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);
367 // Use the default implementation.
368 setOperationAction(ISD::VASTART, MVT::Other, Custom);
369 setOperationAction(ISD::VAARG, MVT::Other, Expand);
370 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
371 setOperationAction(ISD::VAEND, MVT::Other, Expand);
372 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
373 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
374 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
375 // FIXME: Shouldn't need this, since no register is used, but the legalizer
376 // doesn't yet know how to not do that for SjLj.
377 setExceptionSelectorRegister(ARM::R0);
378 if (Subtarget->isThumb())
379 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
381 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
382 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
384 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
385 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
386 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
388 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
390 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
391 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
392 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
394 // We want to custom lower some of our intrinsics.
395 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
397 setOperationAction(ISD::SETCC, MVT::i32, Expand);
398 setOperationAction(ISD::SETCC, MVT::f32, Expand);
399 setOperationAction(ISD::SETCC, MVT::f64, Expand);
400 setOperationAction(ISD::SELECT, MVT::i32, Expand);
401 setOperationAction(ISD::SELECT, MVT::f32, Expand);
402 setOperationAction(ISD::SELECT, MVT::f64, Expand);
403 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
404 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
405 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
407 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
408 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
409 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
410 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
411 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
413 // We don't support sin/cos/fmod/copysign/pow
414 setOperationAction(ISD::FSIN, MVT::f64, Expand);
415 setOperationAction(ISD::FSIN, MVT::f32, Expand);
416 setOperationAction(ISD::FCOS, MVT::f32, Expand);
417 setOperationAction(ISD::FCOS, MVT::f64, Expand);
418 setOperationAction(ISD::FREM, MVT::f64, Expand);
419 setOperationAction(ISD::FREM, MVT::f32, Expand);
420 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
421 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
422 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
424 setOperationAction(ISD::FPOW, MVT::f64, Expand);
425 setOperationAction(ISD::FPOW, MVT::f32, Expand);
427 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
428 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
429 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
430 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
431 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
432 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
435 // We have target-specific dag combine patterns for the following nodes:
436 // ARMISD::FMRRD - No need to call setTargetDAGCombine
437 setTargetDAGCombine(ISD::ADD);
438 setTargetDAGCombine(ISD::SUB);
440 setStackPointerRegisterToSaveRestore(ARM::SP);
441 setSchedulingPreference(SchedulingForRegPressure);
443 // FIXME: If-converter should use instruction latency to determine
444 // profitability rather than relying on fixed limits.
445 if (Subtarget->getCPUString() == "generic") {
446 // Generic (and overly aggressive) if-conversion limits.
447 setIfCvtBlockSizeLimit(10);
448 setIfCvtDupBlockSizeLimit(2);
449 } else if (Subtarget->hasV6Ops()) {
450 setIfCvtBlockSizeLimit(2);
451 setIfCvtDupBlockSizeLimit(1);
453 setIfCvtBlockSizeLimit(3);
454 setIfCvtDupBlockSizeLimit(2);
457 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
458 // Do not enable CodePlacementOpt for now: it currently runs after the
459 // ARMConstantIslandPass and messes up branch relaxation and placement
460 // of constant islands.
461 // benefitFromCodePlacementOpt = true;
464 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
467 case ARMISD::Wrapper: return "ARMISD::Wrapper";
468 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
469 case ARMISD::CALL: return "ARMISD::CALL";
470 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
471 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
472 case ARMISD::tCALL: return "ARMISD::tCALL";
473 case ARMISD::BRCOND: return "ARMISD::BRCOND";
474 case ARMISD::BR_JT: return "ARMISD::BR_JT";
475 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
476 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
477 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
478 case ARMISD::CMP: return "ARMISD::CMP";
479 case ARMISD::CMPZ: return "ARMISD::CMPZ";
480 case ARMISD::CMPFP: return "ARMISD::CMPFP";
481 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
482 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
483 case ARMISD::CMOV: return "ARMISD::CMOV";
484 case ARMISD::CNEG: return "ARMISD::CNEG";
486 case ARMISD::FTOSI: return "ARMISD::FTOSI";
487 case ARMISD::FTOUI: return "ARMISD::FTOUI";
488 case ARMISD::SITOF: return "ARMISD::SITOF";
489 case ARMISD::UITOF: return "ARMISD::UITOF";
491 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
492 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
493 case ARMISD::RRX: return "ARMISD::RRX";
495 case ARMISD::FMRRD: return "ARMISD::FMRRD";
496 case ARMISD::FMDRR: return "ARMISD::FMDRR";
498 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
499 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
501 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
503 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
505 case ARMISD::VCEQ: return "ARMISD::VCEQ";
506 case ARMISD::VCGE: return "ARMISD::VCGE";
507 case ARMISD::VCGEU: return "ARMISD::VCGEU";
508 case ARMISD::VCGT: return "ARMISD::VCGT";
509 case ARMISD::VCGTU: return "ARMISD::VCGTU";
510 case ARMISD::VTST: return "ARMISD::VTST";
512 case ARMISD::VSHL: return "ARMISD::VSHL";
513 case ARMISD::VSHRs: return "ARMISD::VSHRs";
514 case ARMISD::VSHRu: return "ARMISD::VSHRu";
515 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
516 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
517 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
518 case ARMISD::VSHRN: return "ARMISD::VSHRN";
519 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
520 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
521 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
522 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
523 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
524 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
525 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
526 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
527 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
528 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
529 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
530 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
531 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
532 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
533 case ARMISD::VDUP: return "ARMISD::VDUP";
534 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
535 case ARMISD::VEXT: return "ARMISD::VEXT";
536 case ARMISD::VREV64: return "ARMISD::VREV64";
537 case ARMISD::VREV32: return "ARMISD::VREV32";
538 case ARMISD::VREV16: return "ARMISD::VREV16";
539 case ARMISD::VZIP: return "ARMISD::VZIP";
540 case ARMISD::VUZP: return "ARMISD::VUZP";
541 case ARMISD::VTRN: return "ARMISD::VTRN";
545 /// getFunctionAlignment - Return the Log2 alignment of this function.
546 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
547 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
550 //===----------------------------------------------------------------------===//
552 //===----------------------------------------------------------------------===//
554 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
555 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
557 default: llvm_unreachable("Unknown condition code!");
558 case ISD::SETNE: return ARMCC::NE;
559 case ISD::SETEQ: return ARMCC::EQ;
560 case ISD::SETGT: return ARMCC::GT;
561 case ISD::SETGE: return ARMCC::GE;
562 case ISD::SETLT: return ARMCC::LT;
563 case ISD::SETLE: return ARMCC::LE;
564 case ISD::SETUGT: return ARMCC::HI;
565 case ISD::SETUGE: return ARMCC::HS;
566 case ISD::SETULT: return ARMCC::LO;
567 case ISD::SETULE: return ARMCC::LS;
571 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
572 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
573 ARMCC::CondCodes &CondCode2) {
574 CondCode2 = ARMCC::AL;
576 default: llvm_unreachable("Unknown FP condition!");
578 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
580 case ISD::SETOGT: CondCode = ARMCC::GT; break;
582 case ISD::SETOGE: CondCode = ARMCC::GE; break;
583 case ISD::SETOLT: CondCode = ARMCC::MI; break;
584 case ISD::SETOLE: CondCode = ARMCC::LS; break;
585 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
586 case ISD::SETO: CondCode = ARMCC::VC; break;
587 case ISD::SETUO: CondCode = ARMCC::VS; break;
588 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
589 case ISD::SETUGT: CondCode = ARMCC::HI; break;
590 case ISD::SETUGE: CondCode = ARMCC::PL; break;
592 case ISD::SETULT: CondCode = ARMCC::LT; break;
594 case ISD::SETULE: CondCode = ARMCC::LE; break;
596 case ISD::SETUNE: CondCode = ARMCC::NE; break;
600 //===----------------------------------------------------------------------===//
601 // Calling Convention Implementation
602 //===----------------------------------------------------------------------===//
604 #include "ARMGenCallingConv.inc"
606 // APCS f64 is in register pairs, possibly split to stack
607 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
608 CCValAssign::LocInfo &LocInfo,
609 CCState &State, bool CanFail) {
610 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
612 // Try to get the first register.
613 if (unsigned Reg = State.AllocateReg(RegList, 4))
614 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
616 // For the 2nd half of a v2f64, do not fail.
620 // Put the whole thing on the stack.
621 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
622 State.AllocateStack(8, 4),
627 // Try to get the second register.
628 if (unsigned Reg = State.AllocateReg(RegList, 4))
629 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
631 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
632 State.AllocateStack(4, 4),
637 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
638 CCValAssign::LocInfo &LocInfo,
639 ISD::ArgFlagsTy &ArgFlags,
641 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
643 if (LocVT == MVT::v2f64 &&
644 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
646 return true; // we handled it
649 // AAPCS f64 is in aligned register pairs
650 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
651 CCValAssign::LocInfo &LocInfo,
652 CCState &State, bool CanFail) {
653 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
654 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
656 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
658 // For the 2nd half of a v2f64, do not just fail.
662 // Put the whole thing on the stack.
663 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
664 State.AllocateStack(8, 8),
670 for (i = 0; i < 2; ++i)
671 if (HiRegList[i] == Reg)
674 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
675 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
680 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
681 CCValAssign::LocInfo &LocInfo,
682 ISD::ArgFlagsTy &ArgFlags,
684 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
686 if (LocVT == MVT::v2f64 &&
687 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
689 return true; // we handled it
692 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
693 CCValAssign::LocInfo &LocInfo, CCState &State) {
694 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
695 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
697 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
699 return false; // we didn't handle it
702 for (i = 0; i < 2; ++i)
703 if (HiRegList[i] == Reg)
706 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
707 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
712 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
713 CCValAssign::LocInfo &LocInfo,
714 ISD::ArgFlagsTy &ArgFlags,
716 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
718 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
720 return true; // we handled it
723 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
724 CCValAssign::LocInfo &LocInfo,
725 ISD::ArgFlagsTy &ArgFlags,
727 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
731 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
732 /// given CallingConvention value.
733 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
735 bool isVarArg) const {
738 llvm_unreachable("Unsupported calling convention");
740 case CallingConv::Fast:
741 // Use target triple & subtarget features to do actual dispatch.
742 if (Subtarget->isAAPCS_ABI()) {
743 if (Subtarget->hasVFP2() &&
744 FloatABIType == FloatABI::Hard && !isVarArg)
745 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
747 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
749 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
750 case CallingConv::ARM_AAPCS_VFP:
751 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
752 case CallingConv::ARM_AAPCS:
753 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
754 case CallingConv::ARM_APCS:
755 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
759 /// LowerCallResult - Lower the result values of a call into the
760 /// appropriate copies out of appropriate physical registers.
762 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
763 CallingConv::ID CallConv, bool isVarArg,
764 const SmallVectorImpl<ISD::InputArg> &Ins,
765 DebugLoc dl, SelectionDAG &DAG,
766 SmallVectorImpl<SDValue> &InVals) {
768 // Assign locations to each value returned by this call.
769 SmallVector<CCValAssign, 16> RVLocs;
770 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
771 RVLocs, *DAG.getContext());
772 CCInfo.AnalyzeCallResult(Ins,
773 CCAssignFnForNode(CallConv, /* Return*/ true,
776 // Copy all of the result registers out of their specified physreg.
777 for (unsigned i = 0; i != RVLocs.size(); ++i) {
778 CCValAssign VA = RVLocs[i];
781 if (VA.needsCustom()) {
782 // Handle f64 or half of a v2f64.
783 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
785 Chain = Lo.getValue(1);
786 InFlag = Lo.getValue(2);
787 VA = RVLocs[++i]; // skip ahead to next loc
788 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
790 Chain = Hi.getValue(1);
791 InFlag = Hi.getValue(2);
792 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
794 if (VA.getLocVT() == MVT::v2f64) {
795 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
796 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
797 DAG.getConstant(0, MVT::i32));
799 VA = RVLocs[++i]; // skip ahead to next loc
800 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
801 Chain = Lo.getValue(1);
802 InFlag = Lo.getValue(2);
803 VA = RVLocs[++i]; // skip ahead to next loc
804 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
805 Chain = Hi.getValue(1);
806 InFlag = Hi.getValue(2);
807 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
808 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
809 DAG.getConstant(1, MVT::i32));
812 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
814 Chain = Val.getValue(1);
815 InFlag = Val.getValue(2);
818 switch (VA.getLocInfo()) {
819 default: llvm_unreachable("Unknown loc info!");
820 case CCValAssign::Full: break;
821 case CCValAssign::BCvt:
822 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
826 InVals.push_back(Val);
832 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
833 /// by "Src" to address "Dst" of size "Size". Alignment information is
834 /// specified by the specific parameter attribute. The copy will be passed as
835 /// a byval function parameter.
836 /// Sometimes what we are copying is the end of a larger object, the part that
837 /// does not fit in registers.
839 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
840 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
842 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
843 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
844 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
847 /// LowerMemOpCallTo - Store the argument to the stack.
849 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
850 SDValue StackPtr, SDValue Arg,
851 DebugLoc dl, SelectionDAG &DAG,
852 const CCValAssign &VA,
853 ISD::ArgFlagsTy Flags) {
854 unsigned LocMemOffset = VA.getLocMemOffset();
855 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
856 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
857 if (Flags.isByVal()) {
858 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
860 return DAG.getStore(Chain, dl, Arg, PtrOff,
861 PseudoSourceValue::getStack(), LocMemOffset);
864 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
865 SDValue Chain, SDValue &Arg,
866 RegsToPassVector &RegsToPass,
867 CCValAssign &VA, CCValAssign &NextVA,
869 SmallVector<SDValue, 8> &MemOpChains,
870 ISD::ArgFlagsTy Flags) {
872 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
873 DAG.getVTList(MVT::i32, MVT::i32), Arg);
874 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
876 if (NextVA.isRegLoc())
877 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
879 assert(NextVA.isMemLoc());
880 if (StackPtr.getNode() == 0)
881 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
883 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
889 /// LowerCall - Lowering a call into a callseq_start <-
890 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
893 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
894 CallingConv::ID CallConv, bool isVarArg,
896 const SmallVectorImpl<ISD::OutputArg> &Outs,
897 const SmallVectorImpl<ISD::InputArg> &Ins,
898 DebugLoc dl, SelectionDAG &DAG,
899 SmallVectorImpl<SDValue> &InVals) {
901 // Analyze operands of the call, assigning locations to each operand.
902 SmallVector<CCValAssign, 16> ArgLocs;
903 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
905 CCInfo.AnalyzeCallOperands(Outs,
906 CCAssignFnForNode(CallConv, /* Return*/ false,
909 // Get a count of how many bytes are to be pushed on the stack.
910 unsigned NumBytes = CCInfo.getNextStackOffset();
912 // Adjust the stack pointer for the new arguments...
913 // These operations are automatically eliminated by the prolog/epilog pass
914 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
916 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
918 RegsToPassVector RegsToPass;
919 SmallVector<SDValue, 8> MemOpChains;
921 // Walk the register/memloc assignments, inserting copies/loads. In the case
922 // of tail call optimization, arguments are handled later.
923 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
926 CCValAssign &VA = ArgLocs[i];
927 SDValue Arg = Outs[realArgIdx].Val;
928 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
930 // Promote the value if needed.
931 switch (VA.getLocInfo()) {
932 default: llvm_unreachable("Unknown loc info!");
933 case CCValAssign::Full: break;
934 case CCValAssign::SExt:
935 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
937 case CCValAssign::ZExt:
938 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
940 case CCValAssign::AExt:
941 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
943 case CCValAssign::BCvt:
944 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
948 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
949 if (VA.needsCustom()) {
950 if (VA.getLocVT() == MVT::v2f64) {
951 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
952 DAG.getConstant(0, MVT::i32));
953 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
954 DAG.getConstant(1, MVT::i32));
956 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
957 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
959 VA = ArgLocs[++i]; // skip ahead to next loc
961 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
962 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
964 assert(VA.isMemLoc());
965 if (StackPtr.getNode() == 0)
966 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
968 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
969 dl, DAG, VA, Flags));
972 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
973 StackPtr, MemOpChains, Flags);
975 } else if (VA.isRegLoc()) {
976 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
978 assert(VA.isMemLoc());
979 if (StackPtr.getNode() == 0)
980 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
982 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
983 dl, DAG, VA, Flags));
987 if (!MemOpChains.empty())
988 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
989 &MemOpChains[0], MemOpChains.size());
991 // Build a sequence of copy-to-reg nodes chained together with token chain
992 // and flag operands which copy the outgoing args into the appropriate regs.
994 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
995 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
996 RegsToPass[i].second, InFlag);
997 InFlag = Chain.getValue(1);
1000 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1001 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1002 // node so that legalize doesn't hack it.
1003 bool isDirect = false;
1004 bool isARMFunc = false;
1005 bool isLocalARMFunc = false;
1006 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1007 GlobalValue *GV = G->getGlobal();
1009 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1010 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1011 getTargetMachine().getRelocationModel() != Reloc::Static;
1012 isARMFunc = !Subtarget->isThumb() || isStub;
1013 // ARM call to a local ARM function is predicable.
1014 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1015 // tBX takes a register source operand.
1016 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1017 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1020 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1021 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1022 Callee = DAG.getLoad(getPointerTy(), dl,
1023 DAG.getEntryNode(), CPAddr, NULL, 0);
1024 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1025 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1026 getPointerTy(), Callee, PICLabel);
1028 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1029 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1031 bool isStub = Subtarget->isTargetDarwin() &&
1032 getTargetMachine().getRelocationModel() != Reloc::Static;
1033 isARMFunc = !Subtarget->isThumb() || isStub;
1034 // tBX takes a register source operand.
1035 const char *Sym = S->getSymbol();
1036 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1037 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1038 Sym, ARMPCLabelIndex, 4);
1039 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1040 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1041 Callee = DAG.getLoad(getPointerTy(), dl,
1042 DAG.getEntryNode(), CPAddr, NULL, 0);
1043 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1044 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1045 getPointerTy(), Callee, PICLabel);
1047 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1050 // FIXME: handle tail calls differently.
1052 if (Subtarget->isThumb()) {
1053 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1054 CallOpc = ARMISD::CALL_NOLINK;
1056 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1058 CallOpc = (isDirect || Subtarget->hasV5TOps())
1059 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1060 : ARMISD::CALL_NOLINK;
1062 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1063 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1064 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1065 InFlag = Chain.getValue(1);
1068 std::vector<SDValue> Ops;
1069 Ops.push_back(Chain);
1070 Ops.push_back(Callee);
1072 // Add argument registers to the end of the list so that they are known live
1074 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1075 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1076 RegsToPass[i].second.getValueType()));
1078 if (InFlag.getNode())
1079 Ops.push_back(InFlag);
1080 // Returns a chain and a flag for retval copy to use.
1081 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1082 &Ops[0], Ops.size());
1083 InFlag = Chain.getValue(1);
1085 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1086 DAG.getIntPtrConstant(0, true), InFlag);
1088 InFlag = Chain.getValue(1);
1090 // Handle result values, copying them out of physregs into vregs that we
1092 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1097 ARMTargetLowering::LowerReturn(SDValue Chain,
1098 CallingConv::ID CallConv, bool isVarArg,
1099 const SmallVectorImpl<ISD::OutputArg> &Outs,
1100 DebugLoc dl, SelectionDAG &DAG) {
1102 // CCValAssign - represent the assignment of the return value to a location.
1103 SmallVector<CCValAssign, 16> RVLocs;
1105 // CCState - Info about the registers and stack slots.
1106 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1109 // Analyze outgoing return values.
1110 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1113 // If this is the first return lowered for this function, add
1114 // the regs to the liveout set for the function.
1115 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1116 for (unsigned i = 0; i != RVLocs.size(); ++i)
1117 if (RVLocs[i].isRegLoc())
1118 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1123 // Copy the result values into the output registers.
1124 for (unsigned i = 0, realRVLocIdx = 0;
1126 ++i, ++realRVLocIdx) {
1127 CCValAssign &VA = RVLocs[i];
1128 assert(VA.isRegLoc() && "Can only return in registers!");
1130 SDValue Arg = Outs[realRVLocIdx].Val;
1132 switch (VA.getLocInfo()) {
1133 default: llvm_unreachable("Unknown loc info!");
1134 case CCValAssign::Full: break;
1135 case CCValAssign::BCvt:
1136 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1140 if (VA.needsCustom()) {
1141 if (VA.getLocVT() == MVT::v2f64) {
1142 // Extract the first half and return it in two registers.
1143 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1144 DAG.getConstant(0, MVT::i32));
1145 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1146 DAG.getVTList(MVT::i32, MVT::i32), Half);
1148 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1149 Flag = Chain.getValue(1);
1150 VA = RVLocs[++i]; // skip ahead to next loc
1151 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1152 HalfGPRs.getValue(1), Flag);
1153 Flag = Chain.getValue(1);
1154 VA = RVLocs[++i]; // skip ahead to next loc
1156 // Extract the 2nd half and fall through to handle it as an f64 value.
1157 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1158 DAG.getConstant(1, MVT::i32));
1160 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1162 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1163 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1164 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1165 Flag = Chain.getValue(1);
1166 VA = RVLocs[++i]; // skip ahead to next loc
1167 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1170 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1172 // Guarantee that all emitted copies are
1173 // stuck together, avoiding something bad.
1174 Flag = Chain.getValue(1);
1179 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1181 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1186 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1187 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1188 // one of the above mentioned nodes. It has to be wrapped because otherwise
1189 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1190 // be used to form addressing mode. These wrapped nodes will be selected
1192 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1193 EVT PtrVT = Op.getValueType();
1194 // FIXME there is no actual debug info here
1195 DebugLoc dl = Op.getDebugLoc();
1196 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1198 if (CP->isMachineConstantPoolEntry())
1199 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1200 CP->getAlignment());
1202 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1203 CP->getAlignment());
1204 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1207 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1209 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1210 SelectionDAG &DAG) {
1211 DebugLoc dl = GA->getDebugLoc();
1212 EVT PtrVT = getPointerTy();
1213 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1214 ARMConstantPoolValue *CPV =
1215 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1216 ARMCP::CPValue, PCAdj, "tlsgd", true);
1217 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1218 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1219 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1220 SDValue Chain = Argument.getValue(1);
1222 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1223 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1225 // call __tls_get_addr.
1228 Entry.Node = Argument;
1229 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1230 Args.push_back(Entry);
1231 // FIXME: is there useful debug info available here?
1232 std::pair<SDValue, SDValue> CallResult =
1233 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1234 false, false, false, false,
1235 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1236 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1237 return CallResult.first;
1240 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1241 // "local exec" model.
1243 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1244 SelectionDAG &DAG) {
1245 GlobalValue *GV = GA->getGlobal();
1246 DebugLoc dl = GA->getDebugLoc();
1248 SDValue Chain = DAG.getEntryNode();
1249 EVT PtrVT = getPointerTy();
1250 // Get the Thread Pointer
1251 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1253 if (GV->isDeclaration()) {
1254 // initial exec model
1255 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1256 ARMConstantPoolValue *CPV =
1257 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1258 ARMCP::CPValue, PCAdj, "gottpoff", true);
1259 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1260 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1261 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1262 Chain = Offset.getValue(1);
1264 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1265 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1267 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1270 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1271 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1272 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1273 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1276 // The address of the thread local variable is the add of the thread
1277 // pointer with the offset of the variable.
1278 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1282 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1283 // TODO: implement the "local dynamic" model
1284 assert(Subtarget->isTargetELF() &&
1285 "TLS not implemented for non-ELF targets");
1286 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1287 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1288 // otherwise use the "Local Exec" TLS Model
1289 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1290 return LowerToTLSGeneralDynamicModel(GA, DAG);
1292 return LowerToTLSExecModels(GA, DAG);
1295 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1296 SelectionDAG &DAG) {
1297 EVT PtrVT = getPointerTy();
1298 DebugLoc dl = Op.getDebugLoc();
1299 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1300 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1301 if (RelocM == Reloc::PIC_) {
1302 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1303 ARMConstantPoolValue *CPV =
1304 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1305 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1306 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1307 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1309 PseudoSourceValue::getConstantPool(), 0);
1310 SDValue Chain = Result.getValue(1);
1311 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1312 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1314 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1315 PseudoSourceValue::getGOT(), 0);
1318 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1319 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1320 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1321 PseudoSourceValue::getConstantPool(), 0);
1325 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1326 SelectionDAG &DAG) {
1327 EVT PtrVT = getPointerTy();
1328 DebugLoc dl = Op.getDebugLoc();
1329 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1330 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1332 if (RelocM == Reloc::Static)
1333 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1335 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1336 ARMConstantPoolValue *CPV =
1337 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1338 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1340 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1342 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1343 SDValue Chain = Result.getValue(1);
1345 if (RelocM == Reloc::PIC_) {
1346 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1347 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1350 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1351 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1356 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1358 assert(Subtarget->isTargetELF() &&
1359 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1360 EVT PtrVT = getPointerTy();
1361 DebugLoc dl = Op.getDebugLoc();
1362 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1363 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1364 "_GLOBAL_OFFSET_TABLE_",
1365 ARMPCLabelIndex, PCAdj);
1366 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1367 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1368 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1369 PseudoSourceValue::getConstantPool(), 0);
1370 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1371 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1375 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1376 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1377 DebugLoc dl = Op.getDebugLoc();
1379 default: return SDValue(); // Don't custom lower most intrinsics.
1380 case Intrinsic::arm_thread_pointer: {
1381 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1382 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1384 case Intrinsic::eh_sjlj_lsda: {
1385 MachineFunction &MF = DAG.getMachineFunction();
1386 EVT PtrVT = getPointerTy();
1387 DebugLoc dl = Op.getDebugLoc();
1388 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1390 unsigned PCAdj = (RelocM != Reloc::PIC_)
1391 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1392 ARMConstantPoolValue *CPV =
1393 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1394 ARMCP::CPLSDA, PCAdj);
1395 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1396 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1398 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1399 SDValue Chain = Result.getValue(1);
1401 if (RelocM == Reloc::PIC_) {
1402 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1403 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1407 case Intrinsic::eh_sjlj_setjmp:
1408 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1412 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1413 unsigned VarArgsFrameIndex) {
1414 // vastart just stores the address of the VarArgsFrameIndex slot into the
1415 // memory location argument.
1416 DebugLoc dl = Op.getDebugLoc();
1417 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1418 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1419 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1420 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1424 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1425 SDNode *Node = Op.getNode();
1426 DebugLoc dl = Node->getDebugLoc();
1427 EVT VT = Node->getValueType(0);
1428 SDValue Chain = Op.getOperand(0);
1429 SDValue Size = Op.getOperand(1);
1430 SDValue Align = Op.getOperand(2);
1432 // Chain the dynamic stack allocation so that it doesn't modify the stack
1433 // pointer when other instructions are using the stack.
1434 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1436 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1437 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1438 if (AlignVal > StackAlign)
1439 // Do this now since selection pass cannot introduce new target
1440 // independent node.
1441 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1443 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1444 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1445 // do even more horrible hack later.
1446 MachineFunction &MF = DAG.getMachineFunction();
1447 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1448 if (AFI->isThumb1OnlyFunction()) {
1450 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1452 uint32_t Val = C->getZExtValue();
1453 if (Val <= 508 && ((Val & 3) == 0))
1457 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1460 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1461 SDValue Ops1[] = { Chain, Size, Align };
1462 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1463 Chain = Res.getValue(1);
1464 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1465 DAG.getIntPtrConstant(0, true), SDValue());
1466 SDValue Ops2[] = { Res, Chain };
1467 return DAG.getMergeValues(Ops2, 2, dl);
1471 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1472 SDValue &Root, SelectionDAG &DAG,
1474 MachineFunction &MF = DAG.getMachineFunction();
1475 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1477 TargetRegisterClass *RC;
1478 if (AFI->isThumb1OnlyFunction())
1479 RC = ARM::tGPRRegisterClass;
1481 RC = ARM::GPRRegisterClass;
1483 // Transform the arguments stored in physical registers into virtual ones.
1484 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1485 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1488 if (NextVA.isMemLoc()) {
1489 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1490 MachineFrameInfo *MFI = MF.getFrameInfo();
1491 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1493 // Create load node to retrieve arguments from the stack.
1494 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1495 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1497 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1498 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1501 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1505 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1506 CallingConv::ID CallConv, bool isVarArg,
1507 const SmallVectorImpl<ISD::InputArg>
1509 DebugLoc dl, SelectionDAG &DAG,
1510 SmallVectorImpl<SDValue> &InVals) {
1512 MachineFunction &MF = DAG.getMachineFunction();
1513 MachineFrameInfo *MFI = MF.getFrameInfo();
1515 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1517 // Assign locations to all of the incoming arguments.
1518 SmallVector<CCValAssign, 16> ArgLocs;
1519 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1521 CCInfo.AnalyzeFormalArguments(Ins,
1522 CCAssignFnForNode(CallConv, /* Return*/ false,
1525 SmallVector<SDValue, 16> ArgValues;
1527 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1528 CCValAssign &VA = ArgLocs[i];
1530 // Arguments stored in registers.
1531 if (VA.isRegLoc()) {
1532 EVT RegVT = VA.getLocVT();
1535 if (VA.needsCustom()) {
1536 // f64 and vector types are split up into multiple registers or
1537 // combinations of registers and stack slots.
1540 if (VA.getLocVT() == MVT::v2f64) {
1541 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1543 VA = ArgLocs[++i]; // skip ahead to next loc
1544 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1546 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1547 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1548 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1549 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1550 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1552 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1555 TargetRegisterClass *RC;
1557 if (RegVT == MVT::f32)
1558 RC = ARM::SPRRegisterClass;
1559 else if (RegVT == MVT::f64)
1560 RC = ARM::DPRRegisterClass;
1561 else if (RegVT == MVT::v2f64)
1562 RC = ARM::QPRRegisterClass;
1563 else if (RegVT == MVT::i32)
1564 RC = (AFI->isThumb1OnlyFunction() ?
1565 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1567 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1569 // Transform the arguments in physical registers into virtual ones.
1570 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1571 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1574 // If this is an 8 or 16-bit value, it is really passed promoted
1575 // to 32 bits. Insert an assert[sz]ext to capture this, then
1576 // truncate to the right size.
1577 switch (VA.getLocInfo()) {
1578 default: llvm_unreachable("Unknown loc info!");
1579 case CCValAssign::Full: break;
1580 case CCValAssign::BCvt:
1581 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1583 case CCValAssign::SExt:
1584 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1585 DAG.getValueType(VA.getValVT()));
1586 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1588 case CCValAssign::ZExt:
1589 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1590 DAG.getValueType(VA.getValVT()));
1591 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1595 InVals.push_back(ArgValue);
1597 } else { // VA.isRegLoc()
1600 assert(VA.isMemLoc());
1601 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1603 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1604 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1606 // Create load nodes to retrieve arguments from the stack.
1607 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1608 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1614 static const unsigned GPRArgRegs[] = {
1615 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1618 unsigned NumGPRs = CCInfo.getFirstUnallocated
1619 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1621 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1622 unsigned VARegSize = (4 - NumGPRs) * 4;
1623 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1624 unsigned ArgOffset = 0;
1625 if (VARegSaveSize) {
1626 // If this function is vararg, store any remaining integer argument regs
1627 // to their spots on the stack so that they may be loaded by deferencing
1628 // the result of va_next.
1629 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1630 ArgOffset = CCInfo.getNextStackOffset();
1631 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1632 VARegSaveSize - VARegSize);
1633 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1635 SmallVector<SDValue, 4> MemOps;
1636 for (; NumGPRs < 4; ++NumGPRs) {
1637 TargetRegisterClass *RC;
1638 if (AFI->isThumb1OnlyFunction())
1639 RC = ARM::tGPRRegisterClass;
1641 RC = ARM::GPRRegisterClass;
1643 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1644 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1645 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1646 MemOps.push_back(Store);
1647 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1648 DAG.getConstant(4, getPointerTy()));
1650 if (!MemOps.empty())
1651 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1652 &MemOps[0], MemOps.size());
1654 // This will point to the next argument passed via stack.
1655 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1661 /// isFloatingPointZero - Return true if this is +0.0.
1662 static bool isFloatingPointZero(SDValue Op) {
1663 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1664 return CFP->getValueAPF().isPosZero();
1665 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1666 // Maybe this has already been legalized into the constant pool?
1667 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1668 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1669 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1670 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1671 return CFP->getValueAPF().isPosZero();
1677 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1678 return ( isThumb1Only && (C & ~255U) == 0) ||
1679 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1682 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1683 /// the given operands.
1684 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1685 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1687 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1688 unsigned C = RHSC->getZExtValue();
1689 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1690 // Constant does not fit, try adjusting it by one?
1695 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1696 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1697 RHS = DAG.getConstant(C-1, MVT::i32);
1702 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1703 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1704 RHS = DAG.getConstant(C-1, MVT::i32);
1709 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1710 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1711 RHS = DAG.getConstant(C+1, MVT::i32);
1716 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1717 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1718 RHS = DAG.getConstant(C+1, MVT::i32);
1725 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1726 ARMISD::NodeType CompareType;
1729 CompareType = ARMISD::CMP;
1734 CompareType = ARMISD::CMPZ;
1737 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1738 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1741 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1742 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1745 if (!isFloatingPointZero(RHS))
1746 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1748 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1749 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1752 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1753 const ARMSubtarget *ST) {
1754 EVT VT = Op.getValueType();
1755 SDValue LHS = Op.getOperand(0);
1756 SDValue RHS = Op.getOperand(1);
1757 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1758 SDValue TrueVal = Op.getOperand(2);
1759 SDValue FalseVal = Op.getOperand(3);
1760 DebugLoc dl = Op.getDebugLoc();
1762 if (LHS.getValueType() == MVT::i32) {
1764 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1765 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1766 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1769 ARMCC::CondCodes CondCode, CondCode2;
1770 FPCCToARMCC(CC, CondCode, CondCode2);
1772 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1773 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1774 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1775 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1777 if (CondCode2 != ARMCC::AL) {
1778 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1779 // FIXME: Needs another CMP because flag can have but one use.
1780 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1781 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1782 Result, TrueVal, ARMCC2, CCR, Cmp2);
1787 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1788 const ARMSubtarget *ST) {
1789 SDValue Chain = Op.getOperand(0);
1790 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1791 SDValue LHS = Op.getOperand(2);
1792 SDValue RHS = Op.getOperand(3);
1793 SDValue Dest = Op.getOperand(4);
1794 DebugLoc dl = Op.getDebugLoc();
1796 if (LHS.getValueType() == MVT::i32) {
1798 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1799 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1800 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1801 Chain, Dest, ARMCC, CCR,Cmp);
1804 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1805 ARMCC::CondCodes CondCode, CondCode2;
1806 FPCCToARMCC(CC, CondCode, CondCode2);
1808 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1809 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1810 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1811 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1812 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1813 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1814 if (CondCode2 != ARMCC::AL) {
1815 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1816 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1817 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1822 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1823 SDValue Chain = Op.getOperand(0);
1824 SDValue Table = Op.getOperand(1);
1825 SDValue Index = Op.getOperand(2);
1826 DebugLoc dl = Op.getDebugLoc();
1828 EVT PTy = getPointerTy();
1829 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1830 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1831 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1832 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1833 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1834 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1835 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1836 if (Subtarget->isThumb2()) {
1837 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1838 // which does another jump to the destination. This also makes it easier
1839 // to translate it to TBB / TBH later.
1840 // FIXME: This might not work if the function is extremely large.
1841 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1842 Addr, Op.getOperand(2), JTI, UId);
1844 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1845 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1846 Chain = Addr.getValue(1);
1847 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1848 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1850 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1851 Chain = Addr.getValue(1);
1852 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1856 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1857 DebugLoc dl = Op.getDebugLoc();
1859 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1860 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1861 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1864 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1865 EVT VT = Op.getValueType();
1866 DebugLoc dl = Op.getDebugLoc();
1868 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1870 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1871 return DAG.getNode(Opc, dl, VT, Op);
1874 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1875 // Implement fcopysign with a fabs and a conditional fneg.
1876 SDValue Tmp0 = Op.getOperand(0);
1877 SDValue Tmp1 = Op.getOperand(1);
1878 DebugLoc dl = Op.getDebugLoc();
1879 EVT VT = Op.getValueType();
1880 EVT SrcVT = Tmp1.getValueType();
1881 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1882 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1883 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1884 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1885 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1888 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1889 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1890 MFI->setFrameAddressIsTaken(true);
1891 EVT VT = Op.getValueType();
1892 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1893 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1894 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1895 ? ARM::R7 : ARM::R11;
1896 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1898 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1903 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1905 SDValue Dst, SDValue Src,
1906 SDValue Size, unsigned Align,
1908 const Value *DstSV, uint64_t DstSVOff,
1909 const Value *SrcSV, uint64_t SrcSVOff){
1910 // Do repeated 4-byte loads and stores. To be improved.
1911 // This requires 4-byte alignment.
1912 if ((Align & 3) != 0)
1914 // This requires the copy size to be a constant, preferrably
1915 // within a subtarget-specific limit.
1916 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1919 uint64_t SizeVal = ConstantSize->getZExtValue();
1920 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1923 unsigned BytesLeft = SizeVal & 3;
1924 unsigned NumMemOps = SizeVal >> 2;
1925 unsigned EmittedNumMemOps = 0;
1927 unsigned VTSize = 4;
1929 const unsigned MAX_LOADS_IN_LDM = 6;
1930 SDValue TFOps[MAX_LOADS_IN_LDM];
1931 SDValue Loads[MAX_LOADS_IN_LDM];
1932 uint64_t SrcOff = 0, DstOff = 0;
1934 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1935 // same number of stores. The loads and stores will get combined into
1936 // ldm/stm later on.
1937 while (EmittedNumMemOps < NumMemOps) {
1939 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1940 Loads[i] = DAG.getLoad(VT, dl, Chain,
1941 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1942 DAG.getConstant(SrcOff, MVT::i32)),
1943 SrcSV, SrcSVOff + SrcOff);
1944 TFOps[i] = Loads[i].getValue(1);
1947 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1950 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1951 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1952 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1953 DAG.getConstant(DstOff, MVT::i32)),
1954 DstSV, DstSVOff + DstOff);
1957 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1959 EmittedNumMemOps += i;
1965 // Issue loads / stores for the trailing (1 - 3) bytes.
1966 unsigned BytesLeftSave = BytesLeft;
1969 if (BytesLeft >= 2) {
1977 Loads[i] = DAG.getLoad(VT, dl, Chain,
1978 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1979 DAG.getConstant(SrcOff, MVT::i32)),
1980 SrcSV, SrcSVOff + SrcOff);
1981 TFOps[i] = Loads[i].getValue(1);
1984 BytesLeft -= VTSize;
1986 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1989 BytesLeft = BytesLeftSave;
1991 if (BytesLeft >= 2) {
1999 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2000 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2001 DAG.getConstant(DstOff, MVT::i32)),
2002 DstSV, DstSVOff + DstOff);
2005 BytesLeft -= VTSize;
2007 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2010 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2011 SDValue Op = N->getOperand(0);
2012 DebugLoc dl = N->getDebugLoc();
2013 if (N->getValueType(0) == MVT::f64) {
2014 // Turn i64->f64 into FMDRR.
2015 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2016 DAG.getConstant(0, MVT::i32));
2017 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2018 DAG.getConstant(1, MVT::i32));
2019 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2022 // Turn f64->i64 into FMRRD.
2023 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2024 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2026 // Merge the pieces into a single i64 value.
2027 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2030 /// getZeroVector - Returns a vector of specified type with all zero elements.
2032 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2033 assert(VT.isVector() && "Expected a vector type");
2035 // Zero vectors are used to represent vector negation and in those cases
2036 // will be implemented with the NEON VNEG instruction. However, VNEG does
2037 // not support i64 elements, so sometimes the zero vectors will need to be
2038 // explicitly constructed. For those cases, and potentially other uses in
2039 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2040 // to their dest type. This ensures they get CSE'd.
2042 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2043 SmallVector<SDValue, 8> Ops;
2046 if (VT.getSizeInBits() == 64) {
2047 Ops.assign(8, Cst); TVT = MVT::v8i8;
2049 Ops.assign(16, Cst); TVT = MVT::v16i8;
2051 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2053 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2056 /// getOnesVector - Returns a vector of specified type with all bits set.
2058 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2059 assert(VT.isVector() && "Expected a vector type");
2061 // Always build ones vectors as <16 x i32> or <8 x i32> bitcasted to their
2062 // dest type. This ensures they get CSE'd.
2064 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2065 SmallVector<SDValue, 8> Ops;
2068 if (VT.getSizeInBits() == 64) {
2069 Ops.assign(8, Cst); TVT = MVT::v8i8;
2071 Ops.assign(16, Cst); TVT = MVT::v16i8;
2073 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2075 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2078 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2079 const ARMSubtarget *ST) {
2080 EVT VT = N->getValueType(0);
2081 DebugLoc dl = N->getDebugLoc();
2083 // Lower vector shifts on NEON to use VSHL.
2084 if (VT.isVector()) {
2085 assert(ST->hasNEON() && "unexpected vector shift");
2087 // Left shifts translate directly to the vshiftu intrinsic.
2088 if (N->getOpcode() == ISD::SHL)
2089 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2090 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2091 N->getOperand(0), N->getOperand(1));
2093 assert((N->getOpcode() == ISD::SRA ||
2094 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2096 // NEON uses the same intrinsics for both left and right shifts. For
2097 // right shifts, the shift amounts are negative, so negate the vector of
2099 EVT ShiftVT = N->getOperand(1).getValueType();
2100 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2101 getZeroVector(ShiftVT, DAG, dl),
2103 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2104 Intrinsic::arm_neon_vshifts :
2105 Intrinsic::arm_neon_vshiftu);
2106 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2107 DAG.getConstant(vshiftInt, MVT::i32),
2108 N->getOperand(0), NegatedCount);
2111 // We can get here for a node like i32 = ISD::SHL i32, i64
2115 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2116 "Unknown shift to lower!");
2118 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2119 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2120 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2123 // If we are in thumb mode, we don't have RRX.
2124 if (ST->isThumb1Only()) return SDValue();
2126 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2127 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2128 DAG.getConstant(0, MVT::i32));
2129 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2130 DAG.getConstant(1, MVT::i32));
2132 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2133 // captures the result into a carry flag.
2134 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2135 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2137 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2138 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2140 // Merge the pieces into a single i64 value.
2141 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2144 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2145 SDValue TmpOp0, TmpOp1;
2146 bool Invert = false;
2150 SDValue Op0 = Op.getOperand(0);
2151 SDValue Op1 = Op.getOperand(1);
2152 SDValue CC = Op.getOperand(2);
2153 EVT VT = Op.getValueType();
2154 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2155 DebugLoc dl = Op.getDebugLoc();
2157 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2158 switch (SetCCOpcode) {
2159 default: llvm_unreachable("Illegal FP comparison"); break;
2161 case ISD::SETNE: Invert = true; // Fallthrough
2163 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2165 case ISD::SETLT: Swap = true; // Fallthrough
2167 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2169 case ISD::SETLE: Swap = true; // Fallthrough
2171 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2172 case ISD::SETUGE: Swap = true; // Fallthrough
2173 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2174 case ISD::SETUGT: Swap = true; // Fallthrough
2175 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2176 case ISD::SETUEQ: Invert = true; // Fallthrough
2178 // Expand this to (OLT | OGT).
2182 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2183 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2185 case ISD::SETUO: Invert = true; // Fallthrough
2187 // Expand this to (OLT | OGE).
2191 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2192 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2196 // Integer comparisons.
2197 switch (SetCCOpcode) {
2198 default: llvm_unreachable("Illegal integer comparison"); break;
2199 case ISD::SETNE: Invert = true;
2200 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2201 case ISD::SETLT: Swap = true;
2202 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2203 case ISD::SETLE: Swap = true;
2204 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2205 case ISD::SETULT: Swap = true;
2206 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2207 case ISD::SETULE: Swap = true;
2208 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2211 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2212 if (Opc == ARMISD::VCEQ) {
2215 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2217 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2220 // Ignore bitconvert.
2221 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2222 AndOp = AndOp.getOperand(0);
2224 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2226 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2227 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2234 std::swap(Op0, Op1);
2236 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2239 Result = DAG.getNOT(dl, Result, VT);
2244 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2245 /// VMOV instruction, and if so, return the constant being splatted.
2246 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2247 unsigned SplatBitSize, SelectionDAG &DAG) {
2248 switch (SplatBitSize) {
2250 // Any 1-byte value is OK.
2251 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2252 return DAG.getTargetConstant(SplatBits, MVT::i8);
2255 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2256 if ((SplatBits & ~0xff) == 0 ||
2257 (SplatBits & ~0xff00) == 0)
2258 return DAG.getTargetConstant(SplatBits, MVT::i16);
2262 // NEON's 32-bit VMOV supports splat values where:
2263 // * only one byte is nonzero, or
2264 // * the least significant byte is 0xff and the second byte is nonzero, or
2265 // * the least significant 2 bytes are 0xff and the third is nonzero.
2266 if ((SplatBits & ~0xff) == 0 ||
2267 (SplatBits & ~0xff00) == 0 ||
2268 (SplatBits & ~0xff0000) == 0 ||
2269 (SplatBits & ~0xff000000) == 0)
2270 return DAG.getTargetConstant(SplatBits, MVT::i32);
2272 if ((SplatBits & ~0xffff) == 0 &&
2273 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2274 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2276 if ((SplatBits & ~0xffffff) == 0 &&
2277 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2278 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2280 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2281 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2282 // VMOV.I32. A (very) minor optimization would be to replicate the value
2283 // and fall through here to test for a valid 64-bit splat. But, then the
2284 // caller would also need to check and handle the change in size.
2288 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2289 uint64_t BitMask = 0xff;
2291 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2292 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2294 else if ((SplatBits & BitMask) != 0)
2298 return DAG.getTargetConstant(Val, MVT::i64);
2302 llvm_unreachable("unexpected size for isVMOVSplat");
2309 /// getVMOVImm - If this is a build_vector of constants which can be
2310 /// formed by using a VMOV instruction of the specified element size,
2311 /// return the constant being splatted. The ByteSize field indicates the
2312 /// number of bytes of each element [1248].
2313 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2314 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2315 APInt SplatBits, SplatUndef;
2316 unsigned SplatBitSize;
2318 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2319 HasAnyUndefs, ByteSize * 8))
2322 if (SplatBitSize > ByteSize * 8)
2325 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2329 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2330 bool &ReverseVEXT, unsigned &Imm) {
2331 unsigned NumElts = VT.getVectorNumElements();
2332 ReverseVEXT = false;
2335 // If this is a VEXT shuffle, the immediate value is the index of the first
2336 // element. The other shuffle indices must be the successive elements after
2338 unsigned ExpectedElt = Imm;
2339 for (unsigned i = 1; i < NumElts; ++i) {
2340 // Increment the expected index. If it wraps around, it may still be
2341 // a VEXT but the source vectors must be swapped.
2343 if (ExpectedElt == NumElts * 2) {
2348 if (ExpectedElt != static_cast<unsigned>(M[i]))
2352 // Adjust the index value if the source operands will be swapped.
2359 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2360 /// instruction with the specified blocksize. (The order of the elements
2361 /// within each block of the vector is reversed.)
2362 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2363 unsigned BlockSize) {
2364 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2365 "Only possible block sizes for VREV are: 16, 32, 64");
2367 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2371 unsigned NumElts = VT.getVectorNumElements();
2372 unsigned BlockElts = M[0] + 1;
2374 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2377 for (unsigned i = 0; i < NumElts; ++i) {
2378 if ((unsigned) M[i] !=
2379 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2386 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2387 unsigned &WhichResult) {
2388 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2392 unsigned NumElts = VT.getVectorNumElements();
2393 WhichResult = (M[0] == 0 ? 0 : 1);
2394 for (unsigned i = 0; i < NumElts; i += 2) {
2395 if ((unsigned) M[i] != i + WhichResult ||
2396 (unsigned) M[i+1] != i + NumElts + WhichResult)
2402 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2403 unsigned &WhichResult) {
2404 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2408 unsigned NumElts = VT.getVectorNumElements();
2409 WhichResult = (M[0] == 0 ? 0 : 1);
2410 for (unsigned i = 0; i != NumElts; ++i) {
2411 if ((unsigned) M[i] != 2 * i + WhichResult)
2415 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2416 if (VT.is64BitVector() && EltSz == 32)
2422 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2423 unsigned &WhichResult) {
2424 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2428 unsigned NumElts = VT.getVectorNumElements();
2429 WhichResult = (M[0] == 0 ? 0 : 1);
2430 unsigned Idx = WhichResult * NumElts / 2;
2431 for (unsigned i = 0; i != NumElts; i += 2) {
2432 if ((unsigned) M[i] != Idx ||
2433 (unsigned) M[i+1] != Idx + NumElts)
2438 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2439 if (VT.is64BitVector() && EltSz == 32)
2445 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2446 // Canonicalize all-zeros and all-ones vectors.
2447 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2448 if (ConstVal->isNullValue())
2449 return getZeroVector(VT, DAG, dl);
2450 if (ConstVal->isAllOnesValue())
2451 return getOnesVector(VT, DAG, dl);
2454 if (VT.is64BitVector()) {
2455 switch (Val.getValueType().getSizeInBits()) {
2456 case 8: CanonicalVT = MVT::v8i8; break;
2457 case 16: CanonicalVT = MVT::v4i16; break;
2458 case 32: CanonicalVT = MVT::v2i32; break;
2459 case 64: CanonicalVT = MVT::v1i64; break;
2460 default: llvm_unreachable("unexpected splat element type"); break;
2463 assert(VT.is128BitVector() && "unknown splat vector size");
2464 switch (Val.getValueType().getSizeInBits()) {
2465 case 8: CanonicalVT = MVT::v16i8; break;
2466 case 16: CanonicalVT = MVT::v8i16; break;
2467 case 32: CanonicalVT = MVT::v4i32; break;
2468 case 64: CanonicalVT = MVT::v2i64; break;
2469 default: llvm_unreachable("unexpected splat element type"); break;
2473 // Build a canonical splat for this value.
2474 SmallVector<SDValue, 8> Ops;
2475 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2476 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2478 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2481 // If this is a case we can't handle, return null and let the default
2482 // expansion code take care of it.
2483 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2484 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2485 DebugLoc dl = Op.getDebugLoc();
2486 EVT VT = Op.getValueType();
2488 APInt SplatBits, SplatUndef;
2489 unsigned SplatBitSize;
2491 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2492 if (SplatBitSize <= 64) {
2493 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2494 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2496 return BuildSplat(Val, VT, DAG, dl);
2500 // If there are only 2 elements in a 128-bit vector, insert them into an
2501 // undef vector. This handles the common case for 128-bit vector argument
2502 // passing, where the insertions should be translated to subreg accesses
2503 // with no real instructions.
2504 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2505 SDValue Val = DAG.getUNDEF(VT);
2506 SDValue Op0 = Op.getOperand(0);
2507 SDValue Op1 = Op.getOperand(1);
2508 if (Op0.getOpcode() != ISD::UNDEF)
2509 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2510 DAG.getIntPtrConstant(0));
2511 if (Op1.getOpcode() != ISD::UNDEF)
2512 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2513 DAG.getIntPtrConstant(1));
2520 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2521 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2522 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2523 /// are assumed to be legal.
2525 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2527 if (VT.getVectorNumElements() == 4 &&
2528 (VT.is128BitVector() || VT.is64BitVector())) {
2529 unsigned PFIndexes[4];
2530 for (unsigned i = 0; i != 4; ++i) {
2534 PFIndexes[i] = M[i];
2537 // Compute the index in the perfect shuffle table.
2538 unsigned PFTableIndex =
2539 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2540 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2541 unsigned Cost = (PFEntry >> 30);
2548 unsigned Imm, WhichResult;
2550 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2551 isVREVMask(M, VT, 64) ||
2552 isVREVMask(M, VT, 32) ||
2553 isVREVMask(M, VT, 16) ||
2554 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2555 isVTRNMask(M, VT, WhichResult) ||
2556 isVUZPMask(M, VT, WhichResult) ||
2557 isVZIPMask(M, VT, WhichResult));
2560 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2561 /// the specified operations to build the shuffle.
2562 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2563 SDValue RHS, SelectionDAG &DAG,
2565 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2566 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2567 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2570 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2579 OP_VUZPL, // VUZP, left result
2580 OP_VUZPR, // VUZP, right result
2581 OP_VZIPL, // VZIP, left result
2582 OP_VZIPR, // VZIP, right result
2583 OP_VTRNL, // VTRN, left result
2584 OP_VTRNR // VTRN, right result
2587 if (OpNum == OP_COPY) {
2588 if (LHSID == (1*9+2)*9+3) return LHS;
2589 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2593 SDValue OpLHS, OpRHS;
2594 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2595 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2596 EVT VT = OpLHS.getValueType();
2599 default: llvm_unreachable("Unknown shuffle opcode!");
2601 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2606 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2607 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2611 return DAG.getNode(ARMISD::VEXT, dl, VT,
2613 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2616 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2617 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2620 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2621 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2624 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2625 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2629 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2630 SDValue V1 = Op.getOperand(0);
2631 SDValue V2 = Op.getOperand(1);
2632 DebugLoc dl = Op.getDebugLoc();
2633 EVT VT = Op.getValueType();
2634 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2635 SmallVector<int, 8> ShuffleMask;
2637 // Convert shuffles that are directly supported on NEON to target-specific
2638 // DAG nodes, instead of keeping them as shuffles and matching them again
2639 // during code selection. This is more efficient and avoids the possibility
2640 // of inconsistencies between legalization and selection.
2641 // FIXME: floating-point vectors should be canonicalized to integer vectors
2642 // of the same time so that they get CSEd properly.
2643 SVN->getMask(ShuffleMask);
2645 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2646 int Lane = SVN->getSplatIndex();
2647 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2648 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2650 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2651 DAG.getConstant(Lane, MVT::i32));
2656 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2659 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2660 DAG.getConstant(Imm, MVT::i32));
2663 if (isVREVMask(ShuffleMask, VT, 64))
2664 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2665 if (isVREVMask(ShuffleMask, VT, 32))
2666 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2667 if (isVREVMask(ShuffleMask, VT, 16))
2668 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2670 // Check for Neon shuffles that modify both input vectors in place.
2671 // If both results are used, i.e., if there are two shuffles with the same
2672 // source operands and with masks corresponding to both results of one of
2673 // these operations, DAG memoization will ensure that a single node is
2674 // used for both shuffles.
2675 unsigned WhichResult;
2676 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2677 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2678 V1, V2).getValue(WhichResult);
2679 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2680 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2681 V1, V2).getValue(WhichResult);
2682 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2683 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2684 V1, V2).getValue(WhichResult);
2686 // If the shuffle is not directly supported and it has 4 elements, use
2687 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2688 if (VT.getVectorNumElements() == 4 &&
2689 (VT.is128BitVector() || VT.is64BitVector())) {
2690 unsigned PFIndexes[4];
2691 for (unsigned i = 0; i != 4; ++i) {
2692 if (ShuffleMask[i] < 0)
2695 PFIndexes[i] = ShuffleMask[i];
2698 // Compute the index in the perfect shuffle table.
2699 unsigned PFTableIndex =
2700 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2702 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2703 unsigned Cost = (PFEntry >> 30);
2706 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2712 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2713 EVT VT = Op.getValueType();
2714 DebugLoc dl = Op.getDebugLoc();
2715 SDValue Vec = Op.getOperand(0);
2716 SDValue Lane = Op.getOperand(1);
2717 assert(VT == MVT::i32 &&
2718 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
2719 "unexpected type for custom-lowering vector extract");
2720 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2723 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2724 // The only time a CONCAT_VECTORS operation can have legal types is when
2725 // two 64-bit vectors are concatenated to a 128-bit vector.
2726 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2727 "unexpected CONCAT_VECTORS");
2728 DebugLoc dl = Op.getDebugLoc();
2729 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2730 SDValue Op0 = Op.getOperand(0);
2731 SDValue Op1 = Op.getOperand(1);
2732 if (Op0.getOpcode() != ISD::UNDEF)
2733 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2734 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2735 DAG.getIntPtrConstant(0));
2736 if (Op1.getOpcode() != ISD::UNDEF)
2737 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2738 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2739 DAG.getIntPtrConstant(1));
2740 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2743 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2744 switch (Op.getOpcode()) {
2745 default: llvm_unreachable("Don't know how to custom lower this!");
2746 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2747 case ISD::GlobalAddress:
2748 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2749 LowerGlobalAddressELF(Op, DAG);
2750 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2751 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2752 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2753 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2754 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2755 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2756 case ISD::SINT_TO_FP:
2757 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2758 case ISD::FP_TO_SINT:
2759 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2760 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2761 case ISD::RETURNADDR: break;
2762 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2763 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2764 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2765 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2768 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2769 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2770 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2771 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2772 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2773 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2778 /// ReplaceNodeResults - Replace the results of node with an illegal result
2779 /// type with new values built out of custom code.
2780 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2781 SmallVectorImpl<SDValue>&Results,
2782 SelectionDAG &DAG) {
2783 switch (N->getOpcode()) {
2785 llvm_unreachable("Don't know how to custom expand this!");
2787 case ISD::BIT_CONVERT:
2788 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2792 SDValue Res = LowerShift(N, DAG, Subtarget);
2794 Results.push_back(Res);
2800 //===----------------------------------------------------------------------===//
2801 // ARM Scheduler Hooks
2802 //===----------------------------------------------------------------------===//
2805 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2806 MachineBasicBlock *BB,
2807 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
2808 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2809 DebugLoc dl = MI->getDebugLoc();
2810 switch (MI->getOpcode()) {
2812 llvm_unreachable("Unexpected instr type to insert");
2813 case ARM::tMOVCCr_pseudo: {
2814 // To "insert" a SELECT_CC instruction, we actually have to insert the
2815 // diamond control-flow pattern. The incoming instruction knows the
2816 // destination vreg to set, the condition code register to branch on, the
2817 // true/false values to select between, and a branch opcode to use.
2818 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2819 MachineFunction::iterator It = BB;
2825 // cmpTY ccX, r1, r2
2827 // fallthrough --> copy0MBB
2828 MachineBasicBlock *thisMBB = BB;
2829 MachineFunction *F = BB->getParent();
2830 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2831 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2832 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2833 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2834 F->insert(It, copy0MBB);
2835 F->insert(It, sinkMBB);
2836 // Update machine-CFG edges by first adding all successors of the current
2837 // block to the new block which will contain the Phi node for the select.
2838 // Also inform sdisel of the edge changes.
2839 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
2840 E = BB->succ_end(); I != E; ++I) {
2841 EM->insert(std::make_pair(*I, sinkMBB));
2842 sinkMBB->addSuccessor(*I);
2844 // Next, remove all successors of the current block, and add the true
2845 // and fallthrough blocks as its successors.
2846 while (!BB->succ_empty())
2847 BB->removeSuccessor(BB->succ_begin());
2848 BB->addSuccessor(copy0MBB);
2849 BB->addSuccessor(sinkMBB);
2852 // %FalseValue = ...
2853 // # fallthrough to sinkMBB
2856 // Update machine-CFG edges
2857 BB->addSuccessor(sinkMBB);
2860 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2863 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2864 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2865 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2867 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2874 case ARM::t2SUBrSPi_:
2875 case ARM::t2SUBrSPi12_:
2876 case ARM::t2SUBrSPs_: {
2877 MachineFunction *MF = BB->getParent();
2878 unsigned DstReg = MI->getOperand(0).getReg();
2879 unsigned SrcReg = MI->getOperand(1).getReg();
2880 bool DstIsDead = MI->getOperand(0).isDead();
2881 bool SrcIsKill = MI->getOperand(1).isKill();
2883 if (SrcReg != ARM::SP) {
2884 // Copy the source to SP from virtual register.
2885 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2886 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2887 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2888 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2889 .addReg(SrcReg, getKillRegState(SrcIsKill));
2893 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2894 switch (MI->getOpcode()) {
2896 llvm_unreachable("Unexpected pseudo instruction!");
2902 OpOpc = ARM::tADDspr;
2905 OpOpc = ARM::tSUBspi;
2907 case ARM::t2SUBrSPi_:
2908 OpOpc = ARM::t2SUBrSPi;
2909 NeedPred = true; NeedCC = true;
2911 case ARM::t2SUBrSPi12_:
2912 OpOpc = ARM::t2SUBrSPi12;
2915 case ARM::t2SUBrSPs_:
2916 OpOpc = ARM::t2SUBrSPs;
2917 NeedPred = true; NeedCC = true; NeedOp3 = true;
2920 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2921 if (OpOpc == ARM::tAND)
2922 AddDefaultT1CC(MIB);
2923 MIB.addReg(ARM::SP);
2924 MIB.addOperand(MI->getOperand(2));
2926 MIB.addOperand(MI->getOperand(3));
2928 AddDefaultPred(MIB);
2932 // Copy the result from SP to virtual register.
2933 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2934 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2935 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2936 BuildMI(BB, dl, TII->get(CopyOpc))
2937 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
2939 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2945 //===----------------------------------------------------------------------===//
2946 // ARM Optimization Hooks
2947 //===----------------------------------------------------------------------===//
2950 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2951 TargetLowering::DAGCombinerInfo &DCI) {
2952 SelectionDAG &DAG = DCI.DAG;
2953 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2954 EVT VT = N->getValueType(0);
2955 unsigned Opc = N->getOpcode();
2956 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2957 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2958 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2959 ISD::CondCode CC = ISD::SETCC_INVALID;
2962 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2964 SDValue CCOp = Slct.getOperand(0);
2965 if (CCOp.getOpcode() == ISD::SETCC)
2966 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2969 bool DoXform = false;
2971 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2974 if (LHS.getOpcode() == ISD::Constant &&
2975 cast<ConstantSDNode>(LHS)->isNullValue()) {
2977 } else if (CC != ISD::SETCC_INVALID &&
2978 RHS.getOpcode() == ISD::Constant &&
2979 cast<ConstantSDNode>(RHS)->isNullValue()) {
2980 std::swap(LHS, RHS);
2981 SDValue Op0 = Slct.getOperand(0);
2982 EVT OpVT = isSlctCC ? Op0.getValueType() :
2983 Op0.getOperand(0).getValueType();
2984 bool isInt = OpVT.isInteger();
2985 CC = ISD::getSetCCInverse(CC, isInt);
2987 if (!TLI.isCondCodeLegal(CC, OpVT))
2988 return SDValue(); // Inverse operator isn't legal.
2995 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
2997 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
2998 Slct.getOperand(0), Slct.getOperand(1), CC);
2999 SDValue CCOp = Slct.getOperand(0);
3001 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3002 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3003 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3004 CCOp, OtherOp, Result);
3009 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3010 static SDValue PerformADDCombine(SDNode *N,
3011 TargetLowering::DAGCombinerInfo &DCI) {
3012 // added by evan in r37685 with no testcase.
3013 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3015 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3016 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3017 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3018 if (Result.getNode()) return Result;
3020 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3021 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3022 if (Result.getNode()) return Result;
3028 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3029 static SDValue PerformSUBCombine(SDNode *N,
3030 TargetLowering::DAGCombinerInfo &DCI) {
3031 // added by evan in r37685 with no testcase.
3032 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3034 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3035 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3036 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3037 if (Result.getNode()) return Result;
3043 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3044 static SDValue PerformFMRRDCombine(SDNode *N,
3045 TargetLowering::DAGCombinerInfo &DCI) {
3046 // fmrrd(fmdrr x, y) -> x,y
3047 SDValue InDouble = N->getOperand(0);
3048 if (InDouble.getOpcode() == ARMISD::FMDRR)
3049 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3053 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3054 /// operand of a vector shift operation, where all the elements of the
3055 /// build_vector must have the same constant integer value.
3056 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3057 // Ignore bit_converts.
3058 while (Op.getOpcode() == ISD::BIT_CONVERT)
3059 Op = Op.getOperand(0);
3060 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3061 APInt SplatBits, SplatUndef;
3062 unsigned SplatBitSize;
3064 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3065 HasAnyUndefs, ElementBits) ||
3066 SplatBitSize > ElementBits)
3068 Cnt = SplatBits.getSExtValue();
3072 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3073 /// operand of a vector shift left operation. That value must be in the range:
3074 /// 0 <= Value < ElementBits for a left shift; or
3075 /// 0 <= Value <= ElementBits for a long left shift.
3076 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3077 assert(VT.isVector() && "vector shift count is not a vector type");
3078 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3079 if (! getVShiftImm(Op, ElementBits, Cnt))
3081 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3084 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3085 /// operand of a vector shift right operation. For a shift opcode, the value
3086 /// is positive, but for an intrinsic the value count must be negative. The
3087 /// absolute value must be in the range:
3088 /// 1 <= |Value| <= ElementBits for a right shift; or
3089 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3090 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3092 assert(VT.isVector() && "vector shift count is not a vector type");
3093 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3094 if (! getVShiftImm(Op, ElementBits, Cnt))
3098 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3101 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3102 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3103 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3106 // Don't do anything for most intrinsics.
3109 // Vector shifts: check for immediate versions and lower them.
3110 // Note: This is done during DAG combining instead of DAG legalizing because
3111 // the build_vectors for 64-bit vector element shift counts are generally
3112 // not legal, and it is hard to see their values after they get legalized to
3113 // loads from a constant pool.
3114 case Intrinsic::arm_neon_vshifts:
3115 case Intrinsic::arm_neon_vshiftu:
3116 case Intrinsic::arm_neon_vshiftls:
3117 case Intrinsic::arm_neon_vshiftlu:
3118 case Intrinsic::arm_neon_vshiftn:
3119 case Intrinsic::arm_neon_vrshifts:
3120 case Intrinsic::arm_neon_vrshiftu:
3121 case Intrinsic::arm_neon_vrshiftn:
3122 case Intrinsic::arm_neon_vqshifts:
3123 case Intrinsic::arm_neon_vqshiftu:
3124 case Intrinsic::arm_neon_vqshiftsu:
3125 case Intrinsic::arm_neon_vqshiftns:
3126 case Intrinsic::arm_neon_vqshiftnu:
3127 case Intrinsic::arm_neon_vqshiftnsu:
3128 case Intrinsic::arm_neon_vqrshiftns:
3129 case Intrinsic::arm_neon_vqrshiftnu:
3130 case Intrinsic::arm_neon_vqrshiftnsu: {
3131 EVT VT = N->getOperand(1).getValueType();
3133 unsigned VShiftOpc = 0;
3136 case Intrinsic::arm_neon_vshifts:
3137 case Intrinsic::arm_neon_vshiftu:
3138 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3139 VShiftOpc = ARMISD::VSHL;
3142 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3143 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3144 ARMISD::VSHRs : ARMISD::VSHRu);
3149 case Intrinsic::arm_neon_vshiftls:
3150 case Intrinsic::arm_neon_vshiftlu:
3151 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3153 llvm_unreachable("invalid shift count for vshll intrinsic");
3155 case Intrinsic::arm_neon_vrshifts:
3156 case Intrinsic::arm_neon_vrshiftu:
3157 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3161 case Intrinsic::arm_neon_vqshifts:
3162 case Intrinsic::arm_neon_vqshiftu:
3163 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3167 case Intrinsic::arm_neon_vqshiftsu:
3168 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3170 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3172 case Intrinsic::arm_neon_vshiftn:
3173 case Intrinsic::arm_neon_vrshiftn:
3174 case Intrinsic::arm_neon_vqshiftns:
3175 case Intrinsic::arm_neon_vqshiftnu:
3176 case Intrinsic::arm_neon_vqshiftnsu:
3177 case Intrinsic::arm_neon_vqrshiftns:
3178 case Intrinsic::arm_neon_vqrshiftnu:
3179 case Intrinsic::arm_neon_vqrshiftnsu:
3180 // Narrowing shifts require an immediate right shift.
3181 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3183 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3186 llvm_unreachable("unhandled vector shift");
3190 case Intrinsic::arm_neon_vshifts:
3191 case Intrinsic::arm_neon_vshiftu:
3192 // Opcode already set above.
3194 case Intrinsic::arm_neon_vshiftls:
3195 case Intrinsic::arm_neon_vshiftlu:
3196 if (Cnt == VT.getVectorElementType().getSizeInBits())
3197 VShiftOpc = ARMISD::VSHLLi;
3199 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3200 ARMISD::VSHLLs : ARMISD::VSHLLu);
3202 case Intrinsic::arm_neon_vshiftn:
3203 VShiftOpc = ARMISD::VSHRN; break;
3204 case Intrinsic::arm_neon_vrshifts:
3205 VShiftOpc = ARMISD::VRSHRs; break;
3206 case Intrinsic::arm_neon_vrshiftu:
3207 VShiftOpc = ARMISD::VRSHRu; break;
3208 case Intrinsic::arm_neon_vrshiftn:
3209 VShiftOpc = ARMISD::VRSHRN; break;
3210 case Intrinsic::arm_neon_vqshifts:
3211 VShiftOpc = ARMISD::VQSHLs; break;
3212 case Intrinsic::arm_neon_vqshiftu:
3213 VShiftOpc = ARMISD::VQSHLu; break;
3214 case Intrinsic::arm_neon_vqshiftsu:
3215 VShiftOpc = ARMISD::VQSHLsu; break;
3216 case Intrinsic::arm_neon_vqshiftns:
3217 VShiftOpc = ARMISD::VQSHRNs; break;
3218 case Intrinsic::arm_neon_vqshiftnu:
3219 VShiftOpc = ARMISD::VQSHRNu; break;
3220 case Intrinsic::arm_neon_vqshiftnsu:
3221 VShiftOpc = ARMISD::VQSHRNsu; break;
3222 case Intrinsic::arm_neon_vqrshiftns:
3223 VShiftOpc = ARMISD::VQRSHRNs; break;
3224 case Intrinsic::arm_neon_vqrshiftnu:
3225 VShiftOpc = ARMISD::VQRSHRNu; break;
3226 case Intrinsic::arm_neon_vqrshiftnsu:
3227 VShiftOpc = ARMISD::VQRSHRNsu; break;
3230 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3231 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3234 case Intrinsic::arm_neon_vshiftins: {
3235 EVT VT = N->getOperand(1).getValueType();
3237 unsigned VShiftOpc = 0;
3239 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3240 VShiftOpc = ARMISD::VSLI;
3241 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3242 VShiftOpc = ARMISD::VSRI;
3244 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3247 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3248 N->getOperand(1), N->getOperand(2),
3249 DAG.getConstant(Cnt, MVT::i32));
3252 case Intrinsic::arm_neon_vqrshifts:
3253 case Intrinsic::arm_neon_vqrshiftu:
3254 // No immediate versions of these to check for.
3261 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3262 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3263 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3264 /// vector element shift counts are generally not legal, and it is hard to see
3265 /// their values after they get legalized to loads from a constant pool.
3266 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3267 const ARMSubtarget *ST) {
3268 EVT VT = N->getValueType(0);
3270 // Nothing to be done for scalar shifts.
3271 if (! VT.isVector())
3274 assert(ST->hasNEON() && "unexpected vector shift");
3277 switch (N->getOpcode()) {
3278 default: llvm_unreachable("unexpected shift opcode");
3281 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3282 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3283 DAG.getConstant(Cnt, MVT::i32));
3288 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3289 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3290 ARMISD::VSHRs : ARMISD::VSHRu);
3291 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3292 DAG.getConstant(Cnt, MVT::i32));
3298 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3299 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3300 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3301 const ARMSubtarget *ST) {
3302 SDValue N0 = N->getOperand(0);
3304 // Check for sign- and zero-extensions of vector extract operations of 8-
3305 // and 16-bit vector elements. NEON supports these directly. They are
3306 // handled during DAG combining because type legalization will promote them
3307 // to 32-bit types and it is messy to recognize the operations after that.
3308 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3309 SDValue Vec = N0.getOperand(0);
3310 SDValue Lane = N0.getOperand(1);
3311 EVT VT = N->getValueType(0);
3312 EVT EltVT = N0.getValueType();
3313 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3315 if (VT == MVT::i32 &&
3316 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3317 TLI.isTypeLegal(Vec.getValueType())) {
3320 switch (N->getOpcode()) {
3321 default: llvm_unreachable("unexpected opcode");
3322 case ISD::SIGN_EXTEND:
3323 Opc = ARMISD::VGETLANEs;
3325 case ISD::ZERO_EXTEND:
3326 case ISD::ANY_EXTEND:
3327 Opc = ARMISD::VGETLANEu;
3330 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3337 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3338 DAGCombinerInfo &DCI) const {
3339 switch (N->getOpcode()) {
3341 case ISD::ADD: return PerformADDCombine(N, DCI);
3342 case ISD::SUB: return PerformSUBCombine(N, DCI);
3343 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3344 case ISD::INTRINSIC_WO_CHAIN:
3345 return PerformIntrinsicCombine(N, DCI.DAG);
3349 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3350 case ISD::SIGN_EXTEND:
3351 case ISD::ZERO_EXTEND:
3352 case ISD::ANY_EXTEND:
3353 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3358 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3359 if (!Subtarget->hasV6Ops())
3360 // Pre-v6 does not support unaligned mem access.
3362 else if (!Subtarget->hasV6Ops()) {
3363 // v6 may or may not support unaligned mem access.
3364 if (!Subtarget->isTargetDarwin())
3368 switch (VT.getSimpleVT().SimpleTy) {
3375 // FIXME: VLD1 etc with standard alignment is legal.
3379 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3384 switch (VT.getSimpleVT().SimpleTy) {
3385 default: return false;
3400 if ((V & (Scale - 1)) != 0)
3403 return V == (V & ((1LL << 5) - 1));
3406 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3407 const ARMSubtarget *Subtarget) {
3414 switch (VT.getSimpleVT().SimpleTy) {
3415 default: return false;
3420 // + imm12 or - imm8
3422 return V == (V & ((1LL << 8) - 1));
3423 return V == (V & ((1LL << 12) - 1));
3426 // Same as ARM mode. FIXME: NEON?
3427 if (!Subtarget->hasVFP2())
3432 return V == (V & ((1LL << 8) - 1));
3436 /// isLegalAddressImmediate - Return true if the integer value can be used
3437 /// as the offset of the target addressing mode for load / store of the
3439 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3440 const ARMSubtarget *Subtarget) {
3447 if (Subtarget->isThumb1Only())
3448 return isLegalT1AddressImmediate(V, VT);
3449 else if (Subtarget->isThumb2())
3450 return isLegalT2AddressImmediate(V, VT, Subtarget);
3455 switch (VT.getSimpleVT().SimpleTy) {
3456 default: return false;
3461 return V == (V & ((1LL << 12) - 1));
3464 return V == (V & ((1LL << 8) - 1));
3467 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3472 return V == (V & ((1LL << 8) - 1));
3476 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3478 int Scale = AM.Scale;
3482 switch (VT.getSimpleVT().SimpleTy) {
3483 default: return false;
3492 return Scale == 2 || Scale == 4 || Scale == 8;
3495 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3499 // Note, we allow "void" uses (basically, uses that aren't loads or
3500 // stores), because arm allows folding a scale into many arithmetic
3501 // operations. This should be made more precise and revisited later.
3503 // Allow r << imm, but the imm has to be a multiple of two.
3504 if (Scale & 1) return false;
3505 return isPowerOf2_32(Scale);
3509 /// isLegalAddressingMode - Return true if the addressing mode represented
3510 /// by AM is legal for this target, for a load/store of the specified type.
3511 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3512 const Type *Ty) const {
3513 EVT VT = getValueType(Ty, true);
3514 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3517 // Can never fold addr of global into load/store.
3522 case 0: // no scale reg, must be "r+i" or "r", or "i".
3525 if (Subtarget->isThumb1Only())
3529 // ARM doesn't support any R+R*scale+imm addr modes.
3536 if (Subtarget->isThumb2())
3537 return isLegalT2ScaledAddressingMode(AM, VT);
3539 int Scale = AM.Scale;
3540 switch (VT.getSimpleVT().SimpleTy) {
3541 default: return false;
3545 if (Scale < 0) Scale = -Scale;
3549 return isPowerOf2_32(Scale & ~1);
3553 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3558 // Note, we allow "void" uses (basically, uses that aren't loads or
3559 // stores), because arm allows folding a scale into many arithmetic
3560 // operations. This should be made more precise and revisited later.
3562 // Allow r << imm, but the imm has to be a multiple of two.
3563 if (Scale & 1) return false;
3564 return isPowerOf2_32(Scale);
3571 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3572 bool isSEXTLoad, SDValue &Base,
3573 SDValue &Offset, bool &isInc,
3574 SelectionDAG &DAG) {
3575 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3578 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3580 Base = Ptr->getOperand(0);
3581 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3582 int RHSC = (int)RHS->getZExtValue();
3583 if (RHSC < 0 && RHSC > -256) {
3584 assert(Ptr->getOpcode() == ISD::ADD);
3586 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3590 isInc = (Ptr->getOpcode() == ISD::ADD);
3591 Offset = Ptr->getOperand(1);
3593 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3595 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3596 int RHSC = (int)RHS->getZExtValue();
3597 if (RHSC < 0 && RHSC > -0x1000) {
3598 assert(Ptr->getOpcode() == ISD::ADD);
3600 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3601 Base = Ptr->getOperand(0);
3606 if (Ptr->getOpcode() == ISD::ADD) {
3608 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3609 if (ShOpcVal != ARM_AM::no_shift) {
3610 Base = Ptr->getOperand(1);
3611 Offset = Ptr->getOperand(0);
3613 Base = Ptr->getOperand(0);
3614 Offset = Ptr->getOperand(1);
3619 isInc = (Ptr->getOpcode() == ISD::ADD);
3620 Base = Ptr->getOperand(0);
3621 Offset = Ptr->getOperand(1);
3625 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3629 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3630 bool isSEXTLoad, SDValue &Base,
3631 SDValue &Offset, bool &isInc,
3632 SelectionDAG &DAG) {
3633 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3636 Base = Ptr->getOperand(0);
3637 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3638 int RHSC = (int)RHS->getZExtValue();
3639 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3640 assert(Ptr->getOpcode() == ISD::ADD);
3642 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3644 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3645 isInc = Ptr->getOpcode() == ISD::ADD;
3646 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3654 /// getPreIndexedAddressParts - returns true by value, base pointer and
3655 /// offset pointer and addressing mode by reference if the node's address
3656 /// can be legally represented as pre-indexed load / store address.
3658 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3660 ISD::MemIndexedMode &AM,
3661 SelectionDAG &DAG) const {
3662 if (Subtarget->isThumb1Only())
3667 bool isSEXTLoad = false;
3668 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3669 Ptr = LD->getBasePtr();
3670 VT = LD->getMemoryVT();
3671 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3672 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3673 Ptr = ST->getBasePtr();
3674 VT = ST->getMemoryVT();
3679 bool isLegal = false;
3680 if (Subtarget->isThumb2())
3681 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3682 Offset, isInc, DAG);
3684 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3685 Offset, isInc, DAG);
3689 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3693 /// getPostIndexedAddressParts - returns true by value, base pointer and
3694 /// offset pointer and addressing mode by reference if this node can be
3695 /// combined with a load / store to form a post-indexed load / store.
3696 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3699 ISD::MemIndexedMode &AM,
3700 SelectionDAG &DAG) const {
3701 if (Subtarget->isThumb1Only())
3706 bool isSEXTLoad = false;
3707 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3708 VT = LD->getMemoryVT();
3709 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3710 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3711 VT = ST->getMemoryVT();
3716 bool isLegal = false;
3717 if (Subtarget->isThumb2())
3718 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3721 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3726 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3730 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3734 const SelectionDAG &DAG,
3735 unsigned Depth) const {
3736 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3737 switch (Op.getOpcode()) {
3739 case ARMISD::CMOV: {
3740 // Bits are known zero/one if known on the LHS and RHS.
3741 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3742 if (KnownZero == 0 && KnownOne == 0) return;
3744 APInt KnownZeroRHS, KnownOneRHS;
3745 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3746 KnownZeroRHS, KnownOneRHS, Depth+1);
3747 KnownZero &= KnownZeroRHS;
3748 KnownOne &= KnownOneRHS;
3754 //===----------------------------------------------------------------------===//
3755 // ARM Inline Assembly Support
3756 //===----------------------------------------------------------------------===//
3758 /// getConstraintType - Given a constraint letter, return the type of
3759 /// constraint it is for this target.
3760 ARMTargetLowering::ConstraintType
3761 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3762 if (Constraint.size() == 1) {
3763 switch (Constraint[0]) {
3765 case 'l': return C_RegisterClass;
3766 case 'w': return C_RegisterClass;
3769 return TargetLowering::getConstraintType(Constraint);
3772 std::pair<unsigned, const TargetRegisterClass*>
3773 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3775 if (Constraint.size() == 1) {
3776 // GCC RS6000 Constraint Letters
3777 switch (Constraint[0]) {
3779 if (Subtarget->isThumb1Only())
3780 return std::make_pair(0U, ARM::tGPRRegisterClass);
3782 return std::make_pair(0U, ARM::GPRRegisterClass);
3784 return std::make_pair(0U, ARM::GPRRegisterClass);
3787 return std::make_pair(0U, ARM::SPRRegisterClass);
3789 return std::make_pair(0U, ARM::DPRRegisterClass);
3793 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3796 std::vector<unsigned> ARMTargetLowering::
3797 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3799 if (Constraint.size() != 1)
3800 return std::vector<unsigned>();
3802 switch (Constraint[0]) { // GCC ARM Constraint Letters
3805 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3806 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3809 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3810 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3811 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3812 ARM::R12, ARM::LR, 0);
3815 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3816 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3817 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3818 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3819 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3820 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3821 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3822 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3824 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3825 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3826 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3827 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3831 return std::vector<unsigned>();
3834 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3835 /// vector. If it is invalid, don't add anything to Ops.
3836 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3839 std::vector<SDValue>&Ops,
3840 SelectionDAG &DAG) const {
3841 SDValue Result(0, 0);
3843 switch (Constraint) {
3845 case 'I': case 'J': case 'K': case 'L':
3846 case 'M': case 'N': case 'O':
3847 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3851 int64_t CVal64 = C->getSExtValue();
3852 int CVal = (int) CVal64;
3853 // None of these constraints allow values larger than 32 bits. Check
3854 // that the value fits in an int.
3858 switch (Constraint) {
3860 if (Subtarget->isThumb1Only()) {
3861 // This must be a constant between 0 and 255, for ADD
3863 if (CVal >= 0 && CVal <= 255)
3865 } else if (Subtarget->isThumb2()) {
3866 // A constant that can be used as an immediate value in a
3867 // data-processing instruction.
3868 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3871 // A constant that can be used as an immediate value in a
3872 // data-processing instruction.
3873 if (ARM_AM::getSOImmVal(CVal) != -1)
3879 if (Subtarget->isThumb()) { // FIXME thumb2
3880 // This must be a constant between -255 and -1, for negated ADD
3881 // immediates. This can be used in GCC with an "n" modifier that
3882 // prints the negated value, for use with SUB instructions. It is
3883 // not useful otherwise but is implemented for compatibility.
3884 if (CVal >= -255 && CVal <= -1)
3887 // This must be a constant between -4095 and 4095. It is not clear
3888 // what this constraint is intended for. Implemented for
3889 // compatibility with GCC.
3890 if (CVal >= -4095 && CVal <= 4095)
3896 if (Subtarget->isThumb1Only()) {
3897 // A 32-bit value where only one byte has a nonzero value. Exclude
3898 // zero to match GCC. This constraint is used by GCC internally for
3899 // constants that can be loaded with a move/shift combination.
3900 // It is not useful otherwise but is implemented for compatibility.
3901 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3903 } else if (Subtarget->isThumb2()) {
3904 // A constant whose bitwise inverse can be used as an immediate
3905 // value in a data-processing instruction. This can be used in GCC
3906 // with a "B" modifier that prints the inverted value, for use with
3907 // BIC and MVN instructions. It is not useful otherwise but is
3908 // implemented for compatibility.
3909 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
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::getSOImmVal(~CVal) != -1)
3923 if (Subtarget->isThumb1Only()) {
3924 // This must be a constant between -7 and 7,
3925 // for 3-operand ADD/SUB immediate instructions.
3926 if (CVal >= -7 && CVal < 7)
3928 } else if (Subtarget->isThumb2()) {
3929 // A constant whose negation can be used as an immediate value in a
3930 // data-processing instruction. This can be used in GCC with an "n"
3931 // modifier that prints the negated value, for use with SUB
3932 // instructions. It is not useful otherwise but is implemented for
3934 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
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::getSOImmVal(-CVal) != -1)
3948 if (Subtarget->isThumb()) { // FIXME thumb2
3949 // This must be a multiple of 4 between 0 and 1020, for
3950 // ADD sp + immediate.
3951 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3954 // A power of two or a constant between 0 and 32. This is used in
3955 // GCC for the shift amount on shifted register operands, but it is
3956 // useful in general for any shift amounts.
3957 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3963 if (Subtarget->isThumb()) { // FIXME thumb2
3964 // This must be a constant between 0 and 31, for shift amounts.
3965 if (CVal >= 0 && CVal <= 31)
3971 if (Subtarget->isThumb()) { // FIXME thumb2
3972 // This must be a multiple of 4 between -508 and 508, for
3973 // ADD/SUB sp = sp + immediate.
3974 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3979 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3983 if (Result.getNode()) {
3984 Ops.push_back(Result);
3987 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
3992 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3993 // The ARM target isn't yet aware of offsets.
3997 int ARM::getVFPf32Imm(const APFloat &FPImm) {
3998 APInt Imm = FPImm.bitcastToAPInt();
3999 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4000 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4001 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4003 // We can handle 4 bits of mantissa.
4004 // mantissa = (16+UInt(e:f:g:h))/16.
4005 if (Mantissa & 0x7ffff)
4008 if ((Mantissa & 0xf) != Mantissa)
4011 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4012 if (Exp < -3 || Exp > 4)
4014 Exp = ((Exp+3) & 0x7) ^ 4;
4016 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4019 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4020 APInt Imm = FPImm.bitcastToAPInt();
4021 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4022 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4023 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4025 // We can handle 4 bits of mantissa.
4026 // mantissa = (16+UInt(e:f:g:h))/16.
4027 if (Mantissa & 0xffffffffffffLL)
4030 if ((Mantissa & 0xf) != Mantissa)
4033 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4034 if (Exp < -3 || Exp > 4)
4036 Exp = ((Exp+3) & 0x7) ^ 4;
4038 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4041 /// isFPImmLegal - Returns true if the target can instruction select the
4042 /// specified FP immediate natively. If false, the legalizer will
4043 /// materialize the FP immediate as a load from a constant pool.
4044 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4045 if (!Subtarget->hasVFP3())
4048 return ARM::getVFPf32Imm(Imm) != -1;
4050 return ARM::getVFPf64Imm(Imm) != -1;