1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/Instruction.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Type.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/ADT/VectorExtras.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/MathExtras.h"
47 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
48 CCValAssign::LocInfo &LocInfo,
49 ISD::ArgFlagsTy &ArgFlags,
51 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
52 CCValAssign::LocInfo &LocInfo,
53 ISD::ArgFlagsTy &ArgFlags,
55 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
56 CCValAssign::LocInfo &LocInfo,
57 ISD::ArgFlagsTy &ArgFlags,
59 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
60 CCValAssign::LocInfo &LocInfo,
61 ISD::ArgFlagsTy &ArgFlags,
64 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
65 EVT PromotedBitwiseVT) {
66 if (VT != PromotedLdStVT) {
67 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
68 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
69 PromotedLdStVT.getSimpleVT());
71 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
72 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
73 PromotedLdStVT.getSimpleVT());
76 EVT ElemTy = VT.getVectorElementType();
77 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
78 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
79 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
80 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
81 if (ElemTy != MVT::i32) {
82 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
83 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
84 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
85 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
87 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
88 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
89 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
90 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
93 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
94 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
97 // Promote all bit-wise operations.
98 if (VT.isInteger() && VT != PromotedBitwiseVT) {
99 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
100 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
101 PromotedBitwiseVT.getSimpleVT());
102 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
103 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
104 PromotedBitwiseVT.getSimpleVT());
105 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
106 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
107 PromotedBitwiseVT.getSimpleVT());
110 // Neon does not support vector divide/remainder operations.
111 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
112 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
119 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
120 addRegisterClass(VT, ARM::DPRRegisterClass);
121 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
124 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
125 addRegisterClass(VT, ARM::QPRRegisterClass);
126 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
129 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
130 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
131 return new TargetLoweringObjectFileMachO();
132 return new ARMElfTargetObjectFile();
135 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
136 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
137 Subtarget = &TM.getSubtarget<ARMSubtarget>();
139 if (Subtarget->isTargetDarwin()) {
140 // Uses VFP for Thumb libfuncs if available.
141 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
142 // Single-precision floating-point arithmetic.
143 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
144 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
145 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
146 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
148 // Double-precision floating-point arithmetic.
149 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
150 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
151 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
152 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
154 // Single-precision comparisons.
155 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
156 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
157 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
158 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
159 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
160 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
161 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
162 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
164 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
165 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
166 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
167 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
173 // Double-precision comparisons.
174 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
175 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
176 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
177 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
178 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
179 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
180 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
181 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
183 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
184 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
192 // Floating-point to integer conversions.
193 // i64 conversions are done via library routines even when generating VFP
194 // instructions, so use the same ones.
195 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
196 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
197 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
198 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
200 // Conversions between floating types.
201 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
202 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
204 // Integer to floating-point conversions.
205 // i64 conversions are done via library routines even when generating VFP
206 // instructions, so use the same ones.
207 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
208 // e.g., __floatunsidf vs. __floatunssidfvfp.
209 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
210 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
211 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
212 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
216 // These libcalls are not available in 32-bit.
217 setLibcallName(RTLIB::SHL_I128, 0);
218 setLibcallName(RTLIB::SRL_I128, 0);
219 setLibcallName(RTLIB::SRA_I128, 0);
221 // Libcalls should use the AAPCS base standard ABI, even if hard float
222 // is in effect, as per the ARM RTABI specification, section 4.1.2.
223 if (Subtarget->isAAPCS_ABI()) {
224 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
225 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
226 CallingConv::ARM_AAPCS);
230 if (Subtarget->isThumb1Only())
231 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
233 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
234 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
235 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
236 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
238 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
241 if (Subtarget->hasNEON()) {
242 addDRTypeForNEON(MVT::v2f32);
243 addDRTypeForNEON(MVT::v8i8);
244 addDRTypeForNEON(MVT::v4i16);
245 addDRTypeForNEON(MVT::v2i32);
246 addDRTypeForNEON(MVT::v1i64);
248 addQRTypeForNEON(MVT::v4f32);
249 addQRTypeForNEON(MVT::v2f64);
250 addQRTypeForNEON(MVT::v16i8);
251 addQRTypeForNEON(MVT::v8i16);
252 addQRTypeForNEON(MVT::v4i32);
253 addQRTypeForNEON(MVT::v2i64);
255 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
256 // neither Neon nor VFP support any arithmetic operations on it.
257 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
258 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
259 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
260 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
261 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
262 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
263 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
264 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
265 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
266 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
267 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
268 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
269 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
270 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
271 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
272 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
273 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
274 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
275 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
276 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
277 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
278 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
279 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
280 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
282 // Neon does not support some operations on v1i64 and v2i64 types.
283 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
284 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
285 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
286 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
288 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
289 setTargetDAGCombine(ISD::SHL);
290 setTargetDAGCombine(ISD::SRL);
291 setTargetDAGCombine(ISD::SRA);
292 setTargetDAGCombine(ISD::SIGN_EXTEND);
293 setTargetDAGCombine(ISD::ZERO_EXTEND);
294 setTargetDAGCombine(ISD::ANY_EXTEND);
297 computeRegisterProperties();
299 // ARM does not have f32 extending load.
300 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
302 // ARM does not have i1 sign extending load.
303 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
305 // ARM supports all 4 flavors of integer indexed load / store.
306 if (!Subtarget->isThumb1Only()) {
307 for (unsigned im = (unsigned)ISD::PRE_INC;
308 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
309 setIndexedLoadAction(im, MVT::i1, Legal);
310 setIndexedLoadAction(im, MVT::i8, Legal);
311 setIndexedLoadAction(im, MVT::i16, Legal);
312 setIndexedLoadAction(im, MVT::i32, Legal);
313 setIndexedStoreAction(im, MVT::i1, Legal);
314 setIndexedStoreAction(im, MVT::i8, Legal);
315 setIndexedStoreAction(im, MVT::i16, Legal);
316 setIndexedStoreAction(im, MVT::i32, Legal);
320 // i64 operation support.
321 if (Subtarget->isThumb1Only()) {
322 setOperationAction(ISD::MUL, MVT::i64, Expand);
323 setOperationAction(ISD::MULHU, MVT::i32, Expand);
324 setOperationAction(ISD::MULHS, MVT::i32, Expand);
325 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
326 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
328 setOperationAction(ISD::MUL, MVT::i64, Expand);
329 setOperationAction(ISD::MULHU, MVT::i32, Expand);
330 if (!Subtarget->hasV6Ops())
331 setOperationAction(ISD::MULHS, MVT::i32, Expand);
333 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
334 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
335 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
336 setOperationAction(ISD::SRL, MVT::i64, Custom);
337 setOperationAction(ISD::SRA, MVT::i64, Custom);
339 // ARM does not have ROTL.
340 setOperationAction(ISD::ROTL, MVT::i32, Expand);
341 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
342 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
343 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
344 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
346 // Only ARMv6 has BSWAP.
347 if (!Subtarget->hasV6Ops())
348 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
350 // These are expanded into libcalls.
351 setOperationAction(ISD::SDIV, MVT::i32, Expand);
352 setOperationAction(ISD::UDIV, MVT::i32, Expand);
353 setOperationAction(ISD::SREM, MVT::i32, Expand);
354 setOperationAction(ISD::UREM, MVT::i32, Expand);
355 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
356 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
358 // Support label based line numbers.
359 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
360 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
362 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
363 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
364 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
365 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
366 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
368 // Use the default implementation.
369 setOperationAction(ISD::VASTART, MVT::Other, Custom);
370 setOperationAction(ISD::VAARG, MVT::Other, Expand);
371 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
372 setOperationAction(ISD::VAEND, MVT::Other, Expand);
373 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
374 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
375 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
376 // FIXME: Shouldn't need this, since no register is used, but the legalizer
377 // doesn't yet know how to not do that for SjLj.
378 setExceptionSelectorRegister(ARM::R0);
379 if (Subtarget->isThumb())
380 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
382 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
383 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
385 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
386 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
387 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
389 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
391 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
392 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
393 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
395 // We want to custom lower some of our intrinsics.
396 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
398 setOperationAction(ISD::SETCC, MVT::i32, Expand);
399 setOperationAction(ISD::SETCC, MVT::f32, Expand);
400 setOperationAction(ISD::SETCC, MVT::f64, Expand);
401 setOperationAction(ISD::SELECT, MVT::i32, Expand);
402 setOperationAction(ISD::SELECT, MVT::f32, Expand);
403 setOperationAction(ISD::SELECT, MVT::f64, Expand);
404 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
405 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
406 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
408 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
409 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
410 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
411 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
412 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
414 // We don't support sin/cos/fmod/copysign/pow
415 setOperationAction(ISD::FSIN, MVT::f64, Expand);
416 setOperationAction(ISD::FSIN, MVT::f32, Expand);
417 setOperationAction(ISD::FCOS, MVT::f32, Expand);
418 setOperationAction(ISD::FCOS, MVT::f64, Expand);
419 setOperationAction(ISD::FREM, MVT::f64, Expand);
420 setOperationAction(ISD::FREM, MVT::f32, Expand);
421 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
422 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
423 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
425 setOperationAction(ISD::FPOW, MVT::f64, Expand);
426 setOperationAction(ISD::FPOW, MVT::f32, Expand);
428 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
429 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
430 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
431 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
432 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
433 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
436 // We have target-specific dag combine patterns for the following nodes:
437 // ARMISD::FMRRD - No need to call setTargetDAGCombine
438 setTargetDAGCombine(ISD::ADD);
439 setTargetDAGCombine(ISD::SUB);
441 setStackPointerRegisterToSaveRestore(ARM::SP);
442 setSchedulingPreference(SchedulingForRegPressure);
444 // FIXME: If-converter should use instruction latency to determine
445 // profitability rather than relying on fixed limits.
446 if (Subtarget->getCPUString() == "generic") {
447 // Generic (and overly aggressive) if-conversion limits.
448 setIfCvtBlockSizeLimit(10);
449 setIfCvtDupBlockSizeLimit(2);
450 } else if (Subtarget->hasV6Ops()) {
451 setIfCvtBlockSizeLimit(2);
452 setIfCvtDupBlockSizeLimit(1);
454 setIfCvtBlockSizeLimit(3);
455 setIfCvtDupBlockSizeLimit(2);
458 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
459 // Do not enable CodePlacementOpt for now: it currently runs after the
460 // ARMConstantIslandPass and messes up branch relaxation and placement
461 // of constant islands.
462 // benefitFromCodePlacementOpt = true;
465 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
468 case ARMISD::Wrapper: return "ARMISD::Wrapper";
469 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
470 case ARMISD::CALL: return "ARMISD::CALL";
471 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
472 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
473 case ARMISD::tCALL: return "ARMISD::tCALL";
474 case ARMISD::BRCOND: return "ARMISD::BRCOND";
475 case ARMISD::BR_JT: return "ARMISD::BR_JT";
476 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
477 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
478 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
479 case ARMISD::CMP: return "ARMISD::CMP";
480 case ARMISD::CMPZ: return "ARMISD::CMPZ";
481 case ARMISD::CMPFP: return "ARMISD::CMPFP";
482 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
483 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
484 case ARMISD::CMOV: return "ARMISD::CMOV";
485 case ARMISD::CNEG: return "ARMISD::CNEG";
487 case ARMISD::FTOSI: return "ARMISD::FTOSI";
488 case ARMISD::FTOUI: return "ARMISD::FTOUI";
489 case ARMISD::SITOF: return "ARMISD::SITOF";
490 case ARMISD::UITOF: return "ARMISD::UITOF";
492 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
493 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
494 case ARMISD::RRX: return "ARMISD::RRX";
496 case ARMISD::FMRRD: return "ARMISD::FMRRD";
497 case ARMISD::FMDRR: return "ARMISD::FMDRR";
499 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
500 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
502 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
504 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
506 case ARMISD::VCEQ: return "ARMISD::VCEQ";
507 case ARMISD::VCGE: return "ARMISD::VCGE";
508 case ARMISD::VCGEU: return "ARMISD::VCGEU";
509 case ARMISD::VCGT: return "ARMISD::VCGT";
510 case ARMISD::VCGTU: return "ARMISD::VCGTU";
511 case ARMISD::VTST: return "ARMISD::VTST";
513 case ARMISD::VSHL: return "ARMISD::VSHL";
514 case ARMISD::VSHRs: return "ARMISD::VSHRs";
515 case ARMISD::VSHRu: return "ARMISD::VSHRu";
516 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
517 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
518 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
519 case ARMISD::VSHRN: return "ARMISD::VSHRN";
520 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
521 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
522 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
523 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
524 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
525 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
526 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
527 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
528 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
529 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
530 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
531 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
532 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
533 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
534 case ARMISD::VDUP: return "ARMISD::VDUP";
535 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
536 case ARMISD::VEXT: return "ARMISD::VEXT";
537 case ARMISD::VREV64: return "ARMISD::VREV64";
538 case ARMISD::VREV32: return "ARMISD::VREV32";
539 case ARMISD::VREV16: return "ARMISD::VREV16";
540 case ARMISD::VZIP: return "ARMISD::VZIP";
541 case ARMISD::VUZP: return "ARMISD::VUZP";
542 case ARMISD::VTRN: return "ARMISD::VTRN";
546 /// getFunctionAlignment - Return the Log2 alignment of this function.
547 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
548 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
551 //===----------------------------------------------------------------------===//
553 //===----------------------------------------------------------------------===//
555 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
556 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
558 default: llvm_unreachable("Unknown condition code!");
559 case ISD::SETNE: return ARMCC::NE;
560 case ISD::SETEQ: return ARMCC::EQ;
561 case ISD::SETGT: return ARMCC::GT;
562 case ISD::SETGE: return ARMCC::GE;
563 case ISD::SETLT: return ARMCC::LT;
564 case ISD::SETLE: return ARMCC::LE;
565 case ISD::SETUGT: return ARMCC::HI;
566 case ISD::SETUGE: return ARMCC::HS;
567 case ISD::SETULT: return ARMCC::LO;
568 case ISD::SETULE: return ARMCC::LS;
572 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
573 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
574 ARMCC::CondCodes &CondCode2) {
575 CondCode2 = ARMCC::AL;
577 default: llvm_unreachable("Unknown FP condition!");
579 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
581 case ISD::SETOGT: CondCode = ARMCC::GT; break;
583 case ISD::SETOGE: CondCode = ARMCC::GE; break;
584 case ISD::SETOLT: CondCode = ARMCC::MI; break;
585 case ISD::SETOLE: CondCode = ARMCC::LS; break;
586 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
587 case ISD::SETO: CondCode = ARMCC::VC; break;
588 case ISD::SETUO: CondCode = ARMCC::VS; break;
589 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
590 case ISD::SETUGT: CondCode = ARMCC::HI; break;
591 case ISD::SETUGE: CondCode = ARMCC::PL; break;
593 case ISD::SETULT: CondCode = ARMCC::LT; break;
595 case ISD::SETULE: CondCode = ARMCC::LE; break;
597 case ISD::SETUNE: CondCode = ARMCC::NE; break;
601 //===----------------------------------------------------------------------===//
602 // Calling Convention Implementation
603 //===----------------------------------------------------------------------===//
605 #include "ARMGenCallingConv.inc"
607 // APCS f64 is in register pairs, possibly split to stack
608 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
609 CCValAssign::LocInfo &LocInfo,
610 CCState &State, bool CanFail) {
611 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
613 // Try to get the first register.
614 if (unsigned Reg = State.AllocateReg(RegList, 4))
615 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
617 // For the 2nd half of a v2f64, do not fail.
621 // Put the whole thing on the stack.
622 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
623 State.AllocateStack(8, 4),
628 // Try to get the second register.
629 if (unsigned Reg = State.AllocateReg(RegList, 4))
630 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
632 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
633 State.AllocateStack(4, 4),
638 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
639 CCValAssign::LocInfo &LocInfo,
640 ISD::ArgFlagsTy &ArgFlags,
642 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
644 if (LocVT == MVT::v2f64 &&
645 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
647 return true; // we handled it
650 // AAPCS f64 is in aligned register pairs
651 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
652 CCValAssign::LocInfo &LocInfo,
653 CCState &State, bool CanFail) {
654 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
655 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
657 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
659 // For the 2nd half of a v2f64, do not just fail.
663 // Put the whole thing on the stack.
664 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
665 State.AllocateStack(8, 8),
671 for (i = 0; i < 2; ++i)
672 if (HiRegList[i] == Reg)
675 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
676 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
681 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
682 CCValAssign::LocInfo &LocInfo,
683 ISD::ArgFlagsTy &ArgFlags,
685 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
687 if (LocVT == MVT::v2f64 &&
688 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
690 return true; // we handled it
693 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
694 CCValAssign::LocInfo &LocInfo, CCState &State) {
695 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
696 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
698 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
700 return false; // we didn't handle it
703 for (i = 0; i < 2; ++i)
704 if (HiRegList[i] == Reg)
707 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
708 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
713 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
714 CCValAssign::LocInfo &LocInfo,
715 ISD::ArgFlagsTy &ArgFlags,
717 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
719 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
721 return true; // we handled it
724 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
725 CCValAssign::LocInfo &LocInfo,
726 ISD::ArgFlagsTy &ArgFlags,
728 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
732 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
733 /// given CallingConvention value.
734 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
736 bool isVarArg) const {
739 llvm_unreachable("Unsupported calling convention");
741 case CallingConv::Fast:
742 // Use target triple & subtarget features to do actual dispatch.
743 if (Subtarget->isAAPCS_ABI()) {
744 if (Subtarget->hasVFP2() &&
745 FloatABIType == FloatABI::Hard && !isVarArg)
746 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
748 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
750 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
751 case CallingConv::ARM_AAPCS_VFP:
752 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
753 case CallingConv::ARM_AAPCS:
754 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
755 case CallingConv::ARM_APCS:
756 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
760 /// LowerCallResult - Lower the result values of a call into the
761 /// appropriate copies out of appropriate physical registers.
763 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
764 CallingConv::ID CallConv, bool isVarArg,
765 const SmallVectorImpl<ISD::InputArg> &Ins,
766 DebugLoc dl, SelectionDAG &DAG,
767 SmallVectorImpl<SDValue> &InVals) {
769 // Assign locations to each value returned by this call.
770 SmallVector<CCValAssign, 16> RVLocs;
771 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
772 RVLocs, *DAG.getContext());
773 CCInfo.AnalyzeCallResult(Ins,
774 CCAssignFnForNode(CallConv, /* Return*/ true,
777 // Copy all of the result registers out of their specified physreg.
778 for (unsigned i = 0; i != RVLocs.size(); ++i) {
779 CCValAssign VA = RVLocs[i];
782 if (VA.needsCustom()) {
783 // Handle f64 or half of a v2f64.
784 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
786 Chain = Lo.getValue(1);
787 InFlag = Lo.getValue(2);
788 VA = RVLocs[++i]; // skip ahead to next loc
789 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
791 Chain = Hi.getValue(1);
792 InFlag = Hi.getValue(2);
793 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
795 if (VA.getLocVT() == MVT::v2f64) {
796 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
797 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
798 DAG.getConstant(0, MVT::i32));
800 VA = RVLocs[++i]; // skip ahead to next loc
801 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
802 Chain = Lo.getValue(1);
803 InFlag = Lo.getValue(2);
804 VA = RVLocs[++i]; // skip ahead to next loc
805 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
806 Chain = Hi.getValue(1);
807 InFlag = Hi.getValue(2);
808 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
809 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
810 DAG.getConstant(1, MVT::i32));
813 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
815 Chain = Val.getValue(1);
816 InFlag = Val.getValue(2);
819 switch (VA.getLocInfo()) {
820 default: llvm_unreachable("Unknown loc info!");
821 case CCValAssign::Full: break;
822 case CCValAssign::BCvt:
823 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
827 InVals.push_back(Val);
833 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
834 /// by "Src" to address "Dst" of size "Size". Alignment information is
835 /// specified by the specific parameter attribute. The copy will be passed as
836 /// a byval function parameter.
837 /// Sometimes what we are copying is the end of a larger object, the part that
838 /// does not fit in registers.
840 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
841 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
843 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
844 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
845 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
848 /// LowerMemOpCallTo - Store the argument to the stack.
850 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
851 SDValue StackPtr, SDValue Arg,
852 DebugLoc dl, SelectionDAG &DAG,
853 const CCValAssign &VA,
854 ISD::ArgFlagsTy Flags) {
855 unsigned LocMemOffset = VA.getLocMemOffset();
856 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
857 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
858 if (Flags.isByVal()) {
859 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
861 return DAG.getStore(Chain, dl, Arg, PtrOff,
862 PseudoSourceValue::getStack(), LocMemOffset);
865 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
866 SDValue Chain, SDValue &Arg,
867 RegsToPassVector &RegsToPass,
868 CCValAssign &VA, CCValAssign &NextVA,
870 SmallVector<SDValue, 8> &MemOpChains,
871 ISD::ArgFlagsTy Flags) {
873 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
874 DAG.getVTList(MVT::i32, MVT::i32), Arg);
875 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
877 if (NextVA.isRegLoc())
878 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
880 assert(NextVA.isMemLoc());
881 if (StackPtr.getNode() == 0)
882 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
884 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
890 /// LowerCall - Lowering a call into a callseq_start <-
891 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
894 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
895 CallingConv::ID CallConv, bool isVarArg,
897 const SmallVectorImpl<ISD::OutputArg> &Outs,
898 const SmallVectorImpl<ISD::InputArg> &Ins,
899 DebugLoc dl, SelectionDAG &DAG,
900 SmallVectorImpl<SDValue> &InVals) {
902 // Analyze operands of the call, assigning locations to each operand.
903 SmallVector<CCValAssign, 16> ArgLocs;
904 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
906 CCInfo.AnalyzeCallOperands(Outs,
907 CCAssignFnForNode(CallConv, /* Return*/ false,
910 // Get a count of how many bytes are to be pushed on the stack.
911 unsigned NumBytes = CCInfo.getNextStackOffset();
913 // Adjust the stack pointer for the new arguments...
914 // These operations are automatically eliminated by the prolog/epilog pass
915 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
917 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
919 RegsToPassVector RegsToPass;
920 SmallVector<SDValue, 8> MemOpChains;
922 // Walk the register/memloc assignments, inserting copies/loads. In the case
923 // of tail call optimization, arguments are handled later.
924 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
927 CCValAssign &VA = ArgLocs[i];
928 SDValue Arg = Outs[realArgIdx].Val;
929 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
931 // Promote the value if needed.
932 switch (VA.getLocInfo()) {
933 default: llvm_unreachable("Unknown loc info!");
934 case CCValAssign::Full: break;
935 case CCValAssign::SExt:
936 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
938 case CCValAssign::ZExt:
939 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
941 case CCValAssign::AExt:
942 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
944 case CCValAssign::BCvt:
945 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
949 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
950 if (VA.needsCustom()) {
951 if (VA.getLocVT() == MVT::v2f64) {
952 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
953 DAG.getConstant(0, MVT::i32));
954 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
955 DAG.getConstant(1, MVT::i32));
957 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
958 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
960 VA = ArgLocs[++i]; // skip ahead to next loc
962 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
963 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
965 assert(VA.isMemLoc());
966 if (StackPtr.getNode() == 0)
967 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
969 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
970 dl, DAG, VA, Flags));
973 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
974 StackPtr, MemOpChains, Flags);
976 } else if (VA.isRegLoc()) {
977 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
979 assert(VA.isMemLoc());
980 if (StackPtr.getNode() == 0)
981 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
983 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
984 dl, DAG, VA, Flags));
988 if (!MemOpChains.empty())
989 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
990 &MemOpChains[0], MemOpChains.size());
992 // Build a sequence of copy-to-reg nodes chained together with token chain
993 // and flag operands which copy the outgoing args into the appropriate regs.
995 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
996 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
997 RegsToPass[i].second, InFlag);
998 InFlag = Chain.getValue(1);
1001 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1002 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1003 // node so that legalize doesn't hack it.
1004 bool isDirect = false;
1005 bool isARMFunc = false;
1006 bool isLocalARMFunc = false;
1007 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1008 GlobalValue *GV = G->getGlobal();
1010 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1011 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1012 getTargetMachine().getRelocationModel() != Reloc::Static;
1013 isARMFunc = !Subtarget->isThumb() || isStub;
1014 // ARM call to a local ARM function is predicable.
1015 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1016 // tBX takes a register source operand.
1017 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1018 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1021 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1022 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1023 Callee = DAG.getLoad(getPointerTy(), dl,
1024 DAG.getEntryNode(), CPAddr,
1025 PseudoSourceValue::getConstantPool(), 0);
1026 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1027 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1028 getPointerTy(), Callee, PICLabel);
1030 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1031 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1033 bool isStub = Subtarget->isTargetDarwin() &&
1034 getTargetMachine().getRelocationModel() != Reloc::Static;
1035 isARMFunc = !Subtarget->isThumb() || isStub;
1036 // tBX takes a register source operand.
1037 const char *Sym = S->getSymbol();
1038 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1039 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1040 Sym, ARMPCLabelIndex, 4);
1041 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1042 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1043 Callee = DAG.getLoad(getPointerTy(), dl,
1044 DAG.getEntryNode(), CPAddr,
1045 PseudoSourceValue::getConstantPool(), 0);
1046 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1047 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1048 getPointerTy(), Callee, PICLabel);
1050 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1053 // FIXME: handle tail calls differently.
1055 if (Subtarget->isThumb()) {
1056 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1057 CallOpc = ARMISD::CALL_NOLINK;
1059 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1061 CallOpc = (isDirect || Subtarget->hasV5TOps())
1062 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1063 : ARMISD::CALL_NOLINK;
1065 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1066 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1067 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1068 InFlag = Chain.getValue(1);
1071 std::vector<SDValue> Ops;
1072 Ops.push_back(Chain);
1073 Ops.push_back(Callee);
1075 // Add argument registers to the end of the list so that they are known live
1077 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1078 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1079 RegsToPass[i].second.getValueType()));
1081 if (InFlag.getNode())
1082 Ops.push_back(InFlag);
1083 // Returns a chain and a flag for retval copy to use.
1084 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1085 &Ops[0], Ops.size());
1086 InFlag = Chain.getValue(1);
1088 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1089 DAG.getIntPtrConstant(0, true), InFlag);
1091 InFlag = Chain.getValue(1);
1093 // Handle result values, copying them out of physregs into vregs that we
1095 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1100 ARMTargetLowering::LowerReturn(SDValue Chain,
1101 CallingConv::ID CallConv, bool isVarArg,
1102 const SmallVectorImpl<ISD::OutputArg> &Outs,
1103 DebugLoc dl, SelectionDAG &DAG) {
1105 // CCValAssign - represent the assignment of the return value to a location.
1106 SmallVector<CCValAssign, 16> RVLocs;
1108 // CCState - Info about the registers and stack slots.
1109 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1112 // Analyze outgoing return values.
1113 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1116 // If this is the first return lowered for this function, add
1117 // the regs to the liveout set for the function.
1118 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1119 for (unsigned i = 0; i != RVLocs.size(); ++i)
1120 if (RVLocs[i].isRegLoc())
1121 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1126 // Copy the result values into the output registers.
1127 for (unsigned i = 0, realRVLocIdx = 0;
1129 ++i, ++realRVLocIdx) {
1130 CCValAssign &VA = RVLocs[i];
1131 assert(VA.isRegLoc() && "Can only return in registers!");
1133 SDValue Arg = Outs[realRVLocIdx].Val;
1135 switch (VA.getLocInfo()) {
1136 default: llvm_unreachable("Unknown loc info!");
1137 case CCValAssign::Full: break;
1138 case CCValAssign::BCvt:
1139 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1143 if (VA.needsCustom()) {
1144 if (VA.getLocVT() == MVT::v2f64) {
1145 // Extract the first half and return it in two registers.
1146 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1147 DAG.getConstant(0, MVT::i32));
1148 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1149 DAG.getVTList(MVT::i32, MVT::i32), Half);
1151 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1152 Flag = Chain.getValue(1);
1153 VA = RVLocs[++i]; // skip ahead to next loc
1154 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1155 HalfGPRs.getValue(1), Flag);
1156 Flag = Chain.getValue(1);
1157 VA = RVLocs[++i]; // skip ahead to next loc
1159 // Extract the 2nd half and fall through to handle it as an f64 value.
1160 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1161 DAG.getConstant(1, MVT::i32));
1163 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1165 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1166 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1167 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1168 Flag = Chain.getValue(1);
1169 VA = RVLocs[++i]; // skip ahead to next loc
1170 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1173 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1175 // Guarantee that all emitted copies are
1176 // stuck together, avoiding something bad.
1177 Flag = Chain.getValue(1);
1182 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1184 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1189 // ConstantPool, BlockAddress, JumpTable, GlobalAddress, and ExternalSymbol are
1190 // lowered as their target counterpart wrapped in the ARMISD::Wrapper
1191 // node. Suppose N is one of the above mentioned nodes. It has to be wrapped
1192 // because otherwise Select(N) returns N. So the raw TargetGlobalAddress
1193 // nodes, etc. can only be used to form addressing mode. These wrapped nodes
1194 // will be selected into MOVi.
1195 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1196 EVT PtrVT = Op.getValueType();
1197 // FIXME there is no actual debug info here
1198 DebugLoc dl = Op.getDebugLoc();
1199 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1201 if (CP->isMachineConstantPoolEntry())
1202 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1203 CP->getAlignment());
1205 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1206 CP->getAlignment());
1207 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1210 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1211 DebugLoc DL = Op.getDebugLoc();
1212 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1213 SDValue Result = DAG.getBlockAddress(BA, DL, /*isTarget=*/true);
1214 return DAG.getNode(ARMISD::Wrapper, DL, getPointerTy(), Result);
1217 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1219 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1220 SelectionDAG &DAG) {
1221 DebugLoc dl = GA->getDebugLoc();
1222 EVT PtrVT = getPointerTy();
1223 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1224 ARMConstantPoolValue *CPV =
1225 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1226 ARMCP::CPValue, PCAdj, "tlsgd", true);
1227 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1228 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1229 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1230 PseudoSourceValue::getConstantPool(), 0);
1231 SDValue Chain = Argument.getValue(1);
1233 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1234 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1236 // call __tls_get_addr.
1239 Entry.Node = Argument;
1240 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1241 Args.push_back(Entry);
1242 // FIXME: is there useful debug info available here?
1243 std::pair<SDValue, SDValue> CallResult =
1244 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1245 false, false, false, false,
1246 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1247 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1248 return CallResult.first;
1251 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1252 // "local exec" model.
1254 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1255 SelectionDAG &DAG) {
1256 GlobalValue *GV = GA->getGlobal();
1257 DebugLoc dl = GA->getDebugLoc();
1259 SDValue Chain = DAG.getEntryNode();
1260 EVT PtrVT = getPointerTy();
1261 // Get the Thread Pointer
1262 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1264 if (GV->isDeclaration()) {
1265 // initial exec model
1266 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1267 ARMConstantPoolValue *CPV =
1268 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1269 ARMCP::CPValue, PCAdj, "gottpoff", true);
1270 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1271 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1272 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1273 PseudoSourceValue::getConstantPool(), 0);
1274 Chain = Offset.getValue(1);
1276 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1277 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1279 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1280 PseudoSourceValue::getConstantPool(), 0);
1283 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1284 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1285 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1286 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1287 PseudoSourceValue::getConstantPool(), 0);
1290 // The address of the thread local variable is the add of the thread
1291 // pointer with the offset of the variable.
1292 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1296 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1297 // TODO: implement the "local dynamic" model
1298 assert(Subtarget->isTargetELF() &&
1299 "TLS not implemented for non-ELF targets");
1300 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1301 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1302 // otherwise use the "Local Exec" TLS Model
1303 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1304 return LowerToTLSGeneralDynamicModel(GA, DAG);
1306 return LowerToTLSExecModels(GA, DAG);
1309 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1310 SelectionDAG &DAG) {
1311 EVT PtrVT = getPointerTy();
1312 DebugLoc dl = Op.getDebugLoc();
1313 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1314 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1315 if (RelocM == Reloc::PIC_) {
1316 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1317 ARMConstantPoolValue *CPV =
1318 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1319 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1320 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1321 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1323 PseudoSourceValue::getConstantPool(), 0);
1324 SDValue Chain = Result.getValue(1);
1325 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1326 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1328 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1329 PseudoSourceValue::getGOT(), 0);
1332 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1333 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1334 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1335 PseudoSourceValue::getConstantPool(), 0);
1339 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1340 SelectionDAG &DAG) {
1341 EVT PtrVT = getPointerTy();
1342 DebugLoc dl = Op.getDebugLoc();
1343 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1344 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1346 if (RelocM == Reloc::Static)
1347 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1349 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1350 ARMConstantPoolValue *CPV =
1351 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1352 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1354 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1356 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1357 PseudoSourceValue::getConstantPool(), 0);
1358 SDValue Chain = Result.getValue(1);
1360 if (RelocM == Reloc::PIC_) {
1361 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1362 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1365 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1366 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1367 PseudoSourceValue::getGOT(), 0);
1372 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1374 assert(Subtarget->isTargetELF() &&
1375 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1376 EVT PtrVT = getPointerTy();
1377 DebugLoc dl = Op.getDebugLoc();
1378 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1379 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1380 "_GLOBAL_OFFSET_TABLE_",
1381 ARMPCLabelIndex, PCAdj);
1382 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1383 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1384 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1385 PseudoSourceValue::getConstantPool(), 0);
1386 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1387 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1391 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1392 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1393 DebugLoc dl = Op.getDebugLoc();
1395 default: return SDValue(); // Don't custom lower most intrinsics.
1396 case Intrinsic::arm_thread_pointer: {
1397 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1398 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1400 case Intrinsic::eh_sjlj_lsda: {
1401 MachineFunction &MF = DAG.getMachineFunction();
1402 EVT PtrVT = getPointerTy();
1403 DebugLoc dl = Op.getDebugLoc();
1404 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1406 unsigned PCAdj = (RelocM != Reloc::PIC_)
1407 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1408 ARMConstantPoolValue *CPV =
1409 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1410 ARMCP::CPLSDA, PCAdj);
1411 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1412 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1414 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1415 PseudoSourceValue::getConstantPool(), 0);
1416 SDValue Chain = Result.getValue(1);
1418 if (RelocM == Reloc::PIC_) {
1419 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1420 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1424 case Intrinsic::eh_sjlj_setjmp:
1425 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1429 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1430 unsigned VarArgsFrameIndex) {
1431 // vastart just stores the address of the VarArgsFrameIndex slot into the
1432 // memory location argument.
1433 DebugLoc dl = Op.getDebugLoc();
1434 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1435 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1436 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1437 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1441 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1442 SDNode *Node = Op.getNode();
1443 DebugLoc dl = Node->getDebugLoc();
1444 EVT VT = Node->getValueType(0);
1445 SDValue Chain = Op.getOperand(0);
1446 SDValue Size = Op.getOperand(1);
1447 SDValue Align = Op.getOperand(2);
1449 // Chain the dynamic stack allocation so that it doesn't modify the stack
1450 // pointer when other instructions are using the stack.
1451 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1453 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1454 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1455 if (AlignVal > StackAlign)
1456 // Do this now since selection pass cannot introduce new target
1457 // independent node.
1458 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1460 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1461 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1462 // do even more horrible hack later.
1463 MachineFunction &MF = DAG.getMachineFunction();
1464 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1465 if (AFI->isThumb1OnlyFunction()) {
1467 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1469 uint32_t Val = C->getZExtValue();
1470 if (Val <= 508 && ((Val & 3) == 0))
1474 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1477 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1478 SDValue Ops1[] = { Chain, Size, Align };
1479 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1480 Chain = Res.getValue(1);
1481 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1482 DAG.getIntPtrConstant(0, true), SDValue());
1483 SDValue Ops2[] = { Res, Chain };
1484 return DAG.getMergeValues(Ops2, 2, dl);
1488 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1489 SDValue &Root, SelectionDAG &DAG,
1491 MachineFunction &MF = DAG.getMachineFunction();
1492 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1494 TargetRegisterClass *RC;
1495 if (AFI->isThumb1OnlyFunction())
1496 RC = ARM::tGPRRegisterClass;
1498 RC = ARM::GPRRegisterClass;
1500 // Transform the arguments stored in physical registers into virtual ones.
1501 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1502 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1505 if (NextVA.isMemLoc()) {
1506 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1507 MachineFrameInfo *MFI = MF.getFrameInfo();
1508 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1510 // Create load node to retrieve arguments from the stack.
1511 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1512 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1513 PseudoSourceValue::getFixedStack(FI), 0);
1515 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1516 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1519 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1523 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1524 CallingConv::ID CallConv, bool isVarArg,
1525 const SmallVectorImpl<ISD::InputArg>
1527 DebugLoc dl, SelectionDAG &DAG,
1528 SmallVectorImpl<SDValue> &InVals) {
1530 MachineFunction &MF = DAG.getMachineFunction();
1531 MachineFrameInfo *MFI = MF.getFrameInfo();
1533 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1535 // Assign locations to all of the incoming arguments.
1536 SmallVector<CCValAssign, 16> ArgLocs;
1537 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1539 CCInfo.AnalyzeFormalArguments(Ins,
1540 CCAssignFnForNode(CallConv, /* Return*/ false,
1543 SmallVector<SDValue, 16> ArgValues;
1545 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1546 CCValAssign &VA = ArgLocs[i];
1548 // Arguments stored in registers.
1549 if (VA.isRegLoc()) {
1550 EVT RegVT = VA.getLocVT();
1553 if (VA.needsCustom()) {
1554 // f64 and vector types are split up into multiple registers or
1555 // combinations of registers and stack slots.
1558 if (VA.getLocVT() == MVT::v2f64) {
1559 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1561 VA = ArgLocs[++i]; // skip ahead to next loc
1562 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1564 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1565 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1566 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1567 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1568 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1570 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1573 TargetRegisterClass *RC;
1575 if (RegVT == MVT::f32)
1576 RC = ARM::SPRRegisterClass;
1577 else if (RegVT == MVT::f64)
1578 RC = ARM::DPRRegisterClass;
1579 else if (RegVT == MVT::v2f64)
1580 RC = ARM::QPRRegisterClass;
1581 else if (RegVT == MVT::i32)
1582 RC = (AFI->isThumb1OnlyFunction() ?
1583 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1585 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1587 // Transform the arguments in physical registers into virtual ones.
1588 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1589 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1592 // If this is an 8 or 16-bit value, it is really passed promoted
1593 // to 32 bits. Insert an assert[sz]ext to capture this, then
1594 // truncate to the right size.
1595 switch (VA.getLocInfo()) {
1596 default: llvm_unreachable("Unknown loc info!");
1597 case CCValAssign::Full: break;
1598 case CCValAssign::BCvt:
1599 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1601 case CCValAssign::SExt:
1602 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1603 DAG.getValueType(VA.getValVT()));
1604 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1606 case CCValAssign::ZExt:
1607 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1608 DAG.getValueType(VA.getValVT()));
1609 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1613 InVals.push_back(ArgValue);
1615 } else { // VA.isRegLoc()
1618 assert(VA.isMemLoc());
1619 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1621 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1622 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1624 // Create load nodes to retrieve arguments from the stack.
1625 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1626 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1627 PseudoSourceValue::getFixedStack(FI), 0));
1633 static const unsigned GPRArgRegs[] = {
1634 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1637 unsigned NumGPRs = CCInfo.getFirstUnallocated
1638 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1640 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1641 unsigned VARegSize = (4 - NumGPRs) * 4;
1642 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1643 unsigned ArgOffset = CCInfo.getNextStackOffset();
1644 if (VARegSaveSize) {
1645 // If this function is vararg, store any remaining integer argument regs
1646 // to their spots on the stack so that they may be loaded by deferencing
1647 // the result of va_next.
1648 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1649 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1650 VARegSaveSize - VARegSize);
1651 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1653 SmallVector<SDValue, 4> MemOps;
1654 for (; NumGPRs < 4; ++NumGPRs) {
1655 TargetRegisterClass *RC;
1656 if (AFI->isThumb1OnlyFunction())
1657 RC = ARM::tGPRRegisterClass;
1659 RC = ARM::GPRRegisterClass;
1661 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1662 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1663 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1664 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0);
1665 MemOps.push_back(Store);
1666 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1667 DAG.getConstant(4, getPointerTy()));
1669 if (!MemOps.empty())
1670 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1671 &MemOps[0], MemOps.size());
1673 // This will point to the next argument passed via stack.
1674 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1680 /// isFloatingPointZero - Return true if this is +0.0.
1681 static bool isFloatingPointZero(SDValue Op) {
1682 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1683 return CFP->getValueAPF().isPosZero();
1684 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1685 // Maybe this has already been legalized into the constant pool?
1686 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1687 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1688 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1689 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1690 return CFP->getValueAPF().isPosZero();
1696 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1697 return ( isThumb1Only && (C & ~255U) == 0) ||
1698 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1701 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1702 /// the given operands.
1703 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1704 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1706 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1707 unsigned C = RHSC->getZExtValue();
1708 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1709 // Constant does not fit, try adjusting it by one?
1714 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1715 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1716 RHS = DAG.getConstant(C-1, MVT::i32);
1721 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1722 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1723 RHS = DAG.getConstant(C-1, MVT::i32);
1728 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1729 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1730 RHS = DAG.getConstant(C+1, MVT::i32);
1735 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1736 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1737 RHS = DAG.getConstant(C+1, MVT::i32);
1744 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1745 ARMISD::NodeType CompareType;
1748 CompareType = ARMISD::CMP;
1753 CompareType = ARMISD::CMPZ;
1756 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1757 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1760 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1761 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1764 if (!isFloatingPointZero(RHS))
1765 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1767 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1768 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1771 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1772 const ARMSubtarget *ST) {
1773 EVT VT = Op.getValueType();
1774 SDValue LHS = Op.getOperand(0);
1775 SDValue RHS = Op.getOperand(1);
1776 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1777 SDValue TrueVal = Op.getOperand(2);
1778 SDValue FalseVal = Op.getOperand(3);
1779 DebugLoc dl = Op.getDebugLoc();
1781 if (LHS.getValueType() == MVT::i32) {
1783 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1784 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1785 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1788 ARMCC::CondCodes CondCode, CondCode2;
1789 FPCCToARMCC(CC, CondCode, CondCode2);
1791 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1792 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1793 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1794 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1796 if (CondCode2 != ARMCC::AL) {
1797 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1798 // FIXME: Needs another CMP because flag can have but one use.
1799 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1800 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1801 Result, TrueVal, ARMCC2, CCR, Cmp2);
1806 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1807 const ARMSubtarget *ST) {
1808 SDValue Chain = Op.getOperand(0);
1809 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1810 SDValue LHS = Op.getOperand(2);
1811 SDValue RHS = Op.getOperand(3);
1812 SDValue Dest = Op.getOperand(4);
1813 DebugLoc dl = Op.getDebugLoc();
1815 if (LHS.getValueType() == MVT::i32) {
1817 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1818 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1819 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1820 Chain, Dest, ARMCC, CCR,Cmp);
1823 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1824 ARMCC::CondCodes CondCode, CondCode2;
1825 FPCCToARMCC(CC, CondCode, CondCode2);
1827 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1828 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1829 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1830 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1831 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1832 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1833 if (CondCode2 != ARMCC::AL) {
1834 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1835 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1836 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1841 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1842 SDValue Chain = Op.getOperand(0);
1843 SDValue Table = Op.getOperand(1);
1844 SDValue Index = Op.getOperand(2);
1845 DebugLoc dl = Op.getDebugLoc();
1847 EVT PTy = getPointerTy();
1848 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1849 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1850 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1851 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1852 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1853 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1854 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1855 if (Subtarget->isThumb2()) {
1856 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1857 // which does another jump to the destination. This also makes it easier
1858 // to translate it to TBB / TBH later.
1859 // FIXME: This might not work if the function is extremely large.
1860 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1861 Addr, Op.getOperand(2), JTI, UId);
1863 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1864 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1865 PseudoSourceValue::getJumpTable(), 0);
1866 Chain = Addr.getValue(1);
1867 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1868 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1870 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1871 PseudoSourceValue::getJumpTable(), 0);
1872 Chain = Addr.getValue(1);
1873 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1877 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1878 DebugLoc dl = Op.getDebugLoc();
1880 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1881 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1882 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1885 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1886 EVT VT = Op.getValueType();
1887 DebugLoc dl = Op.getDebugLoc();
1889 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1891 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1892 return DAG.getNode(Opc, dl, VT, Op);
1895 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1896 // Implement fcopysign with a fabs and a conditional fneg.
1897 SDValue Tmp0 = Op.getOperand(0);
1898 SDValue Tmp1 = Op.getOperand(1);
1899 DebugLoc dl = Op.getDebugLoc();
1900 EVT VT = Op.getValueType();
1901 EVT SrcVT = Tmp1.getValueType();
1902 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1903 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1904 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1905 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1906 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1909 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1910 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1911 MFI->setFrameAddressIsTaken(true);
1912 EVT VT = Op.getValueType();
1913 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1914 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1915 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1916 ? ARM::R7 : ARM::R11;
1917 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1919 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1924 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1926 SDValue Dst, SDValue Src,
1927 SDValue Size, unsigned Align,
1929 const Value *DstSV, uint64_t DstSVOff,
1930 const Value *SrcSV, uint64_t SrcSVOff){
1931 // Do repeated 4-byte loads and stores. To be improved.
1932 // This requires 4-byte alignment.
1933 if ((Align & 3) != 0)
1935 // This requires the copy size to be a constant, preferrably
1936 // within a subtarget-specific limit.
1937 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1940 uint64_t SizeVal = ConstantSize->getZExtValue();
1941 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1944 unsigned BytesLeft = SizeVal & 3;
1945 unsigned NumMemOps = SizeVal >> 2;
1946 unsigned EmittedNumMemOps = 0;
1948 unsigned VTSize = 4;
1950 const unsigned MAX_LOADS_IN_LDM = 6;
1951 SDValue TFOps[MAX_LOADS_IN_LDM];
1952 SDValue Loads[MAX_LOADS_IN_LDM];
1953 uint64_t SrcOff = 0, DstOff = 0;
1955 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1956 // same number of stores. The loads and stores will get combined into
1957 // ldm/stm later on.
1958 while (EmittedNumMemOps < NumMemOps) {
1960 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1961 Loads[i] = DAG.getLoad(VT, dl, Chain,
1962 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1963 DAG.getConstant(SrcOff, MVT::i32)),
1964 SrcSV, SrcSVOff + SrcOff);
1965 TFOps[i] = Loads[i].getValue(1);
1968 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1971 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1972 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1973 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1974 DAG.getConstant(DstOff, MVT::i32)),
1975 DstSV, DstSVOff + DstOff);
1978 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1980 EmittedNumMemOps += i;
1986 // Issue loads / stores for the trailing (1 - 3) bytes.
1987 unsigned BytesLeftSave = BytesLeft;
1990 if (BytesLeft >= 2) {
1998 Loads[i] = DAG.getLoad(VT, dl, Chain,
1999 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2000 DAG.getConstant(SrcOff, MVT::i32)),
2001 SrcSV, SrcSVOff + SrcOff);
2002 TFOps[i] = Loads[i].getValue(1);
2005 BytesLeft -= VTSize;
2007 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2010 BytesLeft = BytesLeftSave;
2012 if (BytesLeft >= 2) {
2020 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2021 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2022 DAG.getConstant(DstOff, MVT::i32)),
2023 DstSV, DstSVOff + DstOff);
2026 BytesLeft -= VTSize;
2028 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2031 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2032 SDValue Op = N->getOperand(0);
2033 DebugLoc dl = N->getDebugLoc();
2034 if (N->getValueType(0) == MVT::f64) {
2035 // Turn i64->f64 into FMDRR.
2036 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2037 DAG.getConstant(0, MVT::i32));
2038 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2039 DAG.getConstant(1, MVT::i32));
2040 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2043 // Turn f64->i64 into FMRRD.
2044 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2045 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2047 // Merge the pieces into a single i64 value.
2048 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2051 /// getZeroVector - Returns a vector of specified type with all zero elements.
2053 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2054 assert(VT.isVector() && "Expected a vector type");
2056 // Zero vectors are used to represent vector negation and in those cases
2057 // will be implemented with the NEON VNEG instruction. However, VNEG does
2058 // not support i64 elements, so sometimes the zero vectors will need to be
2059 // explicitly constructed. For those cases, and potentially other uses in
2060 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2061 // to their dest type. This ensures they get CSE'd.
2063 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2064 SmallVector<SDValue, 8> Ops;
2067 if (VT.getSizeInBits() == 64) {
2068 Ops.assign(8, Cst); TVT = MVT::v8i8;
2070 Ops.assign(16, Cst); TVT = MVT::v16i8;
2072 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2074 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2077 /// getOnesVector - Returns a vector of specified type with all bits set.
2079 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2080 assert(VT.isVector() && "Expected a vector type");
2082 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2083 // dest type. This ensures they get CSE'd.
2085 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2086 SmallVector<SDValue, 8> Ops;
2089 if (VT.getSizeInBits() == 64) {
2090 Ops.assign(8, Cst); TVT = MVT::v8i8;
2092 Ops.assign(16, Cst); TVT = MVT::v16i8;
2094 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2096 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2099 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2100 const ARMSubtarget *ST) {
2101 EVT VT = N->getValueType(0);
2102 DebugLoc dl = N->getDebugLoc();
2104 // Lower vector shifts on NEON to use VSHL.
2105 if (VT.isVector()) {
2106 assert(ST->hasNEON() && "unexpected vector shift");
2108 // Left shifts translate directly to the vshiftu intrinsic.
2109 if (N->getOpcode() == ISD::SHL)
2110 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2111 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2112 N->getOperand(0), N->getOperand(1));
2114 assert((N->getOpcode() == ISD::SRA ||
2115 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2117 // NEON uses the same intrinsics for both left and right shifts. For
2118 // right shifts, the shift amounts are negative, so negate the vector of
2120 EVT ShiftVT = N->getOperand(1).getValueType();
2121 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2122 getZeroVector(ShiftVT, DAG, dl),
2124 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2125 Intrinsic::arm_neon_vshifts :
2126 Intrinsic::arm_neon_vshiftu);
2127 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2128 DAG.getConstant(vshiftInt, MVT::i32),
2129 N->getOperand(0), NegatedCount);
2132 // We can get here for a node like i32 = ISD::SHL i32, i64
2136 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2137 "Unknown shift to lower!");
2139 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2140 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2141 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2144 // If we are in thumb mode, we don't have RRX.
2145 if (ST->isThumb1Only()) return SDValue();
2147 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2148 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2149 DAG.getConstant(0, MVT::i32));
2150 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2151 DAG.getConstant(1, MVT::i32));
2153 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2154 // captures the result into a carry flag.
2155 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2156 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2158 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2159 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2161 // Merge the pieces into a single i64 value.
2162 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2165 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2166 SDValue TmpOp0, TmpOp1;
2167 bool Invert = false;
2171 SDValue Op0 = Op.getOperand(0);
2172 SDValue Op1 = Op.getOperand(1);
2173 SDValue CC = Op.getOperand(2);
2174 EVT VT = Op.getValueType();
2175 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2176 DebugLoc dl = Op.getDebugLoc();
2178 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2179 switch (SetCCOpcode) {
2180 default: llvm_unreachable("Illegal FP comparison"); break;
2182 case ISD::SETNE: Invert = true; // Fallthrough
2184 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2186 case ISD::SETLT: Swap = true; // Fallthrough
2188 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2190 case ISD::SETLE: Swap = true; // Fallthrough
2192 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2193 case ISD::SETUGE: Swap = true; // Fallthrough
2194 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2195 case ISD::SETUGT: Swap = true; // Fallthrough
2196 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2197 case ISD::SETUEQ: Invert = true; // Fallthrough
2199 // Expand this to (OLT | OGT).
2203 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2204 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2206 case ISD::SETUO: Invert = true; // Fallthrough
2208 // Expand this to (OLT | OGE).
2212 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2213 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2217 // Integer comparisons.
2218 switch (SetCCOpcode) {
2219 default: llvm_unreachable("Illegal integer comparison"); break;
2220 case ISD::SETNE: Invert = true;
2221 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2222 case ISD::SETLT: Swap = true;
2223 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2224 case ISD::SETLE: Swap = true;
2225 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2226 case ISD::SETULT: Swap = true;
2227 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2228 case ISD::SETULE: Swap = true;
2229 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2232 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2233 if (Opc == ARMISD::VCEQ) {
2236 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2238 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2241 // Ignore bitconvert.
2242 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2243 AndOp = AndOp.getOperand(0);
2245 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2247 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2248 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2255 std::swap(Op0, Op1);
2257 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2260 Result = DAG.getNOT(dl, Result, VT);
2265 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2266 /// VMOV instruction, and if so, return the constant being splatted.
2267 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2268 unsigned SplatBitSize, SelectionDAG &DAG) {
2269 switch (SplatBitSize) {
2271 // Any 1-byte value is OK.
2272 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2273 return DAG.getTargetConstant(SplatBits, MVT::i8);
2276 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2277 if ((SplatBits & ~0xff) == 0 ||
2278 (SplatBits & ~0xff00) == 0)
2279 return DAG.getTargetConstant(SplatBits, MVT::i16);
2283 // NEON's 32-bit VMOV supports splat values where:
2284 // * only one byte is nonzero, or
2285 // * the least significant byte is 0xff and the second byte is nonzero, or
2286 // * the least significant 2 bytes are 0xff and the third is nonzero.
2287 if ((SplatBits & ~0xff) == 0 ||
2288 (SplatBits & ~0xff00) == 0 ||
2289 (SplatBits & ~0xff0000) == 0 ||
2290 (SplatBits & ~0xff000000) == 0)
2291 return DAG.getTargetConstant(SplatBits, MVT::i32);
2293 if ((SplatBits & ~0xffff) == 0 &&
2294 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2295 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2297 if ((SplatBits & ~0xffffff) == 0 &&
2298 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2299 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2301 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2302 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2303 // VMOV.I32. A (very) minor optimization would be to replicate the value
2304 // and fall through here to test for a valid 64-bit splat. But, then the
2305 // caller would also need to check and handle the change in size.
2309 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2310 uint64_t BitMask = 0xff;
2312 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2313 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2315 else if ((SplatBits & BitMask) != 0)
2319 return DAG.getTargetConstant(Val, MVT::i64);
2323 llvm_unreachable("unexpected size for isVMOVSplat");
2330 /// getVMOVImm - If this is a build_vector of constants which can be
2331 /// formed by using a VMOV instruction of the specified element size,
2332 /// return the constant being splatted. The ByteSize field indicates the
2333 /// number of bytes of each element [1248].
2334 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2335 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2336 APInt SplatBits, SplatUndef;
2337 unsigned SplatBitSize;
2339 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2340 HasAnyUndefs, ByteSize * 8))
2343 if (SplatBitSize > ByteSize * 8)
2346 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2350 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2351 bool &ReverseVEXT, unsigned &Imm) {
2352 unsigned NumElts = VT.getVectorNumElements();
2353 ReverseVEXT = false;
2356 // If this is a VEXT shuffle, the immediate value is the index of the first
2357 // element. The other shuffle indices must be the successive elements after
2359 unsigned ExpectedElt = Imm;
2360 for (unsigned i = 1; i < NumElts; ++i) {
2361 // Increment the expected index. If it wraps around, it may still be
2362 // a VEXT but the source vectors must be swapped.
2364 if (ExpectedElt == NumElts * 2) {
2369 if (ExpectedElt != static_cast<unsigned>(M[i]))
2373 // Adjust the index value if the source operands will be swapped.
2380 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2381 /// instruction with the specified blocksize. (The order of the elements
2382 /// within each block of the vector is reversed.)
2383 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2384 unsigned BlockSize) {
2385 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2386 "Only possible block sizes for VREV are: 16, 32, 64");
2388 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2392 unsigned NumElts = VT.getVectorNumElements();
2393 unsigned BlockElts = M[0] + 1;
2395 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2398 for (unsigned i = 0; i < NumElts; ++i) {
2399 if ((unsigned) M[i] !=
2400 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2407 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2408 unsigned &WhichResult) {
2409 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2413 unsigned NumElts = VT.getVectorNumElements();
2414 WhichResult = (M[0] == 0 ? 0 : 1);
2415 for (unsigned i = 0; i < NumElts; i += 2) {
2416 if ((unsigned) M[i] != i + WhichResult ||
2417 (unsigned) M[i+1] != i + NumElts + WhichResult)
2423 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2424 unsigned &WhichResult) {
2425 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2429 unsigned NumElts = VT.getVectorNumElements();
2430 WhichResult = (M[0] == 0 ? 0 : 1);
2431 for (unsigned i = 0; i != NumElts; ++i) {
2432 if ((unsigned) M[i] != 2 * i + WhichResult)
2436 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2437 if (VT.is64BitVector() && EltSz == 32)
2443 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2444 unsigned &WhichResult) {
2445 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2449 unsigned NumElts = VT.getVectorNumElements();
2450 WhichResult = (M[0] == 0 ? 0 : 1);
2451 unsigned Idx = WhichResult * NumElts / 2;
2452 for (unsigned i = 0; i != NumElts; i += 2) {
2453 if ((unsigned) M[i] != Idx ||
2454 (unsigned) M[i+1] != Idx + NumElts)
2459 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2460 if (VT.is64BitVector() && EltSz == 32)
2466 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2467 // Canonicalize all-zeros and all-ones vectors.
2468 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2469 if (ConstVal->isNullValue())
2470 return getZeroVector(VT, DAG, dl);
2471 if (ConstVal->isAllOnesValue())
2472 return getOnesVector(VT, DAG, dl);
2475 if (VT.is64BitVector()) {
2476 switch (Val.getValueType().getSizeInBits()) {
2477 case 8: CanonicalVT = MVT::v8i8; break;
2478 case 16: CanonicalVT = MVT::v4i16; break;
2479 case 32: CanonicalVT = MVT::v2i32; break;
2480 case 64: CanonicalVT = MVT::v1i64; break;
2481 default: llvm_unreachable("unexpected splat element type"); break;
2484 assert(VT.is128BitVector() && "unknown splat vector size");
2485 switch (Val.getValueType().getSizeInBits()) {
2486 case 8: CanonicalVT = MVT::v16i8; break;
2487 case 16: CanonicalVT = MVT::v8i16; break;
2488 case 32: CanonicalVT = MVT::v4i32; break;
2489 case 64: CanonicalVT = MVT::v2i64; break;
2490 default: llvm_unreachable("unexpected splat element type"); break;
2494 // Build a canonical splat for this value.
2495 SmallVector<SDValue, 8> Ops;
2496 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2497 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2499 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2502 // If this is a case we can't handle, return null and let the default
2503 // expansion code take care of it.
2504 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2505 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2506 DebugLoc dl = Op.getDebugLoc();
2507 EVT VT = Op.getValueType();
2509 APInt SplatBits, SplatUndef;
2510 unsigned SplatBitSize;
2512 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2513 if (SplatBitSize <= 64) {
2514 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2515 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2517 return BuildSplat(Val, VT, DAG, dl);
2521 // If there are only 2 elements in a 128-bit vector, insert them into an
2522 // undef vector. This handles the common case for 128-bit vector argument
2523 // passing, where the insertions should be translated to subreg accesses
2524 // with no real instructions.
2525 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2526 SDValue Val = DAG.getUNDEF(VT);
2527 SDValue Op0 = Op.getOperand(0);
2528 SDValue Op1 = Op.getOperand(1);
2529 if (Op0.getOpcode() != ISD::UNDEF)
2530 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2531 DAG.getIntPtrConstant(0));
2532 if (Op1.getOpcode() != ISD::UNDEF)
2533 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2534 DAG.getIntPtrConstant(1));
2541 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2542 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2543 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2544 /// are assumed to be legal.
2546 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2548 if (VT.getVectorNumElements() == 4 &&
2549 (VT.is128BitVector() || VT.is64BitVector())) {
2550 unsigned PFIndexes[4];
2551 for (unsigned i = 0; i != 4; ++i) {
2555 PFIndexes[i] = M[i];
2558 // Compute the index in the perfect shuffle table.
2559 unsigned PFTableIndex =
2560 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2561 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2562 unsigned Cost = (PFEntry >> 30);
2569 unsigned Imm, WhichResult;
2571 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2572 isVREVMask(M, VT, 64) ||
2573 isVREVMask(M, VT, 32) ||
2574 isVREVMask(M, VT, 16) ||
2575 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2576 isVTRNMask(M, VT, WhichResult) ||
2577 isVUZPMask(M, VT, WhichResult) ||
2578 isVZIPMask(M, VT, WhichResult));
2581 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2582 /// the specified operations to build the shuffle.
2583 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2584 SDValue RHS, SelectionDAG &DAG,
2586 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2587 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2588 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2591 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2600 OP_VUZPL, // VUZP, left result
2601 OP_VUZPR, // VUZP, right result
2602 OP_VZIPL, // VZIP, left result
2603 OP_VZIPR, // VZIP, right result
2604 OP_VTRNL, // VTRN, left result
2605 OP_VTRNR // VTRN, right result
2608 if (OpNum == OP_COPY) {
2609 if (LHSID == (1*9+2)*9+3) return LHS;
2610 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2614 SDValue OpLHS, OpRHS;
2615 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2616 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2617 EVT VT = OpLHS.getValueType();
2620 default: llvm_unreachable("Unknown shuffle opcode!");
2622 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2627 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2628 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2632 return DAG.getNode(ARMISD::VEXT, dl, VT,
2634 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2637 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2638 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2641 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2642 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2645 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2646 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2650 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2651 SDValue V1 = Op.getOperand(0);
2652 SDValue V2 = Op.getOperand(1);
2653 DebugLoc dl = Op.getDebugLoc();
2654 EVT VT = Op.getValueType();
2655 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2656 SmallVector<int, 8> ShuffleMask;
2658 // Convert shuffles that are directly supported on NEON to target-specific
2659 // DAG nodes, instead of keeping them as shuffles and matching them again
2660 // during code selection. This is more efficient and avoids the possibility
2661 // of inconsistencies between legalization and selection.
2662 // FIXME: floating-point vectors should be canonicalized to integer vectors
2663 // of the same time so that they get CSEd properly.
2664 SVN->getMask(ShuffleMask);
2666 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2667 int Lane = SVN->getSplatIndex();
2668 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2669 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2671 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2672 DAG.getConstant(Lane, MVT::i32));
2677 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2680 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2681 DAG.getConstant(Imm, MVT::i32));
2684 if (isVREVMask(ShuffleMask, VT, 64))
2685 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2686 if (isVREVMask(ShuffleMask, VT, 32))
2687 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2688 if (isVREVMask(ShuffleMask, VT, 16))
2689 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2691 // Check for Neon shuffles that modify both input vectors in place.
2692 // If both results are used, i.e., if there are two shuffles with the same
2693 // source operands and with masks corresponding to both results of one of
2694 // these operations, DAG memoization will ensure that a single node is
2695 // used for both shuffles.
2696 unsigned WhichResult;
2697 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2698 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2699 V1, V2).getValue(WhichResult);
2700 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2701 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2702 V1, V2).getValue(WhichResult);
2703 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2704 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2705 V1, V2).getValue(WhichResult);
2707 // If the shuffle is not directly supported and it has 4 elements, use
2708 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2709 if (VT.getVectorNumElements() == 4 &&
2710 (VT.is128BitVector() || VT.is64BitVector())) {
2711 unsigned PFIndexes[4];
2712 for (unsigned i = 0; i != 4; ++i) {
2713 if (ShuffleMask[i] < 0)
2716 PFIndexes[i] = ShuffleMask[i];
2719 // Compute the index in the perfect shuffle table.
2720 unsigned PFTableIndex =
2721 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2723 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2724 unsigned Cost = (PFEntry >> 30);
2727 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2733 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2734 EVT VT = Op.getValueType();
2735 DebugLoc dl = Op.getDebugLoc();
2736 SDValue Vec = Op.getOperand(0);
2737 SDValue Lane = Op.getOperand(1);
2738 assert(VT == MVT::i32 &&
2739 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
2740 "unexpected type for custom-lowering vector extract");
2741 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2744 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2745 // The only time a CONCAT_VECTORS operation can have legal types is when
2746 // two 64-bit vectors are concatenated to a 128-bit vector.
2747 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2748 "unexpected CONCAT_VECTORS");
2749 DebugLoc dl = Op.getDebugLoc();
2750 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2751 SDValue Op0 = Op.getOperand(0);
2752 SDValue Op1 = Op.getOperand(1);
2753 if (Op0.getOpcode() != ISD::UNDEF)
2754 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2755 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2756 DAG.getIntPtrConstant(0));
2757 if (Op1.getOpcode() != ISD::UNDEF)
2758 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2759 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2760 DAG.getIntPtrConstant(1));
2761 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2764 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2765 switch (Op.getOpcode()) {
2766 default: llvm_unreachable("Don't know how to custom lower this!");
2767 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2768 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
2769 case ISD::GlobalAddress:
2770 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2771 LowerGlobalAddressELF(Op, DAG);
2772 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2773 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2774 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2775 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2776 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2777 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2778 case ISD::SINT_TO_FP:
2779 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2780 case ISD::FP_TO_SINT:
2781 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2782 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2783 case ISD::RETURNADDR: break;
2784 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2785 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2786 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2787 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2790 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2791 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2792 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2793 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2794 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2795 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2800 /// ReplaceNodeResults - Replace the results of node with an illegal result
2801 /// type with new values built out of custom code.
2802 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2803 SmallVectorImpl<SDValue>&Results,
2804 SelectionDAG &DAG) {
2805 switch (N->getOpcode()) {
2807 llvm_unreachable("Don't know how to custom expand this!");
2809 case ISD::BIT_CONVERT:
2810 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2814 SDValue Res = LowerShift(N, DAG, Subtarget);
2816 Results.push_back(Res);
2822 //===----------------------------------------------------------------------===//
2823 // ARM Scheduler Hooks
2824 //===----------------------------------------------------------------------===//
2827 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2828 MachineBasicBlock *BB,
2829 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
2830 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2831 DebugLoc dl = MI->getDebugLoc();
2832 switch (MI->getOpcode()) {
2834 llvm_unreachable("Unexpected instr type to insert");
2835 case ARM::tMOVCCr_pseudo: {
2836 // To "insert" a SELECT_CC instruction, we actually have to insert the
2837 // diamond control-flow pattern. The incoming instruction knows the
2838 // destination vreg to set, the condition code register to branch on, the
2839 // true/false values to select between, and a branch opcode to use.
2840 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2841 MachineFunction::iterator It = BB;
2847 // cmpTY ccX, r1, r2
2849 // fallthrough --> copy0MBB
2850 MachineBasicBlock *thisMBB = BB;
2851 MachineFunction *F = BB->getParent();
2852 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2853 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2854 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2855 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2856 F->insert(It, copy0MBB);
2857 F->insert(It, sinkMBB);
2858 // Update machine-CFG edges by first adding all successors of the current
2859 // block to the new block which will contain the Phi node for the select.
2860 // Also inform sdisel of the edge changes.
2861 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
2862 E = BB->succ_end(); I != E; ++I) {
2863 EM->insert(std::make_pair(*I, sinkMBB));
2864 sinkMBB->addSuccessor(*I);
2866 // Next, remove all successors of the current block, and add the true
2867 // and fallthrough blocks as its successors.
2868 while (!BB->succ_empty())
2869 BB->removeSuccessor(BB->succ_begin());
2870 BB->addSuccessor(copy0MBB);
2871 BB->addSuccessor(sinkMBB);
2874 // %FalseValue = ...
2875 // # fallthrough to sinkMBB
2878 // Update machine-CFG edges
2879 BB->addSuccessor(sinkMBB);
2882 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2885 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2886 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2887 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2889 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2896 case ARM::t2SUBrSPi_:
2897 case ARM::t2SUBrSPi12_:
2898 case ARM::t2SUBrSPs_: {
2899 MachineFunction *MF = BB->getParent();
2900 unsigned DstReg = MI->getOperand(0).getReg();
2901 unsigned SrcReg = MI->getOperand(1).getReg();
2902 bool DstIsDead = MI->getOperand(0).isDead();
2903 bool SrcIsKill = MI->getOperand(1).isKill();
2905 if (SrcReg != ARM::SP) {
2906 // Copy the source to SP from virtual register.
2907 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2908 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2909 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2910 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2911 .addReg(SrcReg, getKillRegState(SrcIsKill));
2915 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2916 switch (MI->getOpcode()) {
2918 llvm_unreachable("Unexpected pseudo instruction!");
2924 OpOpc = ARM::tADDspr;
2927 OpOpc = ARM::tSUBspi;
2929 case ARM::t2SUBrSPi_:
2930 OpOpc = ARM::t2SUBrSPi;
2931 NeedPred = true; NeedCC = true;
2933 case ARM::t2SUBrSPi12_:
2934 OpOpc = ARM::t2SUBrSPi12;
2937 case ARM::t2SUBrSPs_:
2938 OpOpc = ARM::t2SUBrSPs;
2939 NeedPred = true; NeedCC = true; NeedOp3 = true;
2942 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2943 if (OpOpc == ARM::tAND)
2944 AddDefaultT1CC(MIB);
2945 MIB.addReg(ARM::SP);
2946 MIB.addOperand(MI->getOperand(2));
2948 MIB.addOperand(MI->getOperand(3));
2950 AddDefaultPred(MIB);
2954 // Copy the result from SP to virtual register.
2955 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2956 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2957 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2958 BuildMI(BB, dl, TII->get(CopyOpc))
2959 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
2961 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2967 //===----------------------------------------------------------------------===//
2968 // ARM Optimization Hooks
2969 //===----------------------------------------------------------------------===//
2972 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2973 TargetLowering::DAGCombinerInfo &DCI) {
2974 SelectionDAG &DAG = DCI.DAG;
2975 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2976 EVT VT = N->getValueType(0);
2977 unsigned Opc = N->getOpcode();
2978 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2979 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2980 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2981 ISD::CondCode CC = ISD::SETCC_INVALID;
2984 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2986 SDValue CCOp = Slct.getOperand(0);
2987 if (CCOp.getOpcode() == ISD::SETCC)
2988 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2991 bool DoXform = false;
2993 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2996 if (LHS.getOpcode() == ISD::Constant &&
2997 cast<ConstantSDNode>(LHS)->isNullValue()) {
2999 } else if (CC != ISD::SETCC_INVALID &&
3000 RHS.getOpcode() == ISD::Constant &&
3001 cast<ConstantSDNode>(RHS)->isNullValue()) {
3002 std::swap(LHS, RHS);
3003 SDValue Op0 = Slct.getOperand(0);
3004 EVT OpVT = isSlctCC ? Op0.getValueType() :
3005 Op0.getOperand(0).getValueType();
3006 bool isInt = OpVT.isInteger();
3007 CC = ISD::getSetCCInverse(CC, isInt);
3009 if (!TLI.isCondCodeLegal(CC, OpVT))
3010 return SDValue(); // Inverse operator isn't legal.
3017 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3019 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3020 Slct.getOperand(0), Slct.getOperand(1), CC);
3021 SDValue CCOp = Slct.getOperand(0);
3023 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3024 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3025 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3026 CCOp, OtherOp, Result);
3031 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3032 static SDValue PerformADDCombine(SDNode *N,
3033 TargetLowering::DAGCombinerInfo &DCI) {
3034 // added by evan in r37685 with no testcase.
3035 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3037 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3038 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3039 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3040 if (Result.getNode()) return Result;
3042 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3043 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3044 if (Result.getNode()) return Result;
3050 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3051 static SDValue PerformSUBCombine(SDNode *N,
3052 TargetLowering::DAGCombinerInfo &DCI) {
3053 // added by evan in r37685 with no testcase.
3054 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3056 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3057 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3058 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3059 if (Result.getNode()) return Result;
3065 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3066 static SDValue PerformFMRRDCombine(SDNode *N,
3067 TargetLowering::DAGCombinerInfo &DCI) {
3068 // fmrrd(fmdrr x, y) -> x,y
3069 SDValue InDouble = N->getOperand(0);
3070 if (InDouble.getOpcode() == ARMISD::FMDRR)
3071 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3075 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3076 /// operand of a vector shift operation, where all the elements of the
3077 /// build_vector must have the same constant integer value.
3078 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3079 // Ignore bit_converts.
3080 while (Op.getOpcode() == ISD::BIT_CONVERT)
3081 Op = Op.getOperand(0);
3082 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3083 APInt SplatBits, SplatUndef;
3084 unsigned SplatBitSize;
3086 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3087 HasAnyUndefs, ElementBits) ||
3088 SplatBitSize > ElementBits)
3090 Cnt = SplatBits.getSExtValue();
3094 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3095 /// operand of a vector shift left operation. That value must be in the range:
3096 /// 0 <= Value < ElementBits for a left shift; or
3097 /// 0 <= Value <= ElementBits for a long left shift.
3098 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3099 assert(VT.isVector() && "vector shift count is not a vector type");
3100 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3101 if (! getVShiftImm(Op, ElementBits, Cnt))
3103 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3106 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3107 /// operand of a vector shift right operation. For a shift opcode, the value
3108 /// is positive, but for an intrinsic the value count must be negative. The
3109 /// absolute value must be in the range:
3110 /// 1 <= |Value| <= ElementBits for a right shift; or
3111 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3112 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3114 assert(VT.isVector() && "vector shift count is not a vector type");
3115 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3116 if (! getVShiftImm(Op, ElementBits, Cnt))
3120 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3123 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3124 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3125 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3128 // Don't do anything for most intrinsics.
3131 // Vector shifts: check for immediate versions and lower them.
3132 // Note: This is done during DAG combining instead of DAG legalizing because
3133 // the build_vectors for 64-bit vector element shift counts are generally
3134 // not legal, and it is hard to see their values after they get legalized to
3135 // loads from a constant pool.
3136 case Intrinsic::arm_neon_vshifts:
3137 case Intrinsic::arm_neon_vshiftu:
3138 case Intrinsic::arm_neon_vshiftls:
3139 case Intrinsic::arm_neon_vshiftlu:
3140 case Intrinsic::arm_neon_vshiftn:
3141 case Intrinsic::arm_neon_vrshifts:
3142 case Intrinsic::arm_neon_vrshiftu:
3143 case Intrinsic::arm_neon_vrshiftn:
3144 case Intrinsic::arm_neon_vqshifts:
3145 case Intrinsic::arm_neon_vqshiftu:
3146 case Intrinsic::arm_neon_vqshiftsu:
3147 case Intrinsic::arm_neon_vqshiftns:
3148 case Intrinsic::arm_neon_vqshiftnu:
3149 case Intrinsic::arm_neon_vqshiftnsu:
3150 case Intrinsic::arm_neon_vqrshiftns:
3151 case Intrinsic::arm_neon_vqrshiftnu:
3152 case Intrinsic::arm_neon_vqrshiftnsu: {
3153 EVT VT = N->getOperand(1).getValueType();
3155 unsigned VShiftOpc = 0;
3158 case Intrinsic::arm_neon_vshifts:
3159 case Intrinsic::arm_neon_vshiftu:
3160 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3161 VShiftOpc = ARMISD::VSHL;
3164 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3165 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3166 ARMISD::VSHRs : ARMISD::VSHRu);
3171 case Intrinsic::arm_neon_vshiftls:
3172 case Intrinsic::arm_neon_vshiftlu:
3173 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3175 llvm_unreachable("invalid shift count for vshll intrinsic");
3177 case Intrinsic::arm_neon_vrshifts:
3178 case Intrinsic::arm_neon_vrshiftu:
3179 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3183 case Intrinsic::arm_neon_vqshifts:
3184 case Intrinsic::arm_neon_vqshiftu:
3185 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3189 case Intrinsic::arm_neon_vqshiftsu:
3190 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3192 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3194 case Intrinsic::arm_neon_vshiftn:
3195 case Intrinsic::arm_neon_vrshiftn:
3196 case Intrinsic::arm_neon_vqshiftns:
3197 case Intrinsic::arm_neon_vqshiftnu:
3198 case Intrinsic::arm_neon_vqshiftnsu:
3199 case Intrinsic::arm_neon_vqrshiftns:
3200 case Intrinsic::arm_neon_vqrshiftnu:
3201 case Intrinsic::arm_neon_vqrshiftnsu:
3202 // Narrowing shifts require an immediate right shift.
3203 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3205 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3208 llvm_unreachable("unhandled vector shift");
3212 case Intrinsic::arm_neon_vshifts:
3213 case Intrinsic::arm_neon_vshiftu:
3214 // Opcode already set above.
3216 case Intrinsic::arm_neon_vshiftls:
3217 case Intrinsic::arm_neon_vshiftlu:
3218 if (Cnt == VT.getVectorElementType().getSizeInBits())
3219 VShiftOpc = ARMISD::VSHLLi;
3221 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3222 ARMISD::VSHLLs : ARMISD::VSHLLu);
3224 case Intrinsic::arm_neon_vshiftn:
3225 VShiftOpc = ARMISD::VSHRN; break;
3226 case Intrinsic::arm_neon_vrshifts:
3227 VShiftOpc = ARMISD::VRSHRs; break;
3228 case Intrinsic::arm_neon_vrshiftu:
3229 VShiftOpc = ARMISD::VRSHRu; break;
3230 case Intrinsic::arm_neon_vrshiftn:
3231 VShiftOpc = ARMISD::VRSHRN; break;
3232 case Intrinsic::arm_neon_vqshifts:
3233 VShiftOpc = ARMISD::VQSHLs; break;
3234 case Intrinsic::arm_neon_vqshiftu:
3235 VShiftOpc = ARMISD::VQSHLu; break;
3236 case Intrinsic::arm_neon_vqshiftsu:
3237 VShiftOpc = ARMISD::VQSHLsu; break;
3238 case Intrinsic::arm_neon_vqshiftns:
3239 VShiftOpc = ARMISD::VQSHRNs; break;
3240 case Intrinsic::arm_neon_vqshiftnu:
3241 VShiftOpc = ARMISD::VQSHRNu; break;
3242 case Intrinsic::arm_neon_vqshiftnsu:
3243 VShiftOpc = ARMISD::VQSHRNsu; break;
3244 case Intrinsic::arm_neon_vqrshiftns:
3245 VShiftOpc = ARMISD::VQRSHRNs; break;
3246 case Intrinsic::arm_neon_vqrshiftnu:
3247 VShiftOpc = ARMISD::VQRSHRNu; break;
3248 case Intrinsic::arm_neon_vqrshiftnsu:
3249 VShiftOpc = ARMISD::VQRSHRNsu; break;
3252 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3253 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3256 case Intrinsic::arm_neon_vshiftins: {
3257 EVT VT = N->getOperand(1).getValueType();
3259 unsigned VShiftOpc = 0;
3261 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3262 VShiftOpc = ARMISD::VSLI;
3263 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3264 VShiftOpc = ARMISD::VSRI;
3266 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3269 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3270 N->getOperand(1), N->getOperand(2),
3271 DAG.getConstant(Cnt, MVT::i32));
3274 case Intrinsic::arm_neon_vqrshifts:
3275 case Intrinsic::arm_neon_vqrshiftu:
3276 // No immediate versions of these to check for.
3283 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3284 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3285 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3286 /// vector element shift counts are generally not legal, and it is hard to see
3287 /// their values after they get legalized to loads from a constant pool.
3288 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3289 const ARMSubtarget *ST) {
3290 EVT VT = N->getValueType(0);
3292 // Nothing to be done for scalar shifts.
3293 if (! VT.isVector())
3296 assert(ST->hasNEON() && "unexpected vector shift");
3299 switch (N->getOpcode()) {
3300 default: llvm_unreachable("unexpected shift opcode");
3303 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3304 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3305 DAG.getConstant(Cnt, MVT::i32));
3310 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3311 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3312 ARMISD::VSHRs : ARMISD::VSHRu);
3313 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3314 DAG.getConstant(Cnt, MVT::i32));
3320 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3321 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3322 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3323 const ARMSubtarget *ST) {
3324 SDValue N0 = N->getOperand(0);
3326 // Check for sign- and zero-extensions of vector extract operations of 8-
3327 // and 16-bit vector elements. NEON supports these directly. They are
3328 // handled during DAG combining because type legalization will promote them
3329 // to 32-bit types and it is messy to recognize the operations after that.
3330 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3331 SDValue Vec = N0.getOperand(0);
3332 SDValue Lane = N0.getOperand(1);
3333 EVT VT = N->getValueType(0);
3334 EVT EltVT = N0.getValueType();
3335 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3337 if (VT == MVT::i32 &&
3338 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3339 TLI.isTypeLegal(Vec.getValueType())) {
3342 switch (N->getOpcode()) {
3343 default: llvm_unreachable("unexpected opcode");
3344 case ISD::SIGN_EXTEND:
3345 Opc = ARMISD::VGETLANEs;
3347 case ISD::ZERO_EXTEND:
3348 case ISD::ANY_EXTEND:
3349 Opc = ARMISD::VGETLANEu;
3352 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3359 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3360 DAGCombinerInfo &DCI) const {
3361 switch (N->getOpcode()) {
3363 case ISD::ADD: return PerformADDCombine(N, DCI);
3364 case ISD::SUB: return PerformSUBCombine(N, DCI);
3365 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3366 case ISD::INTRINSIC_WO_CHAIN:
3367 return PerformIntrinsicCombine(N, DCI.DAG);
3371 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3372 case ISD::SIGN_EXTEND:
3373 case ISD::ZERO_EXTEND:
3374 case ISD::ANY_EXTEND:
3375 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3380 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3381 if (!Subtarget->hasV6Ops())
3382 // Pre-v6 does not support unaligned mem access.
3384 else if (!Subtarget->hasV6Ops()) {
3385 // v6 may or may not support unaligned mem access.
3386 if (!Subtarget->isTargetDarwin())
3390 switch (VT.getSimpleVT().SimpleTy) {
3397 // FIXME: VLD1 etc with standard alignment is legal.
3401 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3406 switch (VT.getSimpleVT().SimpleTy) {
3407 default: return false;
3422 if ((V & (Scale - 1)) != 0)
3425 return V == (V & ((1LL << 5) - 1));
3428 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3429 const ARMSubtarget *Subtarget) {
3436 switch (VT.getSimpleVT().SimpleTy) {
3437 default: return false;
3442 // + imm12 or - imm8
3444 return V == (V & ((1LL << 8) - 1));
3445 return V == (V & ((1LL << 12) - 1));
3448 // Same as ARM mode. FIXME: NEON?
3449 if (!Subtarget->hasVFP2())
3454 return V == (V & ((1LL << 8) - 1));
3458 /// isLegalAddressImmediate - Return true if the integer value can be used
3459 /// as the offset of the target addressing mode for load / store of the
3461 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3462 const ARMSubtarget *Subtarget) {
3469 if (Subtarget->isThumb1Only())
3470 return isLegalT1AddressImmediate(V, VT);
3471 else if (Subtarget->isThumb2())
3472 return isLegalT2AddressImmediate(V, VT, Subtarget);
3477 switch (VT.getSimpleVT().SimpleTy) {
3478 default: return false;
3483 return V == (V & ((1LL << 12) - 1));
3486 return V == (V & ((1LL << 8) - 1));
3489 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3494 return V == (V & ((1LL << 8) - 1));
3498 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3500 int Scale = AM.Scale;
3504 switch (VT.getSimpleVT().SimpleTy) {
3505 default: return false;
3514 return Scale == 2 || Scale == 4 || Scale == 8;
3517 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3521 // Note, we allow "void" uses (basically, uses that aren't loads or
3522 // stores), because arm allows folding a scale into many arithmetic
3523 // operations. This should be made more precise and revisited later.
3525 // Allow r << imm, but the imm has to be a multiple of two.
3526 if (Scale & 1) return false;
3527 return isPowerOf2_32(Scale);
3531 /// isLegalAddressingMode - Return true if the addressing mode represented
3532 /// by AM is legal for this target, for a load/store of the specified type.
3533 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3534 const Type *Ty) const {
3535 EVT VT = getValueType(Ty, true);
3536 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3539 // Can never fold addr of global into load/store.
3544 case 0: // no scale reg, must be "r+i" or "r", or "i".
3547 if (Subtarget->isThumb1Only())
3551 // ARM doesn't support any R+R*scale+imm addr modes.
3558 if (Subtarget->isThumb2())
3559 return isLegalT2ScaledAddressingMode(AM, VT);
3561 int Scale = AM.Scale;
3562 switch (VT.getSimpleVT().SimpleTy) {
3563 default: return false;
3567 if (Scale < 0) Scale = -Scale;
3571 return isPowerOf2_32(Scale & ~1);
3575 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3580 // Note, we allow "void" uses (basically, uses that aren't loads or
3581 // stores), because arm allows folding a scale into many arithmetic
3582 // operations. This should be made more precise and revisited later.
3584 // Allow r << imm, but the imm has to be a multiple of two.
3585 if (Scale & 1) return false;
3586 return isPowerOf2_32(Scale);
3593 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3594 bool isSEXTLoad, SDValue &Base,
3595 SDValue &Offset, bool &isInc,
3596 SelectionDAG &DAG) {
3597 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3600 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3602 Base = Ptr->getOperand(0);
3603 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3604 int RHSC = (int)RHS->getZExtValue();
3605 if (RHSC < 0 && RHSC > -256) {
3606 assert(Ptr->getOpcode() == ISD::ADD);
3608 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3612 isInc = (Ptr->getOpcode() == ISD::ADD);
3613 Offset = Ptr->getOperand(1);
3615 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3617 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3618 int RHSC = (int)RHS->getZExtValue();
3619 if (RHSC < 0 && RHSC > -0x1000) {
3620 assert(Ptr->getOpcode() == ISD::ADD);
3622 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3623 Base = Ptr->getOperand(0);
3628 if (Ptr->getOpcode() == ISD::ADD) {
3630 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3631 if (ShOpcVal != ARM_AM::no_shift) {
3632 Base = Ptr->getOperand(1);
3633 Offset = Ptr->getOperand(0);
3635 Base = Ptr->getOperand(0);
3636 Offset = Ptr->getOperand(1);
3641 isInc = (Ptr->getOpcode() == ISD::ADD);
3642 Base = Ptr->getOperand(0);
3643 Offset = Ptr->getOperand(1);
3647 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3651 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3652 bool isSEXTLoad, SDValue &Base,
3653 SDValue &Offset, bool &isInc,
3654 SelectionDAG &DAG) {
3655 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3658 Base = Ptr->getOperand(0);
3659 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3660 int RHSC = (int)RHS->getZExtValue();
3661 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3662 assert(Ptr->getOpcode() == ISD::ADD);
3664 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3666 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3667 isInc = Ptr->getOpcode() == ISD::ADD;
3668 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3676 /// getPreIndexedAddressParts - returns true by value, base pointer and
3677 /// offset pointer and addressing mode by reference if the node's address
3678 /// can be legally represented as pre-indexed load / store address.
3680 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3682 ISD::MemIndexedMode &AM,
3683 SelectionDAG &DAG) const {
3684 if (Subtarget->isThumb1Only())
3689 bool isSEXTLoad = false;
3690 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3691 Ptr = LD->getBasePtr();
3692 VT = LD->getMemoryVT();
3693 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3694 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3695 Ptr = ST->getBasePtr();
3696 VT = ST->getMemoryVT();
3701 bool isLegal = false;
3702 if (Subtarget->isThumb2())
3703 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3704 Offset, isInc, DAG);
3706 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3707 Offset, isInc, DAG);
3711 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3715 /// getPostIndexedAddressParts - returns true by value, base pointer and
3716 /// offset pointer and addressing mode by reference if this node can be
3717 /// combined with a load / store to form a post-indexed load / store.
3718 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3721 ISD::MemIndexedMode &AM,
3722 SelectionDAG &DAG) const {
3723 if (Subtarget->isThumb1Only())
3728 bool isSEXTLoad = false;
3729 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3730 VT = LD->getMemoryVT();
3731 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3732 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3733 VT = ST->getMemoryVT();
3738 bool isLegal = false;
3739 if (Subtarget->isThumb2())
3740 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3743 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3748 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3752 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3756 const SelectionDAG &DAG,
3757 unsigned Depth) const {
3758 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3759 switch (Op.getOpcode()) {
3761 case ARMISD::CMOV: {
3762 // Bits are known zero/one if known on the LHS and RHS.
3763 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3764 if (KnownZero == 0 && KnownOne == 0) return;
3766 APInt KnownZeroRHS, KnownOneRHS;
3767 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3768 KnownZeroRHS, KnownOneRHS, Depth+1);
3769 KnownZero &= KnownZeroRHS;
3770 KnownOne &= KnownOneRHS;
3776 //===----------------------------------------------------------------------===//
3777 // ARM Inline Assembly Support
3778 //===----------------------------------------------------------------------===//
3780 /// getConstraintType - Given a constraint letter, return the type of
3781 /// constraint it is for this target.
3782 ARMTargetLowering::ConstraintType
3783 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3784 if (Constraint.size() == 1) {
3785 switch (Constraint[0]) {
3787 case 'l': return C_RegisterClass;
3788 case 'w': return C_RegisterClass;
3791 return TargetLowering::getConstraintType(Constraint);
3794 std::pair<unsigned, const TargetRegisterClass*>
3795 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3797 if (Constraint.size() == 1) {
3798 // GCC RS6000 Constraint Letters
3799 switch (Constraint[0]) {
3801 if (Subtarget->isThumb1Only())
3802 return std::make_pair(0U, ARM::tGPRRegisterClass);
3804 return std::make_pair(0U, ARM::GPRRegisterClass);
3806 return std::make_pair(0U, ARM::GPRRegisterClass);
3809 return std::make_pair(0U, ARM::SPRRegisterClass);
3811 return std::make_pair(0U, ARM::DPRRegisterClass);
3815 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3818 std::vector<unsigned> ARMTargetLowering::
3819 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3821 if (Constraint.size() != 1)
3822 return std::vector<unsigned>();
3824 switch (Constraint[0]) { // GCC ARM Constraint Letters
3827 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3828 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3831 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3832 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3833 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3834 ARM::R12, ARM::LR, 0);
3837 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3838 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3839 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3840 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3841 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3842 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3843 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3844 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3846 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3847 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3848 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3849 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3853 return std::vector<unsigned>();
3856 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3857 /// vector. If it is invalid, don't add anything to Ops.
3858 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3861 std::vector<SDValue>&Ops,
3862 SelectionDAG &DAG) const {
3863 SDValue Result(0, 0);
3865 switch (Constraint) {
3867 case 'I': case 'J': case 'K': case 'L':
3868 case 'M': case 'N': case 'O':
3869 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3873 int64_t CVal64 = C->getSExtValue();
3874 int CVal = (int) CVal64;
3875 // None of these constraints allow values larger than 32 bits. Check
3876 // that the value fits in an int.
3880 switch (Constraint) {
3882 if (Subtarget->isThumb1Only()) {
3883 // This must be a constant between 0 and 255, for ADD
3885 if (CVal >= 0 && CVal <= 255)
3887 } else if (Subtarget->isThumb2()) {
3888 // A constant that can be used as an immediate value in a
3889 // data-processing instruction.
3890 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3893 // A constant that can be used as an immediate value in a
3894 // data-processing instruction.
3895 if (ARM_AM::getSOImmVal(CVal) != -1)
3901 if (Subtarget->isThumb()) { // FIXME thumb2
3902 // This must be a constant between -255 and -1, for negated ADD
3903 // immediates. This can be used in GCC with an "n" modifier that
3904 // prints the negated value, for use with SUB instructions. It is
3905 // not useful otherwise but is implemented for compatibility.
3906 if (CVal >= -255 && CVal <= -1)
3909 // This must be a constant between -4095 and 4095. It is not clear
3910 // what this constraint is intended for. Implemented for
3911 // compatibility with GCC.
3912 if (CVal >= -4095 && CVal <= 4095)
3918 if (Subtarget->isThumb1Only()) {
3919 // A 32-bit value where only one byte has a nonzero value. Exclude
3920 // zero to match GCC. This constraint is used by GCC internally for
3921 // constants that can be loaded with a move/shift combination.
3922 // It is not useful otherwise but is implemented for compatibility.
3923 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3925 } else if (Subtarget->isThumb2()) {
3926 // A constant whose bitwise inverse can be used as an immediate
3927 // value in a data-processing instruction. This can be used in GCC
3928 // with a "B" modifier that prints the inverted value, for use with
3929 // BIC and MVN instructions. It is not useful otherwise but is
3930 // implemented for compatibility.
3931 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3934 // A constant whose bitwise inverse can be used as an immediate
3935 // value in a data-processing instruction. This can be used in GCC
3936 // with a "B" modifier that prints the inverted value, for use with
3937 // BIC and MVN instructions. It is not useful otherwise but is
3938 // implemented for compatibility.
3939 if (ARM_AM::getSOImmVal(~CVal) != -1)
3945 if (Subtarget->isThumb1Only()) {
3946 // This must be a constant between -7 and 7,
3947 // for 3-operand ADD/SUB immediate instructions.
3948 if (CVal >= -7 && CVal < 7)
3950 } else if (Subtarget->isThumb2()) {
3951 // A constant whose negation can be used as an immediate value in a
3952 // data-processing instruction. This can be used in GCC with an "n"
3953 // modifier that prints the negated value, for use with SUB
3954 // instructions. It is not useful otherwise but is implemented for
3956 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3959 // A constant whose negation can be used as an immediate value in a
3960 // data-processing instruction. This can be used in GCC with an "n"
3961 // modifier that prints the negated value, for use with SUB
3962 // instructions. It is not useful otherwise but is implemented for
3964 if (ARM_AM::getSOImmVal(-CVal) != -1)
3970 if (Subtarget->isThumb()) { // FIXME thumb2
3971 // This must be a multiple of 4 between 0 and 1020, for
3972 // ADD sp + immediate.
3973 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3976 // A power of two or a constant between 0 and 32. This is used in
3977 // GCC for the shift amount on shifted register operands, but it is
3978 // useful in general for any shift amounts.
3979 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3985 if (Subtarget->isThumb()) { // FIXME thumb2
3986 // This must be a constant between 0 and 31, for shift amounts.
3987 if (CVal >= 0 && CVal <= 31)
3993 if (Subtarget->isThumb()) { // FIXME thumb2
3994 // This must be a multiple of 4 between -508 and 508, for
3995 // ADD/SUB sp = sp + immediate.
3996 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4001 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4005 if (Result.getNode()) {
4006 Ops.push_back(Result);
4009 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4014 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4015 // The ARM target isn't yet aware of offsets.
4019 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4020 APInt Imm = FPImm.bitcastToAPInt();
4021 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4022 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4023 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4025 // We can handle 4 bits of mantissa.
4026 // mantissa = (16+UInt(e:f:g:h))/16.
4027 if (Mantissa & 0x7ffff)
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 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4042 APInt Imm = FPImm.bitcastToAPInt();
4043 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4044 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4045 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4047 // We can handle 4 bits of mantissa.
4048 // mantissa = (16+UInt(e:f:g:h))/16.
4049 if (Mantissa & 0xffffffffffffLL)
4052 if ((Mantissa & 0xf) != Mantissa)
4055 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4056 if (Exp < -3 || Exp > 4)
4058 Exp = ((Exp+3) & 0x7) ^ 4;
4060 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4063 /// isFPImmLegal - Returns true if the target can instruction select the
4064 /// specified FP immediate natively. If false, the legalizer will
4065 /// materialize the FP immediate as a load from a constant pool.
4066 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4067 if (!Subtarget->hasVFP3())
4070 return ARM::getVFPf32Imm(Imm) != -1;
4072 return ARM::getVFPf64Imm(Imm) != -1;