1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/Instruction.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Type.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/MC/MCSectionMachO.h"
41 #include "llvm/Target/TargetOptions.h"
42 #include "llvm/ADT/VectorExtras.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Dwarf.h"
45 #include "llvm/Support/ErrorHandling.h"
46 #include "llvm/Support/MathExtras.h"
47 #include "llvm/Support/raw_ostream.h"
50 using namespace dwarf;
52 static cl::opt<bool> EnableARMEHTest("enable-arm-eh-test", cl::Hidden,
53 cl::desc("Enable ARM EH Test"));
55 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
56 CCValAssign::LocInfo &LocInfo,
57 ISD::ArgFlagsTy &ArgFlags,
59 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
60 CCValAssign::LocInfo &LocInfo,
61 ISD::ArgFlagsTy &ArgFlags,
63 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
64 CCValAssign::LocInfo &LocInfo,
65 ISD::ArgFlagsTy &ArgFlags,
67 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
68 CCValAssign::LocInfo &LocInfo,
69 ISD::ArgFlagsTy &ArgFlags,
72 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
73 EVT PromotedBitwiseVT) {
74 if (VT != PromotedLdStVT) {
75 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
76 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
77 PromotedLdStVT.getSimpleVT());
79 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
80 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
81 PromotedLdStVT.getSimpleVT());
84 EVT ElemTy = VT.getVectorElementType();
85 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
86 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
87 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
88 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
89 if (ElemTy != MVT::i32) {
90 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
91 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
92 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
95 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
96 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
97 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
98 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
100 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
101 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
102 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
105 // Promote all bit-wise operations.
106 if (VT.isInteger() && VT != PromotedBitwiseVT) {
107 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
108 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
109 PromotedBitwiseVT.getSimpleVT());
110 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
111 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
112 PromotedBitwiseVT.getSimpleVT());
113 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
114 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
115 PromotedBitwiseVT.getSimpleVT());
118 // Neon does not support vector divide/remainder operations.
119 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
121 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
122 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
123 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
124 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
127 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
128 addRegisterClass(VT, ARM::DPRRegisterClass);
129 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
132 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
133 addRegisterClass(VT, ARM::QPRRegisterClass);
134 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
139 // FIXME: This is a test of <rdar://problem/6804645>.
140 class ARMMachOTargetObjectFile : public TargetLoweringObjectFileMachO {
142 virtual void Initialize(MCContext &Ctx, const TargetMachine &TM) {
143 TargetLoweringObjectFileMachO::Initialize(Ctx, TM);
145 // Exception Handling.
146 LSDASection = getMachOSection("__TEXT", "__gcc_except_tab", 0,
147 SectionKind::getReadOnlyWithRel());
150 virtual unsigned getTTypeEncoding() const {
151 return DW_EH_PE_indirect | DW_EH_PE_pcrel | DW_EH_PE_sdata4;
157 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
158 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
160 return new ARMMachOTargetObjectFile();
162 return new TargetLoweringObjectFileMachO();
164 return new ARMElfTargetObjectFile();
167 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
168 : TargetLowering(TM, createTLOF(TM)) {
169 Subtarget = &TM.getSubtarget<ARMSubtarget>();
171 if (Subtarget->isTargetDarwin()) {
172 // Uses VFP for Thumb libfuncs if available.
173 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
174 // Single-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
176 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
177 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
178 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
180 // Double-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
182 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
183 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
184 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
186 // Single-precision comparisons.
187 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
188 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
189 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
190 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
191 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
192 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
193 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
194 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
196 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
205 // Double-precision comparisons.
206 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
207 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
208 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
209 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
210 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
211 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
212 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
213 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
215 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
224 // Floating-point to integer conversions.
225 // i64 conversions are done via library routines even when generating VFP
226 // instructions, so use the same ones.
227 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
228 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
229 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
230 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
232 // Conversions between floating types.
233 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
234 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
236 // Integer to floating-point conversions.
237 // i64 conversions are done via library routines even when generating VFP
238 // instructions, so use the same ones.
239 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
240 // e.g., __floatunsidf vs. __floatunssidfvfp.
241 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
242 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
243 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
244 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
248 // These libcalls are not available in 32-bit.
249 setLibcallName(RTLIB::SHL_I128, 0);
250 setLibcallName(RTLIB::SRL_I128, 0);
251 setLibcallName(RTLIB::SRA_I128, 0);
253 // Libcalls should use the AAPCS base standard ABI, even if hard float
254 // is in effect, as per the ARM RTABI specification, section 4.1.2.
255 if (Subtarget->isAAPCS_ABI()) {
256 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
257 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
258 CallingConv::ARM_AAPCS);
262 if (Subtarget->isThumb1Only())
263 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
265 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
266 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
267 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
268 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
270 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
273 if (Subtarget->hasNEON()) {
274 addDRTypeForNEON(MVT::v2f32);
275 addDRTypeForNEON(MVT::v8i8);
276 addDRTypeForNEON(MVT::v4i16);
277 addDRTypeForNEON(MVT::v2i32);
278 addDRTypeForNEON(MVT::v1i64);
280 addQRTypeForNEON(MVT::v4f32);
281 addQRTypeForNEON(MVT::v2f64);
282 addQRTypeForNEON(MVT::v16i8);
283 addQRTypeForNEON(MVT::v8i16);
284 addQRTypeForNEON(MVT::v4i32);
285 addQRTypeForNEON(MVT::v2i64);
287 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
288 // neither Neon nor VFP support any arithmetic operations on it.
289 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
290 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
291 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
292 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
293 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
294 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
295 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
296 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
297 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
298 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
299 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
300 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
301 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
302 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
303 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
304 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
305 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
306 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
307 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
308 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
309 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
310 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
311 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
312 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
314 // Neon does not support some operations on v1i64 and v2i64 types.
315 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
316 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
317 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
318 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
320 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
321 setTargetDAGCombine(ISD::SHL);
322 setTargetDAGCombine(ISD::SRL);
323 setTargetDAGCombine(ISD::SRA);
324 setTargetDAGCombine(ISD::SIGN_EXTEND);
325 setTargetDAGCombine(ISD::ZERO_EXTEND);
326 setTargetDAGCombine(ISD::ANY_EXTEND);
327 setTargetDAGCombine(ISD::SELECT_CC);
330 computeRegisterProperties();
332 // ARM does not have f32 extending load.
333 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
335 // ARM does not have i1 sign extending load.
336 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
338 // ARM supports all 4 flavors of integer indexed load / store.
339 if (!Subtarget->isThumb1Only()) {
340 for (unsigned im = (unsigned)ISD::PRE_INC;
341 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
342 setIndexedLoadAction(im, MVT::i1, Legal);
343 setIndexedLoadAction(im, MVT::i8, Legal);
344 setIndexedLoadAction(im, MVT::i16, Legal);
345 setIndexedLoadAction(im, MVT::i32, Legal);
346 setIndexedStoreAction(im, MVT::i1, Legal);
347 setIndexedStoreAction(im, MVT::i8, Legal);
348 setIndexedStoreAction(im, MVT::i16, Legal);
349 setIndexedStoreAction(im, MVT::i32, Legal);
353 // i64 operation support.
354 if (Subtarget->isThumb1Only()) {
355 setOperationAction(ISD::MUL, MVT::i64, Expand);
356 setOperationAction(ISD::MULHU, MVT::i32, Expand);
357 setOperationAction(ISD::MULHS, MVT::i32, Expand);
358 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
359 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
361 setOperationAction(ISD::MUL, MVT::i64, Expand);
362 setOperationAction(ISD::MULHU, MVT::i32, Expand);
363 if (!Subtarget->hasV6Ops())
364 setOperationAction(ISD::MULHS, MVT::i32, Expand);
366 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
367 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
368 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
369 setOperationAction(ISD::SRL, MVT::i64, Custom);
370 setOperationAction(ISD::SRA, MVT::i64, Custom);
372 // ARM does not have ROTL.
373 setOperationAction(ISD::ROTL, MVT::i32, Expand);
374 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
375 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
376 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
377 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
379 // Only ARMv6 has BSWAP.
380 if (!Subtarget->hasV6Ops())
381 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
383 // These are expanded into libcalls.
384 setOperationAction(ISD::SDIV, MVT::i32, Expand);
385 setOperationAction(ISD::UDIV, MVT::i32, Expand);
386 setOperationAction(ISD::SREM, MVT::i32, Expand);
387 setOperationAction(ISD::UREM, MVT::i32, Expand);
388 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
389 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
391 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
392 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
393 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
394 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
395 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
397 // Use the default implementation.
398 setOperationAction(ISD::VASTART, MVT::Other, Custom);
399 setOperationAction(ISD::VAARG, MVT::Other, Expand);
400 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
401 setOperationAction(ISD::VAEND, MVT::Other, Expand);
402 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
403 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
404 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
405 // FIXME: Shouldn't need this, since no register is used, but the legalizer
406 // doesn't yet know how to not do that for SjLj.
407 setExceptionSelectorRegister(ARM::R0);
408 if (Subtarget->isThumb())
409 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
411 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
412 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
414 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
415 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
416 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
418 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
420 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
421 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
422 // iff target supports vfp2.
423 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
425 // We want to custom lower some of our intrinsics.
426 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
428 setOperationAction(ISD::SETCC, MVT::i32, Expand);
429 setOperationAction(ISD::SETCC, MVT::f32, Expand);
430 setOperationAction(ISD::SETCC, MVT::f64, Expand);
431 setOperationAction(ISD::SELECT, MVT::i32, Expand);
432 setOperationAction(ISD::SELECT, MVT::f32, Expand);
433 setOperationAction(ISD::SELECT, MVT::f64, Expand);
434 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
435 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
436 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
438 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
439 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
440 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
441 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
442 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
444 // We don't support sin/cos/fmod/copysign/pow
445 setOperationAction(ISD::FSIN, MVT::f64, Expand);
446 setOperationAction(ISD::FSIN, MVT::f32, Expand);
447 setOperationAction(ISD::FCOS, MVT::f32, Expand);
448 setOperationAction(ISD::FCOS, MVT::f64, Expand);
449 setOperationAction(ISD::FREM, MVT::f64, Expand);
450 setOperationAction(ISD::FREM, MVT::f32, Expand);
451 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
452 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
453 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
455 setOperationAction(ISD::FPOW, MVT::f64, Expand);
456 setOperationAction(ISD::FPOW, MVT::f32, Expand);
458 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
459 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
460 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
461 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
462 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
463 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
466 // We have target-specific dag combine patterns for the following nodes:
467 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
468 setTargetDAGCombine(ISD::ADD);
469 setTargetDAGCombine(ISD::SUB);
471 setStackPointerRegisterToSaveRestore(ARM::SP);
472 setSchedulingPreference(SchedulingForRegPressure);
474 // FIXME: If-converter should use instruction latency to determine
475 // profitability rather than relying on fixed limits.
476 if (Subtarget->getCPUString() == "generic") {
477 // Generic (and overly aggressive) if-conversion limits.
478 setIfCvtBlockSizeLimit(10);
479 setIfCvtDupBlockSizeLimit(2);
480 } else if (Subtarget->hasV6Ops()) {
481 setIfCvtBlockSizeLimit(2);
482 setIfCvtDupBlockSizeLimit(1);
484 setIfCvtBlockSizeLimit(3);
485 setIfCvtDupBlockSizeLimit(2);
488 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
489 // Do not enable CodePlacementOpt for now: it currently runs after the
490 // ARMConstantIslandPass and messes up branch relaxation and placement
491 // of constant islands.
492 // benefitFromCodePlacementOpt = true;
495 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
498 case ARMISD::Wrapper: return "ARMISD::Wrapper";
499 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
500 case ARMISD::CALL: return "ARMISD::CALL";
501 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
502 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
503 case ARMISD::tCALL: return "ARMISD::tCALL";
504 case ARMISD::BRCOND: return "ARMISD::BRCOND";
505 case ARMISD::BR_JT: return "ARMISD::BR_JT";
506 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
507 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
508 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
509 case ARMISD::CMP: return "ARMISD::CMP";
510 case ARMISD::CMPZ: return "ARMISD::CMPZ";
511 case ARMISD::CMPFP: return "ARMISD::CMPFP";
512 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
513 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
514 case ARMISD::CMOV: return "ARMISD::CMOV";
515 case ARMISD::CNEG: return "ARMISD::CNEG";
517 case ARMISD::RBIT: return "ARMISD::RBIT";
519 case ARMISD::FTOSI: return "ARMISD::FTOSI";
520 case ARMISD::FTOUI: return "ARMISD::FTOUI";
521 case ARMISD::SITOF: return "ARMISD::SITOF";
522 case ARMISD::UITOF: return "ARMISD::UITOF";
524 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
525 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
526 case ARMISD::RRX: return "ARMISD::RRX";
528 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
529 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
531 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
532 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
534 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
536 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
538 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
539 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
541 case ARMISD::VCEQ: return "ARMISD::VCEQ";
542 case ARMISD::VCGE: return "ARMISD::VCGE";
543 case ARMISD::VCGEU: return "ARMISD::VCGEU";
544 case ARMISD::VCGT: return "ARMISD::VCGT";
545 case ARMISD::VCGTU: return "ARMISD::VCGTU";
546 case ARMISD::VTST: return "ARMISD::VTST";
548 case ARMISD::VSHL: return "ARMISD::VSHL";
549 case ARMISD::VSHRs: return "ARMISD::VSHRs";
550 case ARMISD::VSHRu: return "ARMISD::VSHRu";
551 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
552 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
553 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
554 case ARMISD::VSHRN: return "ARMISD::VSHRN";
555 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
556 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
557 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
558 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
559 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
560 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
561 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
562 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
563 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
564 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
565 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
566 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
567 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
568 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
569 case ARMISD::VDUP: return "ARMISD::VDUP";
570 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
571 case ARMISD::VEXT: return "ARMISD::VEXT";
572 case ARMISD::VREV64: return "ARMISD::VREV64";
573 case ARMISD::VREV32: return "ARMISD::VREV32";
574 case ARMISD::VREV16: return "ARMISD::VREV16";
575 case ARMISD::VZIP: return "ARMISD::VZIP";
576 case ARMISD::VUZP: return "ARMISD::VUZP";
577 case ARMISD::VTRN: return "ARMISD::VTRN";
578 case ARMISD::FMAX: return "ARMISD::FMAX";
579 case ARMISD::FMIN: return "ARMISD::FMIN";
583 /// getFunctionAlignment - Return the Log2 alignment of this function.
584 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
585 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
588 //===----------------------------------------------------------------------===//
590 //===----------------------------------------------------------------------===//
592 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
593 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
595 default: llvm_unreachable("Unknown condition code!");
596 case ISD::SETNE: return ARMCC::NE;
597 case ISD::SETEQ: return ARMCC::EQ;
598 case ISD::SETGT: return ARMCC::GT;
599 case ISD::SETGE: return ARMCC::GE;
600 case ISD::SETLT: return ARMCC::LT;
601 case ISD::SETLE: return ARMCC::LE;
602 case ISD::SETUGT: return ARMCC::HI;
603 case ISD::SETUGE: return ARMCC::HS;
604 case ISD::SETULT: return ARMCC::LO;
605 case ISD::SETULE: return ARMCC::LS;
609 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
610 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
611 ARMCC::CondCodes &CondCode2) {
612 CondCode2 = ARMCC::AL;
614 default: llvm_unreachable("Unknown FP condition!");
616 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
618 case ISD::SETOGT: CondCode = ARMCC::GT; break;
620 case ISD::SETOGE: CondCode = ARMCC::GE; break;
621 case ISD::SETOLT: CondCode = ARMCC::MI; break;
622 case ISD::SETOLE: CondCode = ARMCC::LS; break;
623 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
624 case ISD::SETO: CondCode = ARMCC::VC; break;
625 case ISD::SETUO: CondCode = ARMCC::VS; break;
626 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
627 case ISD::SETUGT: CondCode = ARMCC::HI; break;
628 case ISD::SETUGE: CondCode = ARMCC::PL; break;
630 case ISD::SETULT: CondCode = ARMCC::LT; break;
632 case ISD::SETULE: CondCode = ARMCC::LE; break;
634 case ISD::SETUNE: CondCode = ARMCC::NE; break;
638 //===----------------------------------------------------------------------===//
639 // Calling Convention Implementation
640 //===----------------------------------------------------------------------===//
642 #include "ARMGenCallingConv.inc"
644 // APCS f64 is in register pairs, possibly split to stack
645 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
646 CCValAssign::LocInfo &LocInfo,
647 CCState &State, bool CanFail) {
648 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
650 // Try to get the first register.
651 if (unsigned Reg = State.AllocateReg(RegList, 4))
652 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
654 // For the 2nd half of a v2f64, do not fail.
658 // Put the whole thing on the stack.
659 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
660 State.AllocateStack(8, 4),
665 // Try to get the second register.
666 if (unsigned Reg = State.AllocateReg(RegList, 4))
667 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
669 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
670 State.AllocateStack(4, 4),
675 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
676 CCValAssign::LocInfo &LocInfo,
677 ISD::ArgFlagsTy &ArgFlags,
679 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
681 if (LocVT == MVT::v2f64 &&
682 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
684 return true; // we handled it
687 // AAPCS f64 is in aligned register pairs
688 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
689 CCValAssign::LocInfo &LocInfo,
690 CCState &State, bool CanFail) {
691 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
692 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
694 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
696 // For the 2nd half of a v2f64, do not just fail.
700 // Put the whole thing on the stack.
701 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
702 State.AllocateStack(8, 8),
708 for (i = 0; i < 2; ++i)
709 if (HiRegList[i] == Reg)
712 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
713 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
718 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
719 CCValAssign::LocInfo &LocInfo,
720 ISD::ArgFlagsTy &ArgFlags,
722 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
724 if (LocVT == MVT::v2f64 &&
725 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
727 return true; // we handled it
730 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
731 CCValAssign::LocInfo &LocInfo, CCState &State) {
732 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
733 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
735 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
737 return false; // we didn't handle it
740 for (i = 0; i < 2; ++i)
741 if (HiRegList[i] == Reg)
744 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
745 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
750 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
751 CCValAssign::LocInfo &LocInfo,
752 ISD::ArgFlagsTy &ArgFlags,
754 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
756 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
758 return true; // we handled it
761 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
762 CCValAssign::LocInfo &LocInfo,
763 ISD::ArgFlagsTy &ArgFlags,
765 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
769 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
770 /// given CallingConvention value.
771 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
773 bool isVarArg) const {
776 llvm_unreachable("Unsupported calling convention");
778 case CallingConv::Fast:
779 // Use target triple & subtarget features to do actual dispatch.
780 if (Subtarget->isAAPCS_ABI()) {
781 if (Subtarget->hasVFP2() &&
782 FloatABIType == FloatABI::Hard && !isVarArg)
783 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
785 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
787 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
788 case CallingConv::ARM_AAPCS_VFP:
789 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
790 case CallingConv::ARM_AAPCS:
791 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
792 case CallingConv::ARM_APCS:
793 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
797 /// LowerCallResult - Lower the result values of a call into the
798 /// appropriate copies out of appropriate physical registers.
800 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
801 CallingConv::ID CallConv, bool isVarArg,
802 const SmallVectorImpl<ISD::InputArg> &Ins,
803 DebugLoc dl, SelectionDAG &DAG,
804 SmallVectorImpl<SDValue> &InVals) {
806 // Assign locations to each value returned by this call.
807 SmallVector<CCValAssign, 16> RVLocs;
808 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
809 RVLocs, *DAG.getContext());
810 CCInfo.AnalyzeCallResult(Ins,
811 CCAssignFnForNode(CallConv, /* Return*/ true,
814 // Copy all of the result registers out of their specified physreg.
815 for (unsigned i = 0; i != RVLocs.size(); ++i) {
816 CCValAssign VA = RVLocs[i];
819 if (VA.needsCustom()) {
820 // Handle f64 or half of a v2f64.
821 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
823 Chain = Lo.getValue(1);
824 InFlag = Lo.getValue(2);
825 VA = RVLocs[++i]; // skip ahead to next loc
826 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
828 Chain = Hi.getValue(1);
829 InFlag = Hi.getValue(2);
830 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
832 if (VA.getLocVT() == MVT::v2f64) {
833 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
834 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
835 DAG.getConstant(0, MVT::i32));
837 VA = RVLocs[++i]; // skip ahead to next loc
838 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
839 Chain = Lo.getValue(1);
840 InFlag = Lo.getValue(2);
841 VA = RVLocs[++i]; // skip ahead to next loc
842 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
843 Chain = Hi.getValue(1);
844 InFlag = Hi.getValue(2);
845 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
846 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
847 DAG.getConstant(1, MVT::i32));
850 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
852 Chain = Val.getValue(1);
853 InFlag = Val.getValue(2);
856 switch (VA.getLocInfo()) {
857 default: llvm_unreachable("Unknown loc info!");
858 case CCValAssign::Full: break;
859 case CCValAssign::BCvt:
860 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
864 InVals.push_back(Val);
870 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
871 /// by "Src" to address "Dst" of size "Size". Alignment information is
872 /// specified by the specific parameter attribute. The copy will be passed as
873 /// a byval function parameter.
874 /// Sometimes what we are copying is the end of a larger object, the part that
875 /// does not fit in registers.
877 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
878 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
880 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
881 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
882 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
885 /// LowerMemOpCallTo - Store the argument to the stack.
887 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
888 SDValue StackPtr, SDValue Arg,
889 DebugLoc dl, SelectionDAG &DAG,
890 const CCValAssign &VA,
891 ISD::ArgFlagsTy Flags) {
892 unsigned LocMemOffset = VA.getLocMemOffset();
893 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
894 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
895 if (Flags.isByVal()) {
896 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
898 return DAG.getStore(Chain, dl, Arg, PtrOff,
899 PseudoSourceValue::getStack(), LocMemOffset,
903 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
904 SDValue Chain, SDValue &Arg,
905 RegsToPassVector &RegsToPass,
906 CCValAssign &VA, CCValAssign &NextVA,
908 SmallVector<SDValue, 8> &MemOpChains,
909 ISD::ArgFlagsTy Flags) {
911 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
912 DAG.getVTList(MVT::i32, MVT::i32), Arg);
913 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
915 if (NextVA.isRegLoc())
916 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
918 assert(NextVA.isMemLoc());
919 if (StackPtr.getNode() == 0)
920 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
922 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
928 /// LowerCall - Lowering a call into a callseq_start <-
929 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
932 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
933 CallingConv::ID CallConv, bool isVarArg,
935 const SmallVectorImpl<ISD::OutputArg> &Outs,
936 const SmallVectorImpl<ISD::InputArg> &Ins,
937 DebugLoc dl, SelectionDAG &DAG,
938 SmallVectorImpl<SDValue> &InVals) {
939 // ARM target does not yet support tail call optimization.
942 // Analyze operands of the call, assigning locations to each operand.
943 SmallVector<CCValAssign, 16> ArgLocs;
944 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
946 CCInfo.AnalyzeCallOperands(Outs,
947 CCAssignFnForNode(CallConv, /* Return*/ false,
950 // Get a count of how many bytes are to be pushed on the stack.
951 unsigned NumBytes = CCInfo.getNextStackOffset();
953 // Adjust the stack pointer for the new arguments...
954 // These operations are automatically eliminated by the prolog/epilog pass
955 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
957 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
959 RegsToPassVector RegsToPass;
960 SmallVector<SDValue, 8> MemOpChains;
962 // Walk the register/memloc assignments, inserting copies/loads. In the case
963 // of tail call optimization, arguments are handled later.
964 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
967 CCValAssign &VA = ArgLocs[i];
968 SDValue Arg = Outs[realArgIdx].Val;
969 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
971 // Promote the value if needed.
972 switch (VA.getLocInfo()) {
973 default: llvm_unreachable("Unknown loc info!");
974 case CCValAssign::Full: break;
975 case CCValAssign::SExt:
976 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
978 case CCValAssign::ZExt:
979 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
981 case CCValAssign::AExt:
982 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
984 case CCValAssign::BCvt:
985 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
989 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
990 if (VA.needsCustom()) {
991 if (VA.getLocVT() == MVT::v2f64) {
992 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
993 DAG.getConstant(0, MVT::i32));
994 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
995 DAG.getConstant(1, MVT::i32));
997 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
998 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1000 VA = ArgLocs[++i]; // skip ahead to next loc
1001 if (VA.isRegLoc()) {
1002 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1003 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1005 assert(VA.isMemLoc());
1007 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1008 dl, DAG, VA, Flags));
1011 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1012 StackPtr, MemOpChains, Flags);
1014 } else if (VA.isRegLoc()) {
1015 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1017 assert(VA.isMemLoc());
1019 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1020 dl, DAG, VA, Flags));
1024 if (!MemOpChains.empty())
1025 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1026 &MemOpChains[0], MemOpChains.size());
1028 // Build a sequence of copy-to-reg nodes chained together with token chain
1029 // and flag operands which copy the outgoing args into the appropriate regs.
1031 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1032 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1033 RegsToPass[i].second, InFlag);
1034 InFlag = Chain.getValue(1);
1037 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1038 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1039 // node so that legalize doesn't hack it.
1040 bool isDirect = false;
1041 bool isARMFunc = false;
1042 bool isLocalARMFunc = false;
1043 MachineFunction &MF = DAG.getMachineFunction();
1044 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1045 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1046 GlobalValue *GV = G->getGlobal();
1048 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1049 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1050 getTargetMachine().getRelocationModel() != Reloc::Static;
1051 isARMFunc = !Subtarget->isThumb() || isStub;
1052 // ARM call to a local ARM function is predicable.
1053 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1054 // tBX takes a register source operand.
1055 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1056 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1057 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1060 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1061 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1062 Callee = DAG.getLoad(getPointerTy(), dl,
1063 DAG.getEntryNode(), CPAddr,
1064 PseudoSourceValue::getConstantPool(), 0,
1066 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1067 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1068 getPointerTy(), Callee, PICLabel);
1070 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1071 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1073 bool isStub = Subtarget->isTargetDarwin() &&
1074 getTargetMachine().getRelocationModel() != Reloc::Static;
1075 isARMFunc = !Subtarget->isThumb() || isStub;
1076 // tBX takes a register source operand.
1077 const char *Sym = S->getSymbol();
1078 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1079 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1080 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1081 Sym, ARMPCLabelIndex, 4);
1082 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1083 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1084 Callee = DAG.getLoad(getPointerTy(), dl,
1085 DAG.getEntryNode(), CPAddr,
1086 PseudoSourceValue::getConstantPool(), 0,
1088 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1089 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1090 getPointerTy(), Callee, PICLabel);
1092 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1095 // FIXME: handle tail calls differently.
1097 if (Subtarget->isThumb()) {
1098 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1099 CallOpc = ARMISD::CALL_NOLINK;
1101 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1103 CallOpc = (isDirect || Subtarget->hasV5TOps())
1104 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1105 : ARMISD::CALL_NOLINK;
1107 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1108 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1109 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1110 InFlag = Chain.getValue(1);
1113 std::vector<SDValue> Ops;
1114 Ops.push_back(Chain);
1115 Ops.push_back(Callee);
1117 // Add argument registers to the end of the list so that they are known live
1119 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1120 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1121 RegsToPass[i].second.getValueType()));
1123 if (InFlag.getNode())
1124 Ops.push_back(InFlag);
1125 // Returns a chain and a flag for retval copy to use.
1126 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1127 &Ops[0], Ops.size());
1128 InFlag = Chain.getValue(1);
1130 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1131 DAG.getIntPtrConstant(0, true), InFlag);
1133 InFlag = Chain.getValue(1);
1135 // Handle result values, copying them out of physregs into vregs that we
1137 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1142 ARMTargetLowering::LowerReturn(SDValue Chain,
1143 CallingConv::ID CallConv, bool isVarArg,
1144 const SmallVectorImpl<ISD::OutputArg> &Outs,
1145 DebugLoc dl, SelectionDAG &DAG) {
1147 // CCValAssign - represent the assignment of the return value to a location.
1148 SmallVector<CCValAssign, 16> RVLocs;
1150 // CCState - Info about the registers and stack slots.
1151 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1154 // Analyze outgoing return values.
1155 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1158 // If this is the first return lowered for this function, add
1159 // the regs to the liveout set for the function.
1160 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1161 for (unsigned i = 0; i != RVLocs.size(); ++i)
1162 if (RVLocs[i].isRegLoc())
1163 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1168 // Copy the result values into the output registers.
1169 for (unsigned i = 0, realRVLocIdx = 0;
1171 ++i, ++realRVLocIdx) {
1172 CCValAssign &VA = RVLocs[i];
1173 assert(VA.isRegLoc() && "Can only return in registers!");
1175 SDValue Arg = Outs[realRVLocIdx].Val;
1177 switch (VA.getLocInfo()) {
1178 default: llvm_unreachable("Unknown loc info!");
1179 case CCValAssign::Full: break;
1180 case CCValAssign::BCvt:
1181 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1185 if (VA.needsCustom()) {
1186 if (VA.getLocVT() == MVT::v2f64) {
1187 // Extract the first half and return it in two registers.
1188 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1189 DAG.getConstant(0, MVT::i32));
1190 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1191 DAG.getVTList(MVT::i32, MVT::i32), Half);
1193 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1194 Flag = Chain.getValue(1);
1195 VA = RVLocs[++i]; // skip ahead to next loc
1196 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1197 HalfGPRs.getValue(1), Flag);
1198 Flag = Chain.getValue(1);
1199 VA = RVLocs[++i]; // skip ahead to next loc
1201 // Extract the 2nd half and fall through to handle it as an f64 value.
1202 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1203 DAG.getConstant(1, MVT::i32));
1205 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1207 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1208 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1209 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1210 Flag = Chain.getValue(1);
1211 VA = RVLocs[++i]; // skip ahead to next loc
1212 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1215 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1217 // Guarantee that all emitted copies are
1218 // stuck together, avoiding something bad.
1219 Flag = Chain.getValue(1);
1224 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1226 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1231 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1232 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1233 // one of the above mentioned nodes. It has to be wrapped because otherwise
1234 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1235 // be used to form addressing mode. These wrapped nodes will be selected
1237 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1238 EVT PtrVT = Op.getValueType();
1239 // FIXME there is no actual debug info here
1240 DebugLoc dl = Op.getDebugLoc();
1241 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1243 if (CP->isMachineConstantPoolEntry())
1244 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1245 CP->getAlignment());
1247 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1248 CP->getAlignment());
1249 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1252 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1253 MachineFunction &MF = DAG.getMachineFunction();
1254 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1255 unsigned ARMPCLabelIndex = 0;
1256 DebugLoc DL = Op.getDebugLoc();
1257 EVT PtrVT = getPointerTy();
1258 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1259 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1261 if (RelocM == Reloc::Static) {
1262 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1264 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1265 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1266 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1267 ARMCP::CPBlockAddress,
1269 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1271 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1272 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1273 PseudoSourceValue::getConstantPool(), 0,
1275 if (RelocM == Reloc::Static)
1277 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1278 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1281 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1283 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1284 SelectionDAG &DAG) {
1285 DebugLoc dl = GA->getDebugLoc();
1286 EVT PtrVT = getPointerTy();
1287 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1288 MachineFunction &MF = DAG.getMachineFunction();
1289 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1290 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1291 ARMConstantPoolValue *CPV =
1292 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1293 ARMCP::CPValue, PCAdj, "tlsgd", true);
1294 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1295 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1296 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1297 PseudoSourceValue::getConstantPool(), 0,
1299 SDValue Chain = Argument.getValue(1);
1301 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1302 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1304 // call __tls_get_addr.
1307 Entry.Node = Argument;
1308 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1309 Args.push_back(Entry);
1310 // FIXME: is there useful debug info available here?
1311 std::pair<SDValue, SDValue> CallResult =
1312 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1313 false, false, false, false,
1314 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1315 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1316 return CallResult.first;
1319 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1320 // "local exec" model.
1322 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1323 SelectionDAG &DAG) {
1324 GlobalValue *GV = GA->getGlobal();
1325 DebugLoc dl = GA->getDebugLoc();
1327 SDValue Chain = DAG.getEntryNode();
1328 EVT PtrVT = getPointerTy();
1329 // Get the Thread Pointer
1330 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1332 if (GV->isDeclaration()) {
1333 MachineFunction &MF = DAG.getMachineFunction();
1334 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1335 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1336 // Initial exec model.
1337 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1338 ARMConstantPoolValue *CPV =
1339 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1340 ARMCP::CPValue, PCAdj, "gottpoff", true);
1341 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1342 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1343 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1344 PseudoSourceValue::getConstantPool(), 0,
1346 Chain = Offset.getValue(1);
1348 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1349 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1351 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1352 PseudoSourceValue::getConstantPool(), 0,
1356 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1357 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1358 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1359 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1360 PseudoSourceValue::getConstantPool(), 0,
1364 // The address of the thread local variable is the add of the thread
1365 // pointer with the offset of the variable.
1366 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1370 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1371 // TODO: implement the "local dynamic" model
1372 assert(Subtarget->isTargetELF() &&
1373 "TLS not implemented for non-ELF targets");
1374 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1375 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1376 // otherwise use the "Local Exec" TLS Model
1377 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1378 return LowerToTLSGeneralDynamicModel(GA, DAG);
1380 return LowerToTLSExecModels(GA, DAG);
1383 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1384 SelectionDAG &DAG) {
1385 EVT PtrVT = getPointerTy();
1386 DebugLoc dl = Op.getDebugLoc();
1387 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1388 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1389 if (RelocM == Reloc::PIC_) {
1390 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1391 ARMConstantPoolValue *CPV =
1392 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1393 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1394 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1395 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1397 PseudoSourceValue::getConstantPool(), 0,
1399 SDValue Chain = Result.getValue(1);
1400 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1401 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1403 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1404 PseudoSourceValue::getGOT(), 0,
1408 // If we have T2 ops, we can materialize the address directly via movt/movw
1409 // pair. This is always cheaper.
1410 if (Subtarget->useMovt()) {
1411 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1412 DAG.getTargetGlobalAddress(GV, PtrVT));
1414 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1415 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1416 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1417 PseudoSourceValue::getConstantPool(), 0,
1423 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1424 SelectionDAG &DAG) {
1425 MachineFunction &MF = DAG.getMachineFunction();
1426 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1427 unsigned ARMPCLabelIndex = 0;
1428 EVT PtrVT = getPointerTy();
1429 DebugLoc dl = Op.getDebugLoc();
1430 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1431 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1433 if (RelocM == Reloc::Static)
1434 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1436 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1437 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1438 ARMConstantPoolValue *CPV =
1439 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1440 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1442 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1444 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1445 PseudoSourceValue::getConstantPool(), 0,
1447 SDValue Chain = Result.getValue(1);
1449 if (RelocM == Reloc::PIC_) {
1450 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1451 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1454 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1455 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1456 PseudoSourceValue::getGOT(), 0,
1462 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1464 assert(Subtarget->isTargetELF() &&
1465 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1466 MachineFunction &MF = DAG.getMachineFunction();
1467 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1468 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1469 EVT PtrVT = getPointerTy();
1470 DebugLoc dl = Op.getDebugLoc();
1471 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1472 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1473 "_GLOBAL_OFFSET_TABLE_",
1474 ARMPCLabelIndex, PCAdj);
1475 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1476 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1477 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1478 PseudoSourceValue::getConstantPool(), 0,
1480 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1481 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1485 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1486 const ARMSubtarget *Subtarget) {
1487 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1488 DebugLoc dl = Op.getDebugLoc();
1490 default: return SDValue(); // Don't custom lower most intrinsics.
1491 case Intrinsic::arm_thread_pointer: {
1492 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1493 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1495 case Intrinsic::eh_sjlj_lsda: {
1496 MachineFunction &MF = DAG.getMachineFunction();
1497 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1498 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1499 EVT PtrVT = getPointerTy();
1500 DebugLoc dl = Op.getDebugLoc();
1501 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1503 unsigned PCAdj = (RelocM != Reloc::PIC_)
1504 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1505 ARMConstantPoolValue *CPV =
1506 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1507 ARMCP::CPLSDA, PCAdj);
1508 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1509 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1511 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1512 PseudoSourceValue::getConstantPool(), 0,
1514 SDValue Chain = Result.getValue(1);
1516 if (RelocM == Reloc::PIC_) {
1517 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1518 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1522 case Intrinsic::eh_sjlj_setjmp:
1523 SDValue Val = Subtarget->isThumb() ?
1524 DAG.getCopyFromReg(DAG.getEntryNode(), dl, ARM::SP, MVT::i32) :
1525 DAG.getConstant(0, MVT::i32);
1526 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1),
1531 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1532 const ARMSubtarget *Subtarget) {
1533 DebugLoc dl = Op.getDebugLoc();
1534 SDValue Op5 = Op.getOperand(5);
1536 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1537 if (isDeviceBarrier) {
1538 if (Subtarget->hasV7Ops())
1539 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1541 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1542 DAG.getConstant(0, MVT::i32));
1544 if (Subtarget->hasV7Ops())
1545 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1547 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1548 DAG.getConstant(0, MVT::i32));
1553 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1554 unsigned VarArgsFrameIndex) {
1555 // vastart just stores the address of the VarArgsFrameIndex slot into the
1556 // memory location argument.
1557 DebugLoc dl = Op.getDebugLoc();
1558 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1559 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1560 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1561 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1566 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1567 SDNode *Node = Op.getNode();
1568 DebugLoc dl = Node->getDebugLoc();
1569 EVT VT = Node->getValueType(0);
1570 SDValue Chain = Op.getOperand(0);
1571 SDValue Size = Op.getOperand(1);
1572 SDValue Align = Op.getOperand(2);
1574 // Chain the dynamic stack allocation so that it doesn't modify the stack
1575 // pointer when other instructions are using the stack.
1576 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1578 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1579 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1580 if (AlignVal > StackAlign)
1581 // Do this now since selection pass cannot introduce new target
1582 // independent node.
1583 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1585 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1586 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1587 // do even more horrible hack later.
1588 MachineFunction &MF = DAG.getMachineFunction();
1589 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1590 if (AFI->isThumb1OnlyFunction()) {
1592 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1594 uint32_t Val = C->getZExtValue();
1595 if (Val <= 508 && ((Val & 3) == 0))
1599 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1602 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1603 SDValue Ops1[] = { Chain, Size, Align };
1604 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1605 Chain = Res.getValue(1);
1606 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1607 DAG.getIntPtrConstant(0, true), SDValue());
1608 SDValue Ops2[] = { Res, Chain };
1609 return DAG.getMergeValues(Ops2, 2, dl);
1613 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1614 SDValue &Root, SelectionDAG &DAG,
1616 MachineFunction &MF = DAG.getMachineFunction();
1617 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1619 TargetRegisterClass *RC;
1620 if (AFI->isThumb1OnlyFunction())
1621 RC = ARM::tGPRRegisterClass;
1623 RC = ARM::GPRRegisterClass;
1625 // Transform the arguments stored in physical registers into virtual ones.
1626 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1627 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1630 if (NextVA.isMemLoc()) {
1631 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1632 MachineFrameInfo *MFI = MF.getFrameInfo();
1633 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset(),
1636 // Create load node to retrieve arguments from the stack.
1637 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1638 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1639 PseudoSourceValue::getFixedStack(FI), 0,
1642 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1643 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1646 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1650 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1651 CallingConv::ID CallConv, bool isVarArg,
1652 const SmallVectorImpl<ISD::InputArg>
1654 DebugLoc dl, SelectionDAG &DAG,
1655 SmallVectorImpl<SDValue> &InVals) {
1657 MachineFunction &MF = DAG.getMachineFunction();
1658 MachineFrameInfo *MFI = MF.getFrameInfo();
1660 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1662 // Assign locations to all of the incoming arguments.
1663 SmallVector<CCValAssign, 16> ArgLocs;
1664 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1666 CCInfo.AnalyzeFormalArguments(Ins,
1667 CCAssignFnForNode(CallConv, /* Return*/ false,
1670 SmallVector<SDValue, 16> ArgValues;
1672 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1673 CCValAssign &VA = ArgLocs[i];
1675 // Arguments stored in registers.
1676 if (VA.isRegLoc()) {
1677 EVT RegVT = VA.getLocVT();
1680 if (VA.needsCustom()) {
1681 // f64 and vector types are split up into multiple registers or
1682 // combinations of registers and stack slots.
1685 if (VA.getLocVT() == MVT::v2f64) {
1686 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1688 VA = ArgLocs[++i]; // skip ahead to next loc
1689 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1691 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1692 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1693 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1694 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1695 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1697 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1700 TargetRegisterClass *RC;
1702 if (RegVT == MVT::f32)
1703 RC = ARM::SPRRegisterClass;
1704 else if (RegVT == MVT::f64)
1705 RC = ARM::DPRRegisterClass;
1706 else if (RegVT == MVT::v2f64)
1707 RC = ARM::QPRRegisterClass;
1708 else if (RegVT == MVT::i32)
1709 RC = (AFI->isThumb1OnlyFunction() ?
1710 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1712 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1714 // Transform the arguments in physical registers into virtual ones.
1715 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1716 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1719 // If this is an 8 or 16-bit value, it is really passed promoted
1720 // to 32 bits. Insert an assert[sz]ext to capture this, then
1721 // truncate to the right size.
1722 switch (VA.getLocInfo()) {
1723 default: llvm_unreachable("Unknown loc info!");
1724 case CCValAssign::Full: break;
1725 case CCValAssign::BCvt:
1726 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1728 case CCValAssign::SExt:
1729 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1730 DAG.getValueType(VA.getValVT()));
1731 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1733 case CCValAssign::ZExt:
1734 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1735 DAG.getValueType(VA.getValVT()));
1736 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1740 InVals.push_back(ArgValue);
1742 } else { // VA.isRegLoc()
1745 assert(VA.isMemLoc());
1746 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1748 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1749 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1752 // Create load nodes to retrieve arguments from the stack.
1753 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1754 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1755 PseudoSourceValue::getFixedStack(FI), 0,
1762 static const unsigned GPRArgRegs[] = {
1763 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1766 unsigned NumGPRs = CCInfo.getFirstUnallocated
1767 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1769 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1770 unsigned VARegSize = (4 - NumGPRs) * 4;
1771 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1772 unsigned ArgOffset = CCInfo.getNextStackOffset();
1773 if (VARegSaveSize) {
1774 // If this function is vararg, store any remaining integer argument regs
1775 // to their spots on the stack so that they may be loaded by deferencing
1776 // the result of va_next.
1777 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1778 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1779 VARegSaveSize - VARegSize,
1781 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1783 SmallVector<SDValue, 4> MemOps;
1784 for (; NumGPRs < 4; ++NumGPRs) {
1785 TargetRegisterClass *RC;
1786 if (AFI->isThumb1OnlyFunction())
1787 RC = ARM::tGPRRegisterClass;
1789 RC = ARM::GPRRegisterClass;
1791 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1792 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1793 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1794 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0,
1796 MemOps.push_back(Store);
1797 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1798 DAG.getConstant(4, getPointerTy()));
1800 if (!MemOps.empty())
1801 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1802 &MemOps[0], MemOps.size());
1804 // This will point to the next argument passed via stack.
1805 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
1811 /// isFloatingPointZero - Return true if this is +0.0.
1812 static bool isFloatingPointZero(SDValue Op) {
1813 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1814 return CFP->getValueAPF().isPosZero();
1815 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1816 // Maybe this has already been legalized into the constant pool?
1817 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1818 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1819 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1820 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1821 return CFP->getValueAPF().isPosZero();
1827 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1828 /// the given operands.
1830 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1831 SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
1832 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1833 unsigned C = RHSC->getZExtValue();
1834 if (!isLegalICmpImmediate(C)) {
1835 // Constant does not fit, try adjusting it by one?
1840 if (isLegalICmpImmediate(C-1)) {
1841 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1842 RHS = DAG.getConstant(C-1, MVT::i32);
1847 if (C > 0 && isLegalICmpImmediate(C-1)) {
1848 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1849 RHS = DAG.getConstant(C-1, MVT::i32);
1854 if (isLegalICmpImmediate(C+1)) {
1855 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1856 RHS = DAG.getConstant(C+1, MVT::i32);
1861 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1862 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1863 RHS = DAG.getConstant(C+1, MVT::i32);
1870 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1871 ARMISD::NodeType CompareType;
1874 CompareType = ARMISD::CMP;
1879 CompareType = ARMISD::CMPZ;
1882 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1883 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1886 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1887 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1890 if (!isFloatingPointZero(RHS))
1891 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1893 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1894 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1897 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1898 EVT VT = Op.getValueType();
1899 SDValue LHS = Op.getOperand(0);
1900 SDValue RHS = Op.getOperand(1);
1901 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1902 SDValue TrueVal = Op.getOperand(2);
1903 SDValue FalseVal = Op.getOperand(3);
1904 DebugLoc dl = Op.getDebugLoc();
1906 if (LHS.getValueType() == MVT::i32) {
1908 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1909 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1910 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1913 ARMCC::CondCodes CondCode, CondCode2;
1914 FPCCToARMCC(CC, CondCode, CondCode2);
1916 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1917 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1918 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1919 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1921 if (CondCode2 != ARMCC::AL) {
1922 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1923 // FIXME: Needs another CMP because flag can have but one use.
1924 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1925 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1926 Result, TrueVal, ARMCC2, CCR, Cmp2);
1931 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1932 SDValue Chain = Op.getOperand(0);
1933 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1934 SDValue LHS = Op.getOperand(2);
1935 SDValue RHS = Op.getOperand(3);
1936 SDValue Dest = Op.getOperand(4);
1937 DebugLoc dl = Op.getDebugLoc();
1939 if (LHS.getValueType() == MVT::i32) {
1941 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1942 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1943 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1944 Chain, Dest, ARMCC, CCR,Cmp);
1947 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1948 ARMCC::CondCodes CondCode, CondCode2;
1949 FPCCToARMCC(CC, CondCode, CondCode2);
1951 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1952 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1953 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1954 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1955 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1956 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1957 if (CondCode2 != ARMCC::AL) {
1958 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1959 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1960 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1965 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1966 SDValue Chain = Op.getOperand(0);
1967 SDValue Table = Op.getOperand(1);
1968 SDValue Index = Op.getOperand(2);
1969 DebugLoc dl = Op.getDebugLoc();
1971 EVT PTy = getPointerTy();
1972 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1973 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1974 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1975 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1976 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1977 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1978 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1979 if (Subtarget->isThumb2()) {
1980 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1981 // which does another jump to the destination. This also makes it easier
1982 // to translate it to TBB / TBH later.
1983 // FIXME: This might not work if the function is extremely large.
1984 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1985 Addr, Op.getOperand(2), JTI, UId);
1987 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1988 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1989 PseudoSourceValue::getJumpTable(), 0,
1991 Chain = Addr.getValue(1);
1992 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1993 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1995 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1996 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
1997 Chain = Addr.getValue(1);
1998 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2002 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2003 DebugLoc dl = Op.getDebugLoc();
2005 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
2006 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2007 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2010 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2011 EVT VT = Op.getValueType();
2012 DebugLoc dl = Op.getDebugLoc();
2014 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
2016 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2017 return DAG.getNode(Opc, dl, VT, Op);
2020 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2021 // Implement fcopysign with a fabs and a conditional fneg.
2022 SDValue Tmp0 = Op.getOperand(0);
2023 SDValue Tmp1 = Op.getOperand(1);
2024 DebugLoc dl = Op.getDebugLoc();
2025 EVT VT = Op.getValueType();
2026 EVT SrcVT = Tmp1.getValueType();
2027 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2028 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2029 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2030 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2031 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2034 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
2035 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2036 MFI->setFrameAddressIsTaken(true);
2037 EVT VT = Op.getValueType();
2038 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2039 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2040 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2041 ? ARM::R7 : ARM::R11;
2042 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2044 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2050 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
2052 SDValue Dst, SDValue Src,
2053 SDValue Size, unsigned Align,
2055 const Value *DstSV, uint64_t DstSVOff,
2056 const Value *SrcSV, uint64_t SrcSVOff){
2057 // Do repeated 4-byte loads and stores. To be improved.
2058 // This requires 4-byte alignment.
2059 if ((Align & 3) != 0)
2061 // This requires the copy size to be a constant, preferrably
2062 // within a subtarget-specific limit.
2063 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2066 uint64_t SizeVal = ConstantSize->getZExtValue();
2067 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2070 unsigned BytesLeft = SizeVal & 3;
2071 unsigned NumMemOps = SizeVal >> 2;
2072 unsigned EmittedNumMemOps = 0;
2074 unsigned VTSize = 4;
2076 const unsigned MAX_LOADS_IN_LDM = 6;
2077 SDValue TFOps[MAX_LOADS_IN_LDM];
2078 SDValue Loads[MAX_LOADS_IN_LDM];
2079 uint64_t SrcOff = 0, DstOff = 0;
2081 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2082 // same number of stores. The loads and stores will get combined into
2083 // ldm/stm later on.
2084 while (EmittedNumMemOps < NumMemOps) {
2086 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2087 Loads[i] = DAG.getLoad(VT, dl, Chain,
2088 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2089 DAG.getConstant(SrcOff, MVT::i32)),
2090 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2091 TFOps[i] = Loads[i].getValue(1);
2094 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2097 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2098 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2099 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2100 DAG.getConstant(DstOff, MVT::i32)),
2101 DstSV, DstSVOff + DstOff, false, false, 0);
2104 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2106 EmittedNumMemOps += i;
2112 // Issue loads / stores for the trailing (1 - 3) bytes.
2113 unsigned BytesLeftSave = BytesLeft;
2116 if (BytesLeft >= 2) {
2124 Loads[i] = DAG.getLoad(VT, dl, Chain,
2125 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2126 DAG.getConstant(SrcOff, MVT::i32)),
2127 SrcSV, SrcSVOff + SrcOff, false, false, 0);
2128 TFOps[i] = Loads[i].getValue(1);
2131 BytesLeft -= VTSize;
2133 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2136 BytesLeft = BytesLeftSave;
2138 if (BytesLeft >= 2) {
2146 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2147 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2148 DAG.getConstant(DstOff, MVT::i32)),
2149 DstSV, DstSVOff + DstOff, false, false, 0);
2152 BytesLeft -= VTSize;
2154 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2157 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2158 SDValue Op = N->getOperand(0);
2159 DebugLoc dl = N->getDebugLoc();
2160 if (N->getValueType(0) == MVT::f64) {
2161 // Turn i64->f64 into VMOVDRR.
2162 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2163 DAG.getConstant(0, MVT::i32));
2164 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2165 DAG.getConstant(1, MVT::i32));
2166 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2169 // Turn f64->i64 into VMOVRRD.
2170 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2171 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2173 // Merge the pieces into a single i64 value.
2174 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2177 /// getZeroVector - Returns a vector of specified type with all zero elements.
2179 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2180 assert(VT.isVector() && "Expected a vector type");
2182 // Zero vectors are used to represent vector negation and in those cases
2183 // will be implemented with the NEON VNEG instruction. However, VNEG does
2184 // not support i64 elements, so sometimes the zero vectors will need to be
2185 // explicitly constructed. For those cases, and potentially other uses in
2186 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2187 // to their dest type. This ensures they get CSE'd.
2189 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2190 SmallVector<SDValue, 8> Ops;
2193 if (VT.getSizeInBits() == 64) {
2194 Ops.assign(8, Cst); TVT = MVT::v8i8;
2196 Ops.assign(16, Cst); TVT = MVT::v16i8;
2198 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2200 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2203 /// getOnesVector - Returns a vector of specified type with all bits set.
2205 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2206 assert(VT.isVector() && "Expected a vector type");
2208 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2209 // dest type. This ensures they get CSE'd.
2211 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2212 SmallVector<SDValue, 8> Ops;
2215 if (VT.getSizeInBits() == 64) {
2216 Ops.assign(8, Cst); TVT = MVT::v8i8;
2218 Ops.assign(16, Cst); TVT = MVT::v16i8;
2220 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2222 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2225 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2226 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2227 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
2228 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2229 EVT VT = Op.getValueType();
2230 unsigned VTBits = VT.getSizeInBits();
2231 DebugLoc dl = Op.getDebugLoc();
2232 SDValue ShOpLo = Op.getOperand(0);
2233 SDValue ShOpHi = Op.getOperand(1);
2234 SDValue ShAmt = Op.getOperand(2);
2236 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2238 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2240 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2241 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2242 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2243 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2244 DAG.getConstant(VTBits, MVT::i32));
2245 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2246 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2247 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2249 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2250 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2252 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2253 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2256 SDValue Ops[2] = { Lo, Hi };
2257 return DAG.getMergeValues(Ops, 2, dl);
2260 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2261 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2262 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
2263 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2264 EVT VT = Op.getValueType();
2265 unsigned VTBits = VT.getSizeInBits();
2266 DebugLoc dl = Op.getDebugLoc();
2267 SDValue ShOpLo = Op.getOperand(0);
2268 SDValue ShOpHi = Op.getOperand(1);
2269 SDValue ShAmt = Op.getOperand(2);
2272 assert(Op.getOpcode() == ISD::SHL_PARTS);
2273 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2274 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2275 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2276 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2277 DAG.getConstant(VTBits, MVT::i32));
2278 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2279 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2281 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2282 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2283 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2285 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2286 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2289 SDValue Ops[2] = { Lo, Hi };
2290 return DAG.getMergeValues(Ops, 2, dl);
2293 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2294 const ARMSubtarget *ST) {
2295 EVT VT = N->getValueType(0);
2296 DebugLoc dl = N->getDebugLoc();
2298 if (!ST->hasV6T2Ops())
2301 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2302 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2305 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2306 const ARMSubtarget *ST) {
2307 EVT VT = N->getValueType(0);
2308 DebugLoc dl = N->getDebugLoc();
2310 // Lower vector shifts on NEON to use VSHL.
2311 if (VT.isVector()) {
2312 assert(ST->hasNEON() && "unexpected vector shift");
2314 // Left shifts translate directly to the vshiftu intrinsic.
2315 if (N->getOpcode() == ISD::SHL)
2316 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2317 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2318 N->getOperand(0), N->getOperand(1));
2320 assert((N->getOpcode() == ISD::SRA ||
2321 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2323 // NEON uses the same intrinsics for both left and right shifts. For
2324 // right shifts, the shift amounts are negative, so negate the vector of
2326 EVT ShiftVT = N->getOperand(1).getValueType();
2327 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2328 getZeroVector(ShiftVT, DAG, dl),
2330 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2331 Intrinsic::arm_neon_vshifts :
2332 Intrinsic::arm_neon_vshiftu);
2333 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2334 DAG.getConstant(vshiftInt, MVT::i32),
2335 N->getOperand(0), NegatedCount);
2338 // We can get here for a node like i32 = ISD::SHL i32, i64
2342 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2343 "Unknown shift to lower!");
2345 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2346 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2347 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2350 // If we are in thumb mode, we don't have RRX.
2351 if (ST->isThumb1Only()) return SDValue();
2353 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2354 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2355 DAG.getConstant(0, MVT::i32));
2356 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2357 DAG.getConstant(1, MVT::i32));
2359 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2360 // captures the result into a carry flag.
2361 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2362 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2364 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2365 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2367 // Merge the pieces into a single i64 value.
2368 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2371 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2372 SDValue TmpOp0, TmpOp1;
2373 bool Invert = false;
2377 SDValue Op0 = Op.getOperand(0);
2378 SDValue Op1 = Op.getOperand(1);
2379 SDValue CC = Op.getOperand(2);
2380 EVT VT = Op.getValueType();
2381 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2382 DebugLoc dl = Op.getDebugLoc();
2384 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2385 switch (SetCCOpcode) {
2386 default: llvm_unreachable("Illegal FP comparison"); break;
2388 case ISD::SETNE: Invert = true; // Fallthrough
2390 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2392 case ISD::SETLT: Swap = true; // Fallthrough
2394 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2396 case ISD::SETLE: Swap = true; // Fallthrough
2398 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2399 case ISD::SETUGE: Swap = true; // Fallthrough
2400 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2401 case ISD::SETUGT: Swap = true; // Fallthrough
2402 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2403 case ISD::SETUEQ: Invert = true; // Fallthrough
2405 // Expand this to (OLT | OGT).
2409 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2410 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2412 case ISD::SETUO: Invert = true; // Fallthrough
2414 // Expand this to (OLT | OGE).
2418 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2419 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2423 // Integer comparisons.
2424 switch (SetCCOpcode) {
2425 default: llvm_unreachable("Illegal integer comparison"); break;
2426 case ISD::SETNE: Invert = true;
2427 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2428 case ISD::SETLT: Swap = true;
2429 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2430 case ISD::SETLE: Swap = true;
2431 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2432 case ISD::SETULT: Swap = true;
2433 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2434 case ISD::SETULE: Swap = true;
2435 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2438 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2439 if (Opc == ARMISD::VCEQ) {
2442 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2444 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2447 // Ignore bitconvert.
2448 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2449 AndOp = AndOp.getOperand(0);
2451 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2453 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2454 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2461 std::swap(Op0, Op1);
2463 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2466 Result = DAG.getNOT(dl, Result, VT);
2471 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2472 /// VMOV instruction, and if so, return the constant being splatted.
2473 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2474 unsigned SplatBitSize, SelectionDAG &DAG) {
2475 switch (SplatBitSize) {
2477 // Any 1-byte value is OK.
2478 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2479 return DAG.getTargetConstant(SplatBits, MVT::i8);
2482 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2483 if ((SplatBits & ~0xff) == 0 ||
2484 (SplatBits & ~0xff00) == 0)
2485 return DAG.getTargetConstant(SplatBits, MVT::i16);
2489 // NEON's 32-bit VMOV supports splat values where:
2490 // * only one byte is nonzero, or
2491 // * the least significant byte is 0xff and the second byte is nonzero, or
2492 // * the least significant 2 bytes are 0xff and the third is nonzero.
2493 if ((SplatBits & ~0xff) == 0 ||
2494 (SplatBits & ~0xff00) == 0 ||
2495 (SplatBits & ~0xff0000) == 0 ||
2496 (SplatBits & ~0xff000000) == 0)
2497 return DAG.getTargetConstant(SplatBits, MVT::i32);
2499 if ((SplatBits & ~0xffff) == 0 &&
2500 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2501 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2503 if ((SplatBits & ~0xffffff) == 0 &&
2504 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2505 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2507 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2508 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2509 // VMOV.I32. A (very) minor optimization would be to replicate the value
2510 // and fall through here to test for a valid 64-bit splat. But, then the
2511 // caller would also need to check and handle the change in size.
2515 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2516 uint64_t BitMask = 0xff;
2518 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2519 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2521 else if ((SplatBits & BitMask) != 0)
2525 return DAG.getTargetConstant(Val, MVT::i64);
2529 llvm_unreachable("unexpected size for isVMOVSplat");
2536 /// getVMOVImm - If this is a build_vector of constants which can be
2537 /// formed by using a VMOV instruction of the specified element size,
2538 /// return the constant being splatted. The ByteSize field indicates the
2539 /// number of bytes of each element [1248].
2540 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2541 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2542 APInt SplatBits, SplatUndef;
2543 unsigned SplatBitSize;
2545 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2546 HasAnyUndefs, ByteSize * 8))
2549 if (SplatBitSize > ByteSize * 8)
2552 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2556 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2557 bool &ReverseVEXT, unsigned &Imm) {
2558 unsigned NumElts = VT.getVectorNumElements();
2559 ReverseVEXT = false;
2562 // If this is a VEXT shuffle, the immediate value is the index of the first
2563 // element. The other shuffle indices must be the successive elements after
2565 unsigned ExpectedElt = Imm;
2566 for (unsigned i = 1; i < NumElts; ++i) {
2567 // Increment the expected index. If it wraps around, it may still be
2568 // a VEXT but the source vectors must be swapped.
2570 if (ExpectedElt == NumElts * 2) {
2575 if (ExpectedElt != static_cast<unsigned>(M[i]))
2579 // Adjust the index value if the source operands will be swapped.
2586 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2587 /// instruction with the specified blocksize. (The order of the elements
2588 /// within each block of the vector is reversed.)
2589 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2590 unsigned BlockSize) {
2591 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2592 "Only possible block sizes for VREV are: 16, 32, 64");
2594 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2598 unsigned NumElts = VT.getVectorNumElements();
2599 unsigned BlockElts = M[0] + 1;
2601 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2604 for (unsigned i = 0; i < NumElts; ++i) {
2605 if ((unsigned) M[i] !=
2606 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2613 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2614 unsigned &WhichResult) {
2615 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2619 unsigned NumElts = VT.getVectorNumElements();
2620 WhichResult = (M[0] == 0 ? 0 : 1);
2621 for (unsigned i = 0; i < NumElts; i += 2) {
2622 if ((unsigned) M[i] != i + WhichResult ||
2623 (unsigned) M[i+1] != i + NumElts + WhichResult)
2629 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
2630 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2631 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
2632 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2633 unsigned &WhichResult) {
2634 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2638 unsigned NumElts = VT.getVectorNumElements();
2639 WhichResult = (M[0] == 0 ? 0 : 1);
2640 for (unsigned i = 0; i < NumElts; i += 2) {
2641 if ((unsigned) M[i] != i + WhichResult ||
2642 (unsigned) M[i+1] != i + WhichResult)
2648 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2649 unsigned &WhichResult) {
2650 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2654 unsigned NumElts = VT.getVectorNumElements();
2655 WhichResult = (M[0] == 0 ? 0 : 1);
2656 for (unsigned i = 0; i != NumElts; ++i) {
2657 if ((unsigned) M[i] != 2 * i + WhichResult)
2661 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2662 if (VT.is64BitVector() && EltSz == 32)
2668 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
2669 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2670 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
2671 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2672 unsigned &WhichResult) {
2673 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2677 unsigned Half = VT.getVectorNumElements() / 2;
2678 WhichResult = (M[0] == 0 ? 0 : 1);
2679 for (unsigned j = 0; j != 2; ++j) {
2680 unsigned Idx = WhichResult;
2681 for (unsigned i = 0; i != Half; ++i) {
2682 if ((unsigned) M[i + j * Half] != Idx)
2688 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2689 if (VT.is64BitVector() && EltSz == 32)
2695 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2696 unsigned &WhichResult) {
2697 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2701 unsigned NumElts = VT.getVectorNumElements();
2702 WhichResult = (M[0] == 0 ? 0 : 1);
2703 unsigned Idx = WhichResult * NumElts / 2;
2704 for (unsigned i = 0; i != NumElts; i += 2) {
2705 if ((unsigned) M[i] != Idx ||
2706 (unsigned) M[i+1] != Idx + NumElts)
2711 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2712 if (VT.is64BitVector() && EltSz == 32)
2718 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
2719 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
2720 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
2721 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
2722 unsigned &WhichResult) {
2723 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2727 unsigned NumElts = VT.getVectorNumElements();
2728 WhichResult = (M[0] == 0 ? 0 : 1);
2729 unsigned Idx = WhichResult * NumElts / 2;
2730 for (unsigned i = 0; i != NumElts; i += 2) {
2731 if ((unsigned) M[i] != Idx ||
2732 (unsigned) M[i+1] != Idx)
2737 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2738 if (VT.is64BitVector() && EltSz == 32)
2745 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2746 // Canonicalize all-zeros and all-ones vectors.
2747 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2748 if (ConstVal->isNullValue())
2749 return getZeroVector(VT, DAG, dl);
2750 if (ConstVal->isAllOnesValue())
2751 return getOnesVector(VT, DAG, dl);
2754 if (VT.is64BitVector()) {
2755 switch (Val.getValueType().getSizeInBits()) {
2756 case 8: CanonicalVT = MVT::v8i8; break;
2757 case 16: CanonicalVT = MVT::v4i16; break;
2758 case 32: CanonicalVT = MVT::v2i32; break;
2759 case 64: CanonicalVT = MVT::v1i64; break;
2760 default: llvm_unreachable("unexpected splat element type"); break;
2763 assert(VT.is128BitVector() && "unknown splat vector size");
2764 switch (Val.getValueType().getSizeInBits()) {
2765 case 8: CanonicalVT = MVT::v16i8; break;
2766 case 16: CanonicalVT = MVT::v8i16; break;
2767 case 32: CanonicalVT = MVT::v4i32; break;
2768 case 64: CanonicalVT = MVT::v2i64; break;
2769 default: llvm_unreachable("unexpected splat element type"); break;
2773 // Build a canonical splat for this value.
2774 SmallVector<SDValue, 8> Ops;
2775 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2776 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2778 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2781 // If this is a case we can't handle, return null and let the default
2782 // expansion code take care of it.
2783 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2784 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2785 DebugLoc dl = Op.getDebugLoc();
2786 EVT VT = Op.getValueType();
2788 APInt SplatBits, SplatUndef;
2789 unsigned SplatBitSize;
2791 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2792 if (SplatBitSize <= 64) {
2793 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2794 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2796 return BuildSplat(Val, VT, DAG, dl);
2800 // If there are only 2 elements in a 128-bit vector, insert them into an
2801 // undef vector. This handles the common case for 128-bit vector argument
2802 // passing, where the insertions should be translated to subreg accesses
2803 // with no real instructions.
2804 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2805 SDValue Val = DAG.getUNDEF(VT);
2806 SDValue Op0 = Op.getOperand(0);
2807 SDValue Op1 = Op.getOperand(1);
2808 if (Op0.getOpcode() != ISD::UNDEF)
2809 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2810 DAG.getIntPtrConstant(0));
2811 if (Op1.getOpcode() != ISD::UNDEF)
2812 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2813 DAG.getIntPtrConstant(1));
2820 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2821 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2822 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2823 /// are assumed to be legal.
2825 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2827 if (VT.getVectorNumElements() == 4 &&
2828 (VT.is128BitVector() || VT.is64BitVector())) {
2829 unsigned PFIndexes[4];
2830 for (unsigned i = 0; i != 4; ++i) {
2834 PFIndexes[i] = M[i];
2837 // Compute the index in the perfect shuffle table.
2838 unsigned PFTableIndex =
2839 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2840 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2841 unsigned Cost = (PFEntry >> 30);
2848 unsigned Imm, WhichResult;
2850 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2851 isVREVMask(M, VT, 64) ||
2852 isVREVMask(M, VT, 32) ||
2853 isVREVMask(M, VT, 16) ||
2854 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2855 isVTRNMask(M, VT, WhichResult) ||
2856 isVUZPMask(M, VT, WhichResult) ||
2857 isVZIPMask(M, VT, WhichResult) ||
2858 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
2859 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
2860 isVZIP_v_undef_Mask(M, VT, WhichResult));
2863 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2864 /// the specified operations to build the shuffle.
2865 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2866 SDValue RHS, SelectionDAG &DAG,
2868 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2869 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2870 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2873 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2882 OP_VUZPL, // VUZP, left result
2883 OP_VUZPR, // VUZP, right result
2884 OP_VZIPL, // VZIP, left result
2885 OP_VZIPR, // VZIP, right result
2886 OP_VTRNL, // VTRN, left result
2887 OP_VTRNR // VTRN, right result
2890 if (OpNum == OP_COPY) {
2891 if (LHSID == (1*9+2)*9+3) return LHS;
2892 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2896 SDValue OpLHS, OpRHS;
2897 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2898 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2899 EVT VT = OpLHS.getValueType();
2902 default: llvm_unreachable("Unknown shuffle opcode!");
2904 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2909 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2910 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2914 return DAG.getNode(ARMISD::VEXT, dl, VT,
2916 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2919 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2920 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2923 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2924 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2927 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2928 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2932 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2933 SDValue V1 = Op.getOperand(0);
2934 SDValue V2 = Op.getOperand(1);
2935 DebugLoc dl = Op.getDebugLoc();
2936 EVT VT = Op.getValueType();
2937 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2938 SmallVector<int, 8> ShuffleMask;
2940 // Convert shuffles that are directly supported on NEON to target-specific
2941 // DAG nodes, instead of keeping them as shuffles and matching them again
2942 // during code selection. This is more efficient and avoids the possibility
2943 // of inconsistencies between legalization and selection.
2944 // FIXME: floating-point vectors should be canonicalized to integer vectors
2945 // of the same time so that they get CSEd properly.
2946 SVN->getMask(ShuffleMask);
2948 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2949 int Lane = SVN->getSplatIndex();
2950 // If this is undef splat, generate it via "just" vdup, if possible.
2951 if (Lane == -1) Lane = 0;
2953 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2954 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2956 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2957 DAG.getConstant(Lane, MVT::i32));
2962 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2965 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2966 DAG.getConstant(Imm, MVT::i32));
2969 if (isVREVMask(ShuffleMask, VT, 64))
2970 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2971 if (isVREVMask(ShuffleMask, VT, 32))
2972 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2973 if (isVREVMask(ShuffleMask, VT, 16))
2974 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2976 // Check for Neon shuffles that modify both input vectors in place.
2977 // If both results are used, i.e., if there are two shuffles with the same
2978 // source operands and with masks corresponding to both results of one of
2979 // these operations, DAG memoization will ensure that a single node is
2980 // used for both shuffles.
2981 unsigned WhichResult;
2982 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2983 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2984 V1, V2).getValue(WhichResult);
2985 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2986 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2987 V1, V2).getValue(WhichResult);
2988 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2989 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2990 V1, V2).getValue(WhichResult);
2992 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
2993 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2994 V1, V1).getValue(WhichResult);
2995 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2996 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2997 V1, V1).getValue(WhichResult);
2998 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
2999 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3000 V1, V1).getValue(WhichResult);
3002 // If the shuffle is not directly supported and it has 4 elements, use
3003 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3004 if (VT.getVectorNumElements() == 4 &&
3005 (VT.is128BitVector() || VT.is64BitVector())) {
3006 unsigned PFIndexes[4];
3007 for (unsigned i = 0; i != 4; ++i) {
3008 if (ShuffleMask[i] < 0)
3011 PFIndexes[i] = ShuffleMask[i];
3014 // Compute the index in the perfect shuffle table.
3015 unsigned PFTableIndex =
3016 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3018 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3019 unsigned Cost = (PFEntry >> 30);
3022 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3028 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3029 EVT VT = Op.getValueType();
3030 DebugLoc dl = Op.getDebugLoc();
3031 SDValue Vec = Op.getOperand(0);
3032 SDValue Lane = Op.getOperand(1);
3033 assert(VT == MVT::i32 &&
3034 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3035 "unexpected type for custom-lowering vector extract");
3036 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3039 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3040 // The only time a CONCAT_VECTORS operation can have legal types is when
3041 // two 64-bit vectors are concatenated to a 128-bit vector.
3042 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3043 "unexpected CONCAT_VECTORS");
3044 DebugLoc dl = Op.getDebugLoc();
3045 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3046 SDValue Op0 = Op.getOperand(0);
3047 SDValue Op1 = Op.getOperand(1);
3048 if (Op0.getOpcode() != ISD::UNDEF)
3049 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3050 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3051 DAG.getIntPtrConstant(0));
3052 if (Op1.getOpcode() != ISD::UNDEF)
3053 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3054 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3055 DAG.getIntPtrConstant(1));
3056 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3059 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3060 switch (Op.getOpcode()) {
3061 default: llvm_unreachable("Don't know how to custom lower this!");
3062 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3063 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3064 case ISD::GlobalAddress:
3065 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3066 LowerGlobalAddressELF(Op, DAG);
3067 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3068 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3069 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3070 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3071 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3072 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
3073 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3074 case ISD::SINT_TO_FP:
3075 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3076 case ISD::FP_TO_SINT:
3077 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3078 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3079 case ISD::RETURNADDR: break;
3080 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3081 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3082 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3084 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3087 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3088 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3089 case ISD::SRL_PARTS:
3090 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3091 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3092 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3093 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3094 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3095 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3096 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3101 /// ReplaceNodeResults - Replace the results of node with an illegal result
3102 /// type with new values built out of custom code.
3103 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3104 SmallVectorImpl<SDValue>&Results,
3105 SelectionDAG &DAG) {
3106 switch (N->getOpcode()) {
3108 llvm_unreachable("Don't know how to custom expand this!");
3110 case ISD::BIT_CONVERT:
3111 Results.push_back(ExpandBIT_CONVERT(N, DAG));
3115 SDValue Res = LowerShift(N, DAG, Subtarget);
3117 Results.push_back(Res);
3123 //===----------------------------------------------------------------------===//
3124 // ARM Scheduler Hooks
3125 //===----------------------------------------------------------------------===//
3128 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3129 MachineBasicBlock *BB,
3130 unsigned Size) const {
3131 unsigned dest = MI->getOperand(0).getReg();
3132 unsigned ptr = MI->getOperand(1).getReg();
3133 unsigned oldval = MI->getOperand(2).getReg();
3134 unsigned newval = MI->getOperand(3).getReg();
3135 unsigned scratch = BB->getParent()->getRegInfo()
3136 .createVirtualRegister(ARM::GPRRegisterClass);
3137 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3138 DebugLoc dl = MI->getDebugLoc();
3139 bool isThumb2 = Subtarget->isThumb2();
3141 unsigned ldrOpc, strOpc;
3143 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3145 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3146 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3149 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3150 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3153 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3154 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3158 MachineFunction *MF = BB->getParent();
3159 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3160 MachineFunction::iterator It = BB;
3161 ++It; // insert the new blocks after the current block
3163 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3164 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3165 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3166 MF->insert(It, loop1MBB);
3167 MF->insert(It, loop2MBB);
3168 MF->insert(It, exitMBB);
3169 exitMBB->transferSuccessors(BB);
3173 // fallthrough --> loop1MBB
3174 BB->addSuccessor(loop1MBB);
3177 // ldrex dest, [ptr]
3181 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3182 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3183 .addReg(dest).addReg(oldval));
3184 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3185 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3186 BB->addSuccessor(loop2MBB);
3187 BB->addSuccessor(exitMBB);
3190 // strex scratch, newval, [ptr]
3194 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3196 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3197 .addReg(scratch).addImm(0));
3198 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3199 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3200 BB->addSuccessor(loop1MBB);
3201 BB->addSuccessor(exitMBB);
3207 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3213 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3214 unsigned Size, unsigned BinOpcode) const {
3215 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3216 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3218 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3219 MachineFunction *MF = BB->getParent();
3220 MachineFunction::iterator It = BB;
3223 unsigned dest = MI->getOperand(0).getReg();
3224 unsigned ptr = MI->getOperand(1).getReg();
3225 unsigned incr = MI->getOperand(2).getReg();
3226 DebugLoc dl = MI->getDebugLoc();
3228 bool isThumb2 = Subtarget->isThumb2();
3229 unsigned ldrOpc, strOpc;
3231 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3233 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3234 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3237 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3238 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3241 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3242 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3246 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3247 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3248 MF->insert(It, loopMBB);
3249 MF->insert(It, exitMBB);
3250 exitMBB->transferSuccessors(BB);
3252 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3253 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3254 unsigned scratch2 = (!BinOpcode) ? incr :
3255 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3259 // fallthrough --> loopMBB
3260 BB->addSuccessor(loopMBB);
3264 // <binop> scratch2, dest, incr
3265 // strex scratch, scratch2, ptr
3268 // fallthrough --> exitMBB
3270 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3272 // operand order needs to go the other way for NAND
3273 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3274 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3275 addReg(incr).addReg(dest)).addReg(0);
3277 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3278 addReg(dest).addReg(incr)).addReg(0);
3281 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3283 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3284 .addReg(scratch).addImm(0));
3285 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3286 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3288 BB->addSuccessor(loopMBB);
3289 BB->addSuccessor(exitMBB);
3295 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3301 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3302 MachineBasicBlock *BB,
3303 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
3304 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3305 DebugLoc dl = MI->getDebugLoc();
3306 bool isThumb2 = Subtarget->isThumb2();
3307 switch (MI->getOpcode()) {
3310 llvm_unreachable("Unexpected instr type to insert");
3312 case ARM::ATOMIC_LOAD_ADD_I8:
3313 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3314 case ARM::ATOMIC_LOAD_ADD_I16:
3315 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3316 case ARM::ATOMIC_LOAD_ADD_I32:
3317 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3319 case ARM::ATOMIC_LOAD_AND_I8:
3320 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3321 case ARM::ATOMIC_LOAD_AND_I16:
3322 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3323 case ARM::ATOMIC_LOAD_AND_I32:
3324 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3326 case ARM::ATOMIC_LOAD_OR_I8:
3327 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3328 case ARM::ATOMIC_LOAD_OR_I16:
3329 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3330 case ARM::ATOMIC_LOAD_OR_I32:
3331 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3333 case ARM::ATOMIC_LOAD_XOR_I8:
3334 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3335 case ARM::ATOMIC_LOAD_XOR_I16:
3336 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3337 case ARM::ATOMIC_LOAD_XOR_I32:
3338 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3340 case ARM::ATOMIC_LOAD_NAND_I8:
3341 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3342 case ARM::ATOMIC_LOAD_NAND_I16:
3343 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3344 case ARM::ATOMIC_LOAD_NAND_I32:
3345 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3347 case ARM::ATOMIC_LOAD_SUB_I8:
3348 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3349 case ARM::ATOMIC_LOAD_SUB_I16:
3350 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3351 case ARM::ATOMIC_LOAD_SUB_I32:
3352 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3354 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3355 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3356 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3358 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3359 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3360 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3362 case ARM::tMOVCCr_pseudo: {
3363 // To "insert" a SELECT_CC instruction, we actually have to insert the
3364 // diamond control-flow pattern. The incoming instruction knows the
3365 // destination vreg to set, the condition code register to branch on, the
3366 // true/false values to select between, and a branch opcode to use.
3367 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3368 MachineFunction::iterator It = BB;
3374 // cmpTY ccX, r1, r2
3376 // fallthrough --> copy0MBB
3377 MachineBasicBlock *thisMBB = BB;
3378 MachineFunction *F = BB->getParent();
3379 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3380 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3381 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3382 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3383 F->insert(It, copy0MBB);
3384 F->insert(It, sinkMBB);
3385 // Update machine-CFG edges by first adding all successors of the current
3386 // block to the new block which will contain the Phi node for the select.
3387 // Also inform sdisel of the edge changes.
3388 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3389 E = BB->succ_end(); I != E; ++I) {
3390 EM->insert(std::make_pair(*I, sinkMBB));
3391 sinkMBB->addSuccessor(*I);
3393 // Next, remove all successors of the current block, and add the true
3394 // and fallthrough blocks as its successors.
3395 while (!BB->succ_empty())
3396 BB->removeSuccessor(BB->succ_begin());
3397 BB->addSuccessor(copy0MBB);
3398 BB->addSuccessor(sinkMBB);
3401 // %FalseValue = ...
3402 // # fallthrough to sinkMBB
3405 // Update machine-CFG edges
3406 BB->addSuccessor(sinkMBB);
3409 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3412 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3413 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3414 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3416 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3423 case ARM::t2SUBrSPi_:
3424 case ARM::t2SUBrSPi12_:
3425 case ARM::t2SUBrSPs_: {
3426 MachineFunction *MF = BB->getParent();
3427 unsigned DstReg = MI->getOperand(0).getReg();
3428 unsigned SrcReg = MI->getOperand(1).getReg();
3429 bool DstIsDead = MI->getOperand(0).isDead();
3430 bool SrcIsKill = MI->getOperand(1).isKill();
3432 if (SrcReg != ARM::SP) {
3433 // Copy the source to SP from virtual register.
3434 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3435 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3436 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3437 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3438 .addReg(SrcReg, getKillRegState(SrcIsKill));
3442 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3443 switch (MI->getOpcode()) {
3445 llvm_unreachable("Unexpected pseudo instruction!");
3451 OpOpc = ARM::tADDspr;
3454 OpOpc = ARM::tSUBspi;
3456 case ARM::t2SUBrSPi_:
3457 OpOpc = ARM::t2SUBrSPi;
3458 NeedPred = true; NeedCC = true;
3460 case ARM::t2SUBrSPi12_:
3461 OpOpc = ARM::t2SUBrSPi12;
3464 case ARM::t2SUBrSPs_:
3465 OpOpc = ARM::t2SUBrSPs;
3466 NeedPred = true; NeedCC = true; NeedOp3 = true;
3469 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3470 if (OpOpc == ARM::tAND)
3471 AddDefaultT1CC(MIB);
3472 MIB.addReg(ARM::SP);
3473 MIB.addOperand(MI->getOperand(2));
3475 MIB.addOperand(MI->getOperand(3));
3477 AddDefaultPred(MIB);
3481 // Copy the result from SP to virtual register.
3482 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3483 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3484 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3485 BuildMI(BB, dl, TII->get(CopyOpc))
3486 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3488 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3494 //===----------------------------------------------------------------------===//
3495 // ARM Optimization Hooks
3496 //===----------------------------------------------------------------------===//
3499 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3500 TargetLowering::DAGCombinerInfo &DCI) {
3501 SelectionDAG &DAG = DCI.DAG;
3502 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3503 EVT VT = N->getValueType(0);
3504 unsigned Opc = N->getOpcode();
3505 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3506 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3507 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3508 ISD::CondCode CC = ISD::SETCC_INVALID;
3511 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3513 SDValue CCOp = Slct.getOperand(0);
3514 if (CCOp.getOpcode() == ISD::SETCC)
3515 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3518 bool DoXform = false;
3520 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3523 if (LHS.getOpcode() == ISD::Constant &&
3524 cast<ConstantSDNode>(LHS)->isNullValue()) {
3526 } else if (CC != ISD::SETCC_INVALID &&
3527 RHS.getOpcode() == ISD::Constant &&
3528 cast<ConstantSDNode>(RHS)->isNullValue()) {
3529 std::swap(LHS, RHS);
3530 SDValue Op0 = Slct.getOperand(0);
3531 EVT OpVT = isSlctCC ? Op0.getValueType() :
3532 Op0.getOperand(0).getValueType();
3533 bool isInt = OpVT.isInteger();
3534 CC = ISD::getSetCCInverse(CC, isInt);
3536 if (!TLI.isCondCodeLegal(CC, OpVT))
3537 return SDValue(); // Inverse operator isn't legal.
3544 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3546 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3547 Slct.getOperand(0), Slct.getOperand(1), CC);
3548 SDValue CCOp = Slct.getOperand(0);
3550 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3551 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3552 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3553 CCOp, OtherOp, Result);
3558 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3559 static SDValue PerformADDCombine(SDNode *N,
3560 TargetLowering::DAGCombinerInfo &DCI) {
3561 // added by evan in r37685 with no testcase.
3562 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3564 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3565 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3566 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3567 if (Result.getNode()) return Result;
3569 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3570 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3571 if (Result.getNode()) return Result;
3577 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3578 static SDValue PerformSUBCombine(SDNode *N,
3579 TargetLowering::DAGCombinerInfo &DCI) {
3580 // added by evan in r37685 with no testcase.
3581 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3583 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3584 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3585 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3586 if (Result.getNode()) return Result;
3592 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
3593 /// ARMISD::VMOVRRD.
3594 static SDValue PerformVMOVRRDCombine(SDNode *N,
3595 TargetLowering::DAGCombinerInfo &DCI) {
3596 // fmrrd(fmdrr x, y) -> x,y
3597 SDValue InDouble = N->getOperand(0);
3598 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3599 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3603 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3604 /// operand of a vector shift operation, where all the elements of the
3605 /// build_vector must have the same constant integer value.
3606 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3607 // Ignore bit_converts.
3608 while (Op.getOpcode() == ISD::BIT_CONVERT)
3609 Op = Op.getOperand(0);
3610 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3611 APInt SplatBits, SplatUndef;
3612 unsigned SplatBitSize;
3614 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3615 HasAnyUndefs, ElementBits) ||
3616 SplatBitSize > ElementBits)
3618 Cnt = SplatBits.getSExtValue();
3622 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3623 /// operand of a vector shift left operation. That value must be in the range:
3624 /// 0 <= Value < ElementBits for a left shift; or
3625 /// 0 <= Value <= ElementBits for a long left shift.
3626 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3627 assert(VT.isVector() && "vector shift count is not a vector type");
3628 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3629 if (! getVShiftImm(Op, ElementBits, Cnt))
3631 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3634 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3635 /// operand of a vector shift right operation. For a shift opcode, the value
3636 /// is positive, but for an intrinsic the value count must be negative. The
3637 /// absolute value must be in the range:
3638 /// 1 <= |Value| <= ElementBits for a right shift; or
3639 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3640 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3642 assert(VT.isVector() && "vector shift count is not a vector type");
3643 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3644 if (! getVShiftImm(Op, ElementBits, Cnt))
3648 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3651 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3652 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3653 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3656 // Don't do anything for most intrinsics.
3659 // Vector shifts: check for immediate versions and lower them.
3660 // Note: This is done during DAG combining instead of DAG legalizing because
3661 // the build_vectors for 64-bit vector element shift counts are generally
3662 // not legal, and it is hard to see their values after they get legalized to
3663 // loads from a constant pool.
3664 case Intrinsic::arm_neon_vshifts:
3665 case Intrinsic::arm_neon_vshiftu:
3666 case Intrinsic::arm_neon_vshiftls:
3667 case Intrinsic::arm_neon_vshiftlu:
3668 case Intrinsic::arm_neon_vshiftn:
3669 case Intrinsic::arm_neon_vrshifts:
3670 case Intrinsic::arm_neon_vrshiftu:
3671 case Intrinsic::arm_neon_vrshiftn:
3672 case Intrinsic::arm_neon_vqshifts:
3673 case Intrinsic::arm_neon_vqshiftu:
3674 case Intrinsic::arm_neon_vqshiftsu:
3675 case Intrinsic::arm_neon_vqshiftns:
3676 case Intrinsic::arm_neon_vqshiftnu:
3677 case Intrinsic::arm_neon_vqshiftnsu:
3678 case Intrinsic::arm_neon_vqrshiftns:
3679 case Intrinsic::arm_neon_vqrshiftnu:
3680 case Intrinsic::arm_neon_vqrshiftnsu: {
3681 EVT VT = N->getOperand(1).getValueType();
3683 unsigned VShiftOpc = 0;
3686 case Intrinsic::arm_neon_vshifts:
3687 case Intrinsic::arm_neon_vshiftu:
3688 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3689 VShiftOpc = ARMISD::VSHL;
3692 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3693 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3694 ARMISD::VSHRs : ARMISD::VSHRu);
3699 case Intrinsic::arm_neon_vshiftls:
3700 case Intrinsic::arm_neon_vshiftlu:
3701 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3703 llvm_unreachable("invalid shift count for vshll intrinsic");
3705 case Intrinsic::arm_neon_vrshifts:
3706 case Intrinsic::arm_neon_vrshiftu:
3707 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3711 case Intrinsic::arm_neon_vqshifts:
3712 case Intrinsic::arm_neon_vqshiftu:
3713 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3717 case Intrinsic::arm_neon_vqshiftsu:
3718 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3720 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3722 case Intrinsic::arm_neon_vshiftn:
3723 case Intrinsic::arm_neon_vrshiftn:
3724 case Intrinsic::arm_neon_vqshiftns:
3725 case Intrinsic::arm_neon_vqshiftnu:
3726 case Intrinsic::arm_neon_vqshiftnsu:
3727 case Intrinsic::arm_neon_vqrshiftns:
3728 case Intrinsic::arm_neon_vqrshiftnu:
3729 case Intrinsic::arm_neon_vqrshiftnsu:
3730 // Narrowing shifts require an immediate right shift.
3731 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3733 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3736 llvm_unreachable("unhandled vector shift");
3740 case Intrinsic::arm_neon_vshifts:
3741 case Intrinsic::arm_neon_vshiftu:
3742 // Opcode already set above.
3744 case Intrinsic::arm_neon_vshiftls:
3745 case Intrinsic::arm_neon_vshiftlu:
3746 if (Cnt == VT.getVectorElementType().getSizeInBits())
3747 VShiftOpc = ARMISD::VSHLLi;
3749 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3750 ARMISD::VSHLLs : ARMISD::VSHLLu);
3752 case Intrinsic::arm_neon_vshiftn:
3753 VShiftOpc = ARMISD::VSHRN; break;
3754 case Intrinsic::arm_neon_vrshifts:
3755 VShiftOpc = ARMISD::VRSHRs; break;
3756 case Intrinsic::arm_neon_vrshiftu:
3757 VShiftOpc = ARMISD::VRSHRu; break;
3758 case Intrinsic::arm_neon_vrshiftn:
3759 VShiftOpc = ARMISD::VRSHRN; break;
3760 case Intrinsic::arm_neon_vqshifts:
3761 VShiftOpc = ARMISD::VQSHLs; break;
3762 case Intrinsic::arm_neon_vqshiftu:
3763 VShiftOpc = ARMISD::VQSHLu; break;
3764 case Intrinsic::arm_neon_vqshiftsu:
3765 VShiftOpc = ARMISD::VQSHLsu; break;
3766 case Intrinsic::arm_neon_vqshiftns:
3767 VShiftOpc = ARMISD::VQSHRNs; break;
3768 case Intrinsic::arm_neon_vqshiftnu:
3769 VShiftOpc = ARMISD::VQSHRNu; break;
3770 case Intrinsic::arm_neon_vqshiftnsu:
3771 VShiftOpc = ARMISD::VQSHRNsu; break;
3772 case Intrinsic::arm_neon_vqrshiftns:
3773 VShiftOpc = ARMISD::VQRSHRNs; break;
3774 case Intrinsic::arm_neon_vqrshiftnu:
3775 VShiftOpc = ARMISD::VQRSHRNu; break;
3776 case Intrinsic::arm_neon_vqrshiftnsu:
3777 VShiftOpc = ARMISD::VQRSHRNsu; break;
3780 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3781 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3784 case Intrinsic::arm_neon_vshiftins: {
3785 EVT VT = N->getOperand(1).getValueType();
3787 unsigned VShiftOpc = 0;
3789 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3790 VShiftOpc = ARMISD::VSLI;
3791 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3792 VShiftOpc = ARMISD::VSRI;
3794 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3797 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3798 N->getOperand(1), N->getOperand(2),
3799 DAG.getConstant(Cnt, MVT::i32));
3802 case Intrinsic::arm_neon_vqrshifts:
3803 case Intrinsic::arm_neon_vqrshiftu:
3804 // No immediate versions of these to check for.
3811 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3812 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3813 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3814 /// vector element shift counts are generally not legal, and it is hard to see
3815 /// their values after they get legalized to loads from a constant pool.
3816 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3817 const ARMSubtarget *ST) {
3818 EVT VT = N->getValueType(0);
3820 // Nothing to be done for scalar shifts.
3821 if (! VT.isVector())
3824 assert(ST->hasNEON() && "unexpected vector shift");
3827 switch (N->getOpcode()) {
3828 default: llvm_unreachable("unexpected shift opcode");
3831 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3832 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3833 DAG.getConstant(Cnt, MVT::i32));
3838 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3839 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3840 ARMISD::VSHRs : ARMISD::VSHRu);
3841 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3842 DAG.getConstant(Cnt, MVT::i32));
3848 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3849 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3850 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3851 const ARMSubtarget *ST) {
3852 SDValue N0 = N->getOperand(0);
3854 // Check for sign- and zero-extensions of vector extract operations of 8-
3855 // and 16-bit vector elements. NEON supports these directly. They are
3856 // handled during DAG combining because type legalization will promote them
3857 // to 32-bit types and it is messy to recognize the operations after that.
3858 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3859 SDValue Vec = N0.getOperand(0);
3860 SDValue Lane = N0.getOperand(1);
3861 EVT VT = N->getValueType(0);
3862 EVT EltVT = N0.getValueType();
3863 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3865 if (VT == MVT::i32 &&
3866 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3867 TLI.isTypeLegal(Vec.getValueType())) {
3870 switch (N->getOpcode()) {
3871 default: llvm_unreachable("unexpected opcode");
3872 case ISD::SIGN_EXTEND:
3873 Opc = ARMISD::VGETLANEs;
3875 case ISD::ZERO_EXTEND:
3876 case ISD::ANY_EXTEND:
3877 Opc = ARMISD::VGETLANEu;
3880 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3887 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
3888 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
3889 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
3890 const ARMSubtarget *ST) {
3891 // If the target supports NEON, try to use vmax/vmin instructions for f32
3892 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
3893 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
3894 // a NaN; only do the transformation when it matches that behavior.
3896 // For now only do this when using NEON for FP operations; if using VFP, it
3897 // is not obvious that the benefit outweighs the cost of switching to the
3899 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
3900 N->getValueType(0) != MVT::f32)
3903 SDValue CondLHS = N->getOperand(0);
3904 SDValue CondRHS = N->getOperand(1);
3905 SDValue LHS = N->getOperand(2);
3906 SDValue RHS = N->getOperand(3);
3907 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
3909 unsigned Opcode = 0;
3911 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
3912 IsReversed = false; // x CC y ? x : y
3913 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
3914 IsReversed = true ; // x CC y ? y : x
3928 // If LHS is NaN, an ordered comparison will be false and the result will
3929 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
3930 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3931 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
3932 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3934 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
3935 // will return -0, so vmin can only be used for unsafe math or if one of
3936 // the operands is known to be nonzero.
3937 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
3939 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3941 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
3950 // If LHS is NaN, an ordered comparison will be false and the result will
3951 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
3952 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
3953 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
3954 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
3956 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
3957 // will return +0, so vmax can only be used for unsafe math or if one of
3958 // the operands is known to be nonzero.
3959 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
3961 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
3963 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
3969 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
3972 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3973 DAGCombinerInfo &DCI) const {
3974 switch (N->getOpcode()) {
3976 case ISD::ADD: return PerformADDCombine(N, DCI);
3977 case ISD::SUB: return PerformSUBCombine(N, DCI);
3978 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3979 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
3982 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
3983 case ISD::SIGN_EXTEND:
3984 case ISD::ZERO_EXTEND:
3985 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
3986 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
3991 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3992 if (!Subtarget->hasV6Ops())
3993 // Pre-v6 does not support unaligned mem access.
3996 // v6+ may or may not support unaligned mem access depending on the system
3998 // FIXME: This is pretty conservative. Should we provide cmdline option to
3999 // control the behaviour?
4000 if (!Subtarget->isTargetDarwin())
4004 switch (VT.getSimpleVT().SimpleTy) {
4011 // FIXME: VLD1 etc with standard alignment is legal.
4015 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4020 switch (VT.getSimpleVT().SimpleTy) {
4021 default: return false;
4036 if ((V & (Scale - 1)) != 0)
4039 return V == (V & ((1LL << 5) - 1));
4042 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4043 const ARMSubtarget *Subtarget) {
4050 switch (VT.getSimpleVT().SimpleTy) {
4051 default: return false;
4056 // + imm12 or - imm8
4058 return V == (V & ((1LL << 8) - 1));
4059 return V == (V & ((1LL << 12) - 1));
4062 // Same as ARM mode. FIXME: NEON?
4063 if (!Subtarget->hasVFP2())
4068 return V == (V & ((1LL << 8) - 1));
4072 /// isLegalAddressImmediate - Return true if the integer value can be used
4073 /// as the offset of the target addressing mode for load / store of the
4075 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4076 const ARMSubtarget *Subtarget) {
4083 if (Subtarget->isThumb1Only())
4084 return isLegalT1AddressImmediate(V, VT);
4085 else if (Subtarget->isThumb2())
4086 return isLegalT2AddressImmediate(V, VT, Subtarget);
4091 switch (VT.getSimpleVT().SimpleTy) {
4092 default: return false;
4097 return V == (V & ((1LL << 12) - 1));
4100 return V == (V & ((1LL << 8) - 1));
4103 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4108 return V == (V & ((1LL << 8) - 1));
4112 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4114 int Scale = AM.Scale;
4118 switch (VT.getSimpleVT().SimpleTy) {
4119 default: return false;
4128 return Scale == 2 || Scale == 4 || Scale == 8;
4131 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4135 // Note, we allow "void" uses (basically, uses that aren't loads or
4136 // stores), because arm allows folding a scale into many arithmetic
4137 // operations. This should be made more precise and revisited later.
4139 // Allow r << imm, but the imm has to be a multiple of two.
4140 if (Scale & 1) return false;
4141 return isPowerOf2_32(Scale);
4145 /// isLegalAddressingMode - Return true if the addressing mode represented
4146 /// by AM is legal for this target, for a load/store of the specified type.
4147 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4148 const Type *Ty) const {
4149 EVT VT = getValueType(Ty, true);
4150 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4153 // Can never fold addr of global into load/store.
4158 case 0: // no scale reg, must be "r+i" or "r", or "i".
4161 if (Subtarget->isThumb1Only())
4165 // ARM doesn't support any R+R*scale+imm addr modes.
4172 if (Subtarget->isThumb2())
4173 return isLegalT2ScaledAddressingMode(AM, VT);
4175 int Scale = AM.Scale;
4176 switch (VT.getSimpleVT().SimpleTy) {
4177 default: return false;
4181 if (Scale < 0) Scale = -Scale;
4185 return isPowerOf2_32(Scale & ~1);
4189 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4194 // Note, we allow "void" uses (basically, uses that aren't loads or
4195 // stores), because arm allows folding a scale into many arithmetic
4196 // operations. This should be made more precise and revisited later.
4198 // Allow r << imm, but the imm has to be a multiple of two.
4199 if (Scale & 1) return false;
4200 return isPowerOf2_32(Scale);
4207 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4208 /// icmp immediate, that is the target has icmp instructions which can compare
4209 /// a register against the immediate without having to materialize the
4210 /// immediate into a register.
4211 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4212 if (!Subtarget->isThumb())
4213 return ARM_AM::getSOImmVal(Imm) != -1;
4214 if (Subtarget->isThumb2())
4215 return ARM_AM::getT2SOImmVal(Imm) != -1;
4216 return Imm >= 0 && Imm <= 255;
4219 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4220 bool isSEXTLoad, SDValue &Base,
4221 SDValue &Offset, bool &isInc,
4222 SelectionDAG &DAG) {
4223 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4226 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4228 Base = Ptr->getOperand(0);
4229 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4230 int RHSC = (int)RHS->getZExtValue();
4231 if (RHSC < 0 && RHSC > -256) {
4232 assert(Ptr->getOpcode() == ISD::ADD);
4234 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4238 isInc = (Ptr->getOpcode() == ISD::ADD);
4239 Offset = Ptr->getOperand(1);
4241 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4243 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4244 int RHSC = (int)RHS->getZExtValue();
4245 if (RHSC < 0 && RHSC > -0x1000) {
4246 assert(Ptr->getOpcode() == ISD::ADD);
4248 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4249 Base = Ptr->getOperand(0);
4254 if (Ptr->getOpcode() == ISD::ADD) {
4256 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4257 if (ShOpcVal != ARM_AM::no_shift) {
4258 Base = Ptr->getOperand(1);
4259 Offset = Ptr->getOperand(0);
4261 Base = Ptr->getOperand(0);
4262 Offset = Ptr->getOperand(1);
4267 isInc = (Ptr->getOpcode() == ISD::ADD);
4268 Base = Ptr->getOperand(0);
4269 Offset = Ptr->getOperand(1);
4273 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4277 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4278 bool isSEXTLoad, SDValue &Base,
4279 SDValue &Offset, bool &isInc,
4280 SelectionDAG &DAG) {
4281 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4284 Base = Ptr->getOperand(0);
4285 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4286 int RHSC = (int)RHS->getZExtValue();
4287 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4288 assert(Ptr->getOpcode() == ISD::ADD);
4290 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4292 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4293 isInc = Ptr->getOpcode() == ISD::ADD;
4294 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4302 /// getPreIndexedAddressParts - returns true by value, base pointer and
4303 /// offset pointer and addressing mode by reference if the node's address
4304 /// can be legally represented as pre-indexed load / store address.
4306 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4308 ISD::MemIndexedMode &AM,
4309 SelectionDAG &DAG) const {
4310 if (Subtarget->isThumb1Only())
4315 bool isSEXTLoad = false;
4316 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4317 Ptr = LD->getBasePtr();
4318 VT = LD->getMemoryVT();
4319 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4320 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4321 Ptr = ST->getBasePtr();
4322 VT = ST->getMemoryVT();
4327 bool isLegal = false;
4328 if (Subtarget->isThumb2())
4329 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4330 Offset, isInc, DAG);
4332 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4333 Offset, isInc, DAG);
4337 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4341 /// getPostIndexedAddressParts - returns true by value, base pointer and
4342 /// offset pointer and addressing mode by reference if this node can be
4343 /// combined with a load / store to form a post-indexed load / store.
4344 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4347 ISD::MemIndexedMode &AM,
4348 SelectionDAG &DAG) const {
4349 if (Subtarget->isThumb1Only())
4354 bool isSEXTLoad = false;
4355 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4356 VT = LD->getMemoryVT();
4357 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4358 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4359 VT = ST->getMemoryVT();
4364 bool isLegal = false;
4365 if (Subtarget->isThumb2())
4366 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4369 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4374 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4378 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4382 const SelectionDAG &DAG,
4383 unsigned Depth) const {
4384 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4385 switch (Op.getOpcode()) {
4387 case ARMISD::CMOV: {
4388 // Bits are known zero/one if known on the LHS and RHS.
4389 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4390 if (KnownZero == 0 && KnownOne == 0) return;
4392 APInt KnownZeroRHS, KnownOneRHS;
4393 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4394 KnownZeroRHS, KnownOneRHS, Depth+1);
4395 KnownZero &= KnownZeroRHS;
4396 KnownOne &= KnownOneRHS;
4402 //===----------------------------------------------------------------------===//
4403 // ARM Inline Assembly Support
4404 //===----------------------------------------------------------------------===//
4406 /// getConstraintType - Given a constraint letter, return the type of
4407 /// constraint it is for this target.
4408 ARMTargetLowering::ConstraintType
4409 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4410 if (Constraint.size() == 1) {
4411 switch (Constraint[0]) {
4413 case 'l': return C_RegisterClass;
4414 case 'w': return C_RegisterClass;
4417 return TargetLowering::getConstraintType(Constraint);
4420 std::pair<unsigned, const TargetRegisterClass*>
4421 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4423 if (Constraint.size() == 1) {
4424 // GCC ARM Constraint Letters
4425 switch (Constraint[0]) {
4427 if (Subtarget->isThumb())
4428 return std::make_pair(0U, ARM::tGPRRegisterClass);
4430 return std::make_pair(0U, ARM::GPRRegisterClass);
4432 return std::make_pair(0U, ARM::GPRRegisterClass);
4435 return std::make_pair(0U, ARM::SPRRegisterClass);
4436 if (VT.getSizeInBits() == 64)
4437 return std::make_pair(0U, ARM::DPRRegisterClass);
4438 if (VT.getSizeInBits() == 128)
4439 return std::make_pair(0U, ARM::QPRRegisterClass);
4443 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4446 std::vector<unsigned> ARMTargetLowering::
4447 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4449 if (Constraint.size() != 1)
4450 return std::vector<unsigned>();
4452 switch (Constraint[0]) { // GCC ARM Constraint Letters
4455 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4456 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4459 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4460 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4461 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4462 ARM::R12, ARM::LR, 0);
4465 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4466 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4467 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4468 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4469 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4470 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4471 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4472 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4473 if (VT.getSizeInBits() == 64)
4474 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4475 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4476 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4477 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4478 if (VT.getSizeInBits() == 128)
4479 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4480 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4484 return std::vector<unsigned>();
4487 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4488 /// vector. If it is invalid, don't add anything to Ops.
4489 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4492 std::vector<SDValue>&Ops,
4493 SelectionDAG &DAG) const {
4494 SDValue Result(0, 0);
4496 switch (Constraint) {
4498 case 'I': case 'J': case 'K': case 'L':
4499 case 'M': case 'N': case 'O':
4500 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4504 int64_t CVal64 = C->getSExtValue();
4505 int CVal = (int) CVal64;
4506 // None of these constraints allow values larger than 32 bits. Check
4507 // that the value fits in an int.
4511 switch (Constraint) {
4513 if (Subtarget->isThumb1Only()) {
4514 // This must be a constant between 0 and 255, for ADD
4516 if (CVal >= 0 && CVal <= 255)
4518 } else if (Subtarget->isThumb2()) {
4519 // A constant that can be used as an immediate value in a
4520 // data-processing instruction.
4521 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4524 // A constant that can be used as an immediate value in a
4525 // data-processing instruction.
4526 if (ARM_AM::getSOImmVal(CVal) != -1)
4532 if (Subtarget->isThumb()) { // FIXME thumb2
4533 // This must be a constant between -255 and -1, for negated ADD
4534 // immediates. This can be used in GCC with an "n" modifier that
4535 // prints the negated value, for use with SUB instructions. It is
4536 // not useful otherwise but is implemented for compatibility.
4537 if (CVal >= -255 && CVal <= -1)
4540 // This must be a constant between -4095 and 4095. It is not clear
4541 // what this constraint is intended for. Implemented for
4542 // compatibility with GCC.
4543 if (CVal >= -4095 && CVal <= 4095)
4549 if (Subtarget->isThumb1Only()) {
4550 // A 32-bit value where only one byte has a nonzero value. Exclude
4551 // zero to match GCC. This constraint is used by GCC internally for
4552 // constants that can be loaded with a move/shift combination.
4553 // It is not useful otherwise but is implemented for compatibility.
4554 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4556 } else if (Subtarget->isThumb2()) {
4557 // A constant whose bitwise inverse can be used as an immediate
4558 // value in a data-processing instruction. This can be used in GCC
4559 // with a "B" modifier that prints the inverted value, for use with
4560 // BIC and MVN instructions. It is not useful otherwise but is
4561 // implemented for compatibility.
4562 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4565 // A constant whose bitwise inverse can be used as an immediate
4566 // value in a data-processing instruction. This can be used in GCC
4567 // with a "B" modifier that prints the inverted value, for use with
4568 // BIC and MVN instructions. It is not useful otherwise but is
4569 // implemented for compatibility.
4570 if (ARM_AM::getSOImmVal(~CVal) != -1)
4576 if (Subtarget->isThumb1Only()) {
4577 // This must be a constant between -7 and 7,
4578 // for 3-operand ADD/SUB immediate instructions.
4579 if (CVal >= -7 && CVal < 7)
4581 } else if (Subtarget->isThumb2()) {
4582 // A constant whose negation can be used as an immediate value in a
4583 // data-processing instruction. This can be used in GCC with an "n"
4584 // modifier that prints the negated value, for use with SUB
4585 // instructions. It is not useful otherwise but is implemented for
4587 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4590 // A constant whose negation can be used as an immediate value in a
4591 // data-processing instruction. This can be used in GCC with an "n"
4592 // modifier that prints the negated value, for use with SUB
4593 // instructions. It is not useful otherwise but is implemented for
4595 if (ARM_AM::getSOImmVal(-CVal) != -1)
4601 if (Subtarget->isThumb()) { // FIXME thumb2
4602 // This must be a multiple of 4 between 0 and 1020, for
4603 // ADD sp + immediate.
4604 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4607 // A power of two or a constant between 0 and 32. This is used in
4608 // GCC for the shift amount on shifted register operands, but it is
4609 // useful in general for any shift amounts.
4610 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4616 if (Subtarget->isThumb()) { // FIXME thumb2
4617 // This must be a constant between 0 and 31, for shift amounts.
4618 if (CVal >= 0 && CVal <= 31)
4624 if (Subtarget->isThumb()) { // FIXME thumb2
4625 // This must be a multiple of 4 between -508 and 508, for
4626 // ADD/SUB sp = sp + immediate.
4627 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4632 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4636 if (Result.getNode()) {
4637 Ops.push_back(Result);
4640 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4645 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4646 // The ARM target isn't yet aware of offsets.
4650 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4651 APInt Imm = FPImm.bitcastToAPInt();
4652 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4653 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4654 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4656 // We can handle 4 bits of mantissa.
4657 // mantissa = (16+UInt(e:f:g:h))/16.
4658 if (Mantissa & 0x7ffff)
4661 if ((Mantissa & 0xf) != Mantissa)
4664 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4665 if (Exp < -3 || Exp > 4)
4667 Exp = ((Exp+3) & 0x7) ^ 4;
4669 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4672 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4673 APInt Imm = FPImm.bitcastToAPInt();
4674 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4675 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4676 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4678 // We can handle 4 bits of mantissa.
4679 // mantissa = (16+UInt(e:f:g:h))/16.
4680 if (Mantissa & 0xffffffffffffLL)
4683 if ((Mantissa & 0xf) != Mantissa)
4686 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4687 if (Exp < -3 || Exp > 4)
4689 Exp = ((Exp+3) & 0x7) ^ 4;
4691 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4694 /// isFPImmLegal - Returns true if the target can instruction select the
4695 /// specified FP immediate natively. If false, the legalizer will
4696 /// materialize the FP immediate as a load from a constant pool.
4697 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4698 if (!Subtarget->hasVFP3())
4701 return ARM::getVFPf32Imm(Imm) != -1;
4703 return ARM::getVFPf64Imm(Imm) != -1;