1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/Instruction.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Type.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/ADT/VectorExtras.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
49 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
53 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
54 CCValAssign::LocInfo &LocInfo,
55 ISD::ArgFlagsTy &ArgFlags,
57 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
58 CCValAssign::LocInfo &LocInfo,
59 ISD::ArgFlagsTy &ArgFlags,
61 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
62 CCValAssign::LocInfo &LocInfo,
63 ISD::ArgFlagsTy &ArgFlags,
66 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
67 EVT PromotedBitwiseVT) {
68 if (VT != PromotedLdStVT) {
69 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
70 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
71 PromotedLdStVT.getSimpleVT());
73 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
74 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
75 PromotedLdStVT.getSimpleVT());
78 EVT ElemTy = VT.getVectorElementType();
79 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
80 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
81 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
82 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
83 if (ElemTy != MVT::i32) {
84 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
85 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
86 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
87 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
89 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
90 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
91 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
92 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
94 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
95 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
99 // Promote all bit-wise operations.
100 if (VT.isInteger() && VT != PromotedBitwiseVT) {
101 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
102 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
103 PromotedBitwiseVT.getSimpleVT());
104 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
105 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
106 PromotedBitwiseVT.getSimpleVT());
107 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
108 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
109 PromotedBitwiseVT.getSimpleVT());
112 // Neon does not support vector divide/remainder operations.
113 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
121 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
122 addRegisterClass(VT, ARM::DPRRegisterClass);
123 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
126 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
127 addRegisterClass(VT, ARM::QPRRegisterClass);
128 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
131 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
132 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
133 return new TargetLoweringObjectFileMachO();
134 return new ARMElfTargetObjectFile();
137 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
138 : TargetLowering(TM, createTLOF(TM)) {
139 Subtarget = &TM.getSubtarget<ARMSubtarget>();
141 if (Subtarget->isTargetDarwin()) {
142 // Uses VFP for Thumb libfuncs if available.
143 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
144 // Single-precision floating-point arithmetic.
145 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
146 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
147 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
148 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
150 // Double-precision floating-point arithmetic.
151 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
152 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
153 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
154 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
156 // Single-precision comparisons.
157 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
158 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
159 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
160 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
161 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
162 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
163 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
164 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
166 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
167 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
172 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
173 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
175 // Double-precision comparisons.
176 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
177 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
178 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
179 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
180 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
181 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
182 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
183 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
185 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
194 // Floating-point to integer conversions.
195 // i64 conversions are done via library routines even when generating VFP
196 // instructions, so use the same ones.
197 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
198 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
199 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
200 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
202 // Conversions between floating types.
203 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
204 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
206 // Integer to floating-point conversions.
207 // i64 conversions are done via library routines even when generating VFP
208 // instructions, so use the same ones.
209 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
210 // e.g., __floatunsidf vs. __floatunssidfvfp.
211 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
212 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
213 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
214 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
218 // These libcalls are not available in 32-bit.
219 setLibcallName(RTLIB::SHL_I128, 0);
220 setLibcallName(RTLIB::SRL_I128, 0);
221 setLibcallName(RTLIB::SRA_I128, 0);
223 // Libcalls should use the AAPCS base standard ABI, even if hard float
224 // is in effect, as per the ARM RTABI specification, section 4.1.2.
225 if (Subtarget->isAAPCS_ABI()) {
226 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
227 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
228 CallingConv::ARM_AAPCS);
232 if (Subtarget->isThumb1Only())
233 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
235 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
236 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
237 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
238 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
240 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
243 if (Subtarget->hasNEON()) {
244 addDRTypeForNEON(MVT::v2f32);
245 addDRTypeForNEON(MVT::v8i8);
246 addDRTypeForNEON(MVT::v4i16);
247 addDRTypeForNEON(MVT::v2i32);
248 addDRTypeForNEON(MVT::v1i64);
250 addQRTypeForNEON(MVT::v4f32);
251 addQRTypeForNEON(MVT::v2f64);
252 addQRTypeForNEON(MVT::v16i8);
253 addQRTypeForNEON(MVT::v8i16);
254 addQRTypeForNEON(MVT::v4i32);
255 addQRTypeForNEON(MVT::v2i64);
257 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
258 // neither Neon nor VFP support any arithmetic operations on it.
259 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
260 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
261 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
262 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
263 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
264 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
265 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
266 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
267 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
268 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
269 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
270 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
271 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
272 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
273 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
274 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
275 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
276 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
277 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
278 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
279 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
280 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
281 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
282 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
284 // Neon does not support some operations on v1i64 and v2i64 types.
285 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
286 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
287 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
288 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
290 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
291 setTargetDAGCombine(ISD::SHL);
292 setTargetDAGCombine(ISD::SRL);
293 setTargetDAGCombine(ISD::SRA);
294 setTargetDAGCombine(ISD::SIGN_EXTEND);
295 setTargetDAGCombine(ISD::ZERO_EXTEND);
296 setTargetDAGCombine(ISD::ANY_EXTEND);
297 setTargetDAGCombine(ISD::SELECT_CC);
300 computeRegisterProperties();
302 // ARM does not have f32 extending load.
303 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
305 // ARM does not have i1 sign extending load.
306 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
308 // ARM supports all 4 flavors of integer indexed load / store.
309 if (!Subtarget->isThumb1Only()) {
310 for (unsigned im = (unsigned)ISD::PRE_INC;
311 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
312 setIndexedLoadAction(im, MVT::i1, Legal);
313 setIndexedLoadAction(im, MVT::i8, Legal);
314 setIndexedLoadAction(im, MVT::i16, Legal);
315 setIndexedLoadAction(im, MVT::i32, Legal);
316 setIndexedStoreAction(im, MVT::i1, Legal);
317 setIndexedStoreAction(im, MVT::i8, Legal);
318 setIndexedStoreAction(im, MVT::i16, Legal);
319 setIndexedStoreAction(im, MVT::i32, Legal);
323 // i64 operation support.
324 if (Subtarget->isThumb1Only()) {
325 setOperationAction(ISD::MUL, MVT::i64, Expand);
326 setOperationAction(ISD::MULHU, MVT::i32, Expand);
327 setOperationAction(ISD::MULHS, MVT::i32, Expand);
328 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
329 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
331 setOperationAction(ISD::MUL, MVT::i64, Expand);
332 setOperationAction(ISD::MULHU, MVT::i32, Expand);
333 if (!Subtarget->hasV6Ops())
334 setOperationAction(ISD::MULHS, MVT::i32, Expand);
336 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
337 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
338 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
339 setOperationAction(ISD::SRL, MVT::i64, Custom);
340 setOperationAction(ISD::SRA, MVT::i64, Custom);
342 // ARM does not have ROTL.
343 setOperationAction(ISD::ROTL, MVT::i32, Expand);
344 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
345 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
346 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
347 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
349 // Only ARMv6 has BSWAP.
350 if (!Subtarget->hasV6Ops())
351 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
353 // These are expanded into libcalls.
354 setOperationAction(ISD::SDIV, MVT::i32, Expand);
355 setOperationAction(ISD::UDIV, MVT::i32, Expand);
356 setOperationAction(ISD::SREM, MVT::i32, Expand);
357 setOperationAction(ISD::UREM, MVT::i32, Expand);
358 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
359 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
361 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
362 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
363 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
364 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
365 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
367 // Use the default implementation.
368 setOperationAction(ISD::VASTART, MVT::Other, Custom);
369 setOperationAction(ISD::VAARG, MVT::Other, Expand);
370 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
371 setOperationAction(ISD::VAEND, MVT::Other, Expand);
372 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
373 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
374 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
375 // FIXME: Shouldn't need this, since no register is used, but the legalizer
376 // doesn't yet know how to not do that for SjLj.
377 setExceptionSelectorRegister(ARM::R0);
378 if (Subtarget->isThumb())
379 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
381 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
382 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
384 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
385 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
386 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
388 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
390 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
391 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
392 // iff target supports vfp2.
393 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
395 // We want to custom lower some of our intrinsics.
396 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
398 setOperationAction(ISD::SETCC, MVT::i32, Expand);
399 setOperationAction(ISD::SETCC, MVT::f32, Expand);
400 setOperationAction(ISD::SETCC, MVT::f64, Expand);
401 setOperationAction(ISD::SELECT, MVT::i32, Expand);
402 setOperationAction(ISD::SELECT, MVT::f32, Expand);
403 setOperationAction(ISD::SELECT, MVT::f64, Expand);
404 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
405 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
406 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
408 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
409 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
410 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
411 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
412 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
414 // We don't support sin/cos/fmod/copysign/pow
415 setOperationAction(ISD::FSIN, MVT::f64, Expand);
416 setOperationAction(ISD::FSIN, MVT::f32, Expand);
417 setOperationAction(ISD::FCOS, MVT::f32, Expand);
418 setOperationAction(ISD::FCOS, MVT::f64, Expand);
419 setOperationAction(ISD::FREM, MVT::f64, Expand);
420 setOperationAction(ISD::FREM, MVT::f32, Expand);
421 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
422 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
423 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
425 setOperationAction(ISD::FPOW, MVT::f64, Expand);
426 setOperationAction(ISD::FPOW, MVT::f32, Expand);
428 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
429 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
430 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
431 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
432 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
433 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
436 // We have target-specific dag combine patterns for the following nodes:
437 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
438 setTargetDAGCombine(ISD::ADD);
439 setTargetDAGCombine(ISD::SUB);
441 setStackPointerRegisterToSaveRestore(ARM::SP);
442 setSchedulingPreference(SchedulingForRegPressure);
444 // FIXME: If-converter should use instruction latency to determine
445 // profitability rather than relying on fixed limits.
446 if (Subtarget->getCPUString() == "generic") {
447 // Generic (and overly aggressive) if-conversion limits.
448 setIfCvtBlockSizeLimit(10);
449 setIfCvtDupBlockSizeLimit(2);
450 } else if (Subtarget->hasV6Ops()) {
451 setIfCvtBlockSizeLimit(2);
452 setIfCvtDupBlockSizeLimit(1);
454 setIfCvtBlockSizeLimit(3);
455 setIfCvtDupBlockSizeLimit(2);
458 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
459 // Do not enable CodePlacementOpt for now: it currently runs after the
460 // ARMConstantIslandPass and messes up branch relaxation and placement
461 // of constant islands.
462 // benefitFromCodePlacementOpt = true;
465 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
468 case ARMISD::Wrapper: return "ARMISD::Wrapper";
469 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
470 case ARMISD::CALL: return "ARMISD::CALL";
471 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
472 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
473 case ARMISD::tCALL: return "ARMISD::tCALL";
474 case ARMISD::BRCOND: return "ARMISD::BRCOND";
475 case ARMISD::BR_JT: return "ARMISD::BR_JT";
476 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
477 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
478 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
479 case ARMISD::CMP: return "ARMISD::CMP";
480 case ARMISD::CMPZ: return "ARMISD::CMPZ";
481 case ARMISD::CMPFP: return "ARMISD::CMPFP";
482 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
483 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
484 case ARMISD::CMOV: return "ARMISD::CMOV";
485 case ARMISD::CNEG: return "ARMISD::CNEG";
487 case ARMISD::RBIT: return "ARMISD::RBIT";
489 case ARMISD::FTOSI: return "ARMISD::FTOSI";
490 case ARMISD::FTOUI: return "ARMISD::FTOUI";
491 case ARMISD::SITOF: return "ARMISD::SITOF";
492 case ARMISD::UITOF: return "ARMISD::UITOF";
494 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
495 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
496 case ARMISD::RRX: return "ARMISD::RRX";
498 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
499 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
501 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
502 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
504 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
506 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
508 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
509 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
511 case ARMISD::VCEQ: return "ARMISD::VCEQ";
512 case ARMISD::VCGE: return "ARMISD::VCGE";
513 case ARMISD::VCGEU: return "ARMISD::VCGEU";
514 case ARMISD::VCGT: return "ARMISD::VCGT";
515 case ARMISD::VCGTU: return "ARMISD::VCGTU";
516 case ARMISD::VTST: return "ARMISD::VTST";
518 case ARMISD::VSHL: return "ARMISD::VSHL";
519 case ARMISD::VSHRs: return "ARMISD::VSHRs";
520 case ARMISD::VSHRu: return "ARMISD::VSHRu";
521 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
522 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
523 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
524 case ARMISD::VSHRN: return "ARMISD::VSHRN";
525 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
526 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
527 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
528 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
529 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
530 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
531 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
532 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
533 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
534 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
535 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
536 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
537 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
538 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
539 case ARMISD::VDUP: return "ARMISD::VDUP";
540 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
541 case ARMISD::VEXT: return "ARMISD::VEXT";
542 case ARMISD::VREV64: return "ARMISD::VREV64";
543 case ARMISD::VREV32: return "ARMISD::VREV32";
544 case ARMISD::VREV16: return "ARMISD::VREV16";
545 case ARMISD::VZIP: return "ARMISD::VZIP";
546 case ARMISD::VUZP: return "ARMISD::VUZP";
547 case ARMISD::VTRN: return "ARMISD::VTRN";
548 case ARMISD::FMAX: return "ARMISD::FMAX";
549 case ARMISD::FMIN: return "ARMISD::FMIN";
553 /// getFunctionAlignment - Return the Log2 alignment of this function.
554 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
555 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
558 //===----------------------------------------------------------------------===//
560 //===----------------------------------------------------------------------===//
562 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
563 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
565 default: llvm_unreachable("Unknown condition code!");
566 case ISD::SETNE: return ARMCC::NE;
567 case ISD::SETEQ: return ARMCC::EQ;
568 case ISD::SETGT: return ARMCC::GT;
569 case ISD::SETGE: return ARMCC::GE;
570 case ISD::SETLT: return ARMCC::LT;
571 case ISD::SETLE: return ARMCC::LE;
572 case ISD::SETUGT: return ARMCC::HI;
573 case ISD::SETUGE: return ARMCC::HS;
574 case ISD::SETULT: return ARMCC::LO;
575 case ISD::SETULE: return ARMCC::LS;
579 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
580 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
581 ARMCC::CondCodes &CondCode2) {
582 CondCode2 = ARMCC::AL;
584 default: llvm_unreachable("Unknown FP condition!");
586 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
588 case ISD::SETOGT: CondCode = ARMCC::GT; break;
590 case ISD::SETOGE: CondCode = ARMCC::GE; break;
591 case ISD::SETOLT: CondCode = ARMCC::MI; break;
592 case ISD::SETOLE: CondCode = ARMCC::LS; break;
593 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
594 case ISD::SETO: CondCode = ARMCC::VC; break;
595 case ISD::SETUO: CondCode = ARMCC::VS; break;
596 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
597 case ISD::SETUGT: CondCode = ARMCC::HI; break;
598 case ISD::SETUGE: CondCode = ARMCC::PL; break;
600 case ISD::SETULT: CondCode = ARMCC::LT; break;
602 case ISD::SETULE: CondCode = ARMCC::LE; break;
604 case ISD::SETUNE: CondCode = ARMCC::NE; break;
608 //===----------------------------------------------------------------------===//
609 // Calling Convention Implementation
610 //===----------------------------------------------------------------------===//
612 #include "ARMGenCallingConv.inc"
614 // APCS f64 is in register pairs, possibly split to stack
615 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
616 CCValAssign::LocInfo &LocInfo,
617 CCState &State, bool CanFail) {
618 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
620 // Try to get the first register.
621 if (unsigned Reg = State.AllocateReg(RegList, 4))
622 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
624 // For the 2nd half of a v2f64, do not fail.
628 // Put the whole thing on the stack.
629 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
630 State.AllocateStack(8, 4),
635 // Try to get the second register.
636 if (unsigned Reg = State.AllocateReg(RegList, 4))
637 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
639 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
640 State.AllocateStack(4, 4),
645 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
646 CCValAssign::LocInfo &LocInfo,
647 ISD::ArgFlagsTy &ArgFlags,
649 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
651 if (LocVT == MVT::v2f64 &&
652 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
654 return true; // we handled it
657 // AAPCS f64 is in aligned register pairs
658 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
659 CCValAssign::LocInfo &LocInfo,
660 CCState &State, bool CanFail) {
661 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
662 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
664 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
666 // For the 2nd half of a v2f64, do not just fail.
670 // Put the whole thing on the stack.
671 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
672 State.AllocateStack(8, 8),
678 for (i = 0; i < 2; ++i)
679 if (HiRegList[i] == Reg)
682 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
683 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
688 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
689 CCValAssign::LocInfo &LocInfo,
690 ISD::ArgFlagsTy &ArgFlags,
692 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
694 if (LocVT == MVT::v2f64 &&
695 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
697 return true; // we handled it
700 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
701 CCValAssign::LocInfo &LocInfo, CCState &State) {
702 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
703 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
705 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
707 return false; // we didn't handle it
710 for (i = 0; i < 2; ++i)
711 if (HiRegList[i] == Reg)
714 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
715 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
720 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
721 CCValAssign::LocInfo &LocInfo,
722 ISD::ArgFlagsTy &ArgFlags,
724 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
726 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
728 return true; // we handled it
731 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
732 CCValAssign::LocInfo &LocInfo,
733 ISD::ArgFlagsTy &ArgFlags,
735 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
739 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
740 /// given CallingConvention value.
741 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
743 bool isVarArg) const {
746 llvm_unreachable("Unsupported calling convention");
748 case CallingConv::Fast:
749 // Use target triple & subtarget features to do actual dispatch.
750 if (Subtarget->isAAPCS_ABI()) {
751 if (Subtarget->hasVFP2() &&
752 FloatABIType == FloatABI::Hard && !isVarArg)
753 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
755 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
757 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
758 case CallingConv::ARM_AAPCS_VFP:
759 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
760 case CallingConv::ARM_AAPCS:
761 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
762 case CallingConv::ARM_APCS:
763 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
767 /// LowerCallResult - Lower the result values of a call into the
768 /// appropriate copies out of appropriate physical registers.
770 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
771 CallingConv::ID CallConv, bool isVarArg,
772 const SmallVectorImpl<ISD::InputArg> &Ins,
773 DebugLoc dl, SelectionDAG &DAG,
774 SmallVectorImpl<SDValue> &InVals) {
776 // Assign locations to each value returned by this call.
777 SmallVector<CCValAssign, 16> RVLocs;
778 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
779 RVLocs, *DAG.getContext());
780 CCInfo.AnalyzeCallResult(Ins,
781 CCAssignFnForNode(CallConv, /* Return*/ true,
784 // Copy all of the result registers out of their specified physreg.
785 for (unsigned i = 0; i != RVLocs.size(); ++i) {
786 CCValAssign VA = RVLocs[i];
789 if (VA.needsCustom()) {
790 // Handle f64 or half of a v2f64.
791 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
793 Chain = Lo.getValue(1);
794 InFlag = Lo.getValue(2);
795 VA = RVLocs[++i]; // skip ahead to next loc
796 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
798 Chain = Hi.getValue(1);
799 InFlag = Hi.getValue(2);
800 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
802 if (VA.getLocVT() == MVT::v2f64) {
803 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
804 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
805 DAG.getConstant(0, MVT::i32));
807 VA = RVLocs[++i]; // skip ahead to next loc
808 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
809 Chain = Lo.getValue(1);
810 InFlag = Lo.getValue(2);
811 VA = RVLocs[++i]; // skip ahead to next loc
812 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
813 Chain = Hi.getValue(1);
814 InFlag = Hi.getValue(2);
815 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
816 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
817 DAG.getConstant(1, MVT::i32));
820 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
822 Chain = Val.getValue(1);
823 InFlag = Val.getValue(2);
826 switch (VA.getLocInfo()) {
827 default: llvm_unreachable("Unknown loc info!");
828 case CCValAssign::Full: break;
829 case CCValAssign::BCvt:
830 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
834 InVals.push_back(Val);
840 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
841 /// by "Src" to address "Dst" of size "Size". Alignment information is
842 /// specified by the specific parameter attribute. The copy will be passed as
843 /// a byval function parameter.
844 /// Sometimes what we are copying is the end of a larger object, the part that
845 /// does not fit in registers.
847 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
848 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
850 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
851 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
852 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
855 /// LowerMemOpCallTo - Store the argument to the stack.
857 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
858 SDValue StackPtr, SDValue Arg,
859 DebugLoc dl, SelectionDAG &DAG,
860 const CCValAssign &VA,
861 ISD::ArgFlagsTy Flags) {
862 unsigned LocMemOffset = VA.getLocMemOffset();
863 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
864 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
865 if (Flags.isByVal()) {
866 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
868 return DAG.getStore(Chain, dl, Arg, PtrOff,
869 PseudoSourceValue::getStack(), LocMemOffset,
873 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
874 SDValue Chain, SDValue &Arg,
875 RegsToPassVector &RegsToPass,
876 CCValAssign &VA, CCValAssign &NextVA,
878 SmallVector<SDValue, 8> &MemOpChains,
879 ISD::ArgFlagsTy Flags) {
881 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
882 DAG.getVTList(MVT::i32, MVT::i32), Arg);
883 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
885 if (NextVA.isRegLoc())
886 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
888 assert(NextVA.isMemLoc());
889 if (StackPtr.getNode() == 0)
890 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
892 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
898 /// LowerCall - Lowering a call into a callseq_start <-
899 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
902 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
903 CallingConv::ID CallConv, bool isVarArg,
905 const SmallVectorImpl<ISD::OutputArg> &Outs,
906 const SmallVectorImpl<ISD::InputArg> &Ins,
907 DebugLoc dl, SelectionDAG &DAG,
908 SmallVectorImpl<SDValue> &InVals) {
909 // ARM target does not yet support tail call optimization.
912 // Analyze operands of the call, assigning locations to each operand.
913 SmallVector<CCValAssign, 16> ArgLocs;
914 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
916 CCInfo.AnalyzeCallOperands(Outs,
917 CCAssignFnForNode(CallConv, /* Return*/ false,
920 // Get a count of how many bytes are to be pushed on the stack.
921 unsigned NumBytes = CCInfo.getNextStackOffset();
923 // Adjust the stack pointer for the new arguments...
924 // These operations are automatically eliminated by the prolog/epilog pass
925 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
927 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
929 RegsToPassVector RegsToPass;
930 SmallVector<SDValue, 8> MemOpChains;
932 // Walk the register/memloc assignments, inserting copies/loads. In the case
933 // of tail call optimization, arguments are handled later.
934 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
937 CCValAssign &VA = ArgLocs[i];
938 SDValue Arg = Outs[realArgIdx].Val;
939 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
941 // Promote the value if needed.
942 switch (VA.getLocInfo()) {
943 default: llvm_unreachable("Unknown loc info!");
944 case CCValAssign::Full: break;
945 case CCValAssign::SExt:
946 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
948 case CCValAssign::ZExt:
949 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
951 case CCValAssign::AExt:
952 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
954 case CCValAssign::BCvt:
955 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
959 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
960 if (VA.needsCustom()) {
961 if (VA.getLocVT() == MVT::v2f64) {
962 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
963 DAG.getConstant(0, MVT::i32));
964 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
965 DAG.getConstant(1, MVT::i32));
967 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
968 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
970 VA = ArgLocs[++i]; // skip ahead to next loc
972 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
973 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
975 assert(VA.isMemLoc());
977 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
978 dl, DAG, VA, Flags));
981 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
982 StackPtr, MemOpChains, Flags);
984 } else if (VA.isRegLoc()) {
985 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
987 assert(VA.isMemLoc());
989 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
990 dl, DAG, VA, Flags));
994 if (!MemOpChains.empty())
995 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
996 &MemOpChains[0], MemOpChains.size());
998 // Build a sequence of copy-to-reg nodes chained together with token chain
999 // and flag operands which copy the outgoing args into the appropriate regs.
1001 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1002 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1003 RegsToPass[i].second, InFlag);
1004 InFlag = Chain.getValue(1);
1007 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1008 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1009 // node so that legalize doesn't hack it.
1010 bool isDirect = false;
1011 bool isARMFunc = false;
1012 bool isLocalARMFunc = false;
1013 MachineFunction &MF = DAG.getMachineFunction();
1014 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1015 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1016 GlobalValue *GV = G->getGlobal();
1018 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1019 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1020 getTargetMachine().getRelocationModel() != Reloc::Static;
1021 isARMFunc = !Subtarget->isThumb() || isStub;
1022 // ARM call to a local ARM function is predicable.
1023 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1024 // tBX takes a register source operand.
1025 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1026 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1027 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1030 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1031 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1032 Callee = DAG.getLoad(getPointerTy(), dl,
1033 DAG.getEntryNode(), CPAddr,
1034 PseudoSourceValue::getConstantPool(), 0,
1036 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1037 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1038 getPointerTy(), Callee, PICLabel);
1040 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1041 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1043 bool isStub = Subtarget->isTargetDarwin() &&
1044 getTargetMachine().getRelocationModel() != Reloc::Static;
1045 isARMFunc = !Subtarget->isThumb() || isStub;
1046 // tBX takes a register source operand.
1047 const char *Sym = S->getSymbol();
1048 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1049 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1050 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1051 Sym, ARMPCLabelIndex, 4);
1052 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1053 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1054 Callee = DAG.getLoad(getPointerTy(), dl,
1055 DAG.getEntryNode(), CPAddr,
1056 PseudoSourceValue::getConstantPool(), 0,
1058 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1059 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1060 getPointerTy(), Callee, PICLabel);
1062 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1065 // FIXME: handle tail calls differently.
1067 if (Subtarget->isThumb()) {
1068 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1069 CallOpc = ARMISD::CALL_NOLINK;
1071 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1073 CallOpc = (isDirect || Subtarget->hasV5TOps())
1074 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1075 : ARMISD::CALL_NOLINK;
1077 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1078 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1079 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1080 InFlag = Chain.getValue(1);
1083 std::vector<SDValue> Ops;
1084 Ops.push_back(Chain);
1085 Ops.push_back(Callee);
1087 // Add argument registers to the end of the list so that they are known live
1089 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1090 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1091 RegsToPass[i].second.getValueType()));
1093 if (InFlag.getNode())
1094 Ops.push_back(InFlag);
1095 // Returns a chain and a flag for retval copy to use.
1096 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1097 &Ops[0], Ops.size());
1098 InFlag = Chain.getValue(1);
1100 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1101 DAG.getIntPtrConstant(0, true), InFlag);
1103 InFlag = Chain.getValue(1);
1105 // Handle result values, copying them out of physregs into vregs that we
1107 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1112 ARMTargetLowering::LowerReturn(SDValue Chain,
1113 CallingConv::ID CallConv, bool isVarArg,
1114 const SmallVectorImpl<ISD::OutputArg> &Outs,
1115 DebugLoc dl, SelectionDAG &DAG) {
1117 // CCValAssign - represent the assignment of the return value to a location.
1118 SmallVector<CCValAssign, 16> RVLocs;
1120 // CCState - Info about the registers and stack slots.
1121 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1124 // Analyze outgoing return values.
1125 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1128 // If this is the first return lowered for this function, add
1129 // the regs to the liveout set for the function.
1130 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1131 for (unsigned i = 0; i != RVLocs.size(); ++i)
1132 if (RVLocs[i].isRegLoc())
1133 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1138 // Copy the result values into the output registers.
1139 for (unsigned i = 0, realRVLocIdx = 0;
1141 ++i, ++realRVLocIdx) {
1142 CCValAssign &VA = RVLocs[i];
1143 assert(VA.isRegLoc() && "Can only return in registers!");
1145 SDValue Arg = Outs[realRVLocIdx].Val;
1147 switch (VA.getLocInfo()) {
1148 default: llvm_unreachable("Unknown loc info!");
1149 case CCValAssign::Full: break;
1150 case CCValAssign::BCvt:
1151 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1155 if (VA.needsCustom()) {
1156 if (VA.getLocVT() == MVT::v2f64) {
1157 // Extract the first half and return it in two registers.
1158 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1159 DAG.getConstant(0, MVT::i32));
1160 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1161 DAG.getVTList(MVT::i32, MVT::i32), Half);
1163 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1164 Flag = Chain.getValue(1);
1165 VA = RVLocs[++i]; // skip ahead to next loc
1166 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1167 HalfGPRs.getValue(1), Flag);
1168 Flag = Chain.getValue(1);
1169 VA = RVLocs[++i]; // skip ahead to next loc
1171 // Extract the 2nd half and fall through to handle it as an f64 value.
1172 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1173 DAG.getConstant(1, MVT::i32));
1175 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1177 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1178 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1179 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1180 Flag = Chain.getValue(1);
1181 VA = RVLocs[++i]; // skip ahead to next loc
1182 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1185 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1187 // Guarantee that all emitted copies are
1188 // stuck together, avoiding something bad.
1189 Flag = Chain.getValue(1);
1194 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1196 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1201 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1202 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1203 // one of the above mentioned nodes. It has to be wrapped because otherwise
1204 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1205 // be used to form addressing mode. These wrapped nodes will be selected
1207 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1208 EVT PtrVT = Op.getValueType();
1209 // FIXME there is no actual debug info here
1210 DebugLoc dl = Op.getDebugLoc();
1211 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1213 if (CP->isMachineConstantPoolEntry())
1214 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1215 CP->getAlignment());
1217 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1218 CP->getAlignment());
1219 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1222 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1223 MachineFunction &MF = DAG.getMachineFunction();
1224 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1225 unsigned ARMPCLabelIndex = 0;
1226 DebugLoc DL = Op.getDebugLoc();
1227 EVT PtrVT = getPointerTy();
1228 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1229 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1231 if (RelocM == Reloc::Static) {
1232 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1234 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1235 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1236 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1237 ARMCP::CPBlockAddress,
1239 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1241 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1242 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1243 PseudoSourceValue::getConstantPool(), 0,
1245 if (RelocM == Reloc::Static)
1247 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1248 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1251 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1253 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1254 SelectionDAG &DAG) {
1255 DebugLoc dl = GA->getDebugLoc();
1256 EVT PtrVT = getPointerTy();
1257 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1258 MachineFunction &MF = DAG.getMachineFunction();
1259 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1260 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1261 ARMConstantPoolValue *CPV =
1262 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1263 ARMCP::CPValue, PCAdj, "tlsgd", true);
1264 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1265 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1266 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1267 PseudoSourceValue::getConstantPool(), 0,
1269 SDValue Chain = Argument.getValue(1);
1271 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1272 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1274 // call __tls_get_addr.
1277 Entry.Node = Argument;
1278 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1279 Args.push_back(Entry);
1280 // FIXME: is there useful debug info available here?
1281 std::pair<SDValue, SDValue> CallResult =
1282 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1283 false, false, false, false,
1284 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1285 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1286 return CallResult.first;
1289 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1290 // "local exec" model.
1292 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1293 SelectionDAG &DAG) {
1294 GlobalValue *GV = GA->getGlobal();
1295 DebugLoc dl = GA->getDebugLoc();
1297 SDValue Chain = DAG.getEntryNode();
1298 EVT PtrVT = getPointerTy();
1299 // Get the Thread Pointer
1300 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1302 if (GV->isDeclaration()) {
1303 MachineFunction &MF = DAG.getMachineFunction();
1304 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1305 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1306 // Initial exec model.
1307 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1308 ARMConstantPoolValue *CPV =
1309 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1310 ARMCP::CPValue, PCAdj, "gottpoff", true);
1311 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1312 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1313 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1314 PseudoSourceValue::getConstantPool(), 0,
1316 Chain = Offset.getValue(1);
1318 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1319 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1321 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1322 PseudoSourceValue::getConstantPool(), 0,
1326 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1327 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1328 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1329 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1330 PseudoSourceValue::getConstantPool(), 0,
1334 // The address of the thread local variable is the add of the thread
1335 // pointer with the offset of the variable.
1336 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1340 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1341 // TODO: implement the "local dynamic" model
1342 assert(Subtarget->isTargetELF() &&
1343 "TLS not implemented for non-ELF targets");
1344 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1345 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1346 // otherwise use the "Local Exec" TLS Model
1347 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1348 return LowerToTLSGeneralDynamicModel(GA, DAG);
1350 return LowerToTLSExecModels(GA, DAG);
1353 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1354 SelectionDAG &DAG) {
1355 EVT PtrVT = getPointerTy();
1356 DebugLoc dl = Op.getDebugLoc();
1357 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1358 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1359 if (RelocM == Reloc::PIC_) {
1360 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1361 ARMConstantPoolValue *CPV =
1362 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1363 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1364 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1365 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1367 PseudoSourceValue::getConstantPool(), 0,
1369 SDValue Chain = Result.getValue(1);
1370 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1371 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1373 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1374 PseudoSourceValue::getGOT(), 0,
1378 // If we have T2 ops, we can materialize the address directly via movt/movw
1379 // pair. This is always cheaper.
1380 if (Subtarget->useMovt()) {
1381 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1382 DAG.getTargetGlobalAddress(GV, PtrVT));
1384 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1385 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1386 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1387 PseudoSourceValue::getConstantPool(), 0,
1393 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1394 SelectionDAG &DAG) {
1395 MachineFunction &MF = DAG.getMachineFunction();
1396 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1397 unsigned ARMPCLabelIndex = 0;
1398 EVT PtrVT = getPointerTy();
1399 DebugLoc dl = Op.getDebugLoc();
1400 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1401 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1403 if (RelocM == Reloc::Static)
1404 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1406 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1407 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1408 ARMConstantPoolValue *CPV =
1409 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1410 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1412 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1414 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1415 PseudoSourceValue::getConstantPool(), 0,
1417 SDValue Chain = Result.getValue(1);
1419 if (RelocM == Reloc::PIC_) {
1420 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1421 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1424 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1425 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1426 PseudoSourceValue::getGOT(), 0,
1432 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1434 assert(Subtarget->isTargetELF() &&
1435 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1436 MachineFunction &MF = DAG.getMachineFunction();
1437 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1438 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1439 EVT PtrVT = getPointerTy();
1440 DebugLoc dl = Op.getDebugLoc();
1441 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1442 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1443 "_GLOBAL_OFFSET_TABLE_",
1444 ARMPCLabelIndex, PCAdj);
1445 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1446 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1447 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1448 PseudoSourceValue::getConstantPool(), 0,
1450 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1451 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1455 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1456 const ARMSubtarget *Subtarget) {
1457 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1458 DebugLoc dl = Op.getDebugLoc();
1460 default: return SDValue(); // Don't custom lower most intrinsics.
1461 case Intrinsic::arm_thread_pointer: {
1462 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1463 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1465 case Intrinsic::eh_sjlj_lsda: {
1466 MachineFunction &MF = DAG.getMachineFunction();
1467 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1468 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1469 EVT PtrVT = getPointerTy();
1470 DebugLoc dl = Op.getDebugLoc();
1471 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1473 unsigned PCAdj = (RelocM != Reloc::PIC_)
1474 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1475 ARMConstantPoolValue *CPV =
1476 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1477 ARMCP::CPLSDA, PCAdj);
1478 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1479 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1481 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1482 PseudoSourceValue::getConstantPool(), 0,
1484 SDValue Chain = Result.getValue(1);
1486 if (RelocM == Reloc::PIC_) {
1487 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1488 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1492 case Intrinsic::eh_sjlj_setjmp:
1493 SDValue Val = Subtarget->isThumb() ?
1494 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1495 DAG.getConstant(0, MVT::i32);
1496 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1501 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1502 const ARMSubtarget *Subtarget) {
1503 DebugLoc dl = Op.getDebugLoc();
1504 SDValue Op5 = Op.getOperand(5);
1506 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1507 if (isDeviceBarrier) {
1508 if (Subtarget->hasV7Ops())
1509 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1511 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1512 DAG.getConstant(0, MVT::i32));
1514 if (Subtarget->hasV7Ops())
1515 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1517 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1518 DAG.getConstant(0, MVT::i32));
1523 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1524 unsigned VarArgsFrameIndex) {
1525 // vastart just stores the address of the VarArgsFrameIndex slot into the
1526 // memory location argument.
1527 DebugLoc dl = Op.getDebugLoc();
1528 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1529 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1530 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1531 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1536 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1537 SDNode *Node = Op.getNode();
1538 DebugLoc dl = Node->getDebugLoc();
1539 EVT VT = Node->getValueType(0);
1540 SDValue Chain = Op.getOperand(0);
1541 SDValue Size = Op.getOperand(1);
1542 SDValue Align = Op.getOperand(2);
1544 // Chain the dynamic stack allocation so that it doesn't modify the stack
1545 // pointer when other instructions are using the stack.
1546 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1548 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1549 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1550 if (AlignVal > StackAlign)
1551 // Do this now since selection pass cannot introduce new target
1552 // independent node.
1553 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1555 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1556 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1557 // do even more horrible hack later.
1558 MachineFunction &MF = DAG.getMachineFunction();
1559 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1560 if (AFI->isThumb1OnlyFunction()) {
1562 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1564 uint32_t Val = C->getZExtValue();
1565 if (Val <= 508 && ((Val & 3) == 0))
1569 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1572 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1573 SDValue Ops1[] = { Chain, Size, Align };
1574 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1575 Chain = Res.getValue(1);
1576 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1577 DAG.getIntPtrConstant(0, true), SDValue());
1578 SDValue Ops2[] = { Res, Chain };
1579 return DAG.getMergeValues(Ops2, 2, dl);
1583 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1584 SDValue &Root, SelectionDAG &DAG,
1586 MachineFunction &MF = DAG.getMachineFunction();
1587 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1589 TargetRegisterClass *RC;
1590 if (AFI->isThumb1OnlyFunction())
1591 RC = ARM::tGPRRegisterClass;
1593 RC = ARM::GPRRegisterClass;
1595 // Transform the arguments stored in physical registers into virtual ones.
1596 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1597 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1600 if (NextVA.isMemLoc()) {
1601 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1602 MachineFrameInfo *MFI = MF.getFrameInfo();
1603 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset(),
1606 // Create load node to retrieve arguments from the stack.
1607 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1608 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1609 PseudoSourceValue::getFixedStack(FI), 0,
1612 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1613 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1616 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1620 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1621 CallingConv::ID CallConv, bool isVarArg,
1622 const SmallVectorImpl<ISD::InputArg>
1624 DebugLoc dl, SelectionDAG &DAG,
1625 SmallVectorImpl<SDValue> &InVals) {
1627 MachineFunction &MF = DAG.getMachineFunction();
1628 MachineFrameInfo *MFI = MF.getFrameInfo();
1630 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1632 // Assign locations to all of the incoming arguments.
1633 SmallVector<CCValAssign, 16> ArgLocs;
1634 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1636 CCInfo.AnalyzeFormalArguments(Ins,
1637 CCAssignFnForNode(CallConv, /* Return*/ false,
1640 SmallVector<SDValue, 16> ArgValues;
1642 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1643 CCValAssign &VA = ArgLocs[i];
1645 // Arguments stored in registers.
1646 if (VA.isRegLoc()) {
1647 EVT RegVT = VA.getLocVT();
1650 if (VA.needsCustom()) {
1651 // f64 and vector types are split up into multiple registers or
1652 // combinations of registers and stack slots.
1655 if (VA.getLocVT() == MVT::v2f64) {
1656 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1658 VA = ArgLocs[++i]; // skip ahead to next loc
1659 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1661 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1662 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1663 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1664 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1665 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1667 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1670 TargetRegisterClass *RC;
1672 if (RegVT == MVT::f32)
1673 RC = ARM::SPRRegisterClass;
1674 else if (RegVT == MVT::f64)
1675 RC = ARM::DPRRegisterClass;
1676 else if (RegVT == MVT::v2f64)
1677 RC = ARM::QPRRegisterClass;
1678 else if (RegVT == MVT::i32)
1679 RC = (AFI->isThumb1OnlyFunction() ?
1680 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1682 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1684 // Transform the arguments in physical registers into virtual ones.
1685 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1686 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1689 // If this is an 8 or 16-bit value, it is really passed promoted
1690 // to 32 bits. Insert an assert[sz]ext to capture this, then
1691 // truncate to the right size.
1692 switch (VA.getLocInfo()) {
1693 default: llvm_unreachable("Unknown loc info!");
1694 case CCValAssign::Full: break;
1695 case CCValAssign::BCvt:
1696 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1698 case CCValAssign::SExt:
1699 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1700 DAG.getValueType(VA.getValVT()));
1701 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1703 case CCValAssign::ZExt:
1704 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1705 DAG.getValueType(VA.getValVT()));
1706 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1710 InVals.push_back(ArgValue);
1712 } else { // VA.isRegLoc()
1715 assert(VA.isMemLoc());
1716 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1718 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1719 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1722 // Create load nodes to retrieve arguments from the stack.
1723 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1724 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1725 PseudoSourceValue::getFixedStack(FI), 0,
1732 static const unsigned GPRArgRegs[] = {
1733 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1736 unsigned NumGPRs = CCInfo.getFirstUnallocated
1737 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1739 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1740 unsigned VARegSize = (4 - NumGPRs) * 4;
1741 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1742 unsigned ArgOffset = CCInfo.getNextStackOffset();
1743 if (VARegSaveSize) {
1744 // If this function is vararg, store any remaining integer argument regs
1745 // to their spots on the stack so that they may be loaded by deferencing
1746 // the result of va_next.
1747 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1748 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1749 VARegSaveSize - VARegSize,
1751 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1753 SmallVector<SDValue, 4> MemOps;
1754 for (; NumGPRs < 4; ++NumGPRs) {
1755 TargetRegisterClass *RC;
1756 if (AFI->isThumb1OnlyFunction())
1757 RC = ARM::tGPRRegisterClass;
1759 RC = ARM::GPRRegisterClass;
1761 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1762 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1763 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1764 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0,
1766 MemOps.push_back(Store);
1767 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1768 DAG.getConstant(4, getPointerTy()));
1770 if (!MemOps.empty())
1771 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1772 &MemOps[0], MemOps.size());
1774 // This will point to the next argument passed via stack.
1775 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
1781 /// isFloatingPointZero - Return true if this is +0.0.
1782 static bool isFloatingPointZero(SDValue Op) {
1783 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1784 return CFP->getValueAPF().isPosZero();
1785 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1786 // Maybe this has already been legalized into the constant pool?
1787 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1788 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1789 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1790 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1791 return CFP->getValueAPF().isPosZero();
1797 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1798 /// the given operands.
1800 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1801 SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
1802 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1803 unsigned C = RHSC->getZExtValue();
1804 if (!isLegalICmpImmediate(C)) {
1805 // Constant does not fit, try adjusting it by one?
1810 if (isLegalICmpImmediate(C-1)) {
1811 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1812 RHS = DAG.getConstant(C-1, MVT::i32);
1817 if (C > 0 && isLegalICmpImmediate(C-1)) {
1818 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1819 RHS = DAG.getConstant(C-1, MVT::i32);
1824 if (isLegalICmpImmediate(C+1)) {
1825 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1826 RHS = DAG.getConstant(C+1, MVT::i32);
1831 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1832 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1833 RHS = DAG.getConstant(C+1, MVT::i32);
1840 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1841 ARMISD::NodeType CompareType;
1844 CompareType = ARMISD::CMP;
1849 CompareType = ARMISD::CMPZ;
1852 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1853 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1856 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1857 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1860 if (!isFloatingPointZero(RHS))
1861 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1863 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1864 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1867 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1868 EVT VT = Op.getValueType();
1869 SDValue LHS = Op.getOperand(0);
1870 SDValue RHS = Op.getOperand(1);
1871 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1872 SDValue TrueVal = Op.getOperand(2);
1873 SDValue FalseVal = Op.getOperand(3);
1874 DebugLoc dl = Op.getDebugLoc();
1876 if (LHS.getValueType() == MVT::i32) {
1878 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1879 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1880 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1883 ARMCC::CondCodes CondCode, CondCode2;
1884 FPCCToARMCC(CC, CondCode, CondCode2);
1886 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1887 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1888 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1889 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1891 if (CondCode2 != ARMCC::AL) {
1892 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1893 // FIXME: Needs another CMP because flag can have but one use.
1894 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1895 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1896 Result, TrueVal, ARMCC2, CCR, Cmp2);
1901 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1902 SDValue Chain = Op.getOperand(0);
1903 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1904 SDValue LHS = Op.getOperand(2);
1905 SDValue RHS = Op.getOperand(3);
1906 SDValue Dest = Op.getOperand(4);
1907 DebugLoc dl = Op.getDebugLoc();
1909 if (LHS.getValueType() == MVT::i32) {
1911 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1912 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1913 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1914 Chain, Dest, ARMCC, CCR,Cmp);
1917 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1918 ARMCC::CondCodes CondCode, CondCode2;
1919 FPCCToARMCC(CC, CondCode, CondCode2);
1921 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1922 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1923 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1924 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1925 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1926 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1927 if (CondCode2 != ARMCC::AL) {
1928 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1929 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1930 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1935 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1936 SDValue Chain = Op.getOperand(0);
1937 SDValue Table = Op.getOperand(1);
1938 SDValue Index = Op.getOperand(2);
1939 DebugLoc dl = Op.getDebugLoc();
1941 EVT PTy = getPointerTy();
1942 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1943 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1944 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1945 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1946 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1947 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1948 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1949 if (Subtarget->isThumb2()) {
1950 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1951 // which does another jump to the destination. This also makes it easier
1952 // to translate it to TBB / TBH later.
1953 // FIXME: This might not work if the function is extremely large.
1954 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1955 Addr, Op.getOperand(2), JTI, UId);
1957 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1958 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1959 PseudoSourceValue::getJumpTable(), 0,
1961 Chain = Addr.getValue(1);
1962 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1963 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1965 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1966 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
1967 Chain = Addr.getValue(1);
1968 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1972 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1973 DebugLoc dl = Op.getDebugLoc();
1975 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1976 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1977 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1980 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1981 EVT VT = Op.getValueType();
1982 DebugLoc dl = Op.getDebugLoc();
1984 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1986 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1987 return DAG.getNode(Opc, dl, VT, Op);
1990 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1991 // Implement fcopysign with a fabs and a conditional fneg.
1992 SDValue Tmp0 = Op.getOperand(0);
1993 SDValue Tmp1 = Op.getOperand(1);
1994 DebugLoc dl = Op.getDebugLoc();
1995 EVT VT = Op.getValueType();
1996 EVT SrcVT = Tmp1.getValueType();
1997 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1998 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1999 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2000 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2001 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2004 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
2005 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2006 MFI->setFrameAddressIsTaken(true);
2007 EVT VT = Op.getValueType();
2008 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2009 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2010 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2011 ? ARM::R7 : ARM::R11;
2012 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2014 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2020 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2022 SDValue Dst, SDValue Src,
2023 SDValue Size, unsigned Align,
2025 const Value *DstSV, uint64_t DstSVOff,
2026 const Value *SrcSV, uint64_t SrcSVOff){
2027 // Do repeated 4-byte loads and stores. To be improved.
2028 // This requires 4-byte alignment.
2029 if ((Align & 3) != 0)
2031 // This requires the copy size to be a constant, preferrably
2032 // within a subtarget-specific limit.
2033 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2036 uint64_t SizeVal = ConstantSize->getZExtValue();
2037 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2040 unsigned BytesLeft = SizeVal & 3;
2041 unsigned NumMemOps = SizeVal >> 2;
2042 unsigned EmittedNumMemOps = 0;
2044 unsigned VTSize = 4;
2046 const unsigned MAX_LOADS_IN_LDM = 6;
2047 SDValue TFOps[MAX_LOADS_IN_LDM];
2048 SDValue Loads[MAX_LOADS_IN_LDM];
2049 uint64_t SrcOff = 0, DstOff = 0;
2051 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2052 // same number of stores. The loads and stores will get combined into
2053 // ldm/stm later on.
2054 while (EmittedNumMemOps < NumMemOps) {
2056 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2057 Loads[i] = DAG.getLoad(VT, dl, Chain,
2058 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2059 DAG.getConstant(SrcOff, MVT::i32)),
2060 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2061 TFOps[i] = Loads[i].getValue(1);
2064 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2067 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2068 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2069 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2070 DAG.getConstant(DstOff, MVT::i32)),
2071 DstSV, DstSVOff + DstOff, false, false, 0);
2074 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2076 EmittedNumMemOps += i;
2082 // Issue loads / stores for the trailing (1 - 3) bytes.
2083 unsigned BytesLeftSave = BytesLeft;
2086 if (BytesLeft >= 2) {
2094 Loads[i] = DAG.getLoad(VT, dl, Chain,
2095 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2096 DAG.getConstant(SrcOff, MVT::i32)),
2097 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2098 TFOps[i] = Loads[i].getValue(1);
2101 BytesLeft -= VTSize;
2103 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2106 BytesLeft = BytesLeftSave;
2108 if (BytesLeft >= 2) {
2116 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2117 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2118 DAG.getConstant(DstOff, MVT::i32)),
2119 DstSV, DstSVOff + DstOff, false, false, 0);
2122 BytesLeft -= VTSize;
2124 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2127 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2128 SDValue Op = N->getOperand(0);
2129 DebugLoc dl = N->getDebugLoc();
2130 if (N->getValueType(0) == MVT::f64) {
2131 // Turn i64->f64 into VMOVDRR.
2132 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2133 DAG.getConstant(0, MVT::i32));
2134 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2135 DAG.getConstant(1, MVT::i32));
2136 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2139 // Turn f64->i64 into VMOVRRD.
2140 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2141 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2143 // Merge the pieces into a single i64 value.
2144 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2147 /// getZeroVector - Returns a vector of specified type with all zero elements.
2149 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2150 assert(VT.isVector() && "Expected a vector type");
2152 // Zero vectors are used to represent vector negation and in those cases
2153 // will be implemented with the NEON VNEG instruction. However, VNEG does
2154 // not support i64 elements, so sometimes the zero vectors will need to be
2155 // explicitly constructed. For those cases, and potentially other uses in
2156 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2157 // to their dest type. This ensures they get CSE'd.
2159 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2160 SmallVector<SDValue, 8> Ops;
2163 if (VT.getSizeInBits() == 64) {
2164 Ops.assign(8, Cst); TVT = MVT::v8i8;
2166 Ops.assign(16, Cst); TVT = MVT::v16i8;
2168 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2170 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2173 /// getOnesVector - Returns a vector of specified type with all bits set.
2175 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2176 assert(VT.isVector() && "Expected a vector type");
2178 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2179 // dest type. This ensures they get CSE'd.
2181 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2182 SmallVector<SDValue, 8> Ops;
2185 if (VT.getSizeInBits() == 64) {
2186 Ops.assign(8, Cst); TVT = MVT::v8i8;
2188 Ops.assign(16, Cst); TVT = MVT::v16i8;
2190 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2192 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2195 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2196 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2197 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
2198 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2199 EVT VT = Op.getValueType();
2200 unsigned VTBits = VT.getSizeInBits();
2201 DebugLoc dl = Op.getDebugLoc();
2202 SDValue ShOpLo = Op.getOperand(0);
2203 SDValue ShOpHi = Op.getOperand(1);
2204 SDValue ShAmt = Op.getOperand(2);
2206 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2208 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2210 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2211 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2212 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2213 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2214 DAG.getConstant(VTBits, MVT::i32));
2215 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2216 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2217 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2219 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2220 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2222 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2223 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2226 SDValue Ops[2] = { Lo, Hi };
2227 return DAG.getMergeValues(Ops, 2, dl);
2230 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2231 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2232 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
2233 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2234 EVT VT = Op.getValueType();
2235 unsigned VTBits = VT.getSizeInBits();
2236 DebugLoc dl = Op.getDebugLoc();
2237 SDValue ShOpLo = Op.getOperand(0);
2238 SDValue ShOpHi = Op.getOperand(1);
2239 SDValue ShAmt = Op.getOperand(2);
2242 assert(Op.getOpcode() == ISD::SHL_PARTS);
2243 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2244 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2245 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2246 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2247 DAG.getConstant(VTBits, MVT::i32));
2248 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2249 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2251 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2252 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2253 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2255 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2256 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2259 SDValue Ops[2] = { Lo, Hi };
2260 return DAG.getMergeValues(Ops, 2, dl);
2263 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2264 const ARMSubtarget *ST) {
2265 EVT VT = N->getValueType(0);
2266 DebugLoc dl = N->getDebugLoc();
2268 if (!ST->hasV6T2Ops())
2271 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2272 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2275 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2276 const ARMSubtarget *ST) {
2277 EVT VT = N->getValueType(0);
2278 DebugLoc dl = N->getDebugLoc();
2280 // Lower vector shifts on NEON to use VSHL.
2281 if (VT.isVector()) {
2282 assert(ST->hasNEON() && "unexpected vector shift");
2284 // Left shifts translate directly to the vshiftu intrinsic.
2285 if (N->getOpcode() == ISD::SHL)
2286 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2287 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2288 N->getOperand(0), N->getOperand(1));
2290 assert((N->getOpcode() == ISD::SRA ||
2291 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2293 // NEON uses the same intrinsics for both left and right shifts. For
2294 // right shifts, the shift amounts are negative, so negate the vector of
2296 EVT ShiftVT = N->getOperand(1).getValueType();
2297 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2298 getZeroVector(ShiftVT, DAG, dl),
2300 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2301 Intrinsic::arm_neon_vshifts :
2302 Intrinsic::arm_neon_vshiftu);
2303 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2304 DAG.getConstant(vshiftInt, MVT::i32),
2305 N->getOperand(0), NegatedCount);
2308 // We can get here for a node like i32 = ISD::SHL i32, i64
2312 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2313 "Unknown shift to lower!");
2315 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2316 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2317 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2320 // If we are in thumb mode, we don't have RRX.
2321 if (ST->isThumb1Only()) return SDValue();
2323 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2324 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2325 DAG.getConstant(0, MVT::i32));
2326 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2327 DAG.getConstant(1, MVT::i32));
2329 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2330 // captures the result into a carry flag.
2331 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2332 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2334 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2335 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2337 // Merge the pieces into a single i64 value.
2338 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2341 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2342 SDValue TmpOp0, TmpOp1;
2343 bool Invert = false;
2347 SDValue Op0 = Op.getOperand(0);
2348 SDValue Op1 = Op.getOperand(1);
2349 SDValue CC = Op.getOperand(2);
2350 EVT VT = Op.getValueType();
2351 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2352 DebugLoc dl = Op.getDebugLoc();
2354 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2355 switch (SetCCOpcode) {
2356 default: llvm_unreachable("Illegal FP comparison"); break;
2358 case ISD::SETNE: Invert = true; // Fallthrough
2360 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2362 case ISD::SETLT: Swap = true; // Fallthrough
2364 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2366 case ISD::SETLE: Swap = true; // Fallthrough
2368 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2369 case ISD::SETUGE: Swap = true; // Fallthrough
2370 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2371 case ISD::SETUGT: Swap = true; // Fallthrough
2372 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2373 case ISD::SETUEQ: Invert = true; // Fallthrough
2375 // Expand this to (OLT | OGT).
2379 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2380 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2382 case ISD::SETUO: Invert = true; // Fallthrough
2384 // Expand this to (OLT | OGE).
2388 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2389 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2393 // Integer comparisons.
2394 switch (SetCCOpcode) {
2395 default: llvm_unreachable("Illegal integer comparison"); break;
2396 case ISD::SETNE: Invert = true;
2397 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2398 case ISD::SETLT: Swap = true;
2399 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2400 case ISD::SETLE: Swap = true;
2401 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2402 case ISD::SETULT: Swap = true;
2403 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2404 case ISD::SETULE: Swap = true;
2405 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2408 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2409 if (Opc == ARMISD::VCEQ) {
2412 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2414 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2417 // Ignore bitconvert.
2418 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2419 AndOp = AndOp.getOperand(0);
2421 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2423 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2424 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2431 std::swap(Op0, Op1);
2433 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2436 Result = DAG.getNOT(dl, Result, VT);
2441 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2442 /// VMOV instruction, and if so, return the constant being splatted.
2443 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2444 unsigned SplatBitSize, SelectionDAG &DAG) {
2445 switch (SplatBitSize) {
2447 // Any 1-byte value is OK.
2448 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2449 return DAG.getTargetConstant(SplatBits, MVT::i8);
2452 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2453 if ((SplatBits & ~0xff) == 0 ||
2454 (SplatBits & ~0xff00) == 0)
2455 return DAG.getTargetConstant(SplatBits, MVT::i16);
2459 // NEON's 32-bit VMOV supports splat values where:
2460 // * only one byte is nonzero, or
2461 // * the least significant byte is 0xff and the second byte is nonzero, or
2462 // * the least significant 2 bytes are 0xff and the third is nonzero.
2463 if ((SplatBits & ~0xff) == 0 ||
2464 (SplatBits & ~0xff00) == 0 ||
2465 (SplatBits & ~0xff0000) == 0 ||
2466 (SplatBits & ~0xff000000) == 0)
2467 return DAG.getTargetConstant(SplatBits, MVT::i32);
2469 if ((SplatBits & ~0xffff) == 0 &&
2470 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2471 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2473 if ((SplatBits & ~0xffffff) == 0 &&
2474 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2475 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2477 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2478 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2479 // VMOV.I32. A (very) minor optimization would be to replicate the value
2480 // and fall through here to test for a valid 64-bit splat. But, then the
2481 // caller would also need to check and handle the change in size.
2485 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2486 uint64_t BitMask = 0xff;
2488 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2489 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2491 else if ((SplatBits & BitMask) != 0)
2495 return DAG.getTargetConstant(Val, MVT::i64);
2499 llvm_unreachable("unexpected size for isVMOVSplat");
2506 /// getVMOVImm - If this is a build_vector of constants which can be
2507 /// formed by using a VMOV instruction of the specified element size,
2508 /// return the constant being splatted. The ByteSize field indicates the
2509 /// number of bytes of each element [1248].
2510 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2511 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2512 APInt SplatBits, SplatUndef;
2513 unsigned SplatBitSize;
2515 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2516 HasAnyUndefs, ByteSize * 8))
2519 if (SplatBitSize > ByteSize * 8)
2522 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2526 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2527 bool &ReverseVEXT, unsigned &Imm) {
2528 unsigned NumElts = VT.getVectorNumElements();
2529 ReverseVEXT = false;
2532 // If this is a VEXT shuffle, the immediate value is the index of the first
2533 // element. The other shuffle indices must be the successive elements after
2535 unsigned ExpectedElt = Imm;
2536 for (unsigned i = 1; i < NumElts; ++i) {
2537 // Increment the expected index. If it wraps around, it may still be
2538 // a VEXT but the source vectors must be swapped.
2540 if (ExpectedElt == NumElts * 2) {
2545 if (ExpectedElt != static_cast<unsigned>(M[i]))
2549 // Adjust the index value if the source operands will be swapped.
2556 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2557 /// instruction with the specified blocksize. (The order of the elements
2558 /// within each block of the vector is reversed.)
2559 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2560 unsigned BlockSize) {
2561 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2562 "Only possible block sizes for VREV are: 16, 32, 64");
2564 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2568 unsigned NumElts = VT.getVectorNumElements();
2569 unsigned BlockElts = M[0] + 1;
2571 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2574 for (unsigned i = 0; i < NumElts; ++i) {
2575 if ((unsigned) M[i] !=
2576 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2583 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2584 unsigned &WhichResult) {
2585 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2589 unsigned NumElts = VT.getVectorNumElements();
2590 WhichResult = (M[0] == 0 ? 0 : 1);
2591 for (unsigned i = 0; i < NumElts; i += 2) {
2592 if ((unsigned) M[i] != i + WhichResult ||
2593 (unsigned) M[i+1] != i + NumElts + WhichResult)
2599 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2600 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2601 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2602 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2603 unsigned &WhichResult) {
2604 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2608 unsigned NumElts = VT.getVectorNumElements();
2609 WhichResult = (M[0] == 0 ? 0 : 1);
2610 for (unsigned i = 0; i < NumElts; i += 2) {
2611 if ((unsigned) M[i] != i + WhichResult ||
2612 (unsigned) M[i+1] != i + WhichResult)
2618 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2619 unsigned &WhichResult) {
2620 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2624 unsigned NumElts = VT.getVectorNumElements();
2625 WhichResult = (M[0] == 0 ? 0 : 1);
2626 for (unsigned i = 0; i != NumElts; ++i) {
2627 if ((unsigned) M[i] != 2 * i + WhichResult)
2631 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2632 if (VT.is64BitVector() && EltSz == 32)
2638 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2639 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2640 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2641 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2642 unsigned &WhichResult) {
2643 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2647 unsigned Half = VT.getVectorNumElements() / 2;
2648 WhichResult = (M[0] == 0 ? 0 : 1);
2649 for (unsigned j = 0; j != 2; ++j) {
2650 unsigned Idx = WhichResult;
2651 for (unsigned i = 0; i != Half; ++i) {
2652 if ((unsigned) M[i + j * Half] != Idx)
2658 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2659 if (VT.is64BitVector() && EltSz == 32)
2665 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2666 unsigned &WhichResult) {
2667 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2671 unsigned NumElts = VT.getVectorNumElements();
2672 WhichResult = (M[0] == 0 ? 0 : 1);
2673 unsigned Idx = WhichResult * NumElts / 2;
2674 for (unsigned i = 0; i != NumElts; i += 2) {
2675 if ((unsigned) M[i] != Idx ||
2676 (unsigned) M[i+1] != Idx + NumElts)
2681 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2682 if (VT.is64BitVector() && EltSz == 32)
2688 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2689 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2690 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2691 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2692 unsigned &WhichResult) {
2693 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2697 unsigned NumElts = VT.getVectorNumElements();
2698 WhichResult = (M[0] == 0 ? 0 : 1);
2699 unsigned Idx = WhichResult * NumElts / 2;
2700 for (unsigned i = 0; i != NumElts; i += 2) {
2701 if ((unsigned) M[i] != Idx ||
2702 (unsigned) M[i+1] != Idx)
2707 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2708 if (VT.is64BitVector() && EltSz == 32)
2715 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2716 // Canonicalize all-zeros and all-ones vectors.
2717 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2718 if (ConstVal->isNullValue())
2719 return getZeroVector(VT, DAG, dl);
2720 if (ConstVal->isAllOnesValue())
2721 return getOnesVector(VT, DAG, dl);
2724 if (VT.is64BitVector()) {
2725 switch (Val.getValueType().getSizeInBits()) {
2726 case 8: CanonicalVT = MVT::v8i8; break;
2727 case 16: CanonicalVT = MVT::v4i16; break;
2728 case 32: CanonicalVT = MVT::v2i32; break;
2729 case 64: CanonicalVT = MVT::v1i64; break;
2730 default: llvm_unreachable("unexpected splat element type"); break;
2733 assert(VT.is128BitVector() && "unknown splat vector size");
2734 switch (Val.getValueType().getSizeInBits()) {
2735 case 8: CanonicalVT = MVT::v16i8; break;
2736 case 16: CanonicalVT = MVT::v8i16; break;
2737 case 32: CanonicalVT = MVT::v4i32; break;
2738 case 64: CanonicalVT = MVT::v2i64; break;
2739 default: llvm_unreachable("unexpected splat element type"); break;
2743 // Build a canonical splat for this value.
2744 SmallVector<SDValue, 8> Ops;
2745 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2746 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2748 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2751 // If this is a case we can't handle, return null and let the default
2752 // expansion code take care of it.
2753 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2754 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2755 DebugLoc dl = Op.getDebugLoc();
2756 EVT VT = Op.getValueType();
2758 APInt SplatBits, SplatUndef;
2759 unsigned SplatBitSize;
2761 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2762 if (SplatBitSize <= 64) {
2763 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2764 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2766 return BuildSplat(Val, VT, DAG, dl);
2770 // If there are only 2 elements in a 128-bit vector, insert them into an
2771 // undef vector. This handles the common case for 128-bit vector argument
2772 // passing, where the insertions should be translated to subreg accesses
2773 // with no real instructions.
2774 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2775 SDValue Val = DAG.getUNDEF(VT);
2776 SDValue Op0 = Op.getOperand(0);
2777 SDValue Op1 = Op.getOperand(1);
2778 if (Op0.getOpcode() != ISD::UNDEF)
2779 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2780 DAG.getIntPtrConstant(0));
2781 if (Op1.getOpcode() != ISD::UNDEF)
2782 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2783 DAG.getIntPtrConstant(1));
2790 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2791 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2792 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2793 /// are assumed to be legal.
2795 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2797 if (VT.getVectorNumElements() == 4 &&
2798 (VT.is128BitVector() || VT.is64BitVector())) {
2799 unsigned PFIndexes[4];
2800 for (unsigned i = 0; i != 4; ++i) {
2804 PFIndexes[i] = M[i];
2807 // Compute the index in the perfect shuffle table.
2808 unsigned PFTableIndex =
2809 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2810 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2811 unsigned Cost = (PFEntry >> 30);
2818 unsigned Imm, WhichResult;
2820 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2821 isVREVMask(M, VT, 64) ||
2822 isVREVMask(M, VT, 32) ||
2823 isVREVMask(M, VT, 16) ||
2824 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2825 isVTRNMask(M, VT, WhichResult) ||
2826 isVUZPMask(M, VT, WhichResult) ||
2827 isVZIPMask(M, VT, WhichResult) ||
2828 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2829 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2830 isVZIP_v_undef_Mask(M, VT, WhichResult));
2833 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2834 /// the specified operations to build the shuffle.
2835 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2836 SDValue RHS, SelectionDAG &DAG,
2838 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2839 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2840 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2843 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2852 OP_VUZPL, // VUZP, left result
2853 OP_VUZPR, // VUZP, right result
2854 OP_VZIPL, // VZIP, left result
2855 OP_VZIPR, // VZIP, right result
2856 OP_VTRNL, // VTRN, left result
2857 OP_VTRNR // VTRN, right result
2860 if (OpNum == OP_COPY) {
2861 if (LHSID == (1*9+2)*9+3) return LHS;
2862 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2866 SDValue OpLHS, OpRHS;
2867 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2868 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2869 EVT VT = OpLHS.getValueType();
2872 default: llvm_unreachable("Unknown shuffle opcode!");
2874 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2879 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2880 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2884 return DAG.getNode(ARMISD::VEXT, dl, VT,
2886 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2889 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2890 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2893 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2894 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2897 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2898 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2902 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2903 SDValue V1 = Op.getOperand(0);
2904 SDValue V2 = Op.getOperand(1);
2905 DebugLoc dl = Op.getDebugLoc();
2906 EVT VT = Op.getValueType();
2907 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2908 SmallVector<int, 8> ShuffleMask;
2910 // Convert shuffles that are directly supported on NEON to target-specific
2911 // DAG nodes, instead of keeping them as shuffles and matching them again
2912 // during code selection. This is more efficient and avoids the possibility
2913 // of inconsistencies between legalization and selection.
2914 // FIXME: floating-point vectors should be canonicalized to integer vectors
2915 // of the same time so that they get CSEd properly.
2916 SVN->getMask(ShuffleMask);
2918 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2919 int Lane = SVN->getSplatIndex();
2920 // If this is undef splat, generate it via "just" vdup, if possible.
2921 if (Lane == -1) Lane = 0;
2923 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2924 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2926 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2927 DAG.getConstant(Lane, MVT::i32));
2932 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2935 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2936 DAG.getConstant(Imm, MVT::i32));
2939 if (isVREVMask(ShuffleMask, VT, 64))
2940 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2941 if (isVREVMask(ShuffleMask, VT, 32))
2942 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2943 if (isVREVMask(ShuffleMask, VT, 16))
2944 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2946 // Check for Neon shuffles that modify both input vectors in place.
2947 // If both results are used, i.e., if there are two shuffles with the same
2948 // source operands and with masks corresponding to both results of one of
2949 // these operations, DAG memoization will ensure that a single node is
2950 // used for both shuffles.
2951 unsigned WhichResult;
2952 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2953 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2954 V1, V2).getValue(WhichResult);
2955 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2956 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2957 V1, V2).getValue(WhichResult);
2958 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2959 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2960 V1, V2).getValue(WhichResult);
2962 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
2963 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2964 V1, V1).getValue(WhichResult);
2965 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2966 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2967 V1, V1).getValue(WhichResult);
2968 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2969 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2970 V1, V1).getValue(WhichResult);
2972 // If the shuffle is not directly supported and it has 4 elements, use
2973 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2974 if (VT.getVectorNumElements() == 4 &&
2975 (VT.is128BitVector() || VT.is64BitVector())) {
2976 unsigned PFIndexes[4];
2977 for (unsigned i = 0; i != 4; ++i) {
2978 if (ShuffleMask[i] < 0)
2981 PFIndexes[i] = ShuffleMask[i];
2984 // Compute the index in the perfect shuffle table.
2985 unsigned PFTableIndex =
2986 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2988 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2989 unsigned Cost = (PFEntry >> 30);
2992 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2998 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2999 EVT VT = Op.getValueType();
3000 DebugLoc dl = Op.getDebugLoc();
3001 SDValue Vec = Op.getOperand(0);
3002 SDValue Lane = Op.getOperand(1);
3003 assert(VT == MVT::i32 &&
3004 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3005 "unexpected type for custom-lowering vector extract");
3006 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3009 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3010 // The only time a CONCAT_VECTORS operation can have legal types is when
3011 // two 64-bit vectors are concatenated to a 128-bit vector.
3012 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3013 "unexpected CONCAT_VECTORS");
3014 DebugLoc dl = Op.getDebugLoc();
3015 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3016 SDValue Op0 = Op.getOperand(0);
3017 SDValue Op1 = Op.getOperand(1);
3018 if (Op0.getOpcode() != ISD::UNDEF)
3019 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3020 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3021 DAG.getIntPtrConstant(0));
3022 if (Op1.getOpcode() != ISD::UNDEF)
3023 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3024 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3025 DAG.getIntPtrConstant(1));
3026 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3029 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3030 switch (Op.getOpcode()) {
3031 default: llvm_unreachable("Don't know how to custom lower this!");
3032 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3033 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3034 case ISD::GlobalAddress:
3035 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3036 LowerGlobalAddressELF(Op, DAG);
3037 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3038 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3039 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3040 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3041 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3042 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
3043 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3044 case ISD::SINT_TO_FP:
3045 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3046 case ISD::FP_TO_SINT:
3047 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3048 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3049 case ISD::RETURNADDR: break;
3050 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3051 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3052 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3054 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3057 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3058 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3059 case ISD::SRL_PARTS:
3060 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3061 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3062 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3063 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3064 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3065 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3066 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3071 /// ReplaceNodeResults - Replace the results of node with an illegal result
3072 /// type with new values built out of custom code.
3073 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3074 SmallVectorImpl<SDValue>&Results,
3075 SelectionDAG &DAG) {
3076 switch (N->getOpcode()) {
3078 llvm_unreachable("Don't know how to custom expand this!");
3080 case ISD::BIT_CONVERT:
3081 Results.push_back(ExpandBIT_CONVERT(N, DAG));
3085 SDValue Res = LowerShift(N, DAG, Subtarget);
3087 Results.push_back(Res);
3093 //===----------------------------------------------------------------------===//
3094 // ARM Scheduler Hooks
3095 //===----------------------------------------------------------------------===//
3098 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3099 MachineBasicBlock *BB,
3100 unsigned Size) const {
3101 unsigned dest = MI->getOperand(0).getReg();
3102 unsigned ptr = MI->getOperand(1).getReg();
3103 unsigned oldval = MI->getOperand(2).getReg();
3104 unsigned newval = MI->getOperand(3).getReg();
3105 unsigned scratch = BB->getParent()->getRegInfo()
3106 .createVirtualRegister(ARM::GPRRegisterClass);
3107 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3108 DebugLoc dl = MI->getDebugLoc();
3109 bool isThumb2 = Subtarget->isThumb2();
3111 unsigned ldrOpc, strOpc;
3113 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3115 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3116 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3119 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3120 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3123 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3124 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3128 MachineFunction *MF = BB->getParent();
3129 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3130 MachineFunction::iterator It = BB;
3131 ++It; // insert the new blocks after the current block
3133 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3134 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3135 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3136 MF->insert(It, loop1MBB);
3137 MF->insert(It, loop2MBB);
3138 MF->insert(It, exitMBB);
3139 exitMBB->transferSuccessors(BB);
3143 // fallthrough --> loop1MBB
3144 BB->addSuccessor(loop1MBB);
3147 // ldrex dest, [ptr]
3151 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3152 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3153 .addReg(dest).addReg(oldval));
3154 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3155 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3156 BB->addSuccessor(loop2MBB);
3157 BB->addSuccessor(exitMBB);
3160 // strex scratch, newval, [ptr]
3164 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3166 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3167 .addReg(scratch).addImm(0));
3168 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3169 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3170 BB->addSuccessor(loop1MBB);
3171 BB->addSuccessor(exitMBB);
3177 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3183 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3184 unsigned Size, unsigned BinOpcode) const {
3185 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3186 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3188 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3189 MachineFunction *MF = BB->getParent();
3190 MachineFunction::iterator It = BB;
3193 unsigned dest = MI->getOperand(0).getReg();
3194 unsigned ptr = MI->getOperand(1).getReg();
3195 unsigned incr = MI->getOperand(2).getReg();
3196 DebugLoc dl = MI->getDebugLoc();
3198 bool isThumb2 = Subtarget->isThumb2();
3199 unsigned ldrOpc, strOpc;
3201 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3203 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3204 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3207 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3208 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3211 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3212 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3216 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3217 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3218 MF->insert(It, loopMBB);
3219 MF->insert(It, exitMBB);
3220 exitMBB->transferSuccessors(BB);
3222 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3223 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3224 unsigned scratch2 = (!BinOpcode) ? incr :
3225 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3229 // fallthrough --> loopMBB
3230 BB->addSuccessor(loopMBB);
3234 // <binop> scratch2, dest, incr
3235 // strex scratch, scratch2, ptr
3238 // fallthrough --> exitMBB
3240 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3242 // operand order needs to go the other way for NAND
3243 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3244 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3245 addReg(incr).addReg(dest)).addReg(0);
3247 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3248 addReg(dest).addReg(incr)).addReg(0);
3251 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3253 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3254 .addReg(scratch).addImm(0));
3255 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3256 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3258 BB->addSuccessor(loopMBB);
3259 BB->addSuccessor(exitMBB);
3265 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3271 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3272 MachineBasicBlock *BB,
3273 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
3274 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3275 DebugLoc dl = MI->getDebugLoc();
3276 bool isThumb2 = Subtarget->isThumb2();
3277 switch (MI->getOpcode()) {
3280 llvm_unreachable("Unexpected instr type to insert");
3282 case ARM::ATOMIC_LOAD_ADD_I8:
3283 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3284 case ARM::ATOMIC_LOAD_ADD_I16:
3285 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3286 case ARM::ATOMIC_LOAD_ADD_I32:
3287 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3289 case ARM::ATOMIC_LOAD_AND_I8:
3290 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3291 case ARM::ATOMIC_LOAD_AND_I16:
3292 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3293 case ARM::ATOMIC_LOAD_AND_I32:
3294 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3296 case ARM::ATOMIC_LOAD_OR_I8:
3297 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3298 case ARM::ATOMIC_LOAD_OR_I16:
3299 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3300 case ARM::ATOMIC_LOAD_OR_I32:
3301 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3303 case ARM::ATOMIC_LOAD_XOR_I8:
3304 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3305 case ARM::ATOMIC_LOAD_XOR_I16:
3306 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3307 case ARM::ATOMIC_LOAD_XOR_I32:
3308 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3310 case ARM::ATOMIC_LOAD_NAND_I8:
3311 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3312 case ARM::ATOMIC_LOAD_NAND_I16:
3313 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3314 case ARM::ATOMIC_LOAD_NAND_I32:
3315 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3317 case ARM::ATOMIC_LOAD_SUB_I8:
3318 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3319 case ARM::ATOMIC_LOAD_SUB_I16:
3320 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3321 case ARM::ATOMIC_LOAD_SUB_I32:
3322 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3324 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3325 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3326 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3328 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3329 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3330 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3332 case ARM::tMOVCCr_pseudo: {
3333 // To "insert" a SELECT_CC instruction, we actually have to insert the
3334 // diamond control-flow pattern. The incoming instruction knows the
3335 // destination vreg to set, the condition code register to branch on, the
3336 // true/false values to select between, and a branch opcode to use.
3337 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3338 MachineFunction::iterator It = BB;
3344 // cmpTY ccX, r1, r2
3346 // fallthrough --> copy0MBB
3347 MachineBasicBlock *thisMBB = BB;
3348 MachineFunction *F = BB->getParent();
3349 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3350 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3351 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3352 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3353 F->insert(It, copy0MBB);
3354 F->insert(It, sinkMBB);
3355 // Update machine-CFG edges by first adding all successors of the current
3356 // block to the new block which will contain the Phi node for the select.
3357 // Also inform sdisel of the edge changes.
3358 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3359 E = BB->succ_end(); I != E; ++I) {
3360 EM->insert(std::make_pair(*I, sinkMBB));
3361 sinkMBB->addSuccessor(*I);
3363 // Next, remove all successors of the current block, and add the true
3364 // and fallthrough blocks as its successors.
3365 while (!BB->succ_empty())
3366 BB->removeSuccessor(BB->succ_begin());
3367 BB->addSuccessor(copy0MBB);
3368 BB->addSuccessor(sinkMBB);
3371 // %FalseValue = ...
3372 // # fallthrough to sinkMBB
3375 // Update machine-CFG edges
3376 BB->addSuccessor(sinkMBB);
3379 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3382 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3383 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3384 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3386 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3393 case ARM::t2SUBrSPi_:
3394 case ARM::t2SUBrSPi12_:
3395 case ARM::t2SUBrSPs_: {
3396 MachineFunction *MF = BB->getParent();
3397 unsigned DstReg = MI->getOperand(0).getReg();
3398 unsigned SrcReg = MI->getOperand(1).getReg();
3399 bool DstIsDead = MI->getOperand(0).isDead();
3400 bool SrcIsKill = MI->getOperand(1).isKill();
3402 if (SrcReg != ARM::SP) {
3403 // Copy the source to SP from virtual register.
3404 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3405 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3406 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3407 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3408 .addReg(SrcReg, getKillRegState(SrcIsKill));
3412 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3413 switch (MI->getOpcode()) {
3415 llvm_unreachable("Unexpected pseudo instruction!");
3421 OpOpc = ARM::tADDspr;
3424 OpOpc = ARM::tSUBspi;
3426 case ARM::t2SUBrSPi_:
3427 OpOpc = ARM::t2SUBrSPi;
3428 NeedPred = true; NeedCC = true;
3430 case ARM::t2SUBrSPi12_:
3431 OpOpc = ARM::t2SUBrSPi12;
3434 case ARM::t2SUBrSPs_:
3435 OpOpc = ARM::t2SUBrSPs;
3436 NeedPred = true; NeedCC = true; NeedOp3 = true;
3439 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3440 if (OpOpc == ARM::tAND)
3441 AddDefaultT1CC(MIB);
3442 MIB.addReg(ARM::SP);
3443 MIB.addOperand(MI->getOperand(2));
3445 MIB.addOperand(MI->getOperand(3));
3447 AddDefaultPred(MIB);
3451 // Copy the result from SP to virtual register.
3452 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3453 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3454 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3455 BuildMI(BB, dl, TII->get(CopyOpc))
3456 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3458 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3464 //===----------------------------------------------------------------------===//
3465 // ARM Optimization Hooks
3466 //===----------------------------------------------------------------------===//
3469 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3470 TargetLowering::DAGCombinerInfo &DCI) {
3471 SelectionDAG &DAG = DCI.DAG;
3472 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3473 EVT VT = N->getValueType(0);
3474 unsigned Opc = N->getOpcode();
3475 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3476 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3477 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3478 ISD::CondCode CC = ISD::SETCC_INVALID;
3481 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3483 SDValue CCOp = Slct.getOperand(0);
3484 if (CCOp.getOpcode() == ISD::SETCC)
3485 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3488 bool DoXform = false;
3490 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3493 if (LHS.getOpcode() == ISD::Constant &&
3494 cast<ConstantSDNode>(LHS)->isNullValue()) {
3496 } else if (CC != ISD::SETCC_INVALID &&
3497 RHS.getOpcode() == ISD::Constant &&
3498 cast<ConstantSDNode>(RHS)->isNullValue()) {
3499 std::swap(LHS, RHS);
3500 SDValue Op0 = Slct.getOperand(0);
3501 EVT OpVT = isSlctCC ? Op0.getValueType() :
3502 Op0.getOperand(0).getValueType();
3503 bool isInt = OpVT.isInteger();
3504 CC = ISD::getSetCCInverse(CC, isInt);
3506 if (!TLI.isCondCodeLegal(CC, OpVT))
3507 return SDValue(); // Inverse operator isn't legal.
3514 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3516 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3517 Slct.getOperand(0), Slct.getOperand(1), CC);
3518 SDValue CCOp = Slct.getOperand(0);
3520 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3521 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3522 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3523 CCOp, OtherOp, Result);
3528 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3529 static SDValue PerformADDCombine(SDNode *N,
3530 TargetLowering::DAGCombinerInfo &DCI) {
3531 // added by evan in r37685 with no testcase.
3532 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3534 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3535 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3536 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3537 if (Result.getNode()) return Result;
3539 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3540 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3541 if (Result.getNode()) return Result;
3547 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3548 static SDValue PerformSUBCombine(SDNode *N,
3549 TargetLowering::DAGCombinerInfo &DCI) {
3550 // added by evan in r37685 with no testcase.
3551 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3553 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3554 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3555 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3556 if (Result.getNode()) return Result;
3562 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3563 /// ARMISD::VMOVRRD.
3564 static SDValue PerformVMOVRRDCombine(SDNode *N,
3565 TargetLowering::DAGCombinerInfo &DCI) {
3566 // fmrrd(fmdrr x, y) -> x,y
3567 SDValue InDouble = N->getOperand(0);
3568 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3569 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3573 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3574 /// operand of a vector shift operation, where all the elements of the
3575 /// build_vector must have the same constant integer value.
3576 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3577 // Ignore bit_converts.
3578 while (Op.getOpcode() == ISD::BIT_CONVERT)
3579 Op = Op.getOperand(0);
3580 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3581 APInt SplatBits, SplatUndef;
3582 unsigned SplatBitSize;
3584 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3585 HasAnyUndefs, ElementBits) ||
3586 SplatBitSize > ElementBits)
3588 Cnt = SplatBits.getSExtValue();
3592 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3593 /// operand of a vector shift left operation. That value must be in the range:
3594 /// 0 <= Value < ElementBits for a left shift; or
3595 /// 0 <= Value <= ElementBits for a long left shift.
3596 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3597 assert(VT.isVector() && "vector shift count is not a vector type");
3598 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3599 if (! getVShiftImm(Op, ElementBits, Cnt))
3601 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3604 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3605 /// operand of a vector shift right operation. For a shift opcode, the value
3606 /// is positive, but for an intrinsic the value count must be negative. The
3607 /// absolute value must be in the range:
3608 /// 1 <= |Value| <= ElementBits for a right shift; or
3609 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3610 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3612 assert(VT.isVector() && "vector shift count is not a vector type");
3613 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3614 if (! getVShiftImm(Op, ElementBits, Cnt))
3618 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3621 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3622 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3623 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3626 // Don't do anything for most intrinsics.
3629 // Vector shifts: check for immediate versions and lower them.
3630 // Note: This is done during DAG combining instead of DAG legalizing because
3631 // the build_vectors for 64-bit vector element shift counts are generally
3632 // not legal, and it is hard to see their values after they get legalized to
3633 // loads from a constant pool.
3634 case Intrinsic::arm_neon_vshifts:
3635 case Intrinsic::arm_neon_vshiftu:
3636 case Intrinsic::arm_neon_vshiftls:
3637 case Intrinsic::arm_neon_vshiftlu:
3638 case Intrinsic::arm_neon_vshiftn:
3639 case Intrinsic::arm_neon_vrshifts:
3640 case Intrinsic::arm_neon_vrshiftu:
3641 case Intrinsic::arm_neon_vrshiftn:
3642 case Intrinsic::arm_neon_vqshifts:
3643 case Intrinsic::arm_neon_vqshiftu:
3644 case Intrinsic::arm_neon_vqshiftsu:
3645 case Intrinsic::arm_neon_vqshiftns:
3646 case Intrinsic::arm_neon_vqshiftnu:
3647 case Intrinsic::arm_neon_vqshiftnsu:
3648 case Intrinsic::arm_neon_vqrshiftns:
3649 case Intrinsic::arm_neon_vqrshiftnu:
3650 case Intrinsic::arm_neon_vqrshiftnsu: {
3651 EVT VT = N->getOperand(1).getValueType();
3653 unsigned VShiftOpc = 0;
3656 case Intrinsic::arm_neon_vshifts:
3657 case Intrinsic::arm_neon_vshiftu:
3658 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3659 VShiftOpc = ARMISD::VSHL;
3662 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3663 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3664 ARMISD::VSHRs : ARMISD::VSHRu);
3669 case Intrinsic::arm_neon_vshiftls:
3670 case Intrinsic::arm_neon_vshiftlu:
3671 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3673 llvm_unreachable("invalid shift count for vshll intrinsic");
3675 case Intrinsic::arm_neon_vrshifts:
3676 case Intrinsic::arm_neon_vrshiftu:
3677 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3681 case Intrinsic::arm_neon_vqshifts:
3682 case Intrinsic::arm_neon_vqshiftu:
3683 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3687 case Intrinsic::arm_neon_vqshiftsu:
3688 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3690 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3692 case Intrinsic::arm_neon_vshiftn:
3693 case Intrinsic::arm_neon_vrshiftn:
3694 case Intrinsic::arm_neon_vqshiftns:
3695 case Intrinsic::arm_neon_vqshiftnu:
3696 case Intrinsic::arm_neon_vqshiftnsu:
3697 case Intrinsic::arm_neon_vqrshiftns:
3698 case Intrinsic::arm_neon_vqrshiftnu:
3699 case Intrinsic::arm_neon_vqrshiftnsu:
3700 // Narrowing shifts require an immediate right shift.
3701 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3703 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3706 llvm_unreachable("unhandled vector shift");
3710 case Intrinsic::arm_neon_vshifts:
3711 case Intrinsic::arm_neon_vshiftu:
3712 // Opcode already set above.
3714 case Intrinsic::arm_neon_vshiftls:
3715 case Intrinsic::arm_neon_vshiftlu:
3716 if (Cnt == VT.getVectorElementType().getSizeInBits())
3717 VShiftOpc = ARMISD::VSHLLi;
3719 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3720 ARMISD::VSHLLs : ARMISD::VSHLLu);
3722 case Intrinsic::arm_neon_vshiftn:
3723 VShiftOpc = ARMISD::VSHRN; break;
3724 case Intrinsic::arm_neon_vrshifts:
3725 VShiftOpc = ARMISD::VRSHRs; break;
3726 case Intrinsic::arm_neon_vrshiftu:
3727 VShiftOpc = ARMISD::VRSHRu; break;
3728 case Intrinsic::arm_neon_vrshiftn:
3729 VShiftOpc = ARMISD::VRSHRN; break;
3730 case Intrinsic::arm_neon_vqshifts:
3731 VShiftOpc = ARMISD::VQSHLs; break;
3732 case Intrinsic::arm_neon_vqshiftu:
3733 VShiftOpc = ARMISD::VQSHLu; break;
3734 case Intrinsic::arm_neon_vqshiftsu:
3735 VShiftOpc = ARMISD::VQSHLsu; break;
3736 case Intrinsic::arm_neon_vqshiftns:
3737 VShiftOpc = ARMISD::VQSHRNs; break;
3738 case Intrinsic::arm_neon_vqshiftnu:
3739 VShiftOpc = ARMISD::VQSHRNu; break;
3740 case Intrinsic::arm_neon_vqshiftnsu:
3741 VShiftOpc = ARMISD::VQSHRNsu; break;
3742 case Intrinsic::arm_neon_vqrshiftns:
3743 VShiftOpc = ARMISD::VQRSHRNs; break;
3744 case Intrinsic::arm_neon_vqrshiftnu:
3745 VShiftOpc = ARMISD::VQRSHRNu; break;
3746 case Intrinsic::arm_neon_vqrshiftnsu:
3747 VShiftOpc = ARMISD::VQRSHRNsu; break;
3750 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3751 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3754 case Intrinsic::arm_neon_vshiftins: {
3755 EVT VT = N->getOperand(1).getValueType();
3757 unsigned VShiftOpc = 0;
3759 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3760 VShiftOpc = ARMISD::VSLI;
3761 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3762 VShiftOpc = ARMISD::VSRI;
3764 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3767 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3768 N->getOperand(1), N->getOperand(2),
3769 DAG.getConstant(Cnt, MVT::i32));
3772 case Intrinsic::arm_neon_vqrshifts:
3773 case Intrinsic::arm_neon_vqrshiftu:
3774 // No immediate versions of these to check for.
3781 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3782 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3783 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3784 /// vector element shift counts are generally not legal, and it is hard to see
3785 /// their values after they get legalized to loads from a constant pool.
3786 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3787 const ARMSubtarget *ST) {
3788 EVT VT = N->getValueType(0);
3790 // Nothing to be done for scalar shifts.
3791 if (! VT.isVector())
3794 assert(ST->hasNEON() && "unexpected vector shift");
3797 switch (N->getOpcode()) {
3798 default: llvm_unreachable("unexpected shift opcode");
3801 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3802 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3803 DAG.getConstant(Cnt, MVT::i32));
3808 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3809 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3810 ARMISD::VSHRs : ARMISD::VSHRu);
3811 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3812 DAG.getConstant(Cnt, MVT::i32));
3818 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3819 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3820 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3821 const ARMSubtarget *ST) {
3822 SDValue N0 = N->getOperand(0);
3824 // Check for sign- and zero-extensions of vector extract operations of 8-
3825 // and 16-bit vector elements. NEON supports these directly. They are
3826 // handled during DAG combining because type legalization will promote them
3827 // to 32-bit types and it is messy to recognize the operations after that.
3828 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3829 SDValue Vec = N0.getOperand(0);
3830 SDValue Lane = N0.getOperand(1);
3831 EVT VT = N->getValueType(0);
3832 EVT EltVT = N0.getValueType();
3833 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3835 if (VT == MVT::i32 &&
3836 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3837 TLI.isTypeLegal(Vec.getValueType())) {
3840 switch (N->getOpcode()) {
3841 default: llvm_unreachable("unexpected opcode");
3842 case ISD::SIGN_EXTEND:
3843 Opc = ARMISD::VGETLANEs;
3845 case ISD::ZERO_EXTEND:
3846 case ISD::ANY_EXTEND:
3847 Opc = ARMISD::VGETLANEu;
3850 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3857 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3858 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3859 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3860 const ARMSubtarget *ST) {
3861 // If the target supports NEON, try to use vmax/vmin instructions for f32
3862 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3863 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3864 // a NaN; only do the transformation when it matches that behavior.
3866 // For now only do this when using NEON for FP operations; if using VFP, it
3867 // is not obvious that the benefit outweighs the cost of switching to the
3869 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3870 N->getValueType(0) != MVT::f32)
3873 SDValue CondLHS = N->getOperand(0);
3874 SDValue CondRHS = N->getOperand(1);
3875 SDValue LHS = N->getOperand(2);
3876 SDValue RHS = N->getOperand(3);
3877 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3879 unsigned Opcode = 0;
3881 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3882 IsReversed = false; // x CC y ? x : y
3883 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3884 IsReversed = true ; // x CC y ? y : x
3898 // If LHS is NaN, an ordered comparison will be false and the result will
3899 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3900 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3901 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3902 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3904 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3905 // will return -0, so vmin can only be used for unsafe math or if one of
3906 // the operands is known to be nonzero.
3907 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3909 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3911 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3920 // If LHS is NaN, an ordered comparison will be false and the result will
3921 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3922 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3923 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3924 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3926 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3927 // will return +0, so vmax can only be used for unsafe math or if one of
3928 // the operands is known to be nonzero.
3929 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3931 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3933 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3939 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3942 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3943 DAGCombinerInfo &DCI) const {
3944 switch (N->getOpcode()) {
3946 case ISD::ADD: return PerformADDCombine(N, DCI);
3947 case ISD::SUB: return PerformSUBCombine(N, DCI);
3948 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3949 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
3952 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
3953 case ISD::SIGN_EXTEND:
3954 case ISD::ZERO_EXTEND:
3955 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
3956 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
3961 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3962 if (!Subtarget->hasV6Ops())
3963 // Pre-v6 does not support unaligned mem access.
3966 // v6+ may or may not support unaligned mem access depending on the system
3968 // FIXME: This is pretty conservative. Should we provide cmdline option to
3969 // control the behaviour?
3970 if (!Subtarget->isTargetDarwin())
3974 switch (VT.getSimpleVT().SimpleTy) {
3981 // FIXME: VLD1 etc with standard alignment is legal.
3985 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3990 switch (VT.getSimpleVT().SimpleTy) {
3991 default: return false;
4006 if ((V & (Scale - 1)) != 0)
4009 return V == (V & ((1LL << 5) - 1));
4012 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4013 const ARMSubtarget *Subtarget) {
4020 switch (VT.getSimpleVT().SimpleTy) {
4021 default: return false;
4026 // + imm12 or - imm8
4028 return V == (V & ((1LL << 8) - 1));
4029 return V == (V & ((1LL << 12) - 1));
4032 // Same as ARM mode. FIXME: NEON?
4033 if (!Subtarget->hasVFP2())
4038 return V == (V & ((1LL << 8) - 1));
4042 /// isLegalAddressImmediate - Return true if the integer value can be used
4043 /// as the offset of the target addressing mode for load / store of the
4045 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4046 const ARMSubtarget *Subtarget) {
4053 if (Subtarget->isThumb1Only())
4054 return isLegalT1AddressImmediate(V, VT);
4055 else if (Subtarget->isThumb2())
4056 return isLegalT2AddressImmediate(V, VT, Subtarget);
4061 switch (VT.getSimpleVT().SimpleTy) {
4062 default: return false;
4067 return V == (V & ((1LL << 12) - 1));
4070 return V == (V & ((1LL << 8) - 1));
4073 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4078 return V == (V & ((1LL << 8) - 1));
4082 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4084 int Scale = AM.Scale;
4088 switch (VT.getSimpleVT().SimpleTy) {
4089 default: return false;
4098 return Scale == 2 || Scale == 4 || Scale == 8;
4101 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4105 // Note, we allow "void" uses (basically, uses that aren't loads or
4106 // stores), because arm allows folding a scale into many arithmetic
4107 // operations. This should be made more precise and revisited later.
4109 // Allow r << imm, but the imm has to be a multiple of two.
4110 if (Scale & 1) return false;
4111 return isPowerOf2_32(Scale);
4115 /// isLegalAddressingMode - Return true if the addressing mode represented
4116 /// by AM is legal for this target, for a load/store of the specified type.
4117 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4118 const Type *Ty) const {
4119 EVT VT = getValueType(Ty, true);
4120 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4123 // Can never fold addr of global into load/store.
4128 case 0: // no scale reg, must be "r+i" or "r", or "i".
4131 if (Subtarget->isThumb1Only())
4135 // ARM doesn't support any R+R*scale+imm addr modes.
4142 if (Subtarget->isThumb2())
4143 return isLegalT2ScaledAddressingMode(AM, VT);
4145 int Scale = AM.Scale;
4146 switch (VT.getSimpleVT().SimpleTy) {
4147 default: return false;
4151 if (Scale < 0) Scale = -Scale;
4155 return isPowerOf2_32(Scale & ~1);
4159 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4164 // Note, we allow "void" uses (basically, uses that aren't loads or
4165 // stores), because arm allows folding a scale into many arithmetic
4166 // operations. This should be made more precise and revisited later.
4168 // Allow r << imm, but the imm has to be a multiple of two.
4169 if (Scale & 1) return false;
4170 return isPowerOf2_32(Scale);
4177 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4178 /// icmp immediate, that is the target has icmp instructions which can compare
4179 /// a register against the immediate without having to materialize the
4180 /// immediate into a register.
4181 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4182 if (!Subtarget->isThumb())
4183 return ARM_AM::getSOImmVal(Imm) != -1;
4184 if (Subtarget->isThumb2())
4185 return ARM_AM::getT2SOImmVal(Imm) != -1;
4186 return Imm >= 0 && Imm <= 255;
4189 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4190 bool isSEXTLoad, SDValue &Base,
4191 SDValue &Offset, bool &isInc,
4192 SelectionDAG &DAG) {
4193 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4196 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4198 Base = Ptr->getOperand(0);
4199 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4200 int RHSC = (int)RHS->getZExtValue();
4201 if (RHSC < 0 && RHSC > -256) {
4202 assert(Ptr->getOpcode() == ISD::ADD);
4204 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4208 isInc = (Ptr->getOpcode() == ISD::ADD);
4209 Offset = Ptr->getOperand(1);
4211 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4213 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4214 int RHSC = (int)RHS->getZExtValue();
4215 if (RHSC < 0 && RHSC > -0x1000) {
4216 assert(Ptr->getOpcode() == ISD::ADD);
4218 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4219 Base = Ptr->getOperand(0);
4224 if (Ptr->getOpcode() == ISD::ADD) {
4226 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4227 if (ShOpcVal != ARM_AM::no_shift) {
4228 Base = Ptr->getOperand(1);
4229 Offset = Ptr->getOperand(0);
4231 Base = Ptr->getOperand(0);
4232 Offset = Ptr->getOperand(1);
4237 isInc = (Ptr->getOpcode() == ISD::ADD);
4238 Base = Ptr->getOperand(0);
4239 Offset = Ptr->getOperand(1);
4243 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4247 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4248 bool isSEXTLoad, SDValue &Base,
4249 SDValue &Offset, bool &isInc,
4250 SelectionDAG &DAG) {
4251 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4254 Base = Ptr->getOperand(0);
4255 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4256 int RHSC = (int)RHS->getZExtValue();
4257 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4258 assert(Ptr->getOpcode() == ISD::ADD);
4260 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4262 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4263 isInc = Ptr->getOpcode() == ISD::ADD;
4264 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4272 /// getPreIndexedAddressParts - returns true by value, base pointer and
4273 /// offset pointer and addressing mode by reference if the node's address
4274 /// can be legally represented as pre-indexed load / store address.
4276 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4278 ISD::MemIndexedMode &AM,
4279 SelectionDAG &DAG) const {
4280 if (Subtarget->isThumb1Only())
4285 bool isSEXTLoad = false;
4286 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4287 Ptr = LD->getBasePtr();
4288 VT = LD->getMemoryVT();
4289 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4290 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4291 Ptr = ST->getBasePtr();
4292 VT = ST->getMemoryVT();
4297 bool isLegal = false;
4298 if (Subtarget->isThumb2())
4299 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4300 Offset, isInc, DAG);
4302 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4303 Offset, isInc, DAG);
4307 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4311 /// getPostIndexedAddressParts - returns true by value, base pointer and
4312 /// offset pointer and addressing mode by reference if this node can be
4313 /// combined with a load / store to form a post-indexed load / store.
4314 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4317 ISD::MemIndexedMode &AM,
4318 SelectionDAG &DAG) const {
4319 if (Subtarget->isThumb1Only())
4324 bool isSEXTLoad = false;
4325 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4326 VT = LD->getMemoryVT();
4327 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4328 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4329 VT = ST->getMemoryVT();
4334 bool isLegal = false;
4335 if (Subtarget->isThumb2())
4336 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4339 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4344 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4348 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4352 const SelectionDAG &DAG,
4353 unsigned Depth) const {
4354 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4355 switch (Op.getOpcode()) {
4357 case ARMISD::CMOV: {
4358 // Bits are known zero/one if known on the LHS and RHS.
4359 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4360 if (KnownZero == 0 && KnownOne == 0) return;
4362 APInt KnownZeroRHS, KnownOneRHS;
4363 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4364 KnownZeroRHS, KnownOneRHS, Depth+1);
4365 KnownZero &= KnownZeroRHS;
4366 KnownOne &= KnownOneRHS;
4372 //===----------------------------------------------------------------------===//
4373 // ARM Inline Assembly Support
4374 //===----------------------------------------------------------------------===//
4376 /// getConstraintType - Given a constraint letter, return the type of
4377 /// constraint it is for this target.
4378 ARMTargetLowering::ConstraintType
4379 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4380 if (Constraint.size() == 1) {
4381 switch (Constraint[0]) {
4383 case 'l': return C_RegisterClass;
4384 case 'w': return C_RegisterClass;
4387 return TargetLowering::getConstraintType(Constraint);
4390 std::pair<unsigned, const TargetRegisterClass*>
4391 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4393 if (Constraint.size() == 1) {
4394 // GCC ARM Constraint Letters
4395 switch (Constraint[0]) {
4397 if (Subtarget->isThumb())
4398 return std::make_pair(0U, ARM::tGPRRegisterClass);
4400 return std::make_pair(0U, ARM::GPRRegisterClass);
4402 return std::make_pair(0U, ARM::GPRRegisterClass);
4405 return std::make_pair(0U, ARM::SPRRegisterClass);
4406 if (VT.getSizeInBits() == 64)
4407 return std::make_pair(0U, ARM::DPRRegisterClass);
4408 if (VT.getSizeInBits() == 128)
4409 return std::make_pair(0U, ARM::QPRRegisterClass);
4413 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4416 std::vector<unsigned> ARMTargetLowering::
4417 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4419 if (Constraint.size() != 1)
4420 return std::vector<unsigned>();
4422 switch (Constraint[0]) { // GCC ARM Constraint Letters
4425 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4426 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4429 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4430 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4431 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4432 ARM::R12, ARM::LR, 0);
4435 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4436 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4437 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4438 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4439 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4440 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4441 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4442 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4443 if (VT.getSizeInBits() == 64)
4444 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4445 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4446 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4447 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4448 if (VT.getSizeInBits() == 128)
4449 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4450 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4454 return std::vector<unsigned>();
4457 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4458 /// vector. If it is invalid, don't add anything to Ops.
4459 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4462 std::vector<SDValue>&Ops,
4463 SelectionDAG &DAG) const {
4464 SDValue Result(0, 0);
4466 switch (Constraint) {
4468 case 'I': case 'J': case 'K': case 'L':
4469 case 'M': case 'N': case 'O':
4470 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4474 int64_t CVal64 = C->getSExtValue();
4475 int CVal = (int) CVal64;
4476 // None of these constraints allow values larger than 32 bits. Check
4477 // that the value fits in an int.
4481 switch (Constraint) {
4483 if (Subtarget->isThumb1Only()) {
4484 // This must be a constant between 0 and 255, for ADD
4486 if (CVal >= 0 && CVal <= 255)
4488 } else if (Subtarget->isThumb2()) {
4489 // A constant that can be used as an immediate value in a
4490 // data-processing instruction.
4491 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4494 // A constant that can be used as an immediate value in a
4495 // data-processing instruction.
4496 if (ARM_AM::getSOImmVal(CVal) != -1)
4502 if (Subtarget->isThumb()) { // FIXME thumb2
4503 // This must be a constant between -255 and -1, for negated ADD
4504 // immediates. This can be used in GCC with an "n" modifier that
4505 // prints the negated value, for use with SUB instructions. It is
4506 // not useful otherwise but is implemented for compatibility.
4507 if (CVal >= -255 && CVal <= -1)
4510 // This must be a constant between -4095 and 4095. It is not clear
4511 // what this constraint is intended for. Implemented for
4512 // compatibility with GCC.
4513 if (CVal >= -4095 && CVal <= 4095)
4519 if (Subtarget->isThumb1Only()) {
4520 // A 32-bit value where only one byte has a nonzero value. Exclude
4521 // zero to match GCC. This constraint is used by GCC internally for
4522 // constants that can be loaded with a move/shift combination.
4523 // It is not useful otherwise but is implemented for compatibility.
4524 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4526 } else if (Subtarget->isThumb2()) {
4527 // A constant whose bitwise inverse can be used as an immediate
4528 // value in a data-processing instruction. This can be used in GCC
4529 // with a "B" modifier that prints the inverted value, for use with
4530 // BIC and MVN instructions. It is not useful otherwise but is
4531 // implemented for compatibility.
4532 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4535 // A constant whose bitwise inverse can be used as an immediate
4536 // value in a data-processing instruction. This can be used in GCC
4537 // with a "B" modifier that prints the inverted value, for use with
4538 // BIC and MVN instructions. It is not useful otherwise but is
4539 // implemented for compatibility.
4540 if (ARM_AM::getSOImmVal(~CVal) != -1)
4546 if (Subtarget->isThumb1Only()) {
4547 // This must be a constant between -7 and 7,
4548 // for 3-operand ADD/SUB immediate instructions.
4549 if (CVal >= -7 && CVal < 7)
4551 } else if (Subtarget->isThumb2()) {
4552 // A constant whose negation can be used as an immediate value in a
4553 // data-processing instruction. This can be used in GCC with an "n"
4554 // modifier that prints the negated value, for use with SUB
4555 // instructions. It is not useful otherwise but is implemented for
4557 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4560 // A constant whose negation can be used as an immediate value in a
4561 // data-processing instruction. This can be used in GCC with an "n"
4562 // modifier that prints the negated value, for use with SUB
4563 // instructions. It is not useful otherwise but is implemented for
4565 if (ARM_AM::getSOImmVal(-CVal) != -1)
4571 if (Subtarget->isThumb()) { // FIXME thumb2
4572 // This must be a multiple of 4 between 0 and 1020, for
4573 // ADD sp + immediate.
4574 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4577 // A power of two or a constant between 0 and 32. This is used in
4578 // GCC for the shift amount on shifted register operands, but it is
4579 // useful in general for any shift amounts.
4580 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4586 if (Subtarget->isThumb()) { // FIXME thumb2
4587 // This must be a constant between 0 and 31, for shift amounts.
4588 if (CVal >= 0 && CVal <= 31)
4594 if (Subtarget->isThumb()) { // FIXME thumb2
4595 // This must be a multiple of 4 between -508 and 508, for
4596 // ADD/SUB sp = sp + immediate.
4597 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4602 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4606 if (Result.getNode()) {
4607 Ops.push_back(Result);
4610 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4615 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4616 // The ARM target isn't yet aware of offsets.
4620 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4621 APInt Imm = FPImm.bitcastToAPInt();
4622 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4623 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4624 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4626 // We can handle 4 bits of mantissa.
4627 // mantissa = (16+UInt(e:f:g:h))/16.
4628 if (Mantissa & 0x7ffff)
4631 if ((Mantissa & 0xf) != Mantissa)
4634 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4635 if (Exp < -3 || Exp > 4)
4637 Exp = ((Exp+3) & 0x7) ^ 4;
4639 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4642 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4643 APInt Imm = FPImm.bitcastToAPInt();
4644 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4645 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4646 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4648 // We can handle 4 bits of mantissa.
4649 // mantissa = (16+UInt(e:f:g:h))/16.
4650 if (Mantissa & 0xffffffffffffLL)
4653 if ((Mantissa & 0xf) != Mantissa)
4656 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4657 if (Exp < -3 || Exp > 4)
4659 Exp = ((Exp+3) & 0x7) ^ 4;
4661 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4664 /// isFPImmLegal - Returns true if the target can instruction select the
4665 /// specified FP immediate natively. If false, the legalizer will
4666 /// materialize the FP immediate as a load from a constant pool.
4667 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4668 if (!Subtarget->hasVFP3())
4671 return ARM::getVFPf32Imm(Imm) != -1;
4673 return ARM::getVFPf64Imm(Imm) != -1;