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/Target/TargetOptions.h"
41 #include "llvm/ADT/VectorExtras.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/MathExtras.h"
47 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
48 CCValAssign::LocInfo &LocInfo,
49 ISD::ArgFlagsTy &ArgFlags,
51 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
52 CCValAssign::LocInfo &LocInfo,
53 ISD::ArgFlagsTy &ArgFlags,
55 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
56 CCValAssign::LocInfo &LocInfo,
57 ISD::ArgFlagsTy &ArgFlags,
59 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
60 CCValAssign::LocInfo &LocInfo,
61 ISD::ArgFlagsTy &ArgFlags,
64 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
65 EVT PromotedBitwiseVT) {
66 if (VT != PromotedLdStVT) {
67 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
68 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
69 PromotedLdStVT.getSimpleVT());
71 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
72 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
73 PromotedLdStVT.getSimpleVT());
76 EVT ElemTy = VT.getVectorElementType();
77 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
78 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
79 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
80 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
81 if (ElemTy != MVT::i32) {
82 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
83 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
84 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
85 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
87 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
88 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
89 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
90 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
93 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
94 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
97 // Promote all bit-wise operations.
98 if (VT.isInteger() && VT != PromotedBitwiseVT) {
99 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
100 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
101 PromotedBitwiseVT.getSimpleVT());
102 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
103 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
104 PromotedBitwiseVT.getSimpleVT());
105 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
106 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
107 PromotedBitwiseVT.getSimpleVT());
110 // Neon does not support vector divide/remainder operations.
111 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
112 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
119 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
120 addRegisterClass(VT, ARM::DPRRegisterClass);
121 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
124 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
125 addRegisterClass(VT, ARM::QPRRegisterClass);
126 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
129 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
130 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
131 return new TargetLoweringObjectFileMachO();
132 return new ARMElfTargetObjectFile();
135 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
136 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
137 Subtarget = &TM.getSubtarget<ARMSubtarget>();
139 if (Subtarget->isTargetDarwin()) {
140 // Uses VFP for Thumb libfuncs if available.
141 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
142 // Single-precision floating-point arithmetic.
143 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
144 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
145 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
146 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
148 // Double-precision floating-point arithmetic.
149 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
150 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
151 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
152 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
154 // Single-precision comparisons.
155 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
156 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
157 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
158 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
159 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
160 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
161 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
162 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
164 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
165 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
166 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
167 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
173 // Double-precision comparisons.
174 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
175 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
176 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
177 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
178 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
179 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
180 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
181 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
183 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
192 // Floating-point to integer conversions.
193 // i64 conversions are done via library routines even when generating VFP
194 // instructions, so use the same ones.
195 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
196 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
197 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
198 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
200 // Conversions between floating types.
201 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
202 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
204 // Integer to floating-point conversions.
205 // i64 conversions are done via library routines even when generating VFP
206 // instructions, so use the same ones.
207 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
208 // e.g., __floatunsidf vs. __floatunssidfvfp.
209 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
210 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
211 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
212 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
216 // These libcalls are not available in 32-bit.
217 setLibcallName(RTLIB::SHL_I128, 0);
218 setLibcallName(RTLIB::SRL_I128, 0);
219 setLibcallName(RTLIB::SRA_I128, 0);
221 // Libcalls should use the AAPCS base standard ABI, even if hard float
222 // is in effect, as per the ARM RTABI specification, section 4.1.2.
223 if (Subtarget->isAAPCS_ABI()) {
224 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
225 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
226 CallingConv::ARM_AAPCS);
230 if (Subtarget->isThumb1Only())
231 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
233 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
234 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
235 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
236 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
238 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
241 if (Subtarget->hasNEON()) {
242 addDRTypeForNEON(MVT::v2f32);
243 addDRTypeForNEON(MVT::v8i8);
244 addDRTypeForNEON(MVT::v4i16);
245 addDRTypeForNEON(MVT::v2i32);
246 addDRTypeForNEON(MVT::v1i64);
248 addQRTypeForNEON(MVT::v4f32);
249 addQRTypeForNEON(MVT::v2f64);
250 addQRTypeForNEON(MVT::v16i8);
251 addQRTypeForNEON(MVT::v8i16);
252 addQRTypeForNEON(MVT::v4i32);
253 addQRTypeForNEON(MVT::v2i64);
255 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
256 // neither Neon nor VFP support any arithmetic operations on it.
257 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
258 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
259 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
260 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
261 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
262 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
263 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
264 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
265 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
266 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
267 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
268 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
269 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
270 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
271 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
272 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
273 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
274 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
275 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
276 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
277 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
278 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
279 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
280 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
282 // Neon does not support some operations on v1i64 and v2i64 types.
283 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
284 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
285 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
286 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
288 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
289 setTargetDAGCombine(ISD::SHL);
290 setTargetDAGCombine(ISD::SRL);
291 setTargetDAGCombine(ISD::SRA);
292 setTargetDAGCombine(ISD::SIGN_EXTEND);
293 setTargetDAGCombine(ISD::ZERO_EXTEND);
294 setTargetDAGCombine(ISD::ANY_EXTEND);
297 computeRegisterProperties();
299 // ARM does not have f32 extending load.
300 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
302 // ARM does not have i1 sign extending load.
303 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
305 // ARM supports all 4 flavors of integer indexed load / store.
306 if (!Subtarget->isThumb1Only()) {
307 for (unsigned im = (unsigned)ISD::PRE_INC;
308 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
309 setIndexedLoadAction(im, MVT::i1, Legal);
310 setIndexedLoadAction(im, MVT::i8, Legal);
311 setIndexedLoadAction(im, MVT::i16, Legal);
312 setIndexedLoadAction(im, MVT::i32, Legal);
313 setIndexedStoreAction(im, MVT::i1, Legal);
314 setIndexedStoreAction(im, MVT::i8, Legal);
315 setIndexedStoreAction(im, MVT::i16, Legal);
316 setIndexedStoreAction(im, MVT::i32, Legal);
320 // i64 operation support.
321 if (Subtarget->isThumb1Only()) {
322 setOperationAction(ISD::MUL, MVT::i64, Expand);
323 setOperationAction(ISD::MULHU, MVT::i32, Expand);
324 setOperationAction(ISD::MULHS, MVT::i32, Expand);
325 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
326 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
328 setOperationAction(ISD::MUL, MVT::i64, Expand);
329 setOperationAction(ISD::MULHU, MVT::i32, Expand);
330 if (!Subtarget->hasV6Ops())
331 setOperationAction(ISD::MULHS, MVT::i32, Expand);
333 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
334 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
335 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
336 setOperationAction(ISD::SRL, MVT::i64, Custom);
337 setOperationAction(ISD::SRA, MVT::i64, Custom);
339 // ARM does not have ROTL.
340 setOperationAction(ISD::ROTL, MVT::i32, Expand);
341 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
342 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
343 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
344 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
346 // Only ARMv6 has BSWAP.
347 if (!Subtarget->hasV6Ops())
348 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
350 // These are expanded into libcalls.
351 setOperationAction(ISD::SDIV, MVT::i32, Expand);
352 setOperationAction(ISD::UDIV, MVT::i32, Expand);
353 setOperationAction(ISD::SREM, MVT::i32, Expand);
354 setOperationAction(ISD::UREM, MVT::i32, Expand);
355 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
356 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
358 // Support label based line numbers.
359 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
360 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
362 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
363 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
364 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
365 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
366 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
368 // Use the default implementation.
369 setOperationAction(ISD::VASTART, MVT::Other, Custom);
370 setOperationAction(ISD::VAARG, MVT::Other, Expand);
371 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
372 setOperationAction(ISD::VAEND, MVT::Other, Expand);
373 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
374 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
375 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
376 // FIXME: Shouldn't need this, since no register is used, but the legalizer
377 // doesn't yet know how to not do that for SjLj.
378 setExceptionSelectorRegister(ARM::R0);
379 if (Subtarget->isThumb())
380 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
382 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
383 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
385 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
386 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
387 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
389 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
391 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
392 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
393 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
395 // We want to custom lower some of our intrinsics.
396 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
398 setOperationAction(ISD::SETCC, MVT::i32, Expand);
399 setOperationAction(ISD::SETCC, MVT::f32, Expand);
400 setOperationAction(ISD::SETCC, MVT::f64, Expand);
401 setOperationAction(ISD::SELECT, MVT::i32, Expand);
402 setOperationAction(ISD::SELECT, MVT::f32, Expand);
403 setOperationAction(ISD::SELECT, MVT::f64, Expand);
404 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
405 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
406 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
408 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
409 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
410 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
411 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
412 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
414 // We don't support sin/cos/fmod/copysign/pow
415 setOperationAction(ISD::FSIN, MVT::f64, Expand);
416 setOperationAction(ISD::FSIN, MVT::f32, Expand);
417 setOperationAction(ISD::FCOS, MVT::f32, Expand);
418 setOperationAction(ISD::FCOS, MVT::f64, Expand);
419 setOperationAction(ISD::FREM, MVT::f64, Expand);
420 setOperationAction(ISD::FREM, MVT::f32, Expand);
421 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
422 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
423 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
425 setOperationAction(ISD::FPOW, MVT::f64, Expand);
426 setOperationAction(ISD::FPOW, MVT::f32, Expand);
428 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
429 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
430 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
431 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
432 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
433 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
436 // We have target-specific dag combine patterns for the following nodes:
437 // ARMISD::FMRRD - No need to call setTargetDAGCombine
438 setTargetDAGCombine(ISD::ADD);
439 setTargetDAGCombine(ISD::SUB);
441 setStackPointerRegisterToSaveRestore(ARM::SP);
442 setSchedulingPreference(SchedulingForRegPressure);
444 // FIXME: If-converter should use instruction latency to determine
445 // profitability rather than relying on fixed limits.
446 if (Subtarget->getCPUString() == "generic") {
447 // Generic (and overly aggressive) if-conversion limits.
448 setIfCvtBlockSizeLimit(10);
449 setIfCvtDupBlockSizeLimit(2);
450 } else if (Subtarget->hasV6Ops()) {
451 setIfCvtBlockSizeLimit(2);
452 setIfCvtDupBlockSizeLimit(1);
454 setIfCvtBlockSizeLimit(3);
455 setIfCvtDupBlockSizeLimit(2);
458 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
459 // Do not enable CodePlacementOpt for now: it currently runs after the
460 // ARMConstantIslandPass and messes up branch relaxation and placement
461 // of constant islands.
462 // benefitFromCodePlacementOpt = true;
465 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
468 case ARMISD::Wrapper: return "ARMISD::Wrapper";
469 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
470 case ARMISD::CALL: return "ARMISD::CALL";
471 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
472 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
473 case ARMISD::tCALL: return "ARMISD::tCALL";
474 case ARMISD::BRCOND: return "ARMISD::BRCOND";
475 case ARMISD::BR_JT: return "ARMISD::BR_JT";
476 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
477 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
478 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
479 case ARMISD::CMP: return "ARMISD::CMP";
480 case ARMISD::CMPZ: return "ARMISD::CMPZ";
481 case ARMISD::CMPFP: return "ARMISD::CMPFP";
482 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
483 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
484 case ARMISD::CMOV: return "ARMISD::CMOV";
485 case ARMISD::CNEG: return "ARMISD::CNEG";
487 case ARMISD::FTOSI: return "ARMISD::FTOSI";
488 case ARMISD::FTOUI: return "ARMISD::FTOUI";
489 case ARMISD::SITOF: return "ARMISD::SITOF";
490 case ARMISD::UITOF: return "ARMISD::UITOF";
492 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
493 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
494 case ARMISD::RRX: return "ARMISD::RRX";
496 case ARMISD::FMRRD: return "ARMISD::FMRRD";
497 case ARMISD::FMDRR: return "ARMISD::FMDRR";
499 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
500 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
502 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
504 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
506 case ARMISD::VCEQ: return "ARMISD::VCEQ";
507 case ARMISD::VCGE: return "ARMISD::VCGE";
508 case ARMISD::VCGEU: return "ARMISD::VCGEU";
509 case ARMISD::VCGT: return "ARMISD::VCGT";
510 case ARMISD::VCGTU: return "ARMISD::VCGTU";
511 case ARMISD::VTST: return "ARMISD::VTST";
513 case ARMISD::VSHL: return "ARMISD::VSHL";
514 case ARMISD::VSHRs: return "ARMISD::VSHRs";
515 case ARMISD::VSHRu: return "ARMISD::VSHRu";
516 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
517 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
518 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
519 case ARMISD::VSHRN: return "ARMISD::VSHRN";
520 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
521 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
522 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
523 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
524 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
525 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
526 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
527 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
528 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
529 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
530 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
531 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
532 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
533 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
534 case ARMISD::VDUP: return "ARMISD::VDUP";
535 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
536 case ARMISD::VEXT: return "ARMISD::VEXT";
537 case ARMISD::VREV64: return "ARMISD::VREV64";
538 case ARMISD::VREV32: return "ARMISD::VREV32";
539 case ARMISD::VREV16: return "ARMISD::VREV16";
540 case ARMISD::VZIP: return "ARMISD::VZIP";
541 case ARMISD::VUZP: return "ARMISD::VUZP";
542 case ARMISD::VTRN: return "ARMISD::VTRN";
546 /// getFunctionAlignment - Return the Log2 alignment of this function.
547 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
548 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
551 //===----------------------------------------------------------------------===//
553 //===----------------------------------------------------------------------===//
555 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
556 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
558 default: llvm_unreachable("Unknown condition code!");
559 case ISD::SETNE: return ARMCC::NE;
560 case ISD::SETEQ: return ARMCC::EQ;
561 case ISD::SETGT: return ARMCC::GT;
562 case ISD::SETGE: return ARMCC::GE;
563 case ISD::SETLT: return ARMCC::LT;
564 case ISD::SETLE: return ARMCC::LE;
565 case ISD::SETUGT: return ARMCC::HI;
566 case ISD::SETUGE: return ARMCC::HS;
567 case ISD::SETULT: return ARMCC::LO;
568 case ISD::SETULE: return ARMCC::LS;
572 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
573 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
574 ARMCC::CondCodes &CondCode2) {
575 CondCode2 = ARMCC::AL;
577 default: llvm_unreachable("Unknown FP condition!");
579 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
581 case ISD::SETOGT: CondCode = ARMCC::GT; break;
583 case ISD::SETOGE: CondCode = ARMCC::GE; break;
584 case ISD::SETOLT: CondCode = ARMCC::MI; break;
585 case ISD::SETOLE: CondCode = ARMCC::LS; break;
586 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
587 case ISD::SETO: CondCode = ARMCC::VC; break;
588 case ISD::SETUO: CondCode = ARMCC::VS; break;
589 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
590 case ISD::SETUGT: CondCode = ARMCC::HI; break;
591 case ISD::SETUGE: CondCode = ARMCC::PL; break;
593 case ISD::SETULT: CondCode = ARMCC::LT; break;
595 case ISD::SETULE: CondCode = ARMCC::LE; break;
597 case ISD::SETUNE: CondCode = ARMCC::NE; break;
601 //===----------------------------------------------------------------------===//
602 // Calling Convention Implementation
603 //===----------------------------------------------------------------------===//
605 #include "ARMGenCallingConv.inc"
607 // APCS f64 is in register pairs, possibly split to stack
608 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
609 CCValAssign::LocInfo &LocInfo,
610 CCState &State, bool CanFail) {
611 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
613 // Try to get the first register.
614 if (unsigned Reg = State.AllocateReg(RegList, 4))
615 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
617 // For the 2nd half of a v2f64, do not fail.
621 // Put the whole thing on the stack.
622 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
623 State.AllocateStack(8, 4),
628 // Try to get the second register.
629 if (unsigned Reg = State.AllocateReg(RegList, 4))
630 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
632 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
633 State.AllocateStack(4, 4),
638 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
639 CCValAssign::LocInfo &LocInfo,
640 ISD::ArgFlagsTy &ArgFlags,
642 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
644 if (LocVT == MVT::v2f64 &&
645 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
647 return true; // we handled it
650 // AAPCS f64 is in aligned register pairs
651 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
652 CCValAssign::LocInfo &LocInfo,
653 CCState &State, bool CanFail) {
654 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
655 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
657 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
659 // For the 2nd half of a v2f64, do not just fail.
663 // Put the whole thing on the stack.
664 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
665 State.AllocateStack(8, 8),
671 for (i = 0; i < 2; ++i)
672 if (HiRegList[i] == Reg)
675 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
676 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
681 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
682 CCValAssign::LocInfo &LocInfo,
683 ISD::ArgFlagsTy &ArgFlags,
685 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
687 if (LocVT == MVT::v2f64 &&
688 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
690 return true; // we handled it
693 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
694 CCValAssign::LocInfo &LocInfo, CCState &State) {
695 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
696 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
698 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
700 return false; // we didn't handle it
703 for (i = 0; i < 2; ++i)
704 if (HiRegList[i] == Reg)
707 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
708 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
713 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
714 CCValAssign::LocInfo &LocInfo,
715 ISD::ArgFlagsTy &ArgFlags,
717 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
719 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
721 return true; // we handled it
724 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
725 CCValAssign::LocInfo &LocInfo,
726 ISD::ArgFlagsTy &ArgFlags,
728 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
732 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
733 /// given CallingConvention value.
734 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
736 bool isVarArg) const {
739 llvm_unreachable("Unsupported calling convention");
741 case CallingConv::Fast:
742 // Use target triple & subtarget features to do actual dispatch.
743 if (Subtarget->isAAPCS_ABI()) {
744 if (Subtarget->hasVFP2() &&
745 FloatABIType == FloatABI::Hard && !isVarArg)
746 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
748 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
750 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
751 case CallingConv::ARM_AAPCS_VFP:
752 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
753 case CallingConv::ARM_AAPCS:
754 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
755 case CallingConv::ARM_APCS:
756 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
760 /// LowerCallResult - Lower the result values of a call into the
761 /// appropriate copies out of appropriate physical registers.
763 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
764 CallingConv::ID CallConv, bool isVarArg,
765 const SmallVectorImpl<ISD::InputArg> &Ins,
766 DebugLoc dl, SelectionDAG &DAG,
767 SmallVectorImpl<SDValue> &InVals) {
769 // Assign locations to each value returned by this call.
770 SmallVector<CCValAssign, 16> RVLocs;
771 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
772 RVLocs, *DAG.getContext());
773 CCInfo.AnalyzeCallResult(Ins,
774 CCAssignFnForNode(CallConv, /* Return*/ true,
777 // Copy all of the result registers out of their specified physreg.
778 for (unsigned i = 0; i != RVLocs.size(); ++i) {
779 CCValAssign VA = RVLocs[i];
782 if (VA.needsCustom()) {
783 // Handle f64 or half of a v2f64.
784 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
786 Chain = Lo.getValue(1);
787 InFlag = Lo.getValue(2);
788 VA = RVLocs[++i]; // skip ahead to next loc
789 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
791 Chain = Hi.getValue(1);
792 InFlag = Hi.getValue(2);
793 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
795 if (VA.getLocVT() == MVT::v2f64) {
796 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
797 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
798 DAG.getConstant(0, MVT::i32));
800 VA = RVLocs[++i]; // skip ahead to next loc
801 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
802 Chain = Lo.getValue(1);
803 InFlag = Lo.getValue(2);
804 VA = RVLocs[++i]; // skip ahead to next loc
805 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
806 Chain = Hi.getValue(1);
807 InFlag = Hi.getValue(2);
808 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
809 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
810 DAG.getConstant(1, MVT::i32));
813 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
815 Chain = Val.getValue(1);
816 InFlag = Val.getValue(2);
819 switch (VA.getLocInfo()) {
820 default: llvm_unreachable("Unknown loc info!");
821 case CCValAssign::Full: break;
822 case CCValAssign::BCvt:
823 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
827 InVals.push_back(Val);
833 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
834 /// by "Src" to address "Dst" of size "Size". Alignment information is
835 /// specified by the specific parameter attribute. The copy will be passed as
836 /// a byval function parameter.
837 /// Sometimes what we are copying is the end of a larger object, the part that
838 /// does not fit in registers.
840 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
841 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
843 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
844 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
845 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
848 /// LowerMemOpCallTo - Store the argument to the stack.
850 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
851 SDValue StackPtr, SDValue Arg,
852 DebugLoc dl, SelectionDAG &DAG,
853 const CCValAssign &VA,
854 ISD::ArgFlagsTy Flags) {
855 unsigned LocMemOffset = VA.getLocMemOffset();
856 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
857 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
858 if (Flags.isByVal()) {
859 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
861 return DAG.getStore(Chain, dl, Arg, PtrOff,
862 PseudoSourceValue::getStack(), LocMemOffset);
865 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
866 SDValue Chain, SDValue &Arg,
867 RegsToPassVector &RegsToPass,
868 CCValAssign &VA, CCValAssign &NextVA,
870 SmallVector<SDValue, 8> &MemOpChains,
871 ISD::ArgFlagsTy Flags) {
873 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
874 DAG.getVTList(MVT::i32, MVT::i32), Arg);
875 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
877 if (NextVA.isRegLoc())
878 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
880 assert(NextVA.isMemLoc());
881 if (StackPtr.getNode() == 0)
882 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
884 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
890 /// LowerCall - Lowering a call into a callseq_start <-
891 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
894 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
895 CallingConv::ID CallConv, bool isVarArg,
897 const SmallVectorImpl<ISD::OutputArg> &Outs,
898 const SmallVectorImpl<ISD::InputArg> &Ins,
899 DebugLoc dl, SelectionDAG &DAG,
900 SmallVectorImpl<SDValue> &InVals) {
902 // Analyze operands of the call, assigning locations to each operand.
903 SmallVector<CCValAssign, 16> ArgLocs;
904 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
906 CCInfo.AnalyzeCallOperands(Outs,
907 CCAssignFnForNode(CallConv, /* Return*/ false,
910 // Get a count of how many bytes are to be pushed on the stack.
911 unsigned NumBytes = CCInfo.getNextStackOffset();
913 // Adjust the stack pointer for the new arguments...
914 // These operations are automatically eliminated by the prolog/epilog pass
915 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
917 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
919 RegsToPassVector RegsToPass;
920 SmallVector<SDValue, 8> MemOpChains;
922 // Walk the register/memloc assignments, inserting copies/loads. In the case
923 // of tail call optimization, arguments are handled later.
924 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
927 CCValAssign &VA = ArgLocs[i];
928 SDValue Arg = Outs[realArgIdx].Val;
929 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
931 // Promote the value if needed.
932 switch (VA.getLocInfo()) {
933 default: llvm_unreachable("Unknown loc info!");
934 case CCValAssign::Full: break;
935 case CCValAssign::SExt:
936 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
938 case CCValAssign::ZExt:
939 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
941 case CCValAssign::AExt:
942 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
944 case CCValAssign::BCvt:
945 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
949 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
950 if (VA.needsCustom()) {
951 if (VA.getLocVT() == MVT::v2f64) {
952 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
953 DAG.getConstant(0, MVT::i32));
954 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
955 DAG.getConstant(1, MVT::i32));
957 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
958 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
960 VA = ArgLocs[++i]; // skip ahead to next loc
962 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
963 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
965 assert(VA.isMemLoc());
966 if (StackPtr.getNode() == 0)
967 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
969 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
970 dl, DAG, VA, Flags));
973 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
974 StackPtr, MemOpChains, Flags);
976 } else if (VA.isRegLoc()) {
977 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
979 assert(VA.isMemLoc());
980 if (StackPtr.getNode() == 0)
981 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
983 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
984 dl, DAG, VA, Flags));
988 if (!MemOpChains.empty())
989 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
990 &MemOpChains[0], MemOpChains.size());
992 // Build a sequence of copy-to-reg nodes chained together with token chain
993 // and flag operands which copy the outgoing args into the appropriate regs.
995 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
996 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
997 RegsToPass[i].second, InFlag);
998 InFlag = Chain.getValue(1);
1001 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1002 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1003 // node so that legalize doesn't hack it.
1004 bool isDirect = false;
1005 bool isARMFunc = false;
1006 bool isLocalARMFunc = false;
1007 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1008 GlobalValue *GV = G->getGlobal();
1010 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1011 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1012 getTargetMachine().getRelocationModel() != Reloc::Static;
1013 isARMFunc = !Subtarget->isThumb() || isStub;
1014 // ARM call to a local ARM function is predicable.
1015 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1016 // tBX takes a register source operand.
1017 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1018 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1021 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1022 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1023 Callee = DAG.getLoad(getPointerTy(), dl,
1024 DAG.getEntryNode(), CPAddr, NULL, 0);
1025 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1026 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1027 getPointerTy(), Callee, PICLabel);
1029 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1030 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1032 bool isStub = Subtarget->isTargetDarwin() &&
1033 getTargetMachine().getRelocationModel() != Reloc::Static;
1034 isARMFunc = !Subtarget->isThumb() || isStub;
1035 // tBX takes a register source operand.
1036 const char *Sym = S->getSymbol();
1037 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1038 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1039 Sym, ARMPCLabelIndex, 4);
1040 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1041 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1042 Callee = DAG.getLoad(getPointerTy(), dl,
1043 DAG.getEntryNode(), CPAddr, NULL, 0);
1044 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1045 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1046 getPointerTy(), Callee, PICLabel);
1048 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1051 // FIXME: handle tail calls differently.
1053 if (Subtarget->isThumb()) {
1054 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1055 CallOpc = ARMISD::CALL_NOLINK;
1057 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1059 CallOpc = (isDirect || Subtarget->hasV5TOps())
1060 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1061 : ARMISD::CALL_NOLINK;
1063 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1064 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1065 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1066 InFlag = Chain.getValue(1);
1069 std::vector<SDValue> Ops;
1070 Ops.push_back(Chain);
1071 Ops.push_back(Callee);
1073 // Add argument registers to the end of the list so that they are known live
1075 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1076 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1077 RegsToPass[i].second.getValueType()));
1079 if (InFlag.getNode())
1080 Ops.push_back(InFlag);
1081 // Returns a chain and a flag for retval copy to use.
1082 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1083 &Ops[0], Ops.size());
1084 InFlag = Chain.getValue(1);
1086 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1087 DAG.getIntPtrConstant(0, true), InFlag);
1089 InFlag = Chain.getValue(1);
1091 // Handle result values, copying them out of physregs into vregs that we
1093 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1098 ARMTargetLowering::LowerReturn(SDValue Chain,
1099 CallingConv::ID CallConv, bool isVarArg,
1100 const SmallVectorImpl<ISD::OutputArg> &Outs,
1101 DebugLoc dl, SelectionDAG &DAG) {
1103 // CCValAssign - represent the assignment of the return value to a location.
1104 SmallVector<CCValAssign, 16> RVLocs;
1106 // CCState - Info about the registers and stack slots.
1107 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1110 // Analyze outgoing return values.
1111 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1114 // If this is the first return lowered for this function, add
1115 // the regs to the liveout set for the function.
1116 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1117 for (unsigned i = 0; i != RVLocs.size(); ++i)
1118 if (RVLocs[i].isRegLoc())
1119 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1124 // Copy the result values into the output registers.
1125 for (unsigned i = 0, realRVLocIdx = 0;
1127 ++i, ++realRVLocIdx) {
1128 CCValAssign &VA = RVLocs[i];
1129 assert(VA.isRegLoc() && "Can only return in registers!");
1131 SDValue Arg = Outs[realRVLocIdx].Val;
1133 switch (VA.getLocInfo()) {
1134 default: llvm_unreachable("Unknown loc info!");
1135 case CCValAssign::Full: break;
1136 case CCValAssign::BCvt:
1137 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1141 if (VA.needsCustom()) {
1142 if (VA.getLocVT() == MVT::v2f64) {
1143 // Extract the first half and return it in two registers.
1144 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1145 DAG.getConstant(0, MVT::i32));
1146 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1147 DAG.getVTList(MVT::i32, MVT::i32), Half);
1149 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1150 Flag = Chain.getValue(1);
1151 VA = RVLocs[++i]; // skip ahead to next loc
1152 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1153 HalfGPRs.getValue(1), Flag);
1154 Flag = Chain.getValue(1);
1155 VA = RVLocs[++i]; // skip ahead to next loc
1157 // Extract the 2nd half and fall through to handle it as an f64 value.
1158 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1159 DAG.getConstant(1, MVT::i32));
1161 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1163 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1164 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1165 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1166 Flag = Chain.getValue(1);
1167 VA = RVLocs[++i]; // skip ahead to next loc
1168 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1171 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1173 // Guarantee that all emitted copies are
1174 // stuck together, avoiding something bad.
1175 Flag = Chain.getValue(1);
1180 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1182 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1187 // ConstantPool, BlockAddress, JumpTable, GlobalAddress, and ExternalSymbol are
1188 // lowered as their target counterpart wrapped in the ARMISD::Wrapper
1189 // node. Suppose N is one of the above mentioned nodes. It has to be wrapped
1190 // because otherwise Select(N) returns N. So the raw TargetGlobalAddress
1191 // nodes, etc. can only be used to form addressing mode. These wrapped nodes
1192 // will be selected into MOVi.
1193 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1194 EVT PtrVT = Op.getValueType();
1195 // FIXME there is no actual debug info here
1196 DebugLoc dl = Op.getDebugLoc();
1197 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1199 if (CP->isMachineConstantPoolEntry())
1200 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1201 CP->getAlignment());
1203 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1204 CP->getAlignment());
1205 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1208 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1209 DebugLoc DL = Op.getDebugLoc();
1210 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1211 SDValue Result = DAG.getBlockAddress(BA, DL, /*isTarget=*/true);
1212 return DAG.getNode(ARMISD::Wrapper, DL, getPointerTy(), Result);
1215 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1217 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1218 SelectionDAG &DAG) {
1219 DebugLoc dl = GA->getDebugLoc();
1220 EVT PtrVT = getPointerTy();
1221 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1222 ARMConstantPoolValue *CPV =
1223 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1224 ARMCP::CPValue, PCAdj, "tlsgd", true);
1225 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1226 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1227 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1228 SDValue Chain = Argument.getValue(1);
1230 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1231 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1233 // call __tls_get_addr.
1236 Entry.Node = Argument;
1237 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1238 Args.push_back(Entry);
1239 // FIXME: is there useful debug info available here?
1240 std::pair<SDValue, SDValue> CallResult =
1241 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1242 false, false, false, false,
1243 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1244 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1245 return CallResult.first;
1248 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1249 // "local exec" model.
1251 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1252 SelectionDAG &DAG) {
1253 GlobalValue *GV = GA->getGlobal();
1254 DebugLoc dl = GA->getDebugLoc();
1256 SDValue Chain = DAG.getEntryNode();
1257 EVT PtrVT = getPointerTy();
1258 // Get the Thread Pointer
1259 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1261 if (GV->isDeclaration()) {
1262 // initial exec model
1263 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1264 ARMConstantPoolValue *CPV =
1265 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1266 ARMCP::CPValue, PCAdj, "gottpoff", true);
1267 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1268 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1269 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1270 Chain = Offset.getValue(1);
1272 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1273 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1275 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1278 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1279 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1280 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1281 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1284 // The address of the thread local variable is the add of the thread
1285 // pointer with the offset of the variable.
1286 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1290 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1291 // TODO: implement the "local dynamic" model
1292 assert(Subtarget->isTargetELF() &&
1293 "TLS not implemented for non-ELF targets");
1294 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1295 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1296 // otherwise use the "Local Exec" TLS Model
1297 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1298 return LowerToTLSGeneralDynamicModel(GA, DAG);
1300 return LowerToTLSExecModels(GA, DAG);
1303 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1304 SelectionDAG &DAG) {
1305 EVT PtrVT = getPointerTy();
1306 DebugLoc dl = Op.getDebugLoc();
1307 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1308 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1309 if (RelocM == Reloc::PIC_) {
1310 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1311 ARMConstantPoolValue *CPV =
1312 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1313 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1314 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1315 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1317 PseudoSourceValue::getConstantPool(), 0);
1318 SDValue Chain = Result.getValue(1);
1319 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1320 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1322 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1323 PseudoSourceValue::getGOT(), 0);
1326 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1327 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1328 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1329 PseudoSourceValue::getConstantPool(), 0);
1333 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1334 SelectionDAG &DAG) {
1335 EVT PtrVT = getPointerTy();
1336 DebugLoc dl = Op.getDebugLoc();
1337 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1338 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1340 if (RelocM == Reloc::Static)
1341 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1343 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1344 ARMConstantPoolValue *CPV =
1345 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1346 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1348 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1350 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1351 SDValue Chain = Result.getValue(1);
1353 if (RelocM == Reloc::PIC_) {
1354 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1355 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1358 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1359 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1364 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1366 assert(Subtarget->isTargetELF() &&
1367 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1368 EVT PtrVT = getPointerTy();
1369 DebugLoc dl = Op.getDebugLoc();
1370 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1371 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1372 "_GLOBAL_OFFSET_TABLE_",
1373 ARMPCLabelIndex, PCAdj);
1374 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1375 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1376 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1377 PseudoSourceValue::getConstantPool(), 0);
1378 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1379 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1383 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1384 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1385 DebugLoc dl = Op.getDebugLoc();
1387 default: return SDValue(); // Don't custom lower most intrinsics.
1388 case Intrinsic::arm_thread_pointer: {
1389 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1390 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1392 case Intrinsic::eh_sjlj_lsda: {
1393 MachineFunction &MF = DAG.getMachineFunction();
1394 EVT PtrVT = getPointerTy();
1395 DebugLoc dl = Op.getDebugLoc();
1396 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1398 unsigned PCAdj = (RelocM != Reloc::PIC_)
1399 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1400 ARMConstantPoolValue *CPV =
1401 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1402 ARMCP::CPLSDA, PCAdj);
1403 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1404 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1406 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1407 SDValue Chain = Result.getValue(1);
1409 if (RelocM == Reloc::PIC_) {
1410 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1411 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1415 case Intrinsic::eh_sjlj_setjmp:
1416 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1420 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1421 unsigned VarArgsFrameIndex) {
1422 // vastart just stores the address of the VarArgsFrameIndex slot into the
1423 // memory location argument.
1424 DebugLoc dl = Op.getDebugLoc();
1425 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1426 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1427 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1428 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1432 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1433 SDNode *Node = Op.getNode();
1434 DebugLoc dl = Node->getDebugLoc();
1435 EVT VT = Node->getValueType(0);
1436 SDValue Chain = Op.getOperand(0);
1437 SDValue Size = Op.getOperand(1);
1438 SDValue Align = Op.getOperand(2);
1440 // Chain the dynamic stack allocation so that it doesn't modify the stack
1441 // pointer when other instructions are using the stack.
1442 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1444 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1445 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1446 if (AlignVal > StackAlign)
1447 // Do this now since selection pass cannot introduce new target
1448 // independent node.
1449 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1451 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1452 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1453 // do even more horrible hack later.
1454 MachineFunction &MF = DAG.getMachineFunction();
1455 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1456 if (AFI->isThumb1OnlyFunction()) {
1458 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1460 uint32_t Val = C->getZExtValue();
1461 if (Val <= 508 && ((Val & 3) == 0))
1465 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1468 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1469 SDValue Ops1[] = { Chain, Size, Align };
1470 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1471 Chain = Res.getValue(1);
1472 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1473 DAG.getIntPtrConstant(0, true), SDValue());
1474 SDValue Ops2[] = { Res, Chain };
1475 return DAG.getMergeValues(Ops2, 2, dl);
1479 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1480 SDValue &Root, SelectionDAG &DAG,
1482 MachineFunction &MF = DAG.getMachineFunction();
1483 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1485 TargetRegisterClass *RC;
1486 if (AFI->isThumb1OnlyFunction())
1487 RC = ARM::tGPRRegisterClass;
1489 RC = ARM::GPRRegisterClass;
1491 // Transform the arguments stored in physical registers into virtual ones.
1492 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1493 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1496 if (NextVA.isMemLoc()) {
1497 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1498 MachineFrameInfo *MFI = MF.getFrameInfo();
1499 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1501 // Create load node to retrieve arguments from the stack.
1502 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1503 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1505 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1506 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1509 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1513 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1514 CallingConv::ID CallConv, bool isVarArg,
1515 const SmallVectorImpl<ISD::InputArg>
1517 DebugLoc dl, SelectionDAG &DAG,
1518 SmallVectorImpl<SDValue> &InVals) {
1520 MachineFunction &MF = DAG.getMachineFunction();
1521 MachineFrameInfo *MFI = MF.getFrameInfo();
1523 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1525 // Assign locations to all of the incoming arguments.
1526 SmallVector<CCValAssign, 16> ArgLocs;
1527 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1529 CCInfo.AnalyzeFormalArguments(Ins,
1530 CCAssignFnForNode(CallConv, /* Return*/ false,
1533 SmallVector<SDValue, 16> ArgValues;
1535 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1536 CCValAssign &VA = ArgLocs[i];
1538 // Arguments stored in registers.
1539 if (VA.isRegLoc()) {
1540 EVT RegVT = VA.getLocVT();
1543 if (VA.needsCustom()) {
1544 // f64 and vector types are split up into multiple registers or
1545 // combinations of registers and stack slots.
1548 if (VA.getLocVT() == MVT::v2f64) {
1549 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1551 VA = ArgLocs[++i]; // skip ahead to next loc
1552 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1554 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1555 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1556 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1557 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1558 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1560 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1563 TargetRegisterClass *RC;
1565 if (RegVT == MVT::f32)
1566 RC = ARM::SPRRegisterClass;
1567 else if (RegVT == MVT::f64)
1568 RC = ARM::DPRRegisterClass;
1569 else if (RegVT == MVT::v2f64)
1570 RC = ARM::QPRRegisterClass;
1571 else if (RegVT == MVT::i32)
1572 RC = (AFI->isThumb1OnlyFunction() ?
1573 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1575 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1577 // Transform the arguments in physical registers into virtual ones.
1578 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1579 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1582 // If this is an 8 or 16-bit value, it is really passed promoted
1583 // to 32 bits. Insert an assert[sz]ext to capture this, then
1584 // truncate to the right size.
1585 switch (VA.getLocInfo()) {
1586 default: llvm_unreachable("Unknown loc info!");
1587 case CCValAssign::Full: break;
1588 case CCValAssign::BCvt:
1589 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1591 case CCValAssign::SExt:
1592 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1593 DAG.getValueType(VA.getValVT()));
1594 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1596 case CCValAssign::ZExt:
1597 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1598 DAG.getValueType(VA.getValVT()));
1599 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1603 InVals.push_back(ArgValue);
1605 } else { // VA.isRegLoc()
1608 assert(VA.isMemLoc());
1609 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1611 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1612 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1614 // Create load nodes to retrieve arguments from the stack.
1615 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1616 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1622 static const unsigned GPRArgRegs[] = {
1623 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1626 unsigned NumGPRs = CCInfo.getFirstUnallocated
1627 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1629 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1630 unsigned VARegSize = (4 - NumGPRs) * 4;
1631 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1632 unsigned ArgOffset = CCInfo.getNextStackOffset();
1633 if (VARegSaveSize) {
1634 // If this function is vararg, store any remaining integer argument regs
1635 // to their spots on the stack so that they may be loaded by deferencing
1636 // the result of va_next.
1637 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1638 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1639 VARegSaveSize - VARegSize);
1640 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1642 SmallVector<SDValue, 4> MemOps;
1643 for (; NumGPRs < 4; ++NumGPRs) {
1644 TargetRegisterClass *RC;
1645 if (AFI->isThumb1OnlyFunction())
1646 RC = ARM::tGPRRegisterClass;
1648 RC = ARM::GPRRegisterClass;
1650 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1651 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1652 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1653 MemOps.push_back(Store);
1654 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1655 DAG.getConstant(4, getPointerTy()));
1657 if (!MemOps.empty())
1658 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1659 &MemOps[0], MemOps.size());
1661 // This will point to the next argument passed via stack.
1662 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1668 /// isFloatingPointZero - Return true if this is +0.0.
1669 static bool isFloatingPointZero(SDValue Op) {
1670 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1671 return CFP->getValueAPF().isPosZero();
1672 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1673 // Maybe this has already been legalized into the constant pool?
1674 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1675 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1676 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1677 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1678 return CFP->getValueAPF().isPosZero();
1684 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1685 return ( isThumb1Only && (C & ~255U) == 0) ||
1686 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1689 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1690 /// the given operands.
1691 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1692 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1694 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1695 unsigned C = RHSC->getZExtValue();
1696 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1697 // Constant does not fit, try adjusting it by one?
1702 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1703 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1704 RHS = DAG.getConstant(C-1, MVT::i32);
1709 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1710 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1711 RHS = DAG.getConstant(C-1, MVT::i32);
1716 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1717 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1718 RHS = DAG.getConstant(C+1, MVT::i32);
1723 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1724 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1725 RHS = DAG.getConstant(C+1, MVT::i32);
1732 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1733 ARMISD::NodeType CompareType;
1736 CompareType = ARMISD::CMP;
1741 CompareType = ARMISD::CMPZ;
1744 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1745 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1748 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1749 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1752 if (!isFloatingPointZero(RHS))
1753 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1755 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1756 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1759 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1760 const ARMSubtarget *ST) {
1761 EVT VT = Op.getValueType();
1762 SDValue LHS = Op.getOperand(0);
1763 SDValue RHS = Op.getOperand(1);
1764 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1765 SDValue TrueVal = Op.getOperand(2);
1766 SDValue FalseVal = Op.getOperand(3);
1767 DebugLoc dl = Op.getDebugLoc();
1769 if (LHS.getValueType() == MVT::i32) {
1771 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1772 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1773 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1776 ARMCC::CondCodes CondCode, CondCode2;
1777 FPCCToARMCC(CC, CondCode, CondCode2);
1779 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1780 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1781 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1782 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1784 if (CondCode2 != ARMCC::AL) {
1785 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1786 // FIXME: Needs another CMP because flag can have but one use.
1787 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1788 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1789 Result, TrueVal, ARMCC2, CCR, Cmp2);
1794 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1795 const ARMSubtarget *ST) {
1796 SDValue Chain = Op.getOperand(0);
1797 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1798 SDValue LHS = Op.getOperand(2);
1799 SDValue RHS = Op.getOperand(3);
1800 SDValue Dest = Op.getOperand(4);
1801 DebugLoc dl = Op.getDebugLoc();
1803 if (LHS.getValueType() == MVT::i32) {
1805 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1806 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1807 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1808 Chain, Dest, ARMCC, CCR,Cmp);
1811 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1812 ARMCC::CondCodes CondCode, CondCode2;
1813 FPCCToARMCC(CC, CondCode, CondCode2);
1815 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1816 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1817 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1818 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1819 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1820 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1821 if (CondCode2 != ARMCC::AL) {
1822 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1823 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1824 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1829 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1830 SDValue Chain = Op.getOperand(0);
1831 SDValue Table = Op.getOperand(1);
1832 SDValue Index = Op.getOperand(2);
1833 DebugLoc dl = Op.getDebugLoc();
1835 EVT PTy = getPointerTy();
1836 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1837 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1838 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1839 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1840 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1841 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1842 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1843 if (Subtarget->isThumb2()) {
1844 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1845 // which does another jump to the destination. This also makes it easier
1846 // to translate it to TBB / TBH later.
1847 // FIXME: This might not work if the function is extremely large.
1848 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1849 Addr, Op.getOperand(2), JTI, UId);
1851 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1852 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1853 Chain = Addr.getValue(1);
1854 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1855 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1857 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1858 Chain = Addr.getValue(1);
1859 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1863 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1864 DebugLoc dl = Op.getDebugLoc();
1866 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1867 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1868 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1871 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1872 EVT VT = Op.getValueType();
1873 DebugLoc dl = Op.getDebugLoc();
1875 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1877 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1878 return DAG.getNode(Opc, dl, VT, Op);
1881 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1882 // Implement fcopysign with a fabs and a conditional fneg.
1883 SDValue Tmp0 = Op.getOperand(0);
1884 SDValue Tmp1 = Op.getOperand(1);
1885 DebugLoc dl = Op.getDebugLoc();
1886 EVT VT = Op.getValueType();
1887 EVT SrcVT = Tmp1.getValueType();
1888 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1889 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1890 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1891 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1892 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1895 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1896 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1897 MFI->setFrameAddressIsTaken(true);
1898 EVT VT = Op.getValueType();
1899 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1900 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1901 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1902 ? ARM::R7 : ARM::R11;
1903 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1905 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1910 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1912 SDValue Dst, SDValue Src,
1913 SDValue Size, unsigned Align,
1915 const Value *DstSV, uint64_t DstSVOff,
1916 const Value *SrcSV, uint64_t SrcSVOff){
1917 // Do repeated 4-byte loads and stores. To be improved.
1918 // This requires 4-byte alignment.
1919 if ((Align & 3) != 0)
1921 // This requires the copy size to be a constant, preferrably
1922 // within a subtarget-specific limit.
1923 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1926 uint64_t SizeVal = ConstantSize->getZExtValue();
1927 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1930 unsigned BytesLeft = SizeVal & 3;
1931 unsigned NumMemOps = SizeVal >> 2;
1932 unsigned EmittedNumMemOps = 0;
1934 unsigned VTSize = 4;
1936 const unsigned MAX_LOADS_IN_LDM = 6;
1937 SDValue TFOps[MAX_LOADS_IN_LDM];
1938 SDValue Loads[MAX_LOADS_IN_LDM];
1939 uint64_t SrcOff = 0, DstOff = 0;
1941 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1942 // same number of stores. The loads and stores will get combined into
1943 // ldm/stm later on.
1944 while (EmittedNumMemOps < NumMemOps) {
1946 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1947 Loads[i] = DAG.getLoad(VT, dl, Chain,
1948 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1949 DAG.getConstant(SrcOff, MVT::i32)),
1950 SrcSV, SrcSVOff + SrcOff);
1951 TFOps[i] = Loads[i].getValue(1);
1954 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1957 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1958 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1959 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1960 DAG.getConstant(DstOff, MVT::i32)),
1961 DstSV, DstSVOff + DstOff);
1964 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1966 EmittedNumMemOps += i;
1972 // Issue loads / stores for the trailing (1 - 3) bytes.
1973 unsigned BytesLeftSave = BytesLeft;
1976 if (BytesLeft >= 2) {
1984 Loads[i] = DAG.getLoad(VT, dl, Chain,
1985 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1986 DAG.getConstant(SrcOff, MVT::i32)),
1987 SrcSV, SrcSVOff + SrcOff);
1988 TFOps[i] = Loads[i].getValue(1);
1991 BytesLeft -= VTSize;
1993 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1996 BytesLeft = BytesLeftSave;
1998 if (BytesLeft >= 2) {
2006 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2007 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2008 DAG.getConstant(DstOff, MVT::i32)),
2009 DstSV, DstSVOff + DstOff);
2012 BytesLeft -= VTSize;
2014 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2017 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2018 SDValue Op = N->getOperand(0);
2019 DebugLoc dl = N->getDebugLoc();
2020 if (N->getValueType(0) == MVT::f64) {
2021 // Turn i64->f64 into FMDRR.
2022 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2023 DAG.getConstant(0, MVT::i32));
2024 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2025 DAG.getConstant(1, MVT::i32));
2026 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2029 // Turn f64->i64 into FMRRD.
2030 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2031 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2033 // Merge the pieces into a single i64 value.
2034 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2037 /// getZeroVector - Returns a vector of specified type with all zero elements.
2039 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2040 assert(VT.isVector() && "Expected a vector type");
2042 // Zero vectors are used to represent vector negation and in those cases
2043 // will be implemented with the NEON VNEG instruction. However, VNEG does
2044 // not support i64 elements, so sometimes the zero vectors will need to be
2045 // explicitly constructed. For those cases, and potentially other uses in
2046 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2047 // to their dest type. This ensures they get CSE'd.
2049 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2050 SmallVector<SDValue, 8> Ops;
2053 if (VT.getSizeInBits() == 64) {
2054 Ops.assign(8, Cst); TVT = MVT::v8i8;
2056 Ops.assign(16, Cst); TVT = MVT::v16i8;
2058 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2060 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2063 /// getOnesVector - Returns a vector of specified type with all bits set.
2065 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2066 assert(VT.isVector() && "Expected a vector type");
2068 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2069 // dest type. This ensures they get CSE'd.
2071 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2072 SmallVector<SDValue, 8> Ops;
2075 if (VT.getSizeInBits() == 64) {
2076 Ops.assign(8, Cst); TVT = MVT::v8i8;
2078 Ops.assign(16, Cst); TVT = MVT::v16i8;
2080 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2082 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2085 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2086 const ARMSubtarget *ST) {
2087 EVT VT = N->getValueType(0);
2088 DebugLoc dl = N->getDebugLoc();
2090 // Lower vector shifts on NEON to use VSHL.
2091 if (VT.isVector()) {
2092 assert(ST->hasNEON() && "unexpected vector shift");
2094 // Left shifts translate directly to the vshiftu intrinsic.
2095 if (N->getOpcode() == ISD::SHL)
2096 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2097 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2098 N->getOperand(0), N->getOperand(1));
2100 assert((N->getOpcode() == ISD::SRA ||
2101 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2103 // NEON uses the same intrinsics for both left and right shifts. For
2104 // right shifts, the shift amounts are negative, so negate the vector of
2106 EVT ShiftVT = N->getOperand(1).getValueType();
2107 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2108 getZeroVector(ShiftVT, DAG, dl),
2110 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2111 Intrinsic::arm_neon_vshifts :
2112 Intrinsic::arm_neon_vshiftu);
2113 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2114 DAG.getConstant(vshiftInt, MVT::i32),
2115 N->getOperand(0), NegatedCount);
2118 // We can get here for a node like i32 = ISD::SHL i32, i64
2122 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2123 "Unknown shift to lower!");
2125 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2126 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2127 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2130 // If we are in thumb mode, we don't have RRX.
2131 if (ST->isThumb1Only()) return SDValue();
2133 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2134 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2135 DAG.getConstant(0, MVT::i32));
2136 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2137 DAG.getConstant(1, MVT::i32));
2139 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2140 // captures the result into a carry flag.
2141 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2142 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2144 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2145 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2147 // Merge the pieces into a single i64 value.
2148 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2151 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2152 SDValue TmpOp0, TmpOp1;
2153 bool Invert = false;
2157 SDValue Op0 = Op.getOperand(0);
2158 SDValue Op1 = Op.getOperand(1);
2159 SDValue CC = Op.getOperand(2);
2160 EVT VT = Op.getValueType();
2161 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2162 DebugLoc dl = Op.getDebugLoc();
2164 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2165 switch (SetCCOpcode) {
2166 default: llvm_unreachable("Illegal FP comparison"); break;
2168 case ISD::SETNE: Invert = true; // Fallthrough
2170 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2172 case ISD::SETLT: Swap = true; // Fallthrough
2174 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2176 case ISD::SETLE: Swap = true; // Fallthrough
2178 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2179 case ISD::SETUGE: Swap = true; // Fallthrough
2180 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2181 case ISD::SETUGT: Swap = true; // Fallthrough
2182 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2183 case ISD::SETUEQ: Invert = true; // Fallthrough
2185 // Expand this to (OLT | OGT).
2189 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2190 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2192 case ISD::SETUO: Invert = true; // Fallthrough
2194 // Expand this to (OLT | OGE).
2198 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2199 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2203 // Integer comparisons.
2204 switch (SetCCOpcode) {
2205 default: llvm_unreachable("Illegal integer comparison"); break;
2206 case ISD::SETNE: Invert = true;
2207 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2208 case ISD::SETLT: Swap = true;
2209 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2210 case ISD::SETLE: Swap = true;
2211 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2212 case ISD::SETULT: Swap = true;
2213 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2214 case ISD::SETULE: Swap = true;
2215 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2218 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2219 if (Opc == ARMISD::VCEQ) {
2222 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2224 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2227 // Ignore bitconvert.
2228 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2229 AndOp = AndOp.getOperand(0);
2231 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2233 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2234 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2241 std::swap(Op0, Op1);
2243 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2246 Result = DAG.getNOT(dl, Result, VT);
2251 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2252 /// VMOV instruction, and if so, return the constant being splatted.
2253 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2254 unsigned SplatBitSize, SelectionDAG &DAG) {
2255 switch (SplatBitSize) {
2257 // Any 1-byte value is OK.
2258 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2259 return DAG.getTargetConstant(SplatBits, MVT::i8);
2262 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2263 if ((SplatBits & ~0xff) == 0 ||
2264 (SplatBits & ~0xff00) == 0)
2265 return DAG.getTargetConstant(SplatBits, MVT::i16);
2269 // NEON's 32-bit VMOV supports splat values where:
2270 // * only one byte is nonzero, or
2271 // * the least significant byte is 0xff and the second byte is nonzero, or
2272 // * the least significant 2 bytes are 0xff and the third is nonzero.
2273 if ((SplatBits & ~0xff) == 0 ||
2274 (SplatBits & ~0xff00) == 0 ||
2275 (SplatBits & ~0xff0000) == 0 ||
2276 (SplatBits & ~0xff000000) == 0)
2277 return DAG.getTargetConstant(SplatBits, MVT::i32);
2279 if ((SplatBits & ~0xffff) == 0 &&
2280 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2281 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2283 if ((SplatBits & ~0xffffff) == 0 &&
2284 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2285 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2287 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2288 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2289 // VMOV.I32. A (very) minor optimization would be to replicate the value
2290 // and fall through here to test for a valid 64-bit splat. But, then the
2291 // caller would also need to check and handle the change in size.
2295 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2296 uint64_t BitMask = 0xff;
2298 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2299 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2301 else if ((SplatBits & BitMask) != 0)
2305 return DAG.getTargetConstant(Val, MVT::i64);
2309 llvm_unreachable("unexpected size for isVMOVSplat");
2316 /// getVMOVImm - If this is a build_vector of constants which can be
2317 /// formed by using a VMOV instruction of the specified element size,
2318 /// return the constant being splatted. The ByteSize field indicates the
2319 /// number of bytes of each element [1248].
2320 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2321 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2322 APInt SplatBits, SplatUndef;
2323 unsigned SplatBitSize;
2325 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2326 HasAnyUndefs, ByteSize * 8))
2329 if (SplatBitSize > ByteSize * 8)
2332 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2336 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2337 bool &ReverseVEXT, unsigned &Imm) {
2338 unsigned NumElts = VT.getVectorNumElements();
2339 ReverseVEXT = false;
2342 // If this is a VEXT shuffle, the immediate value is the index of the first
2343 // element. The other shuffle indices must be the successive elements after
2345 unsigned ExpectedElt = Imm;
2346 for (unsigned i = 1; i < NumElts; ++i) {
2347 // Increment the expected index. If it wraps around, it may still be
2348 // a VEXT but the source vectors must be swapped.
2350 if (ExpectedElt == NumElts * 2) {
2355 if (ExpectedElt != static_cast<unsigned>(M[i]))
2359 // Adjust the index value if the source operands will be swapped.
2366 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2367 /// instruction with the specified blocksize. (The order of the elements
2368 /// within each block of the vector is reversed.)
2369 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2370 unsigned BlockSize) {
2371 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2372 "Only possible block sizes for VREV are: 16, 32, 64");
2374 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2378 unsigned NumElts = VT.getVectorNumElements();
2379 unsigned BlockElts = M[0] + 1;
2381 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2384 for (unsigned i = 0; i < NumElts; ++i) {
2385 if ((unsigned) M[i] !=
2386 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2393 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2394 unsigned &WhichResult) {
2395 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2399 unsigned NumElts = VT.getVectorNumElements();
2400 WhichResult = (M[0] == 0 ? 0 : 1);
2401 for (unsigned i = 0; i < NumElts; i += 2) {
2402 if ((unsigned) M[i] != i + WhichResult ||
2403 (unsigned) M[i+1] != i + NumElts + WhichResult)
2409 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2410 unsigned &WhichResult) {
2411 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2415 unsigned NumElts = VT.getVectorNumElements();
2416 WhichResult = (M[0] == 0 ? 0 : 1);
2417 for (unsigned i = 0; i != NumElts; ++i) {
2418 if ((unsigned) M[i] != 2 * i + WhichResult)
2422 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2423 if (VT.is64BitVector() && EltSz == 32)
2429 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2430 unsigned &WhichResult) {
2431 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2435 unsigned NumElts = VT.getVectorNumElements();
2436 WhichResult = (M[0] == 0 ? 0 : 1);
2437 unsigned Idx = WhichResult * NumElts / 2;
2438 for (unsigned i = 0; i != NumElts; i += 2) {
2439 if ((unsigned) M[i] != Idx ||
2440 (unsigned) M[i+1] != Idx + NumElts)
2445 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2446 if (VT.is64BitVector() && EltSz == 32)
2452 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2453 // Canonicalize all-zeros and all-ones vectors.
2454 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2455 if (ConstVal->isNullValue())
2456 return getZeroVector(VT, DAG, dl);
2457 if (ConstVal->isAllOnesValue())
2458 return getOnesVector(VT, DAG, dl);
2461 if (VT.is64BitVector()) {
2462 switch (Val.getValueType().getSizeInBits()) {
2463 case 8: CanonicalVT = MVT::v8i8; break;
2464 case 16: CanonicalVT = MVT::v4i16; break;
2465 case 32: CanonicalVT = MVT::v2i32; break;
2466 case 64: CanonicalVT = MVT::v1i64; break;
2467 default: llvm_unreachable("unexpected splat element type"); break;
2470 assert(VT.is128BitVector() && "unknown splat vector size");
2471 switch (Val.getValueType().getSizeInBits()) {
2472 case 8: CanonicalVT = MVT::v16i8; break;
2473 case 16: CanonicalVT = MVT::v8i16; break;
2474 case 32: CanonicalVT = MVT::v4i32; break;
2475 case 64: CanonicalVT = MVT::v2i64; break;
2476 default: llvm_unreachable("unexpected splat element type"); break;
2480 // Build a canonical splat for this value.
2481 SmallVector<SDValue, 8> Ops;
2482 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2483 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2485 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2488 // If this is a case we can't handle, return null and let the default
2489 // expansion code take care of it.
2490 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2491 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2492 DebugLoc dl = Op.getDebugLoc();
2493 EVT VT = Op.getValueType();
2495 APInt SplatBits, SplatUndef;
2496 unsigned SplatBitSize;
2498 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2499 if (SplatBitSize <= 64) {
2500 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2501 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2503 return BuildSplat(Val, VT, DAG, dl);
2507 // If there are only 2 elements in a 128-bit vector, insert them into an
2508 // undef vector. This handles the common case for 128-bit vector argument
2509 // passing, where the insertions should be translated to subreg accesses
2510 // with no real instructions.
2511 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2512 SDValue Val = DAG.getUNDEF(VT);
2513 SDValue Op0 = Op.getOperand(0);
2514 SDValue Op1 = Op.getOperand(1);
2515 if (Op0.getOpcode() != ISD::UNDEF)
2516 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2517 DAG.getIntPtrConstant(0));
2518 if (Op1.getOpcode() != ISD::UNDEF)
2519 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2520 DAG.getIntPtrConstant(1));
2527 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2528 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2529 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2530 /// are assumed to be legal.
2532 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2534 if (VT.getVectorNumElements() == 4 &&
2535 (VT.is128BitVector() || VT.is64BitVector())) {
2536 unsigned PFIndexes[4];
2537 for (unsigned i = 0; i != 4; ++i) {
2541 PFIndexes[i] = M[i];
2544 // Compute the index in the perfect shuffle table.
2545 unsigned PFTableIndex =
2546 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2547 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2548 unsigned Cost = (PFEntry >> 30);
2555 unsigned Imm, WhichResult;
2557 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2558 isVREVMask(M, VT, 64) ||
2559 isVREVMask(M, VT, 32) ||
2560 isVREVMask(M, VT, 16) ||
2561 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2562 isVTRNMask(M, VT, WhichResult) ||
2563 isVUZPMask(M, VT, WhichResult) ||
2564 isVZIPMask(M, VT, WhichResult));
2567 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2568 /// the specified operations to build the shuffle.
2569 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2570 SDValue RHS, SelectionDAG &DAG,
2572 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2573 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2574 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2577 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2586 OP_VUZPL, // VUZP, left result
2587 OP_VUZPR, // VUZP, right result
2588 OP_VZIPL, // VZIP, left result
2589 OP_VZIPR, // VZIP, right result
2590 OP_VTRNL, // VTRN, left result
2591 OP_VTRNR // VTRN, right result
2594 if (OpNum == OP_COPY) {
2595 if (LHSID == (1*9+2)*9+3) return LHS;
2596 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2600 SDValue OpLHS, OpRHS;
2601 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2602 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2603 EVT VT = OpLHS.getValueType();
2606 default: llvm_unreachable("Unknown shuffle opcode!");
2608 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2613 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2614 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2618 return DAG.getNode(ARMISD::VEXT, dl, VT,
2620 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2623 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2624 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2627 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2628 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2631 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2632 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2636 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2637 SDValue V1 = Op.getOperand(0);
2638 SDValue V2 = Op.getOperand(1);
2639 DebugLoc dl = Op.getDebugLoc();
2640 EVT VT = Op.getValueType();
2641 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2642 SmallVector<int, 8> ShuffleMask;
2644 // Convert shuffles that are directly supported on NEON to target-specific
2645 // DAG nodes, instead of keeping them as shuffles and matching them again
2646 // during code selection. This is more efficient and avoids the possibility
2647 // of inconsistencies between legalization and selection.
2648 // FIXME: floating-point vectors should be canonicalized to integer vectors
2649 // of the same time so that they get CSEd properly.
2650 SVN->getMask(ShuffleMask);
2652 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2653 int Lane = SVN->getSplatIndex();
2654 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2655 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2657 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2658 DAG.getConstant(Lane, MVT::i32));
2663 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2666 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2667 DAG.getConstant(Imm, MVT::i32));
2670 if (isVREVMask(ShuffleMask, VT, 64))
2671 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2672 if (isVREVMask(ShuffleMask, VT, 32))
2673 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2674 if (isVREVMask(ShuffleMask, VT, 16))
2675 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2677 // Check for Neon shuffles that modify both input vectors in place.
2678 // If both results are used, i.e., if there are two shuffles with the same
2679 // source operands and with masks corresponding to both results of one of
2680 // these operations, DAG memoization will ensure that a single node is
2681 // used for both shuffles.
2682 unsigned WhichResult;
2683 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2684 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2685 V1, V2).getValue(WhichResult);
2686 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2687 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2688 V1, V2).getValue(WhichResult);
2689 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2690 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2691 V1, V2).getValue(WhichResult);
2693 // If the shuffle is not directly supported and it has 4 elements, use
2694 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2695 if (VT.getVectorNumElements() == 4 &&
2696 (VT.is128BitVector() || VT.is64BitVector())) {
2697 unsigned PFIndexes[4];
2698 for (unsigned i = 0; i != 4; ++i) {
2699 if (ShuffleMask[i] < 0)
2702 PFIndexes[i] = ShuffleMask[i];
2705 // Compute the index in the perfect shuffle table.
2706 unsigned PFTableIndex =
2707 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2709 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2710 unsigned Cost = (PFEntry >> 30);
2713 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2719 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2720 EVT VT = Op.getValueType();
2721 DebugLoc dl = Op.getDebugLoc();
2722 SDValue Vec = Op.getOperand(0);
2723 SDValue Lane = Op.getOperand(1);
2724 assert(VT == MVT::i32 &&
2725 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
2726 "unexpected type for custom-lowering vector extract");
2727 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2730 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2731 // The only time a CONCAT_VECTORS operation can have legal types is when
2732 // two 64-bit vectors are concatenated to a 128-bit vector.
2733 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2734 "unexpected CONCAT_VECTORS");
2735 DebugLoc dl = Op.getDebugLoc();
2736 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2737 SDValue Op0 = Op.getOperand(0);
2738 SDValue Op1 = Op.getOperand(1);
2739 if (Op0.getOpcode() != ISD::UNDEF)
2740 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2741 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2742 DAG.getIntPtrConstant(0));
2743 if (Op1.getOpcode() != ISD::UNDEF)
2744 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2745 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2746 DAG.getIntPtrConstant(1));
2747 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2750 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2751 switch (Op.getOpcode()) {
2752 default: llvm_unreachable("Don't know how to custom lower this!");
2753 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2754 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
2755 case ISD::GlobalAddress:
2756 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2757 LowerGlobalAddressELF(Op, DAG);
2758 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2759 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2760 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2761 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2762 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2763 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2764 case ISD::SINT_TO_FP:
2765 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2766 case ISD::FP_TO_SINT:
2767 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2768 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2769 case ISD::RETURNADDR: break;
2770 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2771 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2772 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2773 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2776 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2777 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2778 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2779 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2780 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2781 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2786 /// ReplaceNodeResults - Replace the results of node with an illegal result
2787 /// type with new values built out of custom code.
2788 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2789 SmallVectorImpl<SDValue>&Results,
2790 SelectionDAG &DAG) {
2791 switch (N->getOpcode()) {
2793 llvm_unreachable("Don't know how to custom expand this!");
2795 case ISD::BIT_CONVERT:
2796 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2800 SDValue Res = LowerShift(N, DAG, Subtarget);
2802 Results.push_back(Res);
2808 //===----------------------------------------------------------------------===//
2809 // ARM Scheduler Hooks
2810 //===----------------------------------------------------------------------===//
2813 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2814 MachineBasicBlock *BB,
2815 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
2816 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2817 DebugLoc dl = MI->getDebugLoc();
2818 switch (MI->getOpcode()) {
2820 llvm_unreachable("Unexpected instr type to insert");
2821 case ARM::tMOVCCr_pseudo: {
2822 // To "insert" a SELECT_CC instruction, we actually have to insert the
2823 // diamond control-flow pattern. The incoming instruction knows the
2824 // destination vreg to set, the condition code register to branch on, the
2825 // true/false values to select between, and a branch opcode to use.
2826 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2827 MachineFunction::iterator It = BB;
2833 // cmpTY ccX, r1, r2
2835 // fallthrough --> copy0MBB
2836 MachineBasicBlock *thisMBB = BB;
2837 MachineFunction *F = BB->getParent();
2838 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2839 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2840 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2841 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2842 F->insert(It, copy0MBB);
2843 F->insert(It, sinkMBB);
2844 // Update machine-CFG edges by first adding all successors of the current
2845 // block to the new block which will contain the Phi node for the select.
2846 // Also inform sdisel of the edge changes.
2847 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
2848 E = BB->succ_end(); I != E; ++I) {
2849 EM->insert(std::make_pair(*I, sinkMBB));
2850 sinkMBB->addSuccessor(*I);
2852 // Next, remove all successors of the current block, and add the true
2853 // and fallthrough blocks as its successors.
2854 while (!BB->succ_empty())
2855 BB->removeSuccessor(BB->succ_begin());
2856 BB->addSuccessor(copy0MBB);
2857 BB->addSuccessor(sinkMBB);
2860 // %FalseValue = ...
2861 // # fallthrough to sinkMBB
2864 // Update machine-CFG edges
2865 BB->addSuccessor(sinkMBB);
2868 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2871 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2872 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2873 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2875 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2882 case ARM::t2SUBrSPi_:
2883 case ARM::t2SUBrSPi12_:
2884 case ARM::t2SUBrSPs_: {
2885 MachineFunction *MF = BB->getParent();
2886 unsigned DstReg = MI->getOperand(0).getReg();
2887 unsigned SrcReg = MI->getOperand(1).getReg();
2888 bool DstIsDead = MI->getOperand(0).isDead();
2889 bool SrcIsKill = MI->getOperand(1).isKill();
2891 if (SrcReg != ARM::SP) {
2892 // Copy the source to SP from virtual register.
2893 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2894 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2895 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2896 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2897 .addReg(SrcReg, getKillRegState(SrcIsKill));
2901 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2902 switch (MI->getOpcode()) {
2904 llvm_unreachable("Unexpected pseudo instruction!");
2910 OpOpc = ARM::tADDspr;
2913 OpOpc = ARM::tSUBspi;
2915 case ARM::t2SUBrSPi_:
2916 OpOpc = ARM::t2SUBrSPi;
2917 NeedPred = true; NeedCC = true;
2919 case ARM::t2SUBrSPi12_:
2920 OpOpc = ARM::t2SUBrSPi12;
2923 case ARM::t2SUBrSPs_:
2924 OpOpc = ARM::t2SUBrSPs;
2925 NeedPred = true; NeedCC = true; NeedOp3 = true;
2928 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2929 if (OpOpc == ARM::tAND)
2930 AddDefaultT1CC(MIB);
2931 MIB.addReg(ARM::SP);
2932 MIB.addOperand(MI->getOperand(2));
2934 MIB.addOperand(MI->getOperand(3));
2936 AddDefaultPred(MIB);
2940 // Copy the result from SP to virtual register.
2941 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2942 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2943 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2944 BuildMI(BB, dl, TII->get(CopyOpc))
2945 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
2947 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2953 //===----------------------------------------------------------------------===//
2954 // ARM Optimization Hooks
2955 //===----------------------------------------------------------------------===//
2958 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2959 TargetLowering::DAGCombinerInfo &DCI) {
2960 SelectionDAG &DAG = DCI.DAG;
2961 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2962 EVT VT = N->getValueType(0);
2963 unsigned Opc = N->getOpcode();
2964 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2965 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2966 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2967 ISD::CondCode CC = ISD::SETCC_INVALID;
2970 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2972 SDValue CCOp = Slct.getOperand(0);
2973 if (CCOp.getOpcode() == ISD::SETCC)
2974 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2977 bool DoXform = false;
2979 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2982 if (LHS.getOpcode() == ISD::Constant &&
2983 cast<ConstantSDNode>(LHS)->isNullValue()) {
2985 } else if (CC != ISD::SETCC_INVALID &&
2986 RHS.getOpcode() == ISD::Constant &&
2987 cast<ConstantSDNode>(RHS)->isNullValue()) {
2988 std::swap(LHS, RHS);
2989 SDValue Op0 = Slct.getOperand(0);
2990 EVT OpVT = isSlctCC ? Op0.getValueType() :
2991 Op0.getOperand(0).getValueType();
2992 bool isInt = OpVT.isInteger();
2993 CC = ISD::getSetCCInverse(CC, isInt);
2995 if (!TLI.isCondCodeLegal(CC, OpVT))
2996 return SDValue(); // Inverse operator isn't legal.
3003 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3005 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3006 Slct.getOperand(0), Slct.getOperand(1), CC);
3007 SDValue CCOp = Slct.getOperand(0);
3009 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3010 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3011 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3012 CCOp, OtherOp, Result);
3017 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3018 static SDValue PerformADDCombine(SDNode *N,
3019 TargetLowering::DAGCombinerInfo &DCI) {
3020 // added by evan in r37685 with no testcase.
3021 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3023 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3024 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3025 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3026 if (Result.getNode()) return Result;
3028 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3029 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3030 if (Result.getNode()) return Result;
3036 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3037 static SDValue PerformSUBCombine(SDNode *N,
3038 TargetLowering::DAGCombinerInfo &DCI) {
3039 // added by evan in r37685 with no testcase.
3040 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3042 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3043 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3044 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3045 if (Result.getNode()) return Result;
3051 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3052 static SDValue PerformFMRRDCombine(SDNode *N,
3053 TargetLowering::DAGCombinerInfo &DCI) {
3054 // fmrrd(fmdrr x, y) -> x,y
3055 SDValue InDouble = N->getOperand(0);
3056 if (InDouble.getOpcode() == ARMISD::FMDRR)
3057 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3061 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3062 /// operand of a vector shift operation, where all the elements of the
3063 /// build_vector must have the same constant integer value.
3064 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3065 // Ignore bit_converts.
3066 while (Op.getOpcode() == ISD::BIT_CONVERT)
3067 Op = Op.getOperand(0);
3068 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3069 APInt SplatBits, SplatUndef;
3070 unsigned SplatBitSize;
3072 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3073 HasAnyUndefs, ElementBits) ||
3074 SplatBitSize > ElementBits)
3076 Cnt = SplatBits.getSExtValue();
3080 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3081 /// operand of a vector shift left operation. That value must be in the range:
3082 /// 0 <= Value < ElementBits for a left shift; or
3083 /// 0 <= Value <= ElementBits for a long left shift.
3084 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3085 assert(VT.isVector() && "vector shift count is not a vector type");
3086 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3087 if (! getVShiftImm(Op, ElementBits, Cnt))
3089 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3092 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3093 /// operand of a vector shift right operation. For a shift opcode, the value
3094 /// is positive, but for an intrinsic the value count must be negative. The
3095 /// absolute value must be in the range:
3096 /// 1 <= |Value| <= ElementBits for a right shift; or
3097 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3098 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3100 assert(VT.isVector() && "vector shift count is not a vector type");
3101 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3102 if (! getVShiftImm(Op, ElementBits, Cnt))
3106 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3109 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3110 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3111 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3114 // Don't do anything for most intrinsics.
3117 // Vector shifts: check for immediate versions and lower them.
3118 // Note: This is done during DAG combining instead of DAG legalizing because
3119 // the build_vectors for 64-bit vector element shift counts are generally
3120 // not legal, and it is hard to see their values after they get legalized to
3121 // loads from a constant pool.
3122 case Intrinsic::arm_neon_vshifts:
3123 case Intrinsic::arm_neon_vshiftu:
3124 case Intrinsic::arm_neon_vshiftls:
3125 case Intrinsic::arm_neon_vshiftlu:
3126 case Intrinsic::arm_neon_vshiftn:
3127 case Intrinsic::arm_neon_vrshifts:
3128 case Intrinsic::arm_neon_vrshiftu:
3129 case Intrinsic::arm_neon_vrshiftn:
3130 case Intrinsic::arm_neon_vqshifts:
3131 case Intrinsic::arm_neon_vqshiftu:
3132 case Intrinsic::arm_neon_vqshiftsu:
3133 case Intrinsic::arm_neon_vqshiftns:
3134 case Intrinsic::arm_neon_vqshiftnu:
3135 case Intrinsic::arm_neon_vqshiftnsu:
3136 case Intrinsic::arm_neon_vqrshiftns:
3137 case Intrinsic::arm_neon_vqrshiftnu:
3138 case Intrinsic::arm_neon_vqrshiftnsu: {
3139 EVT VT = N->getOperand(1).getValueType();
3141 unsigned VShiftOpc = 0;
3144 case Intrinsic::arm_neon_vshifts:
3145 case Intrinsic::arm_neon_vshiftu:
3146 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3147 VShiftOpc = ARMISD::VSHL;
3150 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3151 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3152 ARMISD::VSHRs : ARMISD::VSHRu);
3157 case Intrinsic::arm_neon_vshiftls:
3158 case Intrinsic::arm_neon_vshiftlu:
3159 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3161 llvm_unreachable("invalid shift count for vshll intrinsic");
3163 case Intrinsic::arm_neon_vrshifts:
3164 case Intrinsic::arm_neon_vrshiftu:
3165 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3169 case Intrinsic::arm_neon_vqshifts:
3170 case Intrinsic::arm_neon_vqshiftu:
3171 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3175 case Intrinsic::arm_neon_vqshiftsu:
3176 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3178 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3180 case Intrinsic::arm_neon_vshiftn:
3181 case Intrinsic::arm_neon_vrshiftn:
3182 case Intrinsic::arm_neon_vqshiftns:
3183 case Intrinsic::arm_neon_vqshiftnu:
3184 case Intrinsic::arm_neon_vqshiftnsu:
3185 case Intrinsic::arm_neon_vqrshiftns:
3186 case Intrinsic::arm_neon_vqrshiftnu:
3187 case Intrinsic::arm_neon_vqrshiftnsu:
3188 // Narrowing shifts require an immediate right shift.
3189 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3191 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3194 llvm_unreachable("unhandled vector shift");
3198 case Intrinsic::arm_neon_vshifts:
3199 case Intrinsic::arm_neon_vshiftu:
3200 // Opcode already set above.
3202 case Intrinsic::arm_neon_vshiftls:
3203 case Intrinsic::arm_neon_vshiftlu:
3204 if (Cnt == VT.getVectorElementType().getSizeInBits())
3205 VShiftOpc = ARMISD::VSHLLi;
3207 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3208 ARMISD::VSHLLs : ARMISD::VSHLLu);
3210 case Intrinsic::arm_neon_vshiftn:
3211 VShiftOpc = ARMISD::VSHRN; break;
3212 case Intrinsic::arm_neon_vrshifts:
3213 VShiftOpc = ARMISD::VRSHRs; break;
3214 case Intrinsic::arm_neon_vrshiftu:
3215 VShiftOpc = ARMISD::VRSHRu; break;
3216 case Intrinsic::arm_neon_vrshiftn:
3217 VShiftOpc = ARMISD::VRSHRN; break;
3218 case Intrinsic::arm_neon_vqshifts:
3219 VShiftOpc = ARMISD::VQSHLs; break;
3220 case Intrinsic::arm_neon_vqshiftu:
3221 VShiftOpc = ARMISD::VQSHLu; break;
3222 case Intrinsic::arm_neon_vqshiftsu:
3223 VShiftOpc = ARMISD::VQSHLsu; break;
3224 case Intrinsic::arm_neon_vqshiftns:
3225 VShiftOpc = ARMISD::VQSHRNs; break;
3226 case Intrinsic::arm_neon_vqshiftnu:
3227 VShiftOpc = ARMISD::VQSHRNu; break;
3228 case Intrinsic::arm_neon_vqshiftnsu:
3229 VShiftOpc = ARMISD::VQSHRNsu; break;
3230 case Intrinsic::arm_neon_vqrshiftns:
3231 VShiftOpc = ARMISD::VQRSHRNs; break;
3232 case Intrinsic::arm_neon_vqrshiftnu:
3233 VShiftOpc = ARMISD::VQRSHRNu; break;
3234 case Intrinsic::arm_neon_vqrshiftnsu:
3235 VShiftOpc = ARMISD::VQRSHRNsu; break;
3238 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3239 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3242 case Intrinsic::arm_neon_vshiftins: {
3243 EVT VT = N->getOperand(1).getValueType();
3245 unsigned VShiftOpc = 0;
3247 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3248 VShiftOpc = ARMISD::VSLI;
3249 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3250 VShiftOpc = ARMISD::VSRI;
3252 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3255 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3256 N->getOperand(1), N->getOperand(2),
3257 DAG.getConstant(Cnt, MVT::i32));
3260 case Intrinsic::arm_neon_vqrshifts:
3261 case Intrinsic::arm_neon_vqrshiftu:
3262 // No immediate versions of these to check for.
3269 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3270 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3271 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3272 /// vector element shift counts are generally not legal, and it is hard to see
3273 /// their values after they get legalized to loads from a constant pool.
3274 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3275 const ARMSubtarget *ST) {
3276 EVT VT = N->getValueType(0);
3278 // Nothing to be done for scalar shifts.
3279 if (! VT.isVector())
3282 assert(ST->hasNEON() && "unexpected vector shift");
3285 switch (N->getOpcode()) {
3286 default: llvm_unreachable("unexpected shift opcode");
3289 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3290 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3291 DAG.getConstant(Cnt, MVT::i32));
3296 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3297 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3298 ARMISD::VSHRs : ARMISD::VSHRu);
3299 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3300 DAG.getConstant(Cnt, MVT::i32));
3306 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3307 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3308 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3309 const ARMSubtarget *ST) {
3310 SDValue N0 = N->getOperand(0);
3312 // Check for sign- and zero-extensions of vector extract operations of 8-
3313 // and 16-bit vector elements. NEON supports these directly. They are
3314 // handled during DAG combining because type legalization will promote them
3315 // to 32-bit types and it is messy to recognize the operations after that.
3316 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3317 SDValue Vec = N0.getOperand(0);
3318 SDValue Lane = N0.getOperand(1);
3319 EVT VT = N->getValueType(0);
3320 EVT EltVT = N0.getValueType();
3321 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3323 if (VT == MVT::i32 &&
3324 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3325 TLI.isTypeLegal(Vec.getValueType())) {
3328 switch (N->getOpcode()) {
3329 default: llvm_unreachable("unexpected opcode");
3330 case ISD::SIGN_EXTEND:
3331 Opc = ARMISD::VGETLANEs;
3333 case ISD::ZERO_EXTEND:
3334 case ISD::ANY_EXTEND:
3335 Opc = ARMISD::VGETLANEu;
3338 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3345 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3346 DAGCombinerInfo &DCI) const {
3347 switch (N->getOpcode()) {
3349 case ISD::ADD: return PerformADDCombine(N, DCI);
3350 case ISD::SUB: return PerformSUBCombine(N, DCI);
3351 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3352 case ISD::INTRINSIC_WO_CHAIN:
3353 return PerformIntrinsicCombine(N, DCI.DAG);
3357 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3358 case ISD::SIGN_EXTEND:
3359 case ISD::ZERO_EXTEND:
3360 case ISD::ANY_EXTEND:
3361 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3366 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3367 if (!Subtarget->hasV6Ops())
3368 // Pre-v6 does not support unaligned mem access.
3370 else if (!Subtarget->hasV6Ops()) {
3371 // v6 may or may not support unaligned mem access.
3372 if (!Subtarget->isTargetDarwin())
3376 switch (VT.getSimpleVT().SimpleTy) {
3383 // FIXME: VLD1 etc with standard alignment is legal.
3387 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3392 switch (VT.getSimpleVT().SimpleTy) {
3393 default: return false;
3408 if ((V & (Scale - 1)) != 0)
3411 return V == (V & ((1LL << 5) - 1));
3414 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3415 const ARMSubtarget *Subtarget) {
3422 switch (VT.getSimpleVT().SimpleTy) {
3423 default: return false;
3428 // + imm12 or - imm8
3430 return V == (V & ((1LL << 8) - 1));
3431 return V == (V & ((1LL << 12) - 1));
3434 // Same as ARM mode. FIXME: NEON?
3435 if (!Subtarget->hasVFP2())
3440 return V == (V & ((1LL << 8) - 1));
3444 /// isLegalAddressImmediate - Return true if the integer value can be used
3445 /// as the offset of the target addressing mode for load / store of the
3447 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3448 const ARMSubtarget *Subtarget) {
3455 if (Subtarget->isThumb1Only())
3456 return isLegalT1AddressImmediate(V, VT);
3457 else if (Subtarget->isThumb2())
3458 return isLegalT2AddressImmediate(V, VT, Subtarget);
3463 switch (VT.getSimpleVT().SimpleTy) {
3464 default: return false;
3469 return V == (V & ((1LL << 12) - 1));
3472 return V == (V & ((1LL << 8) - 1));
3475 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3480 return V == (V & ((1LL << 8) - 1));
3484 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3486 int Scale = AM.Scale;
3490 switch (VT.getSimpleVT().SimpleTy) {
3491 default: return false;
3500 return Scale == 2 || Scale == 4 || Scale == 8;
3503 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3507 // Note, we allow "void" uses (basically, uses that aren't loads or
3508 // stores), because arm allows folding a scale into many arithmetic
3509 // operations. This should be made more precise and revisited later.
3511 // Allow r << imm, but the imm has to be a multiple of two.
3512 if (Scale & 1) return false;
3513 return isPowerOf2_32(Scale);
3517 /// isLegalAddressingMode - Return true if the addressing mode represented
3518 /// by AM is legal for this target, for a load/store of the specified type.
3519 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3520 const Type *Ty) const {
3521 EVT VT = getValueType(Ty, true);
3522 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3525 // Can never fold addr of global into load/store.
3530 case 0: // no scale reg, must be "r+i" or "r", or "i".
3533 if (Subtarget->isThumb1Only())
3537 // ARM doesn't support any R+R*scale+imm addr modes.
3544 if (Subtarget->isThumb2())
3545 return isLegalT2ScaledAddressingMode(AM, VT);
3547 int Scale = AM.Scale;
3548 switch (VT.getSimpleVT().SimpleTy) {
3549 default: return false;
3553 if (Scale < 0) Scale = -Scale;
3557 return isPowerOf2_32(Scale & ~1);
3561 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3566 // Note, we allow "void" uses (basically, uses that aren't loads or
3567 // stores), because arm allows folding a scale into many arithmetic
3568 // operations. This should be made more precise and revisited later.
3570 // Allow r << imm, but the imm has to be a multiple of two.
3571 if (Scale & 1) return false;
3572 return isPowerOf2_32(Scale);
3579 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3580 bool isSEXTLoad, SDValue &Base,
3581 SDValue &Offset, bool &isInc,
3582 SelectionDAG &DAG) {
3583 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3586 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3588 Base = Ptr->getOperand(0);
3589 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3590 int RHSC = (int)RHS->getZExtValue();
3591 if (RHSC < 0 && RHSC > -256) {
3592 assert(Ptr->getOpcode() == ISD::ADD);
3594 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3598 isInc = (Ptr->getOpcode() == ISD::ADD);
3599 Offset = Ptr->getOperand(1);
3601 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3603 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3604 int RHSC = (int)RHS->getZExtValue();
3605 if (RHSC < 0 && RHSC > -0x1000) {
3606 assert(Ptr->getOpcode() == ISD::ADD);
3608 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3609 Base = Ptr->getOperand(0);
3614 if (Ptr->getOpcode() == ISD::ADD) {
3616 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3617 if (ShOpcVal != ARM_AM::no_shift) {
3618 Base = Ptr->getOperand(1);
3619 Offset = Ptr->getOperand(0);
3621 Base = Ptr->getOperand(0);
3622 Offset = Ptr->getOperand(1);
3627 isInc = (Ptr->getOpcode() == ISD::ADD);
3628 Base = Ptr->getOperand(0);
3629 Offset = Ptr->getOperand(1);
3633 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3637 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3638 bool isSEXTLoad, SDValue &Base,
3639 SDValue &Offset, bool &isInc,
3640 SelectionDAG &DAG) {
3641 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3644 Base = Ptr->getOperand(0);
3645 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3646 int RHSC = (int)RHS->getZExtValue();
3647 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3648 assert(Ptr->getOpcode() == ISD::ADD);
3650 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3652 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3653 isInc = Ptr->getOpcode() == ISD::ADD;
3654 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3662 /// getPreIndexedAddressParts - returns true by value, base pointer and
3663 /// offset pointer and addressing mode by reference if the node's address
3664 /// can be legally represented as pre-indexed load / store address.
3666 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3668 ISD::MemIndexedMode &AM,
3669 SelectionDAG &DAG) const {
3670 if (Subtarget->isThumb1Only())
3675 bool isSEXTLoad = false;
3676 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3677 Ptr = LD->getBasePtr();
3678 VT = LD->getMemoryVT();
3679 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3680 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3681 Ptr = ST->getBasePtr();
3682 VT = ST->getMemoryVT();
3687 bool isLegal = false;
3688 if (Subtarget->isThumb2())
3689 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3690 Offset, isInc, DAG);
3692 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3693 Offset, isInc, DAG);
3697 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3701 /// getPostIndexedAddressParts - returns true by value, base pointer and
3702 /// offset pointer and addressing mode by reference if this node can be
3703 /// combined with a load / store to form a post-indexed load / store.
3704 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3707 ISD::MemIndexedMode &AM,
3708 SelectionDAG &DAG) const {
3709 if (Subtarget->isThumb1Only())
3714 bool isSEXTLoad = false;
3715 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3716 VT = LD->getMemoryVT();
3717 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3718 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3719 VT = ST->getMemoryVT();
3724 bool isLegal = false;
3725 if (Subtarget->isThumb2())
3726 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3729 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3734 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3738 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3742 const SelectionDAG &DAG,
3743 unsigned Depth) const {
3744 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3745 switch (Op.getOpcode()) {
3747 case ARMISD::CMOV: {
3748 // Bits are known zero/one if known on the LHS and RHS.
3749 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3750 if (KnownZero == 0 && KnownOne == 0) return;
3752 APInt KnownZeroRHS, KnownOneRHS;
3753 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3754 KnownZeroRHS, KnownOneRHS, Depth+1);
3755 KnownZero &= KnownZeroRHS;
3756 KnownOne &= KnownOneRHS;
3762 //===----------------------------------------------------------------------===//
3763 // ARM Inline Assembly Support
3764 //===----------------------------------------------------------------------===//
3766 /// getConstraintType - Given a constraint letter, return the type of
3767 /// constraint it is for this target.
3768 ARMTargetLowering::ConstraintType
3769 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3770 if (Constraint.size() == 1) {
3771 switch (Constraint[0]) {
3773 case 'l': return C_RegisterClass;
3774 case 'w': return C_RegisterClass;
3777 return TargetLowering::getConstraintType(Constraint);
3780 std::pair<unsigned, const TargetRegisterClass*>
3781 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3783 if (Constraint.size() == 1) {
3784 // GCC RS6000 Constraint Letters
3785 switch (Constraint[0]) {
3787 if (Subtarget->isThumb1Only())
3788 return std::make_pair(0U, ARM::tGPRRegisterClass);
3790 return std::make_pair(0U, ARM::GPRRegisterClass);
3792 return std::make_pair(0U, ARM::GPRRegisterClass);
3795 return std::make_pair(0U, ARM::SPRRegisterClass);
3797 return std::make_pair(0U, ARM::DPRRegisterClass);
3801 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3804 std::vector<unsigned> ARMTargetLowering::
3805 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3807 if (Constraint.size() != 1)
3808 return std::vector<unsigned>();
3810 switch (Constraint[0]) { // GCC ARM Constraint Letters
3813 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3814 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3817 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3818 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3819 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3820 ARM::R12, ARM::LR, 0);
3823 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3824 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3825 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3826 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3827 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3828 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3829 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3830 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3832 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3833 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3834 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3835 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3839 return std::vector<unsigned>();
3842 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3843 /// vector. If it is invalid, don't add anything to Ops.
3844 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3847 std::vector<SDValue>&Ops,
3848 SelectionDAG &DAG) const {
3849 SDValue Result(0, 0);
3851 switch (Constraint) {
3853 case 'I': case 'J': case 'K': case 'L':
3854 case 'M': case 'N': case 'O':
3855 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3859 int64_t CVal64 = C->getSExtValue();
3860 int CVal = (int) CVal64;
3861 // None of these constraints allow values larger than 32 bits. Check
3862 // that the value fits in an int.
3866 switch (Constraint) {
3868 if (Subtarget->isThumb1Only()) {
3869 // This must be a constant between 0 and 255, for ADD
3871 if (CVal >= 0 && CVal <= 255)
3873 } else if (Subtarget->isThumb2()) {
3874 // A constant that can be used as an immediate value in a
3875 // data-processing instruction.
3876 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3879 // A constant that can be used as an immediate value in a
3880 // data-processing instruction.
3881 if (ARM_AM::getSOImmVal(CVal) != -1)
3887 if (Subtarget->isThumb()) { // FIXME thumb2
3888 // This must be a constant between -255 and -1, for negated ADD
3889 // immediates. This can be used in GCC with an "n" modifier that
3890 // prints the negated value, for use with SUB instructions. It is
3891 // not useful otherwise but is implemented for compatibility.
3892 if (CVal >= -255 && CVal <= -1)
3895 // This must be a constant between -4095 and 4095. It is not clear
3896 // what this constraint is intended for. Implemented for
3897 // compatibility with GCC.
3898 if (CVal >= -4095 && CVal <= 4095)
3904 if (Subtarget->isThumb1Only()) {
3905 // A 32-bit value where only one byte has a nonzero value. Exclude
3906 // zero to match GCC. This constraint is used by GCC internally for
3907 // constants that can be loaded with a move/shift combination.
3908 // It is not useful otherwise but is implemented for compatibility.
3909 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3911 } else if (Subtarget->isThumb2()) {
3912 // A constant whose bitwise inverse can be used as an immediate
3913 // value in a data-processing instruction. This can be used in GCC
3914 // with a "B" modifier that prints the inverted value, for use with
3915 // BIC and MVN instructions. It is not useful otherwise but is
3916 // implemented for compatibility.
3917 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3920 // A constant whose bitwise inverse can be used as an immediate
3921 // value in a data-processing instruction. This can be used in GCC
3922 // with a "B" modifier that prints the inverted value, for use with
3923 // BIC and MVN instructions. It is not useful otherwise but is
3924 // implemented for compatibility.
3925 if (ARM_AM::getSOImmVal(~CVal) != -1)
3931 if (Subtarget->isThumb1Only()) {
3932 // This must be a constant between -7 and 7,
3933 // for 3-operand ADD/SUB immediate instructions.
3934 if (CVal >= -7 && CVal < 7)
3936 } else if (Subtarget->isThumb2()) {
3937 // A constant whose negation can be used as an immediate value in a
3938 // data-processing instruction. This can be used in GCC with an "n"
3939 // modifier that prints the negated value, for use with SUB
3940 // instructions. It is not useful otherwise but is implemented for
3942 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3945 // A constant whose negation can be used as an immediate value in a
3946 // data-processing instruction. This can be used in GCC with an "n"
3947 // modifier that prints the negated value, for use with SUB
3948 // instructions. It is not useful otherwise but is implemented for
3950 if (ARM_AM::getSOImmVal(-CVal) != -1)
3956 if (Subtarget->isThumb()) { // FIXME thumb2
3957 // This must be a multiple of 4 between 0 and 1020, for
3958 // ADD sp + immediate.
3959 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3962 // A power of two or a constant between 0 and 32. This is used in
3963 // GCC for the shift amount on shifted register operands, but it is
3964 // useful in general for any shift amounts.
3965 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3971 if (Subtarget->isThumb()) { // FIXME thumb2
3972 // This must be a constant between 0 and 31, for shift amounts.
3973 if (CVal >= 0 && CVal <= 31)
3979 if (Subtarget->isThumb()) { // FIXME thumb2
3980 // This must be a multiple of 4 between -508 and 508, for
3981 // ADD/SUB sp = sp + immediate.
3982 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3987 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3991 if (Result.getNode()) {
3992 Ops.push_back(Result);
3995 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4000 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4001 // The ARM target isn't yet aware of offsets.
4005 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4006 APInt Imm = FPImm.bitcastToAPInt();
4007 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4008 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4009 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4011 // We can handle 4 bits of mantissa.
4012 // mantissa = (16+UInt(e:f:g:h))/16.
4013 if (Mantissa & 0x7ffff)
4016 if ((Mantissa & 0xf) != Mantissa)
4019 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4020 if (Exp < -3 || Exp > 4)
4022 Exp = ((Exp+3) & 0x7) ^ 4;
4024 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4027 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4028 APInt Imm = FPImm.bitcastToAPInt();
4029 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4030 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4031 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4033 // We can handle 4 bits of mantissa.
4034 // mantissa = (16+UInt(e:f:g:h))/16.
4035 if (Mantissa & 0xffffffffffffLL)
4038 if ((Mantissa & 0xf) != Mantissa)
4041 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4042 if (Exp < -3 || Exp > 4)
4044 Exp = ((Exp+3) & 0x7) ^ 4;
4046 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4049 /// isFPImmLegal - Returns true if the target can instruction select the
4050 /// specified FP immediate natively. If false, the legalizer will
4051 /// materialize the FP immediate as a load from a constant pool.
4052 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4053 if (!Subtarget->hasVFP3())
4056 return ARM::getVFPf32Imm(Imm) != -1;
4058 return ARM::getVFPf64Imm(Imm) != -1;