, 16))
3393 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3395 // Check for Neon shuffles that modify both input vectors in place.
3396 // If both results are used, i.e., if there are two shuffles with the same
3397 // source operands and with masks corresponding to both results of one of
3398 // these operations, DAG memoization will ensure that a single node is
3399 // used for both shuffles.
3400 unsigned WhichResult;
3401 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3402 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3403 V1, V2).getValue(WhichResult);
3404 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3405 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3406 V1, V2).getValue(WhichResult);
3407 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3408 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3409 V1, V2).getValue(WhichResult);
3411 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3412 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3413 V1, V1).getValue(WhichResult);
3414 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3415 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3416 V1, V1).getValue(WhichResult);
3417 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3418 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3419 V1, V1).getValue(WhichResult);
3422 // If the shuffle is not directly supported and it has 4 elements, use
3423 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3424 unsigned NumElts = VT.getVectorNumElements();
3426 unsigned PFIndexes[4];
3427 for (unsigned i = 0; i != 4; ++i) {
3428 if (ShuffleMask[i] < 0)
3431 PFIndexes[i] = ShuffleMask[i];
3434 // Compute the index in the perfect shuffle table.
3435 unsigned PFTableIndex =
3436 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3437 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3438 unsigned Cost = (PFEntry >> 30);
3441 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3444 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3445 if (EltSize >= 32) {
3446 // Do the expansion with floating-point types, since that is what the VFP
3447 // registers are defined to use, and since i64 is not legal.
3448 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3449 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3450 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3451 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3452 SmallVector<SDValue, 8> Ops;
3453 for (unsigned i = 0; i < NumElts; ++i) {
3454 if (ShuffleMask[i] < 0)
3455 Ops.push_back(DAG.getUNDEF(EltVT));
3457 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3458 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3459 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3462 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3463 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3469 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3470 EVT VT = Op.getValueType();
3471 DebugLoc dl = Op.getDebugLoc();
3472 SDValue Vec = Op.getOperand(0);
3473 SDValue Lane = Op.getOperand(1);
3474 assert(VT == MVT::i32 &&
3475 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3476 "unexpected type for custom-lowering vector extract");
3477 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3480 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3481 // The only time a CONCAT_VECTORS operation can have legal types is when
3482 // two 64-bit vectors are concatenated to a 128-bit vector.
3483 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3484 "unexpected CONCAT_VECTORS");
3485 DebugLoc dl = Op.getDebugLoc();
3486 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3487 SDValue Op0 = Op.getOperand(0);
3488 SDValue Op1 = Op.getOperand(1);
3489 if (Op0.getOpcode() != ISD::UNDEF)
3490 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3491 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3492 DAG.getIntPtrConstant(0));
3493 if (Op1.getOpcode() != ISD::UNDEF)
3494 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3495 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3496 DAG.getIntPtrConstant(1));
3497 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3500 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3501 switch (Op.getOpcode()) {
3502 default: llvm_unreachable("Don't know how to custom lower this!");
3503 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3504 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3505 case ISD::GlobalAddress:
3506 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3507 LowerGlobalAddressELF(Op, DAG);
3508 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3509 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3510 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3511 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3512 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3513 case ISD::VASTART: return LowerVASTART(Op, DAG);
3514 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3515 case ISD::SINT_TO_FP:
3516 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3517 case ISD::FP_TO_SINT:
3518 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3519 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3520 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3521 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3522 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3523 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3524 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3525 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3527 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3530 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3531 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3532 case ISD::SRL_PARTS:
3533 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3534 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3535 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3536 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3537 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3538 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3539 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3544 /// ReplaceNodeResults - Replace the results of node with an illegal result
3545 /// type with new values built out of custom code.
3546 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3547 SmallVectorImpl<SDValue>&Results,
3548 SelectionDAG &DAG) const {
3550 switch (N->getOpcode()) {
3552 llvm_unreachable("Don't know how to custom expand this!");
3554 case ISD::BIT_CONVERT:
3555 Res = ExpandBIT_CONVERT(N, DAG);
3559 Res = LowerShift(N, DAG, Subtarget);
3563 Results.push_back(Res);
3566 //===----------------------------------------------------------------------===//
3567 // ARM Scheduler Hooks
3568 //===----------------------------------------------------------------------===//
3571 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3572 MachineBasicBlock *BB,
3573 unsigned Size) const {
3574 unsigned dest = MI->getOperand(0).getReg();
3575 unsigned ptr = MI->getOperand(1).getReg();
3576 unsigned oldval = MI->getOperand(2).getReg();
3577 unsigned newval = MI->getOperand(3).getReg();
3578 unsigned scratch = BB->getParent()->getRegInfo()
3579 .createVirtualRegister(ARM::GPRRegisterClass);
3580 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3581 DebugLoc dl = MI->getDebugLoc();
3582 bool isThumb2 = Subtarget->isThumb2();
3584 unsigned ldrOpc, strOpc;
3586 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3588 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3589 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3592 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3593 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3596 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3597 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3601 MachineFunction *MF = BB->getParent();
3602 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3603 MachineFunction::iterator It = BB;
3604 ++It; // insert the new blocks after the current block
3606 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3607 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3608 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3609 MF->insert(It, loop1MBB);
3610 MF->insert(It, loop2MBB);
3611 MF->insert(It, exitMBB);
3612 exitMBB->transferSuccessors(BB);
3616 // fallthrough --> loop1MBB
3617 BB->addSuccessor(loop1MBB);
3620 // ldrex dest, [ptr]
3624 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3625 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3626 .addReg(dest).addReg(oldval));
3627 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3628 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3629 BB->addSuccessor(loop2MBB);
3630 BB->addSuccessor(exitMBB);
3633 // strex scratch, newval, [ptr]
3637 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3639 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3640 .addReg(scratch).addImm(0));
3641 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3642 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3643 BB->addSuccessor(loop1MBB);
3644 BB->addSuccessor(exitMBB);
3650 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3656 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3657 unsigned Size, unsigned BinOpcode) const {
3658 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3659 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3661 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3662 MachineFunction *MF = BB->getParent();
3663 MachineFunction::iterator It = BB;
3666 unsigned dest = MI->getOperand(0).getReg();
3667 unsigned ptr = MI->getOperand(1).getReg();
3668 unsigned incr = MI->getOperand(2).getReg();
3669 DebugLoc dl = MI->getDebugLoc();
3671 bool isThumb2 = Subtarget->isThumb2();
3672 unsigned ldrOpc, strOpc;
3674 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3676 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3677 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3680 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3681 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3684 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3685 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3689 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3690 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3691 MF->insert(It, loopMBB);
3692 MF->insert(It, exitMBB);
3693 exitMBB->transferSuccessors(BB);
3695 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3696 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3697 unsigned scratch2 = (!BinOpcode) ? incr :
3698 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3702 // fallthrough --> loopMBB
3703 BB->addSuccessor(loopMBB);
3707 // <binop> scratch2, dest, incr
3708 // strex scratch, scratch2, ptr
3711 // fallthrough --> exitMBB
3713 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3715 // operand order needs to go the other way for NAND
3716 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3717 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3718 addReg(incr).addReg(dest)).addReg(0);
3720 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3721 addReg(dest).addReg(incr)).addReg(0);
3724 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3726 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3727 .addReg(scratch).addImm(0));
3728 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3729 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3731 BB->addSuccessor(loopMBB);
3732 BB->addSuccessor(exitMBB);
3738 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3744 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3745 MachineBasicBlock *BB) const {
3746 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3747 DebugLoc dl = MI->getDebugLoc();
3748 bool isThumb2 = Subtarget->isThumb2();
3749 switch (MI->getOpcode()) {
3752 llvm_unreachable("Unexpected instr type to insert");
3754 case ARM::ATOMIC_LOAD_ADD_I8:
3755 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3756 case ARM::ATOMIC_LOAD_ADD_I16:
3757 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3758 case ARM::ATOMIC_LOAD_ADD_I32:
3759 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3761 case ARM::ATOMIC_LOAD_AND_I8:
3762 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3763 case ARM::ATOMIC_LOAD_AND_I16:
3764 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3765 case ARM::ATOMIC_LOAD_AND_I32:
3766 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3768 case ARM::ATOMIC_LOAD_OR_I8:
3769 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3770 case ARM::ATOMIC_LOAD_OR_I16:
3771 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3772 case ARM::ATOMIC_LOAD_OR_I32:
3773 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3775 case ARM::ATOMIC_LOAD_XOR_I8:
3776 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3777 case ARM::ATOMIC_LOAD_XOR_I16:
3778 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3779 case ARM::ATOMIC_LOAD_XOR_I32:
3780 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3782 case ARM::ATOMIC_LOAD_NAND_I8:
3783 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3784 case ARM::ATOMIC_LOAD_NAND_I16:
3785 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3786 case ARM::ATOMIC_LOAD_NAND_I32:
3787 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3789 case ARM::ATOMIC_LOAD_SUB_I8:
3790 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3791 case ARM::ATOMIC_LOAD_SUB_I16:
3792 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3793 case ARM::ATOMIC_LOAD_SUB_I32:
3794 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3796 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3797 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3798 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3800 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3801 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3802 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3804 case ARM::tMOVCCr_pseudo: {
3805 // To "insert" a SELECT_CC instruction, we actually have to insert the
3806 // diamond control-flow pattern. The incoming instruction knows the
3807 // destination vreg to set, the condition code register to branch on, the
3808 // true/false values to select between, and a branch opcode to use.
3809 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3810 MachineFunction::iterator It = BB;
3816 // cmpTY ccX, r1, r2
3818 // fallthrough --> copy0MBB
3819 MachineBasicBlock *thisMBB = BB;
3820 MachineFunction *F = BB->getParent();
3821 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3822 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3823 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3824 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3825 F->insert(It, copy0MBB);
3826 F->insert(It, sinkMBB);
3827 // Update machine-CFG edges by first adding all successors of the current
3828 // block to the new block which will contain the Phi node for the select.
3829 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3830 E = BB->succ_end(); I != E; ++I)
3831 sinkMBB->addSuccessor(*I);
3832 // Next, remove all successors of the current block, and add the true
3833 // and fallthrough blocks as its successors.
3834 while (!BB->succ_empty())
3835 BB->removeSuccessor(BB->succ_begin());
3836 BB->addSuccessor(copy0MBB);
3837 BB->addSuccessor(sinkMBB);
3840 // %FalseValue = ...
3841 // # fallthrough to sinkMBB
3844 // Update machine-CFG edges
3845 BB->addSuccessor(sinkMBB);
3848 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3851 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3852 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3853 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3855 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3862 case ARM::t2SUBrSPi_:
3863 case ARM::t2SUBrSPi12_:
3864 case ARM::t2SUBrSPs_: {
3865 MachineFunction *MF = BB->getParent();
3866 unsigned DstReg = MI->getOperand(0).getReg();
3867 unsigned SrcReg = MI->getOperand(1).getReg();
3868 bool DstIsDead = MI->getOperand(0).isDead();
3869 bool SrcIsKill = MI->getOperand(1).isKill();
3871 if (SrcReg != ARM::SP) {
3872 // Copy the source to SP from virtual register.
3873 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3874 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3875 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3876 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3877 .addReg(SrcReg, getKillRegState(SrcIsKill));
3881 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3882 switch (MI->getOpcode()) {
3884 llvm_unreachable("Unexpected pseudo instruction!");
3890 OpOpc = ARM::tADDspr;
3893 OpOpc = ARM::tSUBspi;
3895 case ARM::t2SUBrSPi_:
3896 OpOpc = ARM::t2SUBrSPi;
3897 NeedPred = true; NeedCC = true;
3899 case ARM::t2SUBrSPi12_:
3900 OpOpc = ARM::t2SUBrSPi12;
3903 case ARM::t2SUBrSPs_:
3904 OpOpc = ARM::t2SUBrSPs;
3905 NeedPred = true; NeedCC = true; NeedOp3 = true;
3908 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3909 if (OpOpc == ARM::tAND)
3910 AddDefaultT1CC(MIB);
3911 MIB.addReg(ARM::SP);
3912 MIB.addOperand(MI->getOperand(2));
3914 MIB.addOperand(MI->getOperand(3));
3916 AddDefaultPred(MIB);
3920 // Copy the result from SP to virtual register.
3921 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3922 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3923 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3924 BuildMI(BB, dl, TII->get(CopyOpc))
3925 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3927 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3933 //===----------------------------------------------------------------------===//
3934 // ARM Optimization Hooks
3935 //===----------------------------------------------------------------------===//
3938 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3939 TargetLowering::DAGCombinerInfo &DCI) {
3940 SelectionDAG &DAG = DCI.DAG;
3941 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3942 EVT VT = N->getValueType(0);
3943 unsigned Opc = N->getOpcode();
3944 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3945 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3946 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3947 ISD::CondCode CC = ISD::SETCC_INVALID;
3950 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3952 SDValue CCOp = Slct.getOperand(0);
3953 if (CCOp.getOpcode() == ISD::SETCC)
3954 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3957 bool DoXform = false;
3959 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3962 if (LHS.getOpcode() == ISD::Constant &&
3963 cast<ConstantSDNode>(LHS)->isNullValue()) {
3965 } else if (CC != ISD::SETCC_INVALID &&
3966 RHS.getOpcode() == ISD::Constant &&
3967 cast<ConstantSDNode>(RHS)->isNullValue()) {
3968 std::swap(LHS, RHS);
3969 SDValue Op0 = Slct.getOperand(0);
3970 EVT OpVT = isSlctCC ? Op0.getValueType() :
3971 Op0.getOperand(0).getValueType();
3972 bool isInt = OpVT.isInteger();
3973 CC = ISD::getSetCCInverse(CC, isInt);
3975 if (!TLI.isCondCodeLegal(CC, OpVT))
3976 return SDValue(); // Inverse operator isn't legal.
3983 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3985 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3986 Slct.getOperand(0), Slct.getOperand(1), CC);
3987 SDValue CCOp = Slct.getOperand(0);
3989 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3990 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3991 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3992 CCOp, OtherOp, Result);
3997 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3998 static SDValue PerformADDCombine(SDNode *N,
3999 TargetLowering::DAGCombinerInfo &DCI) {
4000 // added by evan in r37685 with no testcase.
4001 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4003 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
4004 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
4005 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
4006 if (Result.getNode()) return Result;
4008 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4009 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4010 if (Result.getNode()) return Result;
4016 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4017 static SDValue PerformSUBCombine(SDNode *N,
4018 TargetLowering::DAGCombinerInfo &DCI) {
4019 // added by evan in r37685 with no testcase.
4020 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4022 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4023 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4024 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4025 if (Result.getNode()) return Result;
4031 static SDValue PerformMULCombine(SDNode *N,
4032 TargetLowering::DAGCombinerInfo &DCI,
4033 const ARMSubtarget *Subtarget) {
4034 SelectionDAG &DAG = DCI.DAG;
4036 if (Subtarget->isThumb1Only())
4039 if (DAG.getMachineFunction().
4040 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4043 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4046 EVT VT = N->getValueType(0);
4050 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4054 uint64_t MulAmt = C->getZExtValue();
4055 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4056 ShiftAmt = ShiftAmt & (32 - 1);
4057 SDValue V = N->getOperand(0);
4058 DebugLoc DL = N->getDebugLoc();
4061 MulAmt >>= ShiftAmt;
4062 if (isPowerOf2_32(MulAmt - 1)) {
4063 // (mul x, 2^N + 1) => (add (shl x, N), x)
4064 Res = DAG.getNode(ISD::ADD, DL, VT,
4065 V, DAG.getNode(ISD::SHL, DL, VT,
4066 V, DAG.getConstant(Log2_32(MulAmt-1),
4068 } else if (isPowerOf2_32(MulAmt + 1)) {
4069 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4070 Res = DAG.getNode(ISD::SUB, DL, VT,
4071 DAG.getNode(ISD::SHL, DL, VT,
4072 V, DAG.getConstant(Log2_32(MulAmt+1),
4079 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4080 DAG.getConstant(ShiftAmt, MVT::i32));
4082 // Do not add new nodes to DAG combiner worklist.
4083 DCI.CombineTo(N, Res, false);
4087 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4088 /// ARMISD::VMOVRRD.
4089 static SDValue PerformVMOVRRDCombine(SDNode *N,
4090 TargetLowering::DAGCombinerInfo &DCI) {
4091 // fmrrd(fmdrr x, y) -> x,y
4092 SDValue InDouble = N->getOperand(0);
4093 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4094 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4098 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4099 /// operand of a vector shift operation, where all the elements of the
4100 /// build_vector must have the same constant integer value.
4101 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4102 // Ignore bit_converts.
4103 while (Op.getOpcode() == ISD::BIT_CONVERT)
4104 Op = Op.getOperand(0);
4105 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4106 APInt SplatBits, SplatUndef;
4107 unsigned SplatBitSize;
4109 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4110 HasAnyUndefs, ElementBits) ||
4111 SplatBitSize > ElementBits)
4113 Cnt = SplatBits.getSExtValue();
4117 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4118 /// operand of a vector shift left operation. That value must be in the range:
4119 /// 0 <= Value < ElementBits for a left shift; or
4120 /// 0 <= Value <= ElementBits for a long left shift.
4121 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4122 assert(VT.isVector() && "vector shift count is not a vector type");
4123 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4124 if (! getVShiftImm(Op, ElementBits, Cnt))
4126 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4129 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4130 /// operand of a vector shift right operation. For a shift opcode, the value
4131 /// is positive, but for an intrinsic the value count must be negative. The
4132 /// absolute value must be in the range:
4133 /// 1 <= |Value| <= ElementBits for a right shift; or
4134 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4135 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4137 assert(VT.isVector() && "vector shift count is not a vector type");
4138 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4139 if (! getVShiftImm(Op, ElementBits, Cnt))
4143 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4146 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4147 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4148 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4151 // Don't do anything for most intrinsics.
4154 // Vector shifts: check for immediate versions and lower them.
4155 // Note: This is done during DAG combining instead of DAG legalizing because
4156 // the build_vectors for 64-bit vector element shift counts are generally
4157 // not legal, and it is hard to see their values after they get legalized to
4158 // loads from a constant pool.
4159 case Intrinsic::arm_neon_vshifts:
4160 case Intrinsic::arm_neon_vshiftu:
4161 case Intrinsic::arm_neon_vshiftls:
4162 case Intrinsic::arm_neon_vshiftlu:
4163 case Intrinsic::arm_neon_vshiftn:
4164 case Intrinsic::arm_neon_vrshifts:
4165 case Intrinsic::arm_neon_vrshiftu:
4166 case Intrinsic::arm_neon_vrshiftn:
4167 case Intrinsic::arm_neon_vqshifts:
4168 case Intrinsic::arm_neon_vqshiftu:
4169 case Intrinsic::arm_neon_vqshiftsu:
4170 case Intrinsic::arm_neon_vqshiftns:
4171 case Intrinsic::arm_neon_vqshiftnu:
4172 case Intrinsic::arm_neon_vqshiftnsu:
4173 case Intrinsic::arm_neon_vqrshiftns:
4174 case Intrinsic::arm_neon_vqrshiftnu:
4175 case Intrinsic::arm_neon_vqrshiftnsu: {
4176 EVT VT = N->getOperand(1).getValueType();
4178 unsigned VShiftOpc = 0;
4181 case Intrinsic::arm_neon_vshifts:
4182 case Intrinsic::arm_neon_vshiftu:
4183 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4184 VShiftOpc = ARMISD::VSHL;
4187 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4188 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4189 ARMISD::VSHRs : ARMISD::VSHRu);
4194 case Intrinsic::arm_neon_vshiftls:
4195 case Intrinsic::arm_neon_vshiftlu:
4196 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4198 llvm_unreachable("invalid shift count for vshll intrinsic");
4200 case Intrinsic::arm_neon_vrshifts:
4201 case Intrinsic::arm_neon_vrshiftu:
4202 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4206 case Intrinsic::arm_neon_vqshifts:
4207 case Intrinsic::arm_neon_vqshiftu:
4208 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4212 case Intrinsic::arm_neon_vqshiftsu:
4213 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4215 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4217 case Intrinsic::arm_neon_vshiftn:
4218 case Intrinsic::arm_neon_vrshiftn:
4219 case Intrinsic::arm_neon_vqshiftns:
4220 case Intrinsic::arm_neon_vqshiftnu:
4221 case Intrinsic::arm_neon_vqshiftnsu:
4222 case Intrinsic::arm_neon_vqrshiftns:
4223 case Intrinsic::arm_neon_vqrshiftnu:
4224 case Intrinsic::arm_neon_vqrshiftnsu:
4225 // Narrowing shifts require an immediate right shift.
4226 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4228 llvm_unreachable("invalid shift count for narrowing vector shift "
4232 llvm_unreachable("unhandled vector shift");
4236 case Intrinsic::arm_neon_vshifts:
4237 case Intrinsic::arm_neon_vshiftu:
4238 // Opcode already set above.
4240 case Intrinsic::arm_neon_vshiftls:
4241 case Intrinsic::arm_neon_vshiftlu:
4242 if (Cnt == VT.getVectorElementType().getSizeInBits())
4243 VShiftOpc = ARMISD::VSHLLi;
4245 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4246 ARMISD::VSHLLs : ARMISD::VSHLLu);
4248 case Intrinsic::arm_neon_vshiftn:
4249 VShiftOpc = ARMISD::VSHRN; break;
4250 case Intrinsic::arm_neon_vrshifts:
4251 VShiftOpc = ARMISD::VRSHRs; break;
4252 case Intrinsic::arm_neon_vrshiftu:
4253 VShiftOpc = ARMISD::VRSHRu; break;
4254 case Intrinsic::arm_neon_vrshiftn:
4255 VShiftOpc = ARMISD::VRSHRN; break;
4256 case Intrinsic::arm_neon_vqshifts:
4257 VShiftOpc = ARMISD::VQSHLs; break;
4258 case Intrinsic::arm_neon_vqshiftu:
4259 VShiftOpc = ARMISD::VQSHLu; break;
4260 case Intrinsic::arm_neon_vqshiftsu:
4261 VShiftOpc = ARMISD::VQSHLsu; break;
4262 case Intrinsic::arm_neon_vqshiftns:
4263 VShiftOpc = ARMISD::VQSHRNs; break;
4264 case Intrinsic::arm_neon_vqshiftnu:
4265 VShiftOpc = ARMISD::VQSHRNu; break;
4266 case Intrinsic::arm_neon_vqshiftnsu:
4267 VShiftOpc = ARMISD::VQSHRNsu; break;
4268 case Intrinsic::arm_neon_vqrshiftns:
4269 VShiftOpc = ARMISD::VQRSHRNs; break;
4270 case Intrinsic::arm_neon_vqrshiftnu:
4271 VShiftOpc = ARMISD::VQRSHRNu; break;
4272 case Intrinsic::arm_neon_vqrshiftnsu:
4273 VShiftOpc = ARMISD::VQRSHRNsu; break;
4276 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4277 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4280 case Intrinsic::arm_neon_vshiftins: {
4281 EVT VT = N->getOperand(1).getValueType();
4283 unsigned VShiftOpc = 0;
4285 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4286 VShiftOpc = ARMISD::VSLI;
4287 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4288 VShiftOpc = ARMISD::VSRI;
4290 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4293 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4294 N->getOperand(1), N->getOperand(2),
4295 DAG.getConstant(Cnt, MVT::i32));
4298 case Intrinsic::arm_neon_vqrshifts:
4299 case Intrinsic::arm_neon_vqrshiftu:
4300 // No immediate versions of these to check for.
4307 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4308 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4309 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4310 /// vector element shift counts are generally not legal, and it is hard to see
4311 /// their values after they get legalized to loads from a constant pool.
4312 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4313 const ARMSubtarget *ST) {
4314 EVT VT = N->getValueType(0);
4316 // Nothing to be done for scalar shifts.
4317 if (! VT.isVector())
4320 assert(ST->hasNEON() && "unexpected vector shift");
4323 switch (N->getOpcode()) {
4324 default: llvm_unreachable("unexpected shift opcode");
4327 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4328 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4329 DAG.getConstant(Cnt, MVT::i32));
4334 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4335 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4336 ARMISD::VSHRs : ARMISD::VSHRu);
4337 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4338 DAG.getConstant(Cnt, MVT::i32));
4344 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4345 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4346 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4347 const ARMSubtarget *ST) {
4348 SDValue N0 = N->getOperand(0);
4350 // Check for sign- and zero-extensions of vector extract operations of 8-
4351 // and 16-bit vector elements. NEON supports these directly. They are
4352 // handled during DAG combining because type legalization will promote them
4353 // to 32-bit types and it is messy to recognize the operations after that.
4354 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4355 SDValue Vec = N0.getOperand(0);
4356 SDValue Lane = N0.getOperand(1);
4357 EVT VT = N->getValueType(0);
4358 EVT EltVT = N0.getValueType();
4359 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4361 if (VT == MVT::i32 &&
4362 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4363 TLI.isTypeLegal(Vec.getValueType())) {
4366 switch (N->getOpcode()) {
4367 default: llvm_unreachable("unexpected opcode");
4368 case ISD::SIGN_EXTEND:
4369 Opc = ARMISD::VGETLANEs;
4371 case ISD::ZERO_EXTEND:
4372 case ISD::ANY_EXTEND:
4373 Opc = ARMISD::VGETLANEu;
4376 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4383 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4384 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4385 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4386 const ARMSubtarget *ST) {
4387 // If the target supports NEON, try to use vmax/vmin instructions for f32
4388 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4389 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4390 // a NaN; only do the transformation when it matches that behavior.
4392 // For now only do this when using NEON for FP operations; if using VFP, it
4393 // is not obvious that the benefit outweighs the cost of switching to the
4395 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4396 N->getValueType(0) != MVT::f32)
4399 SDValue CondLHS = N->getOperand(0);
4400 SDValue CondRHS = N->getOperand(1);
4401 SDValue LHS = N->getOperand(2);
4402 SDValue RHS = N->getOperand(3);
4403 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4405 unsigned Opcode = 0;
4407 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4408 IsReversed = false; // x CC y ? x : y
4409 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4410 IsReversed = true ; // x CC y ? y : x
4424 // If LHS is NaN, an ordered comparison will be false and the result will
4425 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4426 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4427 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4428 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4430 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4431 // will return -0, so vmin can only be used for unsafe math or if one of
4432 // the operands is known to be nonzero.
4433 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4435 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4437 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4446 // If LHS is NaN, an ordered comparison will be false and the result will
4447 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4448 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4449 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4450 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4452 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4453 // will return +0, so vmax can only be used for unsafe math or if one of
4454 // the operands is known to be nonzero.
4455 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4457 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4459 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4465 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4468 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4469 DAGCombinerInfo &DCI) const {
4470 switch (N->getOpcode()) {
4472 case ISD::ADD: return PerformADDCombine(N, DCI);
4473 case ISD::SUB: return PerformSUBCombine(N, DCI);
4474 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4475 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4476 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4479 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4480 case ISD::SIGN_EXTEND:
4481 case ISD::ZERO_EXTEND:
4482 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4483 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4488 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4489 if (!Subtarget->hasV6Ops())
4490 // Pre-v6 does not support unaligned mem access.
4493 // v6+ may or may not support unaligned mem access depending on the system
4495 // FIXME: This is pretty conservative. Should we provide cmdline option to
4496 // control the behaviour?
4497 if (!Subtarget->isTargetDarwin())
4501 switch (VT.getSimpleVT().SimpleTy) {
4508 // FIXME: VLD1 etc with standard alignment is legal.
4512 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4517 switch (VT.getSimpleVT().SimpleTy) {
4518 default: return false;
4533 if ((V & (Scale - 1)) != 0)
4536 return V == (V & ((1LL << 5) - 1));
4539 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4540 const ARMSubtarget *Subtarget) {
4547 switch (VT.getSimpleVT().SimpleTy) {
4548 default: return false;
4553 // + imm12 or - imm8
4555 return V == (V & ((1LL << 8) - 1));
4556 return V == (V & ((1LL << 12) - 1));
4559 // Same as ARM mode. FIXME: NEON?
4560 if (!Subtarget->hasVFP2())
4565 return V == (V & ((1LL << 8) - 1));
4569 /// isLegalAddressImmediate - Return true if the integer value can be used
4570 /// as the offset of the target addressing mode for load / store of the
4572 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4573 const ARMSubtarget *Subtarget) {
4580 if (Subtarget->isThumb1Only())
4581 return isLegalT1AddressImmediate(V, VT);
4582 else if (Subtarget->isThumb2())
4583 return isLegalT2AddressImmediate(V, VT, Subtarget);
4588 switch (VT.getSimpleVT().SimpleTy) {
4589 default: return false;
4594 return V == (V & ((1LL << 12) - 1));
4597 return V == (V & ((1LL << 8) - 1));
4600 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4605 return V == (V & ((1LL << 8) - 1));
4609 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4611 int Scale = AM.Scale;
4615 switch (VT.getSimpleVT().SimpleTy) {
4616 default: return false;
4625 return Scale == 2 || Scale == 4 || Scale == 8;
4628 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4632 // Note, we allow "void" uses (basically, uses that aren't loads or
4633 // stores), because arm allows folding a scale into many arithmetic
4634 // operations. This should be made more precise and revisited later.
4636 // Allow r << imm, but the imm has to be a multiple of two.
4637 if (Scale & 1) return false;
4638 return isPowerOf2_32(Scale);
4642 /// isLegalAddressingMode - Return true if the addressing mode represented
4643 /// by AM is legal for this target, for a load/store of the specified type.
4644 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4645 const Type *Ty) const {
4646 EVT VT = getValueType(Ty, true);
4647 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4650 // Can never fold addr of global into load/store.
4655 case 0: // no scale reg, must be "r+i" or "r", or "i".
4658 if (Subtarget->isThumb1Only())
4662 // ARM doesn't support any R+R*scale+imm addr modes.
4669 if (Subtarget->isThumb2())
4670 return isLegalT2ScaledAddressingMode(AM, VT);
4672 int Scale = AM.Scale;
4673 switch (VT.getSimpleVT().SimpleTy) {
4674 default: return false;
4678 if (Scale < 0) Scale = -Scale;
4682 return isPowerOf2_32(Scale & ~1);
4686 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4691 // Note, we allow "void" uses (basically, uses that aren't loads or
4692 // stores), because arm allows folding a scale into many arithmetic
4693 // operations. This should be made more precise and revisited later.
4695 // Allow r << imm, but the imm has to be a multiple of two.
4696 if (Scale & 1) return false;
4697 return isPowerOf2_32(Scale);
4704 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4705 /// icmp immediate, that is the target has icmp instructions which can compare
4706 /// a register against the immediate without having to materialize the
4707 /// immediate into a register.
4708 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4709 if (!Subtarget->isThumb())
4710 return ARM_AM::getSOImmVal(Imm) != -1;
4711 if (Subtarget->isThumb2())
4712 return ARM_AM::getT2SOImmVal(Imm) != -1;
4713 return Imm >= 0 && Imm <= 255;
4716 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4717 bool isSEXTLoad, SDValue &Base,
4718 SDValue &Offset, bool &isInc,
4719 SelectionDAG &DAG) {
4720 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4723 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4725 Base = Ptr->getOperand(0);
4726 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4727 int RHSC = (int)RHS->getZExtValue();
4728 if (RHSC < 0 && RHSC > -256) {
4729 assert(Ptr->getOpcode() == ISD::ADD);
4731 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4735 isInc = (Ptr->getOpcode() == ISD::ADD);
4736 Offset = Ptr->getOperand(1);
4738 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4740 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4741 int RHSC = (int)RHS->getZExtValue();
4742 if (RHSC < 0 && RHSC > -0x1000) {
4743 assert(Ptr->getOpcode() == ISD::ADD);
4745 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4746 Base = Ptr->getOperand(0);
4751 if (Ptr->getOpcode() == ISD::ADD) {
4753 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4754 if (ShOpcVal != ARM_AM::no_shift) {
4755 Base = Ptr->getOperand(1);
4756 Offset = Ptr->getOperand(0);
4758 Base = Ptr->getOperand(0);
4759 Offset = Ptr->getOperand(1);
4764 isInc = (Ptr->getOpcode() == ISD::ADD);
4765 Base = Ptr->getOperand(0);
4766 Offset = Ptr->getOperand(1);
4770 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4774 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4775 bool isSEXTLoad, SDValue &Base,
4776 SDValue &Offset, bool &isInc,
4777 SelectionDAG &DAG) {
4778 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4781 Base = Ptr->getOperand(0);
4782 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4783 int RHSC = (int)RHS->getZExtValue();
4784 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4785 assert(Ptr->getOpcode() == ISD::ADD);
4787 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4789 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4790 isInc = Ptr->getOpcode() == ISD::ADD;
4791 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4799 /// getPreIndexedAddressParts - returns true by value, base pointer and
4800 /// offset pointer and addressing mode by reference if the node's address
4801 /// can be legally represented as pre-indexed load / store address.
4803 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4805 ISD::MemIndexedMode &AM,
4806 SelectionDAG &DAG) const {
4807 if (Subtarget->isThumb1Only())
4812 bool isSEXTLoad = false;
4813 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4814 Ptr = LD->getBasePtr();
4815 VT = LD->getMemoryVT();
4816 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4817 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4818 Ptr = ST->getBasePtr();
4819 VT = ST->getMemoryVT();
4824 bool isLegal = false;
4825 if (Subtarget->isThumb2())
4826 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4827 Offset, isInc, DAG);
4829 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4830 Offset, isInc, DAG);
4834 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4838 /// getPostIndexedAddressParts - returns true by value, base pointer and
4839 /// offset pointer and addressing mode by reference if this node can be
4840 /// combined with a load / store to form a post-indexed load / store.
4841 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4844 ISD::MemIndexedMode &AM,
4845 SelectionDAG &DAG) const {
4846 if (Subtarget->isThumb1Only())
4851 bool isSEXTLoad = false;
4852 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4853 VT = LD->getMemoryVT();
4854 Ptr = LD->getBasePtr();
4855 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4856 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4857 VT = ST->getMemoryVT();
4858 Ptr = ST->getBasePtr();
4863 bool isLegal = false;
4864 if (Subtarget->isThumb2())
4865 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4868 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4874 // Swap base ptr and offset to catch more post-index load / store when
4875 // it's legal. In Thumb2 mode, offset must be an immediate.
4876 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4877 !Subtarget->isThumb2())
4878 std::swap(Base, Offset);
4880 // Post-indexed load / store update the base pointer.
4885 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4889 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4893 const SelectionDAG &DAG,
4894 unsigned Depth) const {
4895 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4896 switch (Op.getOpcode()) {
4898 case ARMISD::CMOV: {
4899 // Bits are known zero/one if known on the LHS and RHS.
4900 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4901 if (KnownZero == 0 && KnownOne == 0) return;
4903 APInt KnownZeroRHS, KnownOneRHS;
4904 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4905 KnownZeroRHS, KnownOneRHS, Depth+1);
4906 KnownZero &= KnownZeroRHS;
4907 KnownOne &= KnownOneRHS;
4913 //===----------------------------------------------------------------------===//
4914 // ARM Inline Assembly Support
4915 //===----------------------------------------------------------------------===//
4917 /// getConstraintType - Given a constraint letter, return the type of
4918 /// constraint it is for this target.
4919 ARMTargetLowering::ConstraintType
4920 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4921 if (Constraint.size() == 1) {
4922 switch (Constraint[0]) {
4924 case 'l': return C_RegisterClass;
4925 case 'w': return C_RegisterClass;
4928 return TargetLowering::getConstraintType(Constraint);
4931 std::pair<unsigned, const TargetRegisterClass*>
4932 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4934 if (Constraint.size() == 1) {
4935 // GCC ARM Constraint Letters
4936 switch (Constraint[0]) {
4938 if (Subtarget->isThumb())
4939 return std::make_pair(0U, ARM::tGPRRegisterClass);
4941 return std::make_pair(0U, ARM::GPRRegisterClass);
4943 return std::make_pair(0U, ARM::GPRRegisterClass);
4946 return std::make_pair(0U, ARM::SPRRegisterClass);
4947 if (VT.getSizeInBits() == 64)
4948 return std::make_pair(0U, ARM::DPRRegisterClass);
4949 if (VT.getSizeInBits() == 128)
4950 return std::make_pair(0U, ARM::QPRRegisterClass);
4954 if (StringRef("{cc}").equals_lower(Constraint))
4955 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
4957 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4960 std::vector<unsigned> ARMTargetLowering::
4961 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4963 if (Constraint.size() != 1)
4964 return std::vector<unsigned>();
4966 switch (Constraint[0]) { // GCC ARM Constraint Letters
4969 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4970 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4973 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4974 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4975 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4976 ARM::R12, ARM::LR, 0);
4979 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4980 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4981 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4982 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4983 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4984 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4985 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4986 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4987 if (VT.getSizeInBits() == 64)
4988 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4989 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4990 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4991 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4992 if (VT.getSizeInBits() == 128)
4993 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4994 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4998 return std::vector<unsigned>();
5001 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5002 /// vector. If it is invalid, don't add anything to Ops.
5003 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5006 std::vector<SDValue>&Ops,
5007 SelectionDAG &DAG) const {
5008 SDValue Result(0, 0);
5010 switch (Constraint) {
5012 case 'I': case 'J': case 'K': case 'L':
5013 case 'M': case 'N': case 'O':
5014 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5018 int64_t CVal64 = C->getSExtValue();
5019 int CVal = (int) CVal64;
5020 // None of these constraints allow values larger than 32 bits. Check
5021 // that the value fits in an int.
5025 switch (Constraint) {
5027 if (Subtarget->isThumb1Only()) {
5028 // This must be a constant between 0 and 255, for ADD
5030 if (CVal >= 0 && CVal <= 255)
5032 } else if (Subtarget->isThumb2()) {
5033 // A constant that can be used as an immediate value in a
5034 // data-processing instruction.
5035 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5038 // A constant that can be used as an immediate value in a
5039 // data-processing instruction.
5040 if (ARM_AM::getSOImmVal(CVal) != -1)
5046 if (Subtarget->isThumb()) { // FIXME thumb2
5047 // This must be a constant between -255 and -1, for negated ADD
5048 // immediates. This can be used in GCC with an "n" modifier that
5049 // prints the negated value, for use with SUB instructions. It is
5050 // not useful otherwise but is implemented for compatibility.
5051 if (CVal >= -255 && CVal <= -1)
5054 // This must be a constant between -4095 and 4095. It is not clear
5055 // what this constraint is intended for. Implemented for
5056 // compatibility with GCC.
5057 if (CVal >= -4095 && CVal <= 4095)
5063 if (Subtarget->isThumb1Only()) {
5064 // A 32-bit value where only one byte has a nonzero value. Exclude
5065 // zero to match GCC. This constraint is used by GCC internally for
5066 // constants that can be loaded with a move/shift combination.
5067 // It is not useful otherwise but is implemented for compatibility.
5068 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5070 } else if (Subtarget->isThumb2()) {
5071 // A constant whose bitwise inverse can be used as an immediate
5072 // value in a data-processing instruction. This can be used in GCC
5073 // with a "B" modifier that prints the inverted value, for use with
5074 // BIC and MVN instructions. It is not useful otherwise but is
5075 // implemented for compatibility.
5076 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5079 // A constant whose bitwise inverse can be used as an immediate
5080 // value in a data-processing instruction. This can be used in GCC
5081 // with a "B" modifier that prints the inverted value, for use with
5082 // BIC and MVN instructions. It is not useful otherwise but is
5083 // implemented for compatibility.
5084 if (ARM_AM::getSOImmVal(~CVal) != -1)
5090 if (Subtarget->isThumb1Only()) {
5091 // This must be a constant between -7 and 7,
5092 // for 3-operand ADD/SUB immediate instructions.
5093 if (CVal >= -7 && CVal < 7)
5095 } else if (Subtarget->isThumb2()) {
5096 // A constant whose negation can be used as an immediate value in a
5097 // data-processing instruction. This can be used in GCC with an "n"
5098 // modifier that prints the negated value, for use with SUB
5099 // instructions. It is not useful otherwise but is implemented for
5101 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5104 // A constant whose negation can be used as an immediate value in a
5105 // data-processing instruction. This can be used in GCC with an "n"
5106 // modifier that prints the negated value, for use with SUB
5107 // instructions. It is not useful otherwise but is implemented for
5109 if (ARM_AM::getSOImmVal(-CVal) != -1)
5115 if (Subtarget->isThumb()) { // FIXME thumb2
5116 // This must be a multiple of 4 between 0 and 1020, for
5117 // ADD sp + immediate.
5118 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5121 // A power of two or a constant between 0 and 32. This is used in
5122 // GCC for the shift amount on shifted register operands, but it is
5123 // useful in general for any shift amounts.
5124 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5130 if (Subtarget->isThumb()) { // FIXME thumb2
5131 // This must be a constant between 0 and 31, for shift amounts.
5132 if (CVal >= 0 && CVal <= 31)
5138 if (Subtarget->isThumb()) { // FIXME thumb2
5139 // This must be a multiple of 4 between -508 and 508, for
5140 // ADD/SUB sp = sp + immediate.
5141 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5146 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5150 if (Result.getNode()) {
5151 Ops.push_back(Result);
5154 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
5159 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5160 // The ARM target isn't yet aware of offsets.
5164 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5165 APInt Imm = FPImm.bitcastToAPInt();
5166 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5167 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5168 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5170 // We can handle 4 bits of mantissa.
5171 // mantissa = (16+UInt(e:f:g:h))/16.
5172 if (Mantissa & 0x7ffff)
5175 if ((Mantissa & 0xf) != Mantissa)
5178 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5179 if (Exp < -3 || Exp > 4)
5181 Exp = ((Exp+3) & 0x7) ^ 4;
5183 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5186 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5187 APInt Imm = FPImm.bitcastToAPInt();
5188 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5189 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5190 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5192 // We can handle 4 bits of mantissa.
5193 // mantissa = (16+UInt(e:f:g:h))/16.
5194 if (Mantissa & 0xffffffffffffLL)
5197 if ((Mantissa & 0xf) != Mantissa)
5200 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5201 if (Exp < -3 || Exp > 4)
5203 Exp = ((Exp+3) & 0x7) ^ 4;
5205 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5208 /// isFPImmLegal - Returns true if the target can instruction select the
5209 /// specified FP immediate natively. If false, the legalizer will
5210 /// materialize the FP immediate as a load from a constant pool.
5211 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5212 if (!Subtarget->hasVFP3())
5215 return ARM::getVFPf32Imm(Imm) != -1;
5217 return ARM::getVFPf64Imm(Imm) != -1;