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 // Special handling for half-precision FP.
432 if (!Subtarget->hasFP16()) {
433 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
434 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
438 // We have target-specific dag combine patterns for the following nodes:
439 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
440 setTargetDAGCombine(ISD::ADD);
441 setTargetDAGCombine(ISD::SUB);
443 setStackPointerRegisterToSaveRestore(ARM::SP);
444 setSchedulingPreference(SchedulingForRegPressure);
446 // FIXME: If-converter should use instruction latency to determine
447 // profitability rather than relying on fixed limits.
448 if (Subtarget->getCPUString() == "generic") {
449 // Generic (and overly aggressive) if-conversion limits.
450 setIfCvtBlockSizeLimit(10);
451 setIfCvtDupBlockSizeLimit(2);
452 } else if (Subtarget->hasV6Ops()) {
453 setIfCvtBlockSizeLimit(2);
454 setIfCvtDupBlockSizeLimit(1);
456 setIfCvtBlockSizeLimit(3);
457 setIfCvtDupBlockSizeLimit(2);
460 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
461 // Do not enable CodePlacementOpt for now: it currently runs after the
462 // ARMConstantIslandPass and messes up branch relaxation and placement
463 // of constant islands.
464 // benefitFromCodePlacementOpt = true;
467 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
470 case ARMISD::Wrapper: return "ARMISD::Wrapper";
471 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
472 case ARMISD::CALL: return "ARMISD::CALL";
473 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
474 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
475 case ARMISD::tCALL: return "ARMISD::tCALL";
476 case ARMISD::BRCOND: return "ARMISD::BRCOND";
477 case ARMISD::BR_JT: return "ARMISD::BR_JT";
478 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
479 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
480 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
481 case ARMISD::CMP: return "ARMISD::CMP";
482 case ARMISD::CMPZ: return "ARMISD::CMPZ";
483 case ARMISD::CMPFP: return "ARMISD::CMPFP";
484 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
485 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
486 case ARMISD::CMOV: return "ARMISD::CMOV";
487 case ARMISD::CNEG: return "ARMISD::CNEG";
489 case ARMISD::RBIT: return "ARMISD::RBIT";
491 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
492 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
493 case ARMISD::RRX: return "ARMISD::RRX";
495 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
496 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
498 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
499 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
501 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
503 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
505 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
506 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
508 case ARMISD::VCEQ: return "ARMISD::VCEQ";
509 case ARMISD::VCGE: return "ARMISD::VCGE";
510 case ARMISD::VCGEU: return "ARMISD::VCGEU";
511 case ARMISD::VCGT: return "ARMISD::VCGT";
512 case ARMISD::VCGTU: return "ARMISD::VCGTU";
513 case ARMISD::VTST: return "ARMISD::VTST";
515 case ARMISD::VSHL: return "ARMISD::VSHL";
516 case ARMISD::VSHRs: return "ARMISD::VSHRs";
517 case ARMISD::VSHRu: return "ARMISD::VSHRu";
518 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
519 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
520 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
521 case ARMISD::VSHRN: return "ARMISD::VSHRN";
522 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
523 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
524 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
525 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
526 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
527 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
528 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
529 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
530 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
531 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
532 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
533 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
534 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
535 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
536 case ARMISD::VDUP: return "ARMISD::VDUP";
537 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
538 case ARMISD::VEXT: return "ARMISD::VEXT";
539 case ARMISD::VREV64: return "ARMISD::VREV64";
540 case ARMISD::VREV32: return "ARMISD::VREV32";
541 case ARMISD::VREV16: return "ARMISD::VREV16";
542 case ARMISD::VZIP: return "ARMISD::VZIP";
543 case ARMISD::VUZP: return "ARMISD::VUZP";
544 case ARMISD::VTRN: return "ARMISD::VTRN";
545 case ARMISD::FMAX: return "ARMISD::FMAX";
546 case ARMISD::FMIN: return "ARMISD::FMIN";
550 /// getFunctionAlignment - Return the Log2 alignment of this function.
551 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
552 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
555 //===----------------------------------------------------------------------===//
557 //===----------------------------------------------------------------------===//
559 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
560 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
562 default: llvm_unreachable("Unknown condition code!");
563 case ISD::SETNE: return ARMCC::NE;
564 case ISD::SETEQ: return ARMCC::EQ;
565 case ISD::SETGT: return ARMCC::GT;
566 case ISD::SETGE: return ARMCC::GE;
567 case ISD::SETLT: return ARMCC::LT;
568 case ISD::SETLE: return ARMCC::LE;
569 case ISD::SETUGT: return ARMCC::HI;
570 case ISD::SETUGE: return ARMCC::HS;
571 case ISD::SETULT: return ARMCC::LO;
572 case ISD::SETULE: return ARMCC::LS;
576 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
577 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
578 ARMCC::CondCodes &CondCode2) {
579 CondCode2 = ARMCC::AL;
581 default: llvm_unreachable("Unknown FP condition!");
583 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
585 case ISD::SETOGT: CondCode = ARMCC::GT; break;
587 case ISD::SETOGE: CondCode = ARMCC::GE; break;
588 case ISD::SETOLT: CondCode = ARMCC::MI; break;
589 case ISD::SETOLE: CondCode = ARMCC::LS; break;
590 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
591 case ISD::SETO: CondCode = ARMCC::VC; break;
592 case ISD::SETUO: CondCode = ARMCC::VS; break;
593 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
594 case ISD::SETUGT: CondCode = ARMCC::HI; break;
595 case ISD::SETUGE: CondCode = ARMCC::PL; break;
597 case ISD::SETULT: CondCode = ARMCC::LT; break;
599 case ISD::SETULE: CondCode = ARMCC::LE; break;
601 case ISD::SETUNE: CondCode = ARMCC::NE; break;
605 //===----------------------------------------------------------------------===//
606 // Calling Convention Implementation
607 //===----------------------------------------------------------------------===//
609 #include "ARMGenCallingConv.inc"
611 // APCS f64 is in register pairs, possibly split to stack
612 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
613 CCValAssign::LocInfo &LocInfo,
614 CCState &State, bool CanFail) {
615 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
617 // Try to get the first register.
618 if (unsigned Reg = State.AllocateReg(RegList, 4))
619 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
621 // For the 2nd half of a v2f64, do not fail.
625 // Put the whole thing on the stack.
626 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
627 State.AllocateStack(8, 4),
632 // Try to get the second register.
633 if (unsigned Reg = State.AllocateReg(RegList, 4))
634 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
636 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
637 State.AllocateStack(4, 4),
642 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
643 CCValAssign::LocInfo &LocInfo,
644 ISD::ArgFlagsTy &ArgFlags,
646 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
648 if (LocVT == MVT::v2f64 &&
649 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
651 return true; // we handled it
654 // AAPCS f64 is in aligned register pairs
655 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
656 CCValAssign::LocInfo &LocInfo,
657 CCState &State, bool CanFail) {
658 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
659 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
661 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
663 // For the 2nd half of a v2f64, do not just fail.
667 // Put the whole thing on the stack.
668 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
669 State.AllocateStack(8, 8),
675 for (i = 0; i < 2; ++i)
676 if (HiRegList[i] == Reg)
679 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
680 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
685 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
686 CCValAssign::LocInfo &LocInfo,
687 ISD::ArgFlagsTy &ArgFlags,
689 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
691 if (LocVT == MVT::v2f64 &&
692 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
694 return true; // we handled it
697 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
698 CCValAssign::LocInfo &LocInfo, CCState &State) {
699 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
700 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
702 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
704 return false; // we didn't handle it
707 for (i = 0; i < 2; ++i)
708 if (HiRegList[i] == Reg)
711 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
712 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
717 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
718 CCValAssign::LocInfo &LocInfo,
719 ISD::ArgFlagsTy &ArgFlags,
721 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
723 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
725 return true; // we handled it
728 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
729 CCValAssign::LocInfo &LocInfo,
730 ISD::ArgFlagsTy &ArgFlags,
732 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
736 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
737 /// given CallingConvention value.
738 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
740 bool isVarArg) const {
743 llvm_unreachable("Unsupported calling convention");
745 case CallingConv::Fast:
746 // Use target triple & subtarget features to do actual dispatch.
747 if (Subtarget->isAAPCS_ABI()) {
748 if (Subtarget->hasVFP2() &&
749 FloatABIType == FloatABI::Hard && !isVarArg)
750 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
752 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
754 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
755 case CallingConv::ARM_AAPCS_VFP:
756 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
757 case CallingConv::ARM_AAPCS:
758 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
759 case CallingConv::ARM_APCS:
760 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
764 /// LowerCallResult - Lower the result values of a call into the
765 /// appropriate copies out of appropriate physical registers.
767 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
768 CallingConv::ID CallConv, bool isVarArg,
769 const SmallVectorImpl<ISD::InputArg> &Ins,
770 DebugLoc dl, SelectionDAG &DAG,
771 SmallVectorImpl<SDValue> &InVals) {
773 // Assign locations to each value returned by this call.
774 SmallVector<CCValAssign, 16> RVLocs;
775 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
776 RVLocs, *DAG.getContext());
777 CCInfo.AnalyzeCallResult(Ins,
778 CCAssignFnForNode(CallConv, /* Return*/ true,
781 // Copy all of the result registers out of their specified physreg.
782 for (unsigned i = 0; i != RVLocs.size(); ++i) {
783 CCValAssign VA = RVLocs[i];
786 if (VA.needsCustom()) {
787 // Handle f64 or half of a v2f64.
788 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
790 Chain = Lo.getValue(1);
791 InFlag = Lo.getValue(2);
792 VA = RVLocs[++i]; // skip ahead to next loc
793 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
795 Chain = Hi.getValue(1);
796 InFlag = Hi.getValue(2);
797 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
799 if (VA.getLocVT() == MVT::v2f64) {
800 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
801 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
802 DAG.getConstant(0, MVT::i32));
804 VA = RVLocs[++i]; // skip ahead to next loc
805 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
806 Chain = Lo.getValue(1);
807 InFlag = Lo.getValue(2);
808 VA = RVLocs[++i]; // skip ahead to next loc
809 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
810 Chain = Hi.getValue(1);
811 InFlag = Hi.getValue(2);
812 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
813 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
814 DAG.getConstant(1, MVT::i32));
817 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
819 Chain = Val.getValue(1);
820 InFlag = Val.getValue(2);
823 switch (VA.getLocInfo()) {
824 default: llvm_unreachable("Unknown loc info!");
825 case CCValAssign::Full: break;
826 case CCValAssign::BCvt:
827 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
831 InVals.push_back(Val);
837 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
838 /// by "Src" to address "Dst" of size "Size". Alignment information is
839 /// specified by the specific parameter attribute. The copy will be passed as
840 /// a byval function parameter.
841 /// Sometimes what we are copying is the end of a larger object, the part that
842 /// does not fit in registers.
844 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
845 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
847 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
848 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
849 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
852 /// LowerMemOpCallTo - Store the argument to the stack.
854 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
855 SDValue StackPtr, SDValue Arg,
856 DebugLoc dl, SelectionDAG &DAG,
857 const CCValAssign &VA,
858 ISD::ArgFlagsTy Flags) {
859 unsigned LocMemOffset = VA.getLocMemOffset();
860 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
861 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
862 if (Flags.isByVal()) {
863 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
865 return DAG.getStore(Chain, dl, Arg, PtrOff,
866 PseudoSourceValue::getStack(), LocMemOffset,
870 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
871 SDValue Chain, SDValue &Arg,
872 RegsToPassVector &RegsToPass,
873 CCValAssign &VA, CCValAssign &NextVA,
875 SmallVector<SDValue, 8> &MemOpChains,
876 ISD::ArgFlagsTy Flags) {
878 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
879 DAG.getVTList(MVT::i32, MVT::i32), Arg);
880 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
882 if (NextVA.isRegLoc())
883 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
885 assert(NextVA.isMemLoc());
886 if (StackPtr.getNode() == 0)
887 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
889 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
895 /// LowerCall - Lowering a call into a callseq_start <-
896 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
899 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
900 CallingConv::ID CallConv, bool isVarArg,
902 const SmallVectorImpl<ISD::OutputArg> &Outs,
903 const SmallVectorImpl<ISD::InputArg> &Ins,
904 DebugLoc dl, SelectionDAG &DAG,
905 SmallVectorImpl<SDValue> &InVals) {
906 // ARM target does not yet support tail call optimization.
909 // Analyze operands of the call, assigning locations to each operand.
910 SmallVector<CCValAssign, 16> ArgLocs;
911 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
913 CCInfo.AnalyzeCallOperands(Outs,
914 CCAssignFnForNode(CallConv, /* Return*/ false,
917 // Get a count of how many bytes are to be pushed on the stack.
918 unsigned NumBytes = CCInfo.getNextStackOffset();
920 // Adjust the stack pointer for the new arguments...
921 // These operations are automatically eliminated by the prolog/epilog pass
922 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
924 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
926 RegsToPassVector RegsToPass;
927 SmallVector<SDValue, 8> MemOpChains;
929 // Walk the register/memloc assignments, inserting copies/loads. In the case
930 // of tail call optimization, arguments are handled later.
931 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
934 CCValAssign &VA = ArgLocs[i];
935 SDValue Arg = Outs[realArgIdx].Val;
936 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
938 // Promote the value if needed.
939 switch (VA.getLocInfo()) {
940 default: llvm_unreachable("Unknown loc info!");
941 case CCValAssign::Full: break;
942 case CCValAssign::SExt:
943 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
945 case CCValAssign::ZExt:
946 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
948 case CCValAssign::AExt:
949 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
951 case CCValAssign::BCvt:
952 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
956 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
957 if (VA.needsCustom()) {
958 if (VA.getLocVT() == MVT::v2f64) {
959 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
960 DAG.getConstant(0, MVT::i32));
961 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
962 DAG.getConstant(1, MVT::i32));
964 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
965 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
967 VA = ArgLocs[++i]; // skip ahead to next loc
969 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
970 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
972 assert(VA.isMemLoc());
974 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
975 dl, DAG, VA, Flags));
978 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
979 StackPtr, MemOpChains, Flags);
981 } else if (VA.isRegLoc()) {
982 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
984 assert(VA.isMemLoc());
986 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
987 dl, DAG, VA, Flags));
991 if (!MemOpChains.empty())
992 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
993 &MemOpChains[0], MemOpChains.size());
995 // Build a sequence of copy-to-reg nodes chained together with token chain
996 // and flag operands which copy the outgoing args into the appropriate regs.
998 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
999 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1000 RegsToPass[i].second, InFlag);
1001 InFlag = Chain.getValue(1);
1004 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1005 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1006 // node so that legalize doesn't hack it.
1007 bool isDirect = false;
1008 bool isARMFunc = false;
1009 bool isLocalARMFunc = false;
1010 MachineFunction &MF = DAG.getMachineFunction();
1011 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1012 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1013 GlobalValue *GV = G->getGlobal();
1015 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1016 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1017 getTargetMachine().getRelocationModel() != Reloc::Static;
1018 isARMFunc = !Subtarget->isThumb() || isStub;
1019 // ARM call to a local ARM function is predicable.
1020 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1021 // tBX takes a register source operand.
1022 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1023 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1024 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1027 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1028 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1029 Callee = DAG.getLoad(getPointerTy(), dl,
1030 DAG.getEntryNode(), CPAddr,
1031 PseudoSourceValue::getConstantPool(), 0,
1033 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1034 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1035 getPointerTy(), Callee, PICLabel);
1037 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1038 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1040 bool isStub = Subtarget->isTargetDarwin() &&
1041 getTargetMachine().getRelocationModel() != Reloc::Static;
1042 isARMFunc = !Subtarget->isThumb() || isStub;
1043 // tBX takes a register source operand.
1044 const char *Sym = S->getSymbol();
1045 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1046 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1047 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1048 Sym, ARMPCLabelIndex, 4);
1049 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1050 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1051 Callee = DAG.getLoad(getPointerTy(), dl,
1052 DAG.getEntryNode(), CPAddr,
1053 PseudoSourceValue::getConstantPool(), 0,
1055 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1056 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1057 getPointerTy(), Callee, PICLabel);
1059 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1062 // FIXME: handle tail calls differently.
1064 if (Subtarget->isThumb()) {
1065 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1066 CallOpc = ARMISD::CALL_NOLINK;
1068 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1070 CallOpc = (isDirect || Subtarget->hasV5TOps())
1071 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1072 : ARMISD::CALL_NOLINK;
1074 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1075 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1076 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1077 InFlag = Chain.getValue(1);
1080 std::vector<SDValue> Ops;
1081 Ops.push_back(Chain);
1082 Ops.push_back(Callee);
1084 // Add argument registers to the end of the list so that they are known live
1086 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1087 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1088 RegsToPass[i].second.getValueType()));
1090 if (InFlag.getNode())
1091 Ops.push_back(InFlag);
1092 // Returns a chain and a flag for retval copy to use.
1093 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1094 &Ops[0], Ops.size());
1095 InFlag = Chain.getValue(1);
1097 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1098 DAG.getIntPtrConstant(0, true), InFlag);
1100 InFlag = Chain.getValue(1);
1102 // Handle result values, copying them out of physregs into vregs that we
1104 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1109 ARMTargetLowering::LowerReturn(SDValue Chain,
1110 CallingConv::ID CallConv, bool isVarArg,
1111 const SmallVectorImpl<ISD::OutputArg> &Outs,
1112 DebugLoc dl, SelectionDAG &DAG) {
1114 // CCValAssign - represent the assignment of the return value to a location.
1115 SmallVector<CCValAssign, 16> RVLocs;
1117 // CCState - Info about the registers and stack slots.
1118 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1121 // Analyze outgoing return values.
1122 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1125 // If this is the first return lowered for this function, add
1126 // the regs to the liveout set for the function.
1127 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1128 for (unsigned i = 0; i != RVLocs.size(); ++i)
1129 if (RVLocs[i].isRegLoc())
1130 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1135 // Copy the result values into the output registers.
1136 for (unsigned i = 0, realRVLocIdx = 0;
1138 ++i, ++realRVLocIdx) {
1139 CCValAssign &VA = RVLocs[i];
1140 assert(VA.isRegLoc() && "Can only return in registers!");
1142 SDValue Arg = Outs[realRVLocIdx].Val;
1144 switch (VA.getLocInfo()) {
1145 default: llvm_unreachable("Unknown loc info!");
1146 case CCValAssign::Full: break;
1147 case CCValAssign::BCvt:
1148 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1152 if (VA.needsCustom()) {
1153 if (VA.getLocVT() == MVT::v2f64) {
1154 // Extract the first half and return it in two registers.
1155 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1156 DAG.getConstant(0, MVT::i32));
1157 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1158 DAG.getVTList(MVT::i32, MVT::i32), Half);
1160 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1161 Flag = Chain.getValue(1);
1162 VA = RVLocs[++i]; // skip ahead to next loc
1163 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1164 HalfGPRs.getValue(1), Flag);
1165 Flag = Chain.getValue(1);
1166 VA = RVLocs[++i]; // skip ahead to next loc
1168 // Extract the 2nd half and fall through to handle it as an f64 value.
1169 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1170 DAG.getConstant(1, MVT::i32));
1172 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1174 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1175 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1176 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1177 Flag = Chain.getValue(1);
1178 VA = RVLocs[++i]; // skip ahead to next loc
1179 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1182 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1184 // Guarantee that all emitted copies are
1185 // stuck together, avoiding something bad.
1186 Flag = Chain.getValue(1);
1191 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1193 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1198 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1199 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1200 // one of the above mentioned nodes. It has to be wrapped because otherwise
1201 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1202 // be used to form addressing mode. These wrapped nodes will be selected
1204 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1205 EVT PtrVT = Op.getValueType();
1206 // FIXME there is no actual debug info here
1207 DebugLoc dl = Op.getDebugLoc();
1208 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1210 if (CP->isMachineConstantPoolEntry())
1211 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1212 CP->getAlignment());
1214 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1215 CP->getAlignment());
1216 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1219 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1220 MachineFunction &MF = DAG.getMachineFunction();
1221 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1222 unsigned ARMPCLabelIndex = 0;
1223 DebugLoc DL = Op.getDebugLoc();
1224 EVT PtrVT = getPointerTy();
1225 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1226 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1228 if (RelocM == Reloc::Static) {
1229 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1231 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1232 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1233 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1234 ARMCP::CPBlockAddress,
1236 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1238 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1239 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1240 PseudoSourceValue::getConstantPool(), 0,
1242 if (RelocM == Reloc::Static)
1244 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1245 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1248 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1250 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1251 SelectionDAG &DAG) {
1252 DebugLoc dl = GA->getDebugLoc();
1253 EVT PtrVT = getPointerTy();
1254 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1255 MachineFunction &MF = DAG.getMachineFunction();
1256 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1257 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1258 ARMConstantPoolValue *CPV =
1259 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1260 ARMCP::CPValue, PCAdj, "tlsgd", true);
1261 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1262 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1263 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1264 PseudoSourceValue::getConstantPool(), 0,
1266 SDValue Chain = Argument.getValue(1);
1268 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1269 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1271 // call __tls_get_addr.
1274 Entry.Node = Argument;
1275 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1276 Args.push_back(Entry);
1277 // FIXME: is there useful debug info available here?
1278 std::pair<SDValue, SDValue> CallResult =
1279 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1280 false, false, false, false,
1281 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1282 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1283 return CallResult.first;
1286 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1287 // "local exec" model.
1289 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1290 SelectionDAG &DAG) {
1291 GlobalValue *GV = GA->getGlobal();
1292 DebugLoc dl = GA->getDebugLoc();
1294 SDValue Chain = DAG.getEntryNode();
1295 EVT PtrVT = getPointerTy();
1296 // Get the Thread Pointer
1297 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1299 if (GV->isDeclaration()) {
1300 MachineFunction &MF = DAG.getMachineFunction();
1301 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1302 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1303 // Initial exec model.
1304 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1305 ARMConstantPoolValue *CPV =
1306 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1307 ARMCP::CPValue, PCAdj, "gottpoff", true);
1308 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1309 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1310 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1311 PseudoSourceValue::getConstantPool(), 0,
1313 Chain = Offset.getValue(1);
1315 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1316 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1318 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1319 PseudoSourceValue::getConstantPool(), 0,
1323 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1324 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1325 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1326 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1327 PseudoSourceValue::getConstantPool(), 0,
1331 // The address of the thread local variable is the add of the thread
1332 // pointer with the offset of the variable.
1333 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1337 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1338 // TODO: implement the "local dynamic" model
1339 assert(Subtarget->isTargetELF() &&
1340 "TLS not implemented for non-ELF targets");
1341 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1342 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1343 // otherwise use the "Local Exec" TLS Model
1344 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1345 return LowerToTLSGeneralDynamicModel(GA, DAG);
1347 return LowerToTLSExecModels(GA, DAG);
1350 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1351 SelectionDAG &DAG) {
1352 EVT PtrVT = getPointerTy();
1353 DebugLoc dl = Op.getDebugLoc();
1354 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1355 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1356 if (RelocM == Reloc::PIC_) {
1357 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1358 ARMConstantPoolValue *CPV =
1359 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1360 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1361 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1362 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1364 PseudoSourceValue::getConstantPool(), 0,
1366 SDValue Chain = Result.getValue(1);
1367 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1368 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1370 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1371 PseudoSourceValue::getGOT(), 0,
1375 // If we have T2 ops, we can materialize the address directly via movt/movw
1376 // pair. This is always cheaper.
1377 if (Subtarget->useMovt()) {
1378 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1379 DAG.getTargetGlobalAddress(GV, PtrVT));
1381 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1382 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1383 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1384 PseudoSourceValue::getConstantPool(), 0,
1390 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1391 SelectionDAG &DAG) {
1392 MachineFunction &MF = DAG.getMachineFunction();
1393 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1394 unsigned ARMPCLabelIndex = 0;
1395 EVT PtrVT = getPointerTy();
1396 DebugLoc dl = Op.getDebugLoc();
1397 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1398 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1400 if (RelocM == Reloc::Static)
1401 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1403 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1404 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1405 ARMConstantPoolValue *CPV =
1406 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1407 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1409 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1411 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1412 PseudoSourceValue::getConstantPool(), 0,
1414 SDValue Chain = Result.getValue(1);
1416 if (RelocM == Reloc::PIC_) {
1417 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1418 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1421 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1422 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1423 PseudoSourceValue::getGOT(), 0,
1429 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1431 assert(Subtarget->isTargetELF() &&
1432 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1433 MachineFunction &MF = DAG.getMachineFunction();
1434 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1435 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1436 EVT PtrVT = getPointerTy();
1437 DebugLoc dl = Op.getDebugLoc();
1438 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1439 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1440 "_GLOBAL_OFFSET_TABLE_",
1441 ARMPCLabelIndex, PCAdj);
1442 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1443 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1444 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1445 PseudoSourceValue::getConstantPool(), 0,
1447 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1448 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1452 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1453 const ARMSubtarget *Subtarget) {
1454 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1455 DebugLoc dl = Op.getDebugLoc();
1457 default: return SDValue(); // Don't custom lower most intrinsics.
1458 case Intrinsic::arm_thread_pointer: {
1459 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1460 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1462 case Intrinsic::eh_sjlj_lsda: {
1463 MachineFunction &MF = DAG.getMachineFunction();
1464 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1465 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1466 EVT PtrVT = getPointerTy();
1467 DebugLoc dl = Op.getDebugLoc();
1468 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1470 unsigned PCAdj = (RelocM != Reloc::PIC_)
1471 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1472 ARMConstantPoolValue *CPV =
1473 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1474 ARMCP::CPLSDA, PCAdj);
1475 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1476 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1478 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1479 PseudoSourceValue::getConstantPool(), 0,
1481 SDValue Chain = Result.getValue(1);
1483 if (RelocM == Reloc::PIC_) {
1484 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1485 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1489 case Intrinsic::eh_sjlj_setjmp:
1490 SDValue Val = Subtarget->isThumb() ?
1491 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1492 DAG.getConstant(0, MVT::i32);
1493 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1498 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1499 const ARMSubtarget *Subtarget) {
1500 DebugLoc dl = Op.getDebugLoc();
1501 SDValue Op5 = Op.getOperand(5);
1503 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1504 if (isDeviceBarrier) {
1505 if (Subtarget->hasV7Ops())
1506 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1508 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1509 DAG.getConstant(0, MVT::i32));
1511 if (Subtarget->hasV7Ops())
1512 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1514 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1515 DAG.getConstant(0, MVT::i32));
1520 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1521 unsigned VarArgsFrameIndex) {
1522 // vastart just stores the address of the VarArgsFrameIndex slot into the
1523 // memory location argument.
1524 DebugLoc dl = Op.getDebugLoc();
1525 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1526 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1527 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1528 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1533 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1534 SDNode *Node = Op.getNode();
1535 DebugLoc dl = Node->getDebugLoc();
1536 EVT VT = Node->getValueType(0);
1537 SDValue Chain = Op.getOperand(0);
1538 SDValue Size = Op.getOperand(1);
1539 SDValue Align = Op.getOperand(2);
1541 // Chain the dynamic stack allocation so that it doesn't modify the stack
1542 // pointer when other instructions are using the stack.
1543 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1545 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1546 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1547 if (AlignVal > StackAlign)
1548 // Do this now since selection pass cannot introduce new target
1549 // independent node.
1550 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1552 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1553 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1554 // do even more horrible hack later.
1555 MachineFunction &MF = DAG.getMachineFunction();
1556 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1557 if (AFI->isThumb1OnlyFunction()) {
1559 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1561 uint32_t Val = C->getZExtValue();
1562 if (Val <= 508 && ((Val & 3) == 0))
1566 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1569 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1570 SDValue Ops1[] = { Chain, Size, Align };
1571 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1572 Chain = Res.getValue(1);
1573 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1574 DAG.getIntPtrConstant(0, true), SDValue());
1575 SDValue Ops2[] = { Res, Chain };
1576 return DAG.getMergeValues(Ops2, 2, dl);
1580 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1581 SDValue &Root, SelectionDAG &DAG,
1583 MachineFunction &MF = DAG.getMachineFunction();
1584 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1586 TargetRegisterClass *RC;
1587 if (AFI->isThumb1OnlyFunction())
1588 RC = ARM::tGPRRegisterClass;
1590 RC = ARM::GPRRegisterClass;
1592 // Transform the arguments stored in physical registers into virtual ones.
1593 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1594 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1597 if (NextVA.isMemLoc()) {
1598 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1599 MachineFrameInfo *MFI = MF.getFrameInfo();
1600 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset(),
1603 // Create load node to retrieve arguments from the stack.
1604 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1605 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1606 PseudoSourceValue::getFixedStack(FI), 0,
1609 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1610 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1613 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1617 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1618 CallingConv::ID CallConv, bool isVarArg,
1619 const SmallVectorImpl<ISD::InputArg>
1621 DebugLoc dl, SelectionDAG &DAG,
1622 SmallVectorImpl<SDValue> &InVals) {
1624 MachineFunction &MF = DAG.getMachineFunction();
1625 MachineFrameInfo *MFI = MF.getFrameInfo();
1627 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1629 // Assign locations to all of the incoming arguments.
1630 SmallVector<CCValAssign, 16> ArgLocs;
1631 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1633 CCInfo.AnalyzeFormalArguments(Ins,
1634 CCAssignFnForNode(CallConv, /* Return*/ false,
1637 SmallVector<SDValue, 16> ArgValues;
1639 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1640 CCValAssign &VA = ArgLocs[i];
1642 // Arguments stored in registers.
1643 if (VA.isRegLoc()) {
1644 EVT RegVT = VA.getLocVT();
1647 if (VA.needsCustom()) {
1648 // f64 and vector types are split up into multiple registers or
1649 // combinations of registers and stack slots.
1652 if (VA.getLocVT() == MVT::v2f64) {
1653 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1655 VA = ArgLocs[++i]; // skip ahead to next loc
1656 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1658 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1659 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1660 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1661 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1662 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1664 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1667 TargetRegisterClass *RC;
1669 if (RegVT == MVT::f32)
1670 RC = ARM::SPRRegisterClass;
1671 else if (RegVT == MVT::f64)
1672 RC = ARM::DPRRegisterClass;
1673 else if (RegVT == MVT::v2f64)
1674 RC = ARM::QPRRegisterClass;
1675 else if (RegVT == MVT::i32)
1676 RC = (AFI->isThumb1OnlyFunction() ?
1677 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1679 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1681 // Transform the arguments in physical registers into virtual ones.
1682 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1683 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1686 // If this is an 8 or 16-bit value, it is really passed promoted
1687 // to 32 bits. Insert an assert[sz]ext to capture this, then
1688 // truncate to the right size.
1689 switch (VA.getLocInfo()) {
1690 default: llvm_unreachable("Unknown loc info!");
1691 case CCValAssign::Full: break;
1692 case CCValAssign::BCvt:
1693 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1695 case CCValAssign::SExt:
1696 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1697 DAG.getValueType(VA.getValVT()));
1698 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1700 case CCValAssign::ZExt:
1701 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1702 DAG.getValueType(VA.getValVT()));
1703 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1707 InVals.push_back(ArgValue);
1709 } else { // VA.isRegLoc()
1712 assert(VA.isMemLoc());
1713 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1715 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1716 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1719 // Create load nodes to retrieve arguments from the stack.
1720 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1721 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1722 PseudoSourceValue::getFixedStack(FI), 0,
1729 static const unsigned GPRArgRegs[] = {
1730 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1733 unsigned NumGPRs = CCInfo.getFirstUnallocated
1734 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1736 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1737 unsigned VARegSize = (4 - NumGPRs) * 4;
1738 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1739 unsigned ArgOffset = CCInfo.getNextStackOffset();
1740 if (VARegSaveSize) {
1741 // If this function is vararg, store any remaining integer argument regs
1742 // to their spots on the stack so that they may be loaded by deferencing
1743 // the result of va_next.
1744 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1745 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1746 VARegSaveSize - VARegSize,
1748 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1750 SmallVector<SDValue, 4> MemOps;
1751 for (; NumGPRs < 4; ++NumGPRs) {
1752 TargetRegisterClass *RC;
1753 if (AFI->isThumb1OnlyFunction())
1754 RC = ARM::tGPRRegisterClass;
1756 RC = ARM::GPRRegisterClass;
1758 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1759 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1760 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1761 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0,
1763 MemOps.push_back(Store);
1764 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1765 DAG.getConstant(4, getPointerTy()));
1767 if (!MemOps.empty())
1768 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1769 &MemOps[0], MemOps.size());
1771 // This will point to the next argument passed via stack.
1772 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
1778 /// isFloatingPointZero - Return true if this is +0.0.
1779 static bool isFloatingPointZero(SDValue Op) {
1780 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1781 return CFP->getValueAPF().isPosZero();
1782 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1783 // Maybe this has already been legalized into the constant pool?
1784 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1785 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1786 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1787 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1788 return CFP->getValueAPF().isPosZero();
1794 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1795 /// the given operands.
1797 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1798 SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
1799 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1800 unsigned C = RHSC->getZExtValue();
1801 if (!isLegalICmpImmediate(C)) {
1802 // Constant does not fit, try adjusting it by one?
1807 if (isLegalICmpImmediate(C-1)) {
1808 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1809 RHS = DAG.getConstant(C-1, MVT::i32);
1814 if (C > 0 && isLegalICmpImmediate(C-1)) {
1815 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1816 RHS = DAG.getConstant(C-1, MVT::i32);
1821 if (isLegalICmpImmediate(C+1)) {
1822 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1823 RHS = DAG.getConstant(C+1, MVT::i32);
1828 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1829 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1830 RHS = DAG.getConstant(C+1, MVT::i32);
1837 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1838 ARMISD::NodeType CompareType;
1841 CompareType = ARMISD::CMP;
1846 CompareType = ARMISD::CMPZ;
1849 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1850 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1853 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1854 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1857 if (!isFloatingPointZero(RHS))
1858 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1860 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1861 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1864 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1865 EVT VT = Op.getValueType();
1866 SDValue LHS = Op.getOperand(0);
1867 SDValue RHS = Op.getOperand(1);
1868 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1869 SDValue TrueVal = Op.getOperand(2);
1870 SDValue FalseVal = Op.getOperand(3);
1871 DebugLoc dl = Op.getDebugLoc();
1873 if (LHS.getValueType() == MVT::i32) {
1875 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1876 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1877 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1880 ARMCC::CondCodes CondCode, CondCode2;
1881 FPCCToARMCC(CC, CondCode, CondCode2);
1883 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1884 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1885 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1886 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1888 if (CondCode2 != ARMCC::AL) {
1889 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1890 // FIXME: Needs another CMP because flag can have but one use.
1891 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1892 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1893 Result, TrueVal, ARMCC2, CCR, Cmp2);
1898 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1899 SDValue Chain = Op.getOperand(0);
1900 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1901 SDValue LHS = Op.getOperand(2);
1902 SDValue RHS = Op.getOperand(3);
1903 SDValue Dest = Op.getOperand(4);
1904 DebugLoc dl = Op.getDebugLoc();
1906 if (LHS.getValueType() == MVT::i32) {
1908 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1909 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1910 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1911 Chain, Dest, ARMCC, CCR,Cmp);
1914 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1915 ARMCC::CondCodes CondCode, CondCode2;
1916 FPCCToARMCC(CC, CondCode, CondCode2);
1918 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1919 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1920 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1921 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1922 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1923 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1924 if (CondCode2 != ARMCC::AL) {
1925 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1926 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1927 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1932 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1933 SDValue Chain = Op.getOperand(0);
1934 SDValue Table = Op.getOperand(1);
1935 SDValue Index = Op.getOperand(2);
1936 DebugLoc dl = Op.getDebugLoc();
1938 EVT PTy = getPointerTy();
1939 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1940 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1941 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1942 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1943 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1944 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1945 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1946 if (Subtarget->isThumb2()) {
1947 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1948 // which does another jump to the destination. This also makes it easier
1949 // to translate it to TBB / TBH later.
1950 // FIXME: This might not work if the function is extremely large.
1951 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1952 Addr, Op.getOperand(2), JTI, UId);
1954 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1955 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1956 PseudoSourceValue::getJumpTable(), 0,
1958 Chain = Addr.getValue(1);
1959 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1960 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1962 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1963 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
1964 Chain = Addr.getValue(1);
1965 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1969 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1970 // Implement fcopysign with a fabs and a conditional fneg.
1971 SDValue Tmp0 = Op.getOperand(0);
1972 SDValue Tmp1 = Op.getOperand(1);
1973 DebugLoc dl = Op.getDebugLoc();
1974 EVT VT = Op.getValueType();
1975 EVT SrcVT = Tmp1.getValueType();
1976 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1977 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1978 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1979 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1980 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1983 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1984 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1985 MFI->setFrameAddressIsTaken(true);
1986 EVT VT = Op.getValueType();
1987 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1988 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1989 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1990 ? ARM::R7 : ARM::R11;
1991 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1993 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
1999 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2001 SDValue Dst, SDValue Src,
2002 SDValue Size, unsigned Align,
2004 const Value *DstSV, uint64_t DstSVOff,
2005 const Value *SrcSV, uint64_t SrcSVOff){
2006 // Do repeated 4-byte loads and stores. To be improved.
2007 // This requires 4-byte alignment.
2008 if ((Align & 3) != 0)
2010 // This requires the copy size to be a constant, preferrably
2011 // within a subtarget-specific limit.
2012 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2015 uint64_t SizeVal = ConstantSize->getZExtValue();
2016 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2019 unsigned BytesLeft = SizeVal & 3;
2020 unsigned NumMemOps = SizeVal >> 2;
2021 unsigned EmittedNumMemOps = 0;
2023 unsigned VTSize = 4;
2025 const unsigned MAX_LOADS_IN_LDM = 6;
2026 SDValue TFOps[MAX_LOADS_IN_LDM];
2027 SDValue Loads[MAX_LOADS_IN_LDM];
2028 uint64_t SrcOff = 0, DstOff = 0;
2030 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2031 // same number of stores. The loads and stores will get combined into
2032 // ldm/stm later on.
2033 while (EmittedNumMemOps < NumMemOps) {
2035 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2036 Loads[i] = DAG.getLoad(VT, dl, Chain,
2037 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2038 DAG.getConstant(SrcOff, MVT::i32)),
2039 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2040 TFOps[i] = Loads[i].getValue(1);
2043 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2046 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2047 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2048 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2049 DAG.getConstant(DstOff, MVT::i32)),
2050 DstSV, DstSVOff + DstOff, false, false, 0);
2053 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2055 EmittedNumMemOps += i;
2061 // Issue loads / stores for the trailing (1 - 3) bytes.
2062 unsigned BytesLeftSave = BytesLeft;
2065 if (BytesLeft >= 2) {
2073 Loads[i] = DAG.getLoad(VT, dl, Chain,
2074 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2075 DAG.getConstant(SrcOff, MVT::i32)),
2076 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2077 TFOps[i] = Loads[i].getValue(1);
2080 BytesLeft -= VTSize;
2082 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2085 BytesLeft = BytesLeftSave;
2087 if (BytesLeft >= 2) {
2095 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2096 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2097 DAG.getConstant(DstOff, MVT::i32)),
2098 DstSV, DstSVOff + DstOff, false, false, 0);
2101 BytesLeft -= VTSize;
2103 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2106 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2107 SDValue Op = N->getOperand(0);
2108 DebugLoc dl = N->getDebugLoc();
2109 if (N->getValueType(0) == MVT::f64) {
2110 // Turn i64->f64 into VMOVDRR.
2111 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2112 DAG.getConstant(0, MVT::i32));
2113 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2114 DAG.getConstant(1, MVT::i32));
2115 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2118 // Turn f64->i64 into VMOVRRD.
2119 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2120 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2122 // Merge the pieces into a single i64 value.
2123 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2126 /// getZeroVector - Returns a vector of specified type with all zero elements.
2128 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2129 assert(VT.isVector() && "Expected a vector type");
2131 // Zero vectors are used to represent vector negation and in those cases
2132 // will be implemented with the NEON VNEG instruction. However, VNEG does
2133 // not support i64 elements, so sometimes the zero vectors will need to be
2134 // explicitly constructed. For those cases, and potentially other uses in
2135 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2136 // to their dest type. This ensures they get CSE'd.
2138 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2139 SmallVector<SDValue, 8> Ops;
2142 if (VT.getSizeInBits() == 64) {
2143 Ops.assign(8, Cst); TVT = MVT::v8i8;
2145 Ops.assign(16, Cst); TVT = MVT::v16i8;
2147 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2149 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2152 /// getOnesVector - Returns a vector of specified type with all bits set.
2154 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2155 assert(VT.isVector() && "Expected a vector type");
2157 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2158 // dest type. This ensures they get CSE'd.
2160 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2161 SmallVector<SDValue, 8> Ops;
2164 if (VT.getSizeInBits() == 64) {
2165 Ops.assign(8, Cst); TVT = MVT::v8i8;
2167 Ops.assign(16, Cst); TVT = MVT::v16i8;
2169 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2171 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2174 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2175 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2176 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
2177 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2178 EVT VT = Op.getValueType();
2179 unsigned VTBits = VT.getSizeInBits();
2180 DebugLoc dl = Op.getDebugLoc();
2181 SDValue ShOpLo = Op.getOperand(0);
2182 SDValue ShOpHi = Op.getOperand(1);
2183 SDValue ShAmt = Op.getOperand(2);
2185 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2187 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2189 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2190 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2191 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2192 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2193 DAG.getConstant(VTBits, MVT::i32));
2194 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2195 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2196 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2198 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2199 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2201 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2202 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2205 SDValue Ops[2] = { Lo, Hi };
2206 return DAG.getMergeValues(Ops, 2, dl);
2209 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2210 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2211 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
2212 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2213 EVT VT = Op.getValueType();
2214 unsigned VTBits = VT.getSizeInBits();
2215 DebugLoc dl = Op.getDebugLoc();
2216 SDValue ShOpLo = Op.getOperand(0);
2217 SDValue ShOpHi = Op.getOperand(1);
2218 SDValue ShAmt = Op.getOperand(2);
2221 assert(Op.getOpcode() == ISD::SHL_PARTS);
2222 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2223 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2224 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2225 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2226 DAG.getConstant(VTBits, MVT::i32));
2227 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2228 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2230 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2231 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2232 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2234 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2235 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2238 SDValue Ops[2] = { Lo, Hi };
2239 return DAG.getMergeValues(Ops, 2, dl);
2242 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2243 const ARMSubtarget *ST) {
2244 EVT VT = N->getValueType(0);
2245 DebugLoc dl = N->getDebugLoc();
2247 if (!ST->hasV6T2Ops())
2250 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2251 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2254 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2255 const ARMSubtarget *ST) {
2256 EVT VT = N->getValueType(0);
2257 DebugLoc dl = N->getDebugLoc();
2259 // Lower vector shifts on NEON to use VSHL.
2260 if (VT.isVector()) {
2261 assert(ST->hasNEON() && "unexpected vector shift");
2263 // Left shifts translate directly to the vshiftu intrinsic.
2264 if (N->getOpcode() == ISD::SHL)
2265 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2266 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2267 N->getOperand(0), N->getOperand(1));
2269 assert((N->getOpcode() == ISD::SRA ||
2270 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2272 // NEON uses the same intrinsics for both left and right shifts. For
2273 // right shifts, the shift amounts are negative, so negate the vector of
2275 EVT ShiftVT = N->getOperand(1).getValueType();
2276 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2277 getZeroVector(ShiftVT, DAG, dl),
2279 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2280 Intrinsic::arm_neon_vshifts :
2281 Intrinsic::arm_neon_vshiftu);
2282 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2283 DAG.getConstant(vshiftInt, MVT::i32),
2284 N->getOperand(0), NegatedCount);
2287 // We can get here for a node like i32 = ISD::SHL i32, i64
2291 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2292 "Unknown shift to lower!");
2294 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2295 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2296 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2299 // If we are in thumb mode, we don't have RRX.
2300 if (ST->isThumb1Only()) return SDValue();
2302 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2303 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2304 DAG.getConstant(0, MVT::i32));
2305 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2306 DAG.getConstant(1, MVT::i32));
2308 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2309 // captures the result into a carry flag.
2310 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2311 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2313 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2314 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2316 // Merge the pieces into a single i64 value.
2317 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2320 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2321 SDValue TmpOp0, TmpOp1;
2322 bool Invert = false;
2326 SDValue Op0 = Op.getOperand(0);
2327 SDValue Op1 = Op.getOperand(1);
2328 SDValue CC = Op.getOperand(2);
2329 EVT VT = Op.getValueType();
2330 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2331 DebugLoc dl = Op.getDebugLoc();
2333 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2334 switch (SetCCOpcode) {
2335 default: llvm_unreachable("Illegal FP comparison"); break;
2337 case ISD::SETNE: Invert = true; // Fallthrough
2339 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2341 case ISD::SETLT: Swap = true; // Fallthrough
2343 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2345 case ISD::SETLE: Swap = true; // Fallthrough
2347 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2348 case ISD::SETUGE: Swap = true; // Fallthrough
2349 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2350 case ISD::SETUGT: Swap = true; // Fallthrough
2351 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2352 case ISD::SETUEQ: Invert = true; // Fallthrough
2354 // Expand this to (OLT | OGT).
2358 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2359 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2361 case ISD::SETUO: Invert = true; // Fallthrough
2363 // Expand this to (OLT | OGE).
2367 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2368 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2372 // Integer comparisons.
2373 switch (SetCCOpcode) {
2374 default: llvm_unreachable("Illegal integer comparison"); break;
2375 case ISD::SETNE: Invert = true;
2376 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2377 case ISD::SETLT: Swap = true;
2378 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2379 case ISD::SETLE: Swap = true;
2380 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2381 case ISD::SETULT: Swap = true;
2382 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2383 case ISD::SETULE: Swap = true;
2384 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2387 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2388 if (Opc == ARMISD::VCEQ) {
2391 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2393 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2396 // Ignore bitconvert.
2397 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2398 AndOp = AndOp.getOperand(0);
2400 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2402 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2403 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2410 std::swap(Op0, Op1);
2412 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2415 Result = DAG.getNOT(dl, Result, VT);
2420 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2421 /// VMOV instruction, and if so, return the constant being splatted.
2422 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2423 unsigned SplatBitSize, SelectionDAG &DAG) {
2424 switch (SplatBitSize) {
2426 // Any 1-byte value is OK.
2427 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2428 return DAG.getTargetConstant(SplatBits, MVT::i8);
2431 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2432 if ((SplatBits & ~0xff) == 0 ||
2433 (SplatBits & ~0xff00) == 0)
2434 return DAG.getTargetConstant(SplatBits, MVT::i16);
2438 // NEON's 32-bit VMOV supports splat values where:
2439 // * only one byte is nonzero, or
2440 // * the least significant byte is 0xff and the second byte is nonzero, or
2441 // * the least significant 2 bytes are 0xff and the third is nonzero.
2442 if ((SplatBits & ~0xff) == 0 ||
2443 (SplatBits & ~0xff00) == 0 ||
2444 (SplatBits & ~0xff0000) == 0 ||
2445 (SplatBits & ~0xff000000) == 0)
2446 return DAG.getTargetConstant(SplatBits, MVT::i32);
2448 if ((SplatBits & ~0xffff) == 0 &&
2449 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2450 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2452 if ((SplatBits & ~0xffffff) == 0 &&
2453 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2454 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2456 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2457 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2458 // VMOV.I32. A (very) minor optimization would be to replicate the value
2459 // and fall through here to test for a valid 64-bit splat. But, then the
2460 // caller would also need to check and handle the change in size.
2464 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2465 uint64_t BitMask = 0xff;
2467 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2468 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2470 else if ((SplatBits & BitMask) != 0)
2474 return DAG.getTargetConstant(Val, MVT::i64);
2478 llvm_unreachable("unexpected size for isVMOVSplat");
2485 /// getVMOVImm - If this is a build_vector of constants which can be
2486 /// formed by using a VMOV instruction of the specified element size,
2487 /// return the constant being splatted. The ByteSize field indicates the
2488 /// number of bytes of each element [1248].
2489 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2490 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2491 APInt SplatBits, SplatUndef;
2492 unsigned SplatBitSize;
2494 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2495 HasAnyUndefs, ByteSize * 8))
2498 if (SplatBitSize > ByteSize * 8)
2501 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2505 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2506 bool &ReverseVEXT, unsigned &Imm) {
2507 unsigned NumElts = VT.getVectorNumElements();
2508 ReverseVEXT = false;
2511 // If this is a VEXT shuffle, the immediate value is the index of the first
2512 // element. The other shuffle indices must be the successive elements after
2514 unsigned ExpectedElt = Imm;
2515 for (unsigned i = 1; i < NumElts; ++i) {
2516 // Increment the expected index. If it wraps around, it may still be
2517 // a VEXT but the source vectors must be swapped.
2519 if (ExpectedElt == NumElts * 2) {
2524 if (ExpectedElt != static_cast<unsigned>(M[i]))
2528 // Adjust the index value if the source operands will be swapped.
2535 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2536 /// instruction with the specified blocksize. (The order of the elements
2537 /// within each block of the vector is reversed.)
2538 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2539 unsigned BlockSize) {
2540 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2541 "Only possible block sizes for VREV are: 16, 32, 64");
2543 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2547 unsigned NumElts = VT.getVectorNumElements();
2548 unsigned BlockElts = M[0] + 1;
2550 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2553 for (unsigned i = 0; i < NumElts; ++i) {
2554 if ((unsigned) M[i] !=
2555 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2562 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2563 unsigned &WhichResult) {
2564 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2568 unsigned NumElts = VT.getVectorNumElements();
2569 WhichResult = (M[0] == 0 ? 0 : 1);
2570 for (unsigned i = 0; i < NumElts; i += 2) {
2571 if ((unsigned) M[i] != i + WhichResult ||
2572 (unsigned) M[i+1] != i + NumElts + WhichResult)
2578 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2579 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2580 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2581 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2582 unsigned &WhichResult) {
2583 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2587 unsigned NumElts = VT.getVectorNumElements();
2588 WhichResult = (M[0] == 0 ? 0 : 1);
2589 for (unsigned i = 0; i < NumElts; i += 2) {
2590 if ((unsigned) M[i] != i + WhichResult ||
2591 (unsigned) M[i+1] != i + WhichResult)
2597 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2598 unsigned &WhichResult) {
2599 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2603 unsigned NumElts = VT.getVectorNumElements();
2604 WhichResult = (M[0] == 0 ? 0 : 1);
2605 for (unsigned i = 0; i != NumElts; ++i) {
2606 if ((unsigned) M[i] != 2 * i + WhichResult)
2610 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2611 if (VT.is64BitVector() && EltSz == 32)
2617 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2618 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2619 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2620 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2621 unsigned &WhichResult) {
2622 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2626 unsigned Half = VT.getVectorNumElements() / 2;
2627 WhichResult = (M[0] == 0 ? 0 : 1);
2628 for (unsigned j = 0; j != 2; ++j) {
2629 unsigned Idx = WhichResult;
2630 for (unsigned i = 0; i != Half; ++i) {
2631 if ((unsigned) M[i + j * Half] != Idx)
2637 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2638 if (VT.is64BitVector() && EltSz == 32)
2644 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2645 unsigned &WhichResult) {
2646 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2650 unsigned NumElts = VT.getVectorNumElements();
2651 WhichResult = (M[0] == 0 ? 0 : 1);
2652 unsigned Idx = WhichResult * NumElts / 2;
2653 for (unsigned i = 0; i != NumElts; i += 2) {
2654 if ((unsigned) M[i] != Idx ||
2655 (unsigned) M[i+1] != Idx + NumElts)
2660 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2661 if (VT.is64BitVector() && EltSz == 32)
2667 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2668 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2669 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2670 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2671 unsigned &WhichResult) {
2672 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2676 unsigned NumElts = VT.getVectorNumElements();
2677 WhichResult = (M[0] == 0 ? 0 : 1);
2678 unsigned Idx = WhichResult * NumElts / 2;
2679 for (unsigned i = 0; i != NumElts; i += 2) {
2680 if ((unsigned) M[i] != Idx ||
2681 (unsigned) M[i+1] != Idx)
2686 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2687 if (VT.is64BitVector() && EltSz == 32)
2694 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2695 // Canonicalize all-zeros and all-ones vectors.
2696 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2697 if (ConstVal->isNullValue())
2698 return getZeroVector(VT, DAG, dl);
2699 if (ConstVal->isAllOnesValue())
2700 return getOnesVector(VT, DAG, dl);
2703 if (VT.is64BitVector()) {
2704 switch (Val.getValueType().getSizeInBits()) {
2705 case 8: CanonicalVT = MVT::v8i8; break;
2706 case 16: CanonicalVT = MVT::v4i16; break;
2707 case 32: CanonicalVT = MVT::v2i32; break;
2708 case 64: CanonicalVT = MVT::v1i64; break;
2709 default: llvm_unreachable("unexpected splat element type"); break;
2712 assert(VT.is128BitVector() && "unknown splat vector size");
2713 switch (Val.getValueType().getSizeInBits()) {
2714 case 8: CanonicalVT = MVT::v16i8; break;
2715 case 16: CanonicalVT = MVT::v8i16; break;
2716 case 32: CanonicalVT = MVT::v4i32; break;
2717 case 64: CanonicalVT = MVT::v2i64; break;
2718 default: llvm_unreachable("unexpected splat element type"); break;
2722 // Build a canonical splat for this value.
2723 SmallVector<SDValue, 8> Ops;
2724 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2725 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2727 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2730 // If this is a case we can't handle, return null and let the default
2731 // expansion code take care of it.
2732 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2733 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2734 DebugLoc dl = Op.getDebugLoc();
2735 EVT VT = Op.getValueType();
2737 APInt SplatBits, SplatUndef;
2738 unsigned SplatBitSize;
2740 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2741 if (SplatBitSize <= 64) {
2742 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2743 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2745 return BuildSplat(Val, VT, DAG, dl);
2749 // If there are only 2 elements in a 128-bit vector, insert them into an
2750 // undef vector. This handles the common case for 128-bit vector argument
2751 // passing, where the insertions should be translated to subreg accesses
2752 // with no real instructions.
2753 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2754 SDValue Val = DAG.getUNDEF(VT);
2755 SDValue Op0 = Op.getOperand(0);
2756 SDValue Op1 = Op.getOperand(1);
2757 if (Op0.getOpcode() != ISD::UNDEF)
2758 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2759 DAG.getIntPtrConstant(0));
2760 if (Op1.getOpcode() != ISD::UNDEF)
2761 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2762 DAG.getIntPtrConstant(1));
2769 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2770 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2771 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2772 /// are assumed to be legal.
2774 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2776 if (VT.getVectorNumElements() == 4 &&
2777 (VT.is128BitVector() || VT.is64BitVector())) {
2778 unsigned PFIndexes[4];
2779 for (unsigned i = 0; i != 4; ++i) {
2783 PFIndexes[i] = M[i];
2786 // Compute the index in the perfect shuffle table.
2787 unsigned PFTableIndex =
2788 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2789 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2790 unsigned Cost = (PFEntry >> 30);
2797 unsigned Imm, WhichResult;
2799 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2800 isVREVMask(M, VT, 64) ||
2801 isVREVMask(M, VT, 32) ||
2802 isVREVMask(M, VT, 16) ||
2803 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2804 isVTRNMask(M, VT, WhichResult) ||
2805 isVUZPMask(M, VT, WhichResult) ||
2806 isVZIPMask(M, VT, WhichResult) ||
2807 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2808 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2809 isVZIP_v_undef_Mask(M, VT, WhichResult));
2812 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2813 /// the specified operations to build the shuffle.
2814 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2815 SDValue RHS, SelectionDAG &DAG,
2817 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2818 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2819 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2822 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2831 OP_VUZPL, // VUZP, left result
2832 OP_VUZPR, // VUZP, right result
2833 OP_VZIPL, // VZIP, left result
2834 OP_VZIPR, // VZIP, right result
2835 OP_VTRNL, // VTRN, left result
2836 OP_VTRNR // VTRN, right result
2839 if (OpNum == OP_COPY) {
2840 if (LHSID == (1*9+2)*9+3) return LHS;
2841 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2845 SDValue OpLHS, OpRHS;
2846 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2847 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2848 EVT VT = OpLHS.getValueType();
2851 default: llvm_unreachable("Unknown shuffle opcode!");
2853 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2858 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2859 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2863 return DAG.getNode(ARMISD::VEXT, dl, VT,
2865 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2868 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2869 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2872 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2873 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2876 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2877 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2881 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2882 SDValue V1 = Op.getOperand(0);
2883 SDValue V2 = Op.getOperand(1);
2884 DebugLoc dl = Op.getDebugLoc();
2885 EVT VT = Op.getValueType();
2886 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2887 SmallVector<int, 8> ShuffleMask;
2889 // Convert shuffles that are directly supported on NEON to target-specific
2890 // DAG nodes, instead of keeping them as shuffles and matching them again
2891 // during code selection. This is more efficient and avoids the possibility
2892 // of inconsistencies between legalization and selection.
2893 // FIXME: floating-point vectors should be canonicalized to integer vectors
2894 // of the same time so that they get CSEd properly.
2895 SVN->getMask(ShuffleMask);
2897 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2898 int Lane = SVN->getSplatIndex();
2899 // If this is undef splat, generate it via "just" vdup, if possible.
2900 if (Lane == -1) Lane = 0;
2902 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2903 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2905 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2906 DAG.getConstant(Lane, MVT::i32));
2911 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2914 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2915 DAG.getConstant(Imm, MVT::i32));
2918 if (isVREVMask(ShuffleMask, VT, 64))
2919 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2920 if (isVREVMask(ShuffleMask, VT, 32))
2921 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2922 if (isVREVMask(ShuffleMask, VT, 16))
2923 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2925 // Check for Neon shuffles that modify both input vectors in place.
2926 // If both results are used, i.e., if there are two shuffles with the same
2927 // source operands and with masks corresponding to both results of one of
2928 // these operations, DAG memoization will ensure that a single node is
2929 // used for both shuffles.
2930 unsigned WhichResult;
2931 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2932 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2933 V1, V2).getValue(WhichResult);
2934 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2935 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2936 V1, V2).getValue(WhichResult);
2937 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2938 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2939 V1, V2).getValue(WhichResult);
2941 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
2942 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2943 V1, V1).getValue(WhichResult);
2944 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2945 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2946 V1, V1).getValue(WhichResult);
2947 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2948 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2949 V1, V1).getValue(WhichResult);
2951 // If the shuffle is not directly supported and it has 4 elements, use
2952 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2953 if (VT.getVectorNumElements() == 4 &&
2954 (VT.is128BitVector() || VT.is64BitVector())) {
2955 unsigned PFIndexes[4];
2956 for (unsigned i = 0; i != 4; ++i) {
2957 if (ShuffleMask[i] < 0)
2960 PFIndexes[i] = ShuffleMask[i];
2963 // Compute the index in the perfect shuffle table.
2964 unsigned PFTableIndex =
2965 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2967 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2968 unsigned Cost = (PFEntry >> 30);
2971 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2977 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2978 EVT VT = Op.getValueType();
2979 DebugLoc dl = Op.getDebugLoc();
2980 SDValue Vec = Op.getOperand(0);
2981 SDValue Lane = Op.getOperand(1);
2982 assert(VT == MVT::i32 &&
2983 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
2984 "unexpected type for custom-lowering vector extract");
2985 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2988 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2989 // The only time a CONCAT_VECTORS operation can have legal types is when
2990 // two 64-bit vectors are concatenated to a 128-bit vector.
2991 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2992 "unexpected CONCAT_VECTORS");
2993 DebugLoc dl = Op.getDebugLoc();
2994 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2995 SDValue Op0 = Op.getOperand(0);
2996 SDValue Op1 = Op.getOperand(1);
2997 if (Op0.getOpcode() != ISD::UNDEF)
2998 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2999 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3000 DAG.getIntPtrConstant(0));
3001 if (Op1.getOpcode() != ISD::UNDEF)
3002 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3003 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3004 DAG.getIntPtrConstant(1));
3005 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3008 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3009 switch (Op.getOpcode()) {
3010 default: llvm_unreachable("Don't know how to custom lower this!");
3011 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3012 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3013 case ISD::GlobalAddress:
3014 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3015 LowerGlobalAddressELF(Op, DAG);
3016 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3017 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3018 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3019 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3020 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3021 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
3022 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3023 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3024 case ISD::RETURNADDR: break;
3025 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3026 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3027 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3029 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3032 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3033 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3034 case ISD::SRL_PARTS:
3035 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3036 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3037 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3038 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3039 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3040 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3041 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3046 /// ReplaceNodeResults - Replace the results of node with an illegal result
3047 /// type with new values built out of custom code.
3048 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3049 SmallVectorImpl<SDValue>&Results,
3050 SelectionDAG &DAG) {
3051 switch (N->getOpcode()) {
3053 llvm_unreachable("Don't know how to custom expand this!");
3055 case ISD::BIT_CONVERT:
3056 Results.push_back(ExpandBIT_CONVERT(N, DAG));
3060 SDValue Res = LowerShift(N, DAG, Subtarget);
3062 Results.push_back(Res);
3068 //===----------------------------------------------------------------------===//
3069 // ARM Scheduler Hooks
3070 //===----------------------------------------------------------------------===//
3073 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3074 MachineBasicBlock *BB,
3075 unsigned Size) const {
3076 unsigned dest = MI->getOperand(0).getReg();
3077 unsigned ptr = MI->getOperand(1).getReg();
3078 unsigned oldval = MI->getOperand(2).getReg();
3079 unsigned newval = MI->getOperand(3).getReg();
3080 unsigned scratch = BB->getParent()->getRegInfo()
3081 .createVirtualRegister(ARM::GPRRegisterClass);
3082 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3083 DebugLoc dl = MI->getDebugLoc();
3084 bool isThumb2 = Subtarget->isThumb2();
3086 unsigned ldrOpc, strOpc;
3088 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3090 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3091 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3094 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3095 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3098 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3099 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3103 MachineFunction *MF = BB->getParent();
3104 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3105 MachineFunction::iterator It = BB;
3106 ++It; // insert the new blocks after the current block
3108 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3109 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3110 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3111 MF->insert(It, loop1MBB);
3112 MF->insert(It, loop2MBB);
3113 MF->insert(It, exitMBB);
3114 exitMBB->transferSuccessors(BB);
3118 // fallthrough --> loop1MBB
3119 BB->addSuccessor(loop1MBB);
3122 // ldrex dest, [ptr]
3126 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3127 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3128 .addReg(dest).addReg(oldval));
3129 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3130 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3131 BB->addSuccessor(loop2MBB);
3132 BB->addSuccessor(exitMBB);
3135 // strex scratch, newval, [ptr]
3139 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3141 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3142 .addReg(scratch).addImm(0));
3143 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3144 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3145 BB->addSuccessor(loop1MBB);
3146 BB->addSuccessor(exitMBB);
3152 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3158 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3159 unsigned Size, unsigned BinOpcode) const {
3160 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3161 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3163 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3164 MachineFunction *MF = BB->getParent();
3165 MachineFunction::iterator It = BB;
3168 unsigned dest = MI->getOperand(0).getReg();
3169 unsigned ptr = MI->getOperand(1).getReg();
3170 unsigned incr = MI->getOperand(2).getReg();
3171 DebugLoc dl = MI->getDebugLoc();
3173 bool isThumb2 = Subtarget->isThumb2();
3174 unsigned ldrOpc, strOpc;
3176 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3178 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3179 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3182 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3183 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3186 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3187 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3191 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3192 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3193 MF->insert(It, loopMBB);
3194 MF->insert(It, exitMBB);
3195 exitMBB->transferSuccessors(BB);
3197 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3198 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3199 unsigned scratch2 = (!BinOpcode) ? incr :
3200 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3204 // fallthrough --> loopMBB
3205 BB->addSuccessor(loopMBB);
3209 // <binop> scratch2, dest, incr
3210 // strex scratch, scratch2, ptr
3213 // fallthrough --> exitMBB
3215 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3217 // operand order needs to go the other way for NAND
3218 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3219 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3220 addReg(incr).addReg(dest)).addReg(0);
3222 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3223 addReg(dest).addReg(incr)).addReg(0);
3226 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3228 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3229 .addReg(scratch).addImm(0));
3230 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3231 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3233 BB->addSuccessor(loopMBB);
3234 BB->addSuccessor(exitMBB);
3240 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3246 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3247 MachineBasicBlock *BB,
3248 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
3249 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3250 DebugLoc dl = MI->getDebugLoc();
3251 bool isThumb2 = Subtarget->isThumb2();
3252 switch (MI->getOpcode()) {
3255 llvm_unreachable("Unexpected instr type to insert");
3257 case ARM::ATOMIC_LOAD_ADD_I8:
3258 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3259 case ARM::ATOMIC_LOAD_ADD_I16:
3260 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3261 case ARM::ATOMIC_LOAD_ADD_I32:
3262 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3264 case ARM::ATOMIC_LOAD_AND_I8:
3265 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3266 case ARM::ATOMIC_LOAD_AND_I16:
3267 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3268 case ARM::ATOMIC_LOAD_AND_I32:
3269 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3271 case ARM::ATOMIC_LOAD_OR_I8:
3272 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3273 case ARM::ATOMIC_LOAD_OR_I16:
3274 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3275 case ARM::ATOMIC_LOAD_OR_I32:
3276 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3278 case ARM::ATOMIC_LOAD_XOR_I8:
3279 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3280 case ARM::ATOMIC_LOAD_XOR_I16:
3281 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3282 case ARM::ATOMIC_LOAD_XOR_I32:
3283 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3285 case ARM::ATOMIC_LOAD_NAND_I8:
3286 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3287 case ARM::ATOMIC_LOAD_NAND_I16:
3288 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3289 case ARM::ATOMIC_LOAD_NAND_I32:
3290 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3292 case ARM::ATOMIC_LOAD_SUB_I8:
3293 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3294 case ARM::ATOMIC_LOAD_SUB_I16:
3295 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3296 case ARM::ATOMIC_LOAD_SUB_I32:
3297 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3299 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3300 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3301 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3303 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3304 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3305 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3307 case ARM::tMOVCCr_pseudo: {
3308 // To "insert" a SELECT_CC instruction, we actually have to insert the
3309 // diamond control-flow pattern. The incoming instruction knows the
3310 // destination vreg to set, the condition code register to branch on, the
3311 // true/false values to select between, and a branch opcode to use.
3312 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3313 MachineFunction::iterator It = BB;
3319 // cmpTY ccX, r1, r2
3321 // fallthrough --> copy0MBB
3322 MachineBasicBlock *thisMBB = BB;
3323 MachineFunction *F = BB->getParent();
3324 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3325 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3326 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3327 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3328 F->insert(It, copy0MBB);
3329 F->insert(It, sinkMBB);
3330 // Update machine-CFG edges by first adding all successors of the current
3331 // block to the new block which will contain the Phi node for the select.
3332 // Also inform sdisel of the edge changes.
3333 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3334 E = BB->succ_end(); I != E; ++I) {
3335 EM->insert(std::make_pair(*I, sinkMBB));
3336 sinkMBB->addSuccessor(*I);
3338 // Next, remove all successors of the current block, and add the true
3339 // and fallthrough blocks as its successors.
3340 while (!BB->succ_empty())
3341 BB->removeSuccessor(BB->succ_begin());
3342 BB->addSuccessor(copy0MBB);
3343 BB->addSuccessor(sinkMBB);
3346 // %FalseValue = ...
3347 // # fallthrough to sinkMBB
3350 // Update machine-CFG edges
3351 BB->addSuccessor(sinkMBB);
3354 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3357 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3358 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3359 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3361 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3368 case ARM::t2SUBrSPi_:
3369 case ARM::t2SUBrSPi12_:
3370 case ARM::t2SUBrSPs_: {
3371 MachineFunction *MF = BB->getParent();
3372 unsigned DstReg = MI->getOperand(0).getReg();
3373 unsigned SrcReg = MI->getOperand(1).getReg();
3374 bool DstIsDead = MI->getOperand(0).isDead();
3375 bool SrcIsKill = MI->getOperand(1).isKill();
3377 if (SrcReg != ARM::SP) {
3378 // Copy the source to SP from virtual register.
3379 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3380 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3381 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3382 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3383 .addReg(SrcReg, getKillRegState(SrcIsKill));
3387 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3388 switch (MI->getOpcode()) {
3390 llvm_unreachable("Unexpected pseudo instruction!");
3396 OpOpc = ARM::tADDspr;
3399 OpOpc = ARM::tSUBspi;
3401 case ARM::t2SUBrSPi_:
3402 OpOpc = ARM::t2SUBrSPi;
3403 NeedPred = true; NeedCC = true;
3405 case ARM::t2SUBrSPi12_:
3406 OpOpc = ARM::t2SUBrSPi12;
3409 case ARM::t2SUBrSPs_:
3410 OpOpc = ARM::t2SUBrSPs;
3411 NeedPred = true; NeedCC = true; NeedOp3 = true;
3414 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3415 if (OpOpc == ARM::tAND)
3416 AddDefaultT1CC(MIB);
3417 MIB.addReg(ARM::SP);
3418 MIB.addOperand(MI->getOperand(2));
3420 MIB.addOperand(MI->getOperand(3));
3422 AddDefaultPred(MIB);
3426 // Copy the result from SP to virtual register.
3427 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3428 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3429 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3430 BuildMI(BB, dl, TII->get(CopyOpc))
3431 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3433 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3439 //===----------------------------------------------------------------------===//
3440 // ARM Optimization Hooks
3441 //===----------------------------------------------------------------------===//
3444 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3445 TargetLowering::DAGCombinerInfo &DCI) {
3446 SelectionDAG &DAG = DCI.DAG;
3447 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3448 EVT VT = N->getValueType(0);
3449 unsigned Opc = N->getOpcode();
3450 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3451 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3452 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3453 ISD::CondCode CC = ISD::SETCC_INVALID;
3456 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3458 SDValue CCOp = Slct.getOperand(0);
3459 if (CCOp.getOpcode() == ISD::SETCC)
3460 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3463 bool DoXform = false;
3465 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3468 if (LHS.getOpcode() == ISD::Constant &&
3469 cast<ConstantSDNode>(LHS)->isNullValue()) {
3471 } else if (CC != ISD::SETCC_INVALID &&
3472 RHS.getOpcode() == ISD::Constant &&
3473 cast<ConstantSDNode>(RHS)->isNullValue()) {
3474 std::swap(LHS, RHS);
3475 SDValue Op0 = Slct.getOperand(0);
3476 EVT OpVT = isSlctCC ? Op0.getValueType() :
3477 Op0.getOperand(0).getValueType();
3478 bool isInt = OpVT.isInteger();
3479 CC = ISD::getSetCCInverse(CC, isInt);
3481 if (!TLI.isCondCodeLegal(CC, OpVT))
3482 return SDValue(); // Inverse operator isn't legal.
3489 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3491 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3492 Slct.getOperand(0), Slct.getOperand(1), CC);
3493 SDValue CCOp = Slct.getOperand(0);
3495 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3496 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3497 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3498 CCOp, OtherOp, Result);
3503 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3504 static SDValue PerformADDCombine(SDNode *N,
3505 TargetLowering::DAGCombinerInfo &DCI) {
3506 // added by evan in r37685 with no testcase.
3507 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3509 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3510 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3511 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3512 if (Result.getNode()) return Result;
3514 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3515 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3516 if (Result.getNode()) return Result;
3522 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3523 static SDValue PerformSUBCombine(SDNode *N,
3524 TargetLowering::DAGCombinerInfo &DCI) {
3525 // added by evan in r37685 with no testcase.
3526 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3528 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3529 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3530 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3531 if (Result.getNode()) return Result;
3537 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3538 /// ARMISD::VMOVRRD.
3539 static SDValue PerformVMOVRRDCombine(SDNode *N,
3540 TargetLowering::DAGCombinerInfo &DCI) {
3541 // fmrrd(fmdrr x, y) -> x,y
3542 SDValue InDouble = N->getOperand(0);
3543 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3544 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3548 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3549 /// operand of a vector shift operation, where all the elements of the
3550 /// build_vector must have the same constant integer value.
3551 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3552 // Ignore bit_converts.
3553 while (Op.getOpcode() == ISD::BIT_CONVERT)
3554 Op = Op.getOperand(0);
3555 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3556 APInt SplatBits, SplatUndef;
3557 unsigned SplatBitSize;
3559 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3560 HasAnyUndefs, ElementBits) ||
3561 SplatBitSize > ElementBits)
3563 Cnt = SplatBits.getSExtValue();
3567 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3568 /// operand of a vector shift left operation. That value must be in the range:
3569 /// 0 <= Value < ElementBits for a left shift; or
3570 /// 0 <= Value <= ElementBits for a long left shift.
3571 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3572 assert(VT.isVector() && "vector shift count is not a vector type");
3573 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3574 if (! getVShiftImm(Op, ElementBits, Cnt))
3576 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3579 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3580 /// operand of a vector shift right operation. For a shift opcode, the value
3581 /// is positive, but for an intrinsic the value count must be negative. The
3582 /// absolute value must be in the range:
3583 /// 1 <= |Value| <= ElementBits for a right shift; or
3584 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3585 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3587 assert(VT.isVector() && "vector shift count is not a vector type");
3588 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3589 if (! getVShiftImm(Op, ElementBits, Cnt))
3593 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3596 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3597 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3598 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3601 // Don't do anything for most intrinsics.
3604 // Vector shifts: check for immediate versions and lower them.
3605 // Note: This is done during DAG combining instead of DAG legalizing because
3606 // the build_vectors for 64-bit vector element shift counts are generally
3607 // not legal, and it is hard to see their values after they get legalized to
3608 // loads from a constant pool.
3609 case Intrinsic::arm_neon_vshifts:
3610 case Intrinsic::arm_neon_vshiftu:
3611 case Intrinsic::arm_neon_vshiftls:
3612 case Intrinsic::arm_neon_vshiftlu:
3613 case Intrinsic::arm_neon_vshiftn:
3614 case Intrinsic::arm_neon_vrshifts:
3615 case Intrinsic::arm_neon_vrshiftu:
3616 case Intrinsic::arm_neon_vrshiftn:
3617 case Intrinsic::arm_neon_vqshifts:
3618 case Intrinsic::arm_neon_vqshiftu:
3619 case Intrinsic::arm_neon_vqshiftsu:
3620 case Intrinsic::arm_neon_vqshiftns:
3621 case Intrinsic::arm_neon_vqshiftnu:
3622 case Intrinsic::arm_neon_vqshiftnsu:
3623 case Intrinsic::arm_neon_vqrshiftns:
3624 case Intrinsic::arm_neon_vqrshiftnu:
3625 case Intrinsic::arm_neon_vqrshiftnsu: {
3626 EVT VT = N->getOperand(1).getValueType();
3628 unsigned VShiftOpc = 0;
3631 case Intrinsic::arm_neon_vshifts:
3632 case Intrinsic::arm_neon_vshiftu:
3633 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3634 VShiftOpc = ARMISD::VSHL;
3637 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3638 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3639 ARMISD::VSHRs : ARMISD::VSHRu);
3644 case Intrinsic::arm_neon_vshiftls:
3645 case Intrinsic::arm_neon_vshiftlu:
3646 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3648 llvm_unreachable("invalid shift count for vshll intrinsic");
3650 case Intrinsic::arm_neon_vrshifts:
3651 case Intrinsic::arm_neon_vrshiftu:
3652 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3656 case Intrinsic::arm_neon_vqshifts:
3657 case Intrinsic::arm_neon_vqshiftu:
3658 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3662 case Intrinsic::arm_neon_vqshiftsu:
3663 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3665 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3667 case Intrinsic::arm_neon_vshiftn:
3668 case Intrinsic::arm_neon_vrshiftn:
3669 case Intrinsic::arm_neon_vqshiftns:
3670 case Intrinsic::arm_neon_vqshiftnu:
3671 case Intrinsic::arm_neon_vqshiftnsu:
3672 case Intrinsic::arm_neon_vqrshiftns:
3673 case Intrinsic::arm_neon_vqrshiftnu:
3674 case Intrinsic::arm_neon_vqrshiftnsu:
3675 // Narrowing shifts require an immediate right shift.
3676 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3678 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3681 llvm_unreachable("unhandled vector shift");
3685 case Intrinsic::arm_neon_vshifts:
3686 case Intrinsic::arm_neon_vshiftu:
3687 // Opcode already set above.
3689 case Intrinsic::arm_neon_vshiftls:
3690 case Intrinsic::arm_neon_vshiftlu:
3691 if (Cnt == VT.getVectorElementType().getSizeInBits())
3692 VShiftOpc = ARMISD::VSHLLi;
3694 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3695 ARMISD::VSHLLs : ARMISD::VSHLLu);
3697 case Intrinsic::arm_neon_vshiftn:
3698 VShiftOpc = ARMISD::VSHRN; break;
3699 case Intrinsic::arm_neon_vrshifts:
3700 VShiftOpc = ARMISD::VRSHRs; break;
3701 case Intrinsic::arm_neon_vrshiftu:
3702 VShiftOpc = ARMISD::VRSHRu; break;
3703 case Intrinsic::arm_neon_vrshiftn:
3704 VShiftOpc = ARMISD::VRSHRN; break;
3705 case Intrinsic::arm_neon_vqshifts:
3706 VShiftOpc = ARMISD::VQSHLs; break;
3707 case Intrinsic::arm_neon_vqshiftu:
3708 VShiftOpc = ARMISD::VQSHLu; break;
3709 case Intrinsic::arm_neon_vqshiftsu:
3710 VShiftOpc = ARMISD::VQSHLsu; break;
3711 case Intrinsic::arm_neon_vqshiftns:
3712 VShiftOpc = ARMISD::VQSHRNs; break;
3713 case Intrinsic::arm_neon_vqshiftnu:
3714 VShiftOpc = ARMISD::VQSHRNu; break;
3715 case Intrinsic::arm_neon_vqshiftnsu:
3716 VShiftOpc = ARMISD::VQSHRNsu; break;
3717 case Intrinsic::arm_neon_vqrshiftns:
3718 VShiftOpc = ARMISD::VQRSHRNs; break;
3719 case Intrinsic::arm_neon_vqrshiftnu:
3720 VShiftOpc = ARMISD::VQRSHRNu; break;
3721 case Intrinsic::arm_neon_vqrshiftnsu:
3722 VShiftOpc = ARMISD::VQRSHRNsu; break;
3725 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3726 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3729 case Intrinsic::arm_neon_vshiftins: {
3730 EVT VT = N->getOperand(1).getValueType();
3732 unsigned VShiftOpc = 0;
3734 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3735 VShiftOpc = ARMISD::VSLI;
3736 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3737 VShiftOpc = ARMISD::VSRI;
3739 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3742 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3743 N->getOperand(1), N->getOperand(2),
3744 DAG.getConstant(Cnt, MVT::i32));
3747 case Intrinsic::arm_neon_vqrshifts:
3748 case Intrinsic::arm_neon_vqrshiftu:
3749 // No immediate versions of these to check for.
3756 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3757 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3758 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3759 /// vector element shift counts are generally not legal, and it is hard to see
3760 /// their values after they get legalized to loads from a constant pool.
3761 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3762 const ARMSubtarget *ST) {
3763 EVT VT = N->getValueType(0);
3765 // Nothing to be done for scalar shifts.
3766 if (! VT.isVector())
3769 assert(ST->hasNEON() && "unexpected vector shift");
3772 switch (N->getOpcode()) {
3773 default: llvm_unreachable("unexpected shift opcode");
3776 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3777 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3778 DAG.getConstant(Cnt, MVT::i32));
3783 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3784 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3785 ARMISD::VSHRs : ARMISD::VSHRu);
3786 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3787 DAG.getConstant(Cnt, MVT::i32));
3793 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3794 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3795 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3796 const ARMSubtarget *ST) {
3797 SDValue N0 = N->getOperand(0);
3799 // Check for sign- and zero-extensions of vector extract operations of 8-
3800 // and 16-bit vector elements. NEON supports these directly. They are
3801 // handled during DAG combining because type legalization will promote them
3802 // to 32-bit types and it is messy to recognize the operations after that.
3803 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3804 SDValue Vec = N0.getOperand(0);
3805 SDValue Lane = N0.getOperand(1);
3806 EVT VT = N->getValueType(0);
3807 EVT EltVT = N0.getValueType();
3808 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3810 if (VT == MVT::i32 &&
3811 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3812 TLI.isTypeLegal(Vec.getValueType())) {
3815 switch (N->getOpcode()) {
3816 default: llvm_unreachable("unexpected opcode");
3817 case ISD::SIGN_EXTEND:
3818 Opc = ARMISD::VGETLANEs;
3820 case ISD::ZERO_EXTEND:
3821 case ISD::ANY_EXTEND:
3822 Opc = ARMISD::VGETLANEu;
3825 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3832 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3833 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3834 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3835 const ARMSubtarget *ST) {
3836 // If the target supports NEON, try to use vmax/vmin instructions for f32
3837 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3838 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3839 // a NaN; only do the transformation when it matches that behavior.
3841 // For now only do this when using NEON for FP operations; if using VFP, it
3842 // is not obvious that the benefit outweighs the cost of switching to the
3844 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3845 N->getValueType(0) != MVT::f32)
3848 SDValue CondLHS = N->getOperand(0);
3849 SDValue CondRHS = N->getOperand(1);
3850 SDValue LHS = N->getOperand(2);
3851 SDValue RHS = N->getOperand(3);
3852 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3854 unsigned Opcode = 0;
3856 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3857 IsReversed = false; // x CC y ? x : y
3858 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3859 IsReversed = true ; // x CC y ? y : x
3873 // If LHS is NaN, an ordered comparison will be false and the result will
3874 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3875 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3876 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3877 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3879 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3880 // will return -0, so vmin can only be used for unsafe math or if one of
3881 // the operands is known to be nonzero.
3882 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3884 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3886 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3895 // If LHS is NaN, an ordered comparison will be false and the result will
3896 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3897 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3898 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3899 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3901 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3902 // will return +0, so vmax can only be used for unsafe math or if one of
3903 // the operands is known to be nonzero.
3904 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3906 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3908 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3914 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3917 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3918 DAGCombinerInfo &DCI) const {
3919 switch (N->getOpcode()) {
3921 case ISD::ADD: return PerformADDCombine(N, DCI);
3922 case ISD::SUB: return PerformSUBCombine(N, DCI);
3923 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3924 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
3927 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
3928 case ISD::SIGN_EXTEND:
3929 case ISD::ZERO_EXTEND:
3930 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
3931 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
3936 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3937 if (!Subtarget->hasV6Ops())
3938 // Pre-v6 does not support unaligned mem access.
3941 // v6+ may or may not support unaligned mem access depending on the system
3943 // FIXME: This is pretty conservative. Should we provide cmdline option to
3944 // control the behaviour?
3945 if (!Subtarget->isTargetDarwin())
3949 switch (VT.getSimpleVT().SimpleTy) {
3956 // FIXME: VLD1 etc with standard alignment is legal.
3960 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3965 switch (VT.getSimpleVT().SimpleTy) {
3966 default: return false;
3981 if ((V & (Scale - 1)) != 0)
3984 return V == (V & ((1LL << 5) - 1));
3987 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3988 const ARMSubtarget *Subtarget) {
3995 switch (VT.getSimpleVT().SimpleTy) {
3996 default: return false;
4001 // + imm12 or - imm8
4003 return V == (V & ((1LL << 8) - 1));
4004 return V == (V & ((1LL << 12) - 1));
4007 // Same as ARM mode. FIXME: NEON?
4008 if (!Subtarget->hasVFP2())
4013 return V == (V & ((1LL << 8) - 1));
4017 /// isLegalAddressImmediate - Return true if the integer value can be used
4018 /// as the offset of the target addressing mode for load / store of the
4020 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4021 const ARMSubtarget *Subtarget) {
4028 if (Subtarget->isThumb1Only())
4029 return isLegalT1AddressImmediate(V, VT);
4030 else if (Subtarget->isThumb2())
4031 return isLegalT2AddressImmediate(V, VT, Subtarget);
4036 switch (VT.getSimpleVT().SimpleTy) {
4037 default: return false;
4042 return V == (V & ((1LL << 12) - 1));
4045 return V == (V & ((1LL << 8) - 1));
4048 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4053 return V == (V & ((1LL << 8) - 1));
4057 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4059 int Scale = AM.Scale;
4063 switch (VT.getSimpleVT().SimpleTy) {
4064 default: return false;
4073 return Scale == 2 || Scale == 4 || Scale == 8;
4076 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4080 // Note, we allow "void" uses (basically, uses that aren't loads or
4081 // stores), because arm allows folding a scale into many arithmetic
4082 // operations. This should be made more precise and revisited later.
4084 // Allow r << imm, but the imm has to be a multiple of two.
4085 if (Scale & 1) return false;
4086 return isPowerOf2_32(Scale);
4090 /// isLegalAddressingMode - Return true if the addressing mode represented
4091 /// by AM is legal for this target, for a load/store of the specified type.
4092 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4093 const Type *Ty) const {
4094 EVT VT = getValueType(Ty, true);
4095 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4098 // Can never fold addr of global into load/store.
4103 case 0: // no scale reg, must be "r+i" or "r", or "i".
4106 if (Subtarget->isThumb1Only())
4110 // ARM doesn't support any R+R*scale+imm addr modes.
4117 if (Subtarget->isThumb2())
4118 return isLegalT2ScaledAddressingMode(AM, VT);
4120 int Scale = AM.Scale;
4121 switch (VT.getSimpleVT().SimpleTy) {
4122 default: return false;
4126 if (Scale < 0) Scale = -Scale;
4130 return isPowerOf2_32(Scale & ~1);
4134 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4139 // Note, we allow "void" uses (basically, uses that aren't loads or
4140 // stores), because arm allows folding a scale into many arithmetic
4141 // operations. This should be made more precise and revisited later.
4143 // Allow r << imm, but the imm has to be a multiple of two.
4144 if (Scale & 1) return false;
4145 return isPowerOf2_32(Scale);
4152 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4153 /// icmp immediate, that is the target has icmp instructions which can compare
4154 /// a register against the immediate without having to materialize the
4155 /// immediate into a register.
4156 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4157 if (!Subtarget->isThumb())
4158 return ARM_AM::getSOImmVal(Imm) != -1;
4159 if (Subtarget->isThumb2())
4160 return ARM_AM::getT2SOImmVal(Imm) != -1;
4161 return Imm >= 0 && Imm <= 255;
4164 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4165 bool isSEXTLoad, SDValue &Base,
4166 SDValue &Offset, bool &isInc,
4167 SelectionDAG &DAG) {
4168 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4171 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4173 Base = Ptr->getOperand(0);
4174 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4175 int RHSC = (int)RHS->getZExtValue();
4176 if (RHSC < 0 && RHSC > -256) {
4177 assert(Ptr->getOpcode() == ISD::ADD);
4179 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4183 isInc = (Ptr->getOpcode() == ISD::ADD);
4184 Offset = Ptr->getOperand(1);
4186 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4188 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4189 int RHSC = (int)RHS->getZExtValue();
4190 if (RHSC < 0 && RHSC > -0x1000) {
4191 assert(Ptr->getOpcode() == ISD::ADD);
4193 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4194 Base = Ptr->getOperand(0);
4199 if (Ptr->getOpcode() == ISD::ADD) {
4201 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4202 if (ShOpcVal != ARM_AM::no_shift) {
4203 Base = Ptr->getOperand(1);
4204 Offset = Ptr->getOperand(0);
4206 Base = Ptr->getOperand(0);
4207 Offset = Ptr->getOperand(1);
4212 isInc = (Ptr->getOpcode() == ISD::ADD);
4213 Base = Ptr->getOperand(0);
4214 Offset = Ptr->getOperand(1);
4218 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4222 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4223 bool isSEXTLoad, SDValue &Base,
4224 SDValue &Offset, bool &isInc,
4225 SelectionDAG &DAG) {
4226 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4229 Base = Ptr->getOperand(0);
4230 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4231 int RHSC = (int)RHS->getZExtValue();
4232 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4233 assert(Ptr->getOpcode() == ISD::ADD);
4235 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4237 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4238 isInc = Ptr->getOpcode() == ISD::ADD;
4239 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4247 /// getPreIndexedAddressParts - returns true by value, base pointer and
4248 /// offset pointer and addressing mode by reference if the node's address
4249 /// can be legally represented as pre-indexed load / store address.
4251 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4253 ISD::MemIndexedMode &AM,
4254 SelectionDAG &DAG) const {
4255 if (Subtarget->isThumb1Only())
4260 bool isSEXTLoad = false;
4261 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4262 Ptr = LD->getBasePtr();
4263 VT = LD->getMemoryVT();
4264 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4265 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4266 Ptr = ST->getBasePtr();
4267 VT = ST->getMemoryVT();
4272 bool isLegal = false;
4273 if (Subtarget->isThumb2())
4274 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4275 Offset, isInc, DAG);
4277 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4278 Offset, isInc, DAG);
4282 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4286 /// getPostIndexedAddressParts - returns true by value, base pointer and
4287 /// offset pointer and addressing mode by reference if this node can be
4288 /// combined with a load / store to form a post-indexed load / store.
4289 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4292 ISD::MemIndexedMode &AM,
4293 SelectionDAG &DAG) const {
4294 if (Subtarget->isThumb1Only())
4299 bool isSEXTLoad = false;
4300 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4301 VT = LD->getMemoryVT();
4302 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4303 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4304 VT = ST->getMemoryVT();
4309 bool isLegal = false;
4310 if (Subtarget->isThumb2())
4311 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4314 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4319 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4323 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4327 const SelectionDAG &DAG,
4328 unsigned Depth) const {
4329 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4330 switch (Op.getOpcode()) {
4332 case ARMISD::CMOV: {
4333 // Bits are known zero/one if known on the LHS and RHS.
4334 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4335 if (KnownZero == 0 && KnownOne == 0) return;
4337 APInt KnownZeroRHS, KnownOneRHS;
4338 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4339 KnownZeroRHS, KnownOneRHS, Depth+1);
4340 KnownZero &= KnownZeroRHS;
4341 KnownOne &= KnownOneRHS;
4347 //===----------------------------------------------------------------------===//
4348 // ARM Inline Assembly Support
4349 //===----------------------------------------------------------------------===//
4351 /// getConstraintType - Given a constraint letter, return the type of
4352 /// constraint it is for this target.
4353 ARMTargetLowering::ConstraintType
4354 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4355 if (Constraint.size() == 1) {
4356 switch (Constraint[0]) {
4358 case 'l': return C_RegisterClass;
4359 case 'w': return C_RegisterClass;
4362 return TargetLowering::getConstraintType(Constraint);
4365 std::pair<unsigned, const TargetRegisterClass*>
4366 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4368 if (Constraint.size() == 1) {
4369 // GCC ARM Constraint Letters
4370 switch (Constraint[0]) {
4372 if (Subtarget->isThumb())
4373 return std::make_pair(0U, ARM::tGPRRegisterClass);
4375 return std::make_pair(0U, ARM::GPRRegisterClass);
4377 return std::make_pair(0U, ARM::GPRRegisterClass);
4380 return std::make_pair(0U, ARM::SPRRegisterClass);
4381 if (VT.getSizeInBits() == 64)
4382 return std::make_pair(0U, ARM::DPRRegisterClass);
4383 if (VT.getSizeInBits() == 128)
4384 return std::make_pair(0U, ARM::QPRRegisterClass);
4388 if (StringRef("{cc}").equals_lower(Constraint))
4389 return std::make_pair(0U, ARM::CCRRegisterClass);
4391 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4394 std::vector<unsigned> ARMTargetLowering::
4395 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4397 if (Constraint.size() != 1)
4398 return std::vector<unsigned>();
4400 switch (Constraint[0]) { // GCC ARM Constraint Letters
4403 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4404 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4407 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4408 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4409 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4410 ARM::R12, ARM::LR, 0);
4413 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4414 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4415 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4416 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4417 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4418 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4419 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4420 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4421 if (VT.getSizeInBits() == 64)
4422 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4423 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4424 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4425 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4426 if (VT.getSizeInBits() == 128)
4427 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4428 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4432 return std::vector<unsigned>();
4435 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4436 /// vector. If it is invalid, don't add anything to Ops.
4437 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4440 std::vector<SDValue>&Ops,
4441 SelectionDAG &DAG) const {
4442 SDValue Result(0, 0);
4444 switch (Constraint) {
4446 case 'I': case 'J': case 'K': case 'L':
4447 case 'M': case 'N': case 'O':
4448 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4452 int64_t CVal64 = C->getSExtValue();
4453 int CVal = (int) CVal64;
4454 // None of these constraints allow values larger than 32 bits. Check
4455 // that the value fits in an int.
4459 switch (Constraint) {
4461 if (Subtarget->isThumb1Only()) {
4462 // This must be a constant between 0 and 255, for ADD
4464 if (CVal >= 0 && CVal <= 255)
4466 } else if (Subtarget->isThumb2()) {
4467 // A constant that can be used as an immediate value in a
4468 // data-processing instruction.
4469 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4472 // A constant that can be used as an immediate value in a
4473 // data-processing instruction.
4474 if (ARM_AM::getSOImmVal(CVal) != -1)
4480 if (Subtarget->isThumb()) { // FIXME thumb2
4481 // This must be a constant between -255 and -1, for negated ADD
4482 // immediates. This can be used in GCC with an "n" modifier that
4483 // prints the negated value, for use with SUB instructions. It is
4484 // not useful otherwise but is implemented for compatibility.
4485 if (CVal >= -255 && CVal <= -1)
4488 // This must be a constant between -4095 and 4095. It is not clear
4489 // what this constraint is intended for. Implemented for
4490 // compatibility with GCC.
4491 if (CVal >= -4095 && CVal <= 4095)
4497 if (Subtarget->isThumb1Only()) {
4498 // A 32-bit value where only one byte has a nonzero value. Exclude
4499 // zero to match GCC. This constraint is used by GCC internally for
4500 // constants that can be loaded with a move/shift combination.
4501 // It is not useful otherwise but is implemented for compatibility.
4502 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4504 } else if (Subtarget->isThumb2()) {
4505 // A constant whose bitwise inverse can be used as an immediate
4506 // value in a data-processing instruction. This can be used in GCC
4507 // with a "B" modifier that prints the inverted value, for use with
4508 // BIC and MVN instructions. It is not useful otherwise but is
4509 // implemented for compatibility.
4510 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4513 // A constant whose bitwise inverse can be used as an immediate
4514 // value in a data-processing instruction. This can be used in GCC
4515 // with a "B" modifier that prints the inverted value, for use with
4516 // BIC and MVN instructions. It is not useful otherwise but is
4517 // implemented for compatibility.
4518 if (ARM_AM::getSOImmVal(~CVal) != -1)
4524 if (Subtarget->isThumb1Only()) {
4525 // This must be a constant between -7 and 7,
4526 // for 3-operand ADD/SUB immediate instructions.
4527 if (CVal >= -7 && CVal < 7)
4529 } else if (Subtarget->isThumb2()) {
4530 // A constant whose negation can be used as an immediate value in a
4531 // data-processing instruction. This can be used in GCC with an "n"
4532 // modifier that prints the negated value, for use with SUB
4533 // instructions. It is not useful otherwise but is implemented for
4535 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4538 // A constant whose negation can be used as an immediate value in a
4539 // data-processing instruction. This can be used in GCC with an "n"
4540 // modifier that prints the negated value, for use with SUB
4541 // instructions. It is not useful otherwise but is implemented for
4543 if (ARM_AM::getSOImmVal(-CVal) != -1)
4549 if (Subtarget->isThumb()) { // FIXME thumb2
4550 // This must be a multiple of 4 between 0 and 1020, for
4551 // ADD sp + immediate.
4552 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4555 // A power of two or a constant between 0 and 32. This is used in
4556 // GCC for the shift amount on shifted register operands, but it is
4557 // useful in general for any shift amounts.
4558 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4564 if (Subtarget->isThumb()) { // FIXME thumb2
4565 // This must be a constant between 0 and 31, for shift amounts.
4566 if (CVal >= 0 && CVal <= 31)
4572 if (Subtarget->isThumb()) { // FIXME thumb2
4573 // This must be a multiple of 4 between -508 and 508, for
4574 // ADD/SUB sp = sp + immediate.
4575 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4580 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4584 if (Result.getNode()) {
4585 Ops.push_back(Result);
4588 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4593 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4594 // The ARM target isn't yet aware of offsets.
4598 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4599 APInt Imm = FPImm.bitcastToAPInt();
4600 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4601 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4602 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4604 // We can handle 4 bits of mantissa.
4605 // mantissa = (16+UInt(e:f:g:h))/16.
4606 if (Mantissa & 0x7ffff)
4609 if ((Mantissa & 0xf) != Mantissa)
4612 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4613 if (Exp < -3 || Exp > 4)
4615 Exp = ((Exp+3) & 0x7) ^ 4;
4617 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4620 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4621 APInt Imm = FPImm.bitcastToAPInt();
4622 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4623 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4624 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4626 // We can handle 4 bits of mantissa.
4627 // mantissa = (16+UInt(e:f:g:h))/16.
4628 if (Mantissa & 0xffffffffffffLL)
4631 if ((Mantissa & 0xf) != Mantissa)
4634 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4635 if (Exp < -3 || Exp > 4)
4637 Exp = ((Exp+3) & 0x7) ^ 4;
4639 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4642 /// isFPImmLegal - Returns true if the target can instruction select the
4643 /// specified FP immediate natively. If false, the legalizer will
4644 /// materialize the FP immediate as a load from a constant pool.
4645 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4646 if (!Subtarget->hasVFP3())
4649 return ARM::getVFPf32Imm(Imm) != -1;
4651 return ARM::getVFPf64Imm(Imm) != -1;