1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/Instruction.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Type.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/MC/MCSectionMachO.h"
41 #include "llvm/Target/TargetOptions.h"
42 #include "llvm/ADT/VectorExtras.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
49 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
53 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
54 CCValAssign::LocInfo &LocInfo,
55 ISD::ArgFlagsTy &ArgFlags,
57 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
58 CCValAssign::LocInfo &LocInfo,
59 ISD::ArgFlagsTy &ArgFlags,
61 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
62 CCValAssign::LocInfo &LocInfo,
63 ISD::ArgFlagsTy &ArgFlags,
66 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
67 EVT PromotedBitwiseVT) {
68 if (VT != PromotedLdStVT) {
69 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
70 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
71 PromotedLdStVT.getSimpleVT());
73 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
74 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
75 PromotedLdStVT.getSimpleVT());
78 EVT ElemTy = VT.getVectorElementType();
79 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
80 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
81 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
82 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
83 if (ElemTy != MVT::i32) {
84 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
85 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
86 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
87 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
89 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
90 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
91 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
92 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
94 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
95 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
99 // Promote all bit-wise operations.
100 if (VT.isInteger() && VT != PromotedBitwiseVT) {
101 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
102 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
103 PromotedBitwiseVT.getSimpleVT());
104 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
105 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
106 PromotedBitwiseVT.getSimpleVT());
107 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
108 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
109 PromotedBitwiseVT.getSimpleVT());
112 // Neon does not support vector divide/remainder operations.
113 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
121 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
122 addRegisterClass(VT, ARM::DPRRegisterClass);
123 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
126 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
127 addRegisterClass(VT, ARM::QPRRegisterClass);
128 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
131 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
132 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
133 return new TargetLoweringObjectFileMachO();
135 return new ARMElfTargetObjectFile();
138 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
139 : TargetLowering(TM, createTLOF(TM)) {
140 Subtarget = &TM.getSubtarget<ARMSubtarget>();
142 if (Subtarget->isTargetDarwin()) {
143 // Uses VFP for Thumb libfuncs if available.
144 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
145 // Single-precision floating-point arithmetic.
146 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
147 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
148 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
149 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
151 // Double-precision floating-point arithmetic.
152 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
153 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
154 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
155 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
157 // Single-precision comparisons.
158 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
159 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
160 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
161 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
162 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
163 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
164 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
165 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
167 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
172 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
173 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
174 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
176 // Double-precision comparisons.
177 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
178 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
179 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
180 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
181 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
182 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
183 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
184 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
186 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
195 // Floating-point to integer conversions.
196 // i64 conversions are done via library routines even when generating VFP
197 // instructions, so use the same ones.
198 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
199 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
200 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
201 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
203 // Conversions between floating types.
204 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
205 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
207 // Integer to floating-point conversions.
208 // i64 conversions are done via library routines even when generating VFP
209 // instructions, so use the same ones.
210 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
211 // e.g., __floatunsidf vs. __floatunssidfvfp.
212 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
213 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
214 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
215 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
219 // These libcalls are not available in 32-bit.
220 setLibcallName(RTLIB::SHL_I128, 0);
221 setLibcallName(RTLIB::SRL_I128, 0);
222 setLibcallName(RTLIB::SRA_I128, 0);
224 // Libcalls should use the AAPCS base standard ABI, even if hard float
225 // is in effect, as per the ARM RTABI specification, section 4.1.2.
226 if (Subtarget->isAAPCS_ABI()) {
227 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
228 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
229 CallingConv::ARM_AAPCS);
233 if (Subtarget->isThumb1Only())
234 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
236 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
237 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
238 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
239 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
241 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
244 if (Subtarget->hasNEON()) {
245 addDRTypeForNEON(MVT::v2f32);
246 addDRTypeForNEON(MVT::v8i8);
247 addDRTypeForNEON(MVT::v4i16);
248 addDRTypeForNEON(MVT::v2i32);
249 addDRTypeForNEON(MVT::v1i64);
251 addQRTypeForNEON(MVT::v4f32);
252 addQRTypeForNEON(MVT::v2f64);
253 addQRTypeForNEON(MVT::v16i8);
254 addQRTypeForNEON(MVT::v8i16);
255 addQRTypeForNEON(MVT::v4i32);
256 addQRTypeForNEON(MVT::v2i64);
258 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
259 // neither Neon nor VFP support any arithmetic operations on it.
260 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
261 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
262 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
263 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
264 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
265 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
266 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
267 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
268 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
269 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
270 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
271 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
272 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
273 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
274 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
275 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
276 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
277 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
278 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
279 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
280 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
281 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
282 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
283 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
285 // Neon does not support some operations on v1i64 and v2i64 types.
286 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
287 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
288 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
289 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
291 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
292 setTargetDAGCombine(ISD::SHL);
293 setTargetDAGCombine(ISD::SRL);
294 setTargetDAGCombine(ISD::SRA);
295 setTargetDAGCombine(ISD::SIGN_EXTEND);
296 setTargetDAGCombine(ISD::ZERO_EXTEND);
297 setTargetDAGCombine(ISD::ANY_EXTEND);
298 setTargetDAGCombine(ISD::SELECT_CC);
301 computeRegisterProperties();
303 // ARM does not have f32 extending load.
304 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
306 // ARM does not have i1 sign extending load.
307 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
309 // ARM supports all 4 flavors of integer indexed load / store.
310 if (!Subtarget->isThumb1Only()) {
311 for (unsigned im = (unsigned)ISD::PRE_INC;
312 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
313 setIndexedLoadAction(im, MVT::i1, Legal);
314 setIndexedLoadAction(im, MVT::i8, Legal);
315 setIndexedLoadAction(im, MVT::i16, Legal);
316 setIndexedLoadAction(im, MVT::i32, Legal);
317 setIndexedStoreAction(im, MVT::i1, Legal);
318 setIndexedStoreAction(im, MVT::i8, Legal);
319 setIndexedStoreAction(im, MVT::i16, Legal);
320 setIndexedStoreAction(im, MVT::i32, Legal);
324 // i64 operation support.
325 if (Subtarget->isThumb1Only()) {
326 setOperationAction(ISD::MUL, MVT::i64, Expand);
327 setOperationAction(ISD::MULHU, MVT::i32, Expand);
328 setOperationAction(ISD::MULHS, MVT::i32, Expand);
329 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
330 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
332 setOperationAction(ISD::MUL, MVT::i64, Expand);
333 setOperationAction(ISD::MULHU, MVT::i32, Expand);
334 if (!Subtarget->hasV6Ops())
335 setOperationAction(ISD::MULHS, MVT::i32, Expand);
337 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
338 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
339 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
340 setOperationAction(ISD::SRL, MVT::i64, Custom);
341 setOperationAction(ISD::SRA, MVT::i64, Custom);
343 // ARM does not have ROTL.
344 setOperationAction(ISD::ROTL, MVT::i32, Expand);
345 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
346 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
347 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
348 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
350 // Only ARMv6 has BSWAP.
351 if (!Subtarget->hasV6Ops())
352 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
354 // These are expanded into libcalls.
355 setOperationAction(ISD::SDIV, MVT::i32, Expand);
356 setOperationAction(ISD::UDIV, MVT::i32, Expand);
357 setOperationAction(ISD::SREM, MVT::i32, Expand);
358 setOperationAction(ISD::UREM, MVT::i32, Expand);
359 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
360 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
362 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
363 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
364 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
365 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
366 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
368 // Use the default implementation.
369 setOperationAction(ISD::VASTART, MVT::Other, Custom);
370 setOperationAction(ISD::VAARG, MVT::Other, Expand);
371 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
372 setOperationAction(ISD::VAEND, MVT::Other, Expand);
373 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
374 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
375 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
376 // FIXME: Shouldn't need this, since no register is used, but the legalizer
377 // doesn't yet know how to not do that for SjLj.
378 setExceptionSelectorRegister(ARM::R0);
379 if (Subtarget->isThumb())
380 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
382 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
383 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
385 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
386 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
387 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
389 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
391 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
392 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
393 // iff target supports vfp2.
394 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
396 // We want to custom lower some of our intrinsics.
397 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
399 setOperationAction(ISD::SETCC, MVT::i32, Expand);
400 setOperationAction(ISD::SETCC, MVT::f32, Expand);
401 setOperationAction(ISD::SETCC, MVT::f64, Expand);
402 setOperationAction(ISD::SELECT, MVT::i32, Expand);
403 setOperationAction(ISD::SELECT, MVT::f32, Expand);
404 setOperationAction(ISD::SELECT, MVT::f64, Expand);
405 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
406 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
407 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
409 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
410 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
411 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
412 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
413 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
415 // We don't support sin/cos/fmod/copysign/pow
416 setOperationAction(ISD::FSIN, MVT::f64, Expand);
417 setOperationAction(ISD::FSIN, MVT::f32, Expand);
418 setOperationAction(ISD::FCOS, MVT::f32, Expand);
419 setOperationAction(ISD::FCOS, MVT::f64, Expand);
420 setOperationAction(ISD::FREM, MVT::f64, Expand);
421 setOperationAction(ISD::FREM, MVT::f32, Expand);
422 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
423 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
424 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
426 setOperationAction(ISD::FPOW, MVT::f64, Expand);
427 setOperationAction(ISD::FPOW, MVT::f32, Expand);
429 // Various VFP goodness
430 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
431 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
432 if (Subtarget->hasVFP2()) {
433 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
434 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
435 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
436 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
438 // Special handling for half-precision FP.
439 if (!Subtarget->hasFP16()) {
440 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
441 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
445 // We have target-specific dag combine patterns for the following nodes:
446 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
447 setTargetDAGCombine(ISD::ADD);
448 setTargetDAGCombine(ISD::SUB);
450 setStackPointerRegisterToSaveRestore(ARM::SP);
451 setSchedulingPreference(SchedulingForRegPressure);
453 // FIXME: If-converter should use instruction latency to determine
454 // profitability rather than relying on fixed limits.
455 if (Subtarget->getCPUString() == "generic") {
456 // Generic (and overly aggressive) if-conversion limits.
457 setIfCvtBlockSizeLimit(10);
458 setIfCvtDupBlockSizeLimit(2);
459 } else if (Subtarget->hasV7Ops()) {
460 setIfCvtBlockSizeLimit(3);
461 setIfCvtDupBlockSizeLimit(1);
462 } else if (Subtarget->hasV6Ops()) {
463 setIfCvtBlockSizeLimit(2);
464 setIfCvtDupBlockSizeLimit(1);
466 setIfCvtBlockSizeLimit(3);
467 setIfCvtDupBlockSizeLimit(2);
470 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
471 // Do not enable CodePlacementOpt for now: it currently runs after the
472 // ARMConstantIslandPass and messes up branch relaxation and placement
473 // of constant islands.
474 // benefitFromCodePlacementOpt = true;
477 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
480 case ARMISD::Wrapper: return "ARMISD::Wrapper";
481 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
482 case ARMISD::CALL: return "ARMISD::CALL";
483 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
484 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
485 case ARMISD::tCALL: return "ARMISD::tCALL";
486 case ARMISD::BRCOND: return "ARMISD::BRCOND";
487 case ARMISD::BR_JT: return "ARMISD::BR_JT";
488 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
489 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
490 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
491 case ARMISD::CMP: return "ARMISD::CMP";
492 case ARMISD::CMPZ: return "ARMISD::CMPZ";
493 case ARMISD::CMPFP: return "ARMISD::CMPFP";
494 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
495 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
496 case ARMISD::CMOV: return "ARMISD::CMOV";
497 case ARMISD::CNEG: return "ARMISD::CNEG";
499 case ARMISD::RBIT: return "ARMISD::RBIT";
501 case ARMISD::FTOSI: return "ARMISD::FTOSI";
502 case ARMISD::FTOUI: return "ARMISD::FTOUI";
503 case ARMISD::SITOF: return "ARMISD::SITOF";
504 case ARMISD::UITOF: return "ARMISD::UITOF";
506 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
507 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
508 case ARMISD::RRX: return "ARMISD::RRX";
510 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
511 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
513 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
514 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
516 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
518 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
520 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
521 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
523 case ARMISD::VCEQ: return "ARMISD::VCEQ";
524 case ARMISD::VCGE: return "ARMISD::VCGE";
525 case ARMISD::VCGEU: return "ARMISD::VCGEU";
526 case ARMISD::VCGT: return "ARMISD::VCGT";
527 case ARMISD::VCGTU: return "ARMISD::VCGTU";
528 case ARMISD::VTST: return "ARMISD::VTST";
530 case ARMISD::VSHL: return "ARMISD::VSHL";
531 case ARMISD::VSHRs: return "ARMISD::VSHRs";
532 case ARMISD::VSHRu: return "ARMISD::VSHRu";
533 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
534 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
535 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
536 case ARMISD::VSHRN: return "ARMISD::VSHRN";
537 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
538 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
539 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
540 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
541 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
542 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
543 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
544 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
545 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
546 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
547 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
548 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
549 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
550 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
551 case ARMISD::VDUP: return "ARMISD::VDUP";
552 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
553 case ARMISD::VEXT: return "ARMISD::VEXT";
554 case ARMISD::VREV64: return "ARMISD::VREV64";
555 case ARMISD::VREV32: return "ARMISD::VREV32";
556 case ARMISD::VREV16: return "ARMISD::VREV16";
557 case ARMISD::VZIP: return "ARMISD::VZIP";
558 case ARMISD::VUZP: return "ARMISD::VUZP";
559 case ARMISD::VTRN: return "ARMISD::VTRN";
560 case ARMISD::FMAX: return "ARMISD::FMAX";
561 case ARMISD::FMIN: return "ARMISD::FMIN";
565 /// getFunctionAlignment - Return the Log2 alignment of this function.
566 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
567 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
570 //===----------------------------------------------------------------------===//
572 //===----------------------------------------------------------------------===//
574 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
575 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
577 default: llvm_unreachable("Unknown condition code!");
578 case ISD::SETNE: return ARMCC::NE;
579 case ISD::SETEQ: return ARMCC::EQ;
580 case ISD::SETGT: return ARMCC::GT;
581 case ISD::SETGE: return ARMCC::GE;
582 case ISD::SETLT: return ARMCC::LT;
583 case ISD::SETLE: return ARMCC::LE;
584 case ISD::SETUGT: return ARMCC::HI;
585 case ISD::SETUGE: return ARMCC::HS;
586 case ISD::SETULT: return ARMCC::LO;
587 case ISD::SETULE: return ARMCC::LS;
591 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
592 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
593 ARMCC::CondCodes &CondCode2) {
594 CondCode2 = ARMCC::AL;
596 default: llvm_unreachable("Unknown FP condition!");
598 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
600 case ISD::SETOGT: CondCode = ARMCC::GT; break;
602 case ISD::SETOGE: CondCode = ARMCC::GE; break;
603 case ISD::SETOLT: CondCode = ARMCC::MI; break;
604 case ISD::SETOLE: CondCode = ARMCC::LS; break;
605 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
606 case ISD::SETO: CondCode = ARMCC::VC; break;
607 case ISD::SETUO: CondCode = ARMCC::VS; break;
608 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
609 case ISD::SETUGT: CondCode = ARMCC::HI; break;
610 case ISD::SETUGE: CondCode = ARMCC::PL; break;
612 case ISD::SETULT: CondCode = ARMCC::LT; break;
614 case ISD::SETULE: CondCode = ARMCC::LE; break;
616 case ISD::SETUNE: CondCode = ARMCC::NE; break;
620 //===----------------------------------------------------------------------===//
621 // Calling Convention Implementation
622 //===----------------------------------------------------------------------===//
624 #include "ARMGenCallingConv.inc"
626 // APCS f64 is in register pairs, possibly split to stack
627 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
628 CCValAssign::LocInfo &LocInfo,
629 CCState &State, bool CanFail) {
630 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
632 // Try to get the first register.
633 if (unsigned Reg = State.AllocateReg(RegList, 4))
634 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
636 // For the 2nd half of a v2f64, do not fail.
640 // Put the whole thing on the stack.
641 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
642 State.AllocateStack(8, 4),
647 // Try to get the second register.
648 if (unsigned Reg = State.AllocateReg(RegList, 4))
649 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
651 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
652 State.AllocateStack(4, 4),
657 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
658 CCValAssign::LocInfo &LocInfo,
659 ISD::ArgFlagsTy &ArgFlags,
661 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
663 if (LocVT == MVT::v2f64 &&
664 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
666 return true; // we handled it
669 // AAPCS f64 is in aligned register pairs
670 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
671 CCValAssign::LocInfo &LocInfo,
672 CCState &State, bool CanFail) {
673 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
674 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
676 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
678 // For the 2nd half of a v2f64, do not just fail.
682 // Put the whole thing on the stack.
683 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
684 State.AllocateStack(8, 8),
690 for (i = 0; i < 2; ++i)
691 if (HiRegList[i] == Reg)
694 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
695 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
700 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
701 CCValAssign::LocInfo &LocInfo,
702 ISD::ArgFlagsTy &ArgFlags,
704 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
706 if (LocVT == MVT::v2f64 &&
707 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
709 return true; // we handled it
712 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
713 CCValAssign::LocInfo &LocInfo, CCState &State) {
714 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
715 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
717 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
719 return false; // we didn't handle it
722 for (i = 0; i < 2; ++i)
723 if (HiRegList[i] == Reg)
726 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
727 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
732 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
733 CCValAssign::LocInfo &LocInfo,
734 ISD::ArgFlagsTy &ArgFlags,
736 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
738 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
740 return true; // we handled it
743 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
744 CCValAssign::LocInfo &LocInfo,
745 ISD::ArgFlagsTy &ArgFlags,
747 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
751 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
752 /// given CallingConvention value.
753 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
755 bool isVarArg) const {
758 llvm_unreachable("Unsupported calling convention");
760 case CallingConv::Fast:
761 // Use target triple & subtarget features to do actual dispatch.
762 if (Subtarget->isAAPCS_ABI()) {
763 if (Subtarget->hasVFP2() &&
764 FloatABIType == FloatABI::Hard && !isVarArg)
765 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
767 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
769 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
770 case CallingConv::ARM_AAPCS_VFP:
771 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
772 case CallingConv::ARM_AAPCS:
773 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
774 case CallingConv::ARM_APCS:
775 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
779 /// LowerCallResult - Lower the result values of a call into the
780 /// appropriate copies out of appropriate physical registers.
782 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
783 CallingConv::ID CallConv, bool isVarArg,
784 const SmallVectorImpl<ISD::InputArg> &Ins,
785 DebugLoc dl, SelectionDAG &DAG,
786 SmallVectorImpl<SDValue> &InVals) {
788 // Assign locations to each value returned by this call.
789 SmallVector<CCValAssign, 16> RVLocs;
790 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
791 RVLocs, *DAG.getContext());
792 CCInfo.AnalyzeCallResult(Ins,
793 CCAssignFnForNode(CallConv, /* Return*/ true,
796 // Copy all of the result registers out of their specified physreg.
797 for (unsigned i = 0; i != RVLocs.size(); ++i) {
798 CCValAssign VA = RVLocs[i];
801 if (VA.needsCustom()) {
802 // Handle f64 or half of a v2f64.
803 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
805 Chain = Lo.getValue(1);
806 InFlag = Lo.getValue(2);
807 VA = RVLocs[++i]; // skip ahead to next loc
808 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
810 Chain = Hi.getValue(1);
811 InFlag = Hi.getValue(2);
812 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
814 if (VA.getLocVT() == MVT::v2f64) {
815 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
816 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
817 DAG.getConstant(0, MVT::i32));
819 VA = RVLocs[++i]; // skip ahead to next loc
820 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
821 Chain = Lo.getValue(1);
822 InFlag = Lo.getValue(2);
823 VA = RVLocs[++i]; // skip ahead to next loc
824 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
825 Chain = Hi.getValue(1);
826 InFlag = Hi.getValue(2);
827 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
828 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
829 DAG.getConstant(1, MVT::i32));
832 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
834 Chain = Val.getValue(1);
835 InFlag = Val.getValue(2);
838 switch (VA.getLocInfo()) {
839 default: llvm_unreachable("Unknown loc info!");
840 case CCValAssign::Full: break;
841 case CCValAssign::BCvt:
842 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
846 InVals.push_back(Val);
852 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
853 /// by "Src" to address "Dst" of size "Size". Alignment information is
854 /// specified by the specific parameter attribute. The copy will be passed as
855 /// a byval function parameter.
856 /// Sometimes what we are copying is the end of a larger object, the part that
857 /// does not fit in registers.
859 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
860 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
862 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
863 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
864 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
867 /// LowerMemOpCallTo - Store the argument to the stack.
869 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
870 SDValue StackPtr, SDValue Arg,
871 DebugLoc dl, SelectionDAG &DAG,
872 const CCValAssign &VA,
873 ISD::ArgFlagsTy Flags) {
874 unsigned LocMemOffset = VA.getLocMemOffset();
875 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
876 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
877 if (Flags.isByVal()) {
878 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
880 return DAG.getStore(Chain, dl, Arg, PtrOff,
881 PseudoSourceValue::getStack(), LocMemOffset,
885 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
886 SDValue Chain, SDValue &Arg,
887 RegsToPassVector &RegsToPass,
888 CCValAssign &VA, CCValAssign &NextVA,
890 SmallVector<SDValue, 8> &MemOpChains,
891 ISD::ArgFlagsTy Flags) {
893 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
894 DAG.getVTList(MVT::i32, MVT::i32), Arg);
895 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
897 if (NextVA.isRegLoc())
898 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
900 assert(NextVA.isMemLoc());
901 if (StackPtr.getNode() == 0)
902 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
904 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
910 /// LowerCall - Lowering a call into a callseq_start <-
911 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
914 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
915 CallingConv::ID CallConv, bool isVarArg,
917 const SmallVectorImpl<ISD::OutputArg> &Outs,
918 const SmallVectorImpl<ISD::InputArg> &Ins,
919 DebugLoc dl, SelectionDAG &DAG,
920 SmallVectorImpl<SDValue> &InVals) {
921 // ARM target does not yet support tail call optimization.
924 // Analyze operands of the call, assigning locations to each operand.
925 SmallVector<CCValAssign, 16> ArgLocs;
926 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
928 CCInfo.AnalyzeCallOperands(Outs,
929 CCAssignFnForNode(CallConv, /* Return*/ false,
932 // Get a count of how many bytes are to be pushed on the stack.
933 unsigned NumBytes = CCInfo.getNextStackOffset();
935 // Adjust the stack pointer for the new arguments...
936 // These operations are automatically eliminated by the prolog/epilog pass
937 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
939 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
941 RegsToPassVector RegsToPass;
942 SmallVector<SDValue, 8> MemOpChains;
944 // Walk the register/memloc assignments, inserting copies/loads. In the case
945 // of tail call optimization, arguments are handled later.
946 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
949 CCValAssign &VA = ArgLocs[i];
950 SDValue Arg = Outs[realArgIdx].Val;
951 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
953 // Promote the value if needed.
954 switch (VA.getLocInfo()) {
955 default: llvm_unreachable("Unknown loc info!");
956 case CCValAssign::Full: break;
957 case CCValAssign::SExt:
958 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
960 case CCValAssign::ZExt:
961 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
963 case CCValAssign::AExt:
964 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
966 case CCValAssign::BCvt:
967 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
971 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
972 if (VA.needsCustom()) {
973 if (VA.getLocVT() == MVT::v2f64) {
974 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
975 DAG.getConstant(0, MVT::i32));
976 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
977 DAG.getConstant(1, MVT::i32));
979 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
980 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
982 VA = ArgLocs[++i]; // skip ahead to next loc
984 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
985 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
987 assert(VA.isMemLoc());
989 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
990 dl, DAG, VA, Flags));
993 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
994 StackPtr, MemOpChains, Flags);
996 } else if (VA.isRegLoc()) {
997 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
999 assert(VA.isMemLoc());
1001 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1002 dl, DAG, VA, Flags));
1006 if (!MemOpChains.empty())
1007 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1008 &MemOpChains[0], MemOpChains.size());
1010 // Build a sequence of copy-to-reg nodes chained together with token chain
1011 // and flag operands which copy the outgoing args into the appropriate regs.
1013 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1014 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1015 RegsToPass[i].second, InFlag);
1016 InFlag = Chain.getValue(1);
1019 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1020 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1021 // node so that legalize doesn't hack it.
1022 bool isDirect = false;
1023 bool isARMFunc = false;
1024 bool isLocalARMFunc = false;
1025 MachineFunction &MF = DAG.getMachineFunction();
1026 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1027 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1028 GlobalValue *GV = G->getGlobal();
1030 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1031 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1032 getTargetMachine().getRelocationModel() != Reloc::Static;
1033 isARMFunc = !Subtarget->isThumb() || isStub;
1034 // ARM call to a local ARM function is predicable.
1035 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1036 // tBX takes a register source operand.
1037 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1038 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1039 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1042 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1043 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1044 Callee = DAG.getLoad(getPointerTy(), dl,
1045 DAG.getEntryNode(), CPAddr,
1046 PseudoSourceValue::getConstantPool(), 0,
1048 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1049 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1050 getPointerTy(), Callee, PICLabel);
1052 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1053 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1055 bool isStub = Subtarget->isTargetDarwin() &&
1056 getTargetMachine().getRelocationModel() != Reloc::Static;
1057 isARMFunc = !Subtarget->isThumb() || isStub;
1058 // tBX takes a register source operand.
1059 const char *Sym = S->getSymbol();
1060 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1061 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1062 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1063 Sym, ARMPCLabelIndex, 4);
1064 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1065 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1066 Callee = DAG.getLoad(getPointerTy(), dl,
1067 DAG.getEntryNode(), CPAddr,
1068 PseudoSourceValue::getConstantPool(), 0,
1070 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1071 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1072 getPointerTy(), Callee, PICLabel);
1074 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1077 // FIXME: handle tail calls differently.
1079 if (Subtarget->isThumb()) {
1080 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1081 CallOpc = ARMISD::CALL_NOLINK;
1083 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1085 CallOpc = (isDirect || Subtarget->hasV5TOps())
1086 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1087 : ARMISD::CALL_NOLINK;
1089 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1090 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1091 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1092 InFlag = Chain.getValue(1);
1095 std::vector<SDValue> Ops;
1096 Ops.push_back(Chain);
1097 Ops.push_back(Callee);
1099 // Add argument registers to the end of the list so that they are known live
1101 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1102 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1103 RegsToPass[i].second.getValueType()));
1105 if (InFlag.getNode())
1106 Ops.push_back(InFlag);
1107 // Returns a chain and a flag for retval copy to use.
1108 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1109 &Ops[0], Ops.size());
1110 InFlag = Chain.getValue(1);
1112 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1113 DAG.getIntPtrConstant(0, true), InFlag);
1115 InFlag = Chain.getValue(1);
1117 // Handle result values, copying them out of physregs into vregs that we
1119 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1124 ARMTargetLowering::LowerReturn(SDValue Chain,
1125 CallingConv::ID CallConv, bool isVarArg,
1126 const SmallVectorImpl<ISD::OutputArg> &Outs,
1127 DebugLoc dl, SelectionDAG &DAG) {
1129 // CCValAssign - represent the assignment of the return value to a location.
1130 SmallVector<CCValAssign, 16> RVLocs;
1132 // CCState - Info about the registers and stack slots.
1133 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1136 // Analyze outgoing return values.
1137 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1140 // If this is the first return lowered for this function, add
1141 // the regs to the liveout set for the function.
1142 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1143 for (unsigned i = 0; i != RVLocs.size(); ++i)
1144 if (RVLocs[i].isRegLoc())
1145 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1150 // Copy the result values into the output registers.
1151 for (unsigned i = 0, realRVLocIdx = 0;
1153 ++i, ++realRVLocIdx) {
1154 CCValAssign &VA = RVLocs[i];
1155 assert(VA.isRegLoc() && "Can only return in registers!");
1157 SDValue Arg = Outs[realRVLocIdx].Val;
1159 switch (VA.getLocInfo()) {
1160 default: llvm_unreachable("Unknown loc info!");
1161 case CCValAssign::Full: break;
1162 case CCValAssign::BCvt:
1163 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1167 if (VA.needsCustom()) {
1168 if (VA.getLocVT() == MVT::v2f64) {
1169 // Extract the first half and return it in two registers.
1170 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1171 DAG.getConstant(0, MVT::i32));
1172 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1173 DAG.getVTList(MVT::i32, MVT::i32), Half);
1175 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1176 Flag = Chain.getValue(1);
1177 VA = RVLocs[++i]; // skip ahead to next loc
1178 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1179 HalfGPRs.getValue(1), Flag);
1180 Flag = Chain.getValue(1);
1181 VA = RVLocs[++i]; // skip ahead to next loc
1183 // Extract the 2nd half and fall through to handle it as an f64 value.
1184 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1185 DAG.getConstant(1, MVT::i32));
1187 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1189 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1190 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1191 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1192 Flag = Chain.getValue(1);
1193 VA = RVLocs[++i]; // skip ahead to next loc
1194 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1197 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1199 // Guarantee that all emitted copies are
1200 // stuck together, avoiding something bad.
1201 Flag = Chain.getValue(1);
1206 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1208 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1213 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1214 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1215 // one of the above mentioned nodes. It has to be wrapped because otherwise
1216 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1217 // be used to form addressing mode. These wrapped nodes will be selected
1219 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1220 EVT PtrVT = Op.getValueType();
1221 // FIXME there is no actual debug info here
1222 DebugLoc dl = Op.getDebugLoc();
1223 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1225 if (CP->isMachineConstantPoolEntry())
1226 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1227 CP->getAlignment());
1229 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1230 CP->getAlignment());
1231 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1234 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1235 MachineFunction &MF = DAG.getMachineFunction();
1236 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1237 unsigned ARMPCLabelIndex = 0;
1238 DebugLoc DL = Op.getDebugLoc();
1239 EVT PtrVT = getPointerTy();
1240 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1241 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1243 if (RelocM == Reloc::Static) {
1244 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1246 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1247 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1248 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1249 ARMCP::CPBlockAddress,
1251 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1253 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1254 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1255 PseudoSourceValue::getConstantPool(), 0,
1257 if (RelocM == Reloc::Static)
1259 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1260 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1263 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1265 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1266 SelectionDAG &DAG) {
1267 DebugLoc dl = GA->getDebugLoc();
1268 EVT PtrVT = getPointerTy();
1269 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1270 MachineFunction &MF = DAG.getMachineFunction();
1271 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1272 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1273 ARMConstantPoolValue *CPV =
1274 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1275 ARMCP::CPValue, PCAdj, "tlsgd", true);
1276 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1277 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1278 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1279 PseudoSourceValue::getConstantPool(), 0,
1281 SDValue Chain = Argument.getValue(1);
1283 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1284 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1286 // call __tls_get_addr.
1289 Entry.Node = Argument;
1290 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1291 Args.push_back(Entry);
1292 // FIXME: is there useful debug info available here?
1293 std::pair<SDValue, SDValue> CallResult =
1294 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1295 false, false, false, false,
1296 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1297 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1298 return CallResult.first;
1301 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1302 // "local exec" model.
1304 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1305 SelectionDAG &DAG) {
1306 GlobalValue *GV = GA->getGlobal();
1307 DebugLoc dl = GA->getDebugLoc();
1309 SDValue Chain = DAG.getEntryNode();
1310 EVT PtrVT = getPointerTy();
1311 // Get the Thread Pointer
1312 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1314 if (GV->isDeclaration()) {
1315 MachineFunction &MF = DAG.getMachineFunction();
1316 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1317 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1318 // Initial exec model.
1319 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1320 ARMConstantPoolValue *CPV =
1321 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1322 ARMCP::CPValue, PCAdj, "gottpoff", true);
1323 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1324 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1325 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1326 PseudoSourceValue::getConstantPool(), 0,
1328 Chain = Offset.getValue(1);
1330 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1331 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1333 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1334 PseudoSourceValue::getConstantPool(), 0,
1338 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1339 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1340 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1341 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1342 PseudoSourceValue::getConstantPool(), 0,
1346 // The address of the thread local variable is the add of the thread
1347 // pointer with the offset of the variable.
1348 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1352 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1353 // TODO: implement the "local dynamic" model
1354 assert(Subtarget->isTargetELF() &&
1355 "TLS not implemented for non-ELF targets");
1356 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1357 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1358 // otherwise use the "Local Exec" TLS Model
1359 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1360 return LowerToTLSGeneralDynamicModel(GA, DAG);
1362 return LowerToTLSExecModels(GA, DAG);
1365 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1366 SelectionDAG &DAG) {
1367 EVT PtrVT = getPointerTy();
1368 DebugLoc dl = Op.getDebugLoc();
1369 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1370 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1371 if (RelocM == Reloc::PIC_) {
1372 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1373 ARMConstantPoolValue *CPV =
1374 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1375 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1376 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1377 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1379 PseudoSourceValue::getConstantPool(), 0,
1381 SDValue Chain = Result.getValue(1);
1382 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1383 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1385 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1386 PseudoSourceValue::getGOT(), 0,
1390 // If we have T2 ops, we can materialize the address directly via movt/movw
1391 // pair. This is always cheaper.
1392 if (Subtarget->useMovt()) {
1393 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1394 DAG.getTargetGlobalAddress(GV, PtrVT));
1396 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1397 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1398 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1399 PseudoSourceValue::getConstantPool(), 0,
1405 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1406 SelectionDAG &DAG) {
1407 MachineFunction &MF = DAG.getMachineFunction();
1408 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1409 unsigned ARMPCLabelIndex = 0;
1410 EVT PtrVT = getPointerTy();
1411 DebugLoc dl = Op.getDebugLoc();
1412 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1413 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1415 if (RelocM == Reloc::Static)
1416 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1418 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1419 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1420 ARMConstantPoolValue *CPV =
1421 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1422 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1424 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1426 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1427 PseudoSourceValue::getConstantPool(), 0,
1429 SDValue Chain = Result.getValue(1);
1431 if (RelocM == Reloc::PIC_) {
1432 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1433 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1436 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1437 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1438 PseudoSourceValue::getGOT(), 0,
1444 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1446 assert(Subtarget->isTargetELF() &&
1447 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1448 MachineFunction &MF = DAG.getMachineFunction();
1449 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1450 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1451 EVT PtrVT = getPointerTy();
1452 DebugLoc dl = Op.getDebugLoc();
1453 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1454 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1455 "_GLOBAL_OFFSET_TABLE_",
1456 ARMPCLabelIndex, PCAdj);
1457 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1458 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1459 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1460 PseudoSourceValue::getConstantPool(), 0,
1462 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1463 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1467 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1468 const ARMSubtarget *Subtarget) {
1469 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1470 DebugLoc dl = Op.getDebugLoc();
1472 default: return SDValue(); // Don't custom lower most intrinsics.
1473 case Intrinsic::arm_thread_pointer: {
1474 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1475 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1477 case Intrinsic::eh_sjlj_lsda: {
1478 MachineFunction &MF = DAG.getMachineFunction();
1479 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1480 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1481 EVT PtrVT = getPointerTy();
1482 DebugLoc dl = Op.getDebugLoc();
1483 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1485 unsigned PCAdj = (RelocM != Reloc::PIC_)
1486 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1487 ARMConstantPoolValue *CPV =
1488 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1489 ARMCP::CPLSDA, PCAdj);
1490 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1491 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1493 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1494 PseudoSourceValue::getConstantPool(), 0,
1496 SDValue Chain = Result.getValue(1);
1498 if (RelocM == Reloc::PIC_) {
1499 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1500 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1504 case Intrinsic::eh_sjlj_setjmp:
1505 SDValue Val = Subtarget->isThumb() ?
1506 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1507 DAG.getConstant(0, MVT::i32);
1508 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1513 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1514 const ARMSubtarget *Subtarget) {
1515 DebugLoc dl = Op.getDebugLoc();
1516 SDValue Op5 = Op.getOperand(5);
1518 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1519 if (isDeviceBarrier) {
1520 if (Subtarget->hasV7Ops())
1521 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1523 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1524 DAG.getConstant(0, MVT::i32));
1526 if (Subtarget->hasV7Ops())
1527 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1529 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1530 DAG.getConstant(0, MVT::i32));
1535 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1536 unsigned VarArgsFrameIndex) {
1537 // vastart just stores the address of the VarArgsFrameIndex slot into the
1538 // memory location argument.
1539 DebugLoc dl = Op.getDebugLoc();
1540 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1541 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1542 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1543 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1548 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1549 SDNode *Node = Op.getNode();
1550 DebugLoc dl = Node->getDebugLoc();
1551 EVT VT = Node->getValueType(0);
1552 SDValue Chain = Op.getOperand(0);
1553 SDValue Size = Op.getOperand(1);
1554 SDValue Align = Op.getOperand(2);
1556 // Chain the dynamic stack allocation so that it doesn't modify the stack
1557 // pointer when other instructions are using the stack.
1558 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1560 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1561 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1562 if (AlignVal > StackAlign)
1563 // Do this now since selection pass cannot introduce new target
1564 // independent node.
1565 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1567 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1568 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1569 // do even more horrible hack later.
1570 MachineFunction &MF = DAG.getMachineFunction();
1571 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1572 if (AFI->isThumb1OnlyFunction()) {
1574 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1576 uint32_t Val = C->getZExtValue();
1577 if (Val <= 508 && ((Val & 3) == 0))
1581 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1584 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1585 SDValue Ops1[] = { Chain, Size, Align };
1586 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1587 Chain = Res.getValue(1);
1588 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1589 DAG.getIntPtrConstant(0, true), SDValue());
1590 SDValue Ops2[] = { Res, Chain };
1591 return DAG.getMergeValues(Ops2, 2, dl);
1595 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1596 SDValue &Root, SelectionDAG &DAG,
1598 MachineFunction &MF = DAG.getMachineFunction();
1599 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1601 TargetRegisterClass *RC;
1602 if (AFI->isThumb1OnlyFunction())
1603 RC = ARM::tGPRRegisterClass;
1605 RC = ARM::GPRRegisterClass;
1607 // Transform the arguments stored in physical registers into virtual ones.
1608 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1609 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1612 if (NextVA.isMemLoc()) {
1613 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1614 MachineFrameInfo *MFI = MF.getFrameInfo();
1615 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset(),
1618 // Create load node to retrieve arguments from the stack.
1619 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1620 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1621 PseudoSourceValue::getFixedStack(FI), 0,
1624 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1625 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1628 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1632 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1633 CallingConv::ID CallConv, bool isVarArg,
1634 const SmallVectorImpl<ISD::InputArg>
1636 DebugLoc dl, SelectionDAG &DAG,
1637 SmallVectorImpl<SDValue> &InVals) {
1639 MachineFunction &MF = DAG.getMachineFunction();
1640 MachineFrameInfo *MFI = MF.getFrameInfo();
1642 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1644 // Assign locations to all of the incoming arguments.
1645 SmallVector<CCValAssign, 16> ArgLocs;
1646 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1648 CCInfo.AnalyzeFormalArguments(Ins,
1649 CCAssignFnForNode(CallConv, /* Return*/ false,
1652 SmallVector<SDValue, 16> ArgValues;
1654 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1655 CCValAssign &VA = ArgLocs[i];
1657 // Arguments stored in registers.
1658 if (VA.isRegLoc()) {
1659 EVT RegVT = VA.getLocVT();
1662 if (VA.needsCustom()) {
1663 // f64 and vector types are split up into multiple registers or
1664 // combinations of registers and stack slots.
1667 if (VA.getLocVT() == MVT::v2f64) {
1668 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1670 VA = ArgLocs[++i]; // skip ahead to next loc
1671 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1673 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1674 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1675 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1676 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1677 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1679 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1682 TargetRegisterClass *RC;
1684 if (RegVT == MVT::f32)
1685 RC = ARM::SPRRegisterClass;
1686 else if (RegVT == MVT::f64)
1687 RC = ARM::DPRRegisterClass;
1688 else if (RegVT == MVT::v2f64)
1689 RC = ARM::QPRRegisterClass;
1690 else if (RegVT == MVT::i32)
1691 RC = (AFI->isThumb1OnlyFunction() ?
1692 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1694 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1696 // Transform the arguments in physical registers into virtual ones.
1697 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1698 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1701 // If this is an 8 or 16-bit value, it is really passed promoted
1702 // to 32 bits. Insert an assert[sz]ext to capture this, then
1703 // truncate to the right size.
1704 switch (VA.getLocInfo()) {
1705 default: llvm_unreachable("Unknown loc info!");
1706 case CCValAssign::Full: break;
1707 case CCValAssign::BCvt:
1708 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1710 case CCValAssign::SExt:
1711 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1712 DAG.getValueType(VA.getValVT()));
1713 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1715 case CCValAssign::ZExt:
1716 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1717 DAG.getValueType(VA.getValVT()));
1718 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1722 InVals.push_back(ArgValue);
1724 } else { // VA.isRegLoc()
1727 assert(VA.isMemLoc());
1728 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1730 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1731 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1734 // Create load nodes to retrieve arguments from the stack.
1735 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1736 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1737 PseudoSourceValue::getFixedStack(FI), 0,
1744 static const unsigned GPRArgRegs[] = {
1745 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1748 unsigned NumGPRs = CCInfo.getFirstUnallocated
1749 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1751 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1752 unsigned VARegSize = (4 - NumGPRs) * 4;
1753 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1754 unsigned ArgOffset = CCInfo.getNextStackOffset();
1755 if (VARegSaveSize) {
1756 // If this function is vararg, store any remaining integer argument regs
1757 // to their spots on the stack so that they may be loaded by deferencing
1758 // the result of va_next.
1759 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1760 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1761 VARegSaveSize - VARegSize,
1763 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1765 SmallVector<SDValue, 4> MemOps;
1766 for (; NumGPRs < 4; ++NumGPRs) {
1767 TargetRegisterClass *RC;
1768 if (AFI->isThumb1OnlyFunction())
1769 RC = ARM::tGPRRegisterClass;
1771 RC = ARM::GPRRegisterClass;
1773 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1774 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1775 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1776 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0,
1778 MemOps.push_back(Store);
1779 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1780 DAG.getConstant(4, getPointerTy()));
1782 if (!MemOps.empty())
1783 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1784 &MemOps[0], MemOps.size());
1786 // This will point to the next argument passed via stack.
1787 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
1793 /// isFloatingPointZero - Return true if this is +0.0.
1794 static bool isFloatingPointZero(SDValue Op) {
1795 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1796 return CFP->getValueAPF().isPosZero();
1797 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1798 // Maybe this has already been legalized into the constant pool?
1799 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1800 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1801 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1802 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1803 return CFP->getValueAPF().isPosZero();
1809 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1810 /// the given operands.
1812 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1813 SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
1814 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1815 unsigned C = RHSC->getZExtValue();
1816 if (!isLegalICmpImmediate(C)) {
1817 // Constant does not fit, try adjusting it by one?
1822 if (isLegalICmpImmediate(C-1)) {
1823 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1824 RHS = DAG.getConstant(C-1, MVT::i32);
1829 if (C > 0 && isLegalICmpImmediate(C-1)) {
1830 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1831 RHS = DAG.getConstant(C-1, MVT::i32);
1836 if (isLegalICmpImmediate(C+1)) {
1837 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1838 RHS = DAG.getConstant(C+1, MVT::i32);
1843 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1844 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1845 RHS = DAG.getConstant(C+1, MVT::i32);
1852 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1853 ARMISD::NodeType CompareType;
1856 CompareType = ARMISD::CMP;
1861 CompareType = ARMISD::CMPZ;
1864 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1865 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1868 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1869 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1872 if (!isFloatingPointZero(RHS))
1873 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1875 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1876 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1879 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1880 EVT VT = Op.getValueType();
1881 SDValue LHS = Op.getOperand(0);
1882 SDValue RHS = Op.getOperand(1);
1883 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1884 SDValue TrueVal = Op.getOperand(2);
1885 SDValue FalseVal = Op.getOperand(3);
1886 DebugLoc dl = Op.getDebugLoc();
1888 if (LHS.getValueType() == MVT::i32) {
1890 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1891 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1892 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1895 ARMCC::CondCodes CondCode, CondCode2;
1896 FPCCToARMCC(CC, CondCode, CondCode2);
1898 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1899 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1900 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1901 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1903 if (CondCode2 != ARMCC::AL) {
1904 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1905 // FIXME: Needs another CMP because flag can have but one use.
1906 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1907 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1908 Result, TrueVal, ARMCC2, CCR, Cmp2);
1913 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1914 SDValue Chain = Op.getOperand(0);
1915 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1916 SDValue LHS = Op.getOperand(2);
1917 SDValue RHS = Op.getOperand(3);
1918 SDValue Dest = Op.getOperand(4);
1919 DebugLoc dl = Op.getDebugLoc();
1921 if (LHS.getValueType() == MVT::i32) {
1923 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1924 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1925 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1926 Chain, Dest, ARMCC, CCR,Cmp);
1929 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1930 ARMCC::CondCodes CondCode, CondCode2;
1931 FPCCToARMCC(CC, CondCode, CondCode2);
1933 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1934 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1935 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1936 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1937 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1938 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1939 if (CondCode2 != ARMCC::AL) {
1940 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1941 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1942 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1947 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1948 SDValue Chain = Op.getOperand(0);
1949 SDValue Table = Op.getOperand(1);
1950 SDValue Index = Op.getOperand(2);
1951 DebugLoc dl = Op.getDebugLoc();
1953 EVT PTy = getPointerTy();
1954 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1955 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1956 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1957 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1958 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1959 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1960 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1961 if (Subtarget->isThumb2()) {
1962 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1963 // which does another jump to the destination. This also makes it easier
1964 // to translate it to TBB / TBH later.
1965 // FIXME: This might not work if the function is extremely large.
1966 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1967 Addr, Op.getOperand(2), JTI, UId);
1969 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1970 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1971 PseudoSourceValue::getJumpTable(), 0,
1973 Chain = Addr.getValue(1);
1974 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1975 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1977 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1978 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
1979 Chain = Addr.getValue(1);
1980 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1984 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1985 DebugLoc dl = Op.getDebugLoc();
1988 switch (Op.getOpcode()) {
1990 assert(0 && "Invalid opcode!");
1991 case ISD::FP_TO_SINT:
1992 Opc = ARMISD::FTOSI;
1994 case ISD::FP_TO_UINT:
1995 Opc = ARMISD::FTOUI;
1998 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1999 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2002 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2003 EVT VT = Op.getValueType();
2004 DebugLoc dl = Op.getDebugLoc();
2007 switch (Op.getOpcode()) {
2009 assert(0 && "Invalid opcode!");
2010 case ISD::SINT_TO_FP:
2011 Opc = ARMISD::SITOF;
2013 case ISD::UINT_TO_FP:
2014 Opc = ARMISD::UITOF;
2018 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2019 return DAG.getNode(Opc, dl, VT, Op);
2022 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2023 // Implement fcopysign with a fabs and a conditional fneg.
2024 SDValue Tmp0 = Op.getOperand(0);
2025 SDValue Tmp1 = Op.getOperand(1);
2026 DebugLoc dl = Op.getDebugLoc();
2027 EVT VT = Op.getValueType();
2028 EVT SrcVT = Tmp1.getValueType();
2029 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2030 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2031 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2032 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2033 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2036 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
2037 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2038 MFI->setFrameAddressIsTaken(true);
2039 EVT VT = Op.getValueType();
2040 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2041 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2042 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2043 ? ARM::R7 : ARM::R11;
2044 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2046 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2052 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2054 SDValue Dst, SDValue Src,
2055 SDValue Size, unsigned Align,
2057 const Value *DstSV, uint64_t DstSVOff,
2058 const Value *SrcSV, uint64_t SrcSVOff){
2059 // Do repeated 4-byte loads and stores. To be improved.
2060 // This requires 4-byte alignment.
2061 if ((Align & 3) != 0)
2063 // This requires the copy size to be a constant, preferrably
2064 // within a subtarget-specific limit.
2065 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2068 uint64_t SizeVal = ConstantSize->getZExtValue();
2069 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2072 unsigned BytesLeft = SizeVal & 3;
2073 unsigned NumMemOps = SizeVal >> 2;
2074 unsigned EmittedNumMemOps = 0;
2076 unsigned VTSize = 4;
2078 const unsigned MAX_LOADS_IN_LDM = 6;
2079 SDValue TFOps[MAX_LOADS_IN_LDM];
2080 SDValue Loads[MAX_LOADS_IN_LDM];
2081 uint64_t SrcOff = 0, DstOff = 0;
2083 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2084 // same number of stores. The loads and stores will get combined into
2085 // ldm/stm later on.
2086 while (EmittedNumMemOps < NumMemOps) {
2088 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2089 Loads[i] = DAG.getLoad(VT, dl, Chain,
2090 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2091 DAG.getConstant(SrcOff, MVT::i32)),
2092 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2093 TFOps[i] = Loads[i].getValue(1);
2096 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2099 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2100 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2101 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2102 DAG.getConstant(DstOff, MVT::i32)),
2103 DstSV, DstSVOff + DstOff, false, false, 0);
2106 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2108 EmittedNumMemOps += i;
2114 // Issue loads / stores for the trailing (1 - 3) bytes.
2115 unsigned BytesLeftSave = BytesLeft;
2118 if (BytesLeft >= 2) {
2126 Loads[i] = DAG.getLoad(VT, dl, Chain,
2127 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2128 DAG.getConstant(SrcOff, MVT::i32)),
2129 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2130 TFOps[i] = Loads[i].getValue(1);
2133 BytesLeft -= VTSize;
2135 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2138 BytesLeft = BytesLeftSave;
2140 if (BytesLeft >= 2) {
2148 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2149 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2150 DAG.getConstant(DstOff, MVT::i32)),
2151 DstSV, DstSVOff + DstOff, false, false, 0);
2154 BytesLeft -= VTSize;
2156 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2159 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2160 SDValue Op = N->getOperand(0);
2161 DebugLoc dl = N->getDebugLoc();
2162 if (N->getValueType(0) == MVT::f64) {
2163 // Turn i64->f64 into VMOVDRR.
2164 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2165 DAG.getConstant(0, MVT::i32));
2166 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2167 DAG.getConstant(1, MVT::i32));
2168 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2171 // Turn f64->i64 into VMOVRRD.
2172 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2173 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2175 // Merge the pieces into a single i64 value.
2176 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2179 /// getZeroVector - Returns a vector of specified type with all zero elements.
2181 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2182 assert(VT.isVector() && "Expected a vector type");
2184 // Zero vectors are used to represent vector negation and in those cases
2185 // will be implemented with the NEON VNEG instruction. However, VNEG does
2186 // not support i64 elements, so sometimes the zero vectors will need to be
2187 // explicitly constructed. For those cases, and potentially other uses in
2188 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2189 // to their dest type. This ensures they get CSE'd.
2191 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2192 SmallVector<SDValue, 8> Ops;
2195 if (VT.getSizeInBits() == 64) {
2196 Ops.assign(8, Cst); TVT = MVT::v8i8;
2198 Ops.assign(16, Cst); TVT = MVT::v16i8;
2200 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2202 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2205 /// getOnesVector - Returns a vector of specified type with all bits set.
2207 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2208 assert(VT.isVector() && "Expected a vector type");
2210 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2211 // dest type. This ensures they get CSE'd.
2213 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2214 SmallVector<SDValue, 8> Ops;
2217 if (VT.getSizeInBits() == 64) {
2218 Ops.assign(8, Cst); TVT = MVT::v8i8;
2220 Ops.assign(16, Cst); TVT = MVT::v16i8;
2222 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2224 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2227 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2228 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2229 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
2230 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2231 EVT VT = Op.getValueType();
2232 unsigned VTBits = VT.getSizeInBits();
2233 DebugLoc dl = Op.getDebugLoc();
2234 SDValue ShOpLo = Op.getOperand(0);
2235 SDValue ShOpHi = Op.getOperand(1);
2236 SDValue ShAmt = Op.getOperand(2);
2238 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2240 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2242 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2243 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2244 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2245 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2246 DAG.getConstant(VTBits, MVT::i32));
2247 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2248 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2249 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2251 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2252 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2254 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2255 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2258 SDValue Ops[2] = { Lo, Hi };
2259 return DAG.getMergeValues(Ops, 2, dl);
2262 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2263 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2264 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
2265 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2266 EVT VT = Op.getValueType();
2267 unsigned VTBits = VT.getSizeInBits();
2268 DebugLoc dl = Op.getDebugLoc();
2269 SDValue ShOpLo = Op.getOperand(0);
2270 SDValue ShOpHi = Op.getOperand(1);
2271 SDValue ShAmt = Op.getOperand(2);
2274 assert(Op.getOpcode() == ISD::SHL_PARTS);
2275 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2276 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2277 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2278 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2279 DAG.getConstant(VTBits, MVT::i32));
2280 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2281 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2283 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2284 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2285 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2287 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2288 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2291 SDValue Ops[2] = { Lo, Hi };
2292 return DAG.getMergeValues(Ops, 2, dl);
2295 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2296 const ARMSubtarget *ST) {
2297 EVT VT = N->getValueType(0);
2298 DebugLoc dl = N->getDebugLoc();
2300 if (!ST->hasV6T2Ops())
2303 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2304 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2307 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2308 const ARMSubtarget *ST) {
2309 EVT VT = N->getValueType(0);
2310 DebugLoc dl = N->getDebugLoc();
2312 // Lower vector shifts on NEON to use VSHL.
2313 if (VT.isVector()) {
2314 assert(ST->hasNEON() && "unexpected vector shift");
2316 // Left shifts translate directly to the vshiftu intrinsic.
2317 if (N->getOpcode() == ISD::SHL)
2318 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2319 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2320 N->getOperand(0), N->getOperand(1));
2322 assert((N->getOpcode() == ISD::SRA ||
2323 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2325 // NEON uses the same intrinsics for both left and right shifts. For
2326 // right shifts, the shift amounts are negative, so negate the vector of
2328 EVT ShiftVT = N->getOperand(1).getValueType();
2329 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2330 getZeroVector(ShiftVT, DAG, dl),
2332 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2333 Intrinsic::arm_neon_vshifts :
2334 Intrinsic::arm_neon_vshiftu);
2335 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2336 DAG.getConstant(vshiftInt, MVT::i32),
2337 N->getOperand(0), NegatedCount);
2340 // We can get here for a node like i32 = ISD::SHL i32, i64
2344 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2345 "Unknown shift to lower!");
2347 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2348 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2349 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2352 // If we are in thumb mode, we don't have RRX.
2353 if (ST->isThumb1Only()) return SDValue();
2355 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2356 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2357 DAG.getConstant(0, MVT::i32));
2358 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2359 DAG.getConstant(1, MVT::i32));
2361 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2362 // captures the result into a carry flag.
2363 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2364 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2366 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2367 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2369 // Merge the pieces into a single i64 value.
2370 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2373 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2374 SDValue TmpOp0, TmpOp1;
2375 bool Invert = false;
2379 SDValue Op0 = Op.getOperand(0);
2380 SDValue Op1 = Op.getOperand(1);
2381 SDValue CC = Op.getOperand(2);
2382 EVT VT = Op.getValueType();
2383 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2384 DebugLoc dl = Op.getDebugLoc();
2386 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2387 switch (SetCCOpcode) {
2388 default: llvm_unreachable("Illegal FP comparison"); break;
2390 case ISD::SETNE: Invert = true; // Fallthrough
2392 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2394 case ISD::SETLT: Swap = true; // Fallthrough
2396 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2398 case ISD::SETLE: Swap = true; // Fallthrough
2400 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2401 case ISD::SETUGE: Swap = true; // Fallthrough
2402 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2403 case ISD::SETUGT: Swap = true; // Fallthrough
2404 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2405 case ISD::SETUEQ: Invert = true; // Fallthrough
2407 // Expand this to (OLT | OGT).
2411 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2412 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2414 case ISD::SETUO: Invert = true; // Fallthrough
2416 // Expand this to (OLT | OGE).
2420 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2421 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2425 // Integer comparisons.
2426 switch (SetCCOpcode) {
2427 default: llvm_unreachable("Illegal integer comparison"); break;
2428 case ISD::SETNE: Invert = true;
2429 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2430 case ISD::SETLT: Swap = true;
2431 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2432 case ISD::SETLE: Swap = true;
2433 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2434 case ISD::SETULT: Swap = true;
2435 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2436 case ISD::SETULE: Swap = true;
2437 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2440 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2441 if (Opc == ARMISD::VCEQ) {
2444 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2446 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2449 // Ignore bitconvert.
2450 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2451 AndOp = AndOp.getOperand(0);
2453 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2455 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2456 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2463 std::swap(Op0, Op1);
2465 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2468 Result = DAG.getNOT(dl, Result, VT);
2473 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2474 /// VMOV instruction, and if so, return the constant being splatted.
2475 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2476 unsigned SplatBitSize, SelectionDAG &DAG) {
2477 switch (SplatBitSize) {
2479 // Any 1-byte value is OK.
2480 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2481 return DAG.getTargetConstant(SplatBits, MVT::i8);
2484 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2485 if ((SplatBits & ~0xff) == 0 ||
2486 (SplatBits & ~0xff00) == 0)
2487 return DAG.getTargetConstant(SplatBits, MVT::i16);
2491 // NEON's 32-bit VMOV supports splat values where:
2492 // * only one byte is nonzero, or
2493 // * the least significant byte is 0xff and the second byte is nonzero, or
2494 // * the least significant 2 bytes are 0xff and the third is nonzero.
2495 if ((SplatBits & ~0xff) == 0 ||
2496 (SplatBits & ~0xff00) == 0 ||
2497 (SplatBits & ~0xff0000) == 0 ||
2498 (SplatBits & ~0xff000000) == 0)
2499 return DAG.getTargetConstant(SplatBits, MVT::i32);
2501 if ((SplatBits & ~0xffff) == 0 &&
2502 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2503 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2505 if ((SplatBits & ~0xffffff) == 0 &&
2506 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2507 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2509 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2510 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2511 // VMOV.I32. A (very) minor optimization would be to replicate the value
2512 // and fall through here to test for a valid 64-bit splat. But, then the
2513 // caller would also need to check and handle the change in size.
2517 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2518 uint64_t BitMask = 0xff;
2520 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2521 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2523 else if ((SplatBits & BitMask) != 0)
2527 return DAG.getTargetConstant(Val, MVT::i64);
2531 llvm_unreachable("unexpected size for isVMOVSplat");
2538 /// getVMOVImm - If this is a build_vector of constants which can be
2539 /// formed by using a VMOV instruction of the specified element size,
2540 /// return the constant being splatted. The ByteSize field indicates the
2541 /// number of bytes of each element [1248].
2542 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2543 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2544 APInt SplatBits, SplatUndef;
2545 unsigned SplatBitSize;
2547 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2548 HasAnyUndefs, ByteSize * 8))
2551 if (SplatBitSize > ByteSize * 8)
2554 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2558 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2559 bool &ReverseVEXT, unsigned &Imm) {
2560 unsigned NumElts = VT.getVectorNumElements();
2561 ReverseVEXT = false;
2564 // If this is a VEXT shuffle, the immediate value is the index of the first
2565 // element. The other shuffle indices must be the successive elements after
2567 unsigned ExpectedElt = Imm;
2568 for (unsigned i = 1; i < NumElts; ++i) {
2569 // Increment the expected index. If it wraps around, it may still be
2570 // a VEXT but the source vectors must be swapped.
2572 if (ExpectedElt == NumElts * 2) {
2577 if (ExpectedElt != static_cast<unsigned>(M[i]))
2581 // Adjust the index value if the source operands will be swapped.
2588 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2589 /// instruction with the specified blocksize. (The order of the elements
2590 /// within each block of the vector is reversed.)
2591 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2592 unsigned BlockSize) {
2593 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2594 "Only possible block sizes for VREV are: 16, 32, 64");
2596 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2600 unsigned NumElts = VT.getVectorNumElements();
2601 unsigned BlockElts = M[0] + 1;
2603 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2606 for (unsigned i = 0; i < NumElts; ++i) {
2607 if ((unsigned) M[i] !=
2608 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2615 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2616 unsigned &WhichResult) {
2617 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2621 unsigned NumElts = VT.getVectorNumElements();
2622 WhichResult = (M[0] == 0 ? 0 : 1);
2623 for (unsigned i = 0; i < NumElts; i += 2) {
2624 if ((unsigned) M[i] != i + WhichResult ||
2625 (unsigned) M[i+1] != i + NumElts + WhichResult)
2631 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2632 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2633 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2634 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2635 unsigned &WhichResult) {
2636 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2640 unsigned NumElts = VT.getVectorNumElements();
2641 WhichResult = (M[0] == 0 ? 0 : 1);
2642 for (unsigned i = 0; i < NumElts; i += 2) {
2643 if ((unsigned) M[i] != i + WhichResult ||
2644 (unsigned) M[i+1] != i + WhichResult)
2650 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2651 unsigned &WhichResult) {
2652 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2656 unsigned NumElts = VT.getVectorNumElements();
2657 WhichResult = (M[0] == 0 ? 0 : 1);
2658 for (unsigned i = 0; i != NumElts; ++i) {
2659 if ((unsigned) M[i] != 2 * i + WhichResult)
2663 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2664 if (VT.is64BitVector() && EltSz == 32)
2670 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2671 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2672 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2673 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2674 unsigned &WhichResult) {
2675 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2679 unsigned Half = VT.getVectorNumElements() / 2;
2680 WhichResult = (M[0] == 0 ? 0 : 1);
2681 for (unsigned j = 0; j != 2; ++j) {
2682 unsigned Idx = WhichResult;
2683 for (unsigned i = 0; i != Half; ++i) {
2684 if ((unsigned) M[i + j * Half] != Idx)
2690 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2691 if (VT.is64BitVector() && EltSz == 32)
2697 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2698 unsigned &WhichResult) {
2699 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2703 unsigned NumElts = VT.getVectorNumElements();
2704 WhichResult = (M[0] == 0 ? 0 : 1);
2705 unsigned Idx = WhichResult * NumElts / 2;
2706 for (unsigned i = 0; i != NumElts; i += 2) {
2707 if ((unsigned) M[i] != Idx ||
2708 (unsigned) M[i+1] != Idx + NumElts)
2713 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2714 if (VT.is64BitVector() && EltSz == 32)
2720 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2721 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2722 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2723 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2724 unsigned &WhichResult) {
2725 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2729 unsigned NumElts = VT.getVectorNumElements();
2730 WhichResult = (M[0] == 0 ? 0 : 1);
2731 unsigned Idx = WhichResult * NumElts / 2;
2732 for (unsigned i = 0; i != NumElts; i += 2) {
2733 if ((unsigned) M[i] != Idx ||
2734 (unsigned) M[i+1] != Idx)
2739 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2740 if (VT.is64BitVector() && EltSz == 32)
2747 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2748 // Canonicalize all-zeros and all-ones vectors.
2749 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2750 if (ConstVal->isNullValue())
2751 return getZeroVector(VT, DAG, dl);
2752 if (ConstVal->isAllOnesValue())
2753 return getOnesVector(VT, DAG, dl);
2756 if (VT.is64BitVector()) {
2757 switch (Val.getValueType().getSizeInBits()) {
2758 case 8: CanonicalVT = MVT::v8i8; break;
2759 case 16: CanonicalVT = MVT::v4i16; break;
2760 case 32: CanonicalVT = MVT::v2i32; break;
2761 case 64: CanonicalVT = MVT::v1i64; break;
2762 default: llvm_unreachable("unexpected splat element type"); break;
2765 assert(VT.is128BitVector() && "unknown splat vector size");
2766 switch (Val.getValueType().getSizeInBits()) {
2767 case 8: CanonicalVT = MVT::v16i8; break;
2768 case 16: CanonicalVT = MVT::v8i16; break;
2769 case 32: CanonicalVT = MVT::v4i32; break;
2770 case 64: CanonicalVT = MVT::v2i64; break;
2771 default: llvm_unreachable("unexpected splat element type"); break;
2775 // Build a canonical splat for this value.
2776 SmallVector<SDValue, 8> Ops;
2777 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2778 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2780 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2783 // If this is a case we can't handle, return null and let the default
2784 // expansion code take care of it.
2785 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2786 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2787 DebugLoc dl = Op.getDebugLoc();
2788 EVT VT = Op.getValueType();
2790 APInt SplatBits, SplatUndef;
2791 unsigned SplatBitSize;
2793 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2794 if (SplatBitSize <= 64) {
2795 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2796 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2798 return BuildSplat(Val, VT, DAG, dl);
2802 // If there are only 2 elements in a 128-bit vector, insert them into an
2803 // undef vector. This handles the common case for 128-bit vector argument
2804 // passing, where the insertions should be translated to subreg accesses
2805 // with no real instructions.
2806 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2807 SDValue Val = DAG.getUNDEF(VT);
2808 SDValue Op0 = Op.getOperand(0);
2809 SDValue Op1 = Op.getOperand(1);
2810 if (Op0.getOpcode() != ISD::UNDEF)
2811 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2812 DAG.getIntPtrConstant(0));
2813 if (Op1.getOpcode() != ISD::UNDEF)
2814 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2815 DAG.getIntPtrConstant(1));
2822 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2823 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2824 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2825 /// are assumed to be legal.
2827 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2829 if (VT.getVectorNumElements() == 4 &&
2830 (VT.is128BitVector() || VT.is64BitVector())) {
2831 unsigned PFIndexes[4];
2832 for (unsigned i = 0; i != 4; ++i) {
2836 PFIndexes[i] = M[i];
2839 // Compute the index in the perfect shuffle table.
2840 unsigned PFTableIndex =
2841 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2842 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2843 unsigned Cost = (PFEntry >> 30);
2850 unsigned Imm, WhichResult;
2852 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2853 isVREVMask(M, VT, 64) ||
2854 isVREVMask(M, VT, 32) ||
2855 isVREVMask(M, VT, 16) ||
2856 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2857 isVTRNMask(M, VT, WhichResult) ||
2858 isVUZPMask(M, VT, WhichResult) ||
2859 isVZIPMask(M, VT, WhichResult) ||
2860 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2861 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2862 isVZIP_v_undef_Mask(M, VT, WhichResult));
2865 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2866 /// the specified operations to build the shuffle.
2867 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2868 SDValue RHS, SelectionDAG &DAG,
2870 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2871 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2872 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2875 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2884 OP_VUZPL, // VUZP, left result
2885 OP_VUZPR, // VUZP, right result
2886 OP_VZIPL, // VZIP, left result
2887 OP_VZIPR, // VZIP, right result
2888 OP_VTRNL, // VTRN, left result
2889 OP_VTRNR // VTRN, right result
2892 if (OpNum == OP_COPY) {
2893 if (LHSID == (1*9+2)*9+3) return LHS;
2894 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2898 SDValue OpLHS, OpRHS;
2899 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2900 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2901 EVT VT = OpLHS.getValueType();
2904 default: llvm_unreachable("Unknown shuffle opcode!");
2906 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2911 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2912 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2916 return DAG.getNode(ARMISD::VEXT, dl, VT,
2918 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2921 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2922 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2925 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2926 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2929 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2930 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2934 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2935 SDValue V1 = Op.getOperand(0);
2936 SDValue V2 = Op.getOperand(1);
2937 DebugLoc dl = Op.getDebugLoc();
2938 EVT VT = Op.getValueType();
2939 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2940 SmallVector<int, 8> ShuffleMask;
2942 // Convert shuffles that are directly supported on NEON to target-specific
2943 // DAG nodes, instead of keeping them as shuffles and matching them again
2944 // during code selection. This is more efficient and avoids the possibility
2945 // of inconsistencies between legalization and selection.
2946 // FIXME: floating-point vectors should be canonicalized to integer vectors
2947 // of the same time so that they get CSEd properly.
2948 SVN->getMask(ShuffleMask);
2950 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2951 int Lane = SVN->getSplatIndex();
2952 // If this is undef splat, generate it via "just" vdup, if possible.
2953 if (Lane == -1) Lane = 0;
2955 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2956 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2958 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2959 DAG.getConstant(Lane, MVT::i32));
2964 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2967 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2968 DAG.getConstant(Imm, MVT::i32));
2971 if (isVREVMask(ShuffleMask, VT, 64))
2972 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2973 if (isVREVMask(ShuffleMask, VT, 32))
2974 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2975 if (isVREVMask(ShuffleMask, VT, 16))
2976 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2978 // Check for Neon shuffles that modify both input vectors in place.
2979 // If both results are used, i.e., if there are two shuffles with the same
2980 // source operands and with masks corresponding to both results of one of
2981 // these operations, DAG memoization will ensure that a single node is
2982 // used for both shuffles.
2983 unsigned WhichResult;
2984 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2985 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2986 V1, V2).getValue(WhichResult);
2987 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2988 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2989 V1, V2).getValue(WhichResult);
2990 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2991 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2992 V1, V2).getValue(WhichResult);
2994 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
2995 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2996 V1, V1).getValue(WhichResult);
2997 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2998 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2999 V1, V1).getValue(WhichResult);
3000 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3001 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3002 V1, V1).getValue(WhichResult);
3004 // If the shuffle is not directly supported and it has 4 elements, use
3005 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3006 if (VT.getVectorNumElements() == 4 &&
3007 (VT.is128BitVector() || VT.is64BitVector())) {
3008 unsigned PFIndexes[4];
3009 for (unsigned i = 0; i != 4; ++i) {
3010 if (ShuffleMask[i] < 0)
3013 PFIndexes[i] = ShuffleMask[i];
3016 // Compute the index in the perfect shuffle table.
3017 unsigned PFTableIndex =
3018 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3020 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3021 unsigned Cost = (PFEntry >> 30);
3024 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3030 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3031 EVT VT = Op.getValueType();
3032 DebugLoc dl = Op.getDebugLoc();
3033 SDValue Vec = Op.getOperand(0);
3034 SDValue Lane = Op.getOperand(1);
3035 assert(VT == MVT::i32 &&
3036 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3037 "unexpected type for custom-lowering vector extract");
3038 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3041 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3042 // The only time a CONCAT_VECTORS operation can have legal types is when
3043 // two 64-bit vectors are concatenated to a 128-bit vector.
3044 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3045 "unexpected CONCAT_VECTORS");
3046 DebugLoc dl = Op.getDebugLoc();
3047 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3048 SDValue Op0 = Op.getOperand(0);
3049 SDValue Op1 = Op.getOperand(1);
3050 if (Op0.getOpcode() != ISD::UNDEF)
3051 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3052 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3053 DAG.getIntPtrConstant(0));
3054 if (Op1.getOpcode() != ISD::UNDEF)
3055 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3056 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3057 DAG.getIntPtrConstant(1));
3058 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3061 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3062 switch (Op.getOpcode()) {
3063 default: llvm_unreachable("Don't know how to custom lower this!");
3064 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3065 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3066 case ISD::GlobalAddress:
3067 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3068 LowerGlobalAddressELF(Op, DAG);
3069 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3070 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3071 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3072 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3073 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3074 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
3075 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3076 case ISD::SINT_TO_FP:
3077 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3078 case ISD::FP_TO_SINT:
3079 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3080 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3081 case ISD::RETURNADDR: break;
3082 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3083 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3084 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3086 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3089 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3090 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3091 case ISD::SRL_PARTS:
3092 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3093 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3094 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3095 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3096 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3097 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3098 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3103 /// ReplaceNodeResults - Replace the results of node with an illegal result
3104 /// type with new values built out of custom code.
3105 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3106 SmallVectorImpl<SDValue>&Results,
3107 SelectionDAG &DAG) {
3108 switch (N->getOpcode()) {
3110 llvm_unreachable("Don't know how to custom expand this!");
3112 case ISD::BIT_CONVERT:
3113 Results.push_back(ExpandBIT_CONVERT(N, DAG));
3117 SDValue Res = LowerShift(N, DAG, Subtarget);
3119 Results.push_back(Res);
3125 //===----------------------------------------------------------------------===//
3126 // ARM Scheduler Hooks
3127 //===----------------------------------------------------------------------===//
3130 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3131 MachineBasicBlock *BB,
3132 unsigned Size) const {
3133 unsigned dest = MI->getOperand(0).getReg();
3134 unsigned ptr = MI->getOperand(1).getReg();
3135 unsigned oldval = MI->getOperand(2).getReg();
3136 unsigned newval = MI->getOperand(3).getReg();
3137 unsigned scratch = BB->getParent()->getRegInfo()
3138 .createVirtualRegister(ARM::GPRRegisterClass);
3139 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3140 DebugLoc dl = MI->getDebugLoc();
3141 bool isThumb2 = Subtarget->isThumb2();
3143 unsigned ldrOpc, strOpc;
3145 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3147 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3148 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3151 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3152 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3155 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3156 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3160 MachineFunction *MF = BB->getParent();
3161 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3162 MachineFunction::iterator It = BB;
3163 ++It; // insert the new blocks after the current block
3165 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3166 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3167 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3168 MF->insert(It, loop1MBB);
3169 MF->insert(It, loop2MBB);
3170 MF->insert(It, exitMBB);
3171 exitMBB->transferSuccessors(BB);
3175 // fallthrough --> loop1MBB
3176 BB->addSuccessor(loop1MBB);
3179 // ldrex dest, [ptr]
3183 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3184 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3185 .addReg(dest).addReg(oldval));
3186 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3187 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3188 BB->addSuccessor(loop2MBB);
3189 BB->addSuccessor(exitMBB);
3192 // strex scratch, newval, [ptr]
3196 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3198 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3199 .addReg(scratch).addImm(0));
3200 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3201 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3202 BB->addSuccessor(loop1MBB);
3203 BB->addSuccessor(exitMBB);
3209 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3215 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3216 unsigned Size, unsigned BinOpcode) const {
3217 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3218 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3220 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3221 MachineFunction *MF = BB->getParent();
3222 MachineFunction::iterator It = BB;
3225 unsigned dest = MI->getOperand(0).getReg();
3226 unsigned ptr = MI->getOperand(1).getReg();
3227 unsigned incr = MI->getOperand(2).getReg();
3228 DebugLoc dl = MI->getDebugLoc();
3230 bool isThumb2 = Subtarget->isThumb2();
3231 unsigned ldrOpc, strOpc;
3233 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3235 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3236 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3239 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3240 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3243 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3244 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3248 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3249 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3250 MF->insert(It, loopMBB);
3251 MF->insert(It, exitMBB);
3252 exitMBB->transferSuccessors(BB);
3254 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3255 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3256 unsigned scratch2 = (!BinOpcode) ? incr :
3257 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3261 // fallthrough --> loopMBB
3262 BB->addSuccessor(loopMBB);
3266 // <binop> scratch2, dest, incr
3267 // strex scratch, scratch2, ptr
3270 // fallthrough --> exitMBB
3272 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3274 // operand order needs to go the other way for NAND
3275 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3276 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3277 addReg(incr).addReg(dest)).addReg(0);
3279 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3280 addReg(dest).addReg(incr)).addReg(0);
3283 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3285 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3286 .addReg(scratch).addImm(0));
3287 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3288 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3290 BB->addSuccessor(loopMBB);
3291 BB->addSuccessor(exitMBB);
3297 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3303 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3304 MachineBasicBlock *BB,
3305 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
3306 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3307 DebugLoc dl = MI->getDebugLoc();
3308 bool isThumb2 = Subtarget->isThumb2();
3309 switch (MI->getOpcode()) {
3312 llvm_unreachable("Unexpected instr type to insert");
3314 case ARM::ATOMIC_LOAD_ADD_I8:
3315 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3316 case ARM::ATOMIC_LOAD_ADD_I16:
3317 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3318 case ARM::ATOMIC_LOAD_ADD_I32:
3319 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3321 case ARM::ATOMIC_LOAD_AND_I8:
3322 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3323 case ARM::ATOMIC_LOAD_AND_I16:
3324 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3325 case ARM::ATOMIC_LOAD_AND_I32:
3326 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3328 case ARM::ATOMIC_LOAD_OR_I8:
3329 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3330 case ARM::ATOMIC_LOAD_OR_I16:
3331 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3332 case ARM::ATOMIC_LOAD_OR_I32:
3333 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3335 case ARM::ATOMIC_LOAD_XOR_I8:
3336 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3337 case ARM::ATOMIC_LOAD_XOR_I16:
3338 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3339 case ARM::ATOMIC_LOAD_XOR_I32:
3340 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3342 case ARM::ATOMIC_LOAD_NAND_I8:
3343 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3344 case ARM::ATOMIC_LOAD_NAND_I16:
3345 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3346 case ARM::ATOMIC_LOAD_NAND_I32:
3347 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3349 case ARM::ATOMIC_LOAD_SUB_I8:
3350 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3351 case ARM::ATOMIC_LOAD_SUB_I16:
3352 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3353 case ARM::ATOMIC_LOAD_SUB_I32:
3354 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3356 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3357 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3358 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3360 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3361 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3362 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3364 case ARM::tMOVCCr_pseudo: {
3365 // To "insert" a SELECT_CC instruction, we actually have to insert the
3366 // diamond control-flow pattern. The incoming instruction knows the
3367 // destination vreg to set, the condition code register to branch on, the
3368 // true/false values to select between, and a branch opcode to use.
3369 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3370 MachineFunction::iterator It = BB;
3376 // cmpTY ccX, r1, r2
3378 // fallthrough --> copy0MBB
3379 MachineBasicBlock *thisMBB = BB;
3380 MachineFunction *F = BB->getParent();
3381 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3382 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3383 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3384 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3385 F->insert(It, copy0MBB);
3386 F->insert(It, sinkMBB);
3387 // Update machine-CFG edges by first adding all successors of the current
3388 // block to the new block which will contain the Phi node for the select.
3389 // Also inform sdisel of the edge changes.
3390 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3391 E = BB->succ_end(); I != E; ++I) {
3392 EM->insert(std::make_pair(*I, sinkMBB));
3393 sinkMBB->addSuccessor(*I);
3395 // Next, remove all successors of the current block, and add the true
3396 // and fallthrough blocks as its successors.
3397 while (!BB->succ_empty())
3398 BB->removeSuccessor(BB->succ_begin());
3399 BB->addSuccessor(copy0MBB);
3400 BB->addSuccessor(sinkMBB);
3403 // %FalseValue = ...
3404 // # fallthrough to sinkMBB
3407 // Update machine-CFG edges
3408 BB->addSuccessor(sinkMBB);
3411 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3414 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3415 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3416 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3418 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3425 case ARM::t2SUBrSPi_:
3426 case ARM::t2SUBrSPi12_:
3427 case ARM::t2SUBrSPs_: {
3428 MachineFunction *MF = BB->getParent();
3429 unsigned DstReg = MI->getOperand(0).getReg();
3430 unsigned SrcReg = MI->getOperand(1).getReg();
3431 bool DstIsDead = MI->getOperand(0).isDead();
3432 bool SrcIsKill = MI->getOperand(1).isKill();
3434 if (SrcReg != ARM::SP) {
3435 // Copy the source to SP from virtual register.
3436 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3437 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3438 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3439 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3440 .addReg(SrcReg, getKillRegState(SrcIsKill));
3444 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3445 switch (MI->getOpcode()) {
3447 llvm_unreachable("Unexpected pseudo instruction!");
3453 OpOpc = ARM::tADDspr;
3456 OpOpc = ARM::tSUBspi;
3458 case ARM::t2SUBrSPi_:
3459 OpOpc = ARM::t2SUBrSPi;
3460 NeedPred = true; NeedCC = true;
3462 case ARM::t2SUBrSPi12_:
3463 OpOpc = ARM::t2SUBrSPi12;
3466 case ARM::t2SUBrSPs_:
3467 OpOpc = ARM::t2SUBrSPs;
3468 NeedPred = true; NeedCC = true; NeedOp3 = true;
3471 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3472 if (OpOpc == ARM::tAND)
3473 AddDefaultT1CC(MIB);
3474 MIB.addReg(ARM::SP);
3475 MIB.addOperand(MI->getOperand(2));
3477 MIB.addOperand(MI->getOperand(3));
3479 AddDefaultPred(MIB);
3483 // Copy the result from SP to virtual register.
3484 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3485 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3486 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3487 BuildMI(BB, dl, TII->get(CopyOpc))
3488 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3490 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3496 //===----------------------------------------------------------------------===//
3497 // ARM Optimization Hooks
3498 //===----------------------------------------------------------------------===//
3501 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3502 TargetLowering::DAGCombinerInfo &DCI) {
3503 SelectionDAG &DAG = DCI.DAG;
3504 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3505 EVT VT = N->getValueType(0);
3506 unsigned Opc = N->getOpcode();
3507 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3508 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3509 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3510 ISD::CondCode CC = ISD::SETCC_INVALID;
3513 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3515 SDValue CCOp = Slct.getOperand(0);
3516 if (CCOp.getOpcode() == ISD::SETCC)
3517 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3520 bool DoXform = false;
3522 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3525 if (LHS.getOpcode() == ISD::Constant &&
3526 cast<ConstantSDNode>(LHS)->isNullValue()) {
3528 } else if (CC != ISD::SETCC_INVALID &&
3529 RHS.getOpcode() == ISD::Constant &&
3530 cast<ConstantSDNode>(RHS)->isNullValue()) {
3531 std::swap(LHS, RHS);
3532 SDValue Op0 = Slct.getOperand(0);
3533 EVT OpVT = isSlctCC ? Op0.getValueType() :
3534 Op0.getOperand(0).getValueType();
3535 bool isInt = OpVT.isInteger();
3536 CC = ISD::getSetCCInverse(CC, isInt);
3538 if (!TLI.isCondCodeLegal(CC, OpVT))
3539 return SDValue(); // Inverse operator isn't legal.
3546 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3548 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3549 Slct.getOperand(0), Slct.getOperand(1), CC);
3550 SDValue CCOp = Slct.getOperand(0);
3552 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3553 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3554 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3555 CCOp, OtherOp, Result);
3560 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3561 static SDValue PerformADDCombine(SDNode *N,
3562 TargetLowering::DAGCombinerInfo &DCI) {
3563 // added by evan in r37685 with no testcase.
3564 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3566 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3567 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3568 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3569 if (Result.getNode()) return Result;
3571 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3572 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3573 if (Result.getNode()) return Result;
3579 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3580 static SDValue PerformSUBCombine(SDNode *N,
3581 TargetLowering::DAGCombinerInfo &DCI) {
3582 // added by evan in r37685 with no testcase.
3583 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3585 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3586 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3587 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3588 if (Result.getNode()) return Result;
3594 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3595 /// ARMISD::VMOVRRD.
3596 static SDValue PerformVMOVRRDCombine(SDNode *N,
3597 TargetLowering::DAGCombinerInfo &DCI) {
3598 // fmrrd(fmdrr x, y) -> x,y
3599 SDValue InDouble = N->getOperand(0);
3600 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3601 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3605 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3606 /// operand of a vector shift operation, where all the elements of the
3607 /// build_vector must have the same constant integer value.
3608 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3609 // Ignore bit_converts.
3610 while (Op.getOpcode() == ISD::BIT_CONVERT)
3611 Op = Op.getOperand(0);
3612 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3613 APInt SplatBits, SplatUndef;
3614 unsigned SplatBitSize;
3616 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3617 HasAnyUndefs, ElementBits) ||
3618 SplatBitSize > ElementBits)
3620 Cnt = SplatBits.getSExtValue();
3624 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3625 /// operand of a vector shift left operation. That value must be in the range:
3626 /// 0 <= Value < ElementBits for a left shift; or
3627 /// 0 <= Value <= ElementBits for a long left shift.
3628 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3629 assert(VT.isVector() && "vector shift count is not a vector type");
3630 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3631 if (! getVShiftImm(Op, ElementBits, Cnt))
3633 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3636 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3637 /// operand of a vector shift right operation. For a shift opcode, the value
3638 /// is positive, but for an intrinsic the value count must be negative. The
3639 /// absolute value must be in the range:
3640 /// 1 <= |Value| <= ElementBits for a right shift; or
3641 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3642 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3644 assert(VT.isVector() && "vector shift count is not a vector type");
3645 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3646 if (! getVShiftImm(Op, ElementBits, Cnt))
3650 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3653 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3654 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3655 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3658 // Don't do anything for most intrinsics.
3661 // Vector shifts: check for immediate versions and lower them.
3662 // Note: This is done during DAG combining instead of DAG legalizing because
3663 // the build_vectors for 64-bit vector element shift counts are generally
3664 // not legal, and it is hard to see their values after they get legalized to
3665 // loads from a constant pool.
3666 case Intrinsic::arm_neon_vshifts:
3667 case Intrinsic::arm_neon_vshiftu:
3668 case Intrinsic::arm_neon_vshiftls:
3669 case Intrinsic::arm_neon_vshiftlu:
3670 case Intrinsic::arm_neon_vshiftn:
3671 case Intrinsic::arm_neon_vrshifts:
3672 case Intrinsic::arm_neon_vrshiftu:
3673 case Intrinsic::arm_neon_vrshiftn:
3674 case Intrinsic::arm_neon_vqshifts:
3675 case Intrinsic::arm_neon_vqshiftu:
3676 case Intrinsic::arm_neon_vqshiftsu:
3677 case Intrinsic::arm_neon_vqshiftns:
3678 case Intrinsic::arm_neon_vqshiftnu:
3679 case Intrinsic::arm_neon_vqshiftnsu:
3680 case Intrinsic::arm_neon_vqrshiftns:
3681 case Intrinsic::arm_neon_vqrshiftnu:
3682 case Intrinsic::arm_neon_vqrshiftnsu: {
3683 EVT VT = N->getOperand(1).getValueType();
3685 unsigned VShiftOpc = 0;
3688 case Intrinsic::arm_neon_vshifts:
3689 case Intrinsic::arm_neon_vshiftu:
3690 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3691 VShiftOpc = ARMISD::VSHL;
3694 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3695 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3696 ARMISD::VSHRs : ARMISD::VSHRu);
3701 case Intrinsic::arm_neon_vshiftls:
3702 case Intrinsic::arm_neon_vshiftlu:
3703 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3705 llvm_unreachable("invalid shift count for vshll intrinsic");
3707 case Intrinsic::arm_neon_vrshifts:
3708 case Intrinsic::arm_neon_vrshiftu:
3709 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3713 case Intrinsic::arm_neon_vqshifts:
3714 case Intrinsic::arm_neon_vqshiftu:
3715 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3719 case Intrinsic::arm_neon_vqshiftsu:
3720 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3722 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3724 case Intrinsic::arm_neon_vshiftn:
3725 case Intrinsic::arm_neon_vrshiftn:
3726 case Intrinsic::arm_neon_vqshiftns:
3727 case Intrinsic::arm_neon_vqshiftnu:
3728 case Intrinsic::arm_neon_vqshiftnsu:
3729 case Intrinsic::arm_neon_vqrshiftns:
3730 case Intrinsic::arm_neon_vqrshiftnu:
3731 case Intrinsic::arm_neon_vqrshiftnsu:
3732 // Narrowing shifts require an immediate right shift.
3733 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3735 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3738 llvm_unreachable("unhandled vector shift");
3742 case Intrinsic::arm_neon_vshifts:
3743 case Intrinsic::arm_neon_vshiftu:
3744 // Opcode already set above.
3746 case Intrinsic::arm_neon_vshiftls:
3747 case Intrinsic::arm_neon_vshiftlu:
3748 if (Cnt == VT.getVectorElementType().getSizeInBits())
3749 VShiftOpc = ARMISD::VSHLLi;
3751 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3752 ARMISD::VSHLLs : ARMISD::VSHLLu);
3754 case Intrinsic::arm_neon_vshiftn:
3755 VShiftOpc = ARMISD::VSHRN; break;
3756 case Intrinsic::arm_neon_vrshifts:
3757 VShiftOpc = ARMISD::VRSHRs; break;
3758 case Intrinsic::arm_neon_vrshiftu:
3759 VShiftOpc = ARMISD::VRSHRu; break;
3760 case Intrinsic::arm_neon_vrshiftn:
3761 VShiftOpc = ARMISD::VRSHRN; break;
3762 case Intrinsic::arm_neon_vqshifts:
3763 VShiftOpc = ARMISD::VQSHLs; break;
3764 case Intrinsic::arm_neon_vqshiftu:
3765 VShiftOpc = ARMISD::VQSHLu; break;
3766 case Intrinsic::arm_neon_vqshiftsu:
3767 VShiftOpc = ARMISD::VQSHLsu; break;
3768 case Intrinsic::arm_neon_vqshiftns:
3769 VShiftOpc = ARMISD::VQSHRNs; break;
3770 case Intrinsic::arm_neon_vqshiftnu:
3771 VShiftOpc = ARMISD::VQSHRNu; break;
3772 case Intrinsic::arm_neon_vqshiftnsu:
3773 VShiftOpc = ARMISD::VQSHRNsu; break;
3774 case Intrinsic::arm_neon_vqrshiftns:
3775 VShiftOpc = ARMISD::VQRSHRNs; break;
3776 case Intrinsic::arm_neon_vqrshiftnu:
3777 VShiftOpc = ARMISD::VQRSHRNu; break;
3778 case Intrinsic::arm_neon_vqrshiftnsu:
3779 VShiftOpc = ARMISD::VQRSHRNsu; break;
3782 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3783 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3786 case Intrinsic::arm_neon_vshiftins: {
3787 EVT VT = N->getOperand(1).getValueType();
3789 unsigned VShiftOpc = 0;
3791 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3792 VShiftOpc = ARMISD::VSLI;
3793 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3794 VShiftOpc = ARMISD::VSRI;
3796 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3799 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3800 N->getOperand(1), N->getOperand(2),
3801 DAG.getConstant(Cnt, MVT::i32));
3804 case Intrinsic::arm_neon_vqrshifts:
3805 case Intrinsic::arm_neon_vqrshiftu:
3806 // No immediate versions of these to check for.
3813 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3814 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3815 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3816 /// vector element shift counts are generally not legal, and it is hard to see
3817 /// their values after they get legalized to loads from a constant pool.
3818 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3819 const ARMSubtarget *ST) {
3820 EVT VT = N->getValueType(0);
3822 // Nothing to be done for scalar shifts.
3823 if (! VT.isVector())
3826 assert(ST->hasNEON() && "unexpected vector shift");
3829 switch (N->getOpcode()) {
3830 default: llvm_unreachable("unexpected shift opcode");
3833 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3834 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3835 DAG.getConstant(Cnt, MVT::i32));
3840 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3841 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3842 ARMISD::VSHRs : ARMISD::VSHRu);
3843 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3844 DAG.getConstant(Cnt, MVT::i32));
3850 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3851 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3852 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3853 const ARMSubtarget *ST) {
3854 SDValue N0 = N->getOperand(0);
3856 // Check for sign- and zero-extensions of vector extract operations of 8-
3857 // and 16-bit vector elements. NEON supports these directly. They are
3858 // handled during DAG combining because type legalization will promote them
3859 // to 32-bit types and it is messy to recognize the operations after that.
3860 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3861 SDValue Vec = N0.getOperand(0);
3862 SDValue Lane = N0.getOperand(1);
3863 EVT VT = N->getValueType(0);
3864 EVT EltVT = N0.getValueType();
3865 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3867 if (VT == MVT::i32 &&
3868 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3869 TLI.isTypeLegal(Vec.getValueType())) {
3872 switch (N->getOpcode()) {
3873 default: llvm_unreachable("unexpected opcode");
3874 case ISD::SIGN_EXTEND:
3875 Opc = ARMISD::VGETLANEs;
3877 case ISD::ZERO_EXTEND:
3878 case ISD::ANY_EXTEND:
3879 Opc = ARMISD::VGETLANEu;
3882 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3889 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3890 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3891 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3892 const ARMSubtarget *ST) {
3893 // If the target supports NEON, try to use vmax/vmin instructions for f32
3894 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3895 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3896 // a NaN; only do the transformation when it matches that behavior.
3898 // For now only do this when using NEON for FP operations; if using VFP, it
3899 // is not obvious that the benefit outweighs the cost of switching to the
3901 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3902 N->getValueType(0) != MVT::f32)
3905 SDValue CondLHS = N->getOperand(0);
3906 SDValue CondRHS = N->getOperand(1);
3907 SDValue LHS = N->getOperand(2);
3908 SDValue RHS = N->getOperand(3);
3909 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3911 unsigned Opcode = 0;
3913 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3914 IsReversed = false; // x CC y ? x : y
3915 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3916 IsReversed = true ; // x CC y ? y : x
3930 // If LHS is NaN, an ordered comparison will be false and the result will
3931 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3932 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3933 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3934 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3936 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3937 // will return -0, so vmin can only be used for unsafe math or if one of
3938 // the operands is known to be nonzero.
3939 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3941 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3943 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3952 // If LHS is NaN, an ordered comparison will be false and the result will
3953 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3954 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3955 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3956 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3958 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3959 // will return +0, so vmax can only be used for unsafe math or if one of
3960 // the operands is known to be nonzero.
3961 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3963 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3965 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3971 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3974 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3975 DAGCombinerInfo &DCI) const {
3976 switch (N->getOpcode()) {
3978 case ISD::ADD: return PerformADDCombine(N, DCI);
3979 case ISD::SUB: return PerformSUBCombine(N, DCI);
3980 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3981 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
3984 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
3985 case ISD::SIGN_EXTEND:
3986 case ISD::ZERO_EXTEND:
3987 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
3988 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
3993 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3994 if (!Subtarget->hasV6Ops())
3995 // Pre-v6 does not support unaligned mem access.
3998 // v6+ may or may not support unaligned mem access depending on the system
4000 // FIXME: This is pretty conservative. Should we provide cmdline option to
4001 // control the behaviour?
4002 if (!Subtarget->isTargetDarwin())
4006 switch (VT.getSimpleVT().SimpleTy) {
4013 // FIXME: VLD1 etc with standard alignment is legal.
4017 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4022 switch (VT.getSimpleVT().SimpleTy) {
4023 default: return false;
4038 if ((V & (Scale - 1)) != 0)
4041 return V == (V & ((1LL << 5) - 1));
4044 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4045 const ARMSubtarget *Subtarget) {
4052 switch (VT.getSimpleVT().SimpleTy) {
4053 default: return false;
4058 // + imm12 or - imm8
4060 return V == (V & ((1LL << 8) - 1));
4061 return V == (V & ((1LL << 12) - 1));
4064 // Same as ARM mode. FIXME: NEON?
4065 if (!Subtarget->hasVFP2())
4070 return V == (V & ((1LL << 8) - 1));
4074 /// isLegalAddressImmediate - Return true if the integer value can be used
4075 /// as the offset of the target addressing mode for load / store of the
4077 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4078 const ARMSubtarget *Subtarget) {
4085 if (Subtarget->isThumb1Only())
4086 return isLegalT1AddressImmediate(V, VT);
4087 else if (Subtarget->isThumb2())
4088 return isLegalT2AddressImmediate(V, VT, Subtarget);
4093 switch (VT.getSimpleVT().SimpleTy) {
4094 default: return false;
4099 return V == (V & ((1LL << 12) - 1));
4102 return V == (V & ((1LL << 8) - 1));
4105 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4110 return V == (V & ((1LL << 8) - 1));
4114 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4116 int Scale = AM.Scale;
4120 switch (VT.getSimpleVT().SimpleTy) {
4121 default: return false;
4130 return Scale == 2 || Scale == 4 || Scale == 8;
4133 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4137 // Note, we allow "void" uses (basically, uses that aren't loads or
4138 // stores), because arm allows folding a scale into many arithmetic
4139 // operations. This should be made more precise and revisited later.
4141 // Allow r << imm, but the imm has to be a multiple of two.
4142 if (Scale & 1) return false;
4143 return isPowerOf2_32(Scale);
4147 /// isLegalAddressingMode - Return true if the addressing mode represented
4148 /// by AM is legal for this target, for a load/store of the specified type.
4149 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4150 const Type *Ty) const {
4151 EVT VT = getValueType(Ty, true);
4152 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4155 // Can never fold addr of global into load/store.
4160 case 0: // no scale reg, must be "r+i" or "r", or "i".
4163 if (Subtarget->isThumb1Only())
4167 // ARM doesn't support any R+R*scale+imm addr modes.
4174 if (Subtarget->isThumb2())
4175 return isLegalT2ScaledAddressingMode(AM, VT);
4177 int Scale = AM.Scale;
4178 switch (VT.getSimpleVT().SimpleTy) {
4179 default: return false;
4183 if (Scale < 0) Scale = -Scale;
4187 return isPowerOf2_32(Scale & ~1);
4191 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4196 // Note, we allow "void" uses (basically, uses that aren't loads or
4197 // stores), because arm allows folding a scale into many arithmetic
4198 // operations. This should be made more precise and revisited later.
4200 // Allow r << imm, but the imm has to be a multiple of two.
4201 if (Scale & 1) return false;
4202 return isPowerOf2_32(Scale);
4209 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4210 /// icmp immediate, that is the target has icmp instructions which can compare
4211 /// a register against the immediate without having to materialize the
4212 /// immediate into a register.
4213 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4214 if (!Subtarget->isThumb())
4215 return ARM_AM::getSOImmVal(Imm) != -1;
4216 if (Subtarget->isThumb2())
4217 return ARM_AM::getT2SOImmVal(Imm) != -1;
4218 return Imm >= 0 && Imm <= 255;
4221 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4222 bool isSEXTLoad, SDValue &Base,
4223 SDValue &Offset, bool &isInc,
4224 SelectionDAG &DAG) {
4225 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4228 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4230 Base = Ptr->getOperand(0);
4231 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4232 int RHSC = (int)RHS->getZExtValue();
4233 if (RHSC < 0 && RHSC > -256) {
4234 assert(Ptr->getOpcode() == ISD::ADD);
4236 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4240 isInc = (Ptr->getOpcode() == ISD::ADD);
4241 Offset = Ptr->getOperand(1);
4243 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4245 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4246 int RHSC = (int)RHS->getZExtValue();
4247 if (RHSC < 0 && RHSC > -0x1000) {
4248 assert(Ptr->getOpcode() == ISD::ADD);
4250 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4251 Base = Ptr->getOperand(0);
4256 if (Ptr->getOpcode() == ISD::ADD) {
4258 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4259 if (ShOpcVal != ARM_AM::no_shift) {
4260 Base = Ptr->getOperand(1);
4261 Offset = Ptr->getOperand(0);
4263 Base = Ptr->getOperand(0);
4264 Offset = Ptr->getOperand(1);
4269 isInc = (Ptr->getOpcode() == ISD::ADD);
4270 Base = Ptr->getOperand(0);
4271 Offset = Ptr->getOperand(1);
4275 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4279 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4280 bool isSEXTLoad, SDValue &Base,
4281 SDValue &Offset, bool &isInc,
4282 SelectionDAG &DAG) {
4283 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4286 Base = Ptr->getOperand(0);
4287 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4288 int RHSC = (int)RHS->getZExtValue();
4289 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4290 assert(Ptr->getOpcode() == ISD::ADD);
4292 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4294 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4295 isInc = Ptr->getOpcode() == ISD::ADD;
4296 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4304 /// getPreIndexedAddressParts - returns true by value, base pointer and
4305 /// offset pointer and addressing mode by reference if the node's address
4306 /// can be legally represented as pre-indexed load / store address.
4308 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4310 ISD::MemIndexedMode &AM,
4311 SelectionDAG &DAG) const {
4312 if (Subtarget->isThumb1Only())
4317 bool isSEXTLoad = false;
4318 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4319 Ptr = LD->getBasePtr();
4320 VT = LD->getMemoryVT();
4321 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4322 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4323 Ptr = ST->getBasePtr();
4324 VT = ST->getMemoryVT();
4329 bool isLegal = false;
4330 if (Subtarget->isThumb2())
4331 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4332 Offset, isInc, DAG);
4334 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4335 Offset, isInc, DAG);
4339 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4343 /// getPostIndexedAddressParts - returns true by value, base pointer and
4344 /// offset pointer and addressing mode by reference if this node can be
4345 /// combined with a load / store to form a post-indexed load / store.
4346 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4349 ISD::MemIndexedMode &AM,
4350 SelectionDAG &DAG) const {
4351 if (Subtarget->isThumb1Only())
4356 bool isSEXTLoad = false;
4357 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4358 VT = LD->getMemoryVT();
4359 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4360 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4361 VT = ST->getMemoryVT();
4366 bool isLegal = false;
4367 if (Subtarget->isThumb2())
4368 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4371 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4376 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4380 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4384 const SelectionDAG &DAG,
4385 unsigned Depth) const {
4386 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4387 switch (Op.getOpcode()) {
4389 case ARMISD::CMOV: {
4390 // Bits are known zero/one if known on the LHS and RHS.
4391 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4392 if (KnownZero == 0 && KnownOne == 0) return;
4394 APInt KnownZeroRHS, KnownOneRHS;
4395 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4396 KnownZeroRHS, KnownOneRHS, Depth+1);
4397 KnownZero &= KnownZeroRHS;
4398 KnownOne &= KnownOneRHS;
4404 //===----------------------------------------------------------------------===//
4405 // ARM Inline Assembly Support
4406 //===----------------------------------------------------------------------===//
4408 /// getConstraintType - Given a constraint letter, return the type of
4409 /// constraint it is for this target.
4410 ARMTargetLowering::ConstraintType
4411 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4412 if (Constraint.size() == 1) {
4413 switch (Constraint[0]) {
4415 case 'l': return C_RegisterClass;
4416 case 'w': return C_RegisterClass;
4419 return TargetLowering::getConstraintType(Constraint);
4422 std::pair<unsigned, const TargetRegisterClass*>
4423 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4425 if (Constraint.size() == 1) {
4426 // GCC ARM Constraint Letters
4427 switch (Constraint[0]) {
4429 if (Subtarget->isThumb())
4430 return std::make_pair(0U, ARM::tGPRRegisterClass);
4432 return std::make_pair(0U, ARM::GPRRegisterClass);
4434 return std::make_pair(0U, ARM::GPRRegisterClass);
4437 return std::make_pair(0U, ARM::SPRRegisterClass);
4438 if (VT.getSizeInBits() == 64)
4439 return std::make_pair(0U, ARM::DPRRegisterClass);
4440 if (VT.getSizeInBits() == 128)
4441 return std::make_pair(0U, ARM::QPRRegisterClass);
4445 if (StringRef("{cc}").equals_lower(Constraint))
4446 return std::make_pair(0U, ARM::CCRRegisterClass);
4448 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4451 std::vector<unsigned> ARMTargetLowering::
4452 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4454 if (Constraint.size() != 1)
4455 return std::vector<unsigned>();
4457 switch (Constraint[0]) { // GCC ARM Constraint Letters
4460 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4461 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4464 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4465 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4466 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4467 ARM::R12, ARM::LR, 0);
4470 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4471 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4472 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4473 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4474 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4475 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4476 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4477 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4478 if (VT.getSizeInBits() == 64)
4479 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4480 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4481 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4482 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4483 if (VT.getSizeInBits() == 128)
4484 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4485 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4489 return std::vector<unsigned>();
4492 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4493 /// vector. If it is invalid, don't add anything to Ops.
4494 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4497 std::vector<SDValue>&Ops,
4498 SelectionDAG &DAG) const {
4499 SDValue Result(0, 0);
4501 switch (Constraint) {
4503 case 'I': case 'J': case 'K': case 'L':
4504 case 'M': case 'N': case 'O':
4505 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4509 int64_t CVal64 = C->getSExtValue();
4510 int CVal = (int) CVal64;
4511 // None of these constraints allow values larger than 32 bits. Check
4512 // that the value fits in an int.
4516 switch (Constraint) {
4518 if (Subtarget->isThumb1Only()) {
4519 // This must be a constant between 0 and 255, for ADD
4521 if (CVal >= 0 && CVal <= 255)
4523 } else if (Subtarget->isThumb2()) {
4524 // A constant that can be used as an immediate value in a
4525 // data-processing instruction.
4526 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4529 // A constant that can be used as an immediate value in a
4530 // data-processing instruction.
4531 if (ARM_AM::getSOImmVal(CVal) != -1)
4537 if (Subtarget->isThumb()) { // FIXME thumb2
4538 // This must be a constant between -255 and -1, for negated ADD
4539 // immediates. This can be used in GCC with an "n" modifier that
4540 // prints the negated value, for use with SUB instructions. It is
4541 // not useful otherwise but is implemented for compatibility.
4542 if (CVal >= -255 && CVal <= -1)
4545 // This must be a constant between -4095 and 4095. It is not clear
4546 // what this constraint is intended for. Implemented for
4547 // compatibility with GCC.
4548 if (CVal >= -4095 && CVal <= 4095)
4554 if (Subtarget->isThumb1Only()) {
4555 // A 32-bit value where only one byte has a nonzero value. Exclude
4556 // zero to match GCC. This constraint is used by GCC internally for
4557 // constants that can be loaded with a move/shift combination.
4558 // It is not useful otherwise but is implemented for compatibility.
4559 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4561 } else if (Subtarget->isThumb2()) {
4562 // A constant whose bitwise inverse can be used as an immediate
4563 // value in a data-processing instruction. This can be used in GCC
4564 // with a "B" modifier that prints the inverted value, for use with
4565 // BIC and MVN instructions. It is not useful otherwise but is
4566 // implemented for compatibility.
4567 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4570 // A constant whose bitwise inverse can be used as an immediate
4571 // value in a data-processing instruction. This can be used in GCC
4572 // with a "B" modifier that prints the inverted value, for use with
4573 // BIC and MVN instructions. It is not useful otherwise but is
4574 // implemented for compatibility.
4575 if (ARM_AM::getSOImmVal(~CVal) != -1)
4581 if (Subtarget->isThumb1Only()) {
4582 // This must be a constant between -7 and 7,
4583 // for 3-operand ADD/SUB immediate instructions.
4584 if (CVal >= -7 && CVal < 7)
4586 } else if (Subtarget->isThumb2()) {
4587 // A constant whose negation can be used as an immediate value in a
4588 // data-processing instruction. This can be used in GCC with an "n"
4589 // modifier that prints the negated value, for use with SUB
4590 // instructions. It is not useful otherwise but is implemented for
4592 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4595 // A constant whose negation can be used as an immediate value in a
4596 // data-processing instruction. This can be used in GCC with an "n"
4597 // modifier that prints the negated value, for use with SUB
4598 // instructions. It is not useful otherwise but is implemented for
4600 if (ARM_AM::getSOImmVal(-CVal) != -1)
4606 if (Subtarget->isThumb()) { // FIXME thumb2
4607 // This must be a multiple of 4 between 0 and 1020, for
4608 // ADD sp + immediate.
4609 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4612 // A power of two or a constant between 0 and 32. This is used in
4613 // GCC for the shift amount on shifted register operands, but it is
4614 // useful in general for any shift amounts.
4615 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4621 if (Subtarget->isThumb()) { // FIXME thumb2
4622 // This must be a constant between 0 and 31, for shift amounts.
4623 if (CVal >= 0 && CVal <= 31)
4629 if (Subtarget->isThumb()) { // FIXME thumb2
4630 // This must be a multiple of 4 between -508 and 508, for
4631 // ADD/SUB sp = sp + immediate.
4632 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4637 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4641 if (Result.getNode()) {
4642 Ops.push_back(Result);
4645 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4650 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4651 // The ARM target isn't yet aware of offsets.
4655 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4656 APInt Imm = FPImm.bitcastToAPInt();
4657 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4658 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4659 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4661 // We can handle 4 bits of mantissa.
4662 // mantissa = (16+UInt(e:f:g:h))/16.
4663 if (Mantissa & 0x7ffff)
4666 if ((Mantissa & 0xf) != Mantissa)
4669 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4670 if (Exp < -3 || Exp > 4)
4672 Exp = ((Exp+3) & 0x7) ^ 4;
4674 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4677 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4678 APInt Imm = FPImm.bitcastToAPInt();
4679 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4680 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4681 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4683 // We can handle 4 bits of mantissa.
4684 // mantissa = (16+UInt(e:f:g:h))/16.
4685 if (Mantissa & 0xffffffffffffLL)
4688 if ((Mantissa & 0xf) != Mantissa)
4691 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4692 if (Exp < -3 || Exp > 4)
4694 Exp = ((Exp+3) & 0x7) ^ 4;
4696 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4699 /// isFPImmLegal - Returns true if the target can instruction select the
4700 /// specified FP immediate natively. If false, the legalizer will
4701 /// materialize the FP immediate as a load from a constant pool.
4702 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4703 if (!Subtarget->hasVFP3())
4706 return ARM::getVFPf32Imm(Imm) != -1;
4708 return ARM::getVFPf64Imm(Imm) != -1;