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 "llvm/CallingConv.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instruction.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/CodeGen/CallingConvLower.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/Target/TargetLoweringObjectFile.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(TM);
110 return new TargetLoweringObjectFileELF(true);
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->isThumb1Only() && !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::RET, MVT::Other, Custom);
299 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
300 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
301 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
302 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
304 // Use the default implementation.
305 setOperationAction(ISD::VASTART, MVT::Other, Custom);
306 setOperationAction(ISD::VAARG, MVT::Other, Expand);
307 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
308 setOperationAction(ISD::VAEND, MVT::Other, Expand);
309 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
310 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
311 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
312 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
314 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
315 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
318 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
320 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
321 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
322 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
324 // We want to custom lower some of our intrinsics.
325 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
327 setOperationAction(ISD::SETCC, MVT::i32, Expand);
328 setOperationAction(ISD::SETCC, MVT::f32, Expand);
329 setOperationAction(ISD::SETCC, MVT::f64, Expand);
330 setOperationAction(ISD::SELECT, MVT::i32, Expand);
331 setOperationAction(ISD::SELECT, MVT::f32, Expand);
332 setOperationAction(ISD::SELECT, MVT::f64, Expand);
333 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
334 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
335 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
337 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
338 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
339 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
340 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
341 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
343 // We don't support sin/cos/fmod/copysign/pow
344 setOperationAction(ISD::FSIN, MVT::f64, Expand);
345 setOperationAction(ISD::FSIN, MVT::f32, Expand);
346 setOperationAction(ISD::FCOS, MVT::f32, Expand);
347 setOperationAction(ISD::FCOS, MVT::f64, Expand);
348 setOperationAction(ISD::FREM, MVT::f64, Expand);
349 setOperationAction(ISD::FREM, MVT::f32, Expand);
350 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
351 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
352 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
354 setOperationAction(ISD::FPOW, MVT::f64, Expand);
355 setOperationAction(ISD::FPOW, MVT::f32, Expand);
357 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
358 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
359 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
360 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
361 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
362 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
365 // We have target-specific dag combine patterns for the following nodes:
366 // ARMISD::FMRRD - No need to call setTargetDAGCombine
367 setTargetDAGCombine(ISD::ADD);
368 setTargetDAGCombine(ISD::SUB);
370 setStackPointerRegisterToSaveRestore(ARM::SP);
371 setSchedulingPreference(SchedulingForRegPressure);
372 setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10);
373 setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2);
375 if (!Subtarget->isThumb()) {
376 // Use branch latency information to determine if-conversion limits.
377 // FIXME: If-converter should use instruction latency of the branch being
378 // eliminated to compute the threshold. For ARMv6, the branch "latency"
379 // varies depending on whether it's dynamically or statically predicted
380 // and on whether the destination is in the prefetch buffer.
381 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
382 const InstrItineraryData &InstrItins = Subtarget->getInstrItineraryData();
383 unsigned Latency= InstrItins.getLatency(TII->get(ARM::Bcc).getSchedClass());
385 setIfCvtBlockSizeLimit(Latency-1);
387 setIfCvtDupBlockSizeLimit(Latency-2);
389 setIfCvtBlockSizeLimit(10);
390 setIfCvtDupBlockSizeLimit(2);
394 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
395 // Do not enable CodePlacementOpt for now: it currently runs after the
396 // ARMConstantIslandPass and messes up branch relaxation and placement
397 // of constant islands.
398 // benefitFromCodePlacementOpt = true;
401 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
404 case ARMISD::Wrapper: return "ARMISD::Wrapper";
405 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
406 case ARMISD::CALL: return "ARMISD::CALL";
407 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
408 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
409 case ARMISD::tCALL: return "ARMISD::tCALL";
410 case ARMISD::BRCOND: return "ARMISD::BRCOND";
411 case ARMISD::BR_JT: return "ARMISD::BR_JT";
412 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
413 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
414 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
415 case ARMISD::CMP: return "ARMISD::CMP";
416 case ARMISD::CMPZ: return "ARMISD::CMPZ";
417 case ARMISD::CMPFP: return "ARMISD::CMPFP";
418 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
419 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
420 case ARMISD::CMOV: return "ARMISD::CMOV";
421 case ARMISD::CNEG: return "ARMISD::CNEG";
423 case ARMISD::FTOSI: return "ARMISD::FTOSI";
424 case ARMISD::FTOUI: return "ARMISD::FTOUI";
425 case ARMISD::SITOF: return "ARMISD::SITOF";
426 case ARMISD::UITOF: return "ARMISD::UITOF";
428 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
429 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
430 case ARMISD::RRX: return "ARMISD::RRX";
432 case ARMISD::FMRRD: return "ARMISD::FMRRD";
433 case ARMISD::FMDRR: return "ARMISD::FMDRR";
435 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
437 case ARMISD::VCEQ: return "ARMISD::VCEQ";
438 case ARMISD::VCGE: return "ARMISD::VCGE";
439 case ARMISD::VCGEU: return "ARMISD::VCGEU";
440 case ARMISD::VCGT: return "ARMISD::VCGT";
441 case ARMISD::VCGTU: return "ARMISD::VCGTU";
442 case ARMISD::VTST: return "ARMISD::VTST";
444 case ARMISD::VSHL: return "ARMISD::VSHL";
445 case ARMISD::VSHRs: return "ARMISD::VSHRs";
446 case ARMISD::VSHRu: return "ARMISD::VSHRu";
447 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
448 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
449 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
450 case ARMISD::VSHRN: return "ARMISD::VSHRN";
451 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
452 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
453 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
454 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
455 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
456 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
457 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
458 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
459 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
460 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
461 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
462 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
463 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
464 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
465 case ARMISD::VDUPLANEQ: return "ARMISD::VDUPLANEQ";
469 /// getFunctionAlignment - Return the Log2 alignment of this function.
470 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
471 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
474 //===----------------------------------------------------------------------===//
476 //===----------------------------------------------------------------------===//
478 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
479 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
481 default: llvm_unreachable("Unknown condition code!");
482 case ISD::SETNE: return ARMCC::NE;
483 case ISD::SETEQ: return ARMCC::EQ;
484 case ISD::SETGT: return ARMCC::GT;
485 case ISD::SETGE: return ARMCC::GE;
486 case ISD::SETLT: return ARMCC::LT;
487 case ISD::SETLE: return ARMCC::LE;
488 case ISD::SETUGT: return ARMCC::HI;
489 case ISD::SETUGE: return ARMCC::HS;
490 case ISD::SETULT: return ARMCC::LO;
491 case ISD::SETULE: return ARMCC::LS;
495 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
496 /// returns true if the operands should be inverted to form the proper
498 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
499 ARMCC::CondCodes &CondCode2) {
501 CondCode2 = ARMCC::AL;
503 default: llvm_unreachable("Unknown FP condition!");
505 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
507 case ISD::SETOGT: CondCode = ARMCC::GT; break;
509 case ISD::SETOGE: CondCode = ARMCC::GE; break;
510 case ISD::SETOLT: CondCode = ARMCC::MI; break;
511 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
512 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
513 case ISD::SETO: CondCode = ARMCC::VC; break;
514 case ISD::SETUO: CondCode = ARMCC::VS; break;
515 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
516 case ISD::SETUGT: CondCode = ARMCC::HI; break;
517 case ISD::SETUGE: CondCode = ARMCC::PL; break;
519 case ISD::SETULT: CondCode = ARMCC::LT; break;
521 case ISD::SETULE: CondCode = ARMCC::LE; break;
523 case ISD::SETUNE: CondCode = ARMCC::NE; break;
528 //===----------------------------------------------------------------------===//
529 // Calling Convention Implementation
531 // The lower operations present on calling convention works on this order:
532 // LowerCALL (virt regs --> phys regs, virt regs --> stack)
533 // LowerFORMAL_ARGUMENTS (phys --> virt regs, stack --> virt regs)
534 // LowerRET (virt regs --> phys regs)
535 // LowerCALL (phys regs --> virt regs)
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,
672 llvm_unreachable("Unsupported calling convention");
674 case CallingConv::Fast:
675 // Use target triple & subtarget features to do actual dispatch.
676 if (Subtarget->isAAPCS_ABI()) {
677 if (Subtarget->hasVFP2() &&
678 FloatABIType == FloatABI::Hard)
679 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
681 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
683 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
684 case CallingConv::ARM_AAPCS_VFP:
685 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
686 case CallingConv::ARM_AAPCS:
687 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
688 case CallingConv::ARM_APCS:
689 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
693 /// LowerCallResult - Lower the result values of an ISD::CALL into the
694 /// appropriate copies out of appropriate physical registers. This assumes that
695 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
696 /// being lowered. The returns a SDNode with the same number of values as the
698 SDNode *ARMTargetLowering::
699 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
700 unsigned CallingConv, SelectionDAG &DAG) {
702 DebugLoc dl = TheCall->getDebugLoc();
703 // Assign locations to each value returned by this call.
704 SmallVector<CCValAssign, 16> RVLocs;
705 bool isVarArg = TheCall->isVarArg();
706 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(),
707 RVLocs, *DAG.getContext());
708 CCInfo.AnalyzeCallResult(TheCall,
709 CCAssignFnForNode(CallingConv, /* Return*/ true));
711 SmallVector<SDValue, 8> ResultVals;
713 // Copy all of the result registers out of their specified physreg.
714 for (unsigned i = 0; i != RVLocs.size(); ++i) {
715 CCValAssign VA = RVLocs[i];
718 if (VA.needsCustom()) {
719 // Handle f64 or half of a v2f64.
720 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
722 Chain = Lo.getValue(1);
723 InFlag = Lo.getValue(2);
724 VA = RVLocs[++i]; // skip ahead to next loc
725 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
727 Chain = Hi.getValue(1);
728 InFlag = Hi.getValue(2);
729 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
731 if (VA.getLocVT() == MVT::v2f64) {
732 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
733 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
734 DAG.getConstant(0, MVT::i32));
736 VA = RVLocs[++i]; // skip ahead to next loc
737 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
738 Chain = Lo.getValue(1);
739 InFlag = Lo.getValue(2);
740 VA = RVLocs[++i]; // skip ahead to next loc
741 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
742 Chain = Hi.getValue(1);
743 InFlag = Hi.getValue(2);
744 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
745 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
746 DAG.getConstant(1, MVT::i32));
749 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
751 Chain = Val.getValue(1);
752 InFlag = Val.getValue(2);
755 switch (VA.getLocInfo()) {
756 default: llvm_unreachable("Unknown loc info!");
757 case CCValAssign::Full: break;
758 case CCValAssign::BCvt:
759 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
763 ResultVals.push_back(Val);
766 // Merge everything together with a MERGE_VALUES node.
767 ResultVals.push_back(Chain);
768 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
769 &ResultVals[0], ResultVals.size()).getNode();
772 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
773 /// by "Src" to address "Dst" of size "Size". Alignment information is
774 /// specified by the specific parameter attribute. The copy will be passed as
775 /// a byval function parameter.
776 /// Sometimes what we are copying is the end of a larger object, the part that
777 /// does not fit in registers.
779 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
780 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
782 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
783 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
784 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
787 /// LowerMemOpCallTo - Store the argument to the stack.
789 ARMTargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
790 const SDValue &StackPtr,
791 const CCValAssign &VA, SDValue Chain,
792 SDValue Arg, ISD::ArgFlagsTy Flags) {
793 DebugLoc dl = TheCall->getDebugLoc();
794 unsigned LocMemOffset = VA.getLocMemOffset();
795 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
796 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
797 if (Flags.isByVal()) {
798 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
800 return DAG.getStore(Chain, dl, Arg, PtrOff,
801 PseudoSourceValue::getStack(), LocMemOffset);
804 void ARMTargetLowering::PassF64ArgInRegs(CallSDNode *TheCall, SelectionDAG &DAG,
805 SDValue Chain, SDValue &Arg,
806 RegsToPassVector &RegsToPass,
807 CCValAssign &VA, CCValAssign &NextVA,
809 SmallVector<SDValue, 8> &MemOpChains,
810 ISD::ArgFlagsTy Flags) {
811 DebugLoc dl = TheCall->getDebugLoc();
813 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
814 DAG.getVTList(MVT::i32, MVT::i32), Arg);
815 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
817 if (NextVA.isRegLoc())
818 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
820 assert(NextVA.isMemLoc());
821 if (StackPtr.getNode() == 0)
822 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
824 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, NextVA,
825 Chain, fmrrd.getValue(1), Flags));
829 /// LowerCALL - Lowering a ISD::CALL node into a callseq_start <-
830 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
832 SDValue ARMTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
833 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
834 MVT RetVT = TheCall->getRetValType(0);
835 SDValue Chain = TheCall->getChain();
836 unsigned CC = TheCall->getCallingConv();
837 bool isVarArg = TheCall->isVarArg();
838 SDValue Callee = TheCall->getCallee();
839 DebugLoc dl = TheCall->getDebugLoc();
841 // Analyze operands of the call, assigning locations to each operand.
842 SmallVector<CCValAssign, 16> ArgLocs;
843 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
844 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC, /* Return*/ false));
846 // Get a count of how many bytes are to be pushed on the stack.
847 unsigned NumBytes = CCInfo.getNextStackOffset();
849 // Adjust the stack pointer for the new arguments...
850 // These operations are automatically eliminated by the prolog/epilog pass
851 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
853 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
855 RegsToPassVector RegsToPass;
856 SmallVector<SDValue, 8> MemOpChains;
858 // Walk the register/memloc assignments, inserting copies/loads. In the case
859 // of tail call optimization, arguments are handled later.
860 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
863 CCValAssign &VA = ArgLocs[i];
864 SDValue Arg = TheCall->getArg(realArgIdx);
865 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(realArgIdx);
867 // Promote the value if needed.
868 switch (VA.getLocInfo()) {
869 default: llvm_unreachable("Unknown loc info!");
870 case CCValAssign::Full: break;
871 case CCValAssign::SExt:
872 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
874 case CCValAssign::ZExt:
875 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
877 case CCValAssign::AExt:
878 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
880 case CCValAssign::BCvt:
881 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
885 // f64 and v2f64 are passed in i32 pairs and must be split into pieces
886 if (VA.needsCustom()) {
887 if (VA.getLocVT() == MVT::v2f64) {
888 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
889 DAG.getConstant(0, MVT::i32));
890 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
891 DAG.getConstant(1, MVT::i32));
893 PassF64ArgInRegs(TheCall, DAG, Chain, Op0, RegsToPass,
894 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
896 VA = ArgLocs[++i]; // skip ahead to next loc
898 PassF64ArgInRegs(TheCall, DAG, Chain, Op1, RegsToPass,
899 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
901 assert(VA.isMemLoc());
902 if (StackPtr.getNode() == 0)
903 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
905 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
909 PassF64ArgInRegs(TheCall, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
910 StackPtr, MemOpChains, Flags);
912 } else if (VA.isRegLoc()) {
913 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
915 assert(VA.isMemLoc());
916 if (StackPtr.getNode() == 0)
917 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
919 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
924 if (!MemOpChains.empty())
925 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
926 &MemOpChains[0], MemOpChains.size());
928 // Build a sequence of copy-to-reg nodes chained together with token chain
929 // and flag operands which copy the outgoing args into the appropriate regs.
931 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
932 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
933 RegsToPass[i].second, InFlag);
934 InFlag = Chain.getValue(1);
937 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
938 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
939 // node so that legalize doesn't hack it.
940 bool isDirect = false;
941 bool isARMFunc = false;
942 bool isLocalARMFunc = false;
943 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
944 GlobalValue *GV = G->getGlobal();
946 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
947 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
948 getTargetMachine().getRelocationModel() != Reloc::Static;
949 isARMFunc = !Subtarget->isThumb() || isStub;
950 // ARM call to a local ARM function is predicable.
951 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
952 // tBX takes a register source operand.
953 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
954 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
956 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
957 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
958 Callee = DAG.getLoad(getPointerTy(), dl,
959 DAG.getEntryNode(), CPAddr, NULL, 0);
960 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
961 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
962 getPointerTy(), Callee, PICLabel);
964 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
965 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
967 bool isStub = Subtarget->isTargetDarwin() &&
968 getTargetMachine().getRelocationModel() != Reloc::Static;
969 isARMFunc = !Subtarget->isThumb() || isStub;
970 // tBX takes a register source operand.
971 const char *Sym = S->getSymbol();
972 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
973 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
975 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
976 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
977 Callee = DAG.getLoad(getPointerTy(), dl,
978 DAG.getEntryNode(), CPAddr, NULL, 0);
979 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
980 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
981 getPointerTy(), Callee, PICLabel);
983 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
986 // FIXME: handle tail calls differently.
988 if (Subtarget->isThumb1Only()) {
989 if (!Subtarget->hasV5TOps() && (!isDirect || isARMFunc))
990 CallOpc = ARMISD::CALL_NOLINK;
992 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
994 CallOpc = (isDirect || Subtarget->hasV5TOps())
995 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
996 : ARMISD::CALL_NOLINK;
998 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
999 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1000 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1001 InFlag = Chain.getValue(1);
1004 std::vector<SDValue> Ops;
1005 Ops.push_back(Chain);
1006 Ops.push_back(Callee);
1008 // Add argument registers to the end of the list so that they are known live
1010 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1011 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1012 RegsToPass[i].second.getValueType()));
1014 if (InFlag.getNode())
1015 Ops.push_back(InFlag);
1016 // Returns a chain and a flag for retval copy to use.
1017 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1018 &Ops[0], Ops.size());
1019 InFlag = Chain.getValue(1);
1021 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1022 DAG.getIntPtrConstant(0, true), InFlag);
1023 if (RetVT != MVT::Other)
1024 InFlag = Chain.getValue(1);
1026 // Handle result values, copying them out of physregs into vregs that we
1028 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1032 SDValue ARMTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
1033 // The chain is always operand #0
1034 SDValue Chain = Op.getOperand(0);
1035 DebugLoc dl = Op.getDebugLoc();
1037 // CCValAssign - represent the assignment of the return value to a location.
1038 SmallVector<CCValAssign, 16> RVLocs;
1039 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1040 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1042 // CCState - Info about the registers and stack slots.
1043 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs, *DAG.getContext());
1045 // Analyze return values of ISD::RET.
1046 CCInfo.AnalyzeReturn(Op.getNode(), CCAssignFnForNode(CC, /* Return */ true));
1048 // If this is the first return lowered for this function, add
1049 // the regs to the liveout set for the function.
1050 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1051 for (unsigned i = 0; i != RVLocs.size(); ++i)
1052 if (RVLocs[i].isRegLoc())
1053 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1058 // Copy the result values into the output registers.
1059 for (unsigned i = 0, realRVLocIdx = 0;
1061 ++i, ++realRVLocIdx) {
1062 CCValAssign &VA = RVLocs[i];
1063 assert(VA.isRegLoc() && "Can only return in registers!");
1065 // ISD::RET => ret chain, (regnum1,val1), ...
1066 // So i*2+1 index only the regnums
1067 SDValue Arg = Op.getOperand(realRVLocIdx*2+1);
1069 switch (VA.getLocInfo()) {
1070 default: llvm_unreachable("Unknown loc info!");
1071 case CCValAssign::Full: break;
1072 case CCValAssign::BCvt:
1073 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1077 if (VA.needsCustom()) {
1078 if (VA.getLocVT() == MVT::v2f64) {
1079 // Extract the first half and return it in two registers.
1080 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1081 DAG.getConstant(0, MVT::i32));
1082 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1083 DAG.getVTList(MVT::i32, MVT::i32), Half);
1085 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1086 Flag = Chain.getValue(1);
1087 VA = RVLocs[++i]; // skip ahead to next loc
1088 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1089 HalfGPRs.getValue(1), Flag);
1090 Flag = Chain.getValue(1);
1091 VA = RVLocs[++i]; // skip ahead to next loc
1093 // Extract the 2nd half and fall through to handle it as an f64 value.
1094 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1095 DAG.getConstant(1, MVT::i32));
1097 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1099 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1100 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1101 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1102 Flag = Chain.getValue(1);
1103 VA = RVLocs[++i]; // skip ahead to next loc
1104 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1107 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1109 // Guarantee that all emitted copies are
1110 // stuck together, avoiding something bad.
1111 Flag = Chain.getValue(1);
1116 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1118 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1123 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1124 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1125 // one of the above mentioned nodes. It has to be wrapped because otherwise
1126 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1127 // be used to form addressing mode. These wrapped nodes will be selected
1129 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1130 MVT PtrVT = Op.getValueType();
1131 // FIXME there is no actual debug info here
1132 DebugLoc dl = Op.getDebugLoc();
1133 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1135 if (CP->isMachineConstantPoolEntry())
1136 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1137 CP->getAlignment());
1139 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1140 CP->getAlignment());
1141 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1144 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1146 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1147 SelectionDAG &DAG) {
1148 DebugLoc dl = GA->getDebugLoc();
1149 MVT PtrVT = getPointerTy();
1150 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1151 ARMConstantPoolValue *CPV =
1152 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1153 PCAdj, "tlsgd", true);
1154 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1155 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1156 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1157 SDValue Chain = Argument.getValue(1);
1159 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1160 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1162 // call __tls_get_addr.
1165 Entry.Node = Argument;
1166 Entry.Ty = (const Type *) Type::Int32Ty;
1167 Args.push_back(Entry);
1168 // FIXME: is there useful debug info available here?
1169 std::pair<SDValue, SDValue> CallResult =
1170 LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false, false, false,
1171 0, CallingConv::C, false,
1172 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1173 return CallResult.first;
1176 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1177 // "local exec" model.
1179 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1180 SelectionDAG &DAG) {
1181 GlobalValue *GV = GA->getGlobal();
1182 DebugLoc dl = GA->getDebugLoc();
1184 SDValue Chain = DAG.getEntryNode();
1185 MVT PtrVT = getPointerTy();
1186 // Get the Thread Pointer
1187 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1189 if (GV->isDeclaration()) {
1190 // initial exec model
1191 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1192 ARMConstantPoolValue *CPV =
1193 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1194 PCAdj, "gottpoff", true);
1195 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1196 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1197 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1198 Chain = Offset.getValue(1);
1200 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1201 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1203 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1206 ARMConstantPoolValue *CPV =
1207 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
1208 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1209 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1210 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1213 // The address of the thread local variable is the add of the thread
1214 // pointer with the offset of the variable.
1215 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1219 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1220 // TODO: implement the "local dynamic" model
1221 assert(Subtarget->isTargetELF() &&
1222 "TLS not implemented for non-ELF targets");
1223 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1224 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1225 // otherwise use the "Local Exec" TLS Model
1226 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1227 return LowerToTLSGeneralDynamicModel(GA, DAG);
1229 return LowerToTLSExecModels(GA, DAG);
1232 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1233 SelectionDAG &DAG) {
1234 MVT PtrVT = getPointerTy();
1235 DebugLoc dl = Op.getDebugLoc();
1236 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1237 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1238 if (RelocM == Reloc::PIC_) {
1239 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1240 ARMConstantPoolValue *CPV =
1241 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
1242 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1243 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1244 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1246 SDValue Chain = Result.getValue(1);
1247 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1248 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1250 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1253 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1254 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1255 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1259 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
1260 /// even in non-static mode.
1261 static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
1262 // If symbol visibility is hidden, the extra load is not needed if
1263 // the symbol is definitely defined in the current translation unit.
1264 bool isDecl = GV->isDeclaration() || GV->hasAvailableExternallyLinkage();
1265 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage()))
1267 return RelocM != Reloc::Static && (isDecl || GV->isWeakForLinker());
1270 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1271 SelectionDAG &DAG) {
1272 MVT PtrVT = getPointerTy();
1273 DebugLoc dl = Op.getDebugLoc();
1274 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1275 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1276 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
1278 if (RelocM == Reloc::Static)
1279 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1281 unsigned PCAdj = (RelocM != Reloc::PIC_)
1282 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1283 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
1285 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
1287 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1289 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1291 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1292 SDValue Chain = Result.getValue(1);
1294 if (RelocM == Reloc::PIC_) {
1295 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1296 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1299 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1304 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1306 assert(Subtarget->isTargetELF() &&
1307 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1308 MVT PtrVT = getPointerTy();
1309 DebugLoc dl = Op.getDebugLoc();
1310 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1311 ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_",
1313 ARMCP::CPValue, PCAdj);
1314 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1315 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1316 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1317 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1318 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1322 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1323 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1324 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1325 DebugLoc dl = Op.getDebugLoc();
1327 default: return SDValue(); // Don't custom lower most intrinsics.
1328 case Intrinsic::arm_thread_pointer:
1329 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1330 case Intrinsic::eh_sjlj_setjmp:
1331 SDValue Res = DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32,
1337 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1338 unsigned VarArgsFrameIndex) {
1339 // vastart just stores the address of the VarArgsFrameIndex slot into the
1340 // memory location argument.
1341 DebugLoc dl = Op.getDebugLoc();
1342 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1343 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1344 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1345 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1349 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1350 SDValue &Root, SelectionDAG &DAG,
1352 MachineFunction &MF = DAG.getMachineFunction();
1353 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1355 TargetRegisterClass *RC;
1356 if (AFI->isThumb1OnlyFunction())
1357 RC = ARM::tGPRRegisterClass;
1359 RC = ARM::GPRRegisterClass;
1361 // Transform the arguments stored in physical registers into virtual ones.
1362 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1363 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1366 if (NextVA.isMemLoc()) {
1367 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1368 MachineFrameInfo *MFI = MF.getFrameInfo();
1369 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1371 // Create load node to retrieve arguments from the stack.
1372 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1373 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1375 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1376 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1379 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1383 ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1384 MachineFunction &MF = DAG.getMachineFunction();
1385 MachineFrameInfo *MFI = MF.getFrameInfo();
1387 SDValue Root = Op.getOperand(0);
1388 DebugLoc dl = Op.getDebugLoc();
1389 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1390 unsigned CC = MF.getFunction()->getCallingConv();
1391 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1393 // Assign locations to all of the incoming arguments.
1394 SmallVector<CCValAssign, 16> ArgLocs;
1395 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
1396 CCInfo.AnalyzeFormalArguments(Op.getNode(),
1397 CCAssignFnForNode(CC, /* Return*/ false));
1399 SmallVector<SDValue, 16> ArgValues;
1401 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1402 CCValAssign &VA = ArgLocs[i];
1404 // Arguments stored in registers.
1405 if (VA.isRegLoc()) {
1406 MVT RegVT = VA.getLocVT();
1409 if (VA.needsCustom()) {
1410 // f64 and vector types are split up into multiple registers or
1411 // combinations of registers and stack slots.
1414 if (VA.getLocVT() == MVT::v2f64) {
1415 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1417 VA = ArgLocs[++i]; // skip ahead to next loc
1418 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1420 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1421 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1422 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1423 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1424 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1426 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Root, DAG, dl);
1429 TargetRegisterClass *RC;
1430 if (FloatABIType == FloatABI::Hard && RegVT == MVT::f32)
1431 RC = ARM::SPRRegisterClass;
1432 else if (FloatABIType == FloatABI::Hard && RegVT == MVT::f64)
1433 RC = ARM::DPRRegisterClass;
1434 else if (AFI->isThumb1OnlyFunction())
1435 RC = ARM::tGPRRegisterClass;
1437 RC = ARM::GPRRegisterClass;
1439 assert((RegVT == MVT::i32 || RegVT == MVT::f32 ||
1440 (FloatABIType == FloatABI::Hard && RegVT == MVT::f64)) &&
1441 "RegVT not supported by FORMAL_ARGUMENTS Lowering");
1443 // Transform the arguments in physical registers into virtual ones.
1444 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1445 ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1448 // If this is an 8 or 16-bit value, it is really passed promoted
1449 // to 32 bits. Insert an assert[sz]ext to capture this, then
1450 // truncate to the right size.
1451 switch (VA.getLocInfo()) {
1452 default: llvm_unreachable("Unknown loc info!");
1453 case CCValAssign::Full: break;
1454 case CCValAssign::BCvt:
1455 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1457 case CCValAssign::SExt:
1458 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1459 DAG.getValueType(VA.getValVT()));
1460 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1462 case CCValAssign::ZExt:
1463 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1464 DAG.getValueType(VA.getValVT()));
1465 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1469 ArgValues.push_back(ArgValue);
1471 } else { // VA.isRegLoc()
1474 assert(VA.isMemLoc());
1475 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1477 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1478 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1480 // Create load nodes to retrieve arguments from the stack.
1481 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1482 ArgValues.push_back(DAG.getLoad(VA.getValVT(), dl, Root, FIN, NULL, 0));
1488 static const unsigned GPRArgRegs[] = {
1489 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1492 unsigned NumGPRs = CCInfo.getFirstUnallocated
1493 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1495 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1496 unsigned VARegSize = (4 - NumGPRs) * 4;
1497 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1498 unsigned ArgOffset = 0;
1499 if (VARegSaveSize) {
1500 // If this function is vararg, store any remaining integer argument regs
1501 // to their spots on the stack so that they may be loaded by deferencing
1502 // the result of va_next.
1503 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1504 ArgOffset = CCInfo.getNextStackOffset();
1505 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1506 VARegSaveSize - VARegSize);
1507 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1509 SmallVector<SDValue, 4> MemOps;
1510 for (; NumGPRs < 4; ++NumGPRs) {
1511 TargetRegisterClass *RC;
1512 if (AFI->isThumb1OnlyFunction())
1513 RC = ARM::tGPRRegisterClass;
1515 RC = ARM::GPRRegisterClass;
1517 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1518 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
1519 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1520 MemOps.push_back(Store);
1521 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1522 DAG.getConstant(4, getPointerTy()));
1524 if (!MemOps.empty())
1525 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1526 &MemOps[0], MemOps.size());
1528 // This will point to the next argument passed via stack.
1529 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1532 ArgValues.push_back(Root);
1534 // Return the new list of results.
1535 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1536 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1539 /// isFloatingPointZero - Return true if this is +0.0.
1540 static bool isFloatingPointZero(SDValue Op) {
1541 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1542 return CFP->getValueAPF().isPosZero();
1543 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1544 // Maybe this has already been legalized into the constant pool?
1545 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1546 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1547 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1548 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1549 return CFP->getValueAPF().isPosZero();
1555 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1556 return ( isThumb1Only && (C & ~255U) == 0) ||
1557 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1560 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1561 /// the given operands.
1562 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1563 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1565 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1566 unsigned C = RHSC->getZExtValue();
1567 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1568 // Constant does not fit, try adjusting it by one?
1573 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1574 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1575 RHS = DAG.getConstant(C-1, MVT::i32);
1580 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1581 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1582 RHS = DAG.getConstant(C-1, MVT::i32);
1587 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1588 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1589 RHS = DAG.getConstant(C+1, MVT::i32);
1594 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1595 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1596 RHS = DAG.getConstant(C+1, MVT::i32);
1603 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1604 ARMISD::NodeType CompareType;
1607 CompareType = ARMISD::CMP;
1612 CompareType = ARMISD::CMPZ;
1615 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1616 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1619 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1620 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1623 if (!isFloatingPointZero(RHS))
1624 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1626 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1627 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1630 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1631 const ARMSubtarget *ST) {
1632 MVT VT = Op.getValueType();
1633 SDValue LHS = Op.getOperand(0);
1634 SDValue RHS = Op.getOperand(1);
1635 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1636 SDValue TrueVal = Op.getOperand(2);
1637 SDValue FalseVal = Op.getOperand(3);
1638 DebugLoc dl = Op.getDebugLoc();
1640 if (LHS.getValueType() == MVT::i32) {
1642 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1643 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1644 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1647 ARMCC::CondCodes CondCode, CondCode2;
1648 if (FPCCToARMCC(CC, CondCode, CondCode2))
1649 std::swap(TrueVal, FalseVal);
1651 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1652 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1653 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1654 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1656 if (CondCode2 != ARMCC::AL) {
1657 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1658 // FIXME: Needs another CMP because flag can have but one use.
1659 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1660 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1661 Result, TrueVal, ARMCC2, CCR, Cmp2);
1666 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1667 const ARMSubtarget *ST) {
1668 SDValue Chain = Op.getOperand(0);
1669 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1670 SDValue LHS = Op.getOperand(2);
1671 SDValue RHS = Op.getOperand(3);
1672 SDValue Dest = Op.getOperand(4);
1673 DebugLoc dl = Op.getDebugLoc();
1675 if (LHS.getValueType() == MVT::i32) {
1677 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1678 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1679 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1680 Chain, Dest, ARMCC, CCR,Cmp);
1683 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1684 ARMCC::CondCodes CondCode, CondCode2;
1685 if (FPCCToARMCC(CC, CondCode, CondCode2))
1686 // Swap the LHS/RHS of the comparison if needed.
1687 std::swap(LHS, RHS);
1689 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1690 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1691 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1692 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1693 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1694 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1695 if (CondCode2 != ARMCC::AL) {
1696 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1697 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1698 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1703 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1704 SDValue Chain = Op.getOperand(0);
1705 SDValue Table = Op.getOperand(1);
1706 SDValue Index = Op.getOperand(2);
1707 DebugLoc dl = Op.getDebugLoc();
1709 MVT PTy = getPointerTy();
1710 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1711 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1712 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1713 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1714 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1715 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1716 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1717 if (Subtarget->isThumb2()) {
1718 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1719 // which does another jump to the destination. This also makes it easier
1720 // to translate it to TBB / TBH later.
1721 // FIXME: This might not work if the function is extremely large.
1722 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1723 Addr, Op.getOperand(2), JTI, UId);
1725 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1726 Addr = DAG.getLoad((MVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1727 Chain = Addr.getValue(1);
1728 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1729 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1731 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1732 Chain = Addr.getValue(1);
1733 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1737 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1738 DebugLoc dl = Op.getDebugLoc();
1740 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1741 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1742 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1745 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1746 MVT VT = Op.getValueType();
1747 DebugLoc dl = Op.getDebugLoc();
1749 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1751 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1752 return DAG.getNode(Opc, dl, VT, Op);
1755 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1756 // Implement fcopysign with a fabs and a conditional fneg.
1757 SDValue Tmp0 = Op.getOperand(0);
1758 SDValue Tmp1 = Op.getOperand(1);
1759 DebugLoc dl = Op.getDebugLoc();
1760 MVT VT = Op.getValueType();
1761 MVT SrcVT = Tmp1.getValueType();
1762 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1763 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1764 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1765 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1766 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1769 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1770 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1771 MFI->setFrameAddressIsTaken(true);
1772 MVT VT = Op.getValueType();
1773 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1774 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1775 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1776 ? ARM::R7 : ARM::R11;
1777 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1779 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1784 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1786 SDValue Dst, SDValue Src,
1787 SDValue Size, unsigned Align,
1789 const Value *DstSV, uint64_t DstSVOff,
1790 const Value *SrcSV, uint64_t SrcSVOff){
1791 // Do repeated 4-byte loads and stores. To be improved.
1792 // This requires 4-byte alignment.
1793 if ((Align & 3) != 0)
1795 // This requires the copy size to be a constant, preferrably
1796 // within a subtarget-specific limit.
1797 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1800 uint64_t SizeVal = ConstantSize->getZExtValue();
1801 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1804 unsigned BytesLeft = SizeVal & 3;
1805 unsigned NumMemOps = SizeVal >> 2;
1806 unsigned EmittedNumMemOps = 0;
1808 unsigned VTSize = 4;
1810 const unsigned MAX_LOADS_IN_LDM = 6;
1811 SDValue TFOps[MAX_LOADS_IN_LDM];
1812 SDValue Loads[MAX_LOADS_IN_LDM];
1813 uint64_t SrcOff = 0, DstOff = 0;
1815 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1816 // same number of stores. The loads and stores will get combined into
1817 // ldm/stm later on.
1818 while (EmittedNumMemOps < NumMemOps) {
1820 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1821 Loads[i] = DAG.getLoad(VT, dl, Chain,
1822 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1823 DAG.getConstant(SrcOff, MVT::i32)),
1824 SrcSV, SrcSVOff + SrcOff);
1825 TFOps[i] = Loads[i].getValue(1);
1828 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1831 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1832 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1833 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1834 DAG.getConstant(DstOff, MVT::i32)),
1835 DstSV, DstSVOff + DstOff);
1838 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1840 EmittedNumMemOps += i;
1846 // Issue loads / stores for the trailing (1 - 3) bytes.
1847 unsigned BytesLeftSave = BytesLeft;
1850 if (BytesLeft >= 2) {
1858 Loads[i] = DAG.getLoad(VT, dl, Chain,
1859 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1860 DAG.getConstant(SrcOff, MVT::i32)),
1861 SrcSV, SrcSVOff + SrcOff);
1862 TFOps[i] = Loads[i].getValue(1);
1865 BytesLeft -= VTSize;
1867 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1870 BytesLeft = BytesLeftSave;
1872 if (BytesLeft >= 2) {
1880 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1881 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1882 DAG.getConstant(DstOff, MVT::i32)),
1883 DstSV, DstSVOff + DstOff);
1886 BytesLeft -= VTSize;
1888 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1891 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
1892 SDValue Op = N->getOperand(0);
1893 DebugLoc dl = N->getDebugLoc();
1894 if (N->getValueType(0) == MVT::f64) {
1895 // Turn i64->f64 into FMDRR.
1896 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1897 DAG.getConstant(0, MVT::i32));
1898 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1899 DAG.getConstant(1, MVT::i32));
1900 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
1903 // Turn f64->i64 into FMRRD.
1904 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
1905 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
1907 // Merge the pieces into a single i64 value.
1908 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
1911 /// getZeroVector - Returns a vector of specified type with all zero elements.
1913 static SDValue getZeroVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
1914 assert(VT.isVector() && "Expected a vector type");
1916 // Zero vectors are used to represent vector negation and in those cases
1917 // will be implemented with the NEON VNEG instruction. However, VNEG does
1918 // not support i64 elements, so sometimes the zero vectors will need to be
1919 // explicitly constructed. For those cases, and potentially other uses in
1920 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
1921 // to their dest type. This ensures they get CSE'd.
1923 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
1924 if (VT.getSizeInBits() == 64)
1925 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
1927 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
1929 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
1932 /// getOnesVector - Returns a vector of specified type with all bits set.
1934 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
1935 assert(VT.isVector() && "Expected a vector type");
1937 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
1938 // type. This ensures they get CSE'd.
1940 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
1941 if (VT.getSizeInBits() == 64)
1942 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
1944 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
1946 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
1949 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
1950 const ARMSubtarget *ST) {
1951 MVT VT = N->getValueType(0);
1952 DebugLoc dl = N->getDebugLoc();
1954 // Lower vector shifts on NEON to use VSHL.
1955 if (VT.isVector()) {
1956 assert(ST->hasNEON() && "unexpected vector shift");
1958 // Left shifts translate directly to the vshiftu intrinsic.
1959 if (N->getOpcode() == ISD::SHL)
1960 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
1961 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
1962 N->getOperand(0), N->getOperand(1));
1964 assert((N->getOpcode() == ISD::SRA ||
1965 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
1967 // NEON uses the same intrinsics for both left and right shifts. For
1968 // right shifts, the shift amounts are negative, so negate the vector of
1970 MVT ShiftVT = N->getOperand(1).getValueType();
1971 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
1972 getZeroVector(ShiftVT, DAG, dl),
1974 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
1975 Intrinsic::arm_neon_vshifts :
1976 Intrinsic::arm_neon_vshiftu);
1977 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
1978 DAG.getConstant(vshiftInt, MVT::i32),
1979 N->getOperand(0), NegatedCount);
1982 assert(VT == MVT::i64 &&
1983 (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
1984 "Unknown shift to lower!");
1986 // We only lower SRA, SRL of 1 here, all others use generic lowering.
1987 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
1988 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
1991 // If we are in thumb mode, we don't have RRX.
1992 if (ST->isThumb1Only()) return SDValue();
1994 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
1995 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
1996 DAG.getConstant(0, MVT::i32));
1997 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
1998 DAG.getConstant(1, MVT::i32));
2000 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2001 // captures the result into a carry flag.
2002 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2003 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2005 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2006 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2008 // Merge the pieces into a single i64 value.
2009 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2012 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2013 SDValue TmpOp0, TmpOp1;
2014 bool Invert = false;
2018 SDValue Op0 = Op.getOperand(0);
2019 SDValue Op1 = Op.getOperand(1);
2020 SDValue CC = Op.getOperand(2);
2021 MVT VT = Op.getValueType();
2022 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2023 DebugLoc dl = Op.getDebugLoc();
2025 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2026 switch (SetCCOpcode) {
2027 default: llvm_unreachable("Illegal FP comparison"); break;
2029 case ISD::SETNE: Invert = true; // Fallthrough
2031 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2033 case ISD::SETLT: Swap = true; // Fallthrough
2035 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2037 case ISD::SETLE: Swap = true; // Fallthrough
2039 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2040 case ISD::SETUGE: Swap = true; // Fallthrough
2041 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2042 case ISD::SETUGT: Swap = true; // Fallthrough
2043 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2044 case ISD::SETUEQ: Invert = true; // Fallthrough
2046 // Expand this to (OLT | OGT).
2050 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2051 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2053 case ISD::SETUO: Invert = true; // Fallthrough
2055 // Expand this to (OLT | OGE).
2059 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2060 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2064 // Integer comparisons.
2065 switch (SetCCOpcode) {
2066 default: llvm_unreachable("Illegal integer comparison"); break;
2067 case ISD::SETNE: Invert = true;
2068 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2069 case ISD::SETLT: Swap = true;
2070 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2071 case ISD::SETLE: Swap = true;
2072 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2073 case ISD::SETULT: Swap = true;
2074 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2075 case ISD::SETULE: Swap = true;
2076 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2079 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2080 if (Opc == ARMISD::VCEQ) {
2083 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2085 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2088 // Ignore bitconvert.
2089 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2090 AndOp = AndOp.getOperand(0);
2092 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2094 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2095 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2102 std::swap(Op0, Op1);
2104 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2107 Result = DAG.getNOT(dl, Result, VT);
2112 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2113 /// VMOV instruction, and if so, return the constant being splatted.
2114 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2115 unsigned SplatBitSize, SelectionDAG &DAG) {
2116 switch (SplatBitSize) {
2118 // Any 1-byte value is OK.
2119 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2120 return DAG.getTargetConstant(SplatBits, MVT::i8);
2123 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2124 if ((SplatBits & ~0xff) == 0 ||
2125 (SplatBits & ~0xff00) == 0)
2126 return DAG.getTargetConstant(SplatBits, MVT::i16);
2130 // NEON's 32-bit VMOV supports splat values where:
2131 // * only one byte is nonzero, or
2132 // * the least significant byte is 0xff and the second byte is nonzero, or
2133 // * the least significant 2 bytes are 0xff and the third is nonzero.
2134 if ((SplatBits & ~0xff) == 0 ||
2135 (SplatBits & ~0xff00) == 0 ||
2136 (SplatBits & ~0xff0000) == 0 ||
2137 (SplatBits & ~0xff000000) == 0)
2138 return DAG.getTargetConstant(SplatBits, MVT::i32);
2140 if ((SplatBits & ~0xffff) == 0 &&
2141 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2142 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2144 if ((SplatBits & ~0xffffff) == 0 &&
2145 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2146 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2148 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2149 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2150 // VMOV.I32. A (very) minor optimization would be to replicate the value
2151 // and fall through here to test for a valid 64-bit splat. But, then the
2152 // caller would also need to check and handle the change in size.
2156 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2157 uint64_t BitMask = 0xff;
2159 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2160 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2162 else if ((SplatBits & BitMask) != 0)
2166 return DAG.getTargetConstant(Val, MVT::i64);
2170 llvm_unreachable("unexpected size for isVMOVSplat");
2177 /// getVMOVImm - If this is a build_vector of constants which can be
2178 /// formed by using a VMOV instruction of the specified element size,
2179 /// return the constant being splatted. The ByteSize field indicates the
2180 /// number of bytes of each element [1248].
2181 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2182 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2183 APInt SplatBits, SplatUndef;
2184 unsigned SplatBitSize;
2186 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2187 HasAnyUndefs, ByteSize * 8))
2190 if (SplatBitSize > ByteSize * 8)
2193 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2197 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2198 /// instruction with the specified blocksize. (The order of the elements
2199 /// within each block of the vector is reversed.)
2200 bool ARM::isVREVMask(ShuffleVectorSDNode *N, unsigned BlockSize) {
2201 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2202 "Only possible block sizes for VREV are: 16, 32, 64");
2204 MVT VT = N->getValueType(0);
2205 unsigned NumElts = VT.getVectorNumElements();
2206 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2207 unsigned BlockElts = N->getMaskElt(0) + 1;
2209 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2212 for (unsigned i = 0; i < NumElts; ++i) {
2213 if ((unsigned) N->getMaskElt(i) !=
2214 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2221 static SDValue BuildSplat(SDValue Val, MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2222 // Canonicalize all-zeros and all-ones vectors.
2223 ConstantSDNode *ConstVal = dyn_cast<ConstantSDNode>(Val.getNode());
2224 if (ConstVal->isNullValue())
2225 return getZeroVector(VT, DAG, dl);
2226 if (ConstVal->isAllOnesValue())
2227 return getOnesVector(VT, DAG, dl);
2230 if (VT.is64BitVector()) {
2231 switch (Val.getValueType().getSizeInBits()) {
2232 case 8: CanonicalVT = MVT::v8i8; break;
2233 case 16: CanonicalVT = MVT::v4i16; break;
2234 case 32: CanonicalVT = MVT::v2i32; break;
2235 case 64: CanonicalVT = MVT::v1i64; break;
2236 default: llvm_unreachable("unexpected splat element type"); break;
2239 assert(VT.is128BitVector() && "unknown splat vector size");
2240 switch (Val.getValueType().getSizeInBits()) {
2241 case 8: CanonicalVT = MVT::v16i8; break;
2242 case 16: CanonicalVT = MVT::v8i16; break;
2243 case 32: CanonicalVT = MVT::v4i32; break;
2244 case 64: CanonicalVT = MVT::v2i64; break;
2245 default: llvm_unreachable("unexpected splat element type"); break;
2249 // Build a canonical splat for this value.
2250 SmallVector<SDValue, 8> Ops;
2251 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2252 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2254 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2257 // If this is a case we can't handle, return null and let the default
2258 // expansion code take care of it.
2259 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2260 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
2261 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
2262 DebugLoc dl = Op.getDebugLoc();
2264 APInt SplatBits, SplatUndef;
2265 unsigned SplatBitSize;
2267 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2268 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2269 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2271 return BuildSplat(Val, Op.getValueType(), DAG, dl);
2277 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2281 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
2285 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2286 MVT VT = Op.getValueType();
2287 DebugLoc dl = Op.getDebugLoc();
2288 assert((VT == MVT::i8 || VT == MVT::i16) &&
2289 "unexpected type for custom-lowering vector extract");
2290 SDValue Vec = Op.getOperand(0);
2291 SDValue Lane = Op.getOperand(1);
2292 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2293 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2294 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2297 static SDValue LowerCONCAT_VECTORS(SDValue Op) {
2298 if (Op.getValueType().is128BitVector() && Op.getNumOperands() == 2)
2303 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2304 switch (Op.getOpcode()) {
2305 default: llvm_unreachable("Don't know how to custom lower this!");
2306 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2307 case ISD::GlobalAddress:
2308 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2309 LowerGlobalAddressELF(Op, DAG);
2310 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2311 case ISD::CALL: return LowerCALL(Op, DAG);
2312 case ISD::RET: return LowerRET(Op, DAG);
2313 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2314 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2315 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2316 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2317 case ISD::SINT_TO_FP:
2318 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2319 case ISD::FP_TO_SINT:
2320 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2321 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2322 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
2323 case ISD::RETURNADDR: break;
2324 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2325 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2326 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2327 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2330 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2331 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2332 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2333 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2334 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
2335 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2336 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op);
2341 /// ReplaceNodeResults - Replace the results of node with an illegal result
2342 /// type with new values built out of custom code.
2343 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2344 SmallVectorImpl<SDValue>&Results,
2345 SelectionDAG &DAG) {
2346 switch (N->getOpcode()) {
2348 llvm_unreachable("Don't know how to custom expand this!");
2350 case ISD::BIT_CONVERT:
2351 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2355 SDValue Res = LowerShift(N, DAG, Subtarget);
2357 Results.push_back(Res);
2363 //===----------------------------------------------------------------------===//
2364 // ARM Scheduler Hooks
2365 //===----------------------------------------------------------------------===//
2368 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2369 MachineBasicBlock *BB) const {
2370 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2371 DebugLoc dl = MI->getDebugLoc();
2372 switch (MI->getOpcode()) {
2373 default: assert(false && "Unexpected instr type to insert");
2374 case ARM::tMOVCCr: {
2375 // To "insert" a SELECT_CC instruction, we actually have to insert the
2376 // diamond control-flow pattern. The incoming instruction knows the
2377 // destination vreg to set, the condition code register to branch on, the
2378 // true/false values to select between, and a branch opcode to use.
2379 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2380 MachineFunction::iterator It = BB;
2386 // cmpTY ccX, r1, r2
2388 // fallthrough --> copy0MBB
2389 MachineBasicBlock *thisMBB = BB;
2390 MachineFunction *F = BB->getParent();
2391 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2392 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2393 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2394 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2395 F->insert(It, copy0MBB);
2396 F->insert(It, sinkMBB);
2397 // Update machine-CFG edges by first adding all successors of the current
2398 // block to the new block which will contain the Phi node for the select.
2399 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2400 e = BB->succ_end(); i != e; ++i)
2401 sinkMBB->addSuccessor(*i);
2402 // Next, remove all successors of the current block, and add the true
2403 // and fallthrough blocks as its successors.
2404 while(!BB->succ_empty())
2405 BB->removeSuccessor(BB->succ_begin());
2406 BB->addSuccessor(copy0MBB);
2407 BB->addSuccessor(sinkMBB);
2410 // %FalseValue = ...
2411 // # fallthrough to sinkMBB
2414 // Update machine-CFG edges
2415 BB->addSuccessor(sinkMBB);
2418 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2421 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2422 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2423 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2425 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2431 //===----------------------------------------------------------------------===//
2432 // ARM Optimization Hooks
2433 //===----------------------------------------------------------------------===//
2436 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2437 TargetLowering::DAGCombinerInfo &DCI) {
2438 SelectionDAG &DAG = DCI.DAG;
2439 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2440 MVT VT = N->getValueType(0);
2441 unsigned Opc = N->getOpcode();
2442 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2443 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2444 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2445 ISD::CondCode CC = ISD::SETCC_INVALID;
2448 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2450 SDValue CCOp = Slct.getOperand(0);
2451 if (CCOp.getOpcode() == ISD::SETCC)
2452 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2455 bool DoXform = false;
2457 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2460 if (LHS.getOpcode() == ISD::Constant &&
2461 cast<ConstantSDNode>(LHS)->isNullValue()) {
2463 } else if (CC != ISD::SETCC_INVALID &&
2464 RHS.getOpcode() == ISD::Constant &&
2465 cast<ConstantSDNode>(RHS)->isNullValue()) {
2466 std::swap(LHS, RHS);
2467 SDValue Op0 = Slct.getOperand(0);
2468 MVT OpVT = isSlctCC ? Op0.getValueType() :
2469 Op0.getOperand(0).getValueType();
2470 bool isInt = OpVT.isInteger();
2471 CC = ISD::getSetCCInverse(CC, isInt);
2473 if (!TLI.isCondCodeLegal(CC, OpVT))
2474 return SDValue(); // Inverse operator isn't legal.
2481 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
2483 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
2484 Slct.getOperand(0), Slct.getOperand(1), CC);
2485 SDValue CCOp = Slct.getOperand(0);
2487 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
2488 CCOp.getOperand(0), CCOp.getOperand(1), CC);
2489 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
2490 CCOp, OtherOp, Result);
2495 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
2496 static SDValue PerformADDCombine(SDNode *N,
2497 TargetLowering::DAGCombinerInfo &DCI) {
2498 // added by evan in r37685 with no testcase.
2499 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2501 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
2502 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
2503 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
2504 if (Result.getNode()) return Result;
2506 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2507 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2508 if (Result.getNode()) return Result;
2514 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
2515 static SDValue PerformSUBCombine(SDNode *N,
2516 TargetLowering::DAGCombinerInfo &DCI) {
2517 // added by evan in r37685 with no testcase.
2518 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2520 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
2521 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2522 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2523 if (Result.getNode()) return Result;
2530 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
2531 static SDValue PerformFMRRDCombine(SDNode *N,
2532 TargetLowering::DAGCombinerInfo &DCI) {
2533 // fmrrd(fmdrr x, y) -> x,y
2534 SDValue InDouble = N->getOperand(0);
2535 if (InDouble.getOpcode() == ARMISD::FMDRR)
2536 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
2540 /// getVShiftImm - Check if this is a valid build_vector for the immediate
2541 /// operand of a vector shift operation, where all the elements of the
2542 /// build_vector must have the same constant integer value.
2543 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
2544 // Ignore bit_converts.
2545 while (Op.getOpcode() == ISD::BIT_CONVERT)
2546 Op = Op.getOperand(0);
2547 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
2548 APInt SplatBits, SplatUndef;
2549 unsigned SplatBitSize;
2551 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2552 HasAnyUndefs, ElementBits) ||
2553 SplatBitSize > ElementBits)
2555 Cnt = SplatBits.getSExtValue();
2559 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
2560 /// operand of a vector shift left operation. That value must be in the range:
2561 /// 0 <= Value < ElementBits for a left shift; or
2562 /// 0 <= Value <= ElementBits for a long left shift.
2563 static bool isVShiftLImm(SDValue Op, MVT VT, bool isLong, int64_t &Cnt) {
2564 assert(VT.isVector() && "vector shift count is not a vector type");
2565 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
2566 if (! getVShiftImm(Op, ElementBits, Cnt))
2568 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
2571 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
2572 /// operand of a vector shift right operation. For a shift opcode, the value
2573 /// is positive, but for an intrinsic the value count must be negative. The
2574 /// absolute value must be in the range:
2575 /// 1 <= |Value| <= ElementBits for a right shift; or
2576 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
2577 static bool isVShiftRImm(SDValue Op, MVT VT, bool isNarrow, bool isIntrinsic,
2579 assert(VT.isVector() && "vector shift count is not a vector type");
2580 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
2581 if (! getVShiftImm(Op, ElementBits, Cnt))
2585 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
2588 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
2589 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
2590 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2593 // Don't do anything for most intrinsics.
2596 // Vector shifts: check for immediate versions and lower them.
2597 // Note: This is done during DAG combining instead of DAG legalizing because
2598 // the build_vectors for 64-bit vector element shift counts are generally
2599 // not legal, and it is hard to see their values after they get legalized to
2600 // loads from a constant pool.
2601 case Intrinsic::arm_neon_vshifts:
2602 case Intrinsic::arm_neon_vshiftu:
2603 case Intrinsic::arm_neon_vshiftls:
2604 case Intrinsic::arm_neon_vshiftlu:
2605 case Intrinsic::arm_neon_vshiftn:
2606 case Intrinsic::arm_neon_vrshifts:
2607 case Intrinsic::arm_neon_vrshiftu:
2608 case Intrinsic::arm_neon_vrshiftn:
2609 case Intrinsic::arm_neon_vqshifts:
2610 case Intrinsic::arm_neon_vqshiftu:
2611 case Intrinsic::arm_neon_vqshiftsu:
2612 case Intrinsic::arm_neon_vqshiftns:
2613 case Intrinsic::arm_neon_vqshiftnu:
2614 case Intrinsic::arm_neon_vqshiftnsu:
2615 case Intrinsic::arm_neon_vqrshiftns:
2616 case Intrinsic::arm_neon_vqrshiftnu:
2617 case Intrinsic::arm_neon_vqrshiftnsu: {
2618 MVT VT = N->getOperand(1).getValueType();
2620 unsigned VShiftOpc = 0;
2623 case Intrinsic::arm_neon_vshifts:
2624 case Intrinsic::arm_neon_vshiftu:
2625 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
2626 VShiftOpc = ARMISD::VSHL;
2629 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
2630 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
2631 ARMISD::VSHRs : ARMISD::VSHRu);
2636 case Intrinsic::arm_neon_vshiftls:
2637 case Intrinsic::arm_neon_vshiftlu:
2638 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
2640 llvm_unreachable("invalid shift count for vshll intrinsic");
2642 case Intrinsic::arm_neon_vrshifts:
2643 case Intrinsic::arm_neon_vrshiftu:
2644 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
2648 case Intrinsic::arm_neon_vqshifts:
2649 case Intrinsic::arm_neon_vqshiftu:
2650 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
2654 case Intrinsic::arm_neon_vqshiftsu:
2655 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
2657 llvm_unreachable("invalid shift count for vqshlu intrinsic");
2659 case Intrinsic::arm_neon_vshiftn:
2660 case Intrinsic::arm_neon_vrshiftn:
2661 case Intrinsic::arm_neon_vqshiftns:
2662 case Intrinsic::arm_neon_vqshiftnu:
2663 case Intrinsic::arm_neon_vqshiftnsu:
2664 case Intrinsic::arm_neon_vqrshiftns:
2665 case Intrinsic::arm_neon_vqrshiftnu:
2666 case Intrinsic::arm_neon_vqrshiftnsu:
2667 // Narrowing shifts require an immediate right shift.
2668 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
2670 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
2673 llvm_unreachable("unhandled vector shift");
2677 case Intrinsic::arm_neon_vshifts:
2678 case Intrinsic::arm_neon_vshiftu:
2679 // Opcode already set above.
2681 case Intrinsic::arm_neon_vshiftls:
2682 case Intrinsic::arm_neon_vshiftlu:
2683 if (Cnt == VT.getVectorElementType().getSizeInBits())
2684 VShiftOpc = ARMISD::VSHLLi;
2686 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
2687 ARMISD::VSHLLs : ARMISD::VSHLLu);
2689 case Intrinsic::arm_neon_vshiftn:
2690 VShiftOpc = ARMISD::VSHRN; break;
2691 case Intrinsic::arm_neon_vrshifts:
2692 VShiftOpc = ARMISD::VRSHRs; break;
2693 case Intrinsic::arm_neon_vrshiftu:
2694 VShiftOpc = ARMISD::VRSHRu; break;
2695 case Intrinsic::arm_neon_vrshiftn:
2696 VShiftOpc = ARMISD::VRSHRN; break;
2697 case Intrinsic::arm_neon_vqshifts:
2698 VShiftOpc = ARMISD::VQSHLs; break;
2699 case Intrinsic::arm_neon_vqshiftu:
2700 VShiftOpc = ARMISD::VQSHLu; break;
2701 case Intrinsic::arm_neon_vqshiftsu:
2702 VShiftOpc = ARMISD::VQSHLsu; break;
2703 case Intrinsic::arm_neon_vqshiftns:
2704 VShiftOpc = ARMISD::VQSHRNs; break;
2705 case Intrinsic::arm_neon_vqshiftnu:
2706 VShiftOpc = ARMISD::VQSHRNu; break;
2707 case Intrinsic::arm_neon_vqshiftnsu:
2708 VShiftOpc = ARMISD::VQSHRNsu; break;
2709 case Intrinsic::arm_neon_vqrshiftns:
2710 VShiftOpc = ARMISD::VQRSHRNs; break;
2711 case Intrinsic::arm_neon_vqrshiftnu:
2712 VShiftOpc = ARMISD::VQRSHRNu; break;
2713 case Intrinsic::arm_neon_vqrshiftnsu:
2714 VShiftOpc = ARMISD::VQRSHRNsu; break;
2717 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
2718 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
2721 case Intrinsic::arm_neon_vshiftins: {
2722 MVT VT = N->getOperand(1).getValueType();
2724 unsigned VShiftOpc = 0;
2726 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
2727 VShiftOpc = ARMISD::VSLI;
2728 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
2729 VShiftOpc = ARMISD::VSRI;
2731 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
2734 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
2735 N->getOperand(1), N->getOperand(2),
2736 DAG.getConstant(Cnt, MVT::i32));
2739 case Intrinsic::arm_neon_vqrshifts:
2740 case Intrinsic::arm_neon_vqrshiftu:
2741 // No immediate versions of these to check for.
2748 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
2749 /// lowers them. As with the vector shift intrinsics, this is done during DAG
2750 /// combining instead of DAG legalizing because the build_vectors for 64-bit
2751 /// vector element shift counts are generally not legal, and it is hard to see
2752 /// their values after they get legalized to loads from a constant pool.
2753 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
2754 const ARMSubtarget *ST) {
2755 MVT VT = N->getValueType(0);
2757 // Nothing to be done for scalar shifts.
2758 if (! VT.isVector())
2761 assert(ST->hasNEON() && "unexpected vector shift");
2764 switch (N->getOpcode()) {
2765 default: llvm_unreachable("unexpected shift opcode");
2768 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
2769 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
2770 DAG.getConstant(Cnt, MVT::i32));
2775 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
2776 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
2777 ARMISD::VSHRs : ARMISD::VSHRu);
2778 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
2779 DAG.getConstant(Cnt, MVT::i32));
2785 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
2786 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
2787 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
2788 const ARMSubtarget *ST) {
2789 SDValue N0 = N->getOperand(0);
2791 // Check for sign- and zero-extensions of vector extract operations of 8-
2792 // and 16-bit vector elements. NEON supports these directly. They are
2793 // handled during DAG combining because type legalization will promote them
2794 // to 32-bit types and it is messy to recognize the operations after that.
2795 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
2796 SDValue Vec = N0.getOperand(0);
2797 SDValue Lane = N0.getOperand(1);
2798 MVT VT = N->getValueType(0);
2799 MVT EltVT = N0.getValueType();
2800 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2802 if (VT == MVT::i32 &&
2803 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
2804 TLI.isTypeLegal(Vec.getValueType())) {
2807 switch (N->getOpcode()) {
2808 default: llvm_unreachable("unexpected opcode");
2809 case ISD::SIGN_EXTEND:
2810 Opc = ARMISD::VGETLANEs;
2812 case ISD::ZERO_EXTEND:
2813 case ISD::ANY_EXTEND:
2814 Opc = ARMISD::VGETLANEu;
2817 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
2824 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
2825 DAGCombinerInfo &DCI) const {
2826 switch (N->getOpcode()) {
2828 case ISD::ADD: return PerformADDCombine(N, DCI);
2829 case ISD::SUB: return PerformSUBCombine(N, DCI);
2830 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
2831 case ISD::INTRINSIC_WO_CHAIN:
2832 return PerformIntrinsicCombine(N, DCI.DAG);
2836 return PerformShiftCombine(N, DCI.DAG, Subtarget);
2837 case ISD::SIGN_EXTEND:
2838 case ISD::ZERO_EXTEND:
2839 case ISD::ANY_EXTEND:
2840 return PerformExtendCombine(N, DCI.DAG, Subtarget);
2845 /// isLegalAddressImmediate - Return true if the integer value can be used
2846 /// as the offset of the target addressing mode for load / store of the
2848 static bool isLegalAddressImmediate(int64_t V, MVT VT,
2849 const ARMSubtarget *Subtarget) {
2856 if (Subtarget->isThumb()) { // FIXME for thumb2
2861 switch (VT.getSimpleVT()) {
2862 default: return false;
2877 if ((V & (Scale - 1)) != 0)
2880 return V == (V & ((1LL << 5) - 1));
2885 switch (VT.getSimpleVT()) {
2886 default: return false;
2891 return V == (V & ((1LL << 12) - 1));
2894 return V == (V & ((1LL << 8) - 1));
2897 if (!Subtarget->hasVFP2())
2902 return V == (V & ((1LL << 8) - 1));
2906 /// isLegalAddressingMode - Return true if the addressing mode represented
2907 /// by AM is legal for this target, for a load/store of the specified type.
2908 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
2909 const Type *Ty) const {
2910 MVT VT = getValueType(Ty, true);
2911 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
2914 // Can never fold addr of global into load/store.
2919 case 0: // no scale reg, must be "r+i" or "r", or "i".
2922 if (Subtarget->isThumb()) // FIXME for thumb2
2926 // ARM doesn't support any R+R*scale+imm addr modes.
2933 int Scale = AM.Scale;
2934 switch (VT.getSimpleVT()) {
2935 default: return false;
2940 // This assumes i64 is legalized to a pair of i32. If not (i.e.
2941 // ldrd / strd are used, then its address mode is same as i16.
2943 if (Scale < 0) Scale = -Scale;
2947 return isPowerOf2_32(Scale & ~1);
2950 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
2955 // Note, we allow "void" uses (basically, uses that aren't loads or
2956 // stores), because arm allows folding a scale into many arithmetic
2957 // operations. This should be made more precise and revisited later.
2959 // Allow r << imm, but the imm has to be a multiple of two.
2960 if (AM.Scale & 1) return false;
2961 return isPowerOf2_32(AM.Scale);
2968 static bool getARMIndexedAddressParts(SDNode *Ptr, MVT VT,
2969 bool isSEXTLoad, SDValue &Base,
2970 SDValue &Offset, bool &isInc,
2971 SelectionDAG &DAG) {
2972 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
2975 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
2977 Base = Ptr->getOperand(0);
2978 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
2979 int RHSC = (int)RHS->getZExtValue();
2980 if (RHSC < 0 && RHSC > -256) {
2981 assert(Ptr->getOpcode() == ISD::ADD);
2983 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
2987 isInc = (Ptr->getOpcode() == ISD::ADD);
2988 Offset = Ptr->getOperand(1);
2990 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
2992 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
2993 int RHSC = (int)RHS->getZExtValue();
2994 if (RHSC < 0 && RHSC > -0x1000) {
2995 assert(Ptr->getOpcode() == ISD::ADD);
2997 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
2998 Base = Ptr->getOperand(0);
3003 if (Ptr->getOpcode() == ISD::ADD) {
3005 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3006 if (ShOpcVal != ARM_AM::no_shift) {
3007 Base = Ptr->getOperand(1);
3008 Offset = Ptr->getOperand(0);
3010 Base = Ptr->getOperand(0);
3011 Offset = Ptr->getOperand(1);
3016 isInc = (Ptr->getOpcode() == ISD::ADD);
3017 Base = Ptr->getOperand(0);
3018 Offset = Ptr->getOperand(1);
3022 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3026 static bool getT2IndexedAddressParts(SDNode *Ptr, MVT VT,
3027 bool isSEXTLoad, SDValue &Base,
3028 SDValue &Offset, bool &isInc,
3029 SelectionDAG &DAG) {
3030 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3033 Base = Ptr->getOperand(0);
3034 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3035 int RHSC = (int)RHS->getZExtValue();
3036 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3037 assert(Ptr->getOpcode() == ISD::ADD);
3039 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3041 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3042 isInc = Ptr->getOpcode() == ISD::ADD;
3043 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3051 /// getPreIndexedAddressParts - returns true by value, base pointer and
3052 /// offset pointer and addressing mode by reference if the node's address
3053 /// can be legally represented as pre-indexed load / store address.
3055 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3057 ISD::MemIndexedMode &AM,
3058 SelectionDAG &DAG) const {
3059 if (Subtarget->isThumb1Only())
3064 bool isSEXTLoad = false;
3065 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3066 Ptr = LD->getBasePtr();
3067 VT = LD->getMemoryVT();
3068 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3069 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3070 Ptr = ST->getBasePtr();
3071 VT = ST->getMemoryVT();
3076 bool isLegal = false;
3077 if (Subtarget->isThumb() && Subtarget->hasThumb2())
3078 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3079 Offset, isInc, DAG);
3081 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3082 Offset, isInc, DAG);
3086 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3090 /// getPostIndexedAddressParts - returns true by value, base pointer and
3091 /// offset pointer and addressing mode by reference if this node can be
3092 /// combined with a load / store to form a post-indexed load / store.
3093 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3096 ISD::MemIndexedMode &AM,
3097 SelectionDAG &DAG) const {
3098 if (Subtarget->isThumb1Only())
3103 bool isSEXTLoad = false;
3104 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3105 VT = LD->getMemoryVT();
3106 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3107 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3108 VT = ST->getMemoryVT();
3113 bool isLegal = false;
3114 if (Subtarget->isThumb() && Subtarget->hasThumb2())
3115 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3118 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3123 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3127 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3131 const SelectionDAG &DAG,
3132 unsigned Depth) const {
3133 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3134 switch (Op.getOpcode()) {
3136 case ARMISD::CMOV: {
3137 // Bits are known zero/one if known on the LHS and RHS.
3138 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3139 if (KnownZero == 0 && KnownOne == 0) return;
3141 APInt KnownZeroRHS, KnownOneRHS;
3142 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3143 KnownZeroRHS, KnownOneRHS, Depth+1);
3144 KnownZero &= KnownZeroRHS;
3145 KnownOne &= KnownOneRHS;
3151 //===----------------------------------------------------------------------===//
3152 // ARM Inline Assembly Support
3153 //===----------------------------------------------------------------------===//
3155 /// getConstraintType - Given a constraint letter, return the type of
3156 /// constraint it is for this target.
3157 ARMTargetLowering::ConstraintType
3158 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3159 if (Constraint.size() == 1) {
3160 switch (Constraint[0]) {
3162 case 'l': return C_RegisterClass;
3163 case 'w': return C_RegisterClass;
3166 return TargetLowering::getConstraintType(Constraint);
3169 std::pair<unsigned, const TargetRegisterClass*>
3170 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3172 if (Constraint.size() == 1) {
3173 // GCC RS6000 Constraint Letters
3174 switch (Constraint[0]) {
3176 if (Subtarget->isThumb1Only())
3177 return std::make_pair(0U, ARM::tGPRRegisterClass);
3179 return std::make_pair(0U, ARM::GPRRegisterClass);
3181 return std::make_pair(0U, ARM::GPRRegisterClass);
3184 return std::make_pair(0U, ARM::SPRRegisterClass);
3186 return std::make_pair(0U, ARM::DPRRegisterClass);
3190 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3193 std::vector<unsigned> ARMTargetLowering::
3194 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3196 if (Constraint.size() != 1)
3197 return std::vector<unsigned>();
3199 switch (Constraint[0]) { // GCC ARM Constraint Letters
3202 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3203 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3206 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3207 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3208 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3209 ARM::R12, ARM::LR, 0);
3212 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3213 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3214 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3215 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3216 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3217 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3218 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3219 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3221 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3222 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3223 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3224 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3228 return std::vector<unsigned>();
3231 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3232 /// vector. If it is invalid, don't add anything to Ops.
3233 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3236 std::vector<SDValue>&Ops,
3237 SelectionDAG &DAG) const {
3238 SDValue Result(0, 0);
3240 switch (Constraint) {
3242 case 'I': case 'J': case 'K': case 'L':
3243 case 'M': case 'N': case 'O':
3244 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3248 int64_t CVal64 = C->getSExtValue();
3249 int CVal = (int) CVal64;
3250 // None of these constraints allow values larger than 32 bits. Check
3251 // that the value fits in an int.
3255 switch (Constraint) {
3257 if (Subtarget->isThumb1Only()) {
3258 // This must be a constant between 0 and 255, for ADD
3260 if (CVal >= 0 && CVal <= 255)
3262 } else if (Subtarget->isThumb2()) {
3263 // A constant that can be used as an immediate value in a
3264 // data-processing instruction.
3265 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3268 // A constant that can be used as an immediate value in a
3269 // data-processing instruction.
3270 if (ARM_AM::getSOImmVal(CVal) != -1)
3276 if (Subtarget->isThumb()) { // FIXME thumb2
3277 // This must be a constant between -255 and -1, for negated ADD
3278 // immediates. This can be used in GCC with an "n" modifier that
3279 // prints the negated value, for use with SUB instructions. It is
3280 // not useful otherwise but is implemented for compatibility.
3281 if (CVal >= -255 && CVal <= -1)
3284 // This must be a constant between -4095 and 4095. It is not clear
3285 // what this constraint is intended for. Implemented for
3286 // compatibility with GCC.
3287 if (CVal >= -4095 && CVal <= 4095)
3293 if (Subtarget->isThumb1Only()) {
3294 // A 32-bit value where only one byte has a nonzero value. Exclude
3295 // zero to match GCC. This constraint is used by GCC internally for
3296 // constants that can be loaded with a move/shift combination.
3297 // It is not useful otherwise but is implemented for compatibility.
3298 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3300 } else if (Subtarget->isThumb2()) {
3301 // A constant whose bitwise inverse can be used as an immediate
3302 // value in a data-processing instruction. This can be used in GCC
3303 // with a "B" modifier that prints the inverted value, for use with
3304 // BIC and MVN instructions. It is not useful otherwise but is
3305 // implemented for compatibility.
3306 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3309 // A constant whose bitwise inverse can be used as an immediate
3310 // value in a data-processing instruction. This can be used in GCC
3311 // with a "B" modifier that prints the inverted value, for use with
3312 // BIC and MVN instructions. It is not useful otherwise but is
3313 // implemented for compatibility.
3314 if (ARM_AM::getSOImmVal(~CVal) != -1)
3320 if (Subtarget->isThumb1Only()) {
3321 // This must be a constant between -7 and 7,
3322 // for 3-operand ADD/SUB immediate instructions.
3323 if (CVal >= -7 && CVal < 7)
3325 } else if (Subtarget->isThumb2()) {
3326 // A constant whose negation can be used as an immediate value in a
3327 // data-processing instruction. This can be used in GCC with an "n"
3328 // modifier that prints the negated value, for use with SUB
3329 // instructions. It is not useful otherwise but is implemented for
3331 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3334 // A constant whose negation can be used as an immediate value in a
3335 // data-processing instruction. This can be used in GCC with an "n"
3336 // modifier that prints the negated value, for use with SUB
3337 // instructions. It is not useful otherwise but is implemented for
3339 if (ARM_AM::getSOImmVal(-CVal) != -1)
3345 if (Subtarget->isThumb()) { // FIXME thumb2
3346 // This must be a multiple of 4 between 0 and 1020, for
3347 // ADD sp + immediate.
3348 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3351 // A power of two or a constant between 0 and 32. This is used in
3352 // GCC for the shift amount on shifted register operands, but it is
3353 // useful in general for any shift amounts.
3354 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3360 if (Subtarget->isThumb()) { // FIXME thumb2
3361 // This must be a constant between 0 and 31, for shift amounts.
3362 if (CVal >= 0 && CVal <= 31)
3368 if (Subtarget->isThumb()) { // FIXME thumb2
3369 // This must be a multiple of 4 between -508 and 508, for
3370 // ADD/SUB sp = sp + immediate.
3371 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3376 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3380 if (Result.getNode()) {
3381 Ops.push_back(Result);
3384 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,