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->hasVFP3() && Subtarget->hasFP16()) {
440 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Custom);
441 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Custom);
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->hasV6Ops()) {
460 setIfCvtBlockSizeLimit(2);
461 setIfCvtDupBlockSizeLimit(1);
463 setIfCvtBlockSizeLimit(3);
464 setIfCvtDupBlockSizeLimit(2);
467 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
468 // Do not enable CodePlacementOpt for now: it currently runs after the
469 // ARMConstantIslandPass and messes up branch relaxation and placement
470 // of constant islands.
471 // benefitFromCodePlacementOpt = true;
474 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
477 case ARMISD::Wrapper: return "ARMISD::Wrapper";
478 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
479 case ARMISD::CALL: return "ARMISD::CALL";
480 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
481 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
482 case ARMISD::tCALL: return "ARMISD::tCALL";
483 case ARMISD::BRCOND: return "ARMISD::BRCOND";
484 case ARMISD::BR_JT: return "ARMISD::BR_JT";
485 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
486 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
487 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
488 case ARMISD::CMP: return "ARMISD::CMP";
489 case ARMISD::CMPZ: return "ARMISD::CMPZ";
490 case ARMISD::CMPFP: return "ARMISD::CMPFP";
491 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
492 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
493 case ARMISD::CMOV: return "ARMISD::CMOV";
494 case ARMISD::CNEG: return "ARMISD::CNEG";
496 case ARMISD::RBIT: return "ARMISD::RBIT";
498 case ARMISD::FTOSI: return "ARMISD::FTOSI";
499 case ARMISD::FTOUI: return "ARMISD::FTOUI";
500 case ARMISD::SITOF: return "ARMISD::SITOF";
501 case ARMISD::UITOF: return "ARMISD::UITOF";
502 case ARMISD::F16_TO_F32: return "ARMISD::F16_TO_F32";
503 case ARMISD::F32_TO_F16: return "ARMISD::F32_TO_F16";
505 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
506 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
507 case ARMISD::RRX: return "ARMISD::RRX";
509 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
510 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
512 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
513 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
515 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
517 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
519 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
520 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
522 case ARMISD::VCEQ: return "ARMISD::VCEQ";
523 case ARMISD::VCGE: return "ARMISD::VCGE";
524 case ARMISD::VCGEU: return "ARMISD::VCGEU";
525 case ARMISD::VCGT: return "ARMISD::VCGT";
526 case ARMISD::VCGTU: return "ARMISD::VCGTU";
527 case ARMISD::VTST: return "ARMISD::VTST";
529 case ARMISD::VSHL: return "ARMISD::VSHL";
530 case ARMISD::VSHRs: return "ARMISD::VSHRs";
531 case ARMISD::VSHRu: return "ARMISD::VSHRu";
532 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
533 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
534 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
535 case ARMISD::VSHRN: return "ARMISD::VSHRN";
536 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
537 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
538 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
539 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
540 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
541 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
542 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
543 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
544 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
545 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
546 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
547 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
548 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
549 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
550 case ARMISD::VDUP: return "ARMISD::VDUP";
551 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
552 case ARMISD::VEXT: return "ARMISD::VEXT";
553 case ARMISD::VREV64: return "ARMISD::VREV64";
554 case ARMISD::VREV32: return "ARMISD::VREV32";
555 case ARMISD::VREV16: return "ARMISD::VREV16";
556 case ARMISD::VZIP: return "ARMISD::VZIP";
557 case ARMISD::VUZP: return "ARMISD::VUZP";
558 case ARMISD::VTRN: return "ARMISD::VTRN";
559 case ARMISD::FMAX: return "ARMISD::FMAX";
560 case ARMISD::FMIN: return "ARMISD::FMIN";
564 /// getFunctionAlignment - Return the Log2 alignment of this function.
565 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
566 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
569 //===----------------------------------------------------------------------===//
571 //===----------------------------------------------------------------------===//
573 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
574 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
576 default: llvm_unreachable("Unknown condition code!");
577 case ISD::SETNE: return ARMCC::NE;
578 case ISD::SETEQ: return ARMCC::EQ;
579 case ISD::SETGT: return ARMCC::GT;
580 case ISD::SETGE: return ARMCC::GE;
581 case ISD::SETLT: return ARMCC::LT;
582 case ISD::SETLE: return ARMCC::LE;
583 case ISD::SETUGT: return ARMCC::HI;
584 case ISD::SETUGE: return ARMCC::HS;
585 case ISD::SETULT: return ARMCC::LO;
586 case ISD::SETULE: return ARMCC::LS;
590 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
591 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
592 ARMCC::CondCodes &CondCode2) {
593 CondCode2 = ARMCC::AL;
595 default: llvm_unreachable("Unknown FP condition!");
597 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
599 case ISD::SETOGT: CondCode = ARMCC::GT; break;
601 case ISD::SETOGE: CondCode = ARMCC::GE; break;
602 case ISD::SETOLT: CondCode = ARMCC::MI; break;
603 case ISD::SETOLE: CondCode = ARMCC::LS; break;
604 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
605 case ISD::SETO: CondCode = ARMCC::VC; break;
606 case ISD::SETUO: CondCode = ARMCC::VS; break;
607 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
608 case ISD::SETUGT: CondCode = ARMCC::HI; break;
609 case ISD::SETUGE: CondCode = ARMCC::PL; break;
611 case ISD::SETULT: CondCode = ARMCC::LT; break;
613 case ISD::SETULE: CondCode = ARMCC::LE; break;
615 case ISD::SETUNE: CondCode = ARMCC::NE; break;
619 //===----------------------------------------------------------------------===//
620 // Calling Convention Implementation
621 //===----------------------------------------------------------------------===//
623 #include "ARMGenCallingConv.inc"
625 // APCS f64 is in register pairs, possibly split to stack
626 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
627 CCValAssign::LocInfo &LocInfo,
628 CCState &State, bool CanFail) {
629 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
631 // Try to get the first register.
632 if (unsigned Reg = State.AllocateReg(RegList, 4))
633 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
635 // For the 2nd half of a v2f64, do not fail.
639 // Put the whole thing on the stack.
640 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
641 State.AllocateStack(8, 4),
646 // Try to get the second register.
647 if (unsigned Reg = State.AllocateReg(RegList, 4))
648 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
650 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
651 State.AllocateStack(4, 4),
656 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
657 CCValAssign::LocInfo &LocInfo,
658 ISD::ArgFlagsTy &ArgFlags,
660 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
662 if (LocVT == MVT::v2f64 &&
663 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
665 return true; // we handled it
668 // AAPCS f64 is in aligned register pairs
669 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
670 CCValAssign::LocInfo &LocInfo,
671 CCState &State, bool CanFail) {
672 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
673 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
675 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
677 // For the 2nd half of a v2f64, do not just fail.
681 // Put the whole thing on the stack.
682 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
683 State.AllocateStack(8, 8),
689 for (i = 0; i < 2; ++i)
690 if (HiRegList[i] == Reg)
693 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
694 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
699 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
700 CCValAssign::LocInfo &LocInfo,
701 ISD::ArgFlagsTy &ArgFlags,
703 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
705 if (LocVT == MVT::v2f64 &&
706 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
708 return true; // we handled it
711 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
712 CCValAssign::LocInfo &LocInfo, CCState &State) {
713 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
714 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
716 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
718 return false; // we didn't handle it
721 for (i = 0; i < 2; ++i)
722 if (HiRegList[i] == Reg)
725 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
726 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
731 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
732 CCValAssign::LocInfo &LocInfo,
733 ISD::ArgFlagsTy &ArgFlags,
735 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
737 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
739 return true; // we handled it
742 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
743 CCValAssign::LocInfo &LocInfo,
744 ISD::ArgFlagsTy &ArgFlags,
746 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
750 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
751 /// given CallingConvention value.
752 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
754 bool isVarArg) const {
757 llvm_unreachable("Unsupported calling convention");
759 case CallingConv::Fast:
760 // Use target triple & subtarget features to do actual dispatch.
761 if (Subtarget->isAAPCS_ABI()) {
762 if (Subtarget->hasVFP2() &&
763 FloatABIType == FloatABI::Hard && !isVarArg)
764 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
766 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
768 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
769 case CallingConv::ARM_AAPCS_VFP:
770 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
771 case CallingConv::ARM_AAPCS:
772 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
773 case CallingConv::ARM_APCS:
774 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
778 /// LowerCallResult - Lower the result values of a call into the
779 /// appropriate copies out of appropriate physical registers.
781 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
782 CallingConv::ID CallConv, bool isVarArg,
783 const SmallVectorImpl<ISD::InputArg> &Ins,
784 DebugLoc dl, SelectionDAG &DAG,
785 SmallVectorImpl<SDValue> &InVals) {
787 // Assign locations to each value returned by this call.
788 SmallVector<CCValAssign, 16> RVLocs;
789 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
790 RVLocs, *DAG.getContext());
791 CCInfo.AnalyzeCallResult(Ins,
792 CCAssignFnForNode(CallConv, /* Return*/ true,
795 // Copy all of the result registers out of their specified physreg.
796 for (unsigned i = 0; i != RVLocs.size(); ++i) {
797 CCValAssign VA = RVLocs[i];
800 if (VA.needsCustom()) {
801 // Handle f64 or half of a v2f64.
802 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
804 Chain = Lo.getValue(1);
805 InFlag = Lo.getValue(2);
806 VA = RVLocs[++i]; // skip ahead to next loc
807 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
809 Chain = Hi.getValue(1);
810 InFlag = Hi.getValue(2);
811 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
813 if (VA.getLocVT() == MVT::v2f64) {
814 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
815 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
816 DAG.getConstant(0, MVT::i32));
818 VA = RVLocs[++i]; // skip ahead to next loc
819 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
820 Chain = Lo.getValue(1);
821 InFlag = Lo.getValue(2);
822 VA = RVLocs[++i]; // skip ahead to next loc
823 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
824 Chain = Hi.getValue(1);
825 InFlag = Hi.getValue(2);
826 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
827 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
828 DAG.getConstant(1, MVT::i32));
831 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
833 Chain = Val.getValue(1);
834 InFlag = Val.getValue(2);
837 switch (VA.getLocInfo()) {
838 default: llvm_unreachable("Unknown loc info!");
839 case CCValAssign::Full: break;
840 case CCValAssign::BCvt:
841 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
845 InVals.push_back(Val);
851 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
852 /// by "Src" to address "Dst" of size "Size". Alignment information is
853 /// specified by the specific parameter attribute. The copy will be passed as
854 /// a byval function parameter.
855 /// Sometimes what we are copying is the end of a larger object, the part that
856 /// does not fit in registers.
858 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
859 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
861 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
862 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
863 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
866 /// LowerMemOpCallTo - Store the argument to the stack.
868 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
869 SDValue StackPtr, SDValue Arg,
870 DebugLoc dl, SelectionDAG &DAG,
871 const CCValAssign &VA,
872 ISD::ArgFlagsTy Flags) {
873 unsigned LocMemOffset = VA.getLocMemOffset();
874 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
875 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
876 if (Flags.isByVal()) {
877 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
879 return DAG.getStore(Chain, dl, Arg, PtrOff,
880 PseudoSourceValue::getStack(), LocMemOffset,
884 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
885 SDValue Chain, SDValue &Arg,
886 RegsToPassVector &RegsToPass,
887 CCValAssign &VA, CCValAssign &NextVA,
889 SmallVector<SDValue, 8> &MemOpChains,
890 ISD::ArgFlagsTy Flags) {
892 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
893 DAG.getVTList(MVT::i32, MVT::i32), Arg);
894 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
896 if (NextVA.isRegLoc())
897 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
899 assert(NextVA.isMemLoc());
900 if (StackPtr.getNode() == 0)
901 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
903 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
909 /// LowerCall - Lowering a call into a callseq_start <-
910 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
913 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
914 CallingConv::ID CallConv, bool isVarArg,
916 const SmallVectorImpl<ISD::OutputArg> &Outs,
917 const SmallVectorImpl<ISD::InputArg> &Ins,
918 DebugLoc dl, SelectionDAG &DAG,
919 SmallVectorImpl<SDValue> &InVals) {
920 // ARM target does not yet support tail call optimization.
923 // Analyze operands of the call, assigning locations to each operand.
924 SmallVector<CCValAssign, 16> ArgLocs;
925 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
927 CCInfo.AnalyzeCallOperands(Outs,
928 CCAssignFnForNode(CallConv, /* Return*/ false,
931 // Get a count of how many bytes are to be pushed on the stack.
932 unsigned NumBytes = CCInfo.getNextStackOffset();
934 // Adjust the stack pointer for the new arguments...
935 // These operations are automatically eliminated by the prolog/epilog pass
936 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
938 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
940 RegsToPassVector RegsToPass;
941 SmallVector<SDValue, 8> MemOpChains;
943 // Walk the register/memloc assignments, inserting copies/loads. In the case
944 // of tail call optimization, arguments are handled later.
945 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
948 CCValAssign &VA = ArgLocs[i];
949 SDValue Arg = Outs[realArgIdx].Val;
950 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
952 // Promote the value if needed.
953 switch (VA.getLocInfo()) {
954 default: llvm_unreachable("Unknown loc info!");
955 case CCValAssign::Full: break;
956 case CCValAssign::SExt:
957 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
959 case CCValAssign::ZExt:
960 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
962 case CCValAssign::AExt:
963 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
965 case CCValAssign::BCvt:
966 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
970 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
971 if (VA.needsCustom()) {
972 if (VA.getLocVT() == MVT::v2f64) {
973 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
974 DAG.getConstant(0, MVT::i32));
975 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
976 DAG.getConstant(1, MVT::i32));
978 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
979 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
981 VA = ArgLocs[++i]; // skip ahead to next loc
983 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
984 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
986 assert(VA.isMemLoc());
988 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
989 dl, DAG, VA, Flags));
992 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
993 StackPtr, MemOpChains, Flags);
995 } else if (VA.isRegLoc()) {
996 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
998 assert(VA.isMemLoc());
1000 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1001 dl, DAG, VA, Flags));
1005 if (!MemOpChains.empty())
1006 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1007 &MemOpChains[0], MemOpChains.size());
1009 // Build a sequence of copy-to-reg nodes chained together with token chain
1010 // and flag operands which copy the outgoing args into the appropriate regs.
1012 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1013 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1014 RegsToPass[i].second, InFlag);
1015 InFlag = Chain.getValue(1);
1018 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1019 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1020 // node so that legalize doesn't hack it.
1021 bool isDirect = false;
1022 bool isARMFunc = false;
1023 bool isLocalARMFunc = false;
1024 MachineFunction &MF = DAG.getMachineFunction();
1025 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1026 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1027 GlobalValue *GV = G->getGlobal();
1029 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1030 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1031 getTargetMachine().getRelocationModel() != Reloc::Static;
1032 isARMFunc = !Subtarget->isThumb() || isStub;
1033 // ARM call to a local ARM function is predicable.
1034 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1035 // tBX takes a register source operand.
1036 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1037 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1038 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1041 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1042 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1043 Callee = DAG.getLoad(getPointerTy(), dl,
1044 DAG.getEntryNode(), CPAddr,
1045 PseudoSourceValue::getConstantPool(), 0,
1047 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1048 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1049 getPointerTy(), Callee, PICLabel);
1051 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1052 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1054 bool isStub = Subtarget->isTargetDarwin() &&
1055 getTargetMachine().getRelocationModel() != Reloc::Static;
1056 isARMFunc = !Subtarget->isThumb() || isStub;
1057 // tBX takes a register source operand.
1058 const char *Sym = S->getSymbol();
1059 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1060 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1061 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1062 Sym, ARMPCLabelIndex, 4);
1063 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1064 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1065 Callee = DAG.getLoad(getPointerTy(), dl,
1066 DAG.getEntryNode(), CPAddr,
1067 PseudoSourceValue::getConstantPool(), 0,
1069 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1070 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1071 getPointerTy(), Callee, PICLabel);
1073 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1076 // FIXME: handle tail calls differently.
1078 if (Subtarget->isThumb()) {
1079 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1080 CallOpc = ARMISD::CALL_NOLINK;
1082 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1084 CallOpc = (isDirect || Subtarget->hasV5TOps())
1085 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1086 : ARMISD::CALL_NOLINK;
1088 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1089 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1090 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1091 InFlag = Chain.getValue(1);
1094 std::vector<SDValue> Ops;
1095 Ops.push_back(Chain);
1096 Ops.push_back(Callee);
1098 // Add argument registers to the end of the list so that they are known live
1100 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1101 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1102 RegsToPass[i].second.getValueType()));
1104 if (InFlag.getNode())
1105 Ops.push_back(InFlag);
1106 // Returns a chain and a flag for retval copy to use.
1107 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1108 &Ops[0], Ops.size());
1109 InFlag = Chain.getValue(1);
1111 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1112 DAG.getIntPtrConstant(0, true), InFlag);
1114 InFlag = Chain.getValue(1);
1116 // Handle result values, copying them out of physregs into vregs that we
1118 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1123 ARMTargetLowering::LowerReturn(SDValue Chain,
1124 CallingConv::ID CallConv, bool isVarArg,
1125 const SmallVectorImpl<ISD::OutputArg> &Outs,
1126 DebugLoc dl, SelectionDAG &DAG) {
1128 // CCValAssign - represent the assignment of the return value to a location.
1129 SmallVector<CCValAssign, 16> RVLocs;
1131 // CCState - Info about the registers and stack slots.
1132 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1135 // Analyze outgoing return values.
1136 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1139 // If this is the first return lowered for this function, add
1140 // the regs to the liveout set for the function.
1141 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1142 for (unsigned i = 0; i != RVLocs.size(); ++i)
1143 if (RVLocs[i].isRegLoc())
1144 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1149 // Copy the result values into the output registers.
1150 for (unsigned i = 0, realRVLocIdx = 0;
1152 ++i, ++realRVLocIdx) {
1153 CCValAssign &VA = RVLocs[i];
1154 assert(VA.isRegLoc() && "Can only return in registers!");
1156 SDValue Arg = Outs[realRVLocIdx].Val;
1158 switch (VA.getLocInfo()) {
1159 default: llvm_unreachable("Unknown loc info!");
1160 case CCValAssign::Full: break;
1161 case CCValAssign::BCvt:
1162 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1166 if (VA.needsCustom()) {
1167 if (VA.getLocVT() == MVT::v2f64) {
1168 // Extract the first half and return it in two registers.
1169 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1170 DAG.getConstant(0, MVT::i32));
1171 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1172 DAG.getVTList(MVT::i32, MVT::i32), Half);
1174 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1175 Flag = Chain.getValue(1);
1176 VA = RVLocs[++i]; // skip ahead to next loc
1177 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1178 HalfGPRs.getValue(1), Flag);
1179 Flag = Chain.getValue(1);
1180 VA = RVLocs[++i]; // skip ahead to next loc
1182 // Extract the 2nd half and fall through to handle it as an f64 value.
1183 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1184 DAG.getConstant(1, MVT::i32));
1186 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1188 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1189 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1190 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1191 Flag = Chain.getValue(1);
1192 VA = RVLocs[++i]; // skip ahead to next loc
1193 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1196 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1198 // Guarantee that all emitted copies are
1199 // stuck together, avoiding something bad.
1200 Flag = Chain.getValue(1);
1205 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1207 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1212 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1213 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1214 // one of the above mentioned nodes. It has to be wrapped because otherwise
1215 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1216 // be used to form addressing mode. These wrapped nodes will be selected
1218 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1219 EVT PtrVT = Op.getValueType();
1220 // FIXME there is no actual debug info here
1221 DebugLoc dl = Op.getDebugLoc();
1222 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1224 if (CP->isMachineConstantPoolEntry())
1225 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1226 CP->getAlignment());
1228 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1229 CP->getAlignment());
1230 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1233 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1234 MachineFunction &MF = DAG.getMachineFunction();
1235 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1236 unsigned ARMPCLabelIndex = 0;
1237 DebugLoc DL = Op.getDebugLoc();
1238 EVT PtrVT = getPointerTy();
1239 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1240 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1242 if (RelocM == Reloc::Static) {
1243 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1245 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1246 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1247 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1248 ARMCP::CPBlockAddress,
1250 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1252 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1253 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1254 PseudoSourceValue::getConstantPool(), 0,
1256 if (RelocM == Reloc::Static)
1258 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1259 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1262 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1264 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1265 SelectionDAG &DAG) {
1266 DebugLoc dl = GA->getDebugLoc();
1267 EVT PtrVT = getPointerTy();
1268 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1269 MachineFunction &MF = DAG.getMachineFunction();
1270 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1271 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1272 ARMConstantPoolValue *CPV =
1273 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1274 ARMCP::CPValue, PCAdj, "tlsgd", true);
1275 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1276 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1277 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1278 PseudoSourceValue::getConstantPool(), 0,
1280 SDValue Chain = Argument.getValue(1);
1282 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1283 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1285 // call __tls_get_addr.
1288 Entry.Node = Argument;
1289 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1290 Args.push_back(Entry);
1291 // FIXME: is there useful debug info available here?
1292 std::pair<SDValue, SDValue> CallResult =
1293 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1294 false, false, false, false,
1295 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1296 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1297 return CallResult.first;
1300 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1301 // "local exec" model.
1303 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1304 SelectionDAG &DAG) {
1305 GlobalValue *GV = GA->getGlobal();
1306 DebugLoc dl = GA->getDebugLoc();
1308 SDValue Chain = DAG.getEntryNode();
1309 EVT PtrVT = getPointerTy();
1310 // Get the Thread Pointer
1311 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1313 if (GV->isDeclaration()) {
1314 MachineFunction &MF = DAG.getMachineFunction();
1315 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1316 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1317 // Initial exec model.
1318 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1319 ARMConstantPoolValue *CPV =
1320 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1321 ARMCP::CPValue, PCAdj, "gottpoff", true);
1322 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1323 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1324 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1325 PseudoSourceValue::getConstantPool(), 0,
1327 Chain = Offset.getValue(1);
1329 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1330 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1332 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1333 PseudoSourceValue::getConstantPool(), 0,
1337 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1338 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1339 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1340 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1341 PseudoSourceValue::getConstantPool(), 0,
1345 // The address of the thread local variable is the add of the thread
1346 // pointer with the offset of the variable.
1347 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1351 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1352 // TODO: implement the "local dynamic" model
1353 assert(Subtarget->isTargetELF() &&
1354 "TLS not implemented for non-ELF targets");
1355 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1356 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1357 // otherwise use the "Local Exec" TLS Model
1358 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1359 return LowerToTLSGeneralDynamicModel(GA, DAG);
1361 return LowerToTLSExecModels(GA, DAG);
1364 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1365 SelectionDAG &DAG) {
1366 EVT PtrVT = getPointerTy();
1367 DebugLoc dl = Op.getDebugLoc();
1368 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1369 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1370 if (RelocM == Reloc::PIC_) {
1371 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1372 ARMConstantPoolValue *CPV =
1373 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1374 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1375 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1376 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1378 PseudoSourceValue::getConstantPool(), 0,
1380 SDValue Chain = Result.getValue(1);
1381 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1382 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1384 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1385 PseudoSourceValue::getGOT(), 0,
1389 // If we have T2 ops, we can materialize the address directly via movt/movw
1390 // pair. This is always cheaper.
1391 if (Subtarget->useMovt()) {
1392 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1393 DAG.getTargetGlobalAddress(GV, PtrVT));
1395 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1396 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1397 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1398 PseudoSourceValue::getConstantPool(), 0,
1404 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1405 SelectionDAG &DAG) {
1406 MachineFunction &MF = DAG.getMachineFunction();
1407 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1408 unsigned ARMPCLabelIndex = 0;
1409 EVT PtrVT = getPointerTy();
1410 DebugLoc dl = Op.getDebugLoc();
1411 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1412 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1414 if (RelocM == Reloc::Static)
1415 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1417 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1418 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1419 ARMConstantPoolValue *CPV =
1420 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1421 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1423 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1425 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1426 PseudoSourceValue::getConstantPool(), 0,
1428 SDValue Chain = Result.getValue(1);
1430 if (RelocM == Reloc::PIC_) {
1431 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1432 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1435 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1436 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1437 PseudoSourceValue::getGOT(), 0,
1443 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1445 assert(Subtarget->isTargetELF() &&
1446 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1447 MachineFunction &MF = DAG.getMachineFunction();
1448 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1449 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1450 EVT PtrVT = getPointerTy();
1451 DebugLoc dl = Op.getDebugLoc();
1452 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1453 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1454 "_GLOBAL_OFFSET_TABLE_",
1455 ARMPCLabelIndex, PCAdj);
1456 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1457 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1458 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1459 PseudoSourceValue::getConstantPool(), 0,
1461 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1462 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1466 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1467 const ARMSubtarget *Subtarget) {
1468 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1469 DebugLoc dl = Op.getDebugLoc();
1471 default: return SDValue(); // Don't custom lower most intrinsics.
1472 case Intrinsic::arm_thread_pointer: {
1473 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1474 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1476 case Intrinsic::eh_sjlj_lsda: {
1477 MachineFunction &MF = DAG.getMachineFunction();
1478 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1479 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1480 EVT PtrVT = getPointerTy();
1481 DebugLoc dl = Op.getDebugLoc();
1482 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1484 unsigned PCAdj = (RelocM != Reloc::PIC_)
1485 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1486 ARMConstantPoolValue *CPV =
1487 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1488 ARMCP::CPLSDA, PCAdj);
1489 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1490 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1492 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1493 PseudoSourceValue::getConstantPool(), 0,
1495 SDValue Chain = Result.getValue(1);
1497 if (RelocM == Reloc::PIC_) {
1498 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1499 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1503 case Intrinsic::eh_sjlj_setjmp:
1504 SDValue Val = Subtarget->isThumb() ?
1505 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1506 DAG.getConstant(0, MVT::i32);
1507 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1512 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1513 const ARMSubtarget *Subtarget) {
1514 DebugLoc dl = Op.getDebugLoc();
1515 SDValue Op5 = Op.getOperand(5);
1517 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1518 if (isDeviceBarrier) {
1519 if (Subtarget->hasV7Ops())
1520 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1522 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1523 DAG.getConstant(0, MVT::i32));
1525 if (Subtarget->hasV7Ops())
1526 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1528 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1529 DAG.getConstant(0, MVT::i32));
1534 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1535 unsigned VarArgsFrameIndex) {
1536 // vastart just stores the address of the VarArgsFrameIndex slot into the
1537 // memory location argument.
1538 DebugLoc dl = Op.getDebugLoc();
1539 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1540 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1541 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1542 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1547 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1548 SDNode *Node = Op.getNode();
1549 DebugLoc dl = Node->getDebugLoc();
1550 EVT VT = Node->getValueType(0);
1551 SDValue Chain = Op.getOperand(0);
1552 SDValue Size = Op.getOperand(1);
1553 SDValue Align = Op.getOperand(2);
1555 // Chain the dynamic stack allocation so that it doesn't modify the stack
1556 // pointer when other instructions are using the stack.
1557 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1559 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1560 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1561 if (AlignVal > StackAlign)
1562 // Do this now since selection pass cannot introduce new target
1563 // independent node.
1564 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1566 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1567 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1568 // do even more horrible hack later.
1569 MachineFunction &MF = DAG.getMachineFunction();
1570 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1571 if (AFI->isThumb1OnlyFunction()) {
1573 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1575 uint32_t Val = C->getZExtValue();
1576 if (Val <= 508 && ((Val & 3) == 0))
1580 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1583 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1584 SDValue Ops1[] = { Chain, Size, Align };
1585 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1586 Chain = Res.getValue(1);
1587 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1588 DAG.getIntPtrConstant(0, true), SDValue());
1589 SDValue Ops2[] = { Res, Chain };
1590 return DAG.getMergeValues(Ops2, 2, dl);
1594 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1595 SDValue &Root, SelectionDAG &DAG,
1597 MachineFunction &MF = DAG.getMachineFunction();
1598 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1600 TargetRegisterClass *RC;
1601 if (AFI->isThumb1OnlyFunction())
1602 RC = ARM::tGPRRegisterClass;
1604 RC = ARM::GPRRegisterClass;
1606 // Transform the arguments stored in physical registers into virtual ones.
1607 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1608 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1611 if (NextVA.isMemLoc()) {
1612 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1613 MachineFrameInfo *MFI = MF.getFrameInfo();
1614 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset(),
1617 // Create load node to retrieve arguments from the stack.
1618 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1619 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1620 PseudoSourceValue::getFixedStack(FI), 0,
1623 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1624 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1627 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1631 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1632 CallingConv::ID CallConv, bool isVarArg,
1633 const SmallVectorImpl<ISD::InputArg>
1635 DebugLoc dl, SelectionDAG &DAG,
1636 SmallVectorImpl<SDValue> &InVals) {
1638 MachineFunction &MF = DAG.getMachineFunction();
1639 MachineFrameInfo *MFI = MF.getFrameInfo();
1641 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1643 // Assign locations to all of the incoming arguments.
1644 SmallVector<CCValAssign, 16> ArgLocs;
1645 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1647 CCInfo.AnalyzeFormalArguments(Ins,
1648 CCAssignFnForNode(CallConv, /* Return*/ false,
1651 SmallVector<SDValue, 16> ArgValues;
1653 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1654 CCValAssign &VA = ArgLocs[i];
1656 // Arguments stored in registers.
1657 if (VA.isRegLoc()) {
1658 EVT RegVT = VA.getLocVT();
1661 if (VA.needsCustom()) {
1662 // f64 and vector types are split up into multiple registers or
1663 // combinations of registers and stack slots.
1666 if (VA.getLocVT() == MVT::v2f64) {
1667 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1669 VA = ArgLocs[++i]; // skip ahead to next loc
1670 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1672 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1673 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1674 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1675 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1676 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1678 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1681 TargetRegisterClass *RC;
1683 if (RegVT == MVT::f32)
1684 RC = ARM::SPRRegisterClass;
1685 else if (RegVT == MVT::f64)
1686 RC = ARM::DPRRegisterClass;
1687 else if (RegVT == MVT::v2f64)
1688 RC = ARM::QPRRegisterClass;
1689 else if (RegVT == MVT::i32)
1690 RC = (AFI->isThumb1OnlyFunction() ?
1691 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1693 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1695 // Transform the arguments in physical registers into virtual ones.
1696 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1697 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1700 // If this is an 8 or 16-bit value, it is really passed promoted
1701 // to 32 bits. Insert an assert[sz]ext to capture this, then
1702 // truncate to the right size.
1703 switch (VA.getLocInfo()) {
1704 default: llvm_unreachable("Unknown loc info!");
1705 case CCValAssign::Full: break;
1706 case CCValAssign::BCvt:
1707 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1709 case CCValAssign::SExt:
1710 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1711 DAG.getValueType(VA.getValVT()));
1712 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1714 case CCValAssign::ZExt:
1715 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1716 DAG.getValueType(VA.getValVT()));
1717 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1721 InVals.push_back(ArgValue);
1723 } else { // VA.isRegLoc()
1726 assert(VA.isMemLoc());
1727 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1729 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1730 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1733 // Create load nodes to retrieve arguments from the stack.
1734 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1735 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1736 PseudoSourceValue::getFixedStack(FI), 0,
1743 static const unsigned GPRArgRegs[] = {
1744 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1747 unsigned NumGPRs = CCInfo.getFirstUnallocated
1748 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1750 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1751 unsigned VARegSize = (4 - NumGPRs) * 4;
1752 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1753 unsigned ArgOffset = CCInfo.getNextStackOffset();
1754 if (VARegSaveSize) {
1755 // If this function is vararg, store any remaining integer argument regs
1756 // to their spots on the stack so that they may be loaded by deferencing
1757 // the result of va_next.
1758 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1759 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1760 VARegSaveSize - VARegSize,
1762 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1764 SmallVector<SDValue, 4> MemOps;
1765 for (; NumGPRs < 4; ++NumGPRs) {
1766 TargetRegisterClass *RC;
1767 if (AFI->isThumb1OnlyFunction())
1768 RC = ARM::tGPRRegisterClass;
1770 RC = ARM::GPRRegisterClass;
1772 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1773 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1774 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1775 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0,
1777 MemOps.push_back(Store);
1778 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1779 DAG.getConstant(4, getPointerTy()));
1781 if (!MemOps.empty())
1782 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1783 &MemOps[0], MemOps.size());
1785 // This will point to the next argument passed via stack.
1786 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
1792 /// isFloatingPointZero - Return true if this is +0.0.
1793 static bool isFloatingPointZero(SDValue Op) {
1794 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1795 return CFP->getValueAPF().isPosZero();
1796 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1797 // Maybe this has already been legalized into the constant pool?
1798 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1799 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1800 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1801 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1802 return CFP->getValueAPF().isPosZero();
1808 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1809 /// the given operands.
1811 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1812 SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
1813 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1814 unsigned C = RHSC->getZExtValue();
1815 if (!isLegalICmpImmediate(C)) {
1816 // Constant does not fit, try adjusting it by one?
1821 if (isLegalICmpImmediate(C-1)) {
1822 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1823 RHS = DAG.getConstant(C-1, MVT::i32);
1828 if (C > 0 && isLegalICmpImmediate(C-1)) {
1829 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1830 RHS = DAG.getConstant(C-1, MVT::i32);
1835 if (isLegalICmpImmediate(C+1)) {
1836 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1837 RHS = DAG.getConstant(C+1, MVT::i32);
1842 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1843 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1844 RHS = DAG.getConstant(C+1, MVT::i32);
1851 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1852 ARMISD::NodeType CompareType;
1855 CompareType = ARMISD::CMP;
1860 CompareType = ARMISD::CMPZ;
1863 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1864 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1867 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1868 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1871 if (!isFloatingPointZero(RHS))
1872 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1874 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1875 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1878 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1879 EVT VT = Op.getValueType();
1880 SDValue LHS = Op.getOperand(0);
1881 SDValue RHS = Op.getOperand(1);
1882 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1883 SDValue TrueVal = Op.getOperand(2);
1884 SDValue FalseVal = Op.getOperand(3);
1885 DebugLoc dl = Op.getDebugLoc();
1887 if (LHS.getValueType() == MVT::i32) {
1889 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1890 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1891 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1894 ARMCC::CondCodes CondCode, CondCode2;
1895 FPCCToARMCC(CC, CondCode, CondCode2);
1897 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1898 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1899 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1900 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1902 if (CondCode2 != ARMCC::AL) {
1903 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1904 // FIXME: Needs another CMP because flag can have but one use.
1905 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1906 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1907 Result, TrueVal, ARMCC2, CCR, Cmp2);
1912 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1913 SDValue Chain = Op.getOperand(0);
1914 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1915 SDValue LHS = Op.getOperand(2);
1916 SDValue RHS = Op.getOperand(3);
1917 SDValue Dest = Op.getOperand(4);
1918 DebugLoc dl = Op.getDebugLoc();
1920 if (LHS.getValueType() == MVT::i32) {
1922 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1923 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1924 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1925 Chain, Dest, ARMCC, CCR,Cmp);
1928 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1929 ARMCC::CondCodes CondCode, CondCode2;
1930 FPCCToARMCC(CC, CondCode, CondCode2);
1932 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1933 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1934 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1935 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1936 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1937 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1938 if (CondCode2 != ARMCC::AL) {
1939 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1940 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1941 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1946 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1947 SDValue Chain = Op.getOperand(0);
1948 SDValue Table = Op.getOperand(1);
1949 SDValue Index = Op.getOperand(2);
1950 DebugLoc dl = Op.getDebugLoc();
1952 EVT PTy = getPointerTy();
1953 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1954 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1955 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1956 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1957 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1958 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1959 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1960 if (Subtarget->isThumb2()) {
1961 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1962 // which does another jump to the destination. This also makes it easier
1963 // to translate it to TBB / TBH later.
1964 // FIXME: This might not work if the function is extremely large.
1965 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1966 Addr, Op.getOperand(2), JTI, UId);
1968 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1969 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1970 PseudoSourceValue::getJumpTable(), 0,
1972 Chain = Addr.getValue(1);
1973 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1974 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1976 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1977 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
1978 Chain = Addr.getValue(1);
1979 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1983 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1984 DebugLoc dl = Op.getDebugLoc();
1987 switch (Op.getOpcode()) {
1989 assert(0 && "Invalid opcode!");
1990 case ISD::FP32_TO_FP16:
1991 Opc = ARMISD::F32_TO_F16;
1993 case ISD::FP_TO_SINT:
1994 Opc = ARMISD::FTOSI;
1996 case ISD::FP_TO_UINT:
1997 Opc = ARMISD::FTOUI;
2000 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2001 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2004 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2005 EVT VT = Op.getValueType();
2006 DebugLoc dl = Op.getDebugLoc();
2009 switch (Op.getOpcode()) {
2011 assert(0 && "Invalid opcode!");
2012 case ISD::FP16_TO_FP32:
2013 Opc = ARMISD::F16_TO_F32;
2015 case ISD::SINT_TO_FP:
2016 Opc = ARMISD::SITOF;
2018 case ISD::UINT_TO_FP:
2019 Opc = ARMISD::UITOF;
2023 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2024 return DAG.getNode(Opc, dl, VT, Op);
2027 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2028 // Implement fcopysign with a fabs and a conditional fneg.
2029 SDValue Tmp0 = Op.getOperand(0);
2030 SDValue Tmp1 = Op.getOperand(1);
2031 DebugLoc dl = Op.getDebugLoc();
2032 EVT VT = Op.getValueType();
2033 EVT SrcVT = Tmp1.getValueType();
2034 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2035 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2036 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2037 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2038 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2041 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
2042 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2043 MFI->setFrameAddressIsTaken(true);
2044 EVT VT = Op.getValueType();
2045 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2046 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2047 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2048 ? ARM::R7 : ARM::R11;
2049 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2051 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2057 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2059 SDValue Dst, SDValue Src,
2060 SDValue Size, unsigned Align,
2062 const Value *DstSV, uint64_t DstSVOff,
2063 const Value *SrcSV, uint64_t SrcSVOff){
2064 // Do repeated 4-byte loads and stores. To be improved.
2065 // This requires 4-byte alignment.
2066 if ((Align & 3) != 0)
2068 // This requires the copy size to be a constant, preferrably
2069 // within a subtarget-specific limit.
2070 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2073 uint64_t SizeVal = ConstantSize->getZExtValue();
2074 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2077 unsigned BytesLeft = SizeVal & 3;
2078 unsigned NumMemOps = SizeVal >> 2;
2079 unsigned EmittedNumMemOps = 0;
2081 unsigned VTSize = 4;
2083 const unsigned MAX_LOADS_IN_LDM = 6;
2084 SDValue TFOps[MAX_LOADS_IN_LDM];
2085 SDValue Loads[MAX_LOADS_IN_LDM];
2086 uint64_t SrcOff = 0, DstOff = 0;
2088 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2089 // same number of stores. The loads and stores will get combined into
2090 // ldm/stm later on.
2091 while (EmittedNumMemOps < NumMemOps) {
2093 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2094 Loads[i] = DAG.getLoad(VT, dl, Chain,
2095 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2096 DAG.getConstant(SrcOff, MVT::i32)),
2097 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2098 TFOps[i] = Loads[i].getValue(1);
2101 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2104 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2105 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2106 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2107 DAG.getConstant(DstOff, MVT::i32)),
2108 DstSV, DstSVOff + DstOff, false, false, 0);
2111 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2113 EmittedNumMemOps += i;
2119 // Issue loads / stores for the trailing (1 - 3) bytes.
2120 unsigned BytesLeftSave = BytesLeft;
2123 if (BytesLeft >= 2) {
2131 Loads[i] = DAG.getLoad(VT, dl, Chain,
2132 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2133 DAG.getConstant(SrcOff, MVT::i32)),
2134 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2135 TFOps[i] = Loads[i].getValue(1);
2138 BytesLeft -= VTSize;
2140 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2143 BytesLeft = BytesLeftSave;
2145 if (BytesLeft >= 2) {
2153 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2154 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2155 DAG.getConstant(DstOff, MVT::i32)),
2156 DstSV, DstSVOff + DstOff, false, false, 0);
2159 BytesLeft -= VTSize;
2161 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2164 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2165 SDValue Op = N->getOperand(0);
2166 DebugLoc dl = N->getDebugLoc();
2167 if (N->getValueType(0) == MVT::f64) {
2168 // Turn i64->f64 into VMOVDRR.
2169 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2170 DAG.getConstant(0, MVT::i32));
2171 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2172 DAG.getConstant(1, MVT::i32));
2173 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2176 // Turn f64->i64 into VMOVRRD.
2177 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2178 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2180 // Merge the pieces into a single i64 value.
2181 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2184 /// getZeroVector - Returns a vector of specified type with all zero elements.
2186 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2187 assert(VT.isVector() && "Expected a vector type");
2189 // Zero vectors are used to represent vector negation and in those cases
2190 // will be implemented with the NEON VNEG instruction. However, VNEG does
2191 // not support i64 elements, so sometimes the zero vectors will need to be
2192 // explicitly constructed. For those cases, and potentially other uses in
2193 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2194 // to their dest type. This ensures they get CSE'd.
2196 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2197 SmallVector<SDValue, 8> Ops;
2200 if (VT.getSizeInBits() == 64) {
2201 Ops.assign(8, Cst); TVT = MVT::v8i8;
2203 Ops.assign(16, Cst); TVT = MVT::v16i8;
2205 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2207 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2210 /// getOnesVector - Returns a vector of specified type with all bits set.
2212 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2213 assert(VT.isVector() && "Expected a vector type");
2215 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2216 // dest type. This ensures they get CSE'd.
2218 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2219 SmallVector<SDValue, 8> Ops;
2222 if (VT.getSizeInBits() == 64) {
2223 Ops.assign(8, Cst); TVT = MVT::v8i8;
2225 Ops.assign(16, Cst); TVT = MVT::v16i8;
2227 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2229 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2232 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2233 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2234 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
2235 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2236 EVT VT = Op.getValueType();
2237 unsigned VTBits = VT.getSizeInBits();
2238 DebugLoc dl = Op.getDebugLoc();
2239 SDValue ShOpLo = Op.getOperand(0);
2240 SDValue ShOpHi = Op.getOperand(1);
2241 SDValue ShAmt = Op.getOperand(2);
2243 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2245 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2247 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2248 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2249 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2250 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2251 DAG.getConstant(VTBits, MVT::i32));
2252 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2253 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2254 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2256 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2257 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2259 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2260 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2263 SDValue Ops[2] = { Lo, Hi };
2264 return DAG.getMergeValues(Ops, 2, dl);
2267 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2268 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2269 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
2270 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2271 EVT VT = Op.getValueType();
2272 unsigned VTBits = VT.getSizeInBits();
2273 DebugLoc dl = Op.getDebugLoc();
2274 SDValue ShOpLo = Op.getOperand(0);
2275 SDValue ShOpHi = Op.getOperand(1);
2276 SDValue ShAmt = Op.getOperand(2);
2279 assert(Op.getOpcode() == ISD::SHL_PARTS);
2280 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2281 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2282 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2283 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2284 DAG.getConstant(VTBits, MVT::i32));
2285 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2286 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2288 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2289 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2290 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2292 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2293 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2296 SDValue Ops[2] = { Lo, Hi };
2297 return DAG.getMergeValues(Ops, 2, dl);
2300 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2301 const ARMSubtarget *ST) {
2302 EVT VT = N->getValueType(0);
2303 DebugLoc dl = N->getDebugLoc();
2305 if (!ST->hasV6T2Ops())
2308 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2309 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2312 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2313 const ARMSubtarget *ST) {
2314 EVT VT = N->getValueType(0);
2315 DebugLoc dl = N->getDebugLoc();
2317 // Lower vector shifts on NEON to use VSHL.
2318 if (VT.isVector()) {
2319 assert(ST->hasNEON() && "unexpected vector shift");
2321 // Left shifts translate directly to the vshiftu intrinsic.
2322 if (N->getOpcode() == ISD::SHL)
2323 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2324 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2325 N->getOperand(0), N->getOperand(1));
2327 assert((N->getOpcode() == ISD::SRA ||
2328 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2330 // NEON uses the same intrinsics for both left and right shifts. For
2331 // right shifts, the shift amounts are negative, so negate the vector of
2333 EVT ShiftVT = N->getOperand(1).getValueType();
2334 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2335 getZeroVector(ShiftVT, DAG, dl),
2337 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2338 Intrinsic::arm_neon_vshifts :
2339 Intrinsic::arm_neon_vshiftu);
2340 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2341 DAG.getConstant(vshiftInt, MVT::i32),
2342 N->getOperand(0), NegatedCount);
2345 // We can get here for a node like i32 = ISD::SHL i32, i64
2349 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2350 "Unknown shift to lower!");
2352 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2353 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2354 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2357 // If we are in thumb mode, we don't have RRX.
2358 if (ST->isThumb1Only()) return SDValue();
2360 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2361 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2362 DAG.getConstant(0, MVT::i32));
2363 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2364 DAG.getConstant(1, MVT::i32));
2366 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2367 // captures the result into a carry flag.
2368 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2369 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2371 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2372 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2374 // Merge the pieces into a single i64 value.
2375 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2378 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2379 SDValue TmpOp0, TmpOp1;
2380 bool Invert = false;
2384 SDValue Op0 = Op.getOperand(0);
2385 SDValue Op1 = Op.getOperand(1);
2386 SDValue CC = Op.getOperand(2);
2387 EVT VT = Op.getValueType();
2388 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2389 DebugLoc dl = Op.getDebugLoc();
2391 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2392 switch (SetCCOpcode) {
2393 default: llvm_unreachable("Illegal FP comparison"); break;
2395 case ISD::SETNE: Invert = true; // Fallthrough
2397 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2399 case ISD::SETLT: Swap = true; // Fallthrough
2401 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2403 case ISD::SETLE: Swap = true; // Fallthrough
2405 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2406 case ISD::SETUGE: Swap = true; // Fallthrough
2407 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2408 case ISD::SETUGT: Swap = true; // Fallthrough
2409 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2410 case ISD::SETUEQ: Invert = true; // Fallthrough
2412 // Expand this to (OLT | OGT).
2416 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2417 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2419 case ISD::SETUO: Invert = true; // Fallthrough
2421 // Expand this to (OLT | OGE).
2425 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2426 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2430 // Integer comparisons.
2431 switch (SetCCOpcode) {
2432 default: llvm_unreachable("Illegal integer comparison"); break;
2433 case ISD::SETNE: Invert = true;
2434 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2435 case ISD::SETLT: Swap = true;
2436 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2437 case ISD::SETLE: Swap = true;
2438 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2439 case ISD::SETULT: Swap = true;
2440 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2441 case ISD::SETULE: Swap = true;
2442 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2445 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2446 if (Opc == ARMISD::VCEQ) {
2449 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2451 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2454 // Ignore bitconvert.
2455 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2456 AndOp = AndOp.getOperand(0);
2458 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2460 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2461 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2468 std::swap(Op0, Op1);
2470 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2473 Result = DAG.getNOT(dl, Result, VT);
2478 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2479 /// VMOV instruction, and if so, return the constant being splatted.
2480 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2481 unsigned SplatBitSize, SelectionDAG &DAG) {
2482 switch (SplatBitSize) {
2484 // Any 1-byte value is OK.
2485 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2486 return DAG.getTargetConstant(SplatBits, MVT::i8);
2489 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2490 if ((SplatBits & ~0xff) == 0 ||
2491 (SplatBits & ~0xff00) == 0)
2492 return DAG.getTargetConstant(SplatBits, MVT::i16);
2496 // NEON's 32-bit VMOV supports splat values where:
2497 // * only one byte is nonzero, or
2498 // * the least significant byte is 0xff and the second byte is nonzero, or
2499 // * the least significant 2 bytes are 0xff and the third is nonzero.
2500 if ((SplatBits & ~0xff) == 0 ||
2501 (SplatBits & ~0xff00) == 0 ||
2502 (SplatBits & ~0xff0000) == 0 ||
2503 (SplatBits & ~0xff000000) == 0)
2504 return DAG.getTargetConstant(SplatBits, MVT::i32);
2506 if ((SplatBits & ~0xffff) == 0 &&
2507 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2508 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2510 if ((SplatBits & ~0xffffff) == 0 &&
2511 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2512 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2514 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2515 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2516 // VMOV.I32. A (very) minor optimization would be to replicate the value
2517 // and fall through here to test for a valid 64-bit splat. But, then the
2518 // caller would also need to check and handle the change in size.
2522 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2523 uint64_t BitMask = 0xff;
2525 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2526 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2528 else if ((SplatBits & BitMask) != 0)
2532 return DAG.getTargetConstant(Val, MVT::i64);
2536 llvm_unreachable("unexpected size for isVMOVSplat");
2543 /// getVMOVImm - If this is a build_vector of constants which can be
2544 /// formed by using a VMOV instruction of the specified element size,
2545 /// return the constant being splatted. The ByteSize field indicates the
2546 /// number of bytes of each element [1248].
2547 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2548 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2549 APInt SplatBits, SplatUndef;
2550 unsigned SplatBitSize;
2552 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2553 HasAnyUndefs, ByteSize * 8))
2556 if (SplatBitSize > ByteSize * 8)
2559 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2563 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2564 bool &ReverseVEXT, unsigned &Imm) {
2565 unsigned NumElts = VT.getVectorNumElements();
2566 ReverseVEXT = false;
2569 // If this is a VEXT shuffle, the immediate value is the index of the first
2570 // element. The other shuffle indices must be the successive elements after
2572 unsigned ExpectedElt = Imm;
2573 for (unsigned i = 1; i < NumElts; ++i) {
2574 // Increment the expected index. If it wraps around, it may still be
2575 // a VEXT but the source vectors must be swapped.
2577 if (ExpectedElt == NumElts * 2) {
2582 if (ExpectedElt != static_cast<unsigned>(M[i]))
2586 // Adjust the index value if the source operands will be swapped.
2593 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2594 /// instruction with the specified blocksize. (The order of the elements
2595 /// within each block of the vector is reversed.)
2596 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2597 unsigned BlockSize) {
2598 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2599 "Only possible block sizes for VREV are: 16, 32, 64");
2601 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2605 unsigned NumElts = VT.getVectorNumElements();
2606 unsigned BlockElts = M[0] + 1;
2608 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2611 for (unsigned i = 0; i < NumElts; ++i) {
2612 if ((unsigned) M[i] !=
2613 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2620 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2621 unsigned &WhichResult) {
2622 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2626 unsigned NumElts = VT.getVectorNumElements();
2627 WhichResult = (M[0] == 0 ? 0 : 1);
2628 for (unsigned i = 0; i < NumElts; i += 2) {
2629 if ((unsigned) M[i] != i + WhichResult ||
2630 (unsigned) M[i+1] != i + NumElts + WhichResult)
2636 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2637 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2638 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2639 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2640 unsigned &WhichResult) {
2641 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2645 unsigned NumElts = VT.getVectorNumElements();
2646 WhichResult = (M[0] == 0 ? 0 : 1);
2647 for (unsigned i = 0; i < NumElts; i += 2) {
2648 if ((unsigned) M[i] != i + WhichResult ||
2649 (unsigned) M[i+1] != i + WhichResult)
2655 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2656 unsigned &WhichResult) {
2657 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2661 unsigned NumElts = VT.getVectorNumElements();
2662 WhichResult = (M[0] == 0 ? 0 : 1);
2663 for (unsigned i = 0; i != NumElts; ++i) {
2664 if ((unsigned) M[i] != 2 * i + WhichResult)
2668 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2669 if (VT.is64BitVector() && EltSz == 32)
2675 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2676 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2677 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2678 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2679 unsigned &WhichResult) {
2680 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2684 unsigned Half = VT.getVectorNumElements() / 2;
2685 WhichResult = (M[0] == 0 ? 0 : 1);
2686 for (unsigned j = 0; j != 2; ++j) {
2687 unsigned Idx = WhichResult;
2688 for (unsigned i = 0; i != Half; ++i) {
2689 if ((unsigned) M[i + j * Half] != Idx)
2695 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2696 if (VT.is64BitVector() && EltSz == 32)
2702 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2703 unsigned &WhichResult) {
2704 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2708 unsigned NumElts = VT.getVectorNumElements();
2709 WhichResult = (M[0] == 0 ? 0 : 1);
2710 unsigned Idx = WhichResult * NumElts / 2;
2711 for (unsigned i = 0; i != NumElts; i += 2) {
2712 if ((unsigned) M[i] != Idx ||
2713 (unsigned) M[i+1] != Idx + NumElts)
2718 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2719 if (VT.is64BitVector() && EltSz == 32)
2725 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2726 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2727 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2728 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2729 unsigned &WhichResult) {
2730 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2734 unsigned NumElts = VT.getVectorNumElements();
2735 WhichResult = (M[0] == 0 ? 0 : 1);
2736 unsigned Idx = WhichResult * NumElts / 2;
2737 for (unsigned i = 0; i != NumElts; i += 2) {
2738 if ((unsigned) M[i] != Idx ||
2739 (unsigned) M[i+1] != Idx)
2744 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2745 if (VT.is64BitVector() && EltSz == 32)
2752 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2753 // Canonicalize all-zeros and all-ones vectors.
2754 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2755 if (ConstVal->isNullValue())
2756 return getZeroVector(VT, DAG, dl);
2757 if (ConstVal->isAllOnesValue())
2758 return getOnesVector(VT, DAG, dl);
2761 if (VT.is64BitVector()) {
2762 switch (Val.getValueType().getSizeInBits()) {
2763 case 8: CanonicalVT = MVT::v8i8; break;
2764 case 16: CanonicalVT = MVT::v4i16; break;
2765 case 32: CanonicalVT = MVT::v2i32; break;
2766 case 64: CanonicalVT = MVT::v1i64; break;
2767 default: llvm_unreachable("unexpected splat element type"); break;
2770 assert(VT.is128BitVector() && "unknown splat vector size");
2771 switch (Val.getValueType().getSizeInBits()) {
2772 case 8: CanonicalVT = MVT::v16i8; break;
2773 case 16: CanonicalVT = MVT::v8i16; break;
2774 case 32: CanonicalVT = MVT::v4i32; break;
2775 case 64: CanonicalVT = MVT::v2i64; break;
2776 default: llvm_unreachable("unexpected splat element type"); break;
2780 // Build a canonical splat for this value.
2781 SmallVector<SDValue, 8> Ops;
2782 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2783 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2785 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2788 // If this is a case we can't handle, return null and let the default
2789 // expansion code take care of it.
2790 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2791 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2792 DebugLoc dl = Op.getDebugLoc();
2793 EVT VT = Op.getValueType();
2795 APInt SplatBits, SplatUndef;
2796 unsigned SplatBitSize;
2798 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2799 if (SplatBitSize <= 64) {
2800 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2801 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2803 return BuildSplat(Val, VT, DAG, dl);
2807 // If there are only 2 elements in a 128-bit vector, insert them into an
2808 // undef vector. This handles the common case for 128-bit vector argument
2809 // passing, where the insertions should be translated to subreg accesses
2810 // with no real instructions.
2811 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2812 SDValue Val = DAG.getUNDEF(VT);
2813 SDValue Op0 = Op.getOperand(0);
2814 SDValue Op1 = Op.getOperand(1);
2815 if (Op0.getOpcode() != ISD::UNDEF)
2816 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2817 DAG.getIntPtrConstant(0));
2818 if (Op1.getOpcode() != ISD::UNDEF)
2819 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2820 DAG.getIntPtrConstant(1));
2827 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2828 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2829 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2830 /// are assumed to be legal.
2832 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2834 if (VT.getVectorNumElements() == 4 &&
2835 (VT.is128BitVector() || VT.is64BitVector())) {
2836 unsigned PFIndexes[4];
2837 for (unsigned i = 0; i != 4; ++i) {
2841 PFIndexes[i] = M[i];
2844 // Compute the index in the perfect shuffle table.
2845 unsigned PFTableIndex =
2846 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2847 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2848 unsigned Cost = (PFEntry >> 30);
2855 unsigned Imm, WhichResult;
2857 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2858 isVREVMask(M, VT, 64) ||
2859 isVREVMask(M, VT, 32) ||
2860 isVREVMask(M, VT, 16) ||
2861 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2862 isVTRNMask(M, VT, WhichResult) ||
2863 isVUZPMask(M, VT, WhichResult) ||
2864 isVZIPMask(M, VT, WhichResult) ||
2865 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2866 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2867 isVZIP_v_undef_Mask(M, VT, WhichResult));
2870 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2871 /// the specified operations to build the shuffle.
2872 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2873 SDValue RHS, SelectionDAG &DAG,
2875 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2876 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2877 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2880 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2889 OP_VUZPL, // VUZP, left result
2890 OP_VUZPR, // VUZP, right result
2891 OP_VZIPL, // VZIP, left result
2892 OP_VZIPR, // VZIP, right result
2893 OP_VTRNL, // VTRN, left result
2894 OP_VTRNR // VTRN, right result
2897 if (OpNum == OP_COPY) {
2898 if (LHSID == (1*9+2)*9+3) return LHS;
2899 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2903 SDValue OpLHS, OpRHS;
2904 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2905 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2906 EVT VT = OpLHS.getValueType();
2909 default: llvm_unreachable("Unknown shuffle opcode!");
2911 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2916 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2917 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2921 return DAG.getNode(ARMISD::VEXT, dl, VT,
2923 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2926 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2927 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2930 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2931 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2934 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2935 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2939 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2940 SDValue V1 = Op.getOperand(0);
2941 SDValue V2 = Op.getOperand(1);
2942 DebugLoc dl = Op.getDebugLoc();
2943 EVT VT = Op.getValueType();
2944 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2945 SmallVector<int, 8> ShuffleMask;
2947 // Convert shuffles that are directly supported on NEON to target-specific
2948 // DAG nodes, instead of keeping them as shuffles and matching them again
2949 // during code selection. This is more efficient and avoids the possibility
2950 // of inconsistencies between legalization and selection.
2951 // FIXME: floating-point vectors should be canonicalized to integer vectors
2952 // of the same time so that they get CSEd properly.
2953 SVN->getMask(ShuffleMask);
2955 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2956 int Lane = SVN->getSplatIndex();
2957 // If this is undef splat, generate it via "just" vdup, if possible.
2958 if (Lane == -1) Lane = 0;
2960 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2961 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2963 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2964 DAG.getConstant(Lane, MVT::i32));
2969 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2972 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2973 DAG.getConstant(Imm, MVT::i32));
2976 if (isVREVMask(ShuffleMask, VT, 64))
2977 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2978 if (isVREVMask(ShuffleMask, VT, 32))
2979 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2980 if (isVREVMask(ShuffleMask, VT, 16))
2981 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2983 // Check for Neon shuffles that modify both input vectors in place.
2984 // If both results are used, i.e., if there are two shuffles with the same
2985 // source operands and with masks corresponding to both results of one of
2986 // these operations, DAG memoization will ensure that a single node is
2987 // used for both shuffles.
2988 unsigned WhichResult;
2989 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2990 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2991 V1, V2).getValue(WhichResult);
2992 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2993 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2994 V1, V2).getValue(WhichResult);
2995 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2996 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2997 V1, V2).getValue(WhichResult);
2999 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3000 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3001 V1, V1).getValue(WhichResult);
3002 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3003 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3004 V1, V1).getValue(WhichResult);
3005 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3006 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3007 V1, V1).getValue(WhichResult);
3009 // If the shuffle is not directly supported and it has 4 elements, use
3010 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3011 if (VT.getVectorNumElements() == 4 &&
3012 (VT.is128BitVector() || VT.is64BitVector())) {
3013 unsigned PFIndexes[4];
3014 for (unsigned i = 0; i != 4; ++i) {
3015 if (ShuffleMask[i] < 0)
3018 PFIndexes[i] = ShuffleMask[i];
3021 // Compute the index in the perfect shuffle table.
3022 unsigned PFTableIndex =
3023 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3025 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3026 unsigned Cost = (PFEntry >> 30);
3029 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3035 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3036 EVT VT = Op.getValueType();
3037 DebugLoc dl = Op.getDebugLoc();
3038 SDValue Vec = Op.getOperand(0);
3039 SDValue Lane = Op.getOperand(1);
3040 assert(VT == MVT::i32 &&
3041 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3042 "unexpected type for custom-lowering vector extract");
3043 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3046 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3047 // The only time a CONCAT_VECTORS operation can have legal types is when
3048 // two 64-bit vectors are concatenated to a 128-bit vector.
3049 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3050 "unexpected CONCAT_VECTORS");
3051 DebugLoc dl = Op.getDebugLoc();
3052 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3053 SDValue Op0 = Op.getOperand(0);
3054 SDValue Op1 = Op.getOperand(1);
3055 if (Op0.getOpcode() != ISD::UNDEF)
3056 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3057 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3058 DAG.getIntPtrConstant(0));
3059 if (Op1.getOpcode() != ISD::UNDEF)
3060 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3061 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3062 DAG.getIntPtrConstant(1));
3063 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3066 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3067 switch (Op.getOpcode()) {
3068 default: llvm_unreachable("Don't know how to custom lower this!");
3069 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3070 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3071 case ISD::GlobalAddress:
3072 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3073 LowerGlobalAddressELF(Op, DAG);
3074 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3075 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3076 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3077 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3078 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3079 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
3080 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3081 case ISD::FP16_TO_FP32:
3082 case ISD::SINT_TO_FP:
3083 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3084 case ISD::FP32_TO_FP16:
3085 case ISD::FP_TO_SINT:
3086 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3087 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3088 case ISD::RETURNADDR: break;
3089 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3090 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3091 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3093 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3096 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3097 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3098 case ISD::SRL_PARTS:
3099 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3100 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3101 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3102 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3103 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3104 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3105 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3110 /// ReplaceNodeResults - Replace the results of node with an illegal result
3111 /// type with new values built out of custom code.
3112 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3113 SmallVectorImpl<SDValue>&Results,
3114 SelectionDAG &DAG) {
3115 switch (N->getOpcode()) {
3117 llvm_unreachable("Don't know how to custom expand this!");
3119 case ISD::BIT_CONVERT:
3120 Results.push_back(ExpandBIT_CONVERT(N, DAG));
3124 SDValue Res = LowerShift(N, DAG, Subtarget);
3126 Results.push_back(Res);
3132 //===----------------------------------------------------------------------===//
3133 // ARM Scheduler Hooks
3134 //===----------------------------------------------------------------------===//
3137 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3138 MachineBasicBlock *BB,
3139 unsigned Size) const {
3140 unsigned dest = MI->getOperand(0).getReg();
3141 unsigned ptr = MI->getOperand(1).getReg();
3142 unsigned oldval = MI->getOperand(2).getReg();
3143 unsigned newval = MI->getOperand(3).getReg();
3144 unsigned scratch = BB->getParent()->getRegInfo()
3145 .createVirtualRegister(ARM::GPRRegisterClass);
3146 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3147 DebugLoc dl = MI->getDebugLoc();
3148 bool isThumb2 = Subtarget->isThumb2();
3150 unsigned ldrOpc, strOpc;
3152 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3154 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3155 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3158 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3159 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3162 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3163 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3167 MachineFunction *MF = BB->getParent();
3168 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3169 MachineFunction::iterator It = BB;
3170 ++It; // insert the new blocks after the current block
3172 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3173 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3174 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3175 MF->insert(It, loop1MBB);
3176 MF->insert(It, loop2MBB);
3177 MF->insert(It, exitMBB);
3178 exitMBB->transferSuccessors(BB);
3182 // fallthrough --> loop1MBB
3183 BB->addSuccessor(loop1MBB);
3186 // ldrex dest, [ptr]
3190 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3191 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3192 .addReg(dest).addReg(oldval));
3193 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3194 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3195 BB->addSuccessor(loop2MBB);
3196 BB->addSuccessor(exitMBB);
3199 // strex scratch, newval, [ptr]
3203 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3205 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3206 .addReg(scratch).addImm(0));
3207 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3208 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3209 BB->addSuccessor(loop1MBB);
3210 BB->addSuccessor(exitMBB);
3216 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3222 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3223 unsigned Size, unsigned BinOpcode) const {
3224 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3225 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3227 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3228 MachineFunction *MF = BB->getParent();
3229 MachineFunction::iterator It = BB;
3232 unsigned dest = MI->getOperand(0).getReg();
3233 unsigned ptr = MI->getOperand(1).getReg();
3234 unsigned incr = MI->getOperand(2).getReg();
3235 DebugLoc dl = MI->getDebugLoc();
3237 bool isThumb2 = Subtarget->isThumb2();
3238 unsigned ldrOpc, strOpc;
3240 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3242 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3243 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3246 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3247 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3250 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3251 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3255 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3256 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3257 MF->insert(It, loopMBB);
3258 MF->insert(It, exitMBB);
3259 exitMBB->transferSuccessors(BB);
3261 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3262 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3263 unsigned scratch2 = (!BinOpcode) ? incr :
3264 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3268 // fallthrough --> loopMBB
3269 BB->addSuccessor(loopMBB);
3273 // <binop> scratch2, dest, incr
3274 // strex scratch, scratch2, ptr
3277 // fallthrough --> exitMBB
3279 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3281 // operand order needs to go the other way for NAND
3282 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3283 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3284 addReg(incr).addReg(dest)).addReg(0);
3286 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3287 addReg(dest).addReg(incr)).addReg(0);
3290 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3292 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3293 .addReg(scratch).addImm(0));
3294 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3295 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3297 BB->addSuccessor(loopMBB);
3298 BB->addSuccessor(exitMBB);
3304 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3310 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3311 MachineBasicBlock *BB,
3312 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
3313 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3314 DebugLoc dl = MI->getDebugLoc();
3315 bool isThumb2 = Subtarget->isThumb2();
3316 switch (MI->getOpcode()) {
3319 llvm_unreachable("Unexpected instr type to insert");
3321 case ARM::ATOMIC_LOAD_ADD_I8:
3322 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3323 case ARM::ATOMIC_LOAD_ADD_I16:
3324 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3325 case ARM::ATOMIC_LOAD_ADD_I32:
3326 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3328 case ARM::ATOMIC_LOAD_AND_I8:
3329 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3330 case ARM::ATOMIC_LOAD_AND_I16:
3331 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3332 case ARM::ATOMIC_LOAD_AND_I32:
3333 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3335 case ARM::ATOMIC_LOAD_OR_I8:
3336 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3337 case ARM::ATOMIC_LOAD_OR_I16:
3338 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3339 case ARM::ATOMIC_LOAD_OR_I32:
3340 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3342 case ARM::ATOMIC_LOAD_XOR_I8:
3343 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3344 case ARM::ATOMIC_LOAD_XOR_I16:
3345 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3346 case ARM::ATOMIC_LOAD_XOR_I32:
3347 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3349 case ARM::ATOMIC_LOAD_NAND_I8:
3350 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3351 case ARM::ATOMIC_LOAD_NAND_I16:
3352 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3353 case ARM::ATOMIC_LOAD_NAND_I32:
3354 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3356 case ARM::ATOMIC_LOAD_SUB_I8:
3357 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3358 case ARM::ATOMIC_LOAD_SUB_I16:
3359 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3360 case ARM::ATOMIC_LOAD_SUB_I32:
3361 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3363 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3364 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3365 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3367 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3368 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3369 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3371 case ARM::tMOVCCr_pseudo: {
3372 // To "insert" a SELECT_CC instruction, we actually have to insert the
3373 // diamond control-flow pattern. The incoming instruction knows the
3374 // destination vreg to set, the condition code register to branch on, the
3375 // true/false values to select between, and a branch opcode to use.
3376 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3377 MachineFunction::iterator It = BB;
3383 // cmpTY ccX, r1, r2
3385 // fallthrough --> copy0MBB
3386 MachineBasicBlock *thisMBB = BB;
3387 MachineFunction *F = BB->getParent();
3388 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3389 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3390 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3391 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3392 F->insert(It, copy0MBB);
3393 F->insert(It, sinkMBB);
3394 // Update machine-CFG edges by first adding all successors of the current
3395 // block to the new block which will contain the Phi node for the select.
3396 // Also inform sdisel of the edge changes.
3397 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3398 E = BB->succ_end(); I != E; ++I) {
3399 EM->insert(std::make_pair(*I, sinkMBB));
3400 sinkMBB->addSuccessor(*I);
3402 // Next, remove all successors of the current block, and add the true
3403 // and fallthrough blocks as its successors.
3404 while (!BB->succ_empty())
3405 BB->removeSuccessor(BB->succ_begin());
3406 BB->addSuccessor(copy0MBB);
3407 BB->addSuccessor(sinkMBB);
3410 // %FalseValue = ...
3411 // # fallthrough to sinkMBB
3414 // Update machine-CFG edges
3415 BB->addSuccessor(sinkMBB);
3418 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3421 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3422 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3423 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3425 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3432 case ARM::t2SUBrSPi_:
3433 case ARM::t2SUBrSPi12_:
3434 case ARM::t2SUBrSPs_: {
3435 MachineFunction *MF = BB->getParent();
3436 unsigned DstReg = MI->getOperand(0).getReg();
3437 unsigned SrcReg = MI->getOperand(1).getReg();
3438 bool DstIsDead = MI->getOperand(0).isDead();
3439 bool SrcIsKill = MI->getOperand(1).isKill();
3441 if (SrcReg != ARM::SP) {
3442 // Copy the source to SP from virtual register.
3443 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3444 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3445 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3446 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3447 .addReg(SrcReg, getKillRegState(SrcIsKill));
3451 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3452 switch (MI->getOpcode()) {
3454 llvm_unreachable("Unexpected pseudo instruction!");
3460 OpOpc = ARM::tADDspr;
3463 OpOpc = ARM::tSUBspi;
3465 case ARM::t2SUBrSPi_:
3466 OpOpc = ARM::t2SUBrSPi;
3467 NeedPred = true; NeedCC = true;
3469 case ARM::t2SUBrSPi12_:
3470 OpOpc = ARM::t2SUBrSPi12;
3473 case ARM::t2SUBrSPs_:
3474 OpOpc = ARM::t2SUBrSPs;
3475 NeedPred = true; NeedCC = true; NeedOp3 = true;
3478 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3479 if (OpOpc == ARM::tAND)
3480 AddDefaultT1CC(MIB);
3481 MIB.addReg(ARM::SP);
3482 MIB.addOperand(MI->getOperand(2));
3484 MIB.addOperand(MI->getOperand(3));
3486 AddDefaultPred(MIB);
3490 // Copy the result from SP to virtual register.
3491 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3492 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3493 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3494 BuildMI(BB, dl, TII->get(CopyOpc))
3495 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3497 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3503 //===----------------------------------------------------------------------===//
3504 // ARM Optimization Hooks
3505 //===----------------------------------------------------------------------===//
3508 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3509 TargetLowering::DAGCombinerInfo &DCI) {
3510 SelectionDAG &DAG = DCI.DAG;
3511 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3512 EVT VT = N->getValueType(0);
3513 unsigned Opc = N->getOpcode();
3514 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3515 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3516 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3517 ISD::CondCode CC = ISD::SETCC_INVALID;
3520 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3522 SDValue CCOp = Slct.getOperand(0);
3523 if (CCOp.getOpcode() == ISD::SETCC)
3524 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3527 bool DoXform = false;
3529 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3532 if (LHS.getOpcode() == ISD::Constant &&
3533 cast<ConstantSDNode>(LHS)->isNullValue()) {
3535 } else if (CC != ISD::SETCC_INVALID &&
3536 RHS.getOpcode() == ISD::Constant &&
3537 cast<ConstantSDNode>(RHS)->isNullValue()) {
3538 std::swap(LHS, RHS);
3539 SDValue Op0 = Slct.getOperand(0);
3540 EVT OpVT = isSlctCC ? Op0.getValueType() :
3541 Op0.getOperand(0).getValueType();
3542 bool isInt = OpVT.isInteger();
3543 CC = ISD::getSetCCInverse(CC, isInt);
3545 if (!TLI.isCondCodeLegal(CC, OpVT))
3546 return SDValue(); // Inverse operator isn't legal.
3553 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3555 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3556 Slct.getOperand(0), Slct.getOperand(1), CC);
3557 SDValue CCOp = Slct.getOperand(0);
3559 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3560 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3561 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3562 CCOp, OtherOp, Result);
3567 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3568 static SDValue PerformADDCombine(SDNode *N,
3569 TargetLowering::DAGCombinerInfo &DCI) {
3570 // added by evan in r37685 with no testcase.
3571 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3573 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3574 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3575 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3576 if (Result.getNode()) return Result;
3578 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3579 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3580 if (Result.getNode()) return Result;
3586 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3587 static SDValue PerformSUBCombine(SDNode *N,
3588 TargetLowering::DAGCombinerInfo &DCI) {
3589 // added by evan in r37685 with no testcase.
3590 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3592 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3593 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3594 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3595 if (Result.getNode()) return Result;
3601 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3602 /// ARMISD::VMOVRRD.
3603 static SDValue PerformVMOVRRDCombine(SDNode *N,
3604 TargetLowering::DAGCombinerInfo &DCI) {
3605 // fmrrd(fmdrr x, y) -> x,y
3606 SDValue InDouble = N->getOperand(0);
3607 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3608 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3612 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3613 /// operand of a vector shift operation, where all the elements of the
3614 /// build_vector must have the same constant integer value.
3615 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3616 // Ignore bit_converts.
3617 while (Op.getOpcode() == ISD::BIT_CONVERT)
3618 Op = Op.getOperand(0);
3619 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3620 APInt SplatBits, SplatUndef;
3621 unsigned SplatBitSize;
3623 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3624 HasAnyUndefs, ElementBits) ||
3625 SplatBitSize > ElementBits)
3627 Cnt = SplatBits.getSExtValue();
3631 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3632 /// operand of a vector shift left operation. That value must be in the range:
3633 /// 0 <= Value < ElementBits for a left shift; or
3634 /// 0 <= Value <= ElementBits for a long left shift.
3635 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3636 assert(VT.isVector() && "vector shift count is not a vector type");
3637 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3638 if (! getVShiftImm(Op, ElementBits, Cnt))
3640 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3643 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3644 /// operand of a vector shift right operation. For a shift opcode, the value
3645 /// is positive, but for an intrinsic the value count must be negative. The
3646 /// absolute value must be in the range:
3647 /// 1 <= |Value| <= ElementBits for a right shift; or
3648 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3649 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3651 assert(VT.isVector() && "vector shift count is not a vector type");
3652 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3653 if (! getVShiftImm(Op, ElementBits, Cnt))
3657 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3660 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3661 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3662 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3665 // Don't do anything for most intrinsics.
3668 // Vector shifts: check for immediate versions and lower them.
3669 // Note: This is done during DAG combining instead of DAG legalizing because
3670 // the build_vectors for 64-bit vector element shift counts are generally
3671 // not legal, and it is hard to see their values after they get legalized to
3672 // loads from a constant pool.
3673 case Intrinsic::arm_neon_vshifts:
3674 case Intrinsic::arm_neon_vshiftu:
3675 case Intrinsic::arm_neon_vshiftls:
3676 case Intrinsic::arm_neon_vshiftlu:
3677 case Intrinsic::arm_neon_vshiftn:
3678 case Intrinsic::arm_neon_vrshifts:
3679 case Intrinsic::arm_neon_vrshiftu:
3680 case Intrinsic::arm_neon_vrshiftn:
3681 case Intrinsic::arm_neon_vqshifts:
3682 case Intrinsic::arm_neon_vqshiftu:
3683 case Intrinsic::arm_neon_vqshiftsu:
3684 case Intrinsic::arm_neon_vqshiftns:
3685 case Intrinsic::arm_neon_vqshiftnu:
3686 case Intrinsic::arm_neon_vqshiftnsu:
3687 case Intrinsic::arm_neon_vqrshiftns:
3688 case Intrinsic::arm_neon_vqrshiftnu:
3689 case Intrinsic::arm_neon_vqrshiftnsu: {
3690 EVT VT = N->getOperand(1).getValueType();
3692 unsigned VShiftOpc = 0;
3695 case Intrinsic::arm_neon_vshifts:
3696 case Intrinsic::arm_neon_vshiftu:
3697 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3698 VShiftOpc = ARMISD::VSHL;
3701 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3702 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3703 ARMISD::VSHRs : ARMISD::VSHRu);
3708 case Intrinsic::arm_neon_vshiftls:
3709 case Intrinsic::arm_neon_vshiftlu:
3710 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3712 llvm_unreachable("invalid shift count for vshll intrinsic");
3714 case Intrinsic::arm_neon_vrshifts:
3715 case Intrinsic::arm_neon_vrshiftu:
3716 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3720 case Intrinsic::arm_neon_vqshifts:
3721 case Intrinsic::arm_neon_vqshiftu:
3722 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3726 case Intrinsic::arm_neon_vqshiftsu:
3727 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3729 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3731 case Intrinsic::arm_neon_vshiftn:
3732 case Intrinsic::arm_neon_vrshiftn:
3733 case Intrinsic::arm_neon_vqshiftns:
3734 case Intrinsic::arm_neon_vqshiftnu:
3735 case Intrinsic::arm_neon_vqshiftnsu:
3736 case Intrinsic::arm_neon_vqrshiftns:
3737 case Intrinsic::arm_neon_vqrshiftnu:
3738 case Intrinsic::arm_neon_vqrshiftnsu:
3739 // Narrowing shifts require an immediate right shift.
3740 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3742 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3745 llvm_unreachable("unhandled vector shift");
3749 case Intrinsic::arm_neon_vshifts:
3750 case Intrinsic::arm_neon_vshiftu:
3751 // Opcode already set above.
3753 case Intrinsic::arm_neon_vshiftls:
3754 case Intrinsic::arm_neon_vshiftlu:
3755 if (Cnt == VT.getVectorElementType().getSizeInBits())
3756 VShiftOpc = ARMISD::VSHLLi;
3758 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3759 ARMISD::VSHLLs : ARMISD::VSHLLu);
3761 case Intrinsic::arm_neon_vshiftn:
3762 VShiftOpc = ARMISD::VSHRN; break;
3763 case Intrinsic::arm_neon_vrshifts:
3764 VShiftOpc = ARMISD::VRSHRs; break;
3765 case Intrinsic::arm_neon_vrshiftu:
3766 VShiftOpc = ARMISD::VRSHRu; break;
3767 case Intrinsic::arm_neon_vrshiftn:
3768 VShiftOpc = ARMISD::VRSHRN; break;
3769 case Intrinsic::arm_neon_vqshifts:
3770 VShiftOpc = ARMISD::VQSHLs; break;
3771 case Intrinsic::arm_neon_vqshiftu:
3772 VShiftOpc = ARMISD::VQSHLu; break;
3773 case Intrinsic::arm_neon_vqshiftsu:
3774 VShiftOpc = ARMISD::VQSHLsu; break;
3775 case Intrinsic::arm_neon_vqshiftns:
3776 VShiftOpc = ARMISD::VQSHRNs; break;
3777 case Intrinsic::arm_neon_vqshiftnu:
3778 VShiftOpc = ARMISD::VQSHRNu; break;
3779 case Intrinsic::arm_neon_vqshiftnsu:
3780 VShiftOpc = ARMISD::VQSHRNsu; break;
3781 case Intrinsic::arm_neon_vqrshiftns:
3782 VShiftOpc = ARMISD::VQRSHRNs; break;
3783 case Intrinsic::arm_neon_vqrshiftnu:
3784 VShiftOpc = ARMISD::VQRSHRNu; break;
3785 case Intrinsic::arm_neon_vqrshiftnsu:
3786 VShiftOpc = ARMISD::VQRSHRNsu; break;
3789 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3790 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3793 case Intrinsic::arm_neon_vshiftins: {
3794 EVT VT = N->getOperand(1).getValueType();
3796 unsigned VShiftOpc = 0;
3798 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3799 VShiftOpc = ARMISD::VSLI;
3800 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3801 VShiftOpc = ARMISD::VSRI;
3803 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3806 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3807 N->getOperand(1), N->getOperand(2),
3808 DAG.getConstant(Cnt, MVT::i32));
3811 case Intrinsic::arm_neon_vqrshifts:
3812 case Intrinsic::arm_neon_vqrshiftu:
3813 // No immediate versions of these to check for.
3820 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3821 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3822 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3823 /// vector element shift counts are generally not legal, and it is hard to see
3824 /// their values after they get legalized to loads from a constant pool.
3825 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3826 const ARMSubtarget *ST) {
3827 EVT VT = N->getValueType(0);
3829 // Nothing to be done for scalar shifts.
3830 if (! VT.isVector())
3833 assert(ST->hasNEON() && "unexpected vector shift");
3836 switch (N->getOpcode()) {
3837 default: llvm_unreachable("unexpected shift opcode");
3840 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3841 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3842 DAG.getConstant(Cnt, MVT::i32));
3847 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3848 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3849 ARMISD::VSHRs : ARMISD::VSHRu);
3850 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3851 DAG.getConstant(Cnt, MVT::i32));
3857 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3858 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3859 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3860 const ARMSubtarget *ST) {
3861 SDValue N0 = N->getOperand(0);
3863 // Check for sign- and zero-extensions of vector extract operations of 8-
3864 // and 16-bit vector elements. NEON supports these directly. They are
3865 // handled during DAG combining because type legalization will promote them
3866 // to 32-bit types and it is messy to recognize the operations after that.
3867 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3868 SDValue Vec = N0.getOperand(0);
3869 SDValue Lane = N0.getOperand(1);
3870 EVT VT = N->getValueType(0);
3871 EVT EltVT = N0.getValueType();
3872 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3874 if (VT == MVT::i32 &&
3875 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3876 TLI.isTypeLegal(Vec.getValueType())) {
3879 switch (N->getOpcode()) {
3880 default: llvm_unreachable("unexpected opcode");
3881 case ISD::SIGN_EXTEND:
3882 Opc = ARMISD::VGETLANEs;
3884 case ISD::ZERO_EXTEND:
3885 case ISD::ANY_EXTEND:
3886 Opc = ARMISD::VGETLANEu;
3889 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3896 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3897 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3898 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3899 const ARMSubtarget *ST) {
3900 // If the target supports NEON, try to use vmax/vmin instructions for f32
3901 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3902 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3903 // a NaN; only do the transformation when it matches that behavior.
3905 // For now only do this when using NEON for FP operations; if using VFP, it
3906 // is not obvious that the benefit outweighs the cost of switching to the
3908 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3909 N->getValueType(0) != MVT::f32)
3912 SDValue CondLHS = N->getOperand(0);
3913 SDValue CondRHS = N->getOperand(1);
3914 SDValue LHS = N->getOperand(2);
3915 SDValue RHS = N->getOperand(3);
3916 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3918 unsigned Opcode = 0;
3920 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3921 IsReversed = false; // x CC y ? x : y
3922 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3923 IsReversed = true ; // x CC y ? y : x
3937 // If LHS is NaN, an ordered comparison will be false and the result will
3938 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3939 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3940 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3941 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3943 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3944 // will return -0, so vmin can only be used for unsafe math or if one of
3945 // the operands is known to be nonzero.
3946 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3948 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3950 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3959 // If LHS is NaN, an ordered comparison will be false and the result will
3960 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3961 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3962 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3963 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3965 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3966 // will return +0, so vmax can only be used for unsafe math or if one of
3967 // the operands is known to be nonzero.
3968 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3970 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3972 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3978 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3981 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3982 DAGCombinerInfo &DCI) const {
3983 switch (N->getOpcode()) {
3985 case ISD::ADD: return PerformADDCombine(N, DCI);
3986 case ISD::SUB: return PerformSUBCombine(N, DCI);
3987 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3988 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
3991 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
3992 case ISD::SIGN_EXTEND:
3993 case ISD::ZERO_EXTEND:
3994 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
3995 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4000 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4001 if (!Subtarget->hasV6Ops())
4002 // Pre-v6 does not support unaligned mem access.
4005 // v6+ may or may not support unaligned mem access depending on the system
4007 // FIXME: This is pretty conservative. Should we provide cmdline option to
4008 // control the behaviour?
4009 if (!Subtarget->isTargetDarwin())
4013 switch (VT.getSimpleVT().SimpleTy) {
4020 // FIXME: VLD1 etc with standard alignment is legal.
4024 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4029 switch (VT.getSimpleVT().SimpleTy) {
4030 default: return false;
4045 if ((V & (Scale - 1)) != 0)
4048 return V == (V & ((1LL << 5) - 1));
4051 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4052 const ARMSubtarget *Subtarget) {
4059 switch (VT.getSimpleVT().SimpleTy) {
4060 default: return false;
4065 // + imm12 or - imm8
4067 return V == (V & ((1LL << 8) - 1));
4068 return V == (V & ((1LL << 12) - 1));
4071 // Same as ARM mode. FIXME: NEON?
4072 if (!Subtarget->hasVFP2())
4077 return V == (V & ((1LL << 8) - 1));
4081 /// isLegalAddressImmediate - Return true if the integer value can be used
4082 /// as the offset of the target addressing mode for load / store of the
4084 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4085 const ARMSubtarget *Subtarget) {
4092 if (Subtarget->isThumb1Only())
4093 return isLegalT1AddressImmediate(V, VT);
4094 else if (Subtarget->isThumb2())
4095 return isLegalT2AddressImmediate(V, VT, Subtarget);
4100 switch (VT.getSimpleVT().SimpleTy) {
4101 default: return false;
4106 return V == (V & ((1LL << 12) - 1));
4109 return V == (V & ((1LL << 8) - 1));
4112 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4117 return V == (V & ((1LL << 8) - 1));
4121 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4123 int Scale = AM.Scale;
4127 switch (VT.getSimpleVT().SimpleTy) {
4128 default: return false;
4137 return Scale == 2 || Scale == 4 || Scale == 8;
4140 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4144 // Note, we allow "void" uses (basically, uses that aren't loads or
4145 // stores), because arm allows folding a scale into many arithmetic
4146 // operations. This should be made more precise and revisited later.
4148 // Allow r << imm, but the imm has to be a multiple of two.
4149 if (Scale & 1) return false;
4150 return isPowerOf2_32(Scale);
4154 /// isLegalAddressingMode - Return true if the addressing mode represented
4155 /// by AM is legal for this target, for a load/store of the specified type.
4156 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4157 const Type *Ty) const {
4158 EVT VT = getValueType(Ty, true);
4159 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4162 // Can never fold addr of global into load/store.
4167 case 0: // no scale reg, must be "r+i" or "r", or "i".
4170 if (Subtarget->isThumb1Only())
4174 // ARM doesn't support any R+R*scale+imm addr modes.
4181 if (Subtarget->isThumb2())
4182 return isLegalT2ScaledAddressingMode(AM, VT);
4184 int Scale = AM.Scale;
4185 switch (VT.getSimpleVT().SimpleTy) {
4186 default: return false;
4190 if (Scale < 0) Scale = -Scale;
4194 return isPowerOf2_32(Scale & ~1);
4198 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4203 // Note, we allow "void" uses (basically, uses that aren't loads or
4204 // stores), because arm allows folding a scale into many arithmetic
4205 // operations. This should be made more precise and revisited later.
4207 // Allow r << imm, but the imm has to be a multiple of two.
4208 if (Scale & 1) return false;
4209 return isPowerOf2_32(Scale);
4216 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4217 /// icmp immediate, that is the target has icmp instructions which can compare
4218 /// a register against the immediate without having to materialize the
4219 /// immediate into a register.
4220 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4221 if (!Subtarget->isThumb())
4222 return ARM_AM::getSOImmVal(Imm) != -1;
4223 if (Subtarget->isThumb2())
4224 return ARM_AM::getT2SOImmVal(Imm) != -1;
4225 return Imm >= 0 && Imm <= 255;
4228 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4229 bool isSEXTLoad, SDValue &Base,
4230 SDValue &Offset, bool &isInc,
4231 SelectionDAG &DAG) {
4232 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4235 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4237 Base = Ptr->getOperand(0);
4238 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4239 int RHSC = (int)RHS->getZExtValue();
4240 if (RHSC < 0 && RHSC > -256) {
4241 assert(Ptr->getOpcode() == ISD::ADD);
4243 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4247 isInc = (Ptr->getOpcode() == ISD::ADD);
4248 Offset = Ptr->getOperand(1);
4250 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4252 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4253 int RHSC = (int)RHS->getZExtValue();
4254 if (RHSC < 0 && RHSC > -0x1000) {
4255 assert(Ptr->getOpcode() == ISD::ADD);
4257 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4258 Base = Ptr->getOperand(0);
4263 if (Ptr->getOpcode() == ISD::ADD) {
4265 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4266 if (ShOpcVal != ARM_AM::no_shift) {
4267 Base = Ptr->getOperand(1);
4268 Offset = Ptr->getOperand(0);
4270 Base = Ptr->getOperand(0);
4271 Offset = Ptr->getOperand(1);
4276 isInc = (Ptr->getOpcode() == ISD::ADD);
4277 Base = Ptr->getOperand(0);
4278 Offset = Ptr->getOperand(1);
4282 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4286 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4287 bool isSEXTLoad, SDValue &Base,
4288 SDValue &Offset, bool &isInc,
4289 SelectionDAG &DAG) {
4290 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4293 Base = Ptr->getOperand(0);
4294 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4295 int RHSC = (int)RHS->getZExtValue();
4296 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4297 assert(Ptr->getOpcode() == ISD::ADD);
4299 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4301 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4302 isInc = Ptr->getOpcode() == ISD::ADD;
4303 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4311 /// getPreIndexedAddressParts - returns true by value, base pointer and
4312 /// offset pointer and addressing mode by reference if the node's address
4313 /// can be legally represented as pre-indexed load / store address.
4315 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4317 ISD::MemIndexedMode &AM,
4318 SelectionDAG &DAG) const {
4319 if (Subtarget->isThumb1Only())
4324 bool isSEXTLoad = false;
4325 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4326 Ptr = LD->getBasePtr();
4327 VT = LD->getMemoryVT();
4328 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4329 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4330 Ptr = ST->getBasePtr();
4331 VT = ST->getMemoryVT();
4336 bool isLegal = false;
4337 if (Subtarget->isThumb2())
4338 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4339 Offset, isInc, DAG);
4341 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4342 Offset, isInc, DAG);
4346 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4350 /// getPostIndexedAddressParts - returns true by value, base pointer and
4351 /// offset pointer and addressing mode by reference if this node can be
4352 /// combined with a load / store to form a post-indexed load / store.
4353 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4356 ISD::MemIndexedMode &AM,
4357 SelectionDAG &DAG) const {
4358 if (Subtarget->isThumb1Only())
4363 bool isSEXTLoad = false;
4364 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4365 VT = LD->getMemoryVT();
4366 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4367 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4368 VT = ST->getMemoryVT();
4373 bool isLegal = false;
4374 if (Subtarget->isThumb2())
4375 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4378 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4383 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4387 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4391 const SelectionDAG &DAG,
4392 unsigned Depth) const {
4393 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4394 switch (Op.getOpcode()) {
4396 case ARMISD::CMOV: {
4397 // Bits are known zero/one if known on the LHS and RHS.
4398 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4399 if (KnownZero == 0 && KnownOne == 0) return;
4401 APInt KnownZeroRHS, KnownOneRHS;
4402 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4403 KnownZeroRHS, KnownOneRHS, Depth+1);
4404 KnownZero &= KnownZeroRHS;
4405 KnownOne &= KnownOneRHS;
4411 //===----------------------------------------------------------------------===//
4412 // ARM Inline Assembly Support
4413 //===----------------------------------------------------------------------===//
4415 /// getConstraintType - Given a constraint letter, return the type of
4416 /// constraint it is for this target.
4417 ARMTargetLowering::ConstraintType
4418 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4419 if (Constraint.size() == 1) {
4420 switch (Constraint[0]) {
4422 case 'l': return C_RegisterClass;
4423 case 'w': return C_RegisterClass;
4426 return TargetLowering::getConstraintType(Constraint);
4429 std::pair<unsigned, const TargetRegisterClass*>
4430 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4432 if (Constraint.size() == 1) {
4433 // GCC ARM Constraint Letters
4434 switch (Constraint[0]) {
4436 if (Subtarget->isThumb())
4437 return std::make_pair(0U, ARM::tGPRRegisterClass);
4439 return std::make_pair(0U, ARM::GPRRegisterClass);
4441 return std::make_pair(0U, ARM::GPRRegisterClass);
4444 return std::make_pair(0U, ARM::SPRRegisterClass);
4445 if (VT.getSizeInBits() == 64)
4446 return std::make_pair(0U, ARM::DPRRegisterClass);
4447 if (VT.getSizeInBits() == 128)
4448 return std::make_pair(0U, ARM::QPRRegisterClass);
4452 if (StringRef("{cc}").equals_lower(Constraint))
4453 return std::make_pair(0U, ARM::CCRRegisterClass);
4455 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4458 std::vector<unsigned> ARMTargetLowering::
4459 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4461 if (Constraint.size() != 1)
4462 return std::vector<unsigned>();
4464 switch (Constraint[0]) { // GCC ARM Constraint Letters
4467 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4468 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4471 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4472 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4473 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4474 ARM::R12, ARM::LR, 0);
4477 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4478 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4479 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4480 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4481 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4482 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4483 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4484 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4485 if (VT.getSizeInBits() == 64)
4486 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4487 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4488 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4489 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4490 if (VT.getSizeInBits() == 128)
4491 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4492 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4496 return std::vector<unsigned>();
4499 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4500 /// vector. If it is invalid, don't add anything to Ops.
4501 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4504 std::vector<SDValue>&Ops,
4505 SelectionDAG &DAG) const {
4506 SDValue Result(0, 0);
4508 switch (Constraint) {
4510 case 'I': case 'J': case 'K': case 'L':
4511 case 'M': case 'N': case 'O':
4512 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4516 int64_t CVal64 = C->getSExtValue();
4517 int CVal = (int) CVal64;
4518 // None of these constraints allow values larger than 32 bits. Check
4519 // that the value fits in an int.
4523 switch (Constraint) {
4525 if (Subtarget->isThumb1Only()) {
4526 // This must be a constant between 0 and 255, for ADD
4528 if (CVal >= 0 && CVal <= 255)
4530 } else if (Subtarget->isThumb2()) {
4531 // A constant that can be used as an immediate value in a
4532 // data-processing instruction.
4533 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4536 // A constant that can be used as an immediate value in a
4537 // data-processing instruction.
4538 if (ARM_AM::getSOImmVal(CVal) != -1)
4544 if (Subtarget->isThumb()) { // FIXME thumb2
4545 // This must be a constant between -255 and -1, for negated ADD
4546 // immediates. This can be used in GCC with an "n" modifier that
4547 // prints the negated value, for use with SUB instructions. It is
4548 // not useful otherwise but is implemented for compatibility.
4549 if (CVal >= -255 && CVal <= -1)
4552 // This must be a constant between -4095 and 4095. It is not clear
4553 // what this constraint is intended for. Implemented for
4554 // compatibility with GCC.
4555 if (CVal >= -4095 && CVal <= 4095)
4561 if (Subtarget->isThumb1Only()) {
4562 // A 32-bit value where only one byte has a nonzero value. Exclude
4563 // zero to match GCC. This constraint is used by GCC internally for
4564 // constants that can be loaded with a move/shift combination.
4565 // It is not useful otherwise but is implemented for compatibility.
4566 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4568 } else if (Subtarget->isThumb2()) {
4569 // A constant whose bitwise inverse can be used as an immediate
4570 // value in a data-processing instruction. This can be used in GCC
4571 // with a "B" modifier that prints the inverted value, for use with
4572 // BIC and MVN instructions. It is not useful otherwise but is
4573 // implemented for compatibility.
4574 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4577 // A constant whose bitwise inverse can be used as an immediate
4578 // value in a data-processing instruction. This can be used in GCC
4579 // with a "B" modifier that prints the inverted value, for use with
4580 // BIC and MVN instructions. It is not useful otherwise but is
4581 // implemented for compatibility.
4582 if (ARM_AM::getSOImmVal(~CVal) != -1)
4588 if (Subtarget->isThumb1Only()) {
4589 // This must be a constant between -7 and 7,
4590 // for 3-operand ADD/SUB immediate instructions.
4591 if (CVal >= -7 && CVal < 7)
4593 } else if (Subtarget->isThumb2()) {
4594 // A constant whose negation can be used as an immediate value in a
4595 // data-processing instruction. This can be used in GCC with an "n"
4596 // modifier that prints the negated value, for use with SUB
4597 // instructions. It is not useful otherwise but is implemented for
4599 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4602 // A constant whose negation can be used as an immediate value in a
4603 // data-processing instruction. This can be used in GCC with an "n"
4604 // modifier that prints the negated value, for use with SUB
4605 // instructions. It is not useful otherwise but is implemented for
4607 if (ARM_AM::getSOImmVal(-CVal) != -1)
4613 if (Subtarget->isThumb()) { // FIXME thumb2
4614 // This must be a multiple of 4 between 0 and 1020, for
4615 // ADD sp + immediate.
4616 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4619 // A power of two or a constant between 0 and 32. This is used in
4620 // GCC for the shift amount on shifted register operands, but it is
4621 // useful in general for any shift amounts.
4622 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4628 if (Subtarget->isThumb()) { // FIXME thumb2
4629 // This must be a constant between 0 and 31, for shift amounts.
4630 if (CVal >= 0 && CVal <= 31)
4636 if (Subtarget->isThumb()) { // FIXME thumb2
4637 // This must be a multiple of 4 between -508 and 508, for
4638 // ADD/SUB sp = sp + immediate.
4639 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4644 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4648 if (Result.getNode()) {
4649 Ops.push_back(Result);
4652 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4657 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4658 // The ARM target isn't yet aware of offsets.
4662 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4663 APInt Imm = FPImm.bitcastToAPInt();
4664 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4665 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4666 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4668 // We can handle 4 bits of mantissa.
4669 // mantissa = (16+UInt(e:f:g:h))/16.
4670 if (Mantissa & 0x7ffff)
4673 if ((Mantissa & 0xf) != Mantissa)
4676 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4677 if (Exp < -3 || Exp > 4)
4679 Exp = ((Exp+3) & 0x7) ^ 4;
4681 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4684 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4685 APInt Imm = FPImm.bitcastToAPInt();
4686 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4687 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4688 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4690 // We can handle 4 bits of mantissa.
4691 // mantissa = (16+UInt(e:f:g:h))/16.
4692 if (Mantissa & 0xffffffffffffLL)
4695 if ((Mantissa & 0xf) != Mantissa)
4698 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4699 if (Exp < -3 || Exp > 4)
4701 Exp = ((Exp+3) & 0x7) ^ 4;
4703 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4706 /// isFPImmLegal - Returns true if the target can instruction select the
4707 /// specified FP immediate natively. If false, the legalizer will
4708 /// materialize the FP immediate as a load from a constant pool.
4709 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4710 if (!Subtarget->hasVFP3())
4713 return ARM::getVFPf32Imm(Imm) != -1;
4715 return ARM::getVFPf64Imm(Imm) != -1;