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 "ARMRegisterInfo.h"
21 #include "ARMSubtarget.h"
22 #include "ARMTargetMachine.h"
23 #include "ARMTargetObjectFile.h"
24 #include "llvm/CallingConv.h"
25 #include "llvm/Constants.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instruction.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/CodeGen/CallingConvLower.h"
31 #include "llvm/CodeGen/MachineBasicBlock.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/CodeGen/SelectionDAG.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include "llvm/ADT/VectorExtras.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/MathExtras.h"
44 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
49 CCValAssign::LocInfo &LocInfo,
50 ISD::ArgFlagsTy &ArgFlags,
52 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
53 CCValAssign::LocInfo &LocInfo,
54 ISD::ArgFlagsTy &ArgFlags,
56 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
57 CCValAssign::LocInfo &LocInfo,
58 ISD::ArgFlagsTy &ArgFlags,
61 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
62 MVT PromotedBitwiseVT) {
63 if (VT != PromotedLdStVT) {
64 setOperationAction(ISD::LOAD, VT, Promote);
65 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
67 setOperationAction(ISD::STORE, VT, Promote);
68 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
71 MVT ElemTy = VT.getVectorElementType();
72 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
73 setOperationAction(ISD::VSETCC, VT, Custom);
74 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
75 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
76 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
77 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
78 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
79 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
81 setOperationAction(ISD::SHL, VT, Custom);
82 setOperationAction(ISD::SRA, VT, Custom);
83 setOperationAction(ISD::SRL, VT, Custom);
86 // Promote all bit-wise operations.
87 if (VT.isInteger() && VT != PromotedBitwiseVT) {
88 setOperationAction(ISD::AND, VT, Promote);
89 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
90 setOperationAction(ISD::OR, VT, Promote);
91 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
92 setOperationAction(ISD::XOR, VT, Promote);
93 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
97 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
98 addRegisterClass(VT, ARM::DPRRegisterClass);
99 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
102 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
103 addRegisterClass(VT, ARM::QPRRegisterClass);
104 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
107 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
108 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
109 return new TargetLoweringObjectFileMachO();
110 return new ARMElfTargetObjectFile();
113 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
114 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
115 Subtarget = &TM.getSubtarget<ARMSubtarget>();
117 if (Subtarget->isTargetDarwin()) {
118 // Uses VFP for Thumb libfuncs if available.
119 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
120 // Single-precision floating-point arithmetic.
121 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
122 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
123 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
124 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
126 // Double-precision floating-point arithmetic.
127 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
128 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
129 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
130 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
132 // Single-precision comparisons.
133 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
134 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
135 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
136 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
137 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
138 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
139 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
140 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
142 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
143 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
144 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
145 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
146 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
147 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
148 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
149 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
151 // Double-precision comparisons.
152 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
153 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
154 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
155 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
156 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
157 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
158 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
159 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
161 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
162 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
163 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
164 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
165 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
166 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
167 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
170 // Floating-point to integer conversions.
171 // i64 conversions are done via library routines even when generating VFP
172 // instructions, so use the same ones.
173 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
174 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
175 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
176 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
178 // Conversions between floating types.
179 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
180 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
182 // Integer to floating-point conversions.
183 // i64 conversions are done via library routines even when generating VFP
184 // instructions, so use the same ones.
185 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
186 // e.g., __floatunsidf vs. __floatunssidfvfp.
187 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
188 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
189 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
190 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
194 // These libcalls are not available in 32-bit.
195 setLibcallName(RTLIB::SHL_I128, 0);
196 setLibcallName(RTLIB::SRL_I128, 0);
197 setLibcallName(RTLIB::SRA_I128, 0);
199 if (Subtarget->isThumb1Only())
200 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
202 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
203 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
204 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
205 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
207 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
210 if (Subtarget->hasNEON()) {
211 addDRTypeForNEON(MVT::v2f32);
212 addDRTypeForNEON(MVT::v8i8);
213 addDRTypeForNEON(MVT::v4i16);
214 addDRTypeForNEON(MVT::v2i32);
215 addDRTypeForNEON(MVT::v1i64);
217 addQRTypeForNEON(MVT::v4f32);
218 addQRTypeForNEON(MVT::v2f64);
219 addQRTypeForNEON(MVT::v16i8);
220 addQRTypeForNEON(MVT::v8i16);
221 addQRTypeForNEON(MVT::v4i32);
222 addQRTypeForNEON(MVT::v2i64);
224 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
225 setTargetDAGCombine(ISD::SHL);
226 setTargetDAGCombine(ISD::SRL);
227 setTargetDAGCombine(ISD::SRA);
228 setTargetDAGCombine(ISD::SIGN_EXTEND);
229 setTargetDAGCombine(ISD::ZERO_EXTEND);
230 setTargetDAGCombine(ISD::ANY_EXTEND);
233 computeRegisterProperties();
235 // ARM does not have f32 extending load.
236 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
238 // ARM does not have i1 sign extending load.
239 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
241 // ARM supports all 4 flavors of integer indexed load / store.
242 if (!Subtarget->isThumb1Only()) {
243 for (unsigned im = (unsigned)ISD::PRE_INC;
244 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
245 setIndexedLoadAction(im, MVT::i1, Legal);
246 setIndexedLoadAction(im, MVT::i8, Legal);
247 setIndexedLoadAction(im, MVT::i16, Legal);
248 setIndexedLoadAction(im, MVT::i32, Legal);
249 setIndexedStoreAction(im, MVT::i1, Legal);
250 setIndexedStoreAction(im, MVT::i8, Legal);
251 setIndexedStoreAction(im, MVT::i16, Legal);
252 setIndexedStoreAction(im, MVT::i32, Legal);
256 // i64 operation support.
257 if (Subtarget->isThumb1Only()) {
258 setOperationAction(ISD::MUL, MVT::i64, Expand);
259 setOperationAction(ISD::MULHU, MVT::i32, Expand);
260 setOperationAction(ISD::MULHS, MVT::i32, Expand);
261 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
262 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
264 setOperationAction(ISD::MUL, MVT::i64, Expand);
265 setOperationAction(ISD::MULHU, MVT::i32, Expand);
266 if (!Subtarget->hasV6Ops())
267 setOperationAction(ISD::MULHS, MVT::i32, Expand);
269 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
270 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
271 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
272 setOperationAction(ISD::SRL, MVT::i64, Custom);
273 setOperationAction(ISD::SRA, MVT::i64, Custom);
275 // ARM does not have ROTL.
276 setOperationAction(ISD::ROTL, MVT::i32, Expand);
277 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
278 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
279 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
280 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
282 // Only ARMv6 has BSWAP.
283 if (!Subtarget->hasV6Ops())
284 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
286 // These are expanded into libcalls.
287 setOperationAction(ISD::SDIV, MVT::i32, Expand);
288 setOperationAction(ISD::UDIV, MVT::i32, Expand);
289 setOperationAction(ISD::SREM, MVT::i32, Expand);
290 setOperationAction(ISD::UREM, MVT::i32, Expand);
291 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
292 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
294 // Support label based line numbers.
295 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
296 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
298 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
299 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
300 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
301 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
303 // Use the default implementation.
304 setOperationAction(ISD::VASTART, MVT::Other, Custom);
305 setOperationAction(ISD::VAARG, MVT::Other, Expand);
306 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
307 setOperationAction(ISD::VAEND, MVT::Other, Expand);
308 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
309 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
310 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
311 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
313 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
314 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
315 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
317 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
319 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
320 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
321 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
323 // We want to custom lower some of our intrinsics.
324 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
325 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
326 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
328 setOperationAction(ISD::SETCC, MVT::i32, Expand);
329 setOperationAction(ISD::SETCC, MVT::f32, Expand);
330 setOperationAction(ISD::SETCC, MVT::f64, Expand);
331 setOperationAction(ISD::SELECT, MVT::i32, Expand);
332 setOperationAction(ISD::SELECT, MVT::f32, Expand);
333 setOperationAction(ISD::SELECT, MVT::f64, Expand);
334 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
335 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
336 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
338 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
339 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
340 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
341 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
342 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
344 // We don't support sin/cos/fmod/copysign/pow
345 setOperationAction(ISD::FSIN, MVT::f64, Expand);
346 setOperationAction(ISD::FSIN, MVT::f32, Expand);
347 setOperationAction(ISD::FCOS, MVT::f32, Expand);
348 setOperationAction(ISD::FCOS, MVT::f64, Expand);
349 setOperationAction(ISD::FREM, MVT::f64, Expand);
350 setOperationAction(ISD::FREM, MVT::f32, Expand);
351 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
352 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
353 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
355 setOperationAction(ISD::FPOW, MVT::f64, Expand);
356 setOperationAction(ISD::FPOW, MVT::f32, Expand);
358 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
359 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
360 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
361 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
362 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
363 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
366 // We have target-specific dag combine patterns for the following nodes:
367 // ARMISD::FMRRD - No need to call setTargetDAGCombine
368 setTargetDAGCombine(ISD::ADD);
369 setTargetDAGCombine(ISD::SUB);
371 setStackPointerRegisterToSaveRestore(ARM::SP);
372 setSchedulingPreference(SchedulingForRegPressure);
373 setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10);
374 setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2);
376 if (!Subtarget->isThumb()) {
377 // Use branch latency information to determine if-conversion limits.
378 // FIXME: If-converter should use instruction latency of the branch being
379 // eliminated to compute the threshold. For ARMv6, the branch "latency"
380 // varies depending on whether it's dynamically or statically predicted
381 // and on whether the destination is in the prefetch buffer.
382 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
383 const InstrItineraryData &InstrItins = Subtarget->getInstrItineraryData();
384 unsigned Latency= InstrItins.getLatency(TII->get(ARM::Bcc).getSchedClass());
386 setIfCvtBlockSizeLimit(Latency-1);
388 setIfCvtDupBlockSizeLimit(Latency-2);
390 setIfCvtBlockSizeLimit(10);
391 setIfCvtDupBlockSizeLimit(2);
395 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
396 // Do not enable CodePlacementOpt for now: it currently runs after the
397 // ARMConstantIslandPass and messes up branch relaxation and placement
398 // of constant islands.
399 // benefitFromCodePlacementOpt = true;
402 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
405 case ARMISD::Wrapper: return "ARMISD::Wrapper";
406 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
407 case ARMISD::CALL: return "ARMISD::CALL";
408 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
409 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
410 case ARMISD::tCALL: return "ARMISD::tCALL";
411 case ARMISD::BRCOND: return "ARMISD::BRCOND";
412 case ARMISD::BR_JT: return "ARMISD::BR_JT";
413 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
414 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
415 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
416 case ARMISD::CMP: return "ARMISD::CMP";
417 case ARMISD::CMPZ: return "ARMISD::CMPZ";
418 case ARMISD::CMPFP: return "ARMISD::CMPFP";
419 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
420 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
421 case ARMISD::CMOV: return "ARMISD::CMOV";
422 case ARMISD::CNEG: return "ARMISD::CNEG";
424 case ARMISD::FTOSI: return "ARMISD::FTOSI";
425 case ARMISD::FTOUI: return "ARMISD::FTOUI";
426 case ARMISD::SITOF: return "ARMISD::SITOF";
427 case ARMISD::UITOF: return "ARMISD::UITOF";
429 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
430 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
431 case ARMISD::RRX: return "ARMISD::RRX";
433 case ARMISD::FMRRD: return "ARMISD::FMRRD";
434 case ARMISD::FMDRR: return "ARMISD::FMDRR";
436 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
438 case ARMISD::VCEQ: return "ARMISD::VCEQ";
439 case ARMISD::VCGE: return "ARMISD::VCGE";
440 case ARMISD::VCGEU: return "ARMISD::VCGEU";
441 case ARMISD::VCGT: return "ARMISD::VCGT";
442 case ARMISD::VCGTU: return "ARMISD::VCGTU";
443 case ARMISD::VTST: return "ARMISD::VTST";
445 case ARMISD::VSHL: return "ARMISD::VSHL";
446 case ARMISD::VSHRs: return "ARMISD::VSHRs";
447 case ARMISD::VSHRu: return "ARMISD::VSHRu";
448 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
449 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
450 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
451 case ARMISD::VSHRN: return "ARMISD::VSHRN";
452 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
453 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
454 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
455 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
456 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
457 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
458 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
459 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
460 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
461 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
462 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
463 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
464 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
465 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
466 case ARMISD::VDUPLANEQ: return "ARMISD::VDUPLANEQ";
467 case ARMISD::VLD2D: return "ARMISD::VLD2D";
468 case ARMISD::VLD3D: return "ARMISD::VLD3D";
469 case ARMISD::VLD4D: return "ARMISD::VLD4D";
470 case ARMISD::VST2D: return "ARMISD::VST2D";
471 case ARMISD::VST3D: return "ARMISD::VST3D";
472 case ARMISD::VST4D: return "ARMISD::VST4D";
476 /// getFunctionAlignment - Return the Log2 alignment of this function.
477 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
478 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
481 //===----------------------------------------------------------------------===//
483 //===----------------------------------------------------------------------===//
485 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
486 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
488 default: llvm_unreachable("Unknown condition code!");
489 case ISD::SETNE: return ARMCC::NE;
490 case ISD::SETEQ: return ARMCC::EQ;
491 case ISD::SETGT: return ARMCC::GT;
492 case ISD::SETGE: return ARMCC::GE;
493 case ISD::SETLT: return ARMCC::LT;
494 case ISD::SETLE: return ARMCC::LE;
495 case ISD::SETUGT: return ARMCC::HI;
496 case ISD::SETUGE: return ARMCC::HS;
497 case ISD::SETULT: return ARMCC::LO;
498 case ISD::SETULE: return ARMCC::LS;
502 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
503 /// returns true if the operands should be inverted to form the proper
505 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
506 ARMCC::CondCodes &CondCode2) {
508 CondCode2 = ARMCC::AL;
510 default: llvm_unreachable("Unknown FP condition!");
512 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
514 case ISD::SETOGT: CondCode = ARMCC::GT; break;
516 case ISD::SETOGE: CondCode = ARMCC::GE; break;
517 case ISD::SETOLT: CondCode = ARMCC::MI; break;
518 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
519 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
520 case ISD::SETO: CondCode = ARMCC::VC; break;
521 case ISD::SETUO: CondCode = ARMCC::VS; break;
522 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
523 case ISD::SETUGT: CondCode = ARMCC::HI; break;
524 case ISD::SETUGE: CondCode = ARMCC::PL; break;
526 case ISD::SETULT: CondCode = ARMCC::LT; break;
528 case ISD::SETULE: CondCode = ARMCC::LE; break;
530 case ISD::SETUNE: CondCode = ARMCC::NE; break;
535 //===----------------------------------------------------------------------===//
536 // Calling Convention Implementation
537 //===----------------------------------------------------------------------===//
539 #include "ARMGenCallingConv.inc"
541 // APCS f64 is in register pairs, possibly split to stack
542 static bool f64AssignAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
543 CCValAssign::LocInfo &LocInfo,
544 CCState &State, bool CanFail) {
545 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
547 // Try to get the first register.
548 if (unsigned Reg = State.AllocateReg(RegList, 4))
549 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
551 // For the 2nd half of a v2f64, do not fail.
555 // Put the whole thing on the stack.
556 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
557 State.AllocateStack(8, 4),
562 // Try to get the second register.
563 if (unsigned Reg = State.AllocateReg(RegList, 4))
564 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
566 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
567 State.AllocateStack(4, 4),
572 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
573 CCValAssign::LocInfo &LocInfo,
574 ISD::ArgFlagsTy &ArgFlags,
576 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
578 if (LocVT == MVT::v2f64 &&
579 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
581 return true; // we handled it
584 // AAPCS f64 is in aligned register pairs
585 static bool f64AssignAAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
586 CCValAssign::LocInfo &LocInfo,
587 CCState &State, bool CanFail) {
588 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
589 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
591 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
593 // For the 2nd half of a v2f64, do not just fail.
597 // Put the whole thing on the stack.
598 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
599 State.AllocateStack(8, 8),
605 for (i = 0; i < 2; ++i)
606 if (HiRegList[i] == Reg)
609 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
610 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
615 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
616 CCValAssign::LocInfo &LocInfo,
617 ISD::ArgFlagsTy &ArgFlags,
619 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
621 if (LocVT == MVT::v2f64 &&
622 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
624 return true; // we handled it
627 static bool f64RetAssign(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
628 CCValAssign::LocInfo &LocInfo, CCState &State) {
629 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
630 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
632 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
634 return false; // we didn't handle it
637 for (i = 0; i < 2; ++i)
638 if (HiRegList[i] == Reg)
641 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
642 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
647 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
648 CCValAssign::LocInfo &LocInfo,
649 ISD::ArgFlagsTy &ArgFlags,
651 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
653 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
655 return true; // we handled it
658 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
659 CCValAssign::LocInfo &LocInfo,
660 ISD::ArgFlagsTy &ArgFlags,
662 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
666 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
667 /// given CallingConvention value.
668 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC,
670 bool isVarArg) const {
673 llvm_unreachable("Unsupported calling convention");
675 case CallingConv::Fast:
676 // Use target triple & subtarget features to do actual dispatch.
677 if (Subtarget->isAAPCS_ABI()) {
678 if (Subtarget->hasVFP2() &&
679 FloatABIType == FloatABI::Hard && !isVarArg)
680 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
682 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
684 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
685 case CallingConv::ARM_AAPCS_VFP:
686 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
687 case CallingConv::ARM_AAPCS:
688 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
689 case CallingConv::ARM_APCS:
690 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
694 /// LowerCallResult - Lower the result values of a call into the
695 /// appropriate copies out of appropriate physical registers.
697 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
698 unsigned CallConv, bool isVarArg,
699 const SmallVectorImpl<ISD::InputArg> &Ins,
700 DebugLoc dl, SelectionDAG &DAG,
701 SmallVectorImpl<SDValue> &InVals) {
703 // Assign locations to each value returned by this call.
704 SmallVector<CCValAssign, 16> RVLocs;
705 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
706 RVLocs, *DAG.getContext());
707 CCInfo.AnalyzeCallResult(Ins,
708 CCAssignFnForNode(CallConv, /* Return*/ true,
711 // Copy all of the result registers out of their specified physreg.
712 for (unsigned i = 0; i != RVLocs.size(); ++i) {
713 CCValAssign VA = RVLocs[i];
716 if (VA.needsCustom()) {
717 // Handle f64 or half of a v2f64.
718 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
720 Chain = Lo.getValue(1);
721 InFlag = Lo.getValue(2);
722 VA = RVLocs[++i]; // skip ahead to next loc
723 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
725 Chain = Hi.getValue(1);
726 InFlag = Hi.getValue(2);
727 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
729 if (VA.getLocVT() == MVT::v2f64) {
730 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
731 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
732 DAG.getConstant(0, MVT::i32));
734 VA = RVLocs[++i]; // skip ahead to next loc
735 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
736 Chain = Lo.getValue(1);
737 InFlag = Lo.getValue(2);
738 VA = RVLocs[++i]; // skip ahead to next loc
739 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
740 Chain = Hi.getValue(1);
741 InFlag = Hi.getValue(2);
742 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
743 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
744 DAG.getConstant(1, MVT::i32));
747 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
749 Chain = Val.getValue(1);
750 InFlag = Val.getValue(2);
753 switch (VA.getLocInfo()) {
754 default: llvm_unreachable("Unknown loc info!");
755 case CCValAssign::Full: break;
756 case CCValAssign::BCvt:
757 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
761 InVals.push_back(Val);
767 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
768 /// by "Src" to address "Dst" of size "Size". Alignment information is
769 /// specified by the specific parameter attribute. The copy will be passed as
770 /// a byval function parameter.
771 /// Sometimes what we are copying is the end of a larger object, the part that
772 /// does not fit in registers.
774 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
775 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
777 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
778 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
779 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
782 /// LowerMemOpCallTo - Store the argument to the stack.
784 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
785 SDValue StackPtr, SDValue Arg,
786 DebugLoc dl, SelectionDAG &DAG,
787 const CCValAssign &VA,
788 ISD::ArgFlagsTy Flags) {
789 unsigned LocMemOffset = VA.getLocMemOffset();
790 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
791 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
792 if (Flags.isByVal()) {
793 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
795 return DAG.getStore(Chain, dl, Arg, PtrOff,
796 PseudoSourceValue::getStack(), LocMemOffset);
799 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
800 SDValue Chain, SDValue &Arg,
801 RegsToPassVector &RegsToPass,
802 CCValAssign &VA, CCValAssign &NextVA,
804 SmallVector<SDValue, 8> &MemOpChains,
805 ISD::ArgFlagsTy Flags) {
807 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
808 DAG.getVTList(MVT::i32, MVT::i32), Arg);
809 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
811 if (NextVA.isRegLoc())
812 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
814 assert(NextVA.isMemLoc());
815 if (StackPtr.getNode() == 0)
816 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
818 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
824 /// LowerCall - Lowering a call into a callseq_start <-
825 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
828 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
829 unsigned CallConv, bool isVarArg,
831 const SmallVectorImpl<ISD::OutputArg> &Outs,
832 const SmallVectorImpl<ISD::InputArg> &Ins,
833 DebugLoc dl, SelectionDAG &DAG,
834 SmallVectorImpl<SDValue> &InVals) {
836 // Analyze operands of the call, assigning locations to each operand.
837 SmallVector<CCValAssign, 16> ArgLocs;
838 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
840 CCInfo.AnalyzeCallOperands(Outs,
841 CCAssignFnForNode(CallConv, /* Return*/ false,
844 // Get a count of how many bytes are to be pushed on the stack.
845 unsigned NumBytes = CCInfo.getNextStackOffset();
847 // Adjust the stack pointer for the new arguments...
848 // These operations are automatically eliminated by the prolog/epilog pass
849 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
851 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
853 RegsToPassVector RegsToPass;
854 SmallVector<SDValue, 8> MemOpChains;
856 // Walk the register/memloc assignments, inserting copies/loads. In the case
857 // of tail call optimization, arguments are handled later.
858 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
861 CCValAssign &VA = ArgLocs[i];
862 SDValue Arg = Outs[realArgIdx].Val;
863 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
865 // Promote the value if needed.
866 switch (VA.getLocInfo()) {
867 default: llvm_unreachable("Unknown loc info!");
868 case CCValAssign::Full: break;
869 case CCValAssign::SExt:
870 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
872 case CCValAssign::ZExt:
873 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
875 case CCValAssign::AExt:
876 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
878 case CCValAssign::BCvt:
879 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
883 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
884 if (VA.needsCustom()) {
885 if (VA.getLocVT() == MVT::v2f64) {
886 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
887 DAG.getConstant(0, MVT::i32));
888 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
889 DAG.getConstant(1, MVT::i32));
891 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
892 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
894 VA = ArgLocs[++i]; // skip ahead to next loc
896 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
897 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
899 assert(VA.isMemLoc());
900 if (StackPtr.getNode() == 0)
901 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
903 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
904 dl, DAG, VA, Flags));
907 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
908 StackPtr, MemOpChains, Flags);
910 } else if (VA.isRegLoc()) {
911 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
913 assert(VA.isMemLoc());
914 if (StackPtr.getNode() == 0)
915 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
917 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
918 dl, DAG, VA, Flags));
922 if (!MemOpChains.empty())
923 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
924 &MemOpChains[0], MemOpChains.size());
926 // Build a sequence of copy-to-reg nodes chained together with token chain
927 // and flag operands which copy the outgoing args into the appropriate regs.
929 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
930 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
931 RegsToPass[i].second, InFlag);
932 InFlag = Chain.getValue(1);
935 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
936 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
937 // node so that legalize doesn't hack it.
938 bool isDirect = false;
939 bool isARMFunc = false;
940 bool isLocalARMFunc = false;
941 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
942 GlobalValue *GV = G->getGlobal();
944 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
945 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
946 getTargetMachine().getRelocationModel() != Reloc::Static;
947 isARMFunc = !Subtarget->isThumb() || isStub;
948 // ARM call to a local ARM function is predicable.
949 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
950 // tBX takes a register source operand.
951 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
952 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
954 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
955 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
956 Callee = DAG.getLoad(getPointerTy(), dl,
957 DAG.getEntryNode(), CPAddr, NULL, 0);
958 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
959 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
960 getPointerTy(), Callee, PICLabel);
962 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
963 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
965 bool isStub = Subtarget->isTargetDarwin() &&
966 getTargetMachine().getRelocationModel() != Reloc::Static;
967 isARMFunc = !Subtarget->isThumb() || isStub;
968 // tBX takes a register source operand.
969 const char *Sym = S->getSymbol();
970 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
971 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
973 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
974 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
975 Callee = DAG.getLoad(getPointerTy(), dl,
976 DAG.getEntryNode(), CPAddr, NULL, 0);
977 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
978 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
979 getPointerTy(), Callee, PICLabel);
981 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
984 // FIXME: handle tail calls differently.
986 if (Subtarget->isThumb()) {
987 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
988 CallOpc = ARMISD::CALL_NOLINK;
990 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
992 CallOpc = (isDirect || Subtarget->hasV5TOps())
993 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
994 : ARMISD::CALL_NOLINK;
996 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
997 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
998 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
999 InFlag = Chain.getValue(1);
1002 std::vector<SDValue> Ops;
1003 Ops.push_back(Chain);
1004 Ops.push_back(Callee);
1006 // Add argument registers to the end of the list so that they are known live
1008 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1009 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1010 RegsToPass[i].second.getValueType()));
1012 if (InFlag.getNode())
1013 Ops.push_back(InFlag);
1014 // Returns a chain and a flag for retval copy to use.
1015 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1016 &Ops[0], Ops.size());
1017 InFlag = Chain.getValue(1);
1019 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1020 DAG.getIntPtrConstant(0, true), InFlag);
1022 InFlag = Chain.getValue(1);
1024 // Handle result values, copying them out of physregs into vregs that we
1026 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1031 ARMTargetLowering::LowerReturn(SDValue Chain,
1032 unsigned CallConv, bool isVarArg,
1033 const SmallVectorImpl<ISD::OutputArg> &Outs,
1034 DebugLoc dl, SelectionDAG &DAG) {
1036 // CCValAssign - represent the assignment of the return value to a location.
1037 SmallVector<CCValAssign, 16> RVLocs;
1039 // CCState - Info about the registers and stack slots.
1040 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1043 // Analyze outgoing return values.
1044 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1047 // If this is the first return lowered for this function, add
1048 // the regs to the liveout set for the function.
1049 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1050 for (unsigned i = 0; i != RVLocs.size(); ++i)
1051 if (RVLocs[i].isRegLoc())
1052 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1057 // Copy the result values into the output registers.
1058 for (unsigned i = 0, realRVLocIdx = 0;
1060 ++i, ++realRVLocIdx) {
1061 CCValAssign &VA = RVLocs[i];
1062 assert(VA.isRegLoc() && "Can only return in registers!");
1064 SDValue Arg = Outs[realRVLocIdx].Val;
1066 switch (VA.getLocInfo()) {
1067 default: llvm_unreachable("Unknown loc info!");
1068 case CCValAssign::Full: break;
1069 case CCValAssign::BCvt:
1070 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1074 if (VA.needsCustom()) {
1075 if (VA.getLocVT() == MVT::v2f64) {
1076 // Extract the first half and return it in two registers.
1077 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1078 DAG.getConstant(0, MVT::i32));
1079 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1080 DAG.getVTList(MVT::i32, MVT::i32), Half);
1082 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1083 Flag = Chain.getValue(1);
1084 VA = RVLocs[++i]; // skip ahead to next loc
1085 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1086 HalfGPRs.getValue(1), Flag);
1087 Flag = Chain.getValue(1);
1088 VA = RVLocs[++i]; // skip ahead to next loc
1090 // Extract the 2nd half and fall through to handle it as an f64 value.
1091 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1092 DAG.getConstant(1, MVT::i32));
1094 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1096 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1097 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1098 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1099 Flag = Chain.getValue(1);
1100 VA = RVLocs[++i]; // skip ahead to next loc
1101 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1104 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1106 // Guarantee that all emitted copies are
1107 // stuck together, avoiding something bad.
1108 Flag = Chain.getValue(1);
1113 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1115 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1120 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1121 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1122 // one of the above mentioned nodes. It has to be wrapped because otherwise
1123 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1124 // be used to form addressing mode. These wrapped nodes will be selected
1126 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1127 MVT PtrVT = Op.getValueType();
1128 // FIXME there is no actual debug info here
1129 DebugLoc dl = Op.getDebugLoc();
1130 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1132 if (CP->isMachineConstantPoolEntry())
1133 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1134 CP->getAlignment());
1136 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1137 CP->getAlignment());
1138 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1141 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1143 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1144 SelectionDAG &DAG) {
1145 DebugLoc dl = GA->getDebugLoc();
1146 MVT PtrVT = getPointerTy();
1147 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1148 ARMConstantPoolValue *CPV =
1149 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1150 PCAdj, "tlsgd", true);
1151 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1152 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1153 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1154 SDValue Chain = Argument.getValue(1);
1156 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1157 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1159 // call __tls_get_addr.
1162 Entry.Node = Argument;
1163 Entry.Ty = (const Type *) Type::Int32Ty;
1164 Args.push_back(Entry);
1165 // FIXME: is there useful debug info available here?
1166 std::pair<SDValue, SDValue> CallResult =
1167 LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false, false, false,
1168 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1169 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1170 return CallResult.first;
1173 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1174 // "local exec" model.
1176 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1177 SelectionDAG &DAG) {
1178 GlobalValue *GV = GA->getGlobal();
1179 DebugLoc dl = GA->getDebugLoc();
1181 SDValue Chain = DAG.getEntryNode();
1182 MVT PtrVT = getPointerTy();
1183 // Get the Thread Pointer
1184 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1186 if (GV->isDeclaration()) {
1187 // initial exec model
1188 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1189 ARMConstantPoolValue *CPV =
1190 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1191 PCAdj, "gottpoff", true);
1192 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1193 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1194 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1195 Chain = Offset.getValue(1);
1197 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1198 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1200 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1203 ARMConstantPoolValue *CPV =
1204 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
1205 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1206 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1207 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1210 // The address of the thread local variable is the add of the thread
1211 // pointer with the offset of the variable.
1212 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1216 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1217 // TODO: implement the "local dynamic" model
1218 assert(Subtarget->isTargetELF() &&
1219 "TLS not implemented for non-ELF targets");
1220 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1221 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1222 // otherwise use the "Local Exec" TLS Model
1223 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1224 return LowerToTLSGeneralDynamicModel(GA, DAG);
1226 return LowerToTLSExecModels(GA, DAG);
1229 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1230 SelectionDAG &DAG) {
1231 MVT PtrVT = getPointerTy();
1232 DebugLoc dl = Op.getDebugLoc();
1233 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1234 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1235 if (RelocM == Reloc::PIC_) {
1236 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1237 ARMConstantPoolValue *CPV =
1238 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
1239 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1240 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1241 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1243 SDValue Chain = Result.getValue(1);
1244 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1245 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1247 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1250 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1251 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1252 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1256 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
1257 /// even in non-static mode.
1258 static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
1259 // If symbol visibility is hidden, the extra load is not needed if
1260 // the symbol is definitely defined in the current translation unit.
1261 bool isDecl = GV->isDeclaration() || GV->hasAvailableExternallyLinkage();
1262 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage()))
1264 return RelocM != Reloc::Static && (isDecl || GV->isWeakForLinker());
1267 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1268 SelectionDAG &DAG) {
1269 MVT PtrVT = getPointerTy();
1270 DebugLoc dl = Op.getDebugLoc();
1271 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1272 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1273 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
1275 if (RelocM == Reloc::Static)
1276 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1278 unsigned PCAdj = (RelocM != Reloc::PIC_)
1279 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1280 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
1282 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
1284 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1286 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1288 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1289 SDValue Chain = Result.getValue(1);
1291 if (RelocM == Reloc::PIC_) {
1292 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1293 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1296 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1301 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1303 assert(Subtarget->isTargetELF() &&
1304 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1305 MVT PtrVT = getPointerTy();
1306 DebugLoc dl = Op.getDebugLoc();
1307 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1308 ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_",
1310 ARMCP::CPValue, PCAdj);
1311 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1312 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1313 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1314 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1315 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1318 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
1320 SDNode *Node = Op.getNode();
1321 MVT VT = Node->getValueType(0);
1322 DebugLoc dl = Op.getDebugLoc();
1324 if (!VT.is64BitVector())
1325 return SDValue(); // unimplemented
1327 SDValue Ops[] = { Node->getOperand(0),
1328 Node->getOperand(2) };
1329 return DAG.getNode(Opcode, dl, Node->getVTList(), Ops, 2);
1332 static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG,
1333 unsigned Opcode, unsigned NumVecs) {
1334 SDNode *Node = Op.getNode();
1335 MVT VT = Node->getOperand(3).getValueType();
1336 DebugLoc dl = Op.getDebugLoc();
1338 if (!VT.is64BitVector())
1339 return SDValue(); // unimplemented
1341 SmallVector<SDValue, 6> Ops;
1342 Ops.push_back(Node->getOperand(0));
1343 Ops.push_back(Node->getOperand(2));
1344 for (unsigned N = 0; N < NumVecs; ++N)
1345 Ops.push_back(Node->getOperand(N + 3));
1346 return DAG.getNode(Opcode, dl, MVT::Other, Ops.data(), Ops.size());
1350 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1351 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1353 case Intrinsic::arm_neon_vld2i:
1354 case Intrinsic::arm_neon_vld2f:
1355 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD2D);
1356 case Intrinsic::arm_neon_vld3i:
1357 case Intrinsic::arm_neon_vld3f:
1358 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD3D);
1359 case Intrinsic::arm_neon_vld4i:
1360 case Intrinsic::arm_neon_vld4f:
1361 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD4D);
1362 case Intrinsic::arm_neon_vst2i:
1363 case Intrinsic::arm_neon_vst2f:
1364 return LowerNeonVSTIntrinsic(Op, DAG, ARMISD::VST2D, 2);
1365 case Intrinsic::arm_neon_vst3i:
1366 case Intrinsic::arm_neon_vst3f:
1367 return LowerNeonVSTIntrinsic(Op, DAG, ARMISD::VST3D, 3);
1368 case Intrinsic::arm_neon_vst4i:
1369 case Intrinsic::arm_neon_vst4f:
1370 return LowerNeonVSTIntrinsic(Op, DAG, ARMISD::VST4D, 4);
1371 default: return SDValue(); // Don't custom lower most intrinsics.
1376 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1377 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1378 DebugLoc dl = Op.getDebugLoc();
1380 default: return SDValue(); // Don't custom lower most intrinsics.
1381 case Intrinsic::arm_thread_pointer: {
1382 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1383 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1385 case Intrinsic::eh_sjlj_setjmp:
1386 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1390 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1391 unsigned VarArgsFrameIndex) {
1392 // vastart just stores the address of the VarArgsFrameIndex slot into the
1393 // memory location argument.
1394 DebugLoc dl = Op.getDebugLoc();
1395 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1396 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1397 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1398 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1402 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1403 SDValue &Root, SelectionDAG &DAG,
1405 MachineFunction &MF = DAG.getMachineFunction();
1406 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1408 TargetRegisterClass *RC;
1409 if (AFI->isThumb1OnlyFunction())
1410 RC = ARM::tGPRRegisterClass;
1412 RC = ARM::GPRRegisterClass;
1414 // Transform the arguments stored in physical registers into virtual ones.
1415 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1416 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1419 if (NextVA.isMemLoc()) {
1420 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1421 MachineFrameInfo *MFI = MF.getFrameInfo();
1422 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1424 // Create load node to retrieve arguments from the stack.
1425 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1426 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1428 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1429 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1432 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1436 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1437 unsigned CallConv, bool isVarArg,
1438 const SmallVectorImpl<ISD::InputArg>
1440 DebugLoc dl, SelectionDAG &DAG,
1441 SmallVectorImpl<SDValue> &InVals) {
1443 MachineFunction &MF = DAG.getMachineFunction();
1444 MachineFrameInfo *MFI = MF.getFrameInfo();
1446 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1448 // Assign locations to all of the incoming arguments.
1449 SmallVector<CCValAssign, 16> ArgLocs;
1450 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1452 CCInfo.AnalyzeFormalArguments(Ins,
1453 CCAssignFnForNode(CallConv, /* Return*/ false,
1456 SmallVector<SDValue, 16> ArgValues;
1458 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1459 CCValAssign &VA = ArgLocs[i];
1461 // Arguments stored in registers.
1462 if (VA.isRegLoc()) {
1463 MVT RegVT = VA.getLocVT();
1466 if (VA.needsCustom()) {
1467 // f64 and vector types are split up into multiple registers or
1468 // combinations of registers and stack slots.
1471 if (VA.getLocVT() == MVT::v2f64) {
1472 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1474 VA = ArgLocs[++i]; // skip ahead to next loc
1475 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1477 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1478 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1479 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1480 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1481 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1483 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1486 TargetRegisterClass *RC;
1488 if (RegVT == MVT::f32)
1489 RC = ARM::SPRRegisterClass;
1490 else if (RegVT == MVT::f64)
1491 RC = ARM::DPRRegisterClass;
1492 else if (RegVT == MVT::v2f64)
1493 RC = ARM::QPRRegisterClass;
1494 else if (RegVT == MVT::i32)
1495 RC = (AFI->isThumb1OnlyFunction() ?
1496 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1498 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1500 // Transform the arguments in physical registers into virtual ones.
1501 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1502 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1505 // If this is an 8 or 16-bit value, it is really passed promoted
1506 // to 32 bits. Insert an assert[sz]ext to capture this, then
1507 // truncate to the right size.
1508 switch (VA.getLocInfo()) {
1509 default: llvm_unreachable("Unknown loc info!");
1510 case CCValAssign::Full: break;
1511 case CCValAssign::BCvt:
1512 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1514 case CCValAssign::SExt:
1515 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1516 DAG.getValueType(VA.getValVT()));
1517 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1519 case CCValAssign::ZExt:
1520 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1521 DAG.getValueType(VA.getValVT()));
1522 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1526 InVals.push_back(ArgValue);
1528 } else { // VA.isRegLoc()
1531 assert(VA.isMemLoc());
1532 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1534 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1535 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1537 // Create load nodes to retrieve arguments from the stack.
1538 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1539 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1545 static const unsigned GPRArgRegs[] = {
1546 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1549 unsigned NumGPRs = CCInfo.getFirstUnallocated
1550 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1552 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1553 unsigned VARegSize = (4 - NumGPRs) * 4;
1554 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1555 unsigned ArgOffset = 0;
1556 if (VARegSaveSize) {
1557 // If this function is vararg, store any remaining integer argument regs
1558 // to their spots on the stack so that they may be loaded by deferencing
1559 // the result of va_next.
1560 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1561 ArgOffset = CCInfo.getNextStackOffset();
1562 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1563 VARegSaveSize - VARegSize);
1564 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1566 SmallVector<SDValue, 4> MemOps;
1567 for (; NumGPRs < 4; ++NumGPRs) {
1568 TargetRegisterClass *RC;
1569 if (AFI->isThumb1OnlyFunction())
1570 RC = ARM::tGPRRegisterClass;
1572 RC = ARM::GPRRegisterClass;
1574 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1575 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1576 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1577 MemOps.push_back(Store);
1578 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1579 DAG.getConstant(4, getPointerTy()));
1581 if (!MemOps.empty())
1582 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1583 &MemOps[0], MemOps.size());
1585 // This will point to the next argument passed via stack.
1586 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1592 /// isFloatingPointZero - Return true if this is +0.0.
1593 static bool isFloatingPointZero(SDValue Op) {
1594 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1595 return CFP->getValueAPF().isPosZero();
1596 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1597 // Maybe this has already been legalized into the constant pool?
1598 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1599 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1600 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1601 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1602 return CFP->getValueAPF().isPosZero();
1608 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1609 return ( isThumb1Only && (C & ~255U) == 0) ||
1610 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1613 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1614 /// the given operands.
1615 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1616 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1618 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1619 unsigned C = RHSC->getZExtValue();
1620 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1621 // Constant does not fit, try adjusting it by one?
1626 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1627 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1628 RHS = DAG.getConstant(C-1, MVT::i32);
1633 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1634 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1635 RHS = DAG.getConstant(C-1, MVT::i32);
1640 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1641 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1642 RHS = DAG.getConstant(C+1, MVT::i32);
1647 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1648 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1649 RHS = DAG.getConstant(C+1, MVT::i32);
1656 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1657 ARMISD::NodeType CompareType;
1660 CompareType = ARMISD::CMP;
1665 CompareType = ARMISD::CMPZ;
1668 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1669 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1672 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1673 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1676 if (!isFloatingPointZero(RHS))
1677 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1679 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1680 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1683 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1684 const ARMSubtarget *ST) {
1685 MVT VT = Op.getValueType();
1686 SDValue LHS = Op.getOperand(0);
1687 SDValue RHS = Op.getOperand(1);
1688 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1689 SDValue TrueVal = Op.getOperand(2);
1690 SDValue FalseVal = Op.getOperand(3);
1691 DebugLoc dl = Op.getDebugLoc();
1693 if (LHS.getValueType() == MVT::i32) {
1695 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1696 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1697 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1700 ARMCC::CondCodes CondCode, CondCode2;
1701 if (FPCCToARMCC(CC, CondCode, CondCode2))
1702 std::swap(TrueVal, FalseVal);
1704 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1705 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1706 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1707 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1709 if (CondCode2 != ARMCC::AL) {
1710 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1711 // FIXME: Needs another CMP because flag can have but one use.
1712 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1713 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1714 Result, TrueVal, ARMCC2, CCR, Cmp2);
1719 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1720 const ARMSubtarget *ST) {
1721 SDValue Chain = Op.getOperand(0);
1722 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1723 SDValue LHS = Op.getOperand(2);
1724 SDValue RHS = Op.getOperand(3);
1725 SDValue Dest = Op.getOperand(4);
1726 DebugLoc dl = Op.getDebugLoc();
1728 if (LHS.getValueType() == MVT::i32) {
1730 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1731 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1732 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1733 Chain, Dest, ARMCC, CCR,Cmp);
1736 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1737 ARMCC::CondCodes CondCode, CondCode2;
1738 if (FPCCToARMCC(CC, CondCode, CondCode2))
1739 // Swap the LHS/RHS of the comparison if needed.
1740 std::swap(LHS, RHS);
1742 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1743 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1744 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1745 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1746 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1747 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1748 if (CondCode2 != ARMCC::AL) {
1749 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1750 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1751 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1756 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1757 SDValue Chain = Op.getOperand(0);
1758 SDValue Table = Op.getOperand(1);
1759 SDValue Index = Op.getOperand(2);
1760 DebugLoc dl = Op.getDebugLoc();
1762 MVT PTy = getPointerTy();
1763 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1764 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1765 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1766 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1767 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1768 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1769 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1770 if (Subtarget->isThumb2()) {
1771 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1772 // which does another jump to the destination. This also makes it easier
1773 // to translate it to TBB / TBH later.
1774 // FIXME: This might not work if the function is extremely large.
1775 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1776 Addr, Op.getOperand(2), JTI, UId);
1778 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1779 Addr = DAG.getLoad((MVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1780 Chain = Addr.getValue(1);
1781 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1782 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1784 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1785 Chain = Addr.getValue(1);
1786 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1790 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1791 DebugLoc dl = Op.getDebugLoc();
1793 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1794 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1795 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1798 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1799 MVT VT = Op.getValueType();
1800 DebugLoc dl = Op.getDebugLoc();
1802 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1804 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1805 return DAG.getNode(Opc, dl, VT, Op);
1808 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1809 // Implement fcopysign with a fabs and a conditional fneg.
1810 SDValue Tmp0 = Op.getOperand(0);
1811 SDValue Tmp1 = Op.getOperand(1);
1812 DebugLoc dl = Op.getDebugLoc();
1813 MVT VT = Op.getValueType();
1814 MVT SrcVT = Tmp1.getValueType();
1815 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1816 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1817 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1818 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1819 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1822 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1823 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1824 MFI->setFrameAddressIsTaken(true);
1825 MVT VT = Op.getValueType();
1826 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1827 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1828 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1829 ? ARM::R7 : ARM::R11;
1830 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1832 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1837 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1839 SDValue Dst, SDValue Src,
1840 SDValue Size, unsigned Align,
1842 const Value *DstSV, uint64_t DstSVOff,
1843 const Value *SrcSV, uint64_t SrcSVOff){
1844 // Do repeated 4-byte loads and stores. To be improved.
1845 // This requires 4-byte alignment.
1846 if ((Align & 3) != 0)
1848 // This requires the copy size to be a constant, preferrably
1849 // within a subtarget-specific limit.
1850 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1853 uint64_t SizeVal = ConstantSize->getZExtValue();
1854 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1857 unsigned BytesLeft = SizeVal & 3;
1858 unsigned NumMemOps = SizeVal >> 2;
1859 unsigned EmittedNumMemOps = 0;
1861 unsigned VTSize = 4;
1863 const unsigned MAX_LOADS_IN_LDM = 6;
1864 SDValue TFOps[MAX_LOADS_IN_LDM];
1865 SDValue Loads[MAX_LOADS_IN_LDM];
1866 uint64_t SrcOff = 0, DstOff = 0;
1868 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1869 // same number of stores. The loads and stores will get combined into
1870 // ldm/stm later on.
1871 while (EmittedNumMemOps < NumMemOps) {
1873 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1874 Loads[i] = DAG.getLoad(VT, dl, Chain,
1875 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1876 DAG.getConstant(SrcOff, MVT::i32)),
1877 SrcSV, SrcSVOff + SrcOff);
1878 TFOps[i] = Loads[i].getValue(1);
1881 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1884 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1885 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1886 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1887 DAG.getConstant(DstOff, MVT::i32)),
1888 DstSV, DstSVOff + DstOff);
1891 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1893 EmittedNumMemOps += i;
1899 // Issue loads / stores for the trailing (1 - 3) bytes.
1900 unsigned BytesLeftSave = BytesLeft;
1903 if (BytesLeft >= 2) {
1911 Loads[i] = DAG.getLoad(VT, dl, Chain,
1912 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1913 DAG.getConstant(SrcOff, MVT::i32)),
1914 SrcSV, SrcSVOff + SrcOff);
1915 TFOps[i] = Loads[i].getValue(1);
1918 BytesLeft -= VTSize;
1920 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1923 BytesLeft = BytesLeftSave;
1925 if (BytesLeft >= 2) {
1933 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1934 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1935 DAG.getConstant(DstOff, MVT::i32)),
1936 DstSV, DstSVOff + DstOff);
1939 BytesLeft -= VTSize;
1941 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1944 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
1945 SDValue Op = N->getOperand(0);
1946 DebugLoc dl = N->getDebugLoc();
1947 if (N->getValueType(0) == MVT::f64) {
1948 // Turn i64->f64 into FMDRR.
1949 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1950 DAG.getConstant(0, MVT::i32));
1951 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1952 DAG.getConstant(1, MVT::i32));
1953 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
1956 // Turn f64->i64 into FMRRD.
1957 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
1958 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
1960 // Merge the pieces into a single i64 value.
1961 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
1964 /// getZeroVector - Returns a vector of specified type with all zero elements.
1966 static SDValue getZeroVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
1967 assert(VT.isVector() && "Expected a vector type");
1969 // Zero vectors are used to represent vector negation and in those cases
1970 // will be implemented with the NEON VNEG instruction. However, VNEG does
1971 // not support i64 elements, so sometimes the zero vectors will need to be
1972 // explicitly constructed. For those cases, and potentially other uses in
1973 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
1974 // to their dest type. This ensures they get CSE'd.
1976 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
1977 if (VT.getSizeInBits() == 64)
1978 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
1980 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
1982 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
1985 /// getOnesVector - Returns a vector of specified type with all bits set.
1987 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
1988 assert(VT.isVector() && "Expected a vector type");
1990 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
1991 // type. This ensures they get CSE'd.
1993 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
1994 if (VT.getSizeInBits() == 64)
1995 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
1997 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
1999 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2002 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2003 const ARMSubtarget *ST) {
2004 MVT VT = N->getValueType(0);
2005 DebugLoc dl = N->getDebugLoc();
2007 // Lower vector shifts on NEON to use VSHL.
2008 if (VT.isVector()) {
2009 assert(ST->hasNEON() && "unexpected vector shift");
2011 // Left shifts translate directly to the vshiftu intrinsic.
2012 if (N->getOpcode() == ISD::SHL)
2013 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2014 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2015 N->getOperand(0), N->getOperand(1));
2017 assert((N->getOpcode() == ISD::SRA ||
2018 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2020 // NEON uses the same intrinsics for both left and right shifts. For
2021 // right shifts, the shift amounts are negative, so negate the vector of
2023 MVT ShiftVT = N->getOperand(1).getValueType();
2024 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2025 getZeroVector(ShiftVT, DAG, dl),
2027 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2028 Intrinsic::arm_neon_vshifts :
2029 Intrinsic::arm_neon_vshiftu);
2030 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2031 DAG.getConstant(vshiftInt, MVT::i32),
2032 N->getOperand(0), NegatedCount);
2035 assert(VT == MVT::i64 &&
2036 (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2037 "Unknown shift to lower!");
2039 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2040 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2041 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2044 // If we are in thumb mode, we don't have RRX.
2045 if (ST->isThumb1Only()) return SDValue();
2047 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2048 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2049 DAG.getConstant(0, MVT::i32));
2050 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2051 DAG.getConstant(1, MVT::i32));
2053 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2054 // captures the result into a carry flag.
2055 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2056 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2058 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2059 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2061 // Merge the pieces into a single i64 value.
2062 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2065 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2066 SDValue TmpOp0, TmpOp1;
2067 bool Invert = false;
2071 SDValue Op0 = Op.getOperand(0);
2072 SDValue Op1 = Op.getOperand(1);
2073 SDValue CC = Op.getOperand(2);
2074 MVT VT = Op.getValueType();
2075 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2076 DebugLoc dl = Op.getDebugLoc();
2078 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2079 switch (SetCCOpcode) {
2080 default: llvm_unreachable("Illegal FP comparison"); break;
2082 case ISD::SETNE: Invert = true; // Fallthrough
2084 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2086 case ISD::SETLT: Swap = true; // Fallthrough
2088 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2090 case ISD::SETLE: Swap = true; // Fallthrough
2092 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2093 case ISD::SETUGE: Swap = true; // Fallthrough
2094 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2095 case ISD::SETUGT: Swap = true; // Fallthrough
2096 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2097 case ISD::SETUEQ: Invert = true; // Fallthrough
2099 // Expand this to (OLT | OGT).
2103 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2104 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2106 case ISD::SETUO: Invert = true; // Fallthrough
2108 // Expand this to (OLT | OGE).
2112 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2113 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2117 // Integer comparisons.
2118 switch (SetCCOpcode) {
2119 default: llvm_unreachable("Illegal integer comparison"); break;
2120 case ISD::SETNE: Invert = true;
2121 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2122 case ISD::SETLT: Swap = true;
2123 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2124 case ISD::SETLE: Swap = true;
2125 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2126 case ISD::SETULT: Swap = true;
2127 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2128 case ISD::SETULE: Swap = true;
2129 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2132 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2133 if (Opc == ARMISD::VCEQ) {
2136 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2138 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2141 // Ignore bitconvert.
2142 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2143 AndOp = AndOp.getOperand(0);
2145 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2147 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2148 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2155 std::swap(Op0, Op1);
2157 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2160 Result = DAG.getNOT(dl, Result, VT);
2165 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2166 /// VMOV instruction, and if so, return the constant being splatted.
2167 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2168 unsigned SplatBitSize, SelectionDAG &DAG) {
2169 switch (SplatBitSize) {
2171 // Any 1-byte value is OK.
2172 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2173 return DAG.getTargetConstant(SplatBits, MVT::i8);
2176 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2177 if ((SplatBits & ~0xff) == 0 ||
2178 (SplatBits & ~0xff00) == 0)
2179 return DAG.getTargetConstant(SplatBits, MVT::i16);
2183 // NEON's 32-bit VMOV supports splat values where:
2184 // * only one byte is nonzero, or
2185 // * the least significant byte is 0xff and the second byte is nonzero, or
2186 // * the least significant 2 bytes are 0xff and the third is nonzero.
2187 if ((SplatBits & ~0xff) == 0 ||
2188 (SplatBits & ~0xff00) == 0 ||
2189 (SplatBits & ~0xff0000) == 0 ||
2190 (SplatBits & ~0xff000000) == 0)
2191 return DAG.getTargetConstant(SplatBits, MVT::i32);
2193 if ((SplatBits & ~0xffff) == 0 &&
2194 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2195 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2197 if ((SplatBits & ~0xffffff) == 0 &&
2198 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2199 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2201 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2202 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2203 // VMOV.I32. A (very) minor optimization would be to replicate the value
2204 // and fall through here to test for a valid 64-bit splat. But, then the
2205 // caller would also need to check and handle the change in size.
2209 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2210 uint64_t BitMask = 0xff;
2212 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2213 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2215 else if ((SplatBits & BitMask) != 0)
2219 return DAG.getTargetConstant(Val, MVT::i64);
2223 llvm_unreachable("unexpected size for isVMOVSplat");
2230 /// getVMOVImm - If this is a build_vector of constants which can be
2231 /// formed by using a VMOV instruction of the specified element size,
2232 /// return the constant being splatted. The ByteSize field indicates the
2233 /// number of bytes of each element [1248].
2234 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2235 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2236 APInt SplatBits, SplatUndef;
2237 unsigned SplatBitSize;
2239 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2240 HasAnyUndefs, ByteSize * 8))
2243 if (SplatBitSize > ByteSize * 8)
2246 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2250 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2251 /// instruction with the specified blocksize. (The order of the elements
2252 /// within each block of the vector is reversed.)
2253 bool ARM::isVREVMask(ShuffleVectorSDNode *N, unsigned BlockSize) {
2254 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2255 "Only possible block sizes for VREV are: 16, 32, 64");
2257 MVT VT = N->getValueType(0);
2258 unsigned NumElts = VT.getVectorNumElements();
2259 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2260 unsigned BlockElts = N->getMaskElt(0) + 1;
2262 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2265 for (unsigned i = 0; i < NumElts; ++i) {
2266 if ((unsigned) N->getMaskElt(i) !=
2267 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2274 static SDValue BuildSplat(SDValue Val, MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2275 // Canonicalize all-zeros and all-ones vectors.
2276 ConstantSDNode *ConstVal = dyn_cast<ConstantSDNode>(Val.getNode());
2277 if (ConstVal->isNullValue())
2278 return getZeroVector(VT, DAG, dl);
2279 if (ConstVal->isAllOnesValue())
2280 return getOnesVector(VT, DAG, dl);
2283 if (VT.is64BitVector()) {
2284 switch (Val.getValueType().getSizeInBits()) {
2285 case 8: CanonicalVT = MVT::v8i8; break;
2286 case 16: CanonicalVT = MVT::v4i16; break;
2287 case 32: CanonicalVT = MVT::v2i32; break;
2288 case 64: CanonicalVT = MVT::v1i64; break;
2289 default: llvm_unreachable("unexpected splat element type"); break;
2292 assert(VT.is128BitVector() && "unknown splat vector size");
2293 switch (Val.getValueType().getSizeInBits()) {
2294 case 8: CanonicalVT = MVT::v16i8; break;
2295 case 16: CanonicalVT = MVT::v8i16; break;
2296 case 32: CanonicalVT = MVT::v4i32; break;
2297 case 64: CanonicalVT = MVT::v2i64; break;
2298 default: llvm_unreachable("unexpected splat element type"); break;
2302 // Build a canonical splat for this value.
2303 SmallVector<SDValue, 8> Ops;
2304 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2305 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2307 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2310 // If this is a case we can't handle, return null and let the default
2311 // expansion code take care of it.
2312 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2313 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
2314 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
2315 DebugLoc dl = Op.getDebugLoc();
2316 MVT VT = Op.getValueType();
2318 APInt SplatBits, SplatUndef;
2319 unsigned SplatBitSize;
2321 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2322 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2323 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2325 return BuildSplat(Val, VT, DAG, dl);
2328 // If there are only 2 elements in a 128-bit vector, insert them into an
2329 // undef vector. This handles the common case for 128-bit vector argument
2330 // passing, where the insertions should be translated to subreg accesses
2331 // with no real instructions.
2332 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2333 SDValue Val = DAG.getUNDEF(VT);
2334 SDValue Op0 = Op.getOperand(0);
2335 SDValue Op1 = Op.getOperand(1);
2336 if (Op0.getOpcode() != ISD::UNDEF)
2337 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2338 DAG.getIntPtrConstant(0));
2339 if (Op1.getOpcode() != ISD::UNDEF)
2340 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2341 DAG.getIntPtrConstant(1));
2348 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2352 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
2356 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2357 MVT VT = Op.getValueType();
2358 DebugLoc dl = Op.getDebugLoc();
2359 assert((VT == MVT::i8 || VT == MVT::i16) &&
2360 "unexpected type for custom-lowering vector extract");
2361 SDValue Vec = Op.getOperand(0);
2362 SDValue Lane = Op.getOperand(1);
2363 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2364 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2365 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2368 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2369 // The only time a CONCAT_VECTORS operation can have legal types is when
2370 // two 64-bit vectors are concatenated to a 128-bit vector.
2371 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2372 "unexpected CONCAT_VECTORS");
2373 DebugLoc dl = Op.getDebugLoc();
2374 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2375 SDValue Op0 = Op.getOperand(0);
2376 SDValue Op1 = Op.getOperand(1);
2377 if (Op0.getOpcode() != ISD::UNDEF)
2378 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2379 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2380 DAG.getIntPtrConstant(0));
2381 if (Op1.getOpcode() != ISD::UNDEF)
2382 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2383 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2384 DAG.getIntPtrConstant(1));
2385 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2388 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2389 switch (Op.getOpcode()) {
2390 default: llvm_unreachable("Don't know how to custom lower this!");
2391 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2392 case ISD::GlobalAddress:
2393 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2394 LowerGlobalAddressELF(Op, DAG);
2395 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2396 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2397 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2398 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2399 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2400 case ISD::SINT_TO_FP:
2401 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2402 case ISD::FP_TO_SINT:
2403 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2404 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2405 case ISD::RETURNADDR: break;
2406 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2407 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2408 case ISD::INTRINSIC_VOID:
2409 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2410 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2411 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2414 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2415 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2416 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2417 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2418 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
2419 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2420 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2425 /// ReplaceNodeResults - Replace the results of node with an illegal result
2426 /// type with new values built out of custom code.
2427 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2428 SmallVectorImpl<SDValue>&Results,
2429 SelectionDAG &DAG) {
2430 switch (N->getOpcode()) {
2432 llvm_unreachable("Don't know how to custom expand this!");
2434 case ISD::BIT_CONVERT:
2435 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2439 SDValue Res = LowerShift(N, DAG, Subtarget);
2441 Results.push_back(Res);
2447 //===----------------------------------------------------------------------===//
2448 // ARM Scheduler Hooks
2449 //===----------------------------------------------------------------------===//
2452 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2453 MachineBasicBlock *BB) const {
2454 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2455 DebugLoc dl = MI->getDebugLoc();
2456 switch (MI->getOpcode()) {
2457 default: assert(false && "Unexpected instr type to insert");
2458 case ARM::tMOVCCr: {
2459 // To "insert" a SELECT_CC instruction, we actually have to insert the
2460 // diamond control-flow pattern. The incoming instruction knows the
2461 // destination vreg to set, the condition code register to branch on, the
2462 // true/false values to select between, and a branch opcode to use.
2463 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2464 MachineFunction::iterator It = BB;
2470 // cmpTY ccX, r1, r2
2472 // fallthrough --> copy0MBB
2473 MachineBasicBlock *thisMBB = BB;
2474 MachineFunction *F = BB->getParent();
2475 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2476 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2477 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2478 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2479 F->insert(It, copy0MBB);
2480 F->insert(It, sinkMBB);
2481 // Update machine-CFG edges by first adding all successors of the current
2482 // block to the new block which will contain the Phi node for the select.
2483 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2484 e = BB->succ_end(); i != e; ++i)
2485 sinkMBB->addSuccessor(*i);
2486 // Next, remove all successors of the current block, and add the true
2487 // and fallthrough blocks as its successors.
2488 while(!BB->succ_empty())
2489 BB->removeSuccessor(BB->succ_begin());
2490 BB->addSuccessor(copy0MBB);
2491 BB->addSuccessor(sinkMBB);
2494 // %FalseValue = ...
2495 // # fallthrough to sinkMBB
2498 // Update machine-CFG edges
2499 BB->addSuccessor(sinkMBB);
2502 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2505 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2506 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2507 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2509 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2515 //===----------------------------------------------------------------------===//
2516 // ARM Optimization Hooks
2517 //===----------------------------------------------------------------------===//
2520 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2521 TargetLowering::DAGCombinerInfo &DCI) {
2522 SelectionDAG &DAG = DCI.DAG;
2523 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2524 MVT VT = N->getValueType(0);
2525 unsigned Opc = N->getOpcode();
2526 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2527 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2528 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2529 ISD::CondCode CC = ISD::SETCC_INVALID;
2532 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2534 SDValue CCOp = Slct.getOperand(0);
2535 if (CCOp.getOpcode() == ISD::SETCC)
2536 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2539 bool DoXform = false;
2541 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2544 if (LHS.getOpcode() == ISD::Constant &&
2545 cast<ConstantSDNode>(LHS)->isNullValue()) {
2547 } else if (CC != ISD::SETCC_INVALID &&
2548 RHS.getOpcode() == ISD::Constant &&
2549 cast<ConstantSDNode>(RHS)->isNullValue()) {
2550 std::swap(LHS, RHS);
2551 SDValue Op0 = Slct.getOperand(0);
2552 MVT OpVT = isSlctCC ? Op0.getValueType() :
2553 Op0.getOperand(0).getValueType();
2554 bool isInt = OpVT.isInteger();
2555 CC = ISD::getSetCCInverse(CC, isInt);
2557 if (!TLI.isCondCodeLegal(CC, OpVT))
2558 return SDValue(); // Inverse operator isn't legal.
2565 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
2567 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
2568 Slct.getOperand(0), Slct.getOperand(1), CC);
2569 SDValue CCOp = Slct.getOperand(0);
2571 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
2572 CCOp.getOperand(0), CCOp.getOperand(1), CC);
2573 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
2574 CCOp, OtherOp, Result);
2579 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
2580 static SDValue PerformADDCombine(SDNode *N,
2581 TargetLowering::DAGCombinerInfo &DCI) {
2582 // added by evan in r37685 with no testcase.
2583 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2585 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
2586 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
2587 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
2588 if (Result.getNode()) return Result;
2590 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2591 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2592 if (Result.getNode()) return Result;
2598 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
2599 static SDValue PerformSUBCombine(SDNode *N,
2600 TargetLowering::DAGCombinerInfo &DCI) {
2601 // added by evan in r37685 with no testcase.
2602 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2604 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
2605 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2606 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2607 if (Result.getNode()) return Result;
2614 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
2615 static SDValue PerformFMRRDCombine(SDNode *N,
2616 TargetLowering::DAGCombinerInfo &DCI) {
2617 // fmrrd(fmdrr x, y) -> x,y
2618 SDValue InDouble = N->getOperand(0);
2619 if (InDouble.getOpcode() == ARMISD::FMDRR)
2620 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
2624 /// getVShiftImm - Check if this is a valid build_vector for the immediate
2625 /// operand of a vector shift operation, where all the elements of the
2626 /// build_vector must have the same constant integer value.
2627 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
2628 // Ignore bit_converts.
2629 while (Op.getOpcode() == ISD::BIT_CONVERT)
2630 Op = Op.getOperand(0);
2631 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
2632 APInt SplatBits, SplatUndef;
2633 unsigned SplatBitSize;
2635 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2636 HasAnyUndefs, ElementBits) ||
2637 SplatBitSize > ElementBits)
2639 Cnt = SplatBits.getSExtValue();
2643 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
2644 /// operand of a vector shift left operation. That value must be in the range:
2645 /// 0 <= Value < ElementBits for a left shift; or
2646 /// 0 <= Value <= ElementBits for a long left shift.
2647 static bool isVShiftLImm(SDValue Op, MVT VT, bool isLong, int64_t &Cnt) {
2648 assert(VT.isVector() && "vector shift count is not a vector type");
2649 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
2650 if (! getVShiftImm(Op, ElementBits, Cnt))
2652 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
2655 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
2656 /// operand of a vector shift right operation. For a shift opcode, the value
2657 /// is positive, but for an intrinsic the value count must be negative. The
2658 /// absolute value must be in the range:
2659 /// 1 <= |Value| <= ElementBits for a right shift; or
2660 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
2661 static bool isVShiftRImm(SDValue Op, MVT VT, bool isNarrow, bool isIntrinsic,
2663 assert(VT.isVector() && "vector shift count is not a vector type");
2664 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
2665 if (! getVShiftImm(Op, ElementBits, Cnt))
2669 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
2672 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
2673 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
2674 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2677 // Don't do anything for most intrinsics.
2680 // Vector shifts: check for immediate versions and lower them.
2681 // Note: This is done during DAG combining instead of DAG legalizing because
2682 // the build_vectors for 64-bit vector element shift counts are generally
2683 // not legal, and it is hard to see their values after they get legalized to
2684 // loads from a constant pool.
2685 case Intrinsic::arm_neon_vshifts:
2686 case Intrinsic::arm_neon_vshiftu:
2687 case Intrinsic::arm_neon_vshiftls:
2688 case Intrinsic::arm_neon_vshiftlu:
2689 case Intrinsic::arm_neon_vshiftn:
2690 case Intrinsic::arm_neon_vrshifts:
2691 case Intrinsic::arm_neon_vrshiftu:
2692 case Intrinsic::arm_neon_vrshiftn:
2693 case Intrinsic::arm_neon_vqshifts:
2694 case Intrinsic::arm_neon_vqshiftu:
2695 case Intrinsic::arm_neon_vqshiftsu:
2696 case Intrinsic::arm_neon_vqshiftns:
2697 case Intrinsic::arm_neon_vqshiftnu:
2698 case Intrinsic::arm_neon_vqshiftnsu:
2699 case Intrinsic::arm_neon_vqrshiftns:
2700 case Intrinsic::arm_neon_vqrshiftnu:
2701 case Intrinsic::arm_neon_vqrshiftnsu: {
2702 MVT VT = N->getOperand(1).getValueType();
2704 unsigned VShiftOpc = 0;
2707 case Intrinsic::arm_neon_vshifts:
2708 case Intrinsic::arm_neon_vshiftu:
2709 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
2710 VShiftOpc = ARMISD::VSHL;
2713 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
2714 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
2715 ARMISD::VSHRs : ARMISD::VSHRu);
2720 case Intrinsic::arm_neon_vshiftls:
2721 case Intrinsic::arm_neon_vshiftlu:
2722 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
2724 llvm_unreachable("invalid shift count for vshll intrinsic");
2726 case Intrinsic::arm_neon_vrshifts:
2727 case Intrinsic::arm_neon_vrshiftu:
2728 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
2732 case Intrinsic::arm_neon_vqshifts:
2733 case Intrinsic::arm_neon_vqshiftu:
2734 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
2738 case Intrinsic::arm_neon_vqshiftsu:
2739 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
2741 llvm_unreachable("invalid shift count for vqshlu intrinsic");
2743 case Intrinsic::arm_neon_vshiftn:
2744 case Intrinsic::arm_neon_vrshiftn:
2745 case Intrinsic::arm_neon_vqshiftns:
2746 case Intrinsic::arm_neon_vqshiftnu:
2747 case Intrinsic::arm_neon_vqshiftnsu:
2748 case Intrinsic::arm_neon_vqrshiftns:
2749 case Intrinsic::arm_neon_vqrshiftnu:
2750 case Intrinsic::arm_neon_vqrshiftnsu:
2751 // Narrowing shifts require an immediate right shift.
2752 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
2754 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
2757 llvm_unreachable("unhandled vector shift");
2761 case Intrinsic::arm_neon_vshifts:
2762 case Intrinsic::arm_neon_vshiftu:
2763 // Opcode already set above.
2765 case Intrinsic::arm_neon_vshiftls:
2766 case Intrinsic::arm_neon_vshiftlu:
2767 if (Cnt == VT.getVectorElementType().getSizeInBits())
2768 VShiftOpc = ARMISD::VSHLLi;
2770 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
2771 ARMISD::VSHLLs : ARMISD::VSHLLu);
2773 case Intrinsic::arm_neon_vshiftn:
2774 VShiftOpc = ARMISD::VSHRN; break;
2775 case Intrinsic::arm_neon_vrshifts:
2776 VShiftOpc = ARMISD::VRSHRs; break;
2777 case Intrinsic::arm_neon_vrshiftu:
2778 VShiftOpc = ARMISD::VRSHRu; break;
2779 case Intrinsic::arm_neon_vrshiftn:
2780 VShiftOpc = ARMISD::VRSHRN; break;
2781 case Intrinsic::arm_neon_vqshifts:
2782 VShiftOpc = ARMISD::VQSHLs; break;
2783 case Intrinsic::arm_neon_vqshiftu:
2784 VShiftOpc = ARMISD::VQSHLu; break;
2785 case Intrinsic::arm_neon_vqshiftsu:
2786 VShiftOpc = ARMISD::VQSHLsu; break;
2787 case Intrinsic::arm_neon_vqshiftns:
2788 VShiftOpc = ARMISD::VQSHRNs; break;
2789 case Intrinsic::arm_neon_vqshiftnu:
2790 VShiftOpc = ARMISD::VQSHRNu; break;
2791 case Intrinsic::arm_neon_vqshiftnsu:
2792 VShiftOpc = ARMISD::VQSHRNsu; break;
2793 case Intrinsic::arm_neon_vqrshiftns:
2794 VShiftOpc = ARMISD::VQRSHRNs; break;
2795 case Intrinsic::arm_neon_vqrshiftnu:
2796 VShiftOpc = ARMISD::VQRSHRNu; break;
2797 case Intrinsic::arm_neon_vqrshiftnsu:
2798 VShiftOpc = ARMISD::VQRSHRNsu; break;
2801 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
2802 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
2805 case Intrinsic::arm_neon_vshiftins: {
2806 MVT VT = N->getOperand(1).getValueType();
2808 unsigned VShiftOpc = 0;
2810 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
2811 VShiftOpc = ARMISD::VSLI;
2812 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
2813 VShiftOpc = ARMISD::VSRI;
2815 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
2818 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
2819 N->getOperand(1), N->getOperand(2),
2820 DAG.getConstant(Cnt, MVT::i32));
2823 case Intrinsic::arm_neon_vqrshifts:
2824 case Intrinsic::arm_neon_vqrshiftu:
2825 // No immediate versions of these to check for.
2832 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
2833 /// lowers them. As with the vector shift intrinsics, this is done during DAG
2834 /// combining instead of DAG legalizing because the build_vectors for 64-bit
2835 /// vector element shift counts are generally not legal, and it is hard to see
2836 /// their values after they get legalized to loads from a constant pool.
2837 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
2838 const ARMSubtarget *ST) {
2839 MVT VT = N->getValueType(0);
2841 // Nothing to be done for scalar shifts.
2842 if (! VT.isVector())
2845 assert(ST->hasNEON() && "unexpected vector shift");
2848 switch (N->getOpcode()) {
2849 default: llvm_unreachable("unexpected shift opcode");
2852 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
2853 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
2854 DAG.getConstant(Cnt, MVT::i32));
2859 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
2860 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
2861 ARMISD::VSHRs : ARMISD::VSHRu);
2862 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
2863 DAG.getConstant(Cnt, MVT::i32));
2869 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
2870 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
2871 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
2872 const ARMSubtarget *ST) {
2873 SDValue N0 = N->getOperand(0);
2875 // Check for sign- and zero-extensions of vector extract operations of 8-
2876 // and 16-bit vector elements. NEON supports these directly. They are
2877 // handled during DAG combining because type legalization will promote them
2878 // to 32-bit types and it is messy to recognize the operations after that.
2879 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
2880 SDValue Vec = N0.getOperand(0);
2881 SDValue Lane = N0.getOperand(1);
2882 MVT VT = N->getValueType(0);
2883 MVT EltVT = N0.getValueType();
2884 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2886 if (VT == MVT::i32 &&
2887 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
2888 TLI.isTypeLegal(Vec.getValueType())) {
2891 switch (N->getOpcode()) {
2892 default: llvm_unreachable("unexpected opcode");
2893 case ISD::SIGN_EXTEND:
2894 Opc = ARMISD::VGETLANEs;
2896 case ISD::ZERO_EXTEND:
2897 case ISD::ANY_EXTEND:
2898 Opc = ARMISD::VGETLANEu;
2901 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
2908 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
2909 DAGCombinerInfo &DCI) const {
2910 switch (N->getOpcode()) {
2912 case ISD::ADD: return PerformADDCombine(N, DCI);
2913 case ISD::SUB: return PerformSUBCombine(N, DCI);
2914 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
2915 case ISD::INTRINSIC_WO_CHAIN:
2916 return PerformIntrinsicCombine(N, DCI.DAG);
2920 return PerformShiftCombine(N, DCI.DAG, Subtarget);
2921 case ISD::SIGN_EXTEND:
2922 case ISD::ZERO_EXTEND:
2923 case ISD::ANY_EXTEND:
2924 return PerformExtendCombine(N, DCI.DAG, Subtarget);
2929 /// isLegalAddressImmediate - Return true if the integer value can be used
2930 /// as the offset of the target addressing mode for load / store of the
2932 static bool isLegalAddressImmediate(int64_t V, MVT VT,
2933 const ARMSubtarget *Subtarget) {
2940 if (Subtarget->isThumb()) { // FIXME for thumb2
2945 switch (VT.getSimpleVT()) {
2946 default: return false;
2961 if ((V & (Scale - 1)) != 0)
2964 return V == (V & ((1LL << 5) - 1));
2969 switch (VT.getSimpleVT()) {
2970 default: return false;
2975 return V == (V & ((1LL << 12) - 1));
2978 return V == (V & ((1LL << 8) - 1));
2981 if (!Subtarget->hasVFP2())
2986 return V == (V & ((1LL << 8) - 1));
2990 /// isLegalAddressingMode - Return true if the addressing mode represented
2991 /// by AM is legal for this target, for a load/store of the specified type.
2992 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
2993 const Type *Ty) const {
2994 MVT VT = getValueType(Ty, true);
2995 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
2998 // Can never fold addr of global into load/store.
3003 case 0: // no scale reg, must be "r+i" or "r", or "i".
3006 if (Subtarget->isThumb()) // FIXME for thumb2
3010 // ARM doesn't support any R+R*scale+imm addr modes.
3017 int Scale = AM.Scale;
3018 switch (VT.getSimpleVT()) {
3019 default: return false;
3024 // This assumes i64 is legalized to a pair of i32. If not (i.e.
3025 // ldrd / strd are used, then its address mode is same as i16.
3027 if (Scale < 0) Scale = -Scale;
3031 return isPowerOf2_32(Scale & ~1);
3034 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3039 // Note, we allow "void" uses (basically, uses that aren't loads or
3040 // stores), because arm allows folding a scale into many arithmetic
3041 // operations. This should be made more precise and revisited later.
3043 // Allow r << imm, but the imm has to be a multiple of two.
3044 if (AM.Scale & 1) return false;
3045 return isPowerOf2_32(AM.Scale);
3052 static bool getARMIndexedAddressParts(SDNode *Ptr, MVT VT,
3053 bool isSEXTLoad, SDValue &Base,
3054 SDValue &Offset, bool &isInc,
3055 SelectionDAG &DAG) {
3056 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3059 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3061 Base = Ptr->getOperand(0);
3062 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3063 int RHSC = (int)RHS->getZExtValue();
3064 if (RHSC < 0 && RHSC > -256) {
3065 assert(Ptr->getOpcode() == ISD::ADD);
3067 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3071 isInc = (Ptr->getOpcode() == ISD::ADD);
3072 Offset = Ptr->getOperand(1);
3074 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3076 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3077 int RHSC = (int)RHS->getZExtValue();
3078 if (RHSC < 0 && RHSC > -0x1000) {
3079 assert(Ptr->getOpcode() == ISD::ADD);
3081 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3082 Base = Ptr->getOperand(0);
3087 if (Ptr->getOpcode() == ISD::ADD) {
3089 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3090 if (ShOpcVal != ARM_AM::no_shift) {
3091 Base = Ptr->getOperand(1);
3092 Offset = Ptr->getOperand(0);
3094 Base = Ptr->getOperand(0);
3095 Offset = Ptr->getOperand(1);
3100 isInc = (Ptr->getOpcode() == ISD::ADD);
3101 Base = Ptr->getOperand(0);
3102 Offset = Ptr->getOperand(1);
3106 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3110 static bool getT2IndexedAddressParts(SDNode *Ptr, MVT VT,
3111 bool isSEXTLoad, SDValue &Base,
3112 SDValue &Offset, bool &isInc,
3113 SelectionDAG &DAG) {
3114 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3117 Base = Ptr->getOperand(0);
3118 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3119 int RHSC = (int)RHS->getZExtValue();
3120 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3121 assert(Ptr->getOpcode() == ISD::ADD);
3123 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3125 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3126 isInc = Ptr->getOpcode() == ISD::ADD;
3127 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3135 /// getPreIndexedAddressParts - returns true by value, base pointer and
3136 /// offset pointer and addressing mode by reference if the node's address
3137 /// can be legally represented as pre-indexed load / store address.
3139 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3141 ISD::MemIndexedMode &AM,
3142 SelectionDAG &DAG) const {
3143 if (Subtarget->isThumb1Only())
3148 bool isSEXTLoad = false;
3149 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3150 Ptr = LD->getBasePtr();
3151 VT = LD->getMemoryVT();
3152 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3153 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3154 Ptr = ST->getBasePtr();
3155 VT = ST->getMemoryVT();
3160 bool isLegal = false;
3161 if (Subtarget->isThumb() && Subtarget->hasThumb2())
3162 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3163 Offset, isInc, DAG);
3165 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3166 Offset, isInc, DAG);
3170 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3174 /// getPostIndexedAddressParts - returns true by value, base pointer and
3175 /// offset pointer and addressing mode by reference if this node can be
3176 /// combined with a load / store to form a post-indexed load / store.
3177 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3180 ISD::MemIndexedMode &AM,
3181 SelectionDAG &DAG) const {
3182 if (Subtarget->isThumb1Only())
3187 bool isSEXTLoad = false;
3188 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3189 VT = LD->getMemoryVT();
3190 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3191 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3192 VT = ST->getMemoryVT();
3197 bool isLegal = false;
3198 if (Subtarget->isThumb() && Subtarget->hasThumb2())
3199 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3202 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3207 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3211 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3215 const SelectionDAG &DAG,
3216 unsigned Depth) const {
3217 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3218 switch (Op.getOpcode()) {
3220 case ARMISD::CMOV: {
3221 // Bits are known zero/one if known on the LHS and RHS.
3222 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3223 if (KnownZero == 0 && KnownOne == 0) return;
3225 APInt KnownZeroRHS, KnownOneRHS;
3226 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3227 KnownZeroRHS, KnownOneRHS, Depth+1);
3228 KnownZero &= KnownZeroRHS;
3229 KnownOne &= KnownOneRHS;
3235 //===----------------------------------------------------------------------===//
3236 // ARM Inline Assembly Support
3237 //===----------------------------------------------------------------------===//
3239 /// getConstraintType - Given a constraint letter, return the type of
3240 /// constraint it is for this target.
3241 ARMTargetLowering::ConstraintType
3242 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3243 if (Constraint.size() == 1) {
3244 switch (Constraint[0]) {
3246 case 'l': return C_RegisterClass;
3247 case 'w': return C_RegisterClass;
3250 return TargetLowering::getConstraintType(Constraint);
3253 std::pair<unsigned, const TargetRegisterClass*>
3254 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3256 if (Constraint.size() == 1) {
3257 // GCC RS6000 Constraint Letters
3258 switch (Constraint[0]) {
3260 if (Subtarget->isThumb1Only())
3261 return std::make_pair(0U, ARM::tGPRRegisterClass);
3263 return std::make_pair(0U, ARM::GPRRegisterClass);
3265 return std::make_pair(0U, ARM::GPRRegisterClass);
3268 return std::make_pair(0U, ARM::SPRRegisterClass);
3270 return std::make_pair(0U, ARM::DPRRegisterClass);
3274 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3277 std::vector<unsigned> ARMTargetLowering::
3278 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3280 if (Constraint.size() != 1)
3281 return std::vector<unsigned>();
3283 switch (Constraint[0]) { // GCC ARM Constraint Letters
3286 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3287 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3290 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3291 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3292 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3293 ARM::R12, ARM::LR, 0);
3296 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3297 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3298 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3299 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3300 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3301 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3302 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3303 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3305 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3306 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3307 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3308 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3312 return std::vector<unsigned>();
3315 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3316 /// vector. If it is invalid, don't add anything to Ops.
3317 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3320 std::vector<SDValue>&Ops,
3321 SelectionDAG &DAG) const {
3322 SDValue Result(0, 0);
3324 switch (Constraint) {
3326 case 'I': case 'J': case 'K': case 'L':
3327 case 'M': case 'N': case 'O':
3328 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3332 int64_t CVal64 = C->getSExtValue();
3333 int CVal = (int) CVal64;
3334 // None of these constraints allow values larger than 32 bits. Check
3335 // that the value fits in an int.
3339 switch (Constraint) {
3341 if (Subtarget->isThumb1Only()) {
3342 // This must be a constant between 0 and 255, for ADD
3344 if (CVal >= 0 && CVal <= 255)
3346 } else if (Subtarget->isThumb2()) {
3347 // A constant that can be used as an immediate value in a
3348 // data-processing instruction.
3349 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3352 // A constant that can be used as an immediate value in a
3353 // data-processing instruction.
3354 if (ARM_AM::getSOImmVal(CVal) != -1)
3360 if (Subtarget->isThumb()) { // FIXME thumb2
3361 // This must be a constant between -255 and -1, for negated ADD
3362 // immediates. This can be used in GCC with an "n" modifier that
3363 // prints the negated value, for use with SUB instructions. It is
3364 // not useful otherwise but is implemented for compatibility.
3365 if (CVal >= -255 && CVal <= -1)
3368 // This must be a constant between -4095 and 4095. It is not clear
3369 // what this constraint is intended for. Implemented for
3370 // compatibility with GCC.
3371 if (CVal >= -4095 && CVal <= 4095)
3377 if (Subtarget->isThumb1Only()) {
3378 // A 32-bit value where only one byte has a nonzero value. Exclude
3379 // zero to match GCC. This constraint is used by GCC internally for
3380 // constants that can be loaded with a move/shift combination.
3381 // It is not useful otherwise but is implemented for compatibility.
3382 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3384 } else if (Subtarget->isThumb2()) {
3385 // A constant whose bitwise inverse can be used as an immediate
3386 // value in a data-processing instruction. This can be used in GCC
3387 // with a "B" modifier that prints the inverted value, for use with
3388 // BIC and MVN instructions. It is not useful otherwise but is
3389 // implemented for compatibility.
3390 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3393 // A constant whose bitwise inverse can be used as an immediate
3394 // value in a data-processing instruction. This can be used in GCC
3395 // with a "B" modifier that prints the inverted value, for use with
3396 // BIC and MVN instructions. It is not useful otherwise but is
3397 // implemented for compatibility.
3398 if (ARM_AM::getSOImmVal(~CVal) != -1)
3404 if (Subtarget->isThumb1Only()) {
3405 // This must be a constant between -7 and 7,
3406 // for 3-operand ADD/SUB immediate instructions.
3407 if (CVal >= -7 && CVal < 7)
3409 } else if (Subtarget->isThumb2()) {
3410 // A constant whose negation can be used as an immediate value in a
3411 // data-processing instruction. This can be used in GCC with an "n"
3412 // modifier that prints the negated value, for use with SUB
3413 // instructions. It is not useful otherwise but is implemented for
3415 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3418 // A constant whose negation can be used as an immediate value in a
3419 // data-processing instruction. This can be used in GCC with an "n"
3420 // modifier that prints the negated value, for use with SUB
3421 // instructions. It is not useful otherwise but is implemented for
3423 if (ARM_AM::getSOImmVal(-CVal) != -1)
3429 if (Subtarget->isThumb()) { // FIXME thumb2
3430 // This must be a multiple of 4 between 0 and 1020, for
3431 // ADD sp + immediate.
3432 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3435 // A power of two or a constant between 0 and 32. This is used in
3436 // GCC for the shift amount on shifted register operands, but it is
3437 // useful in general for any shift amounts.
3438 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3444 if (Subtarget->isThumb()) { // FIXME thumb2
3445 // This must be a constant between 0 and 31, for shift amounts.
3446 if (CVal >= 0 && CVal <= 31)
3452 if (Subtarget->isThumb()) { // FIXME thumb2
3453 // This must be a multiple of 4 between -508 and 508, for
3454 // ADD/SUB sp = sp + immediate.
3455 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3460 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3464 if (Result.getNode()) {
3465 Ops.push_back(Result);
3468 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,