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);
93 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
94 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
96 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
97 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
98 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
101 // Promote all bit-wise operations.
102 if (VT.isInteger() && VT != PromotedBitwiseVT) {
103 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
104 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
105 PromotedBitwiseVT.getSimpleVT());
106 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
107 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
108 PromotedBitwiseVT.getSimpleVT());
109 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
110 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
111 PromotedBitwiseVT.getSimpleVT());
114 // Neon does not support vector divide/remainder operations.
115 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
119 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
123 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
124 addRegisterClass(VT, ARM::DPRRegisterClass);
125 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
128 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
129 addRegisterClass(VT, ARM::QPRRegisterClass);
130 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
133 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
134 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
135 return new TargetLoweringObjectFileMachO();
137 return new ARMElfTargetObjectFile();
140 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
141 : TargetLowering(TM, createTLOF(TM)) {
142 Subtarget = &TM.getSubtarget<ARMSubtarget>();
144 if (Subtarget->isTargetDarwin()) {
145 // Uses VFP for Thumb libfuncs if available.
146 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
147 // Single-precision floating-point arithmetic.
148 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
149 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
150 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
151 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
153 // Double-precision floating-point arithmetic.
154 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
155 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
156 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
157 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
159 // Single-precision comparisons.
160 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
161 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
162 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
163 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
164 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
165 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
166 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
167 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
169 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
172 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
173 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
174 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
175 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
176 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
178 // Double-precision comparisons.
179 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
180 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
181 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
182 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
183 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
184 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
185 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
186 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
188 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
197 // Floating-point to integer conversions.
198 // i64 conversions are done via library routines even when generating VFP
199 // instructions, so use the same ones.
200 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
201 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
202 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
203 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
205 // Conversions between floating types.
206 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
207 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
209 // Integer to floating-point conversions.
210 // i64 conversions are done via library routines even when generating VFP
211 // instructions, so use the same ones.
212 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
213 // e.g., __floatunsidf vs. __floatunssidfvfp.
214 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
215 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
216 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
217 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
221 // These libcalls are not available in 32-bit.
222 setLibcallName(RTLIB::SHL_I128, 0);
223 setLibcallName(RTLIB::SRL_I128, 0);
224 setLibcallName(RTLIB::SRA_I128, 0);
226 // Libcalls should use the AAPCS base standard ABI, even if hard float
227 // is in effect, as per the ARM RTABI specification, section 4.1.2.
228 if (Subtarget->isAAPCS_ABI()) {
229 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
230 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
231 CallingConv::ARM_AAPCS);
235 if (Subtarget->isThumb1Only())
236 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
238 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
239 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
240 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
241 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
243 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
246 if (Subtarget->hasNEON()) {
247 addDRTypeForNEON(MVT::v2f32);
248 addDRTypeForNEON(MVT::v8i8);
249 addDRTypeForNEON(MVT::v4i16);
250 addDRTypeForNEON(MVT::v2i32);
251 addDRTypeForNEON(MVT::v1i64);
253 addQRTypeForNEON(MVT::v4f32);
254 addQRTypeForNEON(MVT::v2f64);
255 addQRTypeForNEON(MVT::v16i8);
256 addQRTypeForNEON(MVT::v8i16);
257 addQRTypeForNEON(MVT::v4i32);
258 addQRTypeForNEON(MVT::v2i64);
260 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
261 // neither Neon nor VFP support any arithmetic operations on it.
262 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
263 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
264 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
265 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
266 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
267 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
268 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
269 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
270 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
271 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
272 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
273 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
274 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
275 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
276 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
277 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
278 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
279 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
280 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
281 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
282 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
283 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
284 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
285 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
287 // Neon does not support some operations on v1i64 and v2i64 types.
288 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
289 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
290 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
291 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
293 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
294 setTargetDAGCombine(ISD::SHL);
295 setTargetDAGCombine(ISD::SRL);
296 setTargetDAGCombine(ISD::SRA);
297 setTargetDAGCombine(ISD::SIGN_EXTEND);
298 setTargetDAGCombine(ISD::ZERO_EXTEND);
299 setTargetDAGCombine(ISD::ANY_EXTEND);
300 setTargetDAGCombine(ISD::SELECT_CC);
303 computeRegisterProperties();
305 // ARM does not have f32 extending load.
306 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
308 // ARM does not have i1 sign extending load.
309 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
311 // ARM supports all 4 flavors of integer indexed load / store.
312 if (!Subtarget->isThumb1Only()) {
313 for (unsigned im = (unsigned)ISD::PRE_INC;
314 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
315 setIndexedLoadAction(im, MVT::i1, Legal);
316 setIndexedLoadAction(im, MVT::i8, Legal);
317 setIndexedLoadAction(im, MVT::i16, Legal);
318 setIndexedLoadAction(im, MVT::i32, Legal);
319 setIndexedStoreAction(im, MVT::i1, Legal);
320 setIndexedStoreAction(im, MVT::i8, Legal);
321 setIndexedStoreAction(im, MVT::i16, Legal);
322 setIndexedStoreAction(im, MVT::i32, Legal);
326 // i64 operation support.
327 if (Subtarget->isThumb1Only()) {
328 setOperationAction(ISD::MUL, MVT::i64, Expand);
329 setOperationAction(ISD::MULHU, MVT::i32, Expand);
330 setOperationAction(ISD::MULHS, MVT::i32, Expand);
331 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
332 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
334 setOperationAction(ISD::MUL, MVT::i64, Expand);
335 setOperationAction(ISD::MULHU, MVT::i32, Expand);
336 if (!Subtarget->hasV6Ops())
337 setOperationAction(ISD::MULHS, MVT::i32, Expand);
339 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
340 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
341 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
342 setOperationAction(ISD::SRL, MVT::i64, Custom);
343 setOperationAction(ISD::SRA, MVT::i64, Custom);
345 // ARM does not have ROTL.
346 setOperationAction(ISD::ROTL, MVT::i32, Expand);
347 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
348 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
349 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
350 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
352 // Only ARMv6 has BSWAP.
353 if (!Subtarget->hasV6Ops())
354 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
356 // These are expanded into libcalls.
357 setOperationAction(ISD::SDIV, MVT::i32, Expand);
358 setOperationAction(ISD::UDIV, MVT::i32, Expand);
359 setOperationAction(ISD::SREM, MVT::i32, Expand);
360 setOperationAction(ISD::UREM, MVT::i32, Expand);
361 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
362 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
364 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
365 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
366 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
367 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
368 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
370 // Use the default implementation.
371 setOperationAction(ISD::VASTART, MVT::Other, Custom);
372 setOperationAction(ISD::VAARG, MVT::Other, Expand);
373 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
374 setOperationAction(ISD::VAEND, MVT::Other, Expand);
375 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
376 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
377 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
378 // FIXME: Shouldn't need this, since no register is used, but the legalizer
379 // doesn't yet know how to not do that for SjLj.
380 setExceptionSelectorRegister(ARM::R0);
381 if (Subtarget->isThumb())
382 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
384 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
385 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
387 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
388 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
389 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
391 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
393 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
394 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
395 // iff target supports vfp2.
396 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
398 // We want to custom lower some of our intrinsics.
399 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
401 setOperationAction(ISD::SETCC, MVT::i32, Expand);
402 setOperationAction(ISD::SETCC, MVT::f32, Expand);
403 setOperationAction(ISD::SETCC, MVT::f64, Expand);
404 setOperationAction(ISD::SELECT, MVT::i32, Expand);
405 setOperationAction(ISD::SELECT, MVT::f32, Expand);
406 setOperationAction(ISD::SELECT, MVT::f64, Expand);
407 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
408 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
409 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
411 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
412 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
413 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
414 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
415 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
417 // We don't support sin/cos/fmod/copysign/pow
418 setOperationAction(ISD::FSIN, MVT::f64, Expand);
419 setOperationAction(ISD::FSIN, MVT::f32, Expand);
420 setOperationAction(ISD::FCOS, MVT::f32, Expand);
421 setOperationAction(ISD::FCOS, MVT::f64, Expand);
422 setOperationAction(ISD::FREM, MVT::f64, Expand);
423 setOperationAction(ISD::FREM, MVT::f32, Expand);
424 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
425 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
426 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
428 setOperationAction(ISD::FPOW, MVT::f64, Expand);
429 setOperationAction(ISD::FPOW, MVT::f32, Expand);
431 // Various VFP goodness
432 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
433 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
434 if (Subtarget->hasVFP2()) {
435 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
436 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
437 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
438 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
440 // Special handling for half-precision FP.
441 if (!Subtarget->hasFP16()) {
442 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
443 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
447 // We have target-specific dag combine patterns for the following nodes:
448 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
449 setTargetDAGCombine(ISD::ADD);
450 setTargetDAGCombine(ISD::SUB);
452 setStackPointerRegisterToSaveRestore(ARM::SP);
453 setSchedulingPreference(SchedulingForRegPressure);
455 // FIXME: If-converter should use instruction latency to determine
456 // profitability rather than relying on fixed limits.
457 if (Subtarget->getCPUString() == "generic") {
458 // Generic (and overly aggressive) if-conversion limits.
459 setIfCvtBlockSizeLimit(10);
460 setIfCvtDupBlockSizeLimit(2);
461 } else if (Subtarget->hasV7Ops()) {
462 setIfCvtBlockSizeLimit(3);
463 setIfCvtDupBlockSizeLimit(1);
464 } else if (Subtarget->hasV6Ops()) {
465 setIfCvtBlockSizeLimit(2);
466 setIfCvtDupBlockSizeLimit(1);
468 setIfCvtBlockSizeLimit(3);
469 setIfCvtDupBlockSizeLimit(2);
472 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
473 // Do not enable CodePlacementOpt for now: it currently runs after the
474 // ARMConstantIslandPass and messes up branch relaxation and placement
475 // of constant islands.
476 // benefitFromCodePlacementOpt = true;
479 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
482 case ARMISD::Wrapper: return "ARMISD::Wrapper";
483 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
484 case ARMISD::CALL: return "ARMISD::CALL";
485 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
486 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
487 case ARMISD::tCALL: return "ARMISD::tCALL";
488 case ARMISD::BRCOND: return "ARMISD::BRCOND";
489 case ARMISD::BR_JT: return "ARMISD::BR_JT";
490 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
491 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
492 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
493 case ARMISD::CMP: return "ARMISD::CMP";
494 case ARMISD::CMPZ: return "ARMISD::CMPZ";
495 case ARMISD::CMPFP: return "ARMISD::CMPFP";
496 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
497 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
498 case ARMISD::CMOV: return "ARMISD::CMOV";
499 case ARMISD::CNEG: return "ARMISD::CNEG";
501 case ARMISD::RBIT: return "ARMISD::RBIT";
503 case ARMISD::FTOSI: return "ARMISD::FTOSI";
504 case ARMISD::FTOUI: return "ARMISD::FTOUI";
505 case ARMISD::SITOF: return "ARMISD::SITOF";
506 case ARMISD::UITOF: return "ARMISD::UITOF";
508 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
509 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
510 case ARMISD::RRX: return "ARMISD::RRX";
512 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
513 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
515 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
516 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
518 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
520 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
522 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
523 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
525 case ARMISD::VCEQ: return "ARMISD::VCEQ";
526 case ARMISD::VCGE: return "ARMISD::VCGE";
527 case ARMISD::VCGEU: return "ARMISD::VCGEU";
528 case ARMISD::VCGT: return "ARMISD::VCGT";
529 case ARMISD::VCGTU: return "ARMISD::VCGTU";
530 case ARMISD::VTST: return "ARMISD::VTST";
532 case ARMISD::VSHL: return "ARMISD::VSHL";
533 case ARMISD::VSHRs: return "ARMISD::VSHRs";
534 case ARMISD::VSHRu: return "ARMISD::VSHRu";
535 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
536 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
537 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
538 case ARMISD::VSHRN: return "ARMISD::VSHRN";
539 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
540 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
541 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
542 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
543 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
544 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
545 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
546 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
547 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
548 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
549 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
550 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
551 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
552 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
553 case ARMISD::VDUP: return "ARMISD::VDUP";
554 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
555 case ARMISD::VEXT: return "ARMISD::VEXT";
556 case ARMISD::VREV64: return "ARMISD::VREV64";
557 case ARMISD::VREV32: return "ARMISD::VREV32";
558 case ARMISD::VREV16: return "ARMISD::VREV16";
559 case ARMISD::VZIP: return "ARMISD::VZIP";
560 case ARMISD::VUZP: return "ARMISD::VUZP";
561 case ARMISD::VTRN: return "ARMISD::VTRN";
562 case ARMISD::FMAX: return "ARMISD::FMAX";
563 case ARMISD::FMIN: return "ARMISD::FMIN";
567 /// getFunctionAlignment - Return the Log2 alignment of this function.
568 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
569 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
572 //===----------------------------------------------------------------------===//
574 //===----------------------------------------------------------------------===//
576 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
577 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
579 default: llvm_unreachable("Unknown condition code!");
580 case ISD::SETNE: return ARMCC::NE;
581 case ISD::SETEQ: return ARMCC::EQ;
582 case ISD::SETGT: return ARMCC::GT;
583 case ISD::SETGE: return ARMCC::GE;
584 case ISD::SETLT: return ARMCC::LT;
585 case ISD::SETLE: return ARMCC::LE;
586 case ISD::SETUGT: return ARMCC::HI;
587 case ISD::SETUGE: return ARMCC::HS;
588 case ISD::SETULT: return ARMCC::LO;
589 case ISD::SETULE: return ARMCC::LS;
593 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
594 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
595 ARMCC::CondCodes &CondCode2) {
596 CondCode2 = ARMCC::AL;
598 default: llvm_unreachable("Unknown FP condition!");
600 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
602 case ISD::SETOGT: CondCode = ARMCC::GT; break;
604 case ISD::SETOGE: CondCode = ARMCC::GE; break;
605 case ISD::SETOLT: CondCode = ARMCC::MI; break;
606 case ISD::SETOLE: CondCode = ARMCC::LS; break;
607 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
608 case ISD::SETO: CondCode = ARMCC::VC; break;
609 case ISD::SETUO: CondCode = ARMCC::VS; break;
610 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
611 case ISD::SETUGT: CondCode = ARMCC::HI; break;
612 case ISD::SETUGE: CondCode = ARMCC::PL; break;
614 case ISD::SETULT: CondCode = ARMCC::LT; break;
616 case ISD::SETULE: CondCode = ARMCC::LE; break;
618 case ISD::SETUNE: CondCode = ARMCC::NE; break;
622 //===----------------------------------------------------------------------===//
623 // Calling Convention Implementation
624 //===----------------------------------------------------------------------===//
626 #include "ARMGenCallingConv.inc"
628 // APCS f64 is in register pairs, possibly split to stack
629 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
630 CCValAssign::LocInfo &LocInfo,
631 CCState &State, bool CanFail) {
632 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
634 // Try to get the first register.
635 if (unsigned Reg = State.AllocateReg(RegList, 4))
636 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
638 // For the 2nd half of a v2f64, do not fail.
642 // Put the whole thing on the stack.
643 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
644 State.AllocateStack(8, 4),
649 // Try to get the second register.
650 if (unsigned Reg = State.AllocateReg(RegList, 4))
651 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
653 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
654 State.AllocateStack(4, 4),
659 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
660 CCValAssign::LocInfo &LocInfo,
661 ISD::ArgFlagsTy &ArgFlags,
663 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
665 if (LocVT == MVT::v2f64 &&
666 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
668 return true; // we handled it
671 // AAPCS f64 is in aligned register pairs
672 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
673 CCValAssign::LocInfo &LocInfo,
674 CCState &State, bool CanFail) {
675 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
676 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
678 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
680 // For the 2nd half of a v2f64, do not just fail.
684 // Put the whole thing on the stack.
685 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
686 State.AllocateStack(8, 8),
692 for (i = 0; i < 2; ++i)
693 if (HiRegList[i] == Reg)
696 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
697 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
702 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
703 CCValAssign::LocInfo &LocInfo,
704 ISD::ArgFlagsTy &ArgFlags,
706 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
708 if (LocVT == MVT::v2f64 &&
709 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
711 return true; // we handled it
714 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
715 CCValAssign::LocInfo &LocInfo, CCState &State) {
716 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
717 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
719 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
721 return false; // we didn't handle it
724 for (i = 0; i < 2; ++i)
725 if (HiRegList[i] == Reg)
728 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
729 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
734 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
735 CCValAssign::LocInfo &LocInfo,
736 ISD::ArgFlagsTy &ArgFlags,
738 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
740 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
742 return true; // we handled it
745 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
746 CCValAssign::LocInfo &LocInfo,
747 ISD::ArgFlagsTy &ArgFlags,
749 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
753 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
754 /// given CallingConvention value.
755 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
757 bool isVarArg) const {
760 llvm_unreachable("Unsupported calling convention");
762 case CallingConv::Fast:
763 // Use target triple & subtarget features to do actual dispatch.
764 if (Subtarget->isAAPCS_ABI()) {
765 if (Subtarget->hasVFP2() &&
766 FloatABIType == FloatABI::Hard && !isVarArg)
767 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
769 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
771 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
772 case CallingConv::ARM_AAPCS_VFP:
773 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
774 case CallingConv::ARM_AAPCS:
775 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
776 case CallingConv::ARM_APCS:
777 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
781 /// LowerCallResult - Lower the result values of a call into the
782 /// appropriate copies out of appropriate physical registers.
784 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
785 CallingConv::ID CallConv, bool isVarArg,
786 const SmallVectorImpl<ISD::InputArg> &Ins,
787 DebugLoc dl, SelectionDAG &DAG,
788 SmallVectorImpl<SDValue> &InVals) {
790 // Assign locations to each value returned by this call.
791 SmallVector<CCValAssign, 16> RVLocs;
792 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
793 RVLocs, *DAG.getContext());
794 CCInfo.AnalyzeCallResult(Ins,
795 CCAssignFnForNode(CallConv, /* Return*/ true,
798 // Copy all of the result registers out of their specified physreg.
799 for (unsigned i = 0; i != RVLocs.size(); ++i) {
800 CCValAssign VA = RVLocs[i];
803 if (VA.needsCustom()) {
804 // Handle f64 or half of a v2f64.
805 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
807 Chain = Lo.getValue(1);
808 InFlag = Lo.getValue(2);
809 VA = RVLocs[++i]; // skip ahead to next loc
810 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
812 Chain = Hi.getValue(1);
813 InFlag = Hi.getValue(2);
814 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
816 if (VA.getLocVT() == MVT::v2f64) {
817 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
818 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
819 DAG.getConstant(0, MVT::i32));
821 VA = RVLocs[++i]; // skip ahead to next loc
822 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
823 Chain = Lo.getValue(1);
824 InFlag = Lo.getValue(2);
825 VA = RVLocs[++i]; // skip ahead to next loc
826 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
827 Chain = Hi.getValue(1);
828 InFlag = Hi.getValue(2);
829 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
830 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
831 DAG.getConstant(1, MVT::i32));
834 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
836 Chain = Val.getValue(1);
837 InFlag = Val.getValue(2);
840 switch (VA.getLocInfo()) {
841 default: llvm_unreachable("Unknown loc info!");
842 case CCValAssign::Full: break;
843 case CCValAssign::BCvt:
844 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
848 InVals.push_back(Val);
854 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
855 /// by "Src" to address "Dst" of size "Size". Alignment information is
856 /// specified by the specific parameter attribute. The copy will be passed as
857 /// a byval function parameter.
858 /// Sometimes what we are copying is the end of a larger object, the part that
859 /// does not fit in registers.
861 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
862 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
864 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
865 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
866 /*isVolatile=*/false, /*AlwaysInline=*/false,
870 /// LowerMemOpCallTo - Store the argument to the stack.
872 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
873 SDValue StackPtr, SDValue Arg,
874 DebugLoc dl, SelectionDAG &DAG,
875 const CCValAssign &VA,
876 ISD::ArgFlagsTy Flags) {
877 unsigned LocMemOffset = VA.getLocMemOffset();
878 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
879 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
880 if (Flags.isByVal()) {
881 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
883 return DAG.getStore(Chain, dl, Arg, PtrOff,
884 PseudoSourceValue::getStack(), LocMemOffset,
888 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
889 SDValue Chain, SDValue &Arg,
890 RegsToPassVector &RegsToPass,
891 CCValAssign &VA, CCValAssign &NextVA,
893 SmallVector<SDValue, 8> &MemOpChains,
894 ISD::ArgFlagsTy Flags) {
896 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
897 DAG.getVTList(MVT::i32, MVT::i32), Arg);
898 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
900 if (NextVA.isRegLoc())
901 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
903 assert(NextVA.isMemLoc());
904 if (StackPtr.getNode() == 0)
905 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
907 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
913 /// LowerCall - Lowering a call into a callseq_start <-
914 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
917 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
918 CallingConv::ID CallConv, bool isVarArg,
920 const SmallVectorImpl<ISD::OutputArg> &Outs,
921 const SmallVectorImpl<ISD::InputArg> &Ins,
922 DebugLoc dl, SelectionDAG &DAG,
923 SmallVectorImpl<SDValue> &InVals) {
924 // ARM target does not yet support tail call optimization.
927 // Analyze operands of the call, assigning locations to each operand.
928 SmallVector<CCValAssign, 16> ArgLocs;
929 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
931 CCInfo.AnalyzeCallOperands(Outs,
932 CCAssignFnForNode(CallConv, /* Return*/ false,
935 // Get a count of how many bytes are to be pushed on the stack.
936 unsigned NumBytes = CCInfo.getNextStackOffset();
938 // Adjust the stack pointer for the new arguments...
939 // These operations are automatically eliminated by the prolog/epilog pass
940 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
942 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
944 RegsToPassVector RegsToPass;
945 SmallVector<SDValue, 8> MemOpChains;
947 // Walk the register/memloc assignments, inserting copies/loads. In the case
948 // of tail call optimization, arguments are handled later.
949 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
952 CCValAssign &VA = ArgLocs[i];
953 SDValue Arg = Outs[realArgIdx].Val;
954 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
956 // Promote the value if needed.
957 switch (VA.getLocInfo()) {
958 default: llvm_unreachable("Unknown loc info!");
959 case CCValAssign::Full: break;
960 case CCValAssign::SExt:
961 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
963 case CCValAssign::ZExt:
964 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
966 case CCValAssign::AExt:
967 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
969 case CCValAssign::BCvt:
970 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
974 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
975 if (VA.needsCustom()) {
976 if (VA.getLocVT() == MVT::v2f64) {
977 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
978 DAG.getConstant(0, MVT::i32));
979 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
980 DAG.getConstant(1, MVT::i32));
982 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
983 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
985 VA = ArgLocs[++i]; // skip ahead to next loc
987 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
988 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
990 assert(VA.isMemLoc());
992 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
993 dl, DAG, VA, Flags));
996 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
997 StackPtr, MemOpChains, Flags);
999 } else if (VA.isRegLoc()) {
1000 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1002 assert(VA.isMemLoc());
1004 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1005 dl, DAG, VA, Flags));
1009 if (!MemOpChains.empty())
1010 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1011 &MemOpChains[0], MemOpChains.size());
1013 // Build a sequence of copy-to-reg nodes chained together with token chain
1014 // and flag operands which copy the outgoing args into the appropriate regs.
1016 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1017 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1018 RegsToPass[i].second, InFlag);
1019 InFlag = Chain.getValue(1);
1022 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1023 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1024 // node so that legalize doesn't hack it.
1025 bool isDirect = false;
1026 bool isARMFunc = false;
1027 bool isLocalARMFunc = false;
1028 MachineFunction &MF = DAG.getMachineFunction();
1029 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1030 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1031 GlobalValue *GV = G->getGlobal();
1033 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1034 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1035 getTargetMachine().getRelocationModel() != Reloc::Static;
1036 isARMFunc = !Subtarget->isThumb() || isStub;
1037 // ARM call to a local ARM function is predicable.
1038 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1039 // tBX takes a register source operand.
1040 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1041 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1042 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1045 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1046 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1047 Callee = DAG.getLoad(getPointerTy(), dl,
1048 DAG.getEntryNode(), CPAddr,
1049 PseudoSourceValue::getConstantPool(), 0,
1051 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1052 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1053 getPointerTy(), Callee, PICLabel);
1055 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1056 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1058 bool isStub = Subtarget->isTargetDarwin() &&
1059 getTargetMachine().getRelocationModel() != Reloc::Static;
1060 isARMFunc = !Subtarget->isThumb() || isStub;
1061 // tBX takes a register source operand.
1062 const char *Sym = S->getSymbol();
1063 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1064 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1065 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1066 Sym, ARMPCLabelIndex, 4);
1067 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1068 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1069 Callee = DAG.getLoad(getPointerTy(), dl,
1070 DAG.getEntryNode(), CPAddr,
1071 PseudoSourceValue::getConstantPool(), 0,
1073 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1074 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1075 getPointerTy(), Callee, PICLabel);
1077 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1080 // FIXME: handle tail calls differently.
1082 if (Subtarget->isThumb()) {
1083 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1084 CallOpc = ARMISD::CALL_NOLINK;
1086 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1088 CallOpc = (isDirect || Subtarget->hasV5TOps())
1089 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1090 : ARMISD::CALL_NOLINK;
1092 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1093 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1094 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1095 InFlag = Chain.getValue(1);
1098 std::vector<SDValue> Ops;
1099 Ops.push_back(Chain);
1100 Ops.push_back(Callee);
1102 // Add argument registers to the end of the list so that they are known live
1104 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1105 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1106 RegsToPass[i].second.getValueType()));
1108 if (InFlag.getNode())
1109 Ops.push_back(InFlag);
1110 // Returns a chain and a flag for retval copy to use.
1111 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1112 &Ops[0], Ops.size());
1113 InFlag = Chain.getValue(1);
1115 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1116 DAG.getIntPtrConstant(0, true), InFlag);
1118 InFlag = Chain.getValue(1);
1120 // Handle result values, copying them out of physregs into vregs that we
1122 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1127 ARMTargetLowering::LowerReturn(SDValue Chain,
1128 CallingConv::ID CallConv, bool isVarArg,
1129 const SmallVectorImpl<ISD::OutputArg> &Outs,
1130 DebugLoc dl, SelectionDAG &DAG) {
1132 // CCValAssign - represent the assignment of the return value to a location.
1133 SmallVector<CCValAssign, 16> RVLocs;
1135 // CCState - Info about the registers and stack slots.
1136 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1139 // Analyze outgoing return values.
1140 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1143 // If this is the first return lowered for this function, add
1144 // the regs to the liveout set for the function.
1145 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1146 for (unsigned i = 0; i != RVLocs.size(); ++i)
1147 if (RVLocs[i].isRegLoc())
1148 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1153 // Copy the result values into the output registers.
1154 for (unsigned i = 0, realRVLocIdx = 0;
1156 ++i, ++realRVLocIdx) {
1157 CCValAssign &VA = RVLocs[i];
1158 assert(VA.isRegLoc() && "Can only return in registers!");
1160 SDValue Arg = Outs[realRVLocIdx].Val;
1162 switch (VA.getLocInfo()) {
1163 default: llvm_unreachable("Unknown loc info!");
1164 case CCValAssign::Full: break;
1165 case CCValAssign::BCvt:
1166 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1170 if (VA.needsCustom()) {
1171 if (VA.getLocVT() == MVT::v2f64) {
1172 // Extract the first half and return it in two registers.
1173 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1174 DAG.getConstant(0, MVT::i32));
1175 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1176 DAG.getVTList(MVT::i32, MVT::i32), Half);
1178 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1179 Flag = Chain.getValue(1);
1180 VA = RVLocs[++i]; // skip ahead to next loc
1181 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1182 HalfGPRs.getValue(1), Flag);
1183 Flag = Chain.getValue(1);
1184 VA = RVLocs[++i]; // skip ahead to next loc
1186 // Extract the 2nd half and fall through to handle it as an f64 value.
1187 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1188 DAG.getConstant(1, MVT::i32));
1190 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1192 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1193 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1194 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1195 Flag = Chain.getValue(1);
1196 VA = RVLocs[++i]; // skip ahead to next loc
1197 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1200 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1202 // Guarantee that all emitted copies are
1203 // stuck together, avoiding something bad.
1204 Flag = Chain.getValue(1);
1209 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1211 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1216 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1217 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1218 // one of the above mentioned nodes. It has to be wrapped because otherwise
1219 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1220 // be used to form addressing mode. These wrapped nodes will be selected
1222 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1223 EVT PtrVT = Op.getValueType();
1224 // FIXME there is no actual debug info here
1225 DebugLoc dl = Op.getDebugLoc();
1226 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1228 if (CP->isMachineConstantPoolEntry())
1229 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1230 CP->getAlignment());
1232 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1233 CP->getAlignment());
1234 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1237 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1238 MachineFunction &MF = DAG.getMachineFunction();
1239 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1240 unsigned ARMPCLabelIndex = 0;
1241 DebugLoc DL = Op.getDebugLoc();
1242 EVT PtrVT = getPointerTy();
1243 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1244 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1246 if (RelocM == Reloc::Static) {
1247 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1249 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1250 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1251 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1252 ARMCP::CPBlockAddress,
1254 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1256 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1257 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1258 PseudoSourceValue::getConstantPool(), 0,
1260 if (RelocM == Reloc::Static)
1262 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1263 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1266 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1268 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1269 SelectionDAG &DAG) {
1270 DebugLoc dl = GA->getDebugLoc();
1271 EVT PtrVT = getPointerTy();
1272 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1273 MachineFunction &MF = DAG.getMachineFunction();
1274 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1275 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1276 ARMConstantPoolValue *CPV =
1277 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1278 ARMCP::CPValue, PCAdj, "tlsgd", true);
1279 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1280 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1281 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1282 PseudoSourceValue::getConstantPool(), 0,
1284 SDValue Chain = Argument.getValue(1);
1286 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1287 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1289 // call __tls_get_addr.
1292 Entry.Node = Argument;
1293 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1294 Args.push_back(Entry);
1295 // FIXME: is there useful debug info available here?
1296 std::pair<SDValue, SDValue> CallResult =
1297 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1298 false, false, false, false,
1299 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1300 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1301 return CallResult.first;
1304 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1305 // "local exec" model.
1307 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1308 SelectionDAG &DAG) {
1309 GlobalValue *GV = GA->getGlobal();
1310 DebugLoc dl = GA->getDebugLoc();
1312 SDValue Chain = DAG.getEntryNode();
1313 EVT PtrVT = getPointerTy();
1314 // Get the Thread Pointer
1315 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1317 if (GV->isDeclaration()) {
1318 MachineFunction &MF = DAG.getMachineFunction();
1319 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1320 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1321 // Initial exec model.
1322 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1323 ARMConstantPoolValue *CPV =
1324 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1325 ARMCP::CPValue, PCAdj, "gottpoff", true);
1326 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1327 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1328 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1329 PseudoSourceValue::getConstantPool(), 0,
1331 Chain = Offset.getValue(1);
1333 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1334 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1336 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1337 PseudoSourceValue::getConstantPool(), 0,
1341 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1342 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1343 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1344 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1345 PseudoSourceValue::getConstantPool(), 0,
1349 // The address of the thread local variable is the add of the thread
1350 // pointer with the offset of the variable.
1351 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1355 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1356 // TODO: implement the "local dynamic" model
1357 assert(Subtarget->isTargetELF() &&
1358 "TLS not implemented for non-ELF targets");
1359 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1360 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1361 // otherwise use the "Local Exec" TLS Model
1362 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1363 return LowerToTLSGeneralDynamicModel(GA, DAG);
1365 return LowerToTLSExecModels(GA, DAG);
1368 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1369 SelectionDAG &DAG) {
1370 EVT PtrVT = getPointerTy();
1371 DebugLoc dl = Op.getDebugLoc();
1372 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1373 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1374 if (RelocM == Reloc::PIC_) {
1375 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1376 ARMConstantPoolValue *CPV =
1377 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1378 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1379 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1380 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1382 PseudoSourceValue::getConstantPool(), 0,
1384 SDValue Chain = Result.getValue(1);
1385 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1386 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1388 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1389 PseudoSourceValue::getGOT(), 0,
1393 // If we have T2 ops, we can materialize the address directly via movt/movw
1394 // pair. This is always cheaper.
1395 if (Subtarget->useMovt()) {
1396 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1397 DAG.getTargetGlobalAddress(GV, PtrVT));
1399 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1400 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1401 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1402 PseudoSourceValue::getConstantPool(), 0,
1408 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1409 SelectionDAG &DAG) {
1410 MachineFunction &MF = DAG.getMachineFunction();
1411 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1412 unsigned ARMPCLabelIndex = 0;
1413 EVT PtrVT = getPointerTy();
1414 DebugLoc dl = Op.getDebugLoc();
1415 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1416 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1418 if (RelocM == Reloc::Static)
1419 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1421 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1422 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1423 ARMConstantPoolValue *CPV =
1424 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1425 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1427 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1429 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1430 PseudoSourceValue::getConstantPool(), 0,
1432 SDValue Chain = Result.getValue(1);
1434 if (RelocM == Reloc::PIC_) {
1435 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1436 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1439 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1440 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1441 PseudoSourceValue::getGOT(), 0,
1447 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1449 assert(Subtarget->isTargetELF() &&
1450 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1451 MachineFunction &MF = DAG.getMachineFunction();
1452 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1453 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1454 EVT PtrVT = getPointerTy();
1455 DebugLoc dl = Op.getDebugLoc();
1456 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1457 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1458 "_GLOBAL_OFFSET_TABLE_",
1459 ARMPCLabelIndex, PCAdj);
1460 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1461 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1462 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1463 PseudoSourceValue::getConstantPool(), 0,
1465 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1466 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1470 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1471 const ARMSubtarget *Subtarget) {
1472 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1473 DebugLoc dl = Op.getDebugLoc();
1475 default: return SDValue(); // Don't custom lower most intrinsics.
1476 case Intrinsic::arm_thread_pointer: {
1477 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1478 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1480 case Intrinsic::eh_sjlj_lsda: {
1481 MachineFunction &MF = DAG.getMachineFunction();
1482 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1483 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1484 EVT PtrVT = getPointerTy();
1485 DebugLoc dl = Op.getDebugLoc();
1486 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1488 unsigned PCAdj = (RelocM != Reloc::PIC_)
1489 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1490 ARMConstantPoolValue *CPV =
1491 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1492 ARMCP::CPLSDA, PCAdj);
1493 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1494 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1496 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1497 PseudoSourceValue::getConstantPool(), 0,
1499 SDValue Chain = Result.getValue(1);
1501 if (RelocM == Reloc::PIC_) {
1502 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1503 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1507 case Intrinsic::eh_sjlj_setjmp:
1508 SDValue Val = Subtarget->isThumb() ?
1509 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1510 DAG.getConstant(0, MVT::i32);
1511 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1516 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1517 const ARMSubtarget *Subtarget) {
1518 DebugLoc dl = Op.getDebugLoc();
1519 SDValue Op5 = Op.getOperand(5);
1521 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1522 if (isDeviceBarrier) {
1523 if (Subtarget->hasV7Ops())
1524 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1526 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1527 DAG.getConstant(0, MVT::i32));
1529 if (Subtarget->hasV7Ops())
1530 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1532 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1533 DAG.getConstant(0, MVT::i32));
1538 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1539 unsigned VarArgsFrameIndex) {
1540 // vastart just stores the address of the VarArgsFrameIndex slot into the
1541 // memory location argument.
1542 DebugLoc dl = Op.getDebugLoc();
1543 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1544 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1545 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1546 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1551 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1552 SDNode *Node = Op.getNode();
1553 DebugLoc dl = Node->getDebugLoc();
1554 EVT VT = Node->getValueType(0);
1555 SDValue Chain = Op.getOperand(0);
1556 SDValue Size = Op.getOperand(1);
1557 SDValue Align = Op.getOperand(2);
1559 // Chain the dynamic stack allocation so that it doesn't modify the stack
1560 // pointer when other instructions are using the stack.
1561 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1563 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1564 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1565 if (AlignVal > StackAlign)
1566 // Do this now since selection pass cannot introduce new target
1567 // independent node.
1568 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1570 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1571 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1572 // do even more horrible hack later.
1573 MachineFunction &MF = DAG.getMachineFunction();
1574 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1575 if (AFI->isThumb1OnlyFunction()) {
1577 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1579 uint32_t Val = C->getZExtValue();
1580 if (Val <= 508 && ((Val & 3) == 0))
1584 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1587 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1588 SDValue Ops1[] = { Chain, Size, Align };
1589 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1590 Chain = Res.getValue(1);
1591 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1592 DAG.getIntPtrConstant(0, true), SDValue());
1593 SDValue Ops2[] = { Res, Chain };
1594 return DAG.getMergeValues(Ops2, 2, dl);
1598 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1599 SDValue &Root, SelectionDAG &DAG,
1601 MachineFunction &MF = DAG.getMachineFunction();
1602 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1604 TargetRegisterClass *RC;
1605 if (AFI->isThumb1OnlyFunction())
1606 RC = ARM::tGPRRegisterClass;
1608 RC = ARM::GPRRegisterClass;
1610 // Transform the arguments stored in physical registers into virtual ones.
1611 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1612 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1615 if (NextVA.isMemLoc()) {
1616 MachineFrameInfo *MFI = MF.getFrameInfo();
1617 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1619 // Create load node to retrieve arguments from the stack.
1620 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1621 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1622 PseudoSourceValue::getFixedStack(FI), 0,
1625 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1626 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1629 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1633 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1634 CallingConv::ID CallConv, bool isVarArg,
1635 const SmallVectorImpl<ISD::InputArg>
1637 DebugLoc dl, SelectionDAG &DAG,
1638 SmallVectorImpl<SDValue> &InVals) {
1640 MachineFunction &MF = DAG.getMachineFunction();
1641 MachineFrameInfo *MFI = MF.getFrameInfo();
1643 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1645 // Assign locations to all of the incoming arguments.
1646 SmallVector<CCValAssign, 16> ArgLocs;
1647 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1649 CCInfo.AnalyzeFormalArguments(Ins,
1650 CCAssignFnForNode(CallConv, /* Return*/ false,
1653 SmallVector<SDValue, 16> ArgValues;
1655 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1656 CCValAssign &VA = ArgLocs[i];
1658 // Arguments stored in registers.
1659 if (VA.isRegLoc()) {
1660 EVT RegVT = VA.getLocVT();
1663 if (VA.needsCustom()) {
1664 // f64 and vector types are split up into multiple registers or
1665 // combinations of registers and stack slots.
1666 if (VA.getLocVT() == MVT::v2f64) {
1667 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1669 VA = ArgLocs[++i]; // skip ahead to next loc
1671 if (VA.isMemLoc()) {
1672 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
1674 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1675 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
1676 PseudoSourceValue::getFixedStack(FI), 0,
1679 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1682 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1683 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1684 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1685 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1686 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1688 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1691 TargetRegisterClass *RC;
1693 if (RegVT == MVT::f32)
1694 RC = ARM::SPRRegisterClass;
1695 else if (RegVT == MVT::f64)
1696 RC = ARM::DPRRegisterClass;
1697 else if (RegVT == MVT::v2f64)
1698 RC = ARM::QPRRegisterClass;
1699 else if (RegVT == MVT::i32)
1700 RC = (AFI->isThumb1OnlyFunction() ?
1701 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1703 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1705 // Transform the arguments in physical registers into virtual ones.
1706 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1707 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1710 // If this is an 8 or 16-bit value, it is really passed promoted
1711 // to 32 bits. Insert an assert[sz]ext to capture this, then
1712 // truncate to the right size.
1713 switch (VA.getLocInfo()) {
1714 default: llvm_unreachable("Unknown loc info!");
1715 case CCValAssign::Full: break;
1716 case CCValAssign::BCvt:
1717 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1719 case CCValAssign::SExt:
1720 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1721 DAG.getValueType(VA.getValVT()));
1722 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1724 case CCValAssign::ZExt:
1725 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1726 DAG.getValueType(VA.getValVT()));
1727 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1731 InVals.push_back(ArgValue);
1733 } else { // VA.isRegLoc()
1736 assert(VA.isMemLoc());
1737 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1739 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1740 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1743 // Create load nodes to retrieve arguments from the stack.
1744 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1745 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1746 PseudoSourceValue::getFixedStack(FI), 0,
1753 static const unsigned GPRArgRegs[] = {
1754 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1757 unsigned NumGPRs = CCInfo.getFirstUnallocated
1758 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1760 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1761 unsigned VARegSize = (4 - NumGPRs) * 4;
1762 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1763 unsigned ArgOffset = CCInfo.getNextStackOffset();
1764 if (VARegSaveSize) {
1765 // If this function is vararg, store any remaining integer argument regs
1766 // to their spots on the stack so that they may be loaded by deferencing
1767 // the result of va_next.
1768 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1769 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1770 VARegSaveSize - VARegSize,
1772 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1774 SmallVector<SDValue, 4> MemOps;
1775 for (; NumGPRs < 4; ++NumGPRs) {
1776 TargetRegisterClass *RC;
1777 if (AFI->isThumb1OnlyFunction())
1778 RC = ARM::tGPRRegisterClass;
1780 RC = ARM::GPRRegisterClass;
1782 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1783 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1784 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1785 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0,
1787 MemOps.push_back(Store);
1788 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1789 DAG.getConstant(4, getPointerTy()));
1791 if (!MemOps.empty())
1792 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1793 &MemOps[0], MemOps.size());
1795 // This will point to the next argument passed via stack.
1796 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
1802 /// isFloatingPointZero - Return true if this is +0.0.
1803 static bool isFloatingPointZero(SDValue Op) {
1804 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1805 return CFP->getValueAPF().isPosZero();
1806 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1807 // Maybe this has already been legalized into the constant pool?
1808 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1809 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1810 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1811 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1812 return CFP->getValueAPF().isPosZero();
1818 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1819 /// the given operands.
1821 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1822 SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
1823 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1824 unsigned C = RHSC->getZExtValue();
1825 if (!isLegalICmpImmediate(C)) {
1826 // Constant does not fit, try adjusting it by one?
1831 if (isLegalICmpImmediate(C-1)) {
1832 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1833 RHS = DAG.getConstant(C-1, MVT::i32);
1838 if (C > 0 && isLegalICmpImmediate(C-1)) {
1839 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1840 RHS = DAG.getConstant(C-1, MVT::i32);
1845 if (isLegalICmpImmediate(C+1)) {
1846 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1847 RHS = DAG.getConstant(C+1, MVT::i32);
1852 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1853 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1854 RHS = DAG.getConstant(C+1, MVT::i32);
1861 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1862 ARMISD::NodeType CompareType;
1865 CompareType = ARMISD::CMP;
1870 CompareType = ARMISD::CMPZ;
1873 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1874 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1877 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1878 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1881 if (!isFloatingPointZero(RHS))
1882 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1884 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1885 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1888 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1889 EVT VT = Op.getValueType();
1890 SDValue LHS = Op.getOperand(0);
1891 SDValue RHS = Op.getOperand(1);
1892 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1893 SDValue TrueVal = Op.getOperand(2);
1894 SDValue FalseVal = Op.getOperand(3);
1895 DebugLoc dl = Op.getDebugLoc();
1897 if (LHS.getValueType() == MVT::i32) {
1899 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1900 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1901 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1904 ARMCC::CondCodes CondCode, CondCode2;
1905 FPCCToARMCC(CC, CondCode, CondCode2);
1907 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1908 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1909 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1910 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1912 if (CondCode2 != ARMCC::AL) {
1913 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1914 // FIXME: Needs another CMP because flag can have but one use.
1915 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1916 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1917 Result, TrueVal, ARMCC2, CCR, Cmp2);
1922 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1923 SDValue Chain = Op.getOperand(0);
1924 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1925 SDValue LHS = Op.getOperand(2);
1926 SDValue RHS = Op.getOperand(3);
1927 SDValue Dest = Op.getOperand(4);
1928 DebugLoc dl = Op.getDebugLoc();
1930 if (LHS.getValueType() == MVT::i32) {
1932 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1933 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1934 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1935 Chain, Dest, ARMCC, CCR,Cmp);
1938 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1939 ARMCC::CondCodes CondCode, CondCode2;
1940 FPCCToARMCC(CC, CondCode, CondCode2);
1942 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1943 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1944 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1945 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1946 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1947 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1948 if (CondCode2 != ARMCC::AL) {
1949 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1950 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1951 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1956 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1957 SDValue Chain = Op.getOperand(0);
1958 SDValue Table = Op.getOperand(1);
1959 SDValue Index = Op.getOperand(2);
1960 DebugLoc dl = Op.getDebugLoc();
1962 EVT PTy = getPointerTy();
1963 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1964 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1965 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1966 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1967 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1968 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1969 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1970 if (Subtarget->isThumb2()) {
1971 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1972 // which does another jump to the destination. This also makes it easier
1973 // to translate it to TBB / TBH later.
1974 // FIXME: This might not work if the function is extremely large.
1975 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1976 Addr, Op.getOperand(2), JTI, UId);
1978 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1979 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1980 PseudoSourceValue::getJumpTable(), 0,
1982 Chain = Addr.getValue(1);
1983 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1984 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1986 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1987 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
1988 Chain = Addr.getValue(1);
1989 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1993 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1994 DebugLoc dl = Op.getDebugLoc();
1997 switch (Op.getOpcode()) {
1999 assert(0 && "Invalid opcode!");
2000 case ISD::FP_TO_SINT:
2001 Opc = ARMISD::FTOSI;
2003 case ISD::FP_TO_UINT:
2004 Opc = ARMISD::FTOUI;
2007 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2008 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2011 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2012 EVT VT = Op.getValueType();
2013 DebugLoc dl = Op.getDebugLoc();
2016 switch (Op.getOpcode()) {
2018 assert(0 && "Invalid opcode!");
2019 case ISD::SINT_TO_FP:
2020 Opc = ARMISD::SITOF;
2022 case ISD::UINT_TO_FP:
2023 Opc = ARMISD::UITOF;
2027 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2028 return DAG.getNode(Opc, dl, VT, Op);
2031 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2032 // Implement fcopysign with a fabs and a conditional fneg.
2033 SDValue Tmp0 = Op.getOperand(0);
2034 SDValue Tmp1 = Op.getOperand(1);
2035 DebugLoc dl = Op.getDebugLoc();
2036 EVT VT = Op.getValueType();
2037 EVT SrcVT = Tmp1.getValueType();
2038 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2039 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2040 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2041 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2042 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2045 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
2046 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2047 MFI->setFrameAddressIsTaken(true);
2048 EVT VT = Op.getValueType();
2049 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2050 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2051 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2052 ? ARM::R7 : ARM::R11;
2053 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2055 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2061 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2063 SDValue Dst, SDValue Src,
2064 SDValue Size, unsigned Align,
2065 bool isVolatile, bool AlwaysInline,
2066 const Value *DstSV, uint64_t DstSVOff,
2067 const Value *SrcSV, uint64_t SrcSVOff){
2068 // Do repeated 4-byte loads and stores. To be improved.
2069 // This requires 4-byte alignment.
2070 if ((Align & 3) != 0)
2072 // This requires the copy size to be a constant, preferrably
2073 // within a subtarget-specific limit.
2074 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2077 uint64_t SizeVal = ConstantSize->getZExtValue();
2078 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2081 unsigned BytesLeft = SizeVal & 3;
2082 unsigned NumMemOps = SizeVal >> 2;
2083 unsigned EmittedNumMemOps = 0;
2085 unsigned VTSize = 4;
2087 const unsigned MAX_LOADS_IN_LDM = 6;
2088 SDValue TFOps[MAX_LOADS_IN_LDM];
2089 SDValue Loads[MAX_LOADS_IN_LDM];
2090 uint64_t SrcOff = 0, DstOff = 0;
2092 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2093 // same number of stores. The loads and stores will get combined into
2094 // ldm/stm later on.
2095 while (EmittedNumMemOps < NumMemOps) {
2097 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2098 Loads[i] = DAG.getLoad(VT, dl, Chain,
2099 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2100 DAG.getConstant(SrcOff, MVT::i32)),
2101 SrcSV, SrcSVOff + SrcOff, isVolatile, false, 0);
2102 TFOps[i] = Loads[i].getValue(1);
2105 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2108 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2109 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2110 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2111 DAG.getConstant(DstOff, MVT::i32)),
2112 DstSV, DstSVOff + DstOff, isVolatile, false, 0);
2115 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2117 EmittedNumMemOps += i;
2123 // Issue loads / stores for the trailing (1 - 3) bytes.
2124 unsigned BytesLeftSave = BytesLeft;
2127 if (BytesLeft >= 2) {
2135 Loads[i] = DAG.getLoad(VT, dl, Chain,
2136 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2137 DAG.getConstant(SrcOff, MVT::i32)),
2138 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2139 TFOps[i] = Loads[i].getValue(1);
2142 BytesLeft -= VTSize;
2144 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2147 BytesLeft = BytesLeftSave;
2149 if (BytesLeft >= 2) {
2157 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2158 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2159 DAG.getConstant(DstOff, MVT::i32)),
2160 DstSV, DstSVOff + DstOff, false, false, 0);
2163 BytesLeft -= VTSize;
2165 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2168 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2169 SDValue Op = N->getOperand(0);
2171 // Do not create a VMOVDRR or VMOVRRD node if the operand type is not
2172 // legal. The legalizer won't know what to do with that.
2173 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2174 if (!TLI.isTypeLegal(Op.getValueType()))
2177 DebugLoc dl = N->getDebugLoc();
2178 if (N->getValueType(0) == MVT::f64) {
2179 // Turn i64->f64 into VMOVDRR.
2180 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2181 DAG.getConstant(0, MVT::i32));
2182 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2183 DAG.getConstant(1, MVT::i32));
2184 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2187 // Turn f64->i64 into VMOVRRD.
2188 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2189 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2191 // Merge the pieces into a single i64 value.
2192 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2195 /// getZeroVector - Returns a vector of specified type with all zero elements.
2197 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2198 assert(VT.isVector() && "Expected a vector type");
2200 // Zero vectors are used to represent vector negation and in those cases
2201 // will be implemented with the NEON VNEG instruction. However, VNEG does
2202 // not support i64 elements, so sometimes the zero vectors will need to be
2203 // explicitly constructed. For those cases, and potentially other uses in
2204 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2205 // to their dest type. This ensures they get CSE'd.
2207 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2208 SmallVector<SDValue, 8> Ops;
2211 if (VT.getSizeInBits() == 64) {
2212 Ops.assign(8, Cst); TVT = MVT::v8i8;
2214 Ops.assign(16, Cst); TVT = MVT::v16i8;
2216 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2218 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2221 /// getOnesVector - Returns a vector of specified type with all bits set.
2223 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2224 assert(VT.isVector() && "Expected a vector type");
2226 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2227 // dest type. This ensures they get CSE'd.
2229 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2230 SmallVector<SDValue, 8> Ops;
2233 if (VT.getSizeInBits() == 64) {
2234 Ops.assign(8, Cst); TVT = MVT::v8i8;
2236 Ops.assign(16, Cst); TVT = MVT::v16i8;
2238 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2240 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2243 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2244 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2245 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
2246 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2247 EVT VT = Op.getValueType();
2248 unsigned VTBits = VT.getSizeInBits();
2249 DebugLoc dl = Op.getDebugLoc();
2250 SDValue ShOpLo = Op.getOperand(0);
2251 SDValue ShOpHi = Op.getOperand(1);
2252 SDValue ShAmt = Op.getOperand(2);
2254 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2256 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2258 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2259 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2260 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2261 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2262 DAG.getConstant(VTBits, MVT::i32));
2263 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2264 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2265 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2267 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2268 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2270 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2271 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2274 SDValue Ops[2] = { Lo, Hi };
2275 return DAG.getMergeValues(Ops, 2, dl);
2278 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2279 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2280 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
2281 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2282 EVT VT = Op.getValueType();
2283 unsigned VTBits = VT.getSizeInBits();
2284 DebugLoc dl = Op.getDebugLoc();
2285 SDValue ShOpLo = Op.getOperand(0);
2286 SDValue ShOpHi = Op.getOperand(1);
2287 SDValue ShAmt = Op.getOperand(2);
2290 assert(Op.getOpcode() == ISD::SHL_PARTS);
2291 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2292 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2293 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2294 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2295 DAG.getConstant(VTBits, MVT::i32));
2296 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2297 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2299 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2300 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2301 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2303 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2304 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2307 SDValue Ops[2] = { Lo, Hi };
2308 return DAG.getMergeValues(Ops, 2, dl);
2311 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2312 const ARMSubtarget *ST) {
2313 EVT VT = N->getValueType(0);
2314 DebugLoc dl = N->getDebugLoc();
2316 if (!ST->hasV6T2Ops())
2319 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2320 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2323 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2324 const ARMSubtarget *ST) {
2325 EVT VT = N->getValueType(0);
2326 DebugLoc dl = N->getDebugLoc();
2328 // Lower vector shifts on NEON to use VSHL.
2329 if (VT.isVector()) {
2330 assert(ST->hasNEON() && "unexpected vector shift");
2332 // Left shifts translate directly to the vshiftu intrinsic.
2333 if (N->getOpcode() == ISD::SHL)
2334 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2335 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2336 N->getOperand(0), N->getOperand(1));
2338 assert((N->getOpcode() == ISD::SRA ||
2339 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2341 // NEON uses the same intrinsics for both left and right shifts. For
2342 // right shifts, the shift amounts are negative, so negate the vector of
2344 EVT ShiftVT = N->getOperand(1).getValueType();
2345 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2346 getZeroVector(ShiftVT, DAG, dl),
2348 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2349 Intrinsic::arm_neon_vshifts :
2350 Intrinsic::arm_neon_vshiftu);
2351 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2352 DAG.getConstant(vshiftInt, MVT::i32),
2353 N->getOperand(0), NegatedCount);
2356 // We can get here for a node like i32 = ISD::SHL i32, i64
2360 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2361 "Unknown shift to lower!");
2363 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2364 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2365 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2368 // If we are in thumb mode, we don't have RRX.
2369 if (ST->isThumb1Only()) return SDValue();
2371 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2372 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2373 DAG.getConstant(0, MVT::i32));
2374 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2375 DAG.getConstant(1, MVT::i32));
2377 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2378 // captures the result into a carry flag.
2379 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2380 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2382 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2383 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2385 // Merge the pieces into a single i64 value.
2386 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2389 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2390 SDValue TmpOp0, TmpOp1;
2391 bool Invert = false;
2395 SDValue Op0 = Op.getOperand(0);
2396 SDValue Op1 = Op.getOperand(1);
2397 SDValue CC = Op.getOperand(2);
2398 EVT VT = Op.getValueType();
2399 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2400 DebugLoc dl = Op.getDebugLoc();
2402 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2403 switch (SetCCOpcode) {
2404 default: llvm_unreachable("Illegal FP comparison"); break;
2406 case ISD::SETNE: Invert = true; // Fallthrough
2408 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2410 case ISD::SETLT: Swap = true; // Fallthrough
2412 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2414 case ISD::SETLE: Swap = true; // Fallthrough
2416 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2417 case ISD::SETUGE: Swap = true; // Fallthrough
2418 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2419 case ISD::SETUGT: Swap = true; // Fallthrough
2420 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2421 case ISD::SETUEQ: Invert = true; // Fallthrough
2423 // Expand this to (OLT | OGT).
2427 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2428 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2430 case ISD::SETUO: Invert = true; // Fallthrough
2432 // Expand this to (OLT | OGE).
2436 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2437 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2441 // Integer comparisons.
2442 switch (SetCCOpcode) {
2443 default: llvm_unreachable("Illegal integer comparison"); break;
2444 case ISD::SETNE: Invert = true;
2445 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2446 case ISD::SETLT: Swap = true;
2447 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2448 case ISD::SETLE: Swap = true;
2449 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2450 case ISD::SETULT: Swap = true;
2451 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2452 case ISD::SETULE: Swap = true;
2453 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2456 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2457 if (Opc == ARMISD::VCEQ) {
2460 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2462 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2465 // Ignore bitconvert.
2466 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2467 AndOp = AndOp.getOperand(0);
2469 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2471 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2472 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2479 std::swap(Op0, Op1);
2481 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2484 Result = DAG.getNOT(dl, Result, VT);
2489 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2490 /// VMOV instruction, and if so, return the constant being splatted.
2491 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2492 unsigned SplatBitSize, SelectionDAG &DAG) {
2493 switch (SplatBitSize) {
2495 // Any 1-byte value is OK.
2496 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2497 return DAG.getTargetConstant(SplatBits, MVT::i8);
2500 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2501 if ((SplatBits & ~0xff) == 0 ||
2502 (SplatBits & ~0xff00) == 0)
2503 return DAG.getTargetConstant(SplatBits, MVT::i16);
2507 // NEON's 32-bit VMOV supports splat values where:
2508 // * only one byte is nonzero, or
2509 // * the least significant byte is 0xff and the second byte is nonzero, or
2510 // * the least significant 2 bytes are 0xff and the third is nonzero.
2511 if ((SplatBits & ~0xff) == 0 ||
2512 (SplatBits & ~0xff00) == 0 ||
2513 (SplatBits & ~0xff0000) == 0 ||
2514 (SplatBits & ~0xff000000) == 0)
2515 return DAG.getTargetConstant(SplatBits, MVT::i32);
2517 if ((SplatBits & ~0xffff) == 0 &&
2518 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2519 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2521 if ((SplatBits & ~0xffffff) == 0 &&
2522 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2523 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2525 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2526 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2527 // VMOV.I32. A (very) minor optimization would be to replicate the value
2528 // and fall through here to test for a valid 64-bit splat. But, then the
2529 // caller would also need to check and handle the change in size.
2533 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2534 uint64_t BitMask = 0xff;
2536 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2537 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2539 else if ((SplatBits & BitMask) != 0)
2543 return DAG.getTargetConstant(Val, MVT::i64);
2547 llvm_unreachable("unexpected size for isVMOVSplat");
2554 /// getVMOVImm - If this is a build_vector of constants which can be
2555 /// formed by using a VMOV instruction of the specified element size,
2556 /// return the constant being splatted. The ByteSize field indicates the
2557 /// number of bytes of each element [1248].
2558 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2559 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2560 APInt SplatBits, SplatUndef;
2561 unsigned SplatBitSize;
2563 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2564 HasAnyUndefs, ByteSize * 8))
2567 if (SplatBitSize > ByteSize * 8)
2570 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2574 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2575 bool &ReverseVEXT, unsigned &Imm) {
2576 unsigned NumElts = VT.getVectorNumElements();
2577 ReverseVEXT = false;
2580 // If this is a VEXT shuffle, the immediate value is the index of the first
2581 // element. The other shuffle indices must be the successive elements after
2583 unsigned ExpectedElt = Imm;
2584 for (unsigned i = 1; i < NumElts; ++i) {
2585 // Increment the expected index. If it wraps around, it may still be
2586 // a VEXT but the source vectors must be swapped.
2588 if (ExpectedElt == NumElts * 2) {
2593 if (ExpectedElt != static_cast<unsigned>(M[i]))
2597 // Adjust the index value if the source operands will be swapped.
2604 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2605 /// instruction with the specified blocksize. (The order of the elements
2606 /// within each block of the vector is reversed.)
2607 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2608 unsigned BlockSize) {
2609 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2610 "Only possible block sizes for VREV are: 16, 32, 64");
2612 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2616 unsigned NumElts = VT.getVectorNumElements();
2617 unsigned BlockElts = M[0] + 1;
2619 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2622 for (unsigned i = 0; i < NumElts; ++i) {
2623 if ((unsigned) M[i] !=
2624 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2631 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2632 unsigned &WhichResult) {
2633 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2637 unsigned NumElts = VT.getVectorNumElements();
2638 WhichResult = (M[0] == 0 ? 0 : 1);
2639 for (unsigned i = 0; i < NumElts; i += 2) {
2640 if ((unsigned) M[i] != i + WhichResult ||
2641 (unsigned) M[i+1] != i + NumElts + WhichResult)
2647 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2648 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2649 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2650 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2651 unsigned &WhichResult) {
2652 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2656 unsigned NumElts = VT.getVectorNumElements();
2657 WhichResult = (M[0] == 0 ? 0 : 1);
2658 for (unsigned i = 0; i < NumElts; i += 2) {
2659 if ((unsigned) M[i] != i + WhichResult ||
2660 (unsigned) M[i+1] != i + WhichResult)
2666 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2667 unsigned &WhichResult) {
2668 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2672 unsigned NumElts = VT.getVectorNumElements();
2673 WhichResult = (M[0] == 0 ? 0 : 1);
2674 for (unsigned i = 0; i != NumElts; ++i) {
2675 if ((unsigned) M[i] != 2 * i + WhichResult)
2679 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2680 if (VT.is64BitVector() && EltSz == 32)
2686 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2687 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2688 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2689 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2690 unsigned &WhichResult) {
2691 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2695 unsigned Half = VT.getVectorNumElements() / 2;
2696 WhichResult = (M[0] == 0 ? 0 : 1);
2697 for (unsigned j = 0; j != 2; ++j) {
2698 unsigned Idx = WhichResult;
2699 for (unsigned i = 0; i != Half; ++i) {
2700 if ((unsigned) M[i + j * Half] != Idx)
2706 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2707 if (VT.is64BitVector() && EltSz == 32)
2713 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2714 unsigned &WhichResult) {
2715 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2719 unsigned NumElts = VT.getVectorNumElements();
2720 WhichResult = (M[0] == 0 ? 0 : 1);
2721 unsigned Idx = WhichResult * NumElts / 2;
2722 for (unsigned i = 0; i != NumElts; i += 2) {
2723 if ((unsigned) M[i] != Idx ||
2724 (unsigned) M[i+1] != Idx + NumElts)
2729 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2730 if (VT.is64BitVector() && EltSz == 32)
2736 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2737 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2738 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2739 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2740 unsigned &WhichResult) {
2741 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2745 unsigned NumElts = VT.getVectorNumElements();
2746 WhichResult = (M[0] == 0 ? 0 : 1);
2747 unsigned Idx = WhichResult * NumElts / 2;
2748 for (unsigned i = 0; i != NumElts; i += 2) {
2749 if ((unsigned) M[i] != Idx ||
2750 (unsigned) M[i+1] != Idx)
2755 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2756 if (VT.is64BitVector() && EltSz == 32)
2763 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2764 // Canonicalize all-zeros and all-ones vectors.
2765 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2766 if (ConstVal->isNullValue())
2767 return getZeroVector(VT, DAG, dl);
2768 if (ConstVal->isAllOnesValue())
2769 return getOnesVector(VT, DAG, dl);
2772 if (VT.is64BitVector()) {
2773 switch (Val.getValueType().getSizeInBits()) {
2774 case 8: CanonicalVT = MVT::v8i8; break;
2775 case 16: CanonicalVT = MVT::v4i16; break;
2776 case 32: CanonicalVT = MVT::v2i32; break;
2777 case 64: CanonicalVT = MVT::v1i64; break;
2778 default: llvm_unreachable("unexpected splat element type"); break;
2781 assert(VT.is128BitVector() && "unknown splat vector size");
2782 switch (Val.getValueType().getSizeInBits()) {
2783 case 8: CanonicalVT = MVT::v16i8; break;
2784 case 16: CanonicalVT = MVT::v8i16; break;
2785 case 32: CanonicalVT = MVT::v4i32; break;
2786 case 64: CanonicalVT = MVT::v2i64; break;
2787 default: llvm_unreachable("unexpected splat element type"); break;
2791 // Build a canonical splat for this value.
2792 SmallVector<SDValue, 8> Ops;
2793 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2794 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2796 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2799 // If this is a case we can't handle, return null and let the default
2800 // expansion code take care of it.
2801 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2802 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2803 DebugLoc dl = Op.getDebugLoc();
2804 EVT VT = Op.getValueType();
2806 APInt SplatBits, SplatUndef;
2807 unsigned SplatBitSize;
2809 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2810 if (SplatBitSize <= 64) {
2811 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2812 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2814 return BuildSplat(Val, VT, DAG, dl);
2818 // If there are only 2 elements in a 128-bit vector, insert them into an
2819 // undef vector. This handles the common case for 128-bit vector argument
2820 // passing, where the insertions should be translated to subreg accesses
2821 // with no real instructions.
2822 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2823 SDValue Val = DAG.getUNDEF(VT);
2824 SDValue Op0 = Op.getOperand(0);
2825 SDValue Op1 = Op.getOperand(1);
2826 if (Op0.getOpcode() != ISD::UNDEF)
2827 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2828 DAG.getIntPtrConstant(0));
2829 if (Op1.getOpcode() != ISD::UNDEF)
2830 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2831 DAG.getIntPtrConstant(1));
2838 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2839 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2840 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2841 /// are assumed to be legal.
2843 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2845 if (VT.getVectorNumElements() == 4 &&
2846 (VT.is128BitVector() || VT.is64BitVector())) {
2847 unsigned PFIndexes[4];
2848 for (unsigned i = 0; i != 4; ++i) {
2852 PFIndexes[i] = M[i];
2855 // Compute the index in the perfect shuffle table.
2856 unsigned PFTableIndex =
2857 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2858 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2859 unsigned Cost = (PFEntry >> 30);
2866 unsigned Imm, WhichResult;
2868 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2869 isVREVMask(M, VT, 64) ||
2870 isVREVMask(M, VT, 32) ||
2871 isVREVMask(M, VT, 16) ||
2872 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2873 isVTRNMask(M, VT, WhichResult) ||
2874 isVUZPMask(M, VT, WhichResult) ||
2875 isVZIPMask(M, VT, WhichResult) ||
2876 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2877 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2878 isVZIP_v_undef_Mask(M, VT, WhichResult));
2881 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2882 /// the specified operations to build the shuffle.
2883 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2884 SDValue RHS, SelectionDAG &DAG,
2886 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2887 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2888 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2891 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2900 OP_VUZPL, // VUZP, left result
2901 OP_VUZPR, // VUZP, right result
2902 OP_VZIPL, // VZIP, left result
2903 OP_VZIPR, // VZIP, right result
2904 OP_VTRNL, // VTRN, left result
2905 OP_VTRNR // VTRN, right result
2908 if (OpNum == OP_COPY) {
2909 if (LHSID == (1*9+2)*9+3) return LHS;
2910 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2914 SDValue OpLHS, OpRHS;
2915 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2916 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2917 EVT VT = OpLHS.getValueType();
2920 default: llvm_unreachable("Unknown shuffle opcode!");
2922 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2927 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2928 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2932 return DAG.getNode(ARMISD::VEXT, dl, VT,
2934 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2937 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2938 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2941 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2942 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2945 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2946 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2950 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2951 SDValue V1 = Op.getOperand(0);
2952 SDValue V2 = Op.getOperand(1);
2953 DebugLoc dl = Op.getDebugLoc();
2954 EVT VT = Op.getValueType();
2955 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2956 SmallVector<int, 8> ShuffleMask;
2958 // Convert shuffles that are directly supported on NEON to target-specific
2959 // DAG nodes, instead of keeping them as shuffles and matching them again
2960 // during code selection. This is more efficient and avoids the possibility
2961 // of inconsistencies between legalization and selection.
2962 // FIXME: floating-point vectors should be canonicalized to integer vectors
2963 // of the same time so that they get CSEd properly.
2964 SVN->getMask(ShuffleMask);
2966 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2967 int Lane = SVN->getSplatIndex();
2968 // If this is undef splat, generate it via "just" vdup, if possible.
2969 if (Lane == -1) Lane = 0;
2971 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2972 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2974 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2975 DAG.getConstant(Lane, MVT::i32));
2980 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2983 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2984 DAG.getConstant(Imm, MVT::i32));
2987 if (isVREVMask(ShuffleMask, VT, 64))
2988 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2989 if (isVREVMask(ShuffleMask, VT, 32))
2990 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2991 if (isVREVMask(ShuffleMask, VT, 16))
2992 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2994 // Check for Neon shuffles that modify both input vectors in place.
2995 // If both results are used, i.e., if there are two shuffles with the same
2996 // source operands and with masks corresponding to both results of one of
2997 // these operations, DAG memoization will ensure that a single node is
2998 // used for both shuffles.
2999 unsigned WhichResult;
3000 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3001 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3002 V1, V2).getValue(WhichResult);
3003 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3004 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3005 V1, V2).getValue(WhichResult);
3006 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3007 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3008 V1, V2).getValue(WhichResult);
3010 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3011 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3012 V1, V1).getValue(WhichResult);
3013 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3014 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3015 V1, V1).getValue(WhichResult);
3016 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3017 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3018 V1, V1).getValue(WhichResult);
3020 // If the shuffle is not directly supported and it has 4 elements, use
3021 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3022 if (VT.getVectorNumElements() == 4 &&
3023 (VT.is128BitVector() || VT.is64BitVector())) {
3024 unsigned PFIndexes[4];
3025 for (unsigned i = 0; i != 4; ++i) {
3026 if (ShuffleMask[i] < 0)
3029 PFIndexes[i] = ShuffleMask[i];
3032 // Compute the index in the perfect shuffle table.
3033 unsigned PFTableIndex =
3034 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3036 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3037 unsigned Cost = (PFEntry >> 30);
3040 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3046 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3047 EVT VT = Op.getValueType();
3048 DebugLoc dl = Op.getDebugLoc();
3049 SDValue Vec = Op.getOperand(0);
3050 SDValue Lane = Op.getOperand(1);
3051 assert(VT == MVT::i32 &&
3052 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3053 "unexpected type for custom-lowering vector extract");
3054 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3057 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3058 // The only time a CONCAT_VECTORS operation can have legal types is when
3059 // two 64-bit vectors are concatenated to a 128-bit vector.
3060 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3061 "unexpected CONCAT_VECTORS");
3062 DebugLoc dl = Op.getDebugLoc();
3063 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3064 SDValue Op0 = Op.getOperand(0);
3065 SDValue Op1 = Op.getOperand(1);
3066 if (Op0.getOpcode() != ISD::UNDEF)
3067 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3068 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3069 DAG.getIntPtrConstant(0));
3070 if (Op1.getOpcode() != ISD::UNDEF)
3071 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3072 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3073 DAG.getIntPtrConstant(1));
3074 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3077 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3078 switch (Op.getOpcode()) {
3079 default: llvm_unreachable("Don't know how to custom lower this!");
3080 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3081 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3082 case ISD::GlobalAddress:
3083 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3084 LowerGlobalAddressELF(Op, DAG);
3085 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3086 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3087 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3088 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3089 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3090 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
3091 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3092 case ISD::SINT_TO_FP:
3093 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3094 case ISD::FP_TO_SINT:
3095 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3096 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3097 case ISD::RETURNADDR: break;
3098 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3099 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3100 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3102 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3105 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3106 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3107 case ISD::SRL_PARTS:
3108 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3109 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3110 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3111 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3112 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3113 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3114 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3119 /// ReplaceNodeResults - Replace the results of node with an illegal result
3120 /// type with new values built out of custom code.
3121 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3122 SmallVectorImpl<SDValue>&Results,
3123 SelectionDAG &DAG) {
3125 switch (N->getOpcode()) {
3127 llvm_unreachable("Don't know how to custom expand this!");
3129 case ISD::BIT_CONVERT:
3130 Res = ExpandBIT_CONVERT(N, DAG);
3134 Res = LowerShift(N, DAG, Subtarget);
3138 Results.push_back(Res);
3141 //===----------------------------------------------------------------------===//
3142 // ARM Scheduler Hooks
3143 //===----------------------------------------------------------------------===//
3146 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3147 MachineBasicBlock *BB,
3148 unsigned Size) const {
3149 unsigned dest = MI->getOperand(0).getReg();
3150 unsigned ptr = MI->getOperand(1).getReg();
3151 unsigned oldval = MI->getOperand(2).getReg();
3152 unsigned newval = MI->getOperand(3).getReg();
3153 unsigned scratch = BB->getParent()->getRegInfo()
3154 .createVirtualRegister(ARM::GPRRegisterClass);
3155 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3156 DebugLoc dl = MI->getDebugLoc();
3157 bool isThumb2 = Subtarget->isThumb2();
3159 unsigned ldrOpc, strOpc;
3161 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3163 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3164 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3167 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3168 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3171 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3172 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3176 MachineFunction *MF = BB->getParent();
3177 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3178 MachineFunction::iterator It = BB;
3179 ++It; // insert the new blocks after the current block
3181 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3182 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3183 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3184 MF->insert(It, loop1MBB);
3185 MF->insert(It, loop2MBB);
3186 MF->insert(It, exitMBB);
3187 exitMBB->transferSuccessors(BB);
3191 // fallthrough --> loop1MBB
3192 BB->addSuccessor(loop1MBB);
3195 // ldrex dest, [ptr]
3199 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3200 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3201 .addReg(dest).addReg(oldval));
3202 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3203 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3204 BB->addSuccessor(loop2MBB);
3205 BB->addSuccessor(exitMBB);
3208 // strex scratch, newval, [ptr]
3212 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3214 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3215 .addReg(scratch).addImm(0));
3216 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3217 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3218 BB->addSuccessor(loop1MBB);
3219 BB->addSuccessor(exitMBB);
3225 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3231 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3232 unsigned Size, unsigned BinOpcode) const {
3233 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3234 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3236 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3237 MachineFunction *MF = BB->getParent();
3238 MachineFunction::iterator It = BB;
3241 unsigned dest = MI->getOperand(0).getReg();
3242 unsigned ptr = MI->getOperand(1).getReg();
3243 unsigned incr = MI->getOperand(2).getReg();
3244 DebugLoc dl = MI->getDebugLoc();
3246 bool isThumb2 = Subtarget->isThumb2();
3247 unsigned ldrOpc, strOpc;
3249 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3251 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3252 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3255 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3256 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3259 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3260 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3264 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3265 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3266 MF->insert(It, loopMBB);
3267 MF->insert(It, exitMBB);
3268 exitMBB->transferSuccessors(BB);
3270 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3271 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3272 unsigned scratch2 = (!BinOpcode) ? incr :
3273 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3277 // fallthrough --> loopMBB
3278 BB->addSuccessor(loopMBB);
3282 // <binop> scratch2, dest, incr
3283 // strex scratch, scratch2, ptr
3286 // fallthrough --> exitMBB
3288 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3290 // operand order needs to go the other way for NAND
3291 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3292 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3293 addReg(incr).addReg(dest)).addReg(0);
3295 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3296 addReg(dest).addReg(incr)).addReg(0);
3299 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3301 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3302 .addReg(scratch).addImm(0));
3303 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3304 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3306 BB->addSuccessor(loopMBB);
3307 BB->addSuccessor(exitMBB);
3313 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3319 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3320 MachineBasicBlock *BB,
3321 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
3322 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3323 DebugLoc dl = MI->getDebugLoc();
3324 bool isThumb2 = Subtarget->isThumb2();
3325 switch (MI->getOpcode()) {
3328 llvm_unreachable("Unexpected instr type to insert");
3330 case ARM::ATOMIC_LOAD_ADD_I8:
3331 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3332 case ARM::ATOMIC_LOAD_ADD_I16:
3333 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3334 case ARM::ATOMIC_LOAD_ADD_I32:
3335 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3337 case ARM::ATOMIC_LOAD_AND_I8:
3338 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3339 case ARM::ATOMIC_LOAD_AND_I16:
3340 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3341 case ARM::ATOMIC_LOAD_AND_I32:
3342 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3344 case ARM::ATOMIC_LOAD_OR_I8:
3345 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3346 case ARM::ATOMIC_LOAD_OR_I16:
3347 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3348 case ARM::ATOMIC_LOAD_OR_I32:
3349 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3351 case ARM::ATOMIC_LOAD_XOR_I8:
3352 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3353 case ARM::ATOMIC_LOAD_XOR_I16:
3354 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3355 case ARM::ATOMIC_LOAD_XOR_I32:
3356 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3358 case ARM::ATOMIC_LOAD_NAND_I8:
3359 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3360 case ARM::ATOMIC_LOAD_NAND_I16:
3361 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3362 case ARM::ATOMIC_LOAD_NAND_I32:
3363 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3365 case ARM::ATOMIC_LOAD_SUB_I8:
3366 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3367 case ARM::ATOMIC_LOAD_SUB_I16:
3368 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3369 case ARM::ATOMIC_LOAD_SUB_I32:
3370 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3372 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3373 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3374 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3376 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3377 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3378 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3380 case ARM::tMOVCCr_pseudo: {
3381 // To "insert" a SELECT_CC instruction, we actually have to insert the
3382 // diamond control-flow pattern. The incoming instruction knows the
3383 // destination vreg to set, the condition code register to branch on, the
3384 // true/false values to select between, and a branch opcode to use.
3385 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3386 MachineFunction::iterator It = BB;
3392 // cmpTY ccX, r1, r2
3394 // fallthrough --> copy0MBB
3395 MachineBasicBlock *thisMBB = BB;
3396 MachineFunction *F = BB->getParent();
3397 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3398 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3399 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3400 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3401 F->insert(It, copy0MBB);
3402 F->insert(It, sinkMBB);
3403 // Update machine-CFG edges by first adding all successors of the current
3404 // block to the new block which will contain the Phi node for the select.
3405 // Also inform sdisel of the edge changes.
3406 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3407 E = BB->succ_end(); I != E; ++I) {
3408 EM->insert(std::make_pair(*I, sinkMBB));
3409 sinkMBB->addSuccessor(*I);
3411 // Next, remove all successors of the current block, and add the true
3412 // and fallthrough blocks as its successors.
3413 while (!BB->succ_empty())
3414 BB->removeSuccessor(BB->succ_begin());
3415 BB->addSuccessor(copy0MBB);
3416 BB->addSuccessor(sinkMBB);
3419 // %FalseValue = ...
3420 // # fallthrough to sinkMBB
3423 // Update machine-CFG edges
3424 BB->addSuccessor(sinkMBB);
3427 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3430 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3431 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3432 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3434 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3441 case ARM::t2SUBrSPi_:
3442 case ARM::t2SUBrSPi12_:
3443 case ARM::t2SUBrSPs_: {
3444 MachineFunction *MF = BB->getParent();
3445 unsigned DstReg = MI->getOperand(0).getReg();
3446 unsigned SrcReg = MI->getOperand(1).getReg();
3447 bool DstIsDead = MI->getOperand(0).isDead();
3448 bool SrcIsKill = MI->getOperand(1).isKill();
3450 if (SrcReg != ARM::SP) {
3451 // Copy the source to SP from virtual register.
3452 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3453 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3454 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3455 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3456 .addReg(SrcReg, getKillRegState(SrcIsKill));
3460 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3461 switch (MI->getOpcode()) {
3463 llvm_unreachable("Unexpected pseudo instruction!");
3469 OpOpc = ARM::tADDspr;
3472 OpOpc = ARM::tSUBspi;
3474 case ARM::t2SUBrSPi_:
3475 OpOpc = ARM::t2SUBrSPi;
3476 NeedPred = true; NeedCC = true;
3478 case ARM::t2SUBrSPi12_:
3479 OpOpc = ARM::t2SUBrSPi12;
3482 case ARM::t2SUBrSPs_:
3483 OpOpc = ARM::t2SUBrSPs;
3484 NeedPred = true; NeedCC = true; NeedOp3 = true;
3487 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3488 if (OpOpc == ARM::tAND)
3489 AddDefaultT1CC(MIB);
3490 MIB.addReg(ARM::SP);
3491 MIB.addOperand(MI->getOperand(2));
3493 MIB.addOperand(MI->getOperand(3));
3495 AddDefaultPred(MIB);
3499 // Copy the result from SP to virtual register.
3500 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3501 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3502 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3503 BuildMI(BB, dl, TII->get(CopyOpc))
3504 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3506 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3512 //===----------------------------------------------------------------------===//
3513 // ARM Optimization Hooks
3514 //===----------------------------------------------------------------------===//
3517 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3518 TargetLowering::DAGCombinerInfo &DCI) {
3519 SelectionDAG &DAG = DCI.DAG;
3520 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3521 EVT VT = N->getValueType(0);
3522 unsigned Opc = N->getOpcode();
3523 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3524 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3525 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3526 ISD::CondCode CC = ISD::SETCC_INVALID;
3529 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3531 SDValue CCOp = Slct.getOperand(0);
3532 if (CCOp.getOpcode() == ISD::SETCC)
3533 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3536 bool DoXform = false;
3538 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3541 if (LHS.getOpcode() == ISD::Constant &&
3542 cast<ConstantSDNode>(LHS)->isNullValue()) {
3544 } else if (CC != ISD::SETCC_INVALID &&
3545 RHS.getOpcode() == ISD::Constant &&
3546 cast<ConstantSDNode>(RHS)->isNullValue()) {
3547 std::swap(LHS, RHS);
3548 SDValue Op0 = Slct.getOperand(0);
3549 EVT OpVT = isSlctCC ? Op0.getValueType() :
3550 Op0.getOperand(0).getValueType();
3551 bool isInt = OpVT.isInteger();
3552 CC = ISD::getSetCCInverse(CC, isInt);
3554 if (!TLI.isCondCodeLegal(CC, OpVT))
3555 return SDValue(); // Inverse operator isn't legal.
3562 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3564 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3565 Slct.getOperand(0), Slct.getOperand(1), CC);
3566 SDValue CCOp = Slct.getOperand(0);
3568 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3569 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3570 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3571 CCOp, OtherOp, Result);
3576 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3577 static SDValue PerformADDCombine(SDNode *N,
3578 TargetLowering::DAGCombinerInfo &DCI) {
3579 // added by evan in r37685 with no testcase.
3580 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3582 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3583 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3584 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3585 if (Result.getNode()) return Result;
3587 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3588 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3589 if (Result.getNode()) return Result;
3595 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3596 static SDValue PerformSUBCombine(SDNode *N,
3597 TargetLowering::DAGCombinerInfo &DCI) {
3598 // added by evan in r37685 with no testcase.
3599 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3601 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3602 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3603 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3604 if (Result.getNode()) return Result;
3610 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3611 /// ARMISD::VMOVRRD.
3612 static SDValue PerformVMOVRRDCombine(SDNode *N,
3613 TargetLowering::DAGCombinerInfo &DCI) {
3614 // fmrrd(fmdrr x, y) -> x,y
3615 SDValue InDouble = N->getOperand(0);
3616 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3617 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3621 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3622 /// operand of a vector shift operation, where all the elements of the
3623 /// build_vector must have the same constant integer value.
3624 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3625 // Ignore bit_converts.
3626 while (Op.getOpcode() == ISD::BIT_CONVERT)
3627 Op = Op.getOperand(0);
3628 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3629 APInt SplatBits, SplatUndef;
3630 unsigned SplatBitSize;
3632 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3633 HasAnyUndefs, ElementBits) ||
3634 SplatBitSize > ElementBits)
3636 Cnt = SplatBits.getSExtValue();
3640 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3641 /// operand of a vector shift left operation. That value must be in the range:
3642 /// 0 <= Value < ElementBits for a left shift; or
3643 /// 0 <= Value <= ElementBits for a long left shift.
3644 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3645 assert(VT.isVector() && "vector shift count is not a vector type");
3646 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3647 if (! getVShiftImm(Op, ElementBits, Cnt))
3649 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3652 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3653 /// operand of a vector shift right operation. For a shift opcode, the value
3654 /// is positive, but for an intrinsic the value count must be negative. The
3655 /// absolute value must be in the range:
3656 /// 1 <= |Value| <= ElementBits for a right shift; or
3657 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3658 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3660 assert(VT.isVector() && "vector shift count is not a vector type");
3661 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3662 if (! getVShiftImm(Op, ElementBits, Cnt))
3666 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3669 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3670 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3671 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3674 // Don't do anything for most intrinsics.
3677 // Vector shifts: check for immediate versions and lower them.
3678 // Note: This is done during DAG combining instead of DAG legalizing because
3679 // the build_vectors for 64-bit vector element shift counts are generally
3680 // not legal, and it is hard to see their values after they get legalized to
3681 // loads from a constant pool.
3682 case Intrinsic::arm_neon_vshifts:
3683 case Intrinsic::arm_neon_vshiftu:
3684 case Intrinsic::arm_neon_vshiftls:
3685 case Intrinsic::arm_neon_vshiftlu:
3686 case Intrinsic::arm_neon_vshiftn:
3687 case Intrinsic::arm_neon_vrshifts:
3688 case Intrinsic::arm_neon_vrshiftu:
3689 case Intrinsic::arm_neon_vrshiftn:
3690 case Intrinsic::arm_neon_vqshifts:
3691 case Intrinsic::arm_neon_vqshiftu:
3692 case Intrinsic::arm_neon_vqshiftsu:
3693 case Intrinsic::arm_neon_vqshiftns:
3694 case Intrinsic::arm_neon_vqshiftnu:
3695 case Intrinsic::arm_neon_vqshiftnsu:
3696 case Intrinsic::arm_neon_vqrshiftns:
3697 case Intrinsic::arm_neon_vqrshiftnu:
3698 case Intrinsic::arm_neon_vqrshiftnsu: {
3699 EVT VT = N->getOperand(1).getValueType();
3701 unsigned VShiftOpc = 0;
3704 case Intrinsic::arm_neon_vshifts:
3705 case Intrinsic::arm_neon_vshiftu:
3706 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3707 VShiftOpc = ARMISD::VSHL;
3710 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3711 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3712 ARMISD::VSHRs : ARMISD::VSHRu);
3717 case Intrinsic::arm_neon_vshiftls:
3718 case Intrinsic::arm_neon_vshiftlu:
3719 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3721 llvm_unreachable("invalid shift count for vshll intrinsic");
3723 case Intrinsic::arm_neon_vrshifts:
3724 case Intrinsic::arm_neon_vrshiftu:
3725 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3729 case Intrinsic::arm_neon_vqshifts:
3730 case Intrinsic::arm_neon_vqshiftu:
3731 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3735 case Intrinsic::arm_neon_vqshiftsu:
3736 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3738 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3740 case Intrinsic::arm_neon_vshiftn:
3741 case Intrinsic::arm_neon_vrshiftn:
3742 case Intrinsic::arm_neon_vqshiftns:
3743 case Intrinsic::arm_neon_vqshiftnu:
3744 case Intrinsic::arm_neon_vqshiftnsu:
3745 case Intrinsic::arm_neon_vqrshiftns:
3746 case Intrinsic::arm_neon_vqrshiftnu:
3747 case Intrinsic::arm_neon_vqrshiftnsu:
3748 // Narrowing shifts require an immediate right shift.
3749 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3751 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3754 llvm_unreachable("unhandled vector shift");
3758 case Intrinsic::arm_neon_vshifts:
3759 case Intrinsic::arm_neon_vshiftu:
3760 // Opcode already set above.
3762 case Intrinsic::arm_neon_vshiftls:
3763 case Intrinsic::arm_neon_vshiftlu:
3764 if (Cnt == VT.getVectorElementType().getSizeInBits())
3765 VShiftOpc = ARMISD::VSHLLi;
3767 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3768 ARMISD::VSHLLs : ARMISD::VSHLLu);
3770 case Intrinsic::arm_neon_vshiftn:
3771 VShiftOpc = ARMISD::VSHRN; break;
3772 case Intrinsic::arm_neon_vrshifts:
3773 VShiftOpc = ARMISD::VRSHRs; break;
3774 case Intrinsic::arm_neon_vrshiftu:
3775 VShiftOpc = ARMISD::VRSHRu; break;
3776 case Intrinsic::arm_neon_vrshiftn:
3777 VShiftOpc = ARMISD::VRSHRN; break;
3778 case Intrinsic::arm_neon_vqshifts:
3779 VShiftOpc = ARMISD::VQSHLs; break;
3780 case Intrinsic::arm_neon_vqshiftu:
3781 VShiftOpc = ARMISD::VQSHLu; break;
3782 case Intrinsic::arm_neon_vqshiftsu:
3783 VShiftOpc = ARMISD::VQSHLsu; break;
3784 case Intrinsic::arm_neon_vqshiftns:
3785 VShiftOpc = ARMISD::VQSHRNs; break;
3786 case Intrinsic::arm_neon_vqshiftnu:
3787 VShiftOpc = ARMISD::VQSHRNu; break;
3788 case Intrinsic::arm_neon_vqshiftnsu:
3789 VShiftOpc = ARMISD::VQSHRNsu; break;
3790 case Intrinsic::arm_neon_vqrshiftns:
3791 VShiftOpc = ARMISD::VQRSHRNs; break;
3792 case Intrinsic::arm_neon_vqrshiftnu:
3793 VShiftOpc = ARMISD::VQRSHRNu; break;
3794 case Intrinsic::arm_neon_vqrshiftnsu:
3795 VShiftOpc = ARMISD::VQRSHRNsu; break;
3798 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3799 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3802 case Intrinsic::arm_neon_vshiftins: {
3803 EVT VT = N->getOperand(1).getValueType();
3805 unsigned VShiftOpc = 0;
3807 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3808 VShiftOpc = ARMISD::VSLI;
3809 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3810 VShiftOpc = ARMISD::VSRI;
3812 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3815 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3816 N->getOperand(1), N->getOperand(2),
3817 DAG.getConstant(Cnt, MVT::i32));
3820 case Intrinsic::arm_neon_vqrshifts:
3821 case Intrinsic::arm_neon_vqrshiftu:
3822 // No immediate versions of these to check for.
3829 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3830 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3831 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3832 /// vector element shift counts are generally not legal, and it is hard to see
3833 /// their values after they get legalized to loads from a constant pool.
3834 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3835 const ARMSubtarget *ST) {
3836 EVT VT = N->getValueType(0);
3838 // Nothing to be done for scalar shifts.
3839 if (! VT.isVector())
3842 assert(ST->hasNEON() && "unexpected vector shift");
3845 switch (N->getOpcode()) {
3846 default: llvm_unreachable("unexpected shift opcode");
3849 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3850 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3851 DAG.getConstant(Cnt, MVT::i32));
3856 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3857 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3858 ARMISD::VSHRs : ARMISD::VSHRu);
3859 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3860 DAG.getConstant(Cnt, MVT::i32));
3866 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3867 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3868 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3869 const ARMSubtarget *ST) {
3870 SDValue N0 = N->getOperand(0);
3872 // Check for sign- and zero-extensions of vector extract operations of 8-
3873 // and 16-bit vector elements. NEON supports these directly. They are
3874 // handled during DAG combining because type legalization will promote them
3875 // to 32-bit types and it is messy to recognize the operations after that.
3876 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3877 SDValue Vec = N0.getOperand(0);
3878 SDValue Lane = N0.getOperand(1);
3879 EVT VT = N->getValueType(0);
3880 EVT EltVT = N0.getValueType();
3881 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3883 if (VT == MVT::i32 &&
3884 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3885 TLI.isTypeLegal(Vec.getValueType())) {
3888 switch (N->getOpcode()) {
3889 default: llvm_unreachable("unexpected opcode");
3890 case ISD::SIGN_EXTEND:
3891 Opc = ARMISD::VGETLANEs;
3893 case ISD::ZERO_EXTEND:
3894 case ISD::ANY_EXTEND:
3895 Opc = ARMISD::VGETLANEu;
3898 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3905 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3906 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3907 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3908 const ARMSubtarget *ST) {
3909 // If the target supports NEON, try to use vmax/vmin instructions for f32
3910 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3911 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3912 // a NaN; only do the transformation when it matches that behavior.
3914 // For now only do this when using NEON for FP operations; if using VFP, it
3915 // is not obvious that the benefit outweighs the cost of switching to the
3917 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3918 N->getValueType(0) != MVT::f32)
3921 SDValue CondLHS = N->getOperand(0);
3922 SDValue CondRHS = N->getOperand(1);
3923 SDValue LHS = N->getOperand(2);
3924 SDValue RHS = N->getOperand(3);
3925 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3927 unsigned Opcode = 0;
3929 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3930 IsReversed = false; // x CC y ? x : y
3931 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3932 IsReversed = true ; // x CC y ? y : x
3946 // If LHS is NaN, an ordered comparison will be false and the result will
3947 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3948 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3949 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3950 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3952 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3953 // will return -0, so vmin can only be used for unsafe math or if one of
3954 // the operands is known to be nonzero.
3955 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3957 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3959 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3968 // If LHS is NaN, an ordered comparison will be false and the result will
3969 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3970 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3971 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3972 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3974 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3975 // will return +0, so vmax can only be used for unsafe math or if one of
3976 // the operands is known to be nonzero.
3977 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3979 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3981 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3987 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3990 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3991 DAGCombinerInfo &DCI) const {
3992 switch (N->getOpcode()) {
3994 case ISD::ADD: return PerformADDCombine(N, DCI);
3995 case ISD::SUB: return PerformSUBCombine(N, DCI);
3996 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3997 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4000 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4001 case ISD::SIGN_EXTEND:
4002 case ISD::ZERO_EXTEND:
4003 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4004 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4009 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4010 if (!Subtarget->hasV6Ops())
4011 // Pre-v6 does not support unaligned mem access.
4014 // v6+ may or may not support unaligned mem access depending on the system
4016 // FIXME: This is pretty conservative. Should we provide cmdline option to
4017 // control the behaviour?
4018 if (!Subtarget->isTargetDarwin())
4022 switch (VT.getSimpleVT().SimpleTy) {
4029 // FIXME: VLD1 etc with standard alignment is legal.
4033 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4038 switch (VT.getSimpleVT().SimpleTy) {
4039 default: return false;
4054 if ((V & (Scale - 1)) != 0)
4057 return V == (V & ((1LL << 5) - 1));
4060 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4061 const ARMSubtarget *Subtarget) {
4068 switch (VT.getSimpleVT().SimpleTy) {
4069 default: return false;
4074 // + imm12 or - imm8
4076 return V == (V & ((1LL << 8) - 1));
4077 return V == (V & ((1LL << 12) - 1));
4080 // Same as ARM mode. FIXME: NEON?
4081 if (!Subtarget->hasVFP2())
4086 return V == (V & ((1LL << 8) - 1));
4090 /// isLegalAddressImmediate - Return true if the integer value can be used
4091 /// as the offset of the target addressing mode for load / store of the
4093 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4094 const ARMSubtarget *Subtarget) {
4101 if (Subtarget->isThumb1Only())
4102 return isLegalT1AddressImmediate(V, VT);
4103 else if (Subtarget->isThumb2())
4104 return isLegalT2AddressImmediate(V, VT, Subtarget);
4109 switch (VT.getSimpleVT().SimpleTy) {
4110 default: return false;
4115 return V == (V & ((1LL << 12) - 1));
4118 return V == (V & ((1LL << 8) - 1));
4121 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4126 return V == (V & ((1LL << 8) - 1));
4130 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4132 int Scale = AM.Scale;
4136 switch (VT.getSimpleVT().SimpleTy) {
4137 default: return false;
4146 return Scale == 2 || Scale == 4 || Scale == 8;
4149 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4153 // Note, we allow "void" uses (basically, uses that aren't loads or
4154 // stores), because arm allows folding a scale into many arithmetic
4155 // operations. This should be made more precise and revisited later.
4157 // Allow r << imm, but the imm has to be a multiple of two.
4158 if (Scale & 1) return false;
4159 return isPowerOf2_32(Scale);
4163 /// isLegalAddressingMode - Return true if the addressing mode represented
4164 /// by AM is legal for this target, for a load/store of the specified type.
4165 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4166 const Type *Ty) const {
4167 EVT VT = getValueType(Ty, true);
4168 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4171 // Can never fold addr of global into load/store.
4176 case 0: // no scale reg, must be "r+i" or "r", or "i".
4179 if (Subtarget->isThumb1Only())
4183 // ARM doesn't support any R+R*scale+imm addr modes.
4190 if (Subtarget->isThumb2())
4191 return isLegalT2ScaledAddressingMode(AM, VT);
4193 int Scale = AM.Scale;
4194 switch (VT.getSimpleVT().SimpleTy) {
4195 default: return false;
4199 if (Scale < 0) Scale = -Scale;
4203 return isPowerOf2_32(Scale & ~1);
4207 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4212 // Note, we allow "void" uses (basically, uses that aren't loads or
4213 // stores), because arm allows folding a scale into many arithmetic
4214 // operations. This should be made more precise and revisited later.
4216 // Allow r << imm, but the imm has to be a multiple of two.
4217 if (Scale & 1) return false;
4218 return isPowerOf2_32(Scale);
4225 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4226 /// icmp immediate, that is the target has icmp instructions which can compare
4227 /// a register against the immediate without having to materialize the
4228 /// immediate into a register.
4229 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4230 if (!Subtarget->isThumb())
4231 return ARM_AM::getSOImmVal(Imm) != -1;
4232 if (Subtarget->isThumb2())
4233 return ARM_AM::getT2SOImmVal(Imm) != -1;
4234 return Imm >= 0 && Imm <= 255;
4237 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4238 bool isSEXTLoad, SDValue &Base,
4239 SDValue &Offset, bool &isInc,
4240 SelectionDAG &DAG) {
4241 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4244 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4246 Base = Ptr->getOperand(0);
4247 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4248 int RHSC = (int)RHS->getZExtValue();
4249 if (RHSC < 0 && RHSC > -256) {
4250 assert(Ptr->getOpcode() == ISD::ADD);
4252 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4256 isInc = (Ptr->getOpcode() == ISD::ADD);
4257 Offset = Ptr->getOperand(1);
4259 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4261 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4262 int RHSC = (int)RHS->getZExtValue();
4263 if (RHSC < 0 && RHSC > -0x1000) {
4264 assert(Ptr->getOpcode() == ISD::ADD);
4266 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4267 Base = Ptr->getOperand(0);
4272 if (Ptr->getOpcode() == ISD::ADD) {
4274 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4275 if (ShOpcVal != ARM_AM::no_shift) {
4276 Base = Ptr->getOperand(1);
4277 Offset = Ptr->getOperand(0);
4279 Base = Ptr->getOperand(0);
4280 Offset = Ptr->getOperand(1);
4285 isInc = (Ptr->getOpcode() == ISD::ADD);
4286 Base = Ptr->getOperand(0);
4287 Offset = Ptr->getOperand(1);
4291 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4295 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4296 bool isSEXTLoad, SDValue &Base,
4297 SDValue &Offset, bool &isInc,
4298 SelectionDAG &DAG) {
4299 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4302 Base = Ptr->getOperand(0);
4303 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4304 int RHSC = (int)RHS->getZExtValue();
4305 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4306 assert(Ptr->getOpcode() == ISD::ADD);
4308 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4310 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4311 isInc = Ptr->getOpcode() == ISD::ADD;
4312 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4320 /// getPreIndexedAddressParts - returns true by value, base pointer and
4321 /// offset pointer and addressing mode by reference if the node's address
4322 /// can be legally represented as pre-indexed load / store address.
4324 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4326 ISD::MemIndexedMode &AM,
4327 SelectionDAG &DAG) const {
4328 if (Subtarget->isThumb1Only())
4333 bool isSEXTLoad = false;
4334 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4335 Ptr = LD->getBasePtr();
4336 VT = LD->getMemoryVT();
4337 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4338 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4339 Ptr = ST->getBasePtr();
4340 VT = ST->getMemoryVT();
4345 bool isLegal = false;
4346 if (Subtarget->isThumb2())
4347 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4348 Offset, isInc, DAG);
4350 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4351 Offset, isInc, DAG);
4355 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4359 /// getPostIndexedAddressParts - returns true by value, base pointer and
4360 /// offset pointer and addressing mode by reference if this node can be
4361 /// combined with a load / store to form a post-indexed load / store.
4362 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4365 ISD::MemIndexedMode &AM,
4366 SelectionDAG &DAG) const {
4367 if (Subtarget->isThumb1Only())
4372 bool isSEXTLoad = false;
4373 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4374 VT = LD->getMemoryVT();
4375 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4376 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4377 VT = ST->getMemoryVT();
4382 bool isLegal = false;
4383 if (Subtarget->isThumb2())
4384 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4387 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4392 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4396 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4400 const SelectionDAG &DAG,
4401 unsigned Depth) const {
4402 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4403 switch (Op.getOpcode()) {
4405 case ARMISD::CMOV: {
4406 // Bits are known zero/one if known on the LHS and RHS.
4407 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4408 if (KnownZero == 0 && KnownOne == 0) return;
4410 APInt KnownZeroRHS, KnownOneRHS;
4411 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4412 KnownZeroRHS, KnownOneRHS, Depth+1);
4413 KnownZero &= KnownZeroRHS;
4414 KnownOne &= KnownOneRHS;
4420 //===----------------------------------------------------------------------===//
4421 // ARM Inline Assembly Support
4422 //===----------------------------------------------------------------------===//
4424 /// getConstraintType - Given a constraint letter, return the type of
4425 /// constraint it is for this target.
4426 ARMTargetLowering::ConstraintType
4427 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4428 if (Constraint.size() == 1) {
4429 switch (Constraint[0]) {
4431 case 'l': return C_RegisterClass;
4432 case 'w': return C_RegisterClass;
4435 return TargetLowering::getConstraintType(Constraint);
4438 std::pair<unsigned, const TargetRegisterClass*>
4439 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4441 if (Constraint.size() == 1) {
4442 // GCC ARM Constraint Letters
4443 switch (Constraint[0]) {
4445 if (Subtarget->isThumb())
4446 return std::make_pair(0U, ARM::tGPRRegisterClass);
4448 return std::make_pair(0U, ARM::GPRRegisterClass);
4450 return std::make_pair(0U, ARM::GPRRegisterClass);
4453 return std::make_pair(0U, ARM::SPRRegisterClass);
4454 if (VT.getSizeInBits() == 64)
4455 return std::make_pair(0U, ARM::DPRRegisterClass);
4456 if (VT.getSizeInBits() == 128)
4457 return std::make_pair(0U, ARM::QPRRegisterClass);
4461 if (StringRef("{cc}").equals_lower(Constraint))
4462 return std::make_pair(0U, ARM::CCRRegisterClass);
4464 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4467 std::vector<unsigned> ARMTargetLowering::
4468 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4470 if (Constraint.size() != 1)
4471 return std::vector<unsigned>();
4473 switch (Constraint[0]) { // GCC ARM Constraint Letters
4476 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4477 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4480 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4481 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4482 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4483 ARM::R12, ARM::LR, 0);
4486 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4487 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4488 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4489 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4490 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4491 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4492 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4493 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4494 if (VT.getSizeInBits() == 64)
4495 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4496 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4497 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4498 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4499 if (VT.getSizeInBits() == 128)
4500 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4501 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4505 return std::vector<unsigned>();
4508 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4509 /// vector. If it is invalid, don't add anything to Ops.
4510 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4513 std::vector<SDValue>&Ops,
4514 SelectionDAG &DAG) const {
4515 SDValue Result(0, 0);
4517 switch (Constraint) {
4519 case 'I': case 'J': case 'K': case 'L':
4520 case 'M': case 'N': case 'O':
4521 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4525 int64_t CVal64 = C->getSExtValue();
4526 int CVal = (int) CVal64;
4527 // None of these constraints allow values larger than 32 bits. Check
4528 // that the value fits in an int.
4532 switch (Constraint) {
4534 if (Subtarget->isThumb1Only()) {
4535 // This must be a constant between 0 and 255, for ADD
4537 if (CVal >= 0 && CVal <= 255)
4539 } else if (Subtarget->isThumb2()) {
4540 // A constant that can be used as an immediate value in a
4541 // data-processing instruction.
4542 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4545 // A constant that can be used as an immediate value in a
4546 // data-processing instruction.
4547 if (ARM_AM::getSOImmVal(CVal) != -1)
4553 if (Subtarget->isThumb()) { // FIXME thumb2
4554 // This must be a constant between -255 and -1, for negated ADD
4555 // immediates. This can be used in GCC with an "n" modifier that
4556 // prints the negated value, for use with SUB instructions. It is
4557 // not useful otherwise but is implemented for compatibility.
4558 if (CVal >= -255 && CVal <= -1)
4561 // This must be a constant between -4095 and 4095. It is not clear
4562 // what this constraint is intended for. Implemented for
4563 // compatibility with GCC.
4564 if (CVal >= -4095 && CVal <= 4095)
4570 if (Subtarget->isThumb1Only()) {
4571 // A 32-bit value where only one byte has a nonzero value. Exclude
4572 // zero to match GCC. This constraint is used by GCC internally for
4573 // constants that can be loaded with a move/shift combination.
4574 // It is not useful otherwise but is implemented for compatibility.
4575 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4577 } else if (Subtarget->isThumb2()) {
4578 // A constant whose bitwise inverse can be used as an immediate
4579 // value in a data-processing instruction. This can be used in GCC
4580 // with a "B" modifier that prints the inverted value, for use with
4581 // BIC and MVN instructions. It is not useful otherwise but is
4582 // implemented for compatibility.
4583 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4586 // A constant whose bitwise inverse can be used as an immediate
4587 // value in a data-processing instruction. This can be used in GCC
4588 // with a "B" modifier that prints the inverted value, for use with
4589 // BIC and MVN instructions. It is not useful otherwise but is
4590 // implemented for compatibility.
4591 if (ARM_AM::getSOImmVal(~CVal) != -1)
4597 if (Subtarget->isThumb1Only()) {
4598 // This must be a constant between -7 and 7,
4599 // for 3-operand ADD/SUB immediate instructions.
4600 if (CVal >= -7 && CVal < 7)
4602 } else if (Subtarget->isThumb2()) {
4603 // A constant whose negation can be used as an immediate value in a
4604 // data-processing instruction. This can be used in GCC with an "n"
4605 // modifier that prints the negated value, for use with SUB
4606 // instructions. It is not useful otherwise but is implemented for
4608 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4611 // A constant whose negation can be used as an immediate value in a
4612 // data-processing instruction. This can be used in GCC with an "n"
4613 // modifier that prints the negated value, for use with SUB
4614 // instructions. It is not useful otherwise but is implemented for
4616 if (ARM_AM::getSOImmVal(-CVal) != -1)
4622 if (Subtarget->isThumb()) { // FIXME thumb2
4623 // This must be a multiple of 4 between 0 and 1020, for
4624 // ADD sp + immediate.
4625 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4628 // A power of two or a constant between 0 and 32. This is used in
4629 // GCC for the shift amount on shifted register operands, but it is
4630 // useful in general for any shift amounts.
4631 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4637 if (Subtarget->isThumb()) { // FIXME thumb2
4638 // This must be a constant between 0 and 31, for shift amounts.
4639 if (CVal >= 0 && CVal <= 31)
4645 if (Subtarget->isThumb()) { // FIXME thumb2
4646 // This must be a multiple of 4 between -508 and 508, for
4647 // ADD/SUB sp = sp + immediate.
4648 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4653 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4657 if (Result.getNode()) {
4658 Ops.push_back(Result);
4661 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4666 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4667 // The ARM target isn't yet aware of offsets.
4671 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4672 APInt Imm = FPImm.bitcastToAPInt();
4673 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4674 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4675 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4677 // We can handle 4 bits of mantissa.
4678 // mantissa = (16+UInt(e:f:g:h))/16.
4679 if (Mantissa & 0x7ffff)
4682 if ((Mantissa & 0xf) != Mantissa)
4685 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4686 if (Exp < -3 || Exp > 4)
4688 Exp = ((Exp+3) & 0x7) ^ 4;
4690 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4693 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4694 APInt Imm = FPImm.bitcastToAPInt();
4695 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4696 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4697 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4699 // We can handle 4 bits of mantissa.
4700 // mantissa = (16+UInt(e:f:g:h))/16.
4701 if (Mantissa & 0xffffffffffffLL)
4704 if ((Mantissa & 0xf) != Mantissa)
4707 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4708 if (Exp < -3 || Exp > 4)
4710 Exp = ((Exp+3) & 0x7) ^ 4;
4712 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4715 /// isFPImmLegal - Returns true if the target can instruction select the
4716 /// specified FP immediate natively. If false, the legalizer will
4717 /// materialize the FP immediate as a load from a constant pool.
4718 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4719 if (!Subtarget->hasVFP3())
4722 return ARM::getVFPf32Imm(Imm) != -1;
4724 return ARM::getVFPf64Imm(Imm) != -1;