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/Instruction.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/CodeGen/CallingConvLower.h"
32 #include "llvm/CodeGen/MachineBasicBlock.h"
33 #include "llvm/CodeGen/MachineFrameInfo.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineInstrBuilder.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/CodeGen/SelectionDAG.h"
39 #include "llvm/Target/TargetOptions.h"
40 #include "llvm/ADT/VectorExtras.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/MathExtras.h"
46 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
47 CCValAssign::LocInfo &LocInfo,
48 ISD::ArgFlagsTy &ArgFlags,
50 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
51 CCValAssign::LocInfo &LocInfo,
52 ISD::ArgFlagsTy &ArgFlags,
54 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
55 CCValAssign::LocInfo &LocInfo,
56 ISD::ArgFlagsTy &ArgFlags,
58 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
59 CCValAssign::LocInfo &LocInfo,
60 ISD::ArgFlagsTy &ArgFlags,
63 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
64 EVT PromotedBitwiseVT) {
65 if (VT != PromotedLdStVT) {
66 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
67 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
68 PromotedLdStVT.getSimpleVT());
70 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
71 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
72 PromotedLdStVT.getSimpleVT());
75 EVT ElemTy = VT.getVectorElementType();
76 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
77 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
78 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
79 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
80 if (ElemTy != MVT::i32) {
81 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
82 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
83 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
84 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
86 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
87 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
88 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
89 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
91 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
92 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
93 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
96 // Promote all bit-wise operations.
97 if (VT.isInteger() && VT != PromotedBitwiseVT) {
98 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
99 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
100 PromotedBitwiseVT.getSimpleVT());
101 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
102 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
103 PromotedBitwiseVT.getSimpleVT());
104 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
105 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
106 PromotedBitwiseVT.getSimpleVT());
109 // Neon does not support vector divide/remainder operations.
110 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
111 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
112 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
118 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
119 addRegisterClass(VT, ARM::DPRRegisterClass);
120 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
123 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
124 addRegisterClass(VT, ARM::QPRRegisterClass);
125 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
128 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
129 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
130 return new TargetLoweringObjectFileMachO();
131 return new ARMElfTargetObjectFile();
134 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
135 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
136 Subtarget = &TM.getSubtarget<ARMSubtarget>();
138 if (Subtarget->isTargetDarwin()) {
139 // Uses VFP for Thumb libfuncs if available.
140 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
141 // Single-precision floating-point arithmetic.
142 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
143 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
144 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
145 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
147 // Double-precision floating-point arithmetic.
148 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
149 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
150 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
151 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
153 // Single-precision comparisons.
154 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
155 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
156 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
157 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
158 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
159 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
160 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
161 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
163 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
164 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
165 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
166 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
167 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
172 // Double-precision comparisons.
173 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
174 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
175 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
176 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
177 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
178 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
179 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
180 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
182 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
183 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
191 // Floating-point to integer conversions.
192 // i64 conversions are done via library routines even when generating VFP
193 // instructions, so use the same ones.
194 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
195 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
196 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
197 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
199 // Conversions between floating types.
200 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
201 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
203 // Integer to floating-point conversions.
204 // i64 conversions are done via library routines even when generating VFP
205 // instructions, so use the same ones.
206 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
207 // e.g., __floatunsidf vs. __floatunssidfvfp.
208 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
209 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
210 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
211 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
215 // These libcalls are not available in 32-bit.
216 setLibcallName(RTLIB::SHL_I128, 0);
217 setLibcallName(RTLIB::SRL_I128, 0);
218 setLibcallName(RTLIB::SRA_I128, 0);
220 // Libcalls should use the AAPCS base standard ABI, even if hard float
221 // is in effect, as per the ARM RTABI specification, section 4.1.2.
222 if (Subtarget->isAAPCS_ABI()) {
223 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
224 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
225 CallingConv::ARM_AAPCS);
229 if (Subtarget->isThumb1Only())
230 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
232 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
233 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
234 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
235 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
237 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
240 if (Subtarget->hasNEON()) {
241 addDRTypeForNEON(MVT::v2f32);
242 addDRTypeForNEON(MVT::v8i8);
243 addDRTypeForNEON(MVT::v4i16);
244 addDRTypeForNEON(MVT::v2i32);
245 addDRTypeForNEON(MVT::v1i64);
247 addQRTypeForNEON(MVT::v4f32);
248 addQRTypeForNEON(MVT::v2f64);
249 addQRTypeForNEON(MVT::v16i8);
250 addQRTypeForNEON(MVT::v8i16);
251 addQRTypeForNEON(MVT::v4i32);
252 addQRTypeForNEON(MVT::v2i64);
254 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
255 // neither Neon nor VFP support any arithmetic operations on it.
256 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
257 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
258 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
259 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
260 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
261 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
262 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
263 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
264 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
265 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
266 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
267 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
268 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
269 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
270 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
271 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
272 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
273 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
274 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
275 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
276 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
277 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
278 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
279 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
281 // Neon does not support some operations on v1i64 and v2i64 types.
282 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
283 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
284 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
285 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
287 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
288 setTargetDAGCombine(ISD::SHL);
289 setTargetDAGCombine(ISD::SRL);
290 setTargetDAGCombine(ISD::SRA);
291 setTargetDAGCombine(ISD::SIGN_EXTEND);
292 setTargetDAGCombine(ISD::ZERO_EXTEND);
293 setTargetDAGCombine(ISD::ANY_EXTEND);
296 computeRegisterProperties();
298 // ARM does not have f32 extending load.
299 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
301 // ARM does not have i1 sign extending load.
302 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
304 // ARM supports all 4 flavors of integer indexed load / store.
305 if (!Subtarget->isThumb1Only()) {
306 for (unsigned im = (unsigned)ISD::PRE_INC;
307 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
308 setIndexedLoadAction(im, MVT::i1, Legal);
309 setIndexedLoadAction(im, MVT::i8, Legal);
310 setIndexedLoadAction(im, MVT::i16, Legal);
311 setIndexedLoadAction(im, MVT::i32, Legal);
312 setIndexedStoreAction(im, MVT::i1, Legal);
313 setIndexedStoreAction(im, MVT::i8, Legal);
314 setIndexedStoreAction(im, MVT::i16, Legal);
315 setIndexedStoreAction(im, MVT::i32, Legal);
319 // i64 operation support.
320 if (Subtarget->isThumb1Only()) {
321 setOperationAction(ISD::MUL, MVT::i64, Expand);
322 setOperationAction(ISD::MULHU, MVT::i32, Expand);
323 setOperationAction(ISD::MULHS, MVT::i32, Expand);
324 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
325 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
327 setOperationAction(ISD::MUL, MVT::i64, Expand);
328 setOperationAction(ISD::MULHU, MVT::i32, Expand);
329 if (!Subtarget->hasV6Ops())
330 setOperationAction(ISD::MULHS, MVT::i32, Expand);
332 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
333 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
334 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
335 setOperationAction(ISD::SRL, MVT::i64, Custom);
336 setOperationAction(ISD::SRA, MVT::i64, Custom);
338 // ARM does not have ROTL.
339 setOperationAction(ISD::ROTL, MVT::i32, Expand);
340 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
341 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
342 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
343 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
345 // Only ARMv6 has BSWAP.
346 if (!Subtarget->hasV6Ops())
347 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
349 // These are expanded into libcalls.
350 setOperationAction(ISD::SDIV, MVT::i32, Expand);
351 setOperationAction(ISD::UDIV, MVT::i32, Expand);
352 setOperationAction(ISD::SREM, MVT::i32, Expand);
353 setOperationAction(ISD::UREM, MVT::i32, Expand);
354 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
355 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
357 // Support label based line numbers.
358 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
359 setOperationAction(ISD::DEBUG_LOC, MVT::Other, 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);
366 // Use the default implementation.
367 setOperationAction(ISD::VASTART, MVT::Other, Custom);
368 setOperationAction(ISD::VAARG, MVT::Other, Expand);
369 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
370 setOperationAction(ISD::VAEND, MVT::Other, Expand);
371 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
372 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
373 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
374 // FIXME: Shouldn't need this, since no register is used, but the legalizer
375 // doesn't yet know how to not do that for SjLj.
376 setExceptionSelectorRegister(ARM::R0);
377 if (Subtarget->isThumb())
378 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
380 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
381 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
383 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
384 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
385 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
387 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
389 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
390 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
391 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
393 // We want to custom lower some of our intrinsics.
394 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
395 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
396 setOperationAction(ISD::INTRINSIC_VOID, 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::FMRRD - 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::FTOSI: return "ARMISD::FTOSI";
488 case ARMISD::FTOUI: return "ARMISD::FTOUI";
489 case ARMISD::SITOF: return "ARMISD::SITOF";
490 case ARMISD::UITOF: return "ARMISD::UITOF";
492 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
493 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
494 case ARMISD::RRX: return "ARMISD::RRX";
496 case ARMISD::FMRRD: return "ARMISD::FMRRD";
497 case ARMISD::FMDRR: return "ARMISD::FMDRR";
499 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
501 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
503 case ARMISD::VCEQ: return "ARMISD::VCEQ";
504 case ARMISD::VCGE: return "ARMISD::VCGE";
505 case ARMISD::VCGEU: return "ARMISD::VCGEU";
506 case ARMISD::VCGT: return "ARMISD::VCGT";
507 case ARMISD::VCGTU: return "ARMISD::VCGTU";
508 case ARMISD::VTST: return "ARMISD::VTST";
510 case ARMISD::VSHL: return "ARMISD::VSHL";
511 case ARMISD::VSHRs: return "ARMISD::VSHRs";
512 case ARMISD::VSHRu: return "ARMISD::VSHRu";
513 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
514 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
515 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
516 case ARMISD::VSHRN: return "ARMISD::VSHRN";
517 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
518 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
519 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
520 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
521 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
522 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
523 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
524 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
525 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
526 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
527 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
528 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
529 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
530 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
531 case ARMISD::VDUP: return "ARMISD::VDUP";
532 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
533 case ARMISD::VEXT: return "ARMISD::VEXT";
534 case ARMISD::VREV64: return "ARMISD::VREV64";
535 case ARMISD::VREV32: return "ARMISD::VREV32";
536 case ARMISD::VREV16: return "ARMISD::VREV16";
537 case ARMISD::VZIP: return "ARMISD::VZIP";
538 case ARMISD::VUZP: return "ARMISD::VUZP";
539 case ARMISD::VTRN: return "ARMISD::VTRN";
543 /// getFunctionAlignment - Return the Log2 alignment of this function.
544 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
545 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
548 //===----------------------------------------------------------------------===//
550 //===----------------------------------------------------------------------===//
552 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
553 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
555 default: llvm_unreachable("Unknown condition code!");
556 case ISD::SETNE: return ARMCC::NE;
557 case ISD::SETEQ: return ARMCC::EQ;
558 case ISD::SETGT: return ARMCC::GT;
559 case ISD::SETGE: return ARMCC::GE;
560 case ISD::SETLT: return ARMCC::LT;
561 case ISD::SETLE: return ARMCC::LE;
562 case ISD::SETUGT: return ARMCC::HI;
563 case ISD::SETUGE: return ARMCC::HS;
564 case ISD::SETULT: return ARMCC::LO;
565 case ISD::SETULE: return ARMCC::LS;
569 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
570 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
571 ARMCC::CondCodes &CondCode2) {
572 CondCode2 = ARMCC::AL;
574 default: llvm_unreachable("Unknown FP condition!");
576 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
578 case ISD::SETOGT: CondCode = ARMCC::GT; break;
580 case ISD::SETOGE: CondCode = ARMCC::GE; break;
581 case ISD::SETOLT: CondCode = ARMCC::MI; break;
582 case ISD::SETOLE: CondCode = ARMCC::LS; break;
583 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
584 case ISD::SETO: CondCode = ARMCC::VC; break;
585 case ISD::SETUO: CondCode = ARMCC::VS; break;
586 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
587 case ISD::SETUGT: CondCode = ARMCC::HI; break;
588 case ISD::SETUGE: CondCode = ARMCC::PL; break;
590 case ISD::SETULT: CondCode = ARMCC::LT; break;
592 case ISD::SETULE: CondCode = ARMCC::LE; break;
594 case ISD::SETUNE: CondCode = ARMCC::NE; break;
598 //===----------------------------------------------------------------------===//
599 // Calling Convention Implementation
600 //===----------------------------------------------------------------------===//
602 #include "ARMGenCallingConv.inc"
604 // APCS f64 is in register pairs, possibly split to stack
605 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
606 CCValAssign::LocInfo &LocInfo,
607 CCState &State, bool CanFail) {
608 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
610 // Try to get the first register.
611 if (unsigned Reg = State.AllocateReg(RegList, 4))
612 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
614 // For the 2nd half of a v2f64, do not fail.
618 // Put the whole thing on the stack.
619 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
620 State.AllocateStack(8, 4),
625 // Try to get the second register.
626 if (unsigned Reg = State.AllocateReg(RegList, 4))
627 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
629 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
630 State.AllocateStack(4, 4),
635 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
636 CCValAssign::LocInfo &LocInfo,
637 ISD::ArgFlagsTy &ArgFlags,
639 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
641 if (LocVT == MVT::v2f64 &&
642 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
644 return true; // we handled it
647 // AAPCS f64 is in aligned register pairs
648 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
649 CCValAssign::LocInfo &LocInfo,
650 CCState &State, bool CanFail) {
651 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
652 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
654 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
656 // For the 2nd half of a v2f64, do not just fail.
660 // Put the whole thing on the stack.
661 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
662 State.AllocateStack(8, 8),
668 for (i = 0; i < 2; ++i)
669 if (HiRegList[i] == Reg)
672 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
673 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
678 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
679 CCValAssign::LocInfo &LocInfo,
680 ISD::ArgFlagsTy &ArgFlags,
682 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
684 if (LocVT == MVT::v2f64 &&
685 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
687 return true; // we handled it
690 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
691 CCValAssign::LocInfo &LocInfo, CCState &State) {
692 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
693 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
695 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
697 return false; // we didn't handle it
700 for (i = 0; i < 2; ++i)
701 if (HiRegList[i] == Reg)
704 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
705 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
710 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
711 CCValAssign::LocInfo &LocInfo,
712 ISD::ArgFlagsTy &ArgFlags,
714 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
716 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
718 return true; // we handled it
721 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
722 CCValAssign::LocInfo &LocInfo,
723 ISD::ArgFlagsTy &ArgFlags,
725 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
729 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
730 /// given CallingConvention value.
731 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
733 bool isVarArg) const {
736 llvm_unreachable("Unsupported calling convention");
738 case CallingConv::Fast:
739 // Use target triple & subtarget features to do actual dispatch.
740 if (Subtarget->isAAPCS_ABI()) {
741 if (Subtarget->hasVFP2() &&
742 FloatABIType == FloatABI::Hard && !isVarArg)
743 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
745 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
747 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
748 case CallingConv::ARM_AAPCS_VFP:
749 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
750 case CallingConv::ARM_AAPCS:
751 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
752 case CallingConv::ARM_APCS:
753 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
757 /// LowerCallResult - Lower the result values of a call into the
758 /// appropriate copies out of appropriate physical registers.
760 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
761 CallingConv::ID CallConv, bool isVarArg,
762 const SmallVectorImpl<ISD::InputArg> &Ins,
763 DebugLoc dl, SelectionDAG &DAG,
764 SmallVectorImpl<SDValue> &InVals) {
766 // Assign locations to each value returned by this call.
767 SmallVector<CCValAssign, 16> RVLocs;
768 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
769 RVLocs, *DAG.getContext());
770 CCInfo.AnalyzeCallResult(Ins,
771 CCAssignFnForNode(CallConv, /* Return*/ true,
774 // Copy all of the result registers out of their specified physreg.
775 for (unsigned i = 0; i != RVLocs.size(); ++i) {
776 CCValAssign VA = RVLocs[i];
779 if (VA.needsCustom()) {
780 // Handle f64 or half of a v2f64.
781 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
783 Chain = Lo.getValue(1);
784 InFlag = Lo.getValue(2);
785 VA = RVLocs[++i]; // skip ahead to next loc
786 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
788 Chain = Hi.getValue(1);
789 InFlag = Hi.getValue(2);
790 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
792 if (VA.getLocVT() == MVT::v2f64) {
793 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
794 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
795 DAG.getConstant(0, MVT::i32));
797 VA = RVLocs[++i]; // skip ahead to next loc
798 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
799 Chain = Lo.getValue(1);
800 InFlag = Lo.getValue(2);
801 VA = RVLocs[++i]; // skip ahead to next loc
802 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
803 Chain = Hi.getValue(1);
804 InFlag = Hi.getValue(2);
805 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
806 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
807 DAG.getConstant(1, MVT::i32));
810 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
812 Chain = Val.getValue(1);
813 InFlag = Val.getValue(2);
816 switch (VA.getLocInfo()) {
817 default: llvm_unreachable("Unknown loc info!");
818 case CCValAssign::Full: break;
819 case CCValAssign::BCvt:
820 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
824 InVals.push_back(Val);
830 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
831 /// by "Src" to address "Dst" of size "Size". Alignment information is
832 /// specified by the specific parameter attribute. The copy will be passed as
833 /// a byval function parameter.
834 /// Sometimes what we are copying is the end of a larger object, the part that
835 /// does not fit in registers.
837 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
838 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
840 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
841 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
842 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
845 /// LowerMemOpCallTo - Store the argument to the stack.
847 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
848 SDValue StackPtr, SDValue Arg,
849 DebugLoc dl, SelectionDAG &DAG,
850 const CCValAssign &VA,
851 ISD::ArgFlagsTy Flags) {
852 unsigned LocMemOffset = VA.getLocMemOffset();
853 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
854 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
855 if (Flags.isByVal()) {
856 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
858 return DAG.getStore(Chain, dl, Arg, PtrOff,
859 PseudoSourceValue::getStack(), LocMemOffset);
862 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
863 SDValue Chain, SDValue &Arg,
864 RegsToPassVector &RegsToPass,
865 CCValAssign &VA, CCValAssign &NextVA,
867 SmallVector<SDValue, 8> &MemOpChains,
868 ISD::ArgFlagsTy Flags) {
870 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
871 DAG.getVTList(MVT::i32, MVT::i32), Arg);
872 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
874 if (NextVA.isRegLoc())
875 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
877 assert(NextVA.isMemLoc());
878 if (StackPtr.getNode() == 0)
879 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
881 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
887 /// LowerCall - Lowering a call into a callseq_start <-
888 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
891 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
892 CallingConv::ID CallConv, bool isVarArg,
894 const SmallVectorImpl<ISD::OutputArg> &Outs,
895 const SmallVectorImpl<ISD::InputArg> &Ins,
896 DebugLoc dl, SelectionDAG &DAG,
897 SmallVectorImpl<SDValue> &InVals) {
899 // Analyze operands of the call, assigning locations to each operand.
900 SmallVector<CCValAssign, 16> ArgLocs;
901 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
903 CCInfo.AnalyzeCallOperands(Outs,
904 CCAssignFnForNode(CallConv, /* Return*/ false,
907 // Get a count of how many bytes are to be pushed on the stack.
908 unsigned NumBytes = CCInfo.getNextStackOffset();
910 // Adjust the stack pointer for the new arguments...
911 // These operations are automatically eliminated by the prolog/epilog pass
912 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
914 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
916 RegsToPassVector RegsToPass;
917 SmallVector<SDValue, 8> MemOpChains;
919 // Walk the register/memloc assignments, inserting copies/loads. In the case
920 // of tail call optimization, arguments are handled later.
921 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
924 CCValAssign &VA = ArgLocs[i];
925 SDValue Arg = Outs[realArgIdx].Val;
926 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
928 // Promote the value if needed.
929 switch (VA.getLocInfo()) {
930 default: llvm_unreachable("Unknown loc info!");
931 case CCValAssign::Full: break;
932 case CCValAssign::SExt:
933 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
935 case CCValAssign::ZExt:
936 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
938 case CCValAssign::AExt:
939 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
941 case CCValAssign::BCvt:
942 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
946 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
947 if (VA.needsCustom()) {
948 if (VA.getLocVT() == MVT::v2f64) {
949 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
950 DAG.getConstant(0, MVT::i32));
951 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
952 DAG.getConstant(1, MVT::i32));
954 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
955 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
957 VA = ArgLocs[++i]; // skip ahead to next loc
959 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
960 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
962 assert(VA.isMemLoc());
963 if (StackPtr.getNode() == 0)
964 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
966 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
967 dl, DAG, VA, Flags));
970 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
971 StackPtr, MemOpChains, Flags);
973 } else if (VA.isRegLoc()) {
974 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
976 assert(VA.isMemLoc());
977 if (StackPtr.getNode() == 0)
978 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
980 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
981 dl, DAG, VA, Flags));
985 if (!MemOpChains.empty())
986 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
987 &MemOpChains[0], MemOpChains.size());
989 // Build a sequence of copy-to-reg nodes chained together with token chain
990 // and flag operands which copy the outgoing args into the appropriate regs.
992 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
993 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
994 RegsToPass[i].second, InFlag);
995 InFlag = Chain.getValue(1);
998 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
999 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1000 // node so that legalize doesn't hack it.
1001 bool isDirect = false;
1002 bool isARMFunc = false;
1003 bool isLocalARMFunc = false;
1004 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1005 GlobalValue *GV = G->getGlobal();
1007 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1008 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1009 getTargetMachine().getRelocationModel() != Reloc::Static;
1010 isARMFunc = !Subtarget->isThumb() || isStub;
1011 // ARM call to a local ARM function is predicable.
1012 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1013 // tBX takes a register source operand.
1014 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1015 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1018 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1019 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1020 Callee = DAG.getLoad(getPointerTy(), dl,
1021 DAG.getEntryNode(), CPAddr, NULL, 0);
1022 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1023 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1024 getPointerTy(), Callee, PICLabel);
1026 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1027 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1029 bool isStub = Subtarget->isTargetDarwin() &&
1030 getTargetMachine().getRelocationModel() != Reloc::Static;
1031 isARMFunc = !Subtarget->isThumb() || isStub;
1032 // tBX takes a register source operand.
1033 const char *Sym = S->getSymbol();
1034 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1035 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1036 Sym, ARMPCLabelIndex, 4);
1037 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1038 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1039 Callee = DAG.getLoad(getPointerTy(), dl,
1040 DAG.getEntryNode(), CPAddr, NULL, 0);
1041 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1042 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1043 getPointerTy(), Callee, PICLabel);
1045 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1048 // FIXME: handle tail calls differently.
1050 if (Subtarget->isThumb()) {
1051 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1052 CallOpc = ARMISD::CALL_NOLINK;
1054 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1056 CallOpc = (isDirect || Subtarget->hasV5TOps())
1057 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1058 : ARMISD::CALL_NOLINK;
1060 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1061 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1062 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1063 InFlag = Chain.getValue(1);
1066 std::vector<SDValue> Ops;
1067 Ops.push_back(Chain);
1068 Ops.push_back(Callee);
1070 // Add argument registers to the end of the list so that they are known live
1072 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1073 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1074 RegsToPass[i].second.getValueType()));
1076 if (InFlag.getNode())
1077 Ops.push_back(InFlag);
1078 // Returns a chain and a flag for retval copy to use.
1079 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1080 &Ops[0], Ops.size());
1081 InFlag = Chain.getValue(1);
1083 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1084 DAG.getIntPtrConstant(0, true), InFlag);
1086 InFlag = Chain.getValue(1);
1088 // Handle result values, copying them out of physregs into vregs that we
1090 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1095 ARMTargetLowering::LowerReturn(SDValue Chain,
1096 CallingConv::ID CallConv, bool isVarArg,
1097 const SmallVectorImpl<ISD::OutputArg> &Outs,
1098 DebugLoc dl, SelectionDAG &DAG) {
1100 // CCValAssign - represent the assignment of the return value to a location.
1101 SmallVector<CCValAssign, 16> RVLocs;
1103 // CCState - Info about the registers and stack slots.
1104 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1107 // Analyze outgoing return values.
1108 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1111 // If this is the first return lowered for this function, add
1112 // the regs to the liveout set for the function.
1113 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1114 for (unsigned i = 0; i != RVLocs.size(); ++i)
1115 if (RVLocs[i].isRegLoc())
1116 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1121 // Copy the result values into the output registers.
1122 for (unsigned i = 0, realRVLocIdx = 0;
1124 ++i, ++realRVLocIdx) {
1125 CCValAssign &VA = RVLocs[i];
1126 assert(VA.isRegLoc() && "Can only return in registers!");
1128 SDValue Arg = Outs[realRVLocIdx].Val;
1130 switch (VA.getLocInfo()) {
1131 default: llvm_unreachable("Unknown loc info!");
1132 case CCValAssign::Full: break;
1133 case CCValAssign::BCvt:
1134 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1138 if (VA.needsCustom()) {
1139 if (VA.getLocVT() == MVT::v2f64) {
1140 // Extract the first half and return it in two registers.
1141 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1142 DAG.getConstant(0, MVT::i32));
1143 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1144 DAG.getVTList(MVT::i32, MVT::i32), Half);
1146 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1147 Flag = Chain.getValue(1);
1148 VA = RVLocs[++i]; // skip ahead to next loc
1149 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1150 HalfGPRs.getValue(1), Flag);
1151 Flag = Chain.getValue(1);
1152 VA = RVLocs[++i]; // skip ahead to next loc
1154 // Extract the 2nd half and fall through to handle it as an f64 value.
1155 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1156 DAG.getConstant(1, MVT::i32));
1158 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1160 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1161 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1162 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1163 Flag = Chain.getValue(1);
1164 VA = RVLocs[++i]; // skip ahead to next loc
1165 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1168 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1170 // Guarantee that all emitted copies are
1171 // stuck together, avoiding something bad.
1172 Flag = Chain.getValue(1);
1177 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1179 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1184 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1185 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1186 // one of the above mentioned nodes. It has to be wrapped because otherwise
1187 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1188 // be used to form addressing mode. These wrapped nodes will be selected
1190 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1191 EVT PtrVT = Op.getValueType();
1192 // FIXME there is no actual debug info here
1193 DebugLoc dl = Op.getDebugLoc();
1194 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1196 if (CP->isMachineConstantPoolEntry())
1197 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1198 CP->getAlignment());
1200 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1201 CP->getAlignment());
1202 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1205 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1207 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1208 SelectionDAG &DAG) {
1209 DebugLoc dl = GA->getDebugLoc();
1210 EVT PtrVT = getPointerTy();
1211 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1212 ARMConstantPoolValue *CPV =
1213 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1214 ARMCP::CPValue, PCAdj, "tlsgd", true);
1215 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1216 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1217 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1218 SDValue Chain = Argument.getValue(1);
1220 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1221 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1223 // call __tls_get_addr.
1226 Entry.Node = Argument;
1227 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1228 Args.push_back(Entry);
1229 // FIXME: is there useful debug info available here?
1230 std::pair<SDValue, SDValue> CallResult =
1231 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1232 false, false, false, false,
1233 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1234 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1235 return CallResult.first;
1238 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1239 // "local exec" model.
1241 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1242 SelectionDAG &DAG) {
1243 GlobalValue *GV = GA->getGlobal();
1244 DebugLoc dl = GA->getDebugLoc();
1246 SDValue Chain = DAG.getEntryNode();
1247 EVT PtrVT = getPointerTy();
1248 // Get the Thread Pointer
1249 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1251 if (GV->isDeclaration()) {
1252 // initial exec model
1253 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1254 ARMConstantPoolValue *CPV =
1255 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1256 ARMCP::CPValue, PCAdj, "gottpoff", true);
1257 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1258 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1259 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1260 Chain = Offset.getValue(1);
1262 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1263 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1265 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1268 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1269 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1270 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1271 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1274 // The address of the thread local variable is the add of the thread
1275 // pointer with the offset of the variable.
1276 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1280 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1281 // TODO: implement the "local dynamic" model
1282 assert(Subtarget->isTargetELF() &&
1283 "TLS not implemented for non-ELF targets");
1284 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1285 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1286 // otherwise use the "Local Exec" TLS Model
1287 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1288 return LowerToTLSGeneralDynamicModel(GA, DAG);
1290 return LowerToTLSExecModels(GA, DAG);
1293 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1294 SelectionDAG &DAG) {
1295 EVT PtrVT = getPointerTy();
1296 DebugLoc dl = Op.getDebugLoc();
1297 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1298 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1299 if (RelocM == Reloc::PIC_) {
1300 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1301 ARMConstantPoolValue *CPV =
1302 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1303 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1304 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1305 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1307 PseudoSourceValue::getConstantPool(), 0);
1308 SDValue Chain = Result.getValue(1);
1309 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1310 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1312 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1313 PseudoSourceValue::getGOT(), 0);
1316 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1317 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1318 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1319 PseudoSourceValue::getConstantPool(), 0);
1323 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1324 SelectionDAG &DAG) {
1325 EVT PtrVT = getPointerTy();
1326 DebugLoc dl = Op.getDebugLoc();
1327 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1328 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1330 if (RelocM == Reloc::Static)
1331 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1333 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1334 ARMConstantPoolValue *CPV =
1335 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1336 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1338 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1340 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1341 SDValue Chain = Result.getValue(1);
1343 if (RelocM == Reloc::PIC_) {
1344 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1345 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1348 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1349 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1354 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1356 assert(Subtarget->isTargetELF() &&
1357 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1358 EVT PtrVT = getPointerTy();
1359 DebugLoc dl = Op.getDebugLoc();
1360 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1361 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1362 "_GLOBAL_OFFSET_TABLE_",
1363 ARMPCLabelIndex, PCAdj);
1364 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1365 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1366 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1367 PseudoSourceValue::getConstantPool(), 0);
1368 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1369 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1372 static SDValue LowerNeonVLDLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
1374 // Change the lane number operand to be a TargetConstant; otherwise it
1375 // will be legalized into a register.
1376 SDNode *Node = Op.getNode();
1377 ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
1379 assert(false && "vld lane number must be a constant");
1382 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1383 Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
1384 return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
1387 static SDValue LowerNeonVSTLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
1389 // Change the lane number operand to be a TargetConstant; otherwise it
1390 // will be legalized into a register.
1391 SDNode *Node = Op.getNode();
1392 ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
1394 assert(false && "vst lane number must be a constant");
1397 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1398 Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
1399 return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
1403 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1404 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1406 case Intrinsic::arm_neon_vld2lane:
1407 return LowerNeonVLDLaneIntrinsic(Op, DAG, 2);
1408 case Intrinsic::arm_neon_vld3lane:
1409 return LowerNeonVLDLaneIntrinsic(Op, DAG, 3);
1410 case Intrinsic::arm_neon_vld4lane:
1411 return LowerNeonVLDLaneIntrinsic(Op, DAG, 4);
1412 case Intrinsic::arm_neon_vst2lane:
1413 return LowerNeonVSTLaneIntrinsic(Op, DAG, 2);
1414 case Intrinsic::arm_neon_vst3lane:
1415 return LowerNeonVSTLaneIntrinsic(Op, DAG, 3);
1416 case Intrinsic::arm_neon_vst4lane:
1417 return LowerNeonVSTLaneIntrinsic(Op, DAG, 4);
1418 default: return SDValue(); // Don't custom lower most intrinsics.
1423 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1424 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1425 DebugLoc dl = Op.getDebugLoc();
1427 default: return SDValue(); // Don't custom lower most intrinsics.
1428 case Intrinsic::arm_thread_pointer: {
1429 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1430 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1432 case Intrinsic::eh_sjlj_lsda: {
1433 MachineFunction &MF = DAG.getMachineFunction();
1434 EVT PtrVT = getPointerTy();
1435 DebugLoc dl = Op.getDebugLoc();
1436 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1438 unsigned PCAdj = (RelocM != Reloc::PIC_)
1439 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1440 ARMConstantPoolValue *CPV =
1441 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1442 ARMCP::CPLSDA, PCAdj);
1443 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1444 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1446 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1447 SDValue Chain = Result.getValue(1);
1449 if (RelocM == Reloc::PIC_) {
1450 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1451 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1455 case Intrinsic::eh_sjlj_setjmp:
1456 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1460 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1461 unsigned VarArgsFrameIndex) {
1462 // vastart just stores the address of the VarArgsFrameIndex slot into the
1463 // memory location argument.
1464 DebugLoc dl = Op.getDebugLoc();
1465 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1466 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1467 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1468 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1472 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1473 SDNode *Node = Op.getNode();
1474 DebugLoc dl = Node->getDebugLoc();
1475 EVT VT = Node->getValueType(0);
1476 SDValue Chain = Op.getOperand(0);
1477 SDValue Size = Op.getOperand(1);
1478 SDValue Align = Op.getOperand(2);
1480 // Chain the dynamic stack allocation so that it doesn't modify the stack
1481 // pointer when other instructions are using the stack.
1482 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1484 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1485 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1486 if (AlignVal > StackAlign)
1487 // Do this now since selection pass cannot introduce new target
1488 // independent node.
1489 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1491 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1492 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1493 // do even more horrible hack later.
1494 MachineFunction &MF = DAG.getMachineFunction();
1495 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1496 if (AFI->isThumb1OnlyFunction()) {
1498 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1500 uint32_t Val = C->getZExtValue();
1501 if (Val <= 508 && ((Val & 3) == 0))
1505 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1508 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1509 SDValue Ops1[] = { Chain, Size, Align };
1510 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1511 Chain = Res.getValue(1);
1512 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1513 DAG.getIntPtrConstant(0, true), SDValue());
1514 SDValue Ops2[] = { Res, Chain };
1515 return DAG.getMergeValues(Ops2, 2, dl);
1519 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1520 SDValue &Root, SelectionDAG &DAG,
1522 MachineFunction &MF = DAG.getMachineFunction();
1523 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1525 TargetRegisterClass *RC;
1526 if (AFI->isThumb1OnlyFunction())
1527 RC = ARM::tGPRRegisterClass;
1529 RC = ARM::GPRRegisterClass;
1531 // Transform the arguments stored in physical registers into virtual ones.
1532 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1533 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1536 if (NextVA.isMemLoc()) {
1537 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1538 MachineFrameInfo *MFI = MF.getFrameInfo();
1539 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1541 // Create load node to retrieve arguments from the stack.
1542 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1543 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1545 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1546 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1549 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1553 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1554 CallingConv::ID CallConv, bool isVarArg,
1555 const SmallVectorImpl<ISD::InputArg>
1557 DebugLoc dl, SelectionDAG &DAG,
1558 SmallVectorImpl<SDValue> &InVals) {
1560 MachineFunction &MF = DAG.getMachineFunction();
1561 MachineFrameInfo *MFI = MF.getFrameInfo();
1563 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1565 // Assign locations to all of the incoming arguments.
1566 SmallVector<CCValAssign, 16> ArgLocs;
1567 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1569 CCInfo.AnalyzeFormalArguments(Ins,
1570 CCAssignFnForNode(CallConv, /* Return*/ false,
1573 SmallVector<SDValue, 16> ArgValues;
1575 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1576 CCValAssign &VA = ArgLocs[i];
1578 // Arguments stored in registers.
1579 if (VA.isRegLoc()) {
1580 EVT RegVT = VA.getLocVT();
1583 if (VA.needsCustom()) {
1584 // f64 and vector types are split up into multiple registers or
1585 // combinations of registers and stack slots.
1588 if (VA.getLocVT() == MVT::v2f64) {
1589 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1591 VA = ArgLocs[++i]; // skip ahead to next loc
1592 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1594 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1595 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1596 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1597 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1598 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1600 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1603 TargetRegisterClass *RC;
1605 if (RegVT == MVT::f32)
1606 RC = ARM::SPRRegisterClass;
1607 else if (RegVT == MVT::f64)
1608 RC = ARM::DPRRegisterClass;
1609 else if (RegVT == MVT::v2f64)
1610 RC = ARM::QPRRegisterClass;
1611 else if (RegVT == MVT::i32)
1612 RC = (AFI->isThumb1OnlyFunction() ?
1613 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1615 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1617 // Transform the arguments in physical registers into virtual ones.
1618 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1619 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1622 // If this is an 8 or 16-bit value, it is really passed promoted
1623 // to 32 bits. Insert an assert[sz]ext to capture this, then
1624 // truncate to the right size.
1625 switch (VA.getLocInfo()) {
1626 default: llvm_unreachable("Unknown loc info!");
1627 case CCValAssign::Full: break;
1628 case CCValAssign::BCvt:
1629 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1631 case CCValAssign::SExt:
1632 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1633 DAG.getValueType(VA.getValVT()));
1634 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1636 case CCValAssign::ZExt:
1637 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1638 DAG.getValueType(VA.getValVT()));
1639 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1643 InVals.push_back(ArgValue);
1645 } else { // VA.isRegLoc()
1648 assert(VA.isMemLoc());
1649 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1651 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1652 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1654 // Create load nodes to retrieve arguments from the stack.
1655 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1656 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1662 static const unsigned GPRArgRegs[] = {
1663 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1666 unsigned NumGPRs = CCInfo.getFirstUnallocated
1667 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1669 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1670 unsigned VARegSize = (4 - NumGPRs) * 4;
1671 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1672 unsigned ArgOffset = 0;
1673 if (VARegSaveSize) {
1674 // If this function is vararg, store any remaining integer argument regs
1675 // to their spots on the stack so that they may be loaded by deferencing
1676 // the result of va_next.
1677 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1678 ArgOffset = CCInfo.getNextStackOffset();
1679 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1680 VARegSaveSize - VARegSize);
1681 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1683 SmallVector<SDValue, 4> MemOps;
1684 for (; NumGPRs < 4; ++NumGPRs) {
1685 TargetRegisterClass *RC;
1686 if (AFI->isThumb1OnlyFunction())
1687 RC = ARM::tGPRRegisterClass;
1689 RC = ARM::GPRRegisterClass;
1691 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1692 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1693 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1694 MemOps.push_back(Store);
1695 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1696 DAG.getConstant(4, getPointerTy()));
1698 if (!MemOps.empty())
1699 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1700 &MemOps[0], MemOps.size());
1702 // This will point to the next argument passed via stack.
1703 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1709 /// isFloatingPointZero - Return true if this is +0.0.
1710 static bool isFloatingPointZero(SDValue Op) {
1711 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1712 return CFP->getValueAPF().isPosZero();
1713 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1714 // Maybe this has already been legalized into the constant pool?
1715 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1716 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1717 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1718 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1719 return CFP->getValueAPF().isPosZero();
1725 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1726 return ( isThumb1Only && (C & ~255U) == 0) ||
1727 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1730 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1731 /// the given operands.
1732 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1733 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1735 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1736 unsigned C = RHSC->getZExtValue();
1737 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1738 // Constant does not fit, try adjusting it by one?
1743 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1744 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1745 RHS = DAG.getConstant(C-1, MVT::i32);
1750 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1751 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1752 RHS = DAG.getConstant(C-1, MVT::i32);
1757 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1758 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1759 RHS = DAG.getConstant(C+1, MVT::i32);
1764 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1765 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1766 RHS = DAG.getConstant(C+1, MVT::i32);
1773 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1774 ARMISD::NodeType CompareType;
1777 CompareType = ARMISD::CMP;
1782 CompareType = ARMISD::CMPZ;
1785 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1786 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1789 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1790 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1793 if (!isFloatingPointZero(RHS))
1794 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1796 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1797 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1800 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1801 const ARMSubtarget *ST) {
1802 EVT VT = Op.getValueType();
1803 SDValue LHS = Op.getOperand(0);
1804 SDValue RHS = Op.getOperand(1);
1805 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1806 SDValue TrueVal = Op.getOperand(2);
1807 SDValue FalseVal = Op.getOperand(3);
1808 DebugLoc dl = Op.getDebugLoc();
1810 if (LHS.getValueType() == MVT::i32) {
1812 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1813 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1814 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1817 ARMCC::CondCodes CondCode, CondCode2;
1818 FPCCToARMCC(CC, CondCode, CondCode2);
1820 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1821 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1822 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1823 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1825 if (CondCode2 != ARMCC::AL) {
1826 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1827 // FIXME: Needs another CMP because flag can have but one use.
1828 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1829 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1830 Result, TrueVal, ARMCC2, CCR, Cmp2);
1835 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1836 const ARMSubtarget *ST) {
1837 SDValue Chain = Op.getOperand(0);
1838 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1839 SDValue LHS = Op.getOperand(2);
1840 SDValue RHS = Op.getOperand(3);
1841 SDValue Dest = Op.getOperand(4);
1842 DebugLoc dl = Op.getDebugLoc();
1844 if (LHS.getValueType() == MVT::i32) {
1846 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1847 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1848 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1849 Chain, Dest, ARMCC, CCR,Cmp);
1852 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1853 ARMCC::CondCodes CondCode, CondCode2;
1854 FPCCToARMCC(CC, CondCode, CondCode2);
1856 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1857 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1858 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1859 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1860 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1861 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1862 if (CondCode2 != ARMCC::AL) {
1863 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1864 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1865 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1870 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1871 SDValue Chain = Op.getOperand(0);
1872 SDValue Table = Op.getOperand(1);
1873 SDValue Index = Op.getOperand(2);
1874 DebugLoc dl = Op.getDebugLoc();
1876 EVT PTy = getPointerTy();
1877 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1878 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1879 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1880 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1881 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1882 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1883 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1884 if (Subtarget->isThumb2()) {
1885 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1886 // which does another jump to the destination. This also makes it easier
1887 // to translate it to TBB / TBH later.
1888 // FIXME: This might not work if the function is extremely large.
1889 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1890 Addr, Op.getOperand(2), JTI, UId);
1892 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1893 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1894 Chain = Addr.getValue(1);
1895 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1896 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1898 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1899 Chain = Addr.getValue(1);
1900 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1904 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1905 DebugLoc dl = Op.getDebugLoc();
1907 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1908 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1909 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1912 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1913 EVT VT = Op.getValueType();
1914 DebugLoc dl = Op.getDebugLoc();
1916 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1918 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1919 return DAG.getNode(Opc, dl, VT, Op);
1922 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1923 // Implement fcopysign with a fabs and a conditional fneg.
1924 SDValue Tmp0 = Op.getOperand(0);
1925 SDValue Tmp1 = Op.getOperand(1);
1926 DebugLoc dl = Op.getDebugLoc();
1927 EVT VT = Op.getValueType();
1928 EVT SrcVT = Tmp1.getValueType();
1929 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1930 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1931 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1932 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1933 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1936 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1937 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1938 MFI->setFrameAddressIsTaken(true);
1939 EVT VT = Op.getValueType();
1940 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1941 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1942 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1943 ? ARM::R7 : ARM::R11;
1944 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1946 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1951 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1953 SDValue Dst, SDValue Src,
1954 SDValue Size, unsigned Align,
1956 const Value *DstSV, uint64_t DstSVOff,
1957 const Value *SrcSV, uint64_t SrcSVOff){
1958 // Do repeated 4-byte loads and stores. To be improved.
1959 // This requires 4-byte alignment.
1960 if ((Align & 3) != 0)
1962 // This requires the copy size to be a constant, preferrably
1963 // within a subtarget-specific limit.
1964 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1967 uint64_t SizeVal = ConstantSize->getZExtValue();
1968 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1971 unsigned BytesLeft = SizeVal & 3;
1972 unsigned NumMemOps = SizeVal >> 2;
1973 unsigned EmittedNumMemOps = 0;
1975 unsigned VTSize = 4;
1977 const unsigned MAX_LOADS_IN_LDM = 6;
1978 SDValue TFOps[MAX_LOADS_IN_LDM];
1979 SDValue Loads[MAX_LOADS_IN_LDM];
1980 uint64_t SrcOff = 0, DstOff = 0;
1982 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1983 // same number of stores. The loads and stores will get combined into
1984 // ldm/stm later on.
1985 while (EmittedNumMemOps < NumMemOps) {
1987 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1988 Loads[i] = DAG.getLoad(VT, dl, Chain,
1989 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1990 DAG.getConstant(SrcOff, MVT::i32)),
1991 SrcSV, SrcSVOff + SrcOff);
1992 TFOps[i] = Loads[i].getValue(1);
1995 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1998 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1999 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2000 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2001 DAG.getConstant(DstOff, MVT::i32)),
2002 DstSV, DstSVOff + DstOff);
2005 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2007 EmittedNumMemOps += i;
2013 // Issue loads / stores for the trailing (1 - 3) bytes.
2014 unsigned BytesLeftSave = BytesLeft;
2017 if (BytesLeft >= 2) {
2025 Loads[i] = DAG.getLoad(VT, dl, Chain,
2026 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2027 DAG.getConstant(SrcOff, MVT::i32)),
2028 SrcSV, SrcSVOff + SrcOff);
2029 TFOps[i] = Loads[i].getValue(1);
2032 BytesLeft -= VTSize;
2034 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2037 BytesLeft = BytesLeftSave;
2039 if (BytesLeft >= 2) {
2047 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2048 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2049 DAG.getConstant(DstOff, MVT::i32)),
2050 DstSV, DstSVOff + DstOff);
2053 BytesLeft -= VTSize;
2055 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2058 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2059 SDValue Op = N->getOperand(0);
2060 DebugLoc dl = N->getDebugLoc();
2061 if (N->getValueType(0) == MVT::f64) {
2062 // Turn i64->f64 into FMDRR.
2063 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2064 DAG.getConstant(0, MVT::i32));
2065 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2066 DAG.getConstant(1, MVT::i32));
2067 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2070 // Turn f64->i64 into FMRRD.
2071 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2072 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2074 // Merge the pieces into a single i64 value.
2075 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2078 /// getZeroVector - Returns a vector of specified type with all zero elements.
2080 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2081 assert(VT.isVector() && "Expected a vector type");
2083 // Zero vectors are used to represent vector negation and in those cases
2084 // will be implemented with the NEON VNEG instruction. However, VNEG does
2085 // not support i64 elements, so sometimes the zero vectors will need to be
2086 // explicitly constructed. For those cases, and potentially other uses in
2087 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2088 // to their dest type. This ensures they get CSE'd.
2090 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2091 SmallVector<SDValue, 8> Ops;
2094 if (VT.getSizeInBits() == 64) {
2095 Ops.assign(8, Cst); TVT = MVT::v8i8;
2097 Ops.assign(16, Cst); TVT = MVT::v16i8;
2099 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2101 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2104 /// getOnesVector - Returns a vector of specified type with all bits set.
2106 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2107 assert(VT.isVector() && "Expected a vector type");
2109 // Always build ones vectors as <16 x i32> or <8 x i32> bitcasted to their
2110 // dest type. This ensures they get CSE'd.
2112 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2113 SmallVector<SDValue, 8> Ops;
2116 if (VT.getSizeInBits() == 64) {
2117 Ops.assign(8, Cst); TVT = MVT::v8i8;
2119 Ops.assign(16, Cst); TVT = MVT::v16i8;
2121 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2123 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2126 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2127 const ARMSubtarget *ST) {
2128 EVT VT = N->getValueType(0);
2129 DebugLoc dl = N->getDebugLoc();
2131 // Lower vector shifts on NEON to use VSHL.
2132 if (VT.isVector()) {
2133 assert(ST->hasNEON() && "unexpected vector shift");
2135 // Left shifts translate directly to the vshiftu intrinsic.
2136 if (N->getOpcode() == ISD::SHL)
2137 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2138 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2139 N->getOperand(0), N->getOperand(1));
2141 assert((N->getOpcode() == ISD::SRA ||
2142 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2144 // NEON uses the same intrinsics for both left and right shifts. For
2145 // right shifts, the shift amounts are negative, so negate the vector of
2147 EVT ShiftVT = N->getOperand(1).getValueType();
2148 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2149 getZeroVector(ShiftVT, DAG, dl),
2151 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2152 Intrinsic::arm_neon_vshifts :
2153 Intrinsic::arm_neon_vshiftu);
2154 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2155 DAG.getConstant(vshiftInt, MVT::i32),
2156 N->getOperand(0), NegatedCount);
2159 // We can get here for a node like i32 = ISD::SHL i32, i64
2163 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2164 "Unknown shift to lower!");
2166 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2167 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2168 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2171 // If we are in thumb mode, we don't have RRX.
2172 if (ST->isThumb1Only()) return SDValue();
2174 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2175 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2176 DAG.getConstant(0, MVT::i32));
2177 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2178 DAG.getConstant(1, MVT::i32));
2180 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2181 // captures the result into a carry flag.
2182 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2183 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2185 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2186 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2188 // Merge the pieces into a single i64 value.
2189 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2192 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2193 SDValue TmpOp0, TmpOp1;
2194 bool Invert = false;
2198 SDValue Op0 = Op.getOperand(0);
2199 SDValue Op1 = Op.getOperand(1);
2200 SDValue CC = Op.getOperand(2);
2201 EVT VT = Op.getValueType();
2202 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2203 DebugLoc dl = Op.getDebugLoc();
2205 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2206 switch (SetCCOpcode) {
2207 default: llvm_unreachable("Illegal FP comparison"); break;
2209 case ISD::SETNE: Invert = true; // Fallthrough
2211 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2213 case ISD::SETLT: Swap = true; // Fallthrough
2215 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2217 case ISD::SETLE: Swap = true; // Fallthrough
2219 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2220 case ISD::SETUGE: Swap = true; // Fallthrough
2221 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2222 case ISD::SETUGT: Swap = true; // Fallthrough
2223 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2224 case ISD::SETUEQ: Invert = true; // Fallthrough
2226 // Expand this to (OLT | OGT).
2230 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2231 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2233 case ISD::SETUO: Invert = true; // Fallthrough
2235 // Expand this to (OLT | OGE).
2239 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2240 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2244 // Integer comparisons.
2245 switch (SetCCOpcode) {
2246 default: llvm_unreachable("Illegal integer comparison"); break;
2247 case ISD::SETNE: Invert = true;
2248 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2249 case ISD::SETLT: Swap = true;
2250 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2251 case ISD::SETLE: Swap = true;
2252 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2253 case ISD::SETULT: Swap = true;
2254 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2255 case ISD::SETULE: Swap = true;
2256 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2259 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2260 if (Opc == ARMISD::VCEQ) {
2263 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2265 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2268 // Ignore bitconvert.
2269 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2270 AndOp = AndOp.getOperand(0);
2272 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2274 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2275 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2282 std::swap(Op0, Op1);
2284 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2287 Result = DAG.getNOT(dl, Result, VT);
2292 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2293 /// VMOV instruction, and if so, return the constant being splatted.
2294 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2295 unsigned SplatBitSize, SelectionDAG &DAG) {
2296 switch (SplatBitSize) {
2298 // Any 1-byte value is OK.
2299 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2300 return DAG.getTargetConstant(SplatBits, MVT::i8);
2303 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2304 if ((SplatBits & ~0xff) == 0 ||
2305 (SplatBits & ~0xff00) == 0)
2306 return DAG.getTargetConstant(SplatBits, MVT::i16);
2310 // NEON's 32-bit VMOV supports splat values where:
2311 // * only one byte is nonzero, or
2312 // * the least significant byte is 0xff and the second byte is nonzero, or
2313 // * the least significant 2 bytes are 0xff and the third is nonzero.
2314 if ((SplatBits & ~0xff) == 0 ||
2315 (SplatBits & ~0xff00) == 0 ||
2316 (SplatBits & ~0xff0000) == 0 ||
2317 (SplatBits & ~0xff000000) == 0)
2318 return DAG.getTargetConstant(SplatBits, MVT::i32);
2320 if ((SplatBits & ~0xffff) == 0 &&
2321 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2322 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2324 if ((SplatBits & ~0xffffff) == 0 &&
2325 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2326 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2328 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2329 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2330 // VMOV.I32. A (very) minor optimization would be to replicate the value
2331 // and fall through here to test for a valid 64-bit splat. But, then the
2332 // caller would also need to check and handle the change in size.
2336 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2337 uint64_t BitMask = 0xff;
2339 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2340 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2342 else if ((SplatBits & BitMask) != 0)
2346 return DAG.getTargetConstant(Val, MVT::i64);
2350 llvm_unreachable("unexpected size for isVMOVSplat");
2357 /// getVMOVImm - If this is a build_vector of constants which can be
2358 /// formed by using a VMOV instruction of the specified element size,
2359 /// return the constant being splatted. The ByteSize field indicates the
2360 /// number of bytes of each element [1248].
2361 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2362 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2363 APInt SplatBits, SplatUndef;
2364 unsigned SplatBitSize;
2366 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2367 HasAnyUndefs, ByteSize * 8))
2370 if (SplatBitSize > ByteSize * 8)
2373 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2377 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2378 bool &ReverseVEXT, unsigned &Imm) {
2379 unsigned NumElts = VT.getVectorNumElements();
2380 ReverseVEXT = false;
2383 // If this is a VEXT shuffle, the immediate value is the index of the first
2384 // element. The other shuffle indices must be the successive elements after
2386 unsigned ExpectedElt = Imm;
2387 for (unsigned i = 1; i < NumElts; ++i) {
2388 // Increment the expected index. If it wraps around, it may still be
2389 // a VEXT but the source vectors must be swapped.
2391 if (ExpectedElt == NumElts * 2) {
2396 if (ExpectedElt != static_cast<unsigned>(M[i]))
2400 // Adjust the index value if the source operands will be swapped.
2407 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2408 /// instruction with the specified blocksize. (The order of the elements
2409 /// within each block of the vector is reversed.)
2410 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2411 unsigned BlockSize) {
2412 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2413 "Only possible block sizes for VREV are: 16, 32, 64");
2415 unsigned NumElts = VT.getVectorNumElements();
2416 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2417 unsigned BlockElts = M[0] + 1;
2419 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2422 for (unsigned i = 0; i < NumElts; ++i) {
2423 if ((unsigned) M[i] !=
2424 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2431 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2432 unsigned &WhichResult) {
2433 unsigned NumElts = VT.getVectorNumElements();
2434 WhichResult = (M[0] == 0 ? 0 : 1);
2435 for (unsigned i = 0; i < NumElts; i += 2) {
2436 if ((unsigned) M[i] != i + WhichResult ||
2437 (unsigned) M[i+1] != i + NumElts + WhichResult)
2443 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2444 unsigned &WhichResult) {
2445 unsigned NumElts = VT.getVectorNumElements();
2446 WhichResult = (M[0] == 0 ? 0 : 1);
2447 for (unsigned i = 0; i != NumElts; ++i) {
2448 if ((unsigned) M[i] != 2 * i + WhichResult)
2452 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2453 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2459 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2460 unsigned &WhichResult) {
2461 unsigned NumElts = VT.getVectorNumElements();
2462 WhichResult = (M[0] == 0 ? 0 : 1);
2463 unsigned Idx = WhichResult * NumElts / 2;
2464 for (unsigned i = 0; i != NumElts; i += 2) {
2465 if ((unsigned) M[i] != Idx ||
2466 (unsigned) M[i+1] != Idx + NumElts)
2471 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2472 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2478 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2479 // Canonicalize all-zeros and all-ones vectors.
2480 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2481 if (ConstVal->isNullValue())
2482 return getZeroVector(VT, DAG, dl);
2483 if (ConstVal->isAllOnesValue())
2484 return getOnesVector(VT, DAG, dl);
2487 if (VT.is64BitVector()) {
2488 switch (Val.getValueType().getSizeInBits()) {
2489 case 8: CanonicalVT = MVT::v8i8; break;
2490 case 16: CanonicalVT = MVT::v4i16; break;
2491 case 32: CanonicalVT = MVT::v2i32; break;
2492 case 64: CanonicalVT = MVT::v1i64; break;
2493 default: llvm_unreachable("unexpected splat element type"); break;
2496 assert(VT.is128BitVector() && "unknown splat vector size");
2497 switch (Val.getValueType().getSizeInBits()) {
2498 case 8: CanonicalVT = MVT::v16i8; break;
2499 case 16: CanonicalVT = MVT::v8i16; break;
2500 case 32: CanonicalVT = MVT::v4i32; break;
2501 case 64: CanonicalVT = MVT::v2i64; break;
2502 default: llvm_unreachable("unexpected splat element type"); break;
2506 // Build a canonical splat for this value.
2507 SmallVector<SDValue, 8> Ops;
2508 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2509 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2511 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2514 // If this is a case we can't handle, return null and let the default
2515 // expansion code take care of it.
2516 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2517 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2518 DebugLoc dl = Op.getDebugLoc();
2519 EVT VT = Op.getValueType();
2521 APInt SplatBits, SplatUndef;
2522 unsigned SplatBitSize;
2524 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2525 if (SplatBitSize <= 64) {
2526 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2527 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2529 return BuildSplat(Val, VT, DAG, dl);
2533 // If there are only 2 elements in a 128-bit vector, insert them into an
2534 // undef vector. This handles the common case for 128-bit vector argument
2535 // passing, where the insertions should be translated to subreg accesses
2536 // with no real instructions.
2537 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2538 SDValue Val = DAG.getUNDEF(VT);
2539 SDValue Op0 = Op.getOperand(0);
2540 SDValue Op1 = Op.getOperand(1);
2541 if (Op0.getOpcode() != ISD::UNDEF)
2542 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2543 DAG.getIntPtrConstant(0));
2544 if (Op1.getOpcode() != ISD::UNDEF)
2545 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2546 DAG.getIntPtrConstant(1));
2553 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2554 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2555 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2556 /// are assumed to be legal.
2558 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2560 if (VT.getVectorNumElements() == 4 &&
2561 (VT.is128BitVector() || VT.is64BitVector())) {
2562 unsigned PFIndexes[4];
2563 for (unsigned i = 0; i != 4; ++i) {
2567 PFIndexes[i] = M[i];
2570 // Compute the index in the perfect shuffle table.
2571 unsigned PFTableIndex =
2572 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2573 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2574 unsigned Cost = (PFEntry >> 30);
2581 unsigned Imm, WhichResult;
2583 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2584 isVREVMask(M, VT, 64) ||
2585 isVREVMask(M, VT, 32) ||
2586 isVREVMask(M, VT, 16) ||
2587 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2588 isVTRNMask(M, VT, WhichResult) ||
2589 isVUZPMask(M, VT, WhichResult) ||
2590 isVZIPMask(M, VT, WhichResult));
2593 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2594 /// the specified operations to build the shuffle.
2595 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2596 SDValue RHS, SelectionDAG &DAG,
2598 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2599 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2600 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2603 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2612 OP_VUZPL, // VUZP, left result
2613 OP_VUZPR, // VUZP, right result
2614 OP_VZIPL, // VZIP, left result
2615 OP_VZIPR, // VZIP, right result
2616 OP_VTRNL, // VTRN, left result
2617 OP_VTRNR // VTRN, right result
2620 if (OpNum == OP_COPY) {
2621 if (LHSID == (1*9+2)*9+3) return LHS;
2622 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2626 SDValue OpLHS, OpRHS;
2627 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2628 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2629 EVT VT = OpLHS.getValueType();
2632 default: llvm_unreachable("Unknown shuffle opcode!");
2634 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2639 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2640 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2644 return DAG.getNode(ARMISD::VEXT, dl, VT,
2646 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2649 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2650 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2653 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2654 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2657 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2658 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2662 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2663 SDValue V1 = Op.getOperand(0);
2664 SDValue V2 = Op.getOperand(1);
2665 DebugLoc dl = Op.getDebugLoc();
2666 EVT VT = Op.getValueType();
2667 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2668 SmallVector<int, 8> ShuffleMask;
2670 // Convert shuffles that are directly supported on NEON to target-specific
2671 // DAG nodes, instead of keeping them as shuffles and matching them again
2672 // during code selection. This is more efficient and avoids the possibility
2673 // of inconsistencies between legalization and selection.
2674 // FIXME: floating-point vectors should be canonicalized to integer vectors
2675 // of the same time so that they get CSEd properly.
2676 SVN->getMask(ShuffleMask);
2678 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2679 int Lane = SVN->getSplatIndex();
2680 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2681 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2683 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2684 DAG.getConstant(Lane, MVT::i32));
2689 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2692 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2693 DAG.getConstant(Imm, MVT::i32));
2696 if (isVREVMask(ShuffleMask, VT, 64))
2697 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2698 if (isVREVMask(ShuffleMask, VT, 32))
2699 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2700 if (isVREVMask(ShuffleMask, VT, 16))
2701 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2703 // Check for Neon shuffles that modify both input vectors in place.
2704 // If both results are used, i.e., if there are two shuffles with the same
2705 // source operands and with masks corresponding to both results of one of
2706 // these operations, DAG memoization will ensure that a single node is
2707 // used for both shuffles.
2708 unsigned WhichResult;
2709 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2710 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2711 V1, V2).getValue(WhichResult);
2712 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2713 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2714 V1, V2).getValue(WhichResult);
2715 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2716 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2717 V1, V2).getValue(WhichResult);
2719 // If the shuffle is not directly supported and it has 4 elements, use
2720 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2721 if (VT.getVectorNumElements() == 4 &&
2722 (VT.is128BitVector() || VT.is64BitVector())) {
2723 unsigned PFIndexes[4];
2724 for (unsigned i = 0; i != 4; ++i) {
2725 if (ShuffleMask[i] < 0)
2728 PFIndexes[i] = ShuffleMask[i];
2731 // Compute the index in the perfect shuffle table.
2732 unsigned PFTableIndex =
2733 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2735 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2736 unsigned Cost = (PFEntry >> 30);
2739 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2745 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2746 EVT VT = Op.getValueType();
2747 DebugLoc dl = Op.getDebugLoc();
2748 SDValue Vec = Op.getOperand(0);
2749 SDValue Lane = Op.getOperand(1);
2751 // FIXME: This is invalid for 8 and 16-bit elements - the information about
2752 // sign / zero extension is lost!
2753 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2754 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2756 if (VT.bitsLT(MVT::i32))
2757 Op = DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2758 else if (VT.bitsGT(MVT::i32))
2759 Op = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op);
2764 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2765 // The only time a CONCAT_VECTORS operation can have legal types is when
2766 // two 64-bit vectors are concatenated to a 128-bit vector.
2767 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2768 "unexpected CONCAT_VECTORS");
2769 DebugLoc dl = Op.getDebugLoc();
2770 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2771 SDValue Op0 = Op.getOperand(0);
2772 SDValue Op1 = Op.getOperand(1);
2773 if (Op0.getOpcode() != ISD::UNDEF)
2774 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2775 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2776 DAG.getIntPtrConstant(0));
2777 if (Op1.getOpcode() != ISD::UNDEF)
2778 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2779 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2780 DAG.getIntPtrConstant(1));
2781 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2784 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2785 switch (Op.getOpcode()) {
2786 default: llvm_unreachable("Don't know how to custom lower this!");
2787 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2788 case ISD::GlobalAddress:
2789 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2790 LowerGlobalAddressELF(Op, DAG);
2791 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2792 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2793 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2794 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2795 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2796 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2797 case ISD::SINT_TO_FP:
2798 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2799 case ISD::FP_TO_SINT:
2800 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2801 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2802 case ISD::RETURNADDR: break;
2803 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2804 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2805 case ISD::INTRINSIC_VOID:
2806 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2807 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2808 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2811 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2812 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2813 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2814 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2815 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2816 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2821 /// ReplaceNodeResults - Replace the results of node with an illegal result
2822 /// type with new values built out of custom code.
2823 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2824 SmallVectorImpl<SDValue>&Results,
2825 SelectionDAG &DAG) {
2826 switch (N->getOpcode()) {
2828 llvm_unreachable("Don't know how to custom expand this!");
2830 case ISD::BIT_CONVERT:
2831 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2835 SDValue Res = LowerShift(N, DAG, Subtarget);
2837 Results.push_back(Res);
2843 //===----------------------------------------------------------------------===//
2844 // ARM Scheduler Hooks
2845 //===----------------------------------------------------------------------===//
2848 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2849 MachineBasicBlock *BB,
2850 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
2851 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2852 DebugLoc dl = MI->getDebugLoc();
2853 switch (MI->getOpcode()) {
2855 llvm_unreachable("Unexpected instr type to insert");
2856 case ARM::tMOVCCr_pseudo: {
2857 // To "insert" a SELECT_CC instruction, we actually have to insert the
2858 // diamond control-flow pattern. The incoming instruction knows the
2859 // destination vreg to set, the condition code register to branch on, the
2860 // true/false values to select between, and a branch opcode to use.
2861 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2862 MachineFunction::iterator It = BB;
2868 // cmpTY ccX, r1, r2
2870 // fallthrough --> copy0MBB
2871 MachineBasicBlock *thisMBB = BB;
2872 MachineFunction *F = BB->getParent();
2873 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2874 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2875 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2876 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2877 F->insert(It, copy0MBB);
2878 F->insert(It, sinkMBB);
2879 // Update machine-CFG edges by first adding all successors of the current
2880 // block to the new block which will contain the Phi node for the select.
2881 // Also inform sdisel of the edge changes.
2882 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
2883 E = BB->succ_end(); I != E; ++I) {
2884 EM->insert(std::make_pair(*I, sinkMBB));
2885 sinkMBB->addSuccessor(*I);
2887 // Next, remove all successors of the current block, and add the true
2888 // and fallthrough blocks as its successors.
2889 while (!BB->succ_empty())
2890 BB->removeSuccessor(BB->succ_begin());
2891 BB->addSuccessor(copy0MBB);
2892 BB->addSuccessor(sinkMBB);
2895 // %FalseValue = ...
2896 // # fallthrough to sinkMBB
2899 // Update machine-CFG edges
2900 BB->addSuccessor(sinkMBB);
2903 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2906 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2907 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2908 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2910 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2917 case ARM::t2SUBrSPi_:
2918 case ARM::t2SUBrSPi12_:
2919 case ARM::t2SUBrSPs_: {
2920 MachineFunction *MF = BB->getParent();
2921 unsigned DstReg = MI->getOperand(0).getReg();
2922 unsigned SrcReg = MI->getOperand(1).getReg();
2923 bool DstIsDead = MI->getOperand(0).isDead();
2924 bool SrcIsKill = MI->getOperand(1).isKill();
2926 if (SrcReg != ARM::SP) {
2927 // Copy the source to SP from virtual register.
2928 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2929 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2930 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2931 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2932 .addReg(SrcReg, getKillRegState(SrcIsKill));
2936 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2937 switch (MI->getOpcode()) {
2939 llvm_unreachable("Unexpected pseudo instruction!");
2945 OpOpc = ARM::tADDspr;
2948 OpOpc = ARM::tSUBspi;
2950 case ARM::t2SUBrSPi_:
2951 OpOpc = ARM::t2SUBrSPi;
2952 NeedPred = true; NeedCC = true;
2954 case ARM::t2SUBrSPi12_:
2955 OpOpc = ARM::t2SUBrSPi12;
2958 case ARM::t2SUBrSPs_:
2959 OpOpc = ARM::t2SUBrSPs;
2960 NeedPred = true; NeedCC = true; NeedOp3 = true;
2963 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2964 if (OpOpc == ARM::tAND)
2965 AddDefaultT1CC(MIB);
2966 MIB.addReg(ARM::SP);
2967 MIB.addOperand(MI->getOperand(2));
2969 MIB.addOperand(MI->getOperand(3));
2971 AddDefaultPred(MIB);
2975 // Copy the result from SP to virtual register.
2976 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2977 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2978 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2979 BuildMI(BB, dl, TII->get(CopyOpc))
2980 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
2982 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2988 //===----------------------------------------------------------------------===//
2989 // ARM Optimization Hooks
2990 //===----------------------------------------------------------------------===//
2993 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2994 TargetLowering::DAGCombinerInfo &DCI) {
2995 SelectionDAG &DAG = DCI.DAG;
2996 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2997 EVT VT = N->getValueType(0);
2998 unsigned Opc = N->getOpcode();
2999 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3000 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3001 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3002 ISD::CondCode CC = ISD::SETCC_INVALID;
3005 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3007 SDValue CCOp = Slct.getOperand(0);
3008 if (CCOp.getOpcode() == ISD::SETCC)
3009 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3012 bool DoXform = false;
3014 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3017 if (LHS.getOpcode() == ISD::Constant &&
3018 cast<ConstantSDNode>(LHS)->isNullValue()) {
3020 } else if (CC != ISD::SETCC_INVALID &&
3021 RHS.getOpcode() == ISD::Constant &&
3022 cast<ConstantSDNode>(RHS)->isNullValue()) {
3023 std::swap(LHS, RHS);
3024 SDValue Op0 = Slct.getOperand(0);
3025 EVT OpVT = isSlctCC ? Op0.getValueType() :
3026 Op0.getOperand(0).getValueType();
3027 bool isInt = OpVT.isInteger();
3028 CC = ISD::getSetCCInverse(CC, isInt);
3030 if (!TLI.isCondCodeLegal(CC, OpVT))
3031 return SDValue(); // Inverse operator isn't legal.
3038 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3040 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3041 Slct.getOperand(0), Slct.getOperand(1), CC);
3042 SDValue CCOp = Slct.getOperand(0);
3044 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3045 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3046 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3047 CCOp, OtherOp, Result);
3052 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3053 static SDValue PerformADDCombine(SDNode *N,
3054 TargetLowering::DAGCombinerInfo &DCI) {
3055 // added by evan in r37685 with no testcase.
3056 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3058 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3059 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3060 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3061 if (Result.getNode()) return Result;
3063 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3064 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3065 if (Result.getNode()) return Result;
3071 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3072 static SDValue PerformSUBCombine(SDNode *N,
3073 TargetLowering::DAGCombinerInfo &DCI) {
3074 // added by evan in r37685 with no testcase.
3075 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3077 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3078 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3079 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3080 if (Result.getNode()) return Result;
3087 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3088 static SDValue PerformFMRRDCombine(SDNode *N,
3089 TargetLowering::DAGCombinerInfo &DCI) {
3090 // fmrrd(fmdrr x, y) -> x,y
3091 SDValue InDouble = N->getOperand(0);
3092 if (InDouble.getOpcode() == ARMISD::FMDRR)
3093 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3097 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3098 /// operand of a vector shift operation, where all the elements of the
3099 /// build_vector must have the same constant integer value.
3100 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3101 // Ignore bit_converts.
3102 while (Op.getOpcode() == ISD::BIT_CONVERT)
3103 Op = Op.getOperand(0);
3104 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3105 APInt SplatBits, SplatUndef;
3106 unsigned SplatBitSize;
3108 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3109 HasAnyUndefs, ElementBits) ||
3110 SplatBitSize > ElementBits)
3112 Cnt = SplatBits.getSExtValue();
3116 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3117 /// operand of a vector shift left operation. That value must be in the range:
3118 /// 0 <= Value < ElementBits for a left shift; or
3119 /// 0 <= Value <= ElementBits for a long left shift.
3120 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3121 assert(VT.isVector() && "vector shift count is not a vector type");
3122 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3123 if (! getVShiftImm(Op, ElementBits, Cnt))
3125 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3128 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3129 /// operand of a vector shift right operation. For a shift opcode, the value
3130 /// is positive, but for an intrinsic the value count must be negative. The
3131 /// absolute value must be in the range:
3132 /// 1 <= |Value| <= ElementBits for a right shift; or
3133 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3134 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3136 assert(VT.isVector() && "vector shift count is not a vector type");
3137 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3138 if (! getVShiftImm(Op, ElementBits, Cnt))
3142 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3145 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3146 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3147 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3150 // Don't do anything for most intrinsics.
3153 // Vector shifts: check for immediate versions and lower them.
3154 // Note: This is done during DAG combining instead of DAG legalizing because
3155 // the build_vectors for 64-bit vector element shift counts are generally
3156 // not legal, and it is hard to see their values after they get legalized to
3157 // loads from a constant pool.
3158 case Intrinsic::arm_neon_vshifts:
3159 case Intrinsic::arm_neon_vshiftu:
3160 case Intrinsic::arm_neon_vshiftls:
3161 case Intrinsic::arm_neon_vshiftlu:
3162 case Intrinsic::arm_neon_vshiftn:
3163 case Intrinsic::arm_neon_vrshifts:
3164 case Intrinsic::arm_neon_vrshiftu:
3165 case Intrinsic::arm_neon_vrshiftn:
3166 case Intrinsic::arm_neon_vqshifts:
3167 case Intrinsic::arm_neon_vqshiftu:
3168 case Intrinsic::arm_neon_vqshiftsu:
3169 case Intrinsic::arm_neon_vqshiftns:
3170 case Intrinsic::arm_neon_vqshiftnu:
3171 case Intrinsic::arm_neon_vqshiftnsu:
3172 case Intrinsic::arm_neon_vqrshiftns:
3173 case Intrinsic::arm_neon_vqrshiftnu:
3174 case Intrinsic::arm_neon_vqrshiftnsu: {
3175 EVT VT = N->getOperand(1).getValueType();
3177 unsigned VShiftOpc = 0;
3180 case Intrinsic::arm_neon_vshifts:
3181 case Intrinsic::arm_neon_vshiftu:
3182 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3183 VShiftOpc = ARMISD::VSHL;
3186 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3187 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3188 ARMISD::VSHRs : ARMISD::VSHRu);
3193 case Intrinsic::arm_neon_vshiftls:
3194 case Intrinsic::arm_neon_vshiftlu:
3195 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3197 llvm_unreachable("invalid shift count for vshll intrinsic");
3199 case Intrinsic::arm_neon_vrshifts:
3200 case Intrinsic::arm_neon_vrshiftu:
3201 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3205 case Intrinsic::arm_neon_vqshifts:
3206 case Intrinsic::arm_neon_vqshiftu:
3207 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3211 case Intrinsic::arm_neon_vqshiftsu:
3212 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3214 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3216 case Intrinsic::arm_neon_vshiftn:
3217 case Intrinsic::arm_neon_vrshiftn:
3218 case Intrinsic::arm_neon_vqshiftns:
3219 case Intrinsic::arm_neon_vqshiftnu:
3220 case Intrinsic::arm_neon_vqshiftnsu:
3221 case Intrinsic::arm_neon_vqrshiftns:
3222 case Intrinsic::arm_neon_vqrshiftnu:
3223 case Intrinsic::arm_neon_vqrshiftnsu:
3224 // Narrowing shifts require an immediate right shift.
3225 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3227 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3230 llvm_unreachable("unhandled vector shift");
3234 case Intrinsic::arm_neon_vshifts:
3235 case Intrinsic::arm_neon_vshiftu:
3236 // Opcode already set above.
3238 case Intrinsic::arm_neon_vshiftls:
3239 case Intrinsic::arm_neon_vshiftlu:
3240 if (Cnt == VT.getVectorElementType().getSizeInBits())
3241 VShiftOpc = ARMISD::VSHLLi;
3243 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3244 ARMISD::VSHLLs : ARMISD::VSHLLu);
3246 case Intrinsic::arm_neon_vshiftn:
3247 VShiftOpc = ARMISD::VSHRN; break;
3248 case Intrinsic::arm_neon_vrshifts:
3249 VShiftOpc = ARMISD::VRSHRs; break;
3250 case Intrinsic::arm_neon_vrshiftu:
3251 VShiftOpc = ARMISD::VRSHRu; break;
3252 case Intrinsic::arm_neon_vrshiftn:
3253 VShiftOpc = ARMISD::VRSHRN; break;
3254 case Intrinsic::arm_neon_vqshifts:
3255 VShiftOpc = ARMISD::VQSHLs; break;
3256 case Intrinsic::arm_neon_vqshiftu:
3257 VShiftOpc = ARMISD::VQSHLu; break;
3258 case Intrinsic::arm_neon_vqshiftsu:
3259 VShiftOpc = ARMISD::VQSHLsu; break;
3260 case Intrinsic::arm_neon_vqshiftns:
3261 VShiftOpc = ARMISD::VQSHRNs; break;
3262 case Intrinsic::arm_neon_vqshiftnu:
3263 VShiftOpc = ARMISD::VQSHRNu; break;
3264 case Intrinsic::arm_neon_vqshiftnsu:
3265 VShiftOpc = ARMISD::VQSHRNsu; break;
3266 case Intrinsic::arm_neon_vqrshiftns:
3267 VShiftOpc = ARMISD::VQRSHRNs; break;
3268 case Intrinsic::arm_neon_vqrshiftnu:
3269 VShiftOpc = ARMISD::VQRSHRNu; break;
3270 case Intrinsic::arm_neon_vqrshiftnsu:
3271 VShiftOpc = ARMISD::VQRSHRNsu; break;
3274 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3275 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3278 case Intrinsic::arm_neon_vshiftins: {
3279 EVT VT = N->getOperand(1).getValueType();
3281 unsigned VShiftOpc = 0;
3283 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3284 VShiftOpc = ARMISD::VSLI;
3285 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3286 VShiftOpc = ARMISD::VSRI;
3288 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3291 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3292 N->getOperand(1), N->getOperand(2),
3293 DAG.getConstant(Cnt, MVT::i32));
3296 case Intrinsic::arm_neon_vqrshifts:
3297 case Intrinsic::arm_neon_vqrshiftu:
3298 // No immediate versions of these to check for.
3305 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3306 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3307 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3308 /// vector element shift counts are generally not legal, and it is hard to see
3309 /// their values after they get legalized to loads from a constant pool.
3310 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3311 const ARMSubtarget *ST) {
3312 EVT VT = N->getValueType(0);
3314 // Nothing to be done for scalar shifts.
3315 if (! VT.isVector())
3318 assert(ST->hasNEON() && "unexpected vector shift");
3321 switch (N->getOpcode()) {
3322 default: llvm_unreachable("unexpected shift opcode");
3325 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3326 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3327 DAG.getConstant(Cnt, MVT::i32));
3332 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3333 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3334 ARMISD::VSHRs : ARMISD::VSHRu);
3335 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3336 DAG.getConstant(Cnt, MVT::i32));
3342 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3343 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3344 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3345 const ARMSubtarget *ST) {
3346 SDValue N0 = N->getOperand(0);
3348 // Check for sign- and zero-extensions of vector extract operations of 8-
3349 // and 16-bit vector elements. NEON supports these directly. They are
3350 // handled during DAG combining because type legalization will promote them
3351 // to 32-bit types and it is messy to recognize the operations after that.
3352 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3353 SDValue Vec = N0.getOperand(0);
3354 SDValue Lane = N0.getOperand(1);
3355 EVT VT = N->getValueType(0);
3356 EVT EltVT = N0.getValueType();
3357 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3359 if (VT == MVT::i32 &&
3360 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3361 TLI.isTypeLegal(Vec.getValueType())) {
3364 switch (N->getOpcode()) {
3365 default: llvm_unreachable("unexpected opcode");
3366 case ISD::SIGN_EXTEND:
3367 Opc = ARMISD::VGETLANEs;
3369 case ISD::ZERO_EXTEND:
3370 case ISD::ANY_EXTEND:
3371 Opc = ARMISD::VGETLANEu;
3374 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3381 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3382 DAGCombinerInfo &DCI) const {
3383 switch (N->getOpcode()) {
3385 case ISD::ADD: return PerformADDCombine(N, DCI);
3386 case ISD::SUB: return PerformSUBCombine(N, DCI);
3387 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3388 case ISD::INTRINSIC_WO_CHAIN:
3389 return PerformIntrinsicCombine(N, DCI.DAG);
3393 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3394 case ISD::SIGN_EXTEND:
3395 case ISD::ZERO_EXTEND:
3396 case ISD::ANY_EXTEND:
3397 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3402 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3403 if (!Subtarget->hasV6Ops())
3404 // Pre-v6 does not support unaligned mem access.
3406 else if (!Subtarget->hasV6Ops()) {
3407 // v6 may or may not support unaligned mem access.
3408 if (!Subtarget->isTargetDarwin())
3412 switch (VT.getSimpleVT().SimpleTy) {
3419 // FIXME: VLD1 etc with standard alignment is legal.
3423 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3428 switch (VT.getSimpleVT().SimpleTy) {
3429 default: return false;
3444 if ((V & (Scale - 1)) != 0)
3447 return V == (V & ((1LL << 5) - 1));
3450 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3451 const ARMSubtarget *Subtarget) {
3458 switch (VT.getSimpleVT().SimpleTy) {
3459 default: return false;
3464 // + imm12 or - imm8
3466 return V == (V & ((1LL << 8) - 1));
3467 return V == (V & ((1LL << 12) - 1));
3470 // Same as ARM mode. FIXME: NEON?
3471 if (!Subtarget->hasVFP2())
3476 return V == (V & ((1LL << 8) - 1));
3480 /// isLegalAddressImmediate - Return true if the integer value can be used
3481 /// as the offset of the target addressing mode for load / store of the
3483 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3484 const ARMSubtarget *Subtarget) {
3491 if (Subtarget->isThumb1Only())
3492 return isLegalT1AddressImmediate(V, VT);
3493 else if (Subtarget->isThumb2())
3494 return isLegalT2AddressImmediate(V, VT, Subtarget);
3499 switch (VT.getSimpleVT().SimpleTy) {
3500 default: return false;
3505 return V == (V & ((1LL << 12) - 1));
3508 return V == (V & ((1LL << 8) - 1));
3511 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3516 return V == (V & ((1LL << 8) - 1));
3520 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3522 int Scale = AM.Scale;
3526 switch (VT.getSimpleVT().SimpleTy) {
3527 default: return false;
3536 return Scale == 2 || Scale == 4 || Scale == 8;
3539 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3543 // Note, we allow "void" uses (basically, uses that aren't loads or
3544 // stores), because arm allows folding a scale into many arithmetic
3545 // operations. This should be made more precise and revisited later.
3547 // Allow r << imm, but the imm has to be a multiple of two.
3548 if (Scale & 1) return false;
3549 return isPowerOf2_32(Scale);
3553 /// isLegalAddressingMode - Return true if the addressing mode represented
3554 /// by AM is legal for this target, for a load/store of the specified type.
3555 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3556 const Type *Ty) const {
3557 EVT VT = getValueType(Ty, true);
3558 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3561 // Can never fold addr of global into load/store.
3566 case 0: // no scale reg, must be "r+i" or "r", or "i".
3569 if (Subtarget->isThumb1Only())
3573 // ARM doesn't support any R+R*scale+imm addr modes.
3580 if (Subtarget->isThumb2())
3581 return isLegalT2ScaledAddressingMode(AM, VT);
3583 int Scale = AM.Scale;
3584 switch (VT.getSimpleVT().SimpleTy) {
3585 default: return false;
3589 if (Scale < 0) Scale = -Scale;
3593 return isPowerOf2_32(Scale & ~1);
3597 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3602 // Note, we allow "void" uses (basically, uses that aren't loads or
3603 // stores), because arm allows folding a scale into many arithmetic
3604 // operations. This should be made more precise and revisited later.
3606 // Allow r << imm, but the imm has to be a multiple of two.
3607 if (Scale & 1) return false;
3608 return isPowerOf2_32(Scale);
3615 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3616 bool isSEXTLoad, SDValue &Base,
3617 SDValue &Offset, bool &isInc,
3618 SelectionDAG &DAG) {
3619 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3622 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3624 Base = Ptr->getOperand(0);
3625 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3626 int RHSC = (int)RHS->getZExtValue();
3627 if (RHSC < 0 && RHSC > -256) {
3628 assert(Ptr->getOpcode() == ISD::ADD);
3630 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3634 isInc = (Ptr->getOpcode() == ISD::ADD);
3635 Offset = Ptr->getOperand(1);
3637 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3639 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3640 int RHSC = (int)RHS->getZExtValue();
3641 if (RHSC < 0 && RHSC > -0x1000) {
3642 assert(Ptr->getOpcode() == ISD::ADD);
3644 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3645 Base = Ptr->getOperand(0);
3650 if (Ptr->getOpcode() == ISD::ADD) {
3652 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3653 if (ShOpcVal != ARM_AM::no_shift) {
3654 Base = Ptr->getOperand(1);
3655 Offset = Ptr->getOperand(0);
3657 Base = Ptr->getOperand(0);
3658 Offset = Ptr->getOperand(1);
3663 isInc = (Ptr->getOpcode() == ISD::ADD);
3664 Base = Ptr->getOperand(0);
3665 Offset = Ptr->getOperand(1);
3669 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3673 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3674 bool isSEXTLoad, SDValue &Base,
3675 SDValue &Offset, bool &isInc,
3676 SelectionDAG &DAG) {
3677 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3680 Base = Ptr->getOperand(0);
3681 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3682 int RHSC = (int)RHS->getZExtValue();
3683 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3684 assert(Ptr->getOpcode() == ISD::ADD);
3686 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3688 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3689 isInc = Ptr->getOpcode() == ISD::ADD;
3690 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3698 /// getPreIndexedAddressParts - returns true by value, base pointer and
3699 /// offset pointer and addressing mode by reference if the node's address
3700 /// can be legally represented as pre-indexed load / store address.
3702 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3704 ISD::MemIndexedMode &AM,
3705 SelectionDAG &DAG) const {
3706 if (Subtarget->isThumb1Only())
3711 bool isSEXTLoad = false;
3712 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3713 Ptr = LD->getBasePtr();
3714 VT = LD->getMemoryVT();
3715 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3716 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3717 Ptr = ST->getBasePtr();
3718 VT = ST->getMemoryVT();
3723 bool isLegal = false;
3724 if (Subtarget->isThumb2())
3725 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3726 Offset, isInc, DAG);
3728 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3729 Offset, isInc, DAG);
3733 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3737 /// getPostIndexedAddressParts - returns true by value, base pointer and
3738 /// offset pointer and addressing mode by reference if this node can be
3739 /// combined with a load / store to form a post-indexed load / store.
3740 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3743 ISD::MemIndexedMode &AM,
3744 SelectionDAG &DAG) const {
3745 if (Subtarget->isThumb1Only())
3750 bool isSEXTLoad = false;
3751 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3752 VT = LD->getMemoryVT();
3753 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3754 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3755 VT = ST->getMemoryVT();
3760 bool isLegal = false;
3761 if (Subtarget->isThumb2())
3762 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3765 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3770 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3774 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3778 const SelectionDAG &DAG,
3779 unsigned Depth) const {
3780 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3781 switch (Op.getOpcode()) {
3783 case ARMISD::CMOV: {
3784 // Bits are known zero/one if known on the LHS and RHS.
3785 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3786 if (KnownZero == 0 && KnownOne == 0) return;
3788 APInt KnownZeroRHS, KnownOneRHS;
3789 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3790 KnownZeroRHS, KnownOneRHS, Depth+1);
3791 KnownZero &= KnownZeroRHS;
3792 KnownOne &= KnownOneRHS;
3798 //===----------------------------------------------------------------------===//
3799 // ARM Inline Assembly Support
3800 //===----------------------------------------------------------------------===//
3802 /// getConstraintType - Given a constraint letter, return the type of
3803 /// constraint it is for this target.
3804 ARMTargetLowering::ConstraintType
3805 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3806 if (Constraint.size() == 1) {
3807 switch (Constraint[0]) {
3809 case 'l': return C_RegisterClass;
3810 case 'w': return C_RegisterClass;
3813 return TargetLowering::getConstraintType(Constraint);
3816 std::pair<unsigned, const TargetRegisterClass*>
3817 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3819 if (Constraint.size() == 1) {
3820 // GCC RS6000 Constraint Letters
3821 switch (Constraint[0]) {
3823 if (Subtarget->isThumb1Only())
3824 return std::make_pair(0U, ARM::tGPRRegisterClass);
3826 return std::make_pair(0U, ARM::GPRRegisterClass);
3828 return std::make_pair(0U, ARM::GPRRegisterClass);
3831 return std::make_pair(0U, ARM::SPRRegisterClass);
3833 return std::make_pair(0U, ARM::DPRRegisterClass);
3837 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3840 std::vector<unsigned> ARMTargetLowering::
3841 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3843 if (Constraint.size() != 1)
3844 return std::vector<unsigned>();
3846 switch (Constraint[0]) { // GCC ARM Constraint Letters
3849 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3850 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3853 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3854 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3855 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3856 ARM::R12, ARM::LR, 0);
3859 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3860 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3861 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3862 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3863 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3864 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3865 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3866 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3868 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3869 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3870 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3871 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3875 return std::vector<unsigned>();
3878 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3879 /// vector. If it is invalid, don't add anything to Ops.
3880 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3883 std::vector<SDValue>&Ops,
3884 SelectionDAG &DAG) const {
3885 SDValue Result(0, 0);
3887 switch (Constraint) {
3889 case 'I': case 'J': case 'K': case 'L':
3890 case 'M': case 'N': case 'O':
3891 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3895 int64_t CVal64 = C->getSExtValue();
3896 int CVal = (int) CVal64;
3897 // None of these constraints allow values larger than 32 bits. Check
3898 // that the value fits in an int.
3902 switch (Constraint) {
3904 if (Subtarget->isThumb1Only()) {
3905 // This must be a constant between 0 and 255, for ADD
3907 if (CVal >= 0 && CVal <= 255)
3909 } else if (Subtarget->isThumb2()) {
3910 // A constant that can be used as an immediate value in a
3911 // data-processing instruction.
3912 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3915 // A constant that can be used as an immediate value in a
3916 // data-processing instruction.
3917 if (ARM_AM::getSOImmVal(CVal) != -1)
3923 if (Subtarget->isThumb()) { // FIXME thumb2
3924 // This must be a constant between -255 and -1, for negated ADD
3925 // immediates. This can be used in GCC with an "n" modifier that
3926 // prints the negated value, for use with SUB instructions. It is
3927 // not useful otherwise but is implemented for compatibility.
3928 if (CVal >= -255 && CVal <= -1)
3931 // This must be a constant between -4095 and 4095. It is not clear
3932 // what this constraint is intended for. Implemented for
3933 // compatibility with GCC.
3934 if (CVal >= -4095 && CVal <= 4095)
3940 if (Subtarget->isThumb1Only()) {
3941 // A 32-bit value where only one byte has a nonzero value. Exclude
3942 // zero to match GCC. This constraint is used by GCC internally for
3943 // constants that can be loaded with a move/shift combination.
3944 // It is not useful otherwise but is implemented for compatibility.
3945 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3947 } else if (Subtarget->isThumb2()) {
3948 // A constant whose bitwise inverse can be used as an immediate
3949 // value in a data-processing instruction. This can be used in GCC
3950 // with a "B" modifier that prints the inverted value, for use with
3951 // BIC and MVN instructions. It is not useful otherwise but is
3952 // implemented for compatibility.
3953 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3956 // A constant whose bitwise inverse can be used as an immediate
3957 // value in a data-processing instruction. This can be used in GCC
3958 // with a "B" modifier that prints the inverted value, for use with
3959 // BIC and MVN instructions. It is not useful otherwise but is
3960 // implemented for compatibility.
3961 if (ARM_AM::getSOImmVal(~CVal) != -1)
3967 if (Subtarget->isThumb1Only()) {
3968 // This must be a constant between -7 and 7,
3969 // for 3-operand ADD/SUB immediate instructions.
3970 if (CVal >= -7 && CVal < 7)
3972 } else if (Subtarget->isThumb2()) {
3973 // A constant whose negation can be used as an immediate value in a
3974 // data-processing instruction. This can be used in GCC with an "n"
3975 // modifier that prints the negated value, for use with SUB
3976 // instructions. It is not useful otherwise but is implemented for
3978 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3981 // A constant whose negation can be used as an immediate value in a
3982 // data-processing instruction. This can be used in GCC with an "n"
3983 // modifier that prints the negated value, for use with SUB
3984 // instructions. It is not useful otherwise but is implemented for
3986 if (ARM_AM::getSOImmVal(-CVal) != -1)
3992 if (Subtarget->isThumb()) { // FIXME thumb2
3993 // This must be a multiple of 4 between 0 and 1020, for
3994 // ADD sp + immediate.
3995 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3998 // A power of two or a constant between 0 and 32. This is used in
3999 // GCC for the shift amount on shifted register operands, but it is
4000 // useful in general for any shift amounts.
4001 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4007 if (Subtarget->isThumb()) { // FIXME thumb2
4008 // This must be a constant between 0 and 31, for shift amounts.
4009 if (CVal >= 0 && CVal <= 31)
4015 if (Subtarget->isThumb()) { // FIXME thumb2
4016 // This must be a multiple of 4 between -508 and 508, for
4017 // ADD/SUB sp = sp + immediate.
4018 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4023 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4027 if (Result.getNode()) {
4028 Ops.push_back(Result);
4031 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4036 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4037 // The ARM target isn't yet aware of offsets.