1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMRegisterInfo.h"
21 #include "ARMSubtarget.h"
22 #include "ARMTargetMachine.h"
23 #include "ARMTargetObjectFile.h"
24 #include "llvm/CallingConv.h"
25 #include "llvm/Constants.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instruction.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/CodeGen/CallingConvLower.h"
31 #include "llvm/CodeGen/MachineBasicBlock.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/CodeGen/SelectionDAG.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include "llvm/ADT/VectorExtras.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/MathExtras.h"
44 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
49 CCValAssign::LocInfo &LocInfo,
50 ISD::ArgFlagsTy &ArgFlags,
52 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
53 CCValAssign::LocInfo &LocInfo,
54 ISD::ArgFlagsTy &ArgFlags,
56 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
57 CCValAssign::LocInfo &LocInfo,
58 ISD::ArgFlagsTy &ArgFlags,
61 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
62 MVT PromotedBitwiseVT) {
63 if (VT != PromotedLdStVT) {
64 setOperationAction(ISD::LOAD, VT, Promote);
65 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
67 setOperationAction(ISD::STORE, VT, Promote);
68 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
71 MVT ElemTy = VT.getVectorElementType();
72 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
73 setOperationAction(ISD::VSETCC, VT, Custom);
74 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
75 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
76 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
77 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
78 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
79 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
81 setOperationAction(ISD::SHL, VT, Custom);
82 setOperationAction(ISD::SRA, VT, Custom);
83 setOperationAction(ISD::SRL, VT, Custom);
86 // Promote all bit-wise operations.
87 if (VT.isInteger() && VT != PromotedBitwiseVT) {
88 setOperationAction(ISD::AND, VT, Promote);
89 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
90 setOperationAction(ISD::OR, VT, Promote);
91 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
92 setOperationAction(ISD::XOR, VT, Promote);
93 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
97 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
98 addRegisterClass(VT, ARM::DPRRegisterClass);
99 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
102 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
103 addRegisterClass(VT, ARM::QPRRegisterClass);
104 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
107 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
108 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
109 return new TargetLoweringObjectFileMachO();
110 return new ARMElfTargetObjectFile();
113 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
114 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
115 Subtarget = &TM.getSubtarget<ARMSubtarget>();
117 if (Subtarget->isTargetDarwin()) {
118 // Uses VFP for Thumb libfuncs if available.
119 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
120 // Single-precision floating-point arithmetic.
121 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
122 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
123 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
124 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
126 // Double-precision floating-point arithmetic.
127 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
128 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
129 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
130 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
132 // Single-precision comparisons.
133 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
134 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
135 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
136 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
137 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
138 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
139 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
140 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
142 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
143 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
144 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
145 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
146 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
147 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
148 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
149 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
151 // Double-precision comparisons.
152 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
153 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
154 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
155 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
156 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
157 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
158 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
159 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
161 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
162 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
163 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
164 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
165 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
166 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
167 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
170 // Floating-point to integer conversions.
171 // i64 conversions are done via library routines even when generating VFP
172 // instructions, so use the same ones.
173 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
174 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
175 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
176 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
178 // Conversions between floating types.
179 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
180 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
182 // Integer to floating-point conversions.
183 // i64 conversions are done via library routines even when generating VFP
184 // instructions, so use the same ones.
185 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
186 // e.g., __floatunsidf vs. __floatunssidfvfp.
187 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
188 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
189 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
190 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
194 // These libcalls are not available in 32-bit.
195 setLibcallName(RTLIB::SHL_I128, 0);
196 setLibcallName(RTLIB::SRL_I128, 0);
197 setLibcallName(RTLIB::SRA_I128, 0);
199 if (Subtarget->isThumb1Only())
200 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
202 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
203 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
204 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
205 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
207 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
210 if (Subtarget->hasNEON()) {
211 addDRTypeForNEON(MVT::v2f32);
212 addDRTypeForNEON(MVT::v8i8);
213 addDRTypeForNEON(MVT::v4i16);
214 addDRTypeForNEON(MVT::v2i32);
215 addDRTypeForNEON(MVT::v1i64);
217 addQRTypeForNEON(MVT::v4f32);
218 addQRTypeForNEON(MVT::v2f64);
219 addQRTypeForNEON(MVT::v16i8);
220 addQRTypeForNEON(MVT::v8i16);
221 addQRTypeForNEON(MVT::v4i32);
222 addQRTypeForNEON(MVT::v2i64);
224 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
225 setTargetDAGCombine(ISD::SHL);
226 setTargetDAGCombine(ISD::SRL);
227 setTargetDAGCombine(ISD::SRA);
228 setTargetDAGCombine(ISD::SIGN_EXTEND);
229 setTargetDAGCombine(ISD::ZERO_EXTEND);
230 setTargetDAGCombine(ISD::ANY_EXTEND);
233 computeRegisterProperties();
235 // ARM does not have f32 extending load.
236 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
238 // ARM does not have i1 sign extending load.
239 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
241 // ARM supports all 4 flavors of integer indexed load / store.
242 if (!Subtarget->isThumb1Only()) {
243 for (unsigned im = (unsigned)ISD::PRE_INC;
244 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
245 setIndexedLoadAction(im, MVT::i1, Legal);
246 setIndexedLoadAction(im, MVT::i8, Legal);
247 setIndexedLoadAction(im, MVT::i16, Legal);
248 setIndexedLoadAction(im, MVT::i32, Legal);
249 setIndexedStoreAction(im, MVT::i1, Legal);
250 setIndexedStoreAction(im, MVT::i8, Legal);
251 setIndexedStoreAction(im, MVT::i16, Legal);
252 setIndexedStoreAction(im, MVT::i32, Legal);
256 // i64 operation support.
257 if (Subtarget->isThumb1Only()) {
258 setOperationAction(ISD::MUL, MVT::i64, Expand);
259 setOperationAction(ISD::MULHU, MVT::i32, Expand);
260 setOperationAction(ISD::MULHS, MVT::i32, Expand);
261 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
262 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
264 setOperationAction(ISD::MUL, MVT::i64, Expand);
265 setOperationAction(ISD::MULHU, MVT::i32, Expand);
266 if (!Subtarget->hasV6Ops())
267 setOperationAction(ISD::MULHS, MVT::i32, Expand);
269 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
270 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
271 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
272 setOperationAction(ISD::SRL, MVT::i64, Custom);
273 setOperationAction(ISD::SRA, MVT::i64, Custom);
275 // ARM does not have ROTL.
276 setOperationAction(ISD::ROTL, MVT::i32, Expand);
277 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
278 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
279 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
280 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
282 // Only ARMv6 has BSWAP.
283 if (!Subtarget->hasV6Ops())
284 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
286 // These are expanded into libcalls.
287 setOperationAction(ISD::SDIV, MVT::i32, Expand);
288 setOperationAction(ISD::UDIV, MVT::i32, Expand);
289 setOperationAction(ISD::SREM, MVT::i32, Expand);
290 setOperationAction(ISD::UREM, MVT::i32, Expand);
291 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
292 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
294 // Support label based line numbers.
295 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
296 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
298 setOperationAction(ISD::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);
326 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
328 setOperationAction(ISD::SETCC, MVT::i32, Expand);
329 setOperationAction(ISD::SETCC, MVT::f32, Expand);
330 setOperationAction(ISD::SETCC, MVT::f64, Expand);
331 setOperationAction(ISD::SELECT, MVT::i32, Expand);
332 setOperationAction(ISD::SELECT, MVT::f32, Expand);
333 setOperationAction(ISD::SELECT, MVT::f64, Expand);
334 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
335 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
336 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
338 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
339 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
340 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
341 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
342 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
344 // We don't support sin/cos/fmod/copysign/pow
345 setOperationAction(ISD::FSIN, MVT::f64, Expand);
346 setOperationAction(ISD::FSIN, MVT::f32, Expand);
347 setOperationAction(ISD::FCOS, MVT::f32, Expand);
348 setOperationAction(ISD::FCOS, MVT::f64, Expand);
349 setOperationAction(ISD::FREM, MVT::f64, Expand);
350 setOperationAction(ISD::FREM, MVT::f32, Expand);
351 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
352 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
353 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
355 setOperationAction(ISD::FPOW, MVT::f64, Expand);
356 setOperationAction(ISD::FPOW, MVT::f32, Expand);
358 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
359 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
360 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
361 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
362 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
363 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
366 // We have target-specific dag combine patterns for the following nodes:
367 // ARMISD::FMRRD - No need to call setTargetDAGCombine
368 setTargetDAGCombine(ISD::ADD);
369 setTargetDAGCombine(ISD::SUB);
371 setStackPointerRegisterToSaveRestore(ARM::SP);
372 setSchedulingPreference(SchedulingForRegPressure);
373 setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10);
374 setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2);
376 if (!Subtarget->isThumb()) {
377 // Use branch latency information to determine if-conversion limits.
378 // FIXME: If-converter should use instruction latency of the branch being
379 // eliminated to compute the threshold. For ARMv6, the branch "latency"
380 // varies depending on whether it's dynamically or statically predicted
381 // and on whether the destination is in the prefetch buffer.
382 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
383 const InstrItineraryData &InstrItins = Subtarget->getInstrItineraryData();
384 unsigned Latency= InstrItins.getLatency(TII->get(ARM::Bcc).getSchedClass());
386 setIfCvtBlockSizeLimit(Latency-1);
388 setIfCvtDupBlockSizeLimit(Latency-2);
390 setIfCvtBlockSizeLimit(10);
391 setIfCvtDupBlockSizeLimit(2);
395 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
396 // Do not enable CodePlacementOpt for now: it currently runs after the
397 // ARMConstantIslandPass and messes up branch relaxation and placement
398 // of constant islands.
399 // benefitFromCodePlacementOpt = true;
402 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
405 case ARMISD::Wrapper: return "ARMISD::Wrapper";
406 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
407 case ARMISD::CALL: return "ARMISD::CALL";
408 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
409 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
410 case ARMISD::tCALL: return "ARMISD::tCALL";
411 case ARMISD::BRCOND: return "ARMISD::BRCOND";
412 case ARMISD::BR_JT: return "ARMISD::BR_JT";
413 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
414 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
415 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
416 case ARMISD::CMP: return "ARMISD::CMP";
417 case ARMISD::CMPZ: return "ARMISD::CMPZ";
418 case ARMISD::CMPFP: return "ARMISD::CMPFP";
419 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
420 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
421 case ARMISD::CMOV: return "ARMISD::CMOV";
422 case ARMISD::CNEG: return "ARMISD::CNEG";
424 case ARMISD::FTOSI: return "ARMISD::FTOSI";
425 case ARMISD::FTOUI: return "ARMISD::FTOUI";
426 case ARMISD::SITOF: return "ARMISD::SITOF";
427 case ARMISD::UITOF: return "ARMISD::UITOF";
429 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
430 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
431 case ARMISD::RRX: return "ARMISD::RRX";
433 case ARMISD::FMRRD: return "ARMISD::FMRRD";
434 case ARMISD::FMDRR: return "ARMISD::FMDRR";
436 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
438 case ARMISD::VCEQ: return "ARMISD::VCEQ";
439 case ARMISD::VCGE: return "ARMISD::VCGE";
440 case ARMISD::VCGEU: return "ARMISD::VCGEU";
441 case ARMISD::VCGT: return "ARMISD::VCGT";
442 case ARMISD::VCGTU: return "ARMISD::VCGTU";
443 case ARMISD::VTST: return "ARMISD::VTST";
445 case ARMISD::VSHL: return "ARMISD::VSHL";
446 case ARMISD::VSHRs: return "ARMISD::VSHRs";
447 case ARMISD::VSHRu: return "ARMISD::VSHRu";
448 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
449 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
450 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
451 case ARMISD::VSHRN: return "ARMISD::VSHRN";
452 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
453 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
454 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
455 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
456 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
457 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
458 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
459 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
460 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
461 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
462 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
463 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
464 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
465 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
466 case ARMISD::VDUPLANEQ: return "ARMISD::VDUPLANEQ";
467 case ARMISD::VLD2D: return "ARMISD::VLD2D";
468 case ARMISD::VLD3D: return "ARMISD::VLD3D";
469 case ARMISD::VLD4D: return "ARMISD::VLD4D";
473 /// getFunctionAlignment - Return the Log2 alignment of this function.
474 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
475 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
478 //===----------------------------------------------------------------------===//
480 //===----------------------------------------------------------------------===//
482 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
483 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
485 default: llvm_unreachable("Unknown condition code!");
486 case ISD::SETNE: return ARMCC::NE;
487 case ISD::SETEQ: return ARMCC::EQ;
488 case ISD::SETGT: return ARMCC::GT;
489 case ISD::SETGE: return ARMCC::GE;
490 case ISD::SETLT: return ARMCC::LT;
491 case ISD::SETLE: return ARMCC::LE;
492 case ISD::SETUGT: return ARMCC::HI;
493 case ISD::SETUGE: return ARMCC::HS;
494 case ISD::SETULT: return ARMCC::LO;
495 case ISD::SETULE: return ARMCC::LS;
499 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
500 /// returns true if the operands should be inverted to form the proper
502 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
503 ARMCC::CondCodes &CondCode2) {
505 CondCode2 = ARMCC::AL;
507 default: llvm_unreachable("Unknown FP condition!");
509 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
511 case ISD::SETOGT: CondCode = ARMCC::GT; break;
513 case ISD::SETOGE: CondCode = ARMCC::GE; break;
514 case ISD::SETOLT: CondCode = ARMCC::MI; break;
515 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
516 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
517 case ISD::SETO: CondCode = ARMCC::VC; break;
518 case ISD::SETUO: CondCode = ARMCC::VS; break;
519 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
520 case ISD::SETUGT: CondCode = ARMCC::HI; break;
521 case ISD::SETUGE: CondCode = ARMCC::PL; break;
523 case ISD::SETULT: CondCode = ARMCC::LT; break;
525 case ISD::SETULE: CondCode = ARMCC::LE; break;
527 case ISD::SETUNE: CondCode = ARMCC::NE; break;
532 //===----------------------------------------------------------------------===//
533 // Calling Convention Implementation
535 // The lower operations present on calling convention works on this order:
536 // LowerCALL (virt regs --> phys regs, virt regs --> stack)
537 // LowerFORMAL_ARGUMENTS (phys --> virt regs, stack --> virt regs)
538 // LowerRET (virt regs --> phys regs)
539 // LowerCALL (phys regs --> virt regs)
541 //===----------------------------------------------------------------------===//
543 #include "ARMGenCallingConv.inc"
545 // APCS f64 is in register pairs, possibly split to stack
546 static bool f64AssignAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
547 CCValAssign::LocInfo &LocInfo,
548 CCState &State, bool CanFail) {
549 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
551 // Try to get the first register.
552 if (unsigned Reg = State.AllocateReg(RegList, 4))
553 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
555 // For the 2nd half of a v2f64, do not fail.
559 // Put the whole thing on the stack.
560 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
561 State.AllocateStack(8, 4),
566 // Try to get the second register.
567 if (unsigned Reg = State.AllocateReg(RegList, 4))
568 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
570 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
571 State.AllocateStack(4, 4),
576 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
577 CCValAssign::LocInfo &LocInfo,
578 ISD::ArgFlagsTy &ArgFlags,
580 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
582 if (LocVT == MVT::v2f64 &&
583 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
585 return true; // we handled it
588 // AAPCS f64 is in aligned register pairs
589 static bool f64AssignAAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
590 CCValAssign::LocInfo &LocInfo,
591 CCState &State, bool CanFail) {
592 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
593 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
595 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
597 // For the 2nd half of a v2f64, do not just fail.
601 // Put the whole thing on the stack.
602 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
603 State.AllocateStack(8, 8),
609 for (i = 0; i < 2; ++i)
610 if (HiRegList[i] == Reg)
613 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
614 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
619 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
620 CCValAssign::LocInfo &LocInfo,
621 ISD::ArgFlagsTy &ArgFlags,
623 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
625 if (LocVT == MVT::v2f64 &&
626 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
628 return true; // we handled it
631 static bool f64RetAssign(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
632 CCValAssign::LocInfo &LocInfo, CCState &State) {
633 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
634 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
636 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
638 return false; // we didn't handle it
641 for (i = 0; i < 2; ++i)
642 if (HiRegList[i] == Reg)
645 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
646 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
651 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
652 CCValAssign::LocInfo &LocInfo,
653 ISD::ArgFlagsTy &ArgFlags,
655 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
657 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
659 return true; // we handled it
662 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
663 CCValAssign::LocInfo &LocInfo,
664 ISD::ArgFlagsTy &ArgFlags,
666 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
670 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
671 /// given CallingConvention value.
672 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC,
676 llvm_unreachable("Unsupported calling convention");
678 case CallingConv::Fast:
679 // Use target triple & subtarget features to do actual dispatch.
680 if (Subtarget->isAAPCS_ABI()) {
681 if (Subtarget->hasVFP2() &&
682 FloatABIType == FloatABI::Hard)
683 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
685 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
687 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
688 case CallingConv::ARM_AAPCS_VFP:
689 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
690 case CallingConv::ARM_AAPCS:
691 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
692 case CallingConv::ARM_APCS:
693 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
697 /// LowerCallResult - Lower the result values of an ISD::CALL into the
698 /// appropriate copies out of appropriate physical registers. This assumes that
699 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
700 /// being lowered. The returns a SDNode with the same number of values as the
702 SDNode *ARMTargetLowering::
703 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
704 unsigned CallingConv, SelectionDAG &DAG) {
706 DebugLoc dl = TheCall->getDebugLoc();
707 // Assign locations to each value returned by this call.
708 SmallVector<CCValAssign, 16> RVLocs;
709 bool isVarArg = TheCall->isVarArg();
710 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(),
711 RVLocs, *DAG.getContext());
712 CCInfo.AnalyzeCallResult(TheCall,
713 CCAssignFnForNode(CallingConv, /* Return*/ true));
715 SmallVector<SDValue, 8> ResultVals;
717 // Copy all of the result registers out of their specified physreg.
718 for (unsigned i = 0; i != RVLocs.size(); ++i) {
719 CCValAssign VA = RVLocs[i];
722 if (VA.needsCustom()) {
723 // Handle f64 or half of a v2f64.
724 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
726 Chain = Lo.getValue(1);
727 InFlag = Lo.getValue(2);
728 VA = RVLocs[++i]; // skip ahead to next loc
729 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
731 Chain = Hi.getValue(1);
732 InFlag = Hi.getValue(2);
733 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
735 if (VA.getLocVT() == MVT::v2f64) {
736 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
737 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
738 DAG.getConstant(0, MVT::i32));
740 VA = RVLocs[++i]; // skip ahead to next loc
741 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
742 Chain = Lo.getValue(1);
743 InFlag = Lo.getValue(2);
744 VA = RVLocs[++i]; // skip ahead to next loc
745 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
746 Chain = Hi.getValue(1);
747 InFlag = Hi.getValue(2);
748 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
749 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
750 DAG.getConstant(1, MVT::i32));
753 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
755 Chain = Val.getValue(1);
756 InFlag = Val.getValue(2);
759 switch (VA.getLocInfo()) {
760 default: llvm_unreachable("Unknown loc info!");
761 case CCValAssign::Full: break;
762 case CCValAssign::BCvt:
763 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
767 ResultVals.push_back(Val);
770 // Merge everything together with a MERGE_VALUES node.
771 ResultVals.push_back(Chain);
772 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
773 &ResultVals[0], ResultVals.size()).getNode();
776 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
777 /// by "Src" to address "Dst" of size "Size". Alignment information is
778 /// specified by the specific parameter attribute. The copy will be passed as
779 /// a byval function parameter.
780 /// Sometimes what we are copying is the end of a larger object, the part that
781 /// does not fit in registers.
783 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
784 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
786 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
787 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
788 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
791 /// LowerMemOpCallTo - Store the argument to the stack.
793 ARMTargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
794 const SDValue &StackPtr,
795 const CCValAssign &VA, SDValue Chain,
796 SDValue Arg, ISD::ArgFlagsTy Flags) {
797 DebugLoc dl = TheCall->getDebugLoc();
798 unsigned LocMemOffset = VA.getLocMemOffset();
799 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
800 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
801 if (Flags.isByVal()) {
802 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
804 return DAG.getStore(Chain, dl, Arg, PtrOff,
805 PseudoSourceValue::getStack(), LocMemOffset);
808 void ARMTargetLowering::PassF64ArgInRegs(CallSDNode *TheCall, SelectionDAG &DAG,
809 SDValue Chain, SDValue &Arg,
810 RegsToPassVector &RegsToPass,
811 CCValAssign &VA, CCValAssign &NextVA,
813 SmallVector<SDValue, 8> &MemOpChains,
814 ISD::ArgFlagsTy Flags) {
815 DebugLoc dl = TheCall->getDebugLoc();
817 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
818 DAG.getVTList(MVT::i32, MVT::i32), Arg);
819 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
821 if (NextVA.isRegLoc())
822 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
824 assert(NextVA.isMemLoc());
825 if (StackPtr.getNode() == 0)
826 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
828 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, NextVA,
829 Chain, fmrrd.getValue(1), Flags));
833 /// LowerCALL - Lowering a ISD::CALL node into a callseq_start <-
834 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
836 SDValue ARMTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
837 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
838 MVT RetVT = TheCall->getRetValType(0);
839 SDValue Chain = TheCall->getChain();
840 unsigned CC = TheCall->getCallingConv();
841 bool isVarArg = TheCall->isVarArg();
842 SDValue Callee = TheCall->getCallee();
843 DebugLoc dl = TheCall->getDebugLoc();
845 // Analyze operands of the call, assigning locations to each operand.
846 SmallVector<CCValAssign, 16> ArgLocs;
847 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
848 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC, /* Return*/ false));
850 // Get a count of how many bytes are to be pushed on the stack.
851 unsigned NumBytes = CCInfo.getNextStackOffset();
853 // Adjust the stack pointer for the new arguments...
854 // These operations are automatically eliminated by the prolog/epilog pass
855 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
857 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
859 RegsToPassVector RegsToPass;
860 SmallVector<SDValue, 8> MemOpChains;
862 // Walk the register/memloc assignments, inserting copies/loads. In the case
863 // of tail call optimization, arguments are handled later.
864 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
867 CCValAssign &VA = ArgLocs[i];
868 SDValue Arg = TheCall->getArg(realArgIdx);
869 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(realArgIdx);
871 // Promote the value if needed.
872 switch (VA.getLocInfo()) {
873 default: llvm_unreachable("Unknown loc info!");
874 case CCValAssign::Full: break;
875 case CCValAssign::SExt:
876 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
878 case CCValAssign::ZExt:
879 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
881 case CCValAssign::AExt:
882 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
884 case CCValAssign::BCvt:
885 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
889 // f64 and v2f64 are passed in i32 pairs and must be split into pieces
890 if (VA.needsCustom()) {
891 if (VA.getLocVT() == MVT::v2f64) {
892 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
893 DAG.getConstant(0, MVT::i32));
894 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
895 DAG.getConstant(1, MVT::i32));
897 PassF64ArgInRegs(TheCall, DAG, Chain, Op0, RegsToPass,
898 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
900 VA = ArgLocs[++i]; // skip ahead to next loc
902 PassF64ArgInRegs(TheCall, DAG, Chain, Op1, RegsToPass,
903 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
905 assert(VA.isMemLoc());
906 if (StackPtr.getNode() == 0)
907 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
909 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
913 PassF64ArgInRegs(TheCall, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
914 StackPtr, MemOpChains, Flags);
916 } else if (VA.isRegLoc()) {
917 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
919 assert(VA.isMemLoc());
920 if (StackPtr.getNode() == 0)
921 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
923 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
928 if (!MemOpChains.empty())
929 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
930 &MemOpChains[0], MemOpChains.size());
932 // Build a sequence of copy-to-reg nodes chained together with token chain
933 // and flag operands which copy the outgoing args into the appropriate regs.
935 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
936 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
937 RegsToPass[i].second, InFlag);
938 InFlag = Chain.getValue(1);
941 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
942 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
943 // node so that legalize doesn't hack it.
944 bool isDirect = false;
945 bool isARMFunc = false;
946 bool isLocalARMFunc = false;
947 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
948 GlobalValue *GV = G->getGlobal();
950 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
951 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
952 getTargetMachine().getRelocationModel() != Reloc::Static;
953 isARMFunc = !Subtarget->isThumb() || isStub;
954 // ARM call to a local ARM function is predicable.
955 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
956 // tBX takes a register source operand.
957 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
958 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
960 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
961 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
962 Callee = DAG.getLoad(getPointerTy(), dl,
963 DAG.getEntryNode(), CPAddr, NULL, 0);
964 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
965 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
966 getPointerTy(), Callee, PICLabel);
968 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
969 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
971 bool isStub = Subtarget->isTargetDarwin() &&
972 getTargetMachine().getRelocationModel() != Reloc::Static;
973 isARMFunc = !Subtarget->isThumb() || isStub;
974 // tBX takes a register source operand.
975 const char *Sym = S->getSymbol();
976 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
977 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
979 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
980 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
981 Callee = DAG.getLoad(getPointerTy(), dl,
982 DAG.getEntryNode(), CPAddr, NULL, 0);
983 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
984 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
985 getPointerTy(), Callee, PICLabel);
987 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
990 // FIXME: handle tail calls differently.
992 if (Subtarget->isThumb()) {
993 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
994 CallOpc = ARMISD::CALL_NOLINK;
996 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
998 CallOpc = (isDirect || Subtarget->hasV5TOps())
999 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1000 : ARMISD::CALL_NOLINK;
1002 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1003 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1004 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1005 InFlag = Chain.getValue(1);
1008 std::vector<SDValue> Ops;
1009 Ops.push_back(Chain);
1010 Ops.push_back(Callee);
1012 // Add argument registers to the end of the list so that they are known live
1014 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1015 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1016 RegsToPass[i].second.getValueType()));
1018 if (InFlag.getNode())
1019 Ops.push_back(InFlag);
1020 // Returns a chain and a flag for retval copy to use.
1021 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1022 &Ops[0], Ops.size());
1023 InFlag = Chain.getValue(1);
1025 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1026 DAG.getIntPtrConstant(0, true), InFlag);
1027 if (RetVT != MVT::Other)
1028 InFlag = Chain.getValue(1);
1030 // Handle result values, copying them out of physregs into vregs that we
1032 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
1036 SDValue ARMTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
1037 // The chain is always operand #0
1038 SDValue Chain = Op.getOperand(0);
1039 DebugLoc dl = Op.getDebugLoc();
1041 // CCValAssign - represent the assignment of the return value to a location.
1042 SmallVector<CCValAssign, 16> RVLocs;
1043 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1044 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1046 // CCState - Info about the registers and stack slots.
1047 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs, *DAG.getContext());
1049 // Analyze return values of ISD::RET.
1050 CCInfo.AnalyzeReturn(Op.getNode(), CCAssignFnForNode(CC, /* Return */ true));
1052 // If this is the first return lowered for this function, add
1053 // the regs to the liveout set for the function.
1054 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1055 for (unsigned i = 0; i != RVLocs.size(); ++i)
1056 if (RVLocs[i].isRegLoc())
1057 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1062 // Copy the result values into the output registers.
1063 for (unsigned i = 0, realRVLocIdx = 0;
1065 ++i, ++realRVLocIdx) {
1066 CCValAssign &VA = RVLocs[i];
1067 assert(VA.isRegLoc() && "Can only return in registers!");
1069 // ISD::RET => ret chain, (regnum1,val1), ...
1070 // So i*2+1 index only the regnums
1071 SDValue Arg = Op.getOperand(realRVLocIdx*2+1);
1073 switch (VA.getLocInfo()) {
1074 default: llvm_unreachable("Unknown loc info!");
1075 case CCValAssign::Full: break;
1076 case CCValAssign::BCvt:
1077 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1081 if (VA.needsCustom()) {
1082 if (VA.getLocVT() == MVT::v2f64) {
1083 // Extract the first half and return it in two registers.
1084 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1085 DAG.getConstant(0, MVT::i32));
1086 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1087 DAG.getVTList(MVT::i32, MVT::i32), Half);
1089 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1090 Flag = Chain.getValue(1);
1091 VA = RVLocs[++i]; // skip ahead to next loc
1092 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1093 HalfGPRs.getValue(1), Flag);
1094 Flag = Chain.getValue(1);
1095 VA = RVLocs[++i]; // skip ahead to next loc
1097 // Extract the 2nd half and fall through to handle it as an f64 value.
1098 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1099 DAG.getConstant(1, MVT::i32));
1101 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1103 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1104 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1105 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1106 Flag = Chain.getValue(1);
1107 VA = RVLocs[++i]; // skip ahead to next loc
1108 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1111 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1113 // Guarantee that all emitted copies are
1114 // stuck together, avoiding something bad.
1115 Flag = Chain.getValue(1);
1120 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1122 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1127 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1128 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1129 // one of the above mentioned nodes. It has to be wrapped because otherwise
1130 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1131 // be used to form addressing mode. These wrapped nodes will be selected
1133 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1134 MVT PtrVT = Op.getValueType();
1135 // FIXME there is no actual debug info here
1136 DebugLoc dl = Op.getDebugLoc();
1137 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1139 if (CP->isMachineConstantPoolEntry())
1140 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1141 CP->getAlignment());
1143 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1144 CP->getAlignment());
1145 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1148 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1150 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1151 SelectionDAG &DAG) {
1152 DebugLoc dl = GA->getDebugLoc();
1153 MVT PtrVT = getPointerTy();
1154 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1155 ARMConstantPoolValue *CPV =
1156 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1157 PCAdj, "tlsgd", true);
1158 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1159 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1160 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1161 SDValue Chain = Argument.getValue(1);
1163 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1164 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1166 // call __tls_get_addr.
1169 Entry.Node = Argument;
1170 Entry.Ty = (const Type *) Type::Int32Ty;
1171 Args.push_back(Entry);
1172 // FIXME: is there useful debug info available here?
1173 std::pair<SDValue, SDValue> CallResult =
1174 LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false, false, false,
1175 0, CallingConv::C, false,
1176 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1177 return CallResult.first;
1180 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1181 // "local exec" model.
1183 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1184 SelectionDAG &DAG) {
1185 GlobalValue *GV = GA->getGlobal();
1186 DebugLoc dl = GA->getDebugLoc();
1188 SDValue Chain = DAG.getEntryNode();
1189 MVT PtrVT = getPointerTy();
1190 // Get the Thread Pointer
1191 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1193 if (GV->isDeclaration()) {
1194 // initial exec model
1195 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1196 ARMConstantPoolValue *CPV =
1197 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1198 PCAdj, "gottpoff", true);
1199 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1200 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1201 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1202 Chain = Offset.getValue(1);
1204 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1205 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1207 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1210 ARMConstantPoolValue *CPV =
1211 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
1212 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1213 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1214 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1217 // The address of the thread local variable is the add of the thread
1218 // pointer with the offset of the variable.
1219 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1223 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1224 // TODO: implement the "local dynamic" model
1225 assert(Subtarget->isTargetELF() &&
1226 "TLS not implemented for non-ELF targets");
1227 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1228 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1229 // otherwise use the "Local Exec" TLS Model
1230 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1231 return LowerToTLSGeneralDynamicModel(GA, DAG);
1233 return LowerToTLSExecModels(GA, DAG);
1236 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1237 SelectionDAG &DAG) {
1238 MVT PtrVT = getPointerTy();
1239 DebugLoc dl = Op.getDebugLoc();
1240 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1241 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1242 if (RelocM == Reloc::PIC_) {
1243 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1244 ARMConstantPoolValue *CPV =
1245 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
1246 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1247 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1248 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1250 SDValue Chain = Result.getValue(1);
1251 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1252 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1254 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1257 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1258 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1259 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1263 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
1264 /// even in non-static mode.
1265 static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
1266 // If symbol visibility is hidden, the extra load is not needed if
1267 // the symbol is definitely defined in the current translation unit.
1268 bool isDecl = GV->isDeclaration() || GV->hasAvailableExternallyLinkage();
1269 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage()))
1271 return RelocM != Reloc::Static && (isDecl || GV->isWeakForLinker());
1274 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1275 SelectionDAG &DAG) {
1276 MVT PtrVT = getPointerTy();
1277 DebugLoc dl = Op.getDebugLoc();
1278 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1279 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1280 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
1282 if (RelocM == Reloc::Static)
1283 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1285 unsigned PCAdj = (RelocM != Reloc::PIC_)
1286 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1287 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
1289 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
1291 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1293 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1295 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1296 SDValue Chain = Result.getValue(1);
1298 if (RelocM == Reloc::PIC_) {
1299 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1300 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1303 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1308 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1310 assert(Subtarget->isTargetELF() &&
1311 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1312 MVT PtrVT = getPointerTy();
1313 DebugLoc dl = Op.getDebugLoc();
1314 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1315 ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_",
1317 ARMCP::CPValue, PCAdj);
1318 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1319 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1320 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1321 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1322 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1325 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
1326 unsigned Opcode, unsigned NumVecs) {
1327 SDNode *Node = Op.getNode();
1328 MVT VT = Node->getValueType(0);
1329 DebugLoc dl = Op.getDebugLoc();
1331 if (!VT.is64BitVector())
1332 return SDValue(); // unimplemented
1334 SDValue Ops[] = { Node->getOperand(0),
1335 Node->getOperand(1) };
1336 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
1337 SDValue Result = DAG.getNode(Opcode, dl, Tys, Ops, 2);
1339 static const unsigned VLDRegs[] = {
1340 ARM::D0, ARM::D1, ARM::D2, ARM::D3
1343 SmallVector<SDValue, 4> ResultVals;
1344 SDValue Chain = Result.getValue(0);
1345 SDValue Flag = Result.getValue(1);
1346 for (unsigned N = 0; N < NumVecs; ++N) {
1347 Chain = DAG.getCopyFromReg(Chain, dl, VLDRegs[N], VT, Flag).getValue(1);
1348 ResultVals.push_back(Chain.getValue(0));
1349 Flag = Chain.getValue(2);
1351 ResultVals.push_back(Chain);
1352 return DAG.getNode(ISD::MERGE_VALUES, dl, Node->getVTList(),
1353 ResultVals.data(), NumVecs + 1);
1357 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1358 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1360 case Intrinsic::arm_neon_vld2i:
1361 case Intrinsic::arm_neon_vld2f:
1362 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD2D, 2);
1363 case Intrinsic::arm_neon_vld3i:
1364 case Intrinsic::arm_neon_vld3f:
1365 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD3D, 3);
1366 case Intrinsic::arm_neon_vld4i:
1367 case Intrinsic::arm_neon_vld4f:
1368 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD4D, 4);
1369 case Intrinsic::arm_neon_vst2i:
1370 case Intrinsic::arm_neon_vst2f:
1371 case Intrinsic::arm_neon_vst3i:
1372 case Intrinsic::arm_neon_vst3f:
1373 case Intrinsic::arm_neon_vst4i:
1374 case Intrinsic::arm_neon_vst4f:
1375 default: return SDValue(); // Don't custom lower most intrinsics.
1380 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1381 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1382 DebugLoc dl = Op.getDebugLoc();
1384 default: return SDValue(); // Don't custom lower most intrinsics.
1385 case Intrinsic::arm_thread_pointer: {
1386 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1387 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1389 case Intrinsic::eh_sjlj_setjmp:
1390 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1394 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1395 unsigned VarArgsFrameIndex) {
1396 // vastart just stores the address of the VarArgsFrameIndex slot into the
1397 // memory location argument.
1398 DebugLoc dl = Op.getDebugLoc();
1399 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1400 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1401 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1402 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1406 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1407 SDValue &Root, SelectionDAG &DAG,
1409 MachineFunction &MF = DAG.getMachineFunction();
1410 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1412 TargetRegisterClass *RC;
1413 if (AFI->isThumb1OnlyFunction())
1414 RC = ARM::tGPRRegisterClass;
1416 RC = ARM::GPRRegisterClass;
1418 // Transform the arguments stored in physical registers into virtual ones.
1419 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1420 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1423 if (NextVA.isMemLoc()) {
1424 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1425 MachineFrameInfo *MFI = MF.getFrameInfo();
1426 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1428 // Create load node to retrieve arguments from the stack.
1429 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1430 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1432 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1433 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1436 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1440 ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1441 MachineFunction &MF = DAG.getMachineFunction();
1442 MachineFrameInfo *MFI = MF.getFrameInfo();
1444 SDValue Root = Op.getOperand(0);
1445 DebugLoc dl = Op.getDebugLoc();
1446 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1447 unsigned CC = MF.getFunction()->getCallingConv();
1448 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1450 // Assign locations to all of the incoming arguments.
1451 SmallVector<CCValAssign, 16> ArgLocs;
1452 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
1453 CCInfo.AnalyzeFormalArguments(Op.getNode(),
1454 CCAssignFnForNode(CC, /* Return*/ false));
1456 SmallVector<SDValue, 16> ArgValues;
1458 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1459 CCValAssign &VA = ArgLocs[i];
1461 // Arguments stored in registers.
1462 if (VA.isRegLoc()) {
1463 MVT RegVT = VA.getLocVT();
1466 if (VA.needsCustom()) {
1467 // f64 and vector types are split up into multiple registers or
1468 // combinations of registers and stack slots.
1471 if (VA.getLocVT() == MVT::v2f64) {
1472 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1474 VA = ArgLocs[++i]; // skip ahead to next loc
1475 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1477 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1478 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1479 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1480 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1481 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1483 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Root, DAG, dl);
1486 TargetRegisterClass *RC;
1487 if (FloatABIType == FloatABI::Hard && RegVT == MVT::f32)
1488 RC = ARM::SPRRegisterClass;
1489 else if (FloatABIType == FloatABI::Hard && RegVT == MVT::f64)
1490 RC = ARM::DPRRegisterClass;
1491 else if (AFI->isThumb1OnlyFunction())
1492 RC = ARM::tGPRRegisterClass;
1494 RC = ARM::GPRRegisterClass;
1496 assert((RegVT == MVT::i32 || RegVT == MVT::f32 ||
1497 (FloatABIType == FloatABI::Hard && RegVT == MVT::f64)) &&
1498 "RegVT not supported by FORMAL_ARGUMENTS Lowering");
1500 // Transform the arguments in physical registers into virtual ones.
1501 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1502 ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1505 // If this is an 8 or 16-bit value, it is really passed promoted
1506 // to 32 bits. Insert an assert[sz]ext to capture this, then
1507 // truncate to the right size.
1508 switch (VA.getLocInfo()) {
1509 default: llvm_unreachable("Unknown loc info!");
1510 case CCValAssign::Full: break;
1511 case CCValAssign::BCvt:
1512 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1514 case CCValAssign::SExt:
1515 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1516 DAG.getValueType(VA.getValVT()));
1517 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1519 case CCValAssign::ZExt:
1520 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1521 DAG.getValueType(VA.getValVT()));
1522 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1526 ArgValues.push_back(ArgValue);
1528 } else { // VA.isRegLoc()
1531 assert(VA.isMemLoc());
1532 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1534 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1535 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1537 // Create load nodes to retrieve arguments from the stack.
1538 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1539 ArgValues.push_back(DAG.getLoad(VA.getValVT(), dl, Root, FIN, NULL, 0));
1545 static const unsigned GPRArgRegs[] = {
1546 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1549 unsigned NumGPRs = CCInfo.getFirstUnallocated
1550 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1552 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1553 unsigned VARegSize = (4 - NumGPRs) * 4;
1554 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1555 unsigned ArgOffset = 0;
1556 if (VARegSaveSize) {
1557 // If this function is vararg, store any remaining integer argument regs
1558 // to their spots on the stack so that they may be loaded by deferencing
1559 // the result of va_next.
1560 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1561 ArgOffset = CCInfo.getNextStackOffset();
1562 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1563 VARegSaveSize - VARegSize);
1564 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1566 SmallVector<SDValue, 4> MemOps;
1567 for (; NumGPRs < 4; ++NumGPRs) {
1568 TargetRegisterClass *RC;
1569 if (AFI->isThumb1OnlyFunction())
1570 RC = ARM::tGPRRegisterClass;
1572 RC = ARM::GPRRegisterClass;
1574 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1575 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
1576 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1577 MemOps.push_back(Store);
1578 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1579 DAG.getConstant(4, getPointerTy()));
1581 if (!MemOps.empty())
1582 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1583 &MemOps[0], MemOps.size());
1585 // This will point to the next argument passed via stack.
1586 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1589 ArgValues.push_back(Root);
1591 // Return the new list of results.
1592 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1593 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1596 /// isFloatingPointZero - Return true if this is +0.0.
1597 static bool isFloatingPointZero(SDValue Op) {
1598 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1599 return CFP->getValueAPF().isPosZero();
1600 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1601 // Maybe this has already been legalized into the constant pool?
1602 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1603 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1604 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1605 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1606 return CFP->getValueAPF().isPosZero();
1612 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1613 return ( isThumb1Only && (C & ~255U) == 0) ||
1614 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1617 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1618 /// the given operands.
1619 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1620 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1622 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1623 unsigned C = RHSC->getZExtValue();
1624 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1625 // Constant does not fit, try adjusting it by one?
1630 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1631 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1632 RHS = DAG.getConstant(C-1, MVT::i32);
1637 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1638 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1639 RHS = DAG.getConstant(C-1, MVT::i32);
1644 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1645 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1646 RHS = DAG.getConstant(C+1, MVT::i32);
1651 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1652 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1653 RHS = DAG.getConstant(C+1, MVT::i32);
1660 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1661 ARMISD::NodeType CompareType;
1664 CompareType = ARMISD::CMP;
1669 CompareType = ARMISD::CMPZ;
1672 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1673 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1676 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1677 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1680 if (!isFloatingPointZero(RHS))
1681 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1683 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1684 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1687 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1688 const ARMSubtarget *ST) {
1689 MVT VT = Op.getValueType();
1690 SDValue LHS = Op.getOperand(0);
1691 SDValue RHS = Op.getOperand(1);
1692 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1693 SDValue TrueVal = Op.getOperand(2);
1694 SDValue FalseVal = Op.getOperand(3);
1695 DebugLoc dl = Op.getDebugLoc();
1697 if (LHS.getValueType() == MVT::i32) {
1699 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1700 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1701 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1704 ARMCC::CondCodes CondCode, CondCode2;
1705 if (FPCCToARMCC(CC, CondCode, CondCode2))
1706 std::swap(TrueVal, FalseVal);
1708 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1709 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1710 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1711 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1713 if (CondCode2 != ARMCC::AL) {
1714 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1715 // FIXME: Needs another CMP because flag can have but one use.
1716 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1717 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1718 Result, TrueVal, ARMCC2, CCR, Cmp2);
1723 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1724 const ARMSubtarget *ST) {
1725 SDValue Chain = Op.getOperand(0);
1726 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1727 SDValue LHS = Op.getOperand(2);
1728 SDValue RHS = Op.getOperand(3);
1729 SDValue Dest = Op.getOperand(4);
1730 DebugLoc dl = Op.getDebugLoc();
1732 if (LHS.getValueType() == MVT::i32) {
1734 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1735 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1736 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1737 Chain, Dest, ARMCC, CCR,Cmp);
1740 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1741 ARMCC::CondCodes CondCode, CondCode2;
1742 if (FPCCToARMCC(CC, CondCode, CondCode2))
1743 // Swap the LHS/RHS of the comparison if needed.
1744 std::swap(LHS, RHS);
1746 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1747 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1748 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1749 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1750 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1751 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1752 if (CondCode2 != ARMCC::AL) {
1753 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1754 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1755 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1760 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1761 SDValue Chain = Op.getOperand(0);
1762 SDValue Table = Op.getOperand(1);
1763 SDValue Index = Op.getOperand(2);
1764 DebugLoc dl = Op.getDebugLoc();
1766 MVT PTy = getPointerTy();
1767 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1768 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1769 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1770 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1771 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1772 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1773 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1774 if (Subtarget->isThumb2()) {
1775 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1776 // which does another jump to the destination. This also makes it easier
1777 // to translate it to TBB / TBH later.
1778 // FIXME: This might not work if the function is extremely large.
1779 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1780 Addr, Op.getOperand(2), JTI, UId);
1782 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1783 Addr = DAG.getLoad((MVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1784 Chain = Addr.getValue(1);
1785 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1786 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1788 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1789 Chain = Addr.getValue(1);
1790 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1794 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1795 DebugLoc dl = Op.getDebugLoc();
1797 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1798 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1799 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1802 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1803 MVT VT = Op.getValueType();
1804 DebugLoc dl = Op.getDebugLoc();
1806 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1808 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1809 return DAG.getNode(Opc, dl, VT, Op);
1812 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1813 // Implement fcopysign with a fabs and a conditional fneg.
1814 SDValue Tmp0 = Op.getOperand(0);
1815 SDValue Tmp1 = Op.getOperand(1);
1816 DebugLoc dl = Op.getDebugLoc();
1817 MVT VT = Op.getValueType();
1818 MVT SrcVT = Tmp1.getValueType();
1819 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1820 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1821 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1822 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1823 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1826 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1827 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1828 MFI->setFrameAddressIsTaken(true);
1829 MVT VT = Op.getValueType();
1830 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1831 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1832 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1833 ? ARM::R7 : ARM::R11;
1834 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1836 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1841 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1843 SDValue Dst, SDValue Src,
1844 SDValue Size, unsigned Align,
1846 const Value *DstSV, uint64_t DstSVOff,
1847 const Value *SrcSV, uint64_t SrcSVOff){
1848 // Do repeated 4-byte loads and stores. To be improved.
1849 // This requires 4-byte alignment.
1850 if ((Align & 3) != 0)
1852 // This requires the copy size to be a constant, preferrably
1853 // within a subtarget-specific limit.
1854 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1857 uint64_t SizeVal = ConstantSize->getZExtValue();
1858 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1861 unsigned BytesLeft = SizeVal & 3;
1862 unsigned NumMemOps = SizeVal >> 2;
1863 unsigned EmittedNumMemOps = 0;
1865 unsigned VTSize = 4;
1867 const unsigned MAX_LOADS_IN_LDM = 6;
1868 SDValue TFOps[MAX_LOADS_IN_LDM];
1869 SDValue Loads[MAX_LOADS_IN_LDM];
1870 uint64_t SrcOff = 0, DstOff = 0;
1872 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1873 // same number of stores. The loads and stores will get combined into
1874 // ldm/stm later on.
1875 while (EmittedNumMemOps < NumMemOps) {
1877 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1878 Loads[i] = DAG.getLoad(VT, dl, Chain,
1879 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1880 DAG.getConstant(SrcOff, MVT::i32)),
1881 SrcSV, SrcSVOff + SrcOff);
1882 TFOps[i] = Loads[i].getValue(1);
1885 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1888 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1889 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1890 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1891 DAG.getConstant(DstOff, MVT::i32)),
1892 DstSV, DstSVOff + DstOff);
1895 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1897 EmittedNumMemOps += i;
1903 // Issue loads / stores for the trailing (1 - 3) bytes.
1904 unsigned BytesLeftSave = BytesLeft;
1907 if (BytesLeft >= 2) {
1915 Loads[i] = DAG.getLoad(VT, dl, Chain,
1916 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1917 DAG.getConstant(SrcOff, MVT::i32)),
1918 SrcSV, SrcSVOff + SrcOff);
1919 TFOps[i] = Loads[i].getValue(1);
1922 BytesLeft -= VTSize;
1924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1927 BytesLeft = BytesLeftSave;
1929 if (BytesLeft >= 2) {
1937 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1938 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1939 DAG.getConstant(DstOff, MVT::i32)),
1940 DstSV, DstSVOff + DstOff);
1943 BytesLeft -= VTSize;
1945 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1948 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
1949 SDValue Op = N->getOperand(0);
1950 DebugLoc dl = N->getDebugLoc();
1951 if (N->getValueType(0) == MVT::f64) {
1952 // Turn i64->f64 into FMDRR.
1953 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1954 DAG.getConstant(0, MVT::i32));
1955 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1956 DAG.getConstant(1, MVT::i32));
1957 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
1960 // Turn f64->i64 into FMRRD.
1961 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
1962 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
1964 // Merge the pieces into a single i64 value.
1965 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
1968 /// getZeroVector - Returns a vector of specified type with all zero elements.
1970 static SDValue getZeroVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
1971 assert(VT.isVector() && "Expected a vector type");
1973 // Zero vectors are used to represent vector negation and in those cases
1974 // will be implemented with the NEON VNEG instruction. However, VNEG does
1975 // not support i64 elements, so sometimes the zero vectors will need to be
1976 // explicitly constructed. For those cases, and potentially other uses in
1977 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
1978 // to their dest type. This ensures they get CSE'd.
1980 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
1981 if (VT.getSizeInBits() == 64)
1982 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
1984 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
1986 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
1989 /// getOnesVector - Returns a vector of specified type with all bits set.
1991 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
1992 assert(VT.isVector() && "Expected a vector type");
1994 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
1995 // type. This ensures they get CSE'd.
1997 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
1998 if (VT.getSizeInBits() == 64)
1999 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2001 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2003 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2006 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2007 const ARMSubtarget *ST) {
2008 MVT VT = N->getValueType(0);
2009 DebugLoc dl = N->getDebugLoc();
2011 // Lower vector shifts on NEON to use VSHL.
2012 if (VT.isVector()) {
2013 assert(ST->hasNEON() && "unexpected vector shift");
2015 // Left shifts translate directly to the vshiftu intrinsic.
2016 if (N->getOpcode() == ISD::SHL)
2017 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2018 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2019 N->getOperand(0), N->getOperand(1));
2021 assert((N->getOpcode() == ISD::SRA ||
2022 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2024 // NEON uses the same intrinsics for both left and right shifts. For
2025 // right shifts, the shift amounts are negative, so negate the vector of
2027 MVT ShiftVT = N->getOperand(1).getValueType();
2028 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2029 getZeroVector(ShiftVT, DAG, dl),
2031 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2032 Intrinsic::arm_neon_vshifts :
2033 Intrinsic::arm_neon_vshiftu);
2034 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2035 DAG.getConstant(vshiftInt, MVT::i32),
2036 N->getOperand(0), NegatedCount);
2039 assert(VT == MVT::i64 &&
2040 (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2041 "Unknown shift to lower!");
2043 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2044 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2045 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2048 // If we are in thumb mode, we don't have RRX.
2049 if (ST->isThumb1Only()) return SDValue();
2051 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2052 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2053 DAG.getConstant(0, MVT::i32));
2054 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2055 DAG.getConstant(1, MVT::i32));
2057 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2058 // captures the result into a carry flag.
2059 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2060 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2062 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2063 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2065 // Merge the pieces into a single i64 value.
2066 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2069 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2070 SDValue TmpOp0, TmpOp1;
2071 bool Invert = false;
2075 SDValue Op0 = Op.getOperand(0);
2076 SDValue Op1 = Op.getOperand(1);
2077 SDValue CC = Op.getOperand(2);
2078 MVT VT = Op.getValueType();
2079 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2080 DebugLoc dl = Op.getDebugLoc();
2082 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2083 switch (SetCCOpcode) {
2084 default: llvm_unreachable("Illegal FP comparison"); break;
2086 case ISD::SETNE: Invert = true; // Fallthrough
2088 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2090 case ISD::SETLT: Swap = true; // Fallthrough
2092 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2094 case ISD::SETLE: Swap = true; // Fallthrough
2096 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2097 case ISD::SETUGE: Swap = true; // Fallthrough
2098 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2099 case ISD::SETUGT: Swap = true; // Fallthrough
2100 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2101 case ISD::SETUEQ: Invert = true; // Fallthrough
2103 // Expand this to (OLT | OGT).
2107 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2108 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2110 case ISD::SETUO: Invert = true; // Fallthrough
2112 // Expand this to (OLT | OGE).
2116 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2117 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2121 // Integer comparisons.
2122 switch (SetCCOpcode) {
2123 default: llvm_unreachable("Illegal integer comparison"); break;
2124 case ISD::SETNE: Invert = true;
2125 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2126 case ISD::SETLT: Swap = true;
2127 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2128 case ISD::SETLE: Swap = true;
2129 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2130 case ISD::SETULT: Swap = true;
2131 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2132 case ISD::SETULE: Swap = true;
2133 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2136 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2137 if (Opc == ARMISD::VCEQ) {
2140 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2142 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2145 // Ignore bitconvert.
2146 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2147 AndOp = AndOp.getOperand(0);
2149 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2151 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2152 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2159 std::swap(Op0, Op1);
2161 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2164 Result = DAG.getNOT(dl, Result, VT);
2169 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2170 /// VMOV instruction, and if so, return the constant being splatted.
2171 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2172 unsigned SplatBitSize, SelectionDAG &DAG) {
2173 switch (SplatBitSize) {
2175 // Any 1-byte value is OK.
2176 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2177 return DAG.getTargetConstant(SplatBits, MVT::i8);
2180 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2181 if ((SplatBits & ~0xff) == 0 ||
2182 (SplatBits & ~0xff00) == 0)
2183 return DAG.getTargetConstant(SplatBits, MVT::i16);
2187 // NEON's 32-bit VMOV supports splat values where:
2188 // * only one byte is nonzero, or
2189 // * the least significant byte is 0xff and the second byte is nonzero, or
2190 // * the least significant 2 bytes are 0xff and the third is nonzero.
2191 if ((SplatBits & ~0xff) == 0 ||
2192 (SplatBits & ~0xff00) == 0 ||
2193 (SplatBits & ~0xff0000) == 0 ||
2194 (SplatBits & ~0xff000000) == 0)
2195 return DAG.getTargetConstant(SplatBits, MVT::i32);
2197 if ((SplatBits & ~0xffff) == 0 &&
2198 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2199 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2201 if ((SplatBits & ~0xffffff) == 0 &&
2202 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2203 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2205 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2206 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2207 // VMOV.I32. A (very) minor optimization would be to replicate the value
2208 // and fall through here to test for a valid 64-bit splat. But, then the
2209 // caller would also need to check and handle the change in size.
2213 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2214 uint64_t BitMask = 0xff;
2216 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2217 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2219 else if ((SplatBits & BitMask) != 0)
2223 return DAG.getTargetConstant(Val, MVT::i64);
2227 llvm_unreachable("unexpected size for isVMOVSplat");
2234 /// getVMOVImm - If this is a build_vector of constants which can be
2235 /// formed by using a VMOV instruction of the specified element size,
2236 /// return the constant being splatted. The ByteSize field indicates the
2237 /// number of bytes of each element [1248].
2238 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2239 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2240 APInt SplatBits, SplatUndef;
2241 unsigned SplatBitSize;
2243 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2244 HasAnyUndefs, ByteSize * 8))
2247 if (SplatBitSize > ByteSize * 8)
2250 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2254 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2255 /// instruction with the specified blocksize. (The order of the elements
2256 /// within each block of the vector is reversed.)
2257 bool ARM::isVREVMask(ShuffleVectorSDNode *N, unsigned BlockSize) {
2258 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2259 "Only possible block sizes for VREV are: 16, 32, 64");
2261 MVT VT = N->getValueType(0);
2262 unsigned NumElts = VT.getVectorNumElements();
2263 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2264 unsigned BlockElts = N->getMaskElt(0) + 1;
2266 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2269 for (unsigned i = 0; i < NumElts; ++i) {
2270 if ((unsigned) N->getMaskElt(i) !=
2271 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2278 static SDValue BuildSplat(SDValue Val, MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2279 // Canonicalize all-zeros and all-ones vectors.
2280 ConstantSDNode *ConstVal = dyn_cast<ConstantSDNode>(Val.getNode());
2281 if (ConstVal->isNullValue())
2282 return getZeroVector(VT, DAG, dl);
2283 if (ConstVal->isAllOnesValue())
2284 return getOnesVector(VT, DAG, dl);
2287 if (VT.is64BitVector()) {
2288 switch (Val.getValueType().getSizeInBits()) {
2289 case 8: CanonicalVT = MVT::v8i8; break;
2290 case 16: CanonicalVT = MVT::v4i16; break;
2291 case 32: CanonicalVT = MVT::v2i32; break;
2292 case 64: CanonicalVT = MVT::v1i64; break;
2293 default: llvm_unreachable("unexpected splat element type"); break;
2296 assert(VT.is128BitVector() && "unknown splat vector size");
2297 switch (Val.getValueType().getSizeInBits()) {
2298 case 8: CanonicalVT = MVT::v16i8; break;
2299 case 16: CanonicalVT = MVT::v8i16; break;
2300 case 32: CanonicalVT = MVT::v4i32; break;
2301 case 64: CanonicalVT = MVT::v2i64; break;
2302 default: llvm_unreachable("unexpected splat element type"); break;
2306 // Build a canonical splat for this value.
2307 SmallVector<SDValue, 8> Ops;
2308 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2309 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2311 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2314 // If this is a case we can't handle, return null and let the default
2315 // expansion code take care of it.
2316 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2317 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
2318 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
2319 DebugLoc dl = Op.getDebugLoc();
2320 MVT VT = Op.getValueType();
2322 APInt SplatBits, SplatUndef;
2323 unsigned SplatBitSize;
2325 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2326 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2327 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2329 return BuildSplat(Val, VT, DAG, dl);
2332 // If there are only 2 elements in a 128-bit vector, insert them into an
2333 // undef vector. This handles the common case for 128-bit vector argument
2334 // passing, where the insertions should be translated to subreg accesses
2335 // with no real instructions.
2336 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2337 SDValue Val = DAG.getUNDEF(VT);
2338 SDValue Op0 = Op.getOperand(0);
2339 SDValue Op1 = Op.getOperand(1);
2340 if (Op0.getOpcode() != ISD::UNDEF)
2341 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2342 DAG.getIntPtrConstant(0));
2343 if (Op1.getOpcode() != ISD::UNDEF)
2344 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2345 DAG.getIntPtrConstant(1));
2352 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2356 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
2360 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2361 MVT VT = Op.getValueType();
2362 DebugLoc dl = Op.getDebugLoc();
2363 assert((VT == MVT::i8 || VT == MVT::i16) &&
2364 "unexpected type for custom-lowering vector extract");
2365 SDValue Vec = Op.getOperand(0);
2366 SDValue Lane = Op.getOperand(1);
2367 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2368 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2369 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2372 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2373 // The only time a CONCAT_VECTORS operation can have legal types is when
2374 // two 64-bit vectors are concatenated to a 128-bit vector.
2375 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2376 "unexpected CONCAT_VECTORS");
2377 DebugLoc dl = Op.getDebugLoc();
2378 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2379 SDValue Op0 = Op.getOperand(0);
2380 SDValue Op1 = Op.getOperand(1);
2381 if (Op0.getOpcode() != ISD::UNDEF)
2382 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2383 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2384 DAG.getIntPtrConstant(0));
2385 if (Op1.getOpcode() != ISD::UNDEF)
2386 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2387 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2388 DAG.getIntPtrConstant(1));
2389 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2392 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2393 switch (Op.getOpcode()) {
2394 default: llvm_unreachable("Don't know how to custom lower this!");
2395 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2396 case ISD::GlobalAddress:
2397 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2398 LowerGlobalAddressELF(Op, DAG);
2399 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2400 case ISD::CALL: return LowerCALL(Op, DAG);
2401 case ISD::RET: return LowerRET(Op, DAG);
2402 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2403 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2404 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2405 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2406 case ISD::SINT_TO_FP:
2407 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2408 case ISD::FP_TO_SINT:
2409 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2410 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2411 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
2412 case ISD::RETURNADDR: break;
2413 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2414 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2415 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2416 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2417 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2420 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2421 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2422 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2423 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2424 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
2425 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2426 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2431 /// ReplaceNodeResults - Replace the results of node with an illegal result
2432 /// type with new values built out of custom code.
2433 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2434 SmallVectorImpl<SDValue>&Results,
2435 SelectionDAG &DAG) {
2436 switch (N->getOpcode()) {
2438 llvm_unreachable("Don't know how to custom expand this!");
2440 case ISD::BIT_CONVERT:
2441 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2445 SDValue Res = LowerShift(N, DAG, Subtarget);
2447 Results.push_back(Res);
2453 //===----------------------------------------------------------------------===//
2454 // ARM Scheduler Hooks
2455 //===----------------------------------------------------------------------===//
2458 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2459 MachineBasicBlock *BB) const {
2460 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2461 DebugLoc dl = MI->getDebugLoc();
2462 switch (MI->getOpcode()) {
2463 default: assert(false && "Unexpected instr type to insert");
2464 case ARM::tMOVCCr: {
2465 // To "insert" a SELECT_CC instruction, we actually have to insert the
2466 // diamond control-flow pattern. The incoming instruction knows the
2467 // destination vreg to set, the condition code register to branch on, the
2468 // true/false values to select between, and a branch opcode to use.
2469 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2470 MachineFunction::iterator It = BB;
2476 // cmpTY ccX, r1, r2
2478 // fallthrough --> copy0MBB
2479 MachineBasicBlock *thisMBB = BB;
2480 MachineFunction *F = BB->getParent();
2481 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2482 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2483 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2484 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2485 F->insert(It, copy0MBB);
2486 F->insert(It, sinkMBB);
2487 // Update machine-CFG edges by first adding all successors of the current
2488 // block to the new block which will contain the Phi node for the select.
2489 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2490 e = BB->succ_end(); i != e; ++i)
2491 sinkMBB->addSuccessor(*i);
2492 // Next, remove all successors of the current block, and add the true
2493 // and fallthrough blocks as its successors.
2494 while(!BB->succ_empty())
2495 BB->removeSuccessor(BB->succ_begin());
2496 BB->addSuccessor(copy0MBB);
2497 BB->addSuccessor(sinkMBB);
2500 // %FalseValue = ...
2501 // # fallthrough to sinkMBB
2504 // Update machine-CFG edges
2505 BB->addSuccessor(sinkMBB);
2508 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2511 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2512 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2513 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2515 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2521 //===----------------------------------------------------------------------===//
2522 // ARM Optimization Hooks
2523 //===----------------------------------------------------------------------===//
2526 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2527 TargetLowering::DAGCombinerInfo &DCI) {
2528 SelectionDAG &DAG = DCI.DAG;
2529 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2530 MVT VT = N->getValueType(0);
2531 unsigned Opc = N->getOpcode();
2532 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2533 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2534 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2535 ISD::CondCode CC = ISD::SETCC_INVALID;
2538 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2540 SDValue CCOp = Slct.getOperand(0);
2541 if (CCOp.getOpcode() == ISD::SETCC)
2542 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2545 bool DoXform = false;
2547 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2550 if (LHS.getOpcode() == ISD::Constant &&
2551 cast<ConstantSDNode>(LHS)->isNullValue()) {
2553 } else if (CC != ISD::SETCC_INVALID &&
2554 RHS.getOpcode() == ISD::Constant &&
2555 cast<ConstantSDNode>(RHS)->isNullValue()) {
2556 std::swap(LHS, RHS);
2557 SDValue Op0 = Slct.getOperand(0);
2558 MVT OpVT = isSlctCC ? Op0.getValueType() :
2559 Op0.getOperand(0).getValueType();
2560 bool isInt = OpVT.isInteger();
2561 CC = ISD::getSetCCInverse(CC, isInt);
2563 if (!TLI.isCondCodeLegal(CC, OpVT))
2564 return SDValue(); // Inverse operator isn't legal.
2571 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
2573 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
2574 Slct.getOperand(0), Slct.getOperand(1), CC);
2575 SDValue CCOp = Slct.getOperand(0);
2577 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
2578 CCOp.getOperand(0), CCOp.getOperand(1), CC);
2579 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
2580 CCOp, OtherOp, Result);
2585 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
2586 static SDValue PerformADDCombine(SDNode *N,
2587 TargetLowering::DAGCombinerInfo &DCI) {
2588 // added by evan in r37685 with no testcase.
2589 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2591 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
2592 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
2593 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
2594 if (Result.getNode()) return Result;
2596 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2597 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2598 if (Result.getNode()) return Result;
2604 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
2605 static SDValue PerformSUBCombine(SDNode *N,
2606 TargetLowering::DAGCombinerInfo &DCI) {
2607 // added by evan in r37685 with no testcase.
2608 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2610 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
2611 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2612 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2613 if (Result.getNode()) return Result;
2620 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
2621 static SDValue PerformFMRRDCombine(SDNode *N,
2622 TargetLowering::DAGCombinerInfo &DCI) {
2623 // fmrrd(fmdrr x, y) -> x,y
2624 SDValue InDouble = N->getOperand(0);
2625 if (InDouble.getOpcode() == ARMISD::FMDRR)
2626 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
2630 /// getVShiftImm - Check if this is a valid build_vector for the immediate
2631 /// operand of a vector shift operation, where all the elements of the
2632 /// build_vector must have the same constant integer value.
2633 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
2634 // Ignore bit_converts.
2635 while (Op.getOpcode() == ISD::BIT_CONVERT)
2636 Op = Op.getOperand(0);
2637 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
2638 APInt SplatBits, SplatUndef;
2639 unsigned SplatBitSize;
2641 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2642 HasAnyUndefs, ElementBits) ||
2643 SplatBitSize > ElementBits)
2645 Cnt = SplatBits.getSExtValue();
2649 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
2650 /// operand of a vector shift left operation. That value must be in the range:
2651 /// 0 <= Value < ElementBits for a left shift; or
2652 /// 0 <= Value <= ElementBits for a long left shift.
2653 static bool isVShiftLImm(SDValue Op, MVT VT, bool isLong, int64_t &Cnt) {
2654 assert(VT.isVector() && "vector shift count is not a vector type");
2655 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
2656 if (! getVShiftImm(Op, ElementBits, Cnt))
2658 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
2661 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
2662 /// operand of a vector shift right operation. For a shift opcode, the value
2663 /// is positive, but for an intrinsic the value count must be negative. The
2664 /// absolute value must be in the range:
2665 /// 1 <= |Value| <= ElementBits for a right shift; or
2666 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
2667 static bool isVShiftRImm(SDValue Op, MVT VT, bool isNarrow, bool isIntrinsic,
2669 assert(VT.isVector() && "vector shift count is not a vector type");
2670 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
2671 if (! getVShiftImm(Op, ElementBits, Cnt))
2675 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
2678 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
2679 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
2680 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2683 // Don't do anything for most intrinsics.
2686 // Vector shifts: check for immediate versions and lower them.
2687 // Note: This is done during DAG combining instead of DAG legalizing because
2688 // the build_vectors for 64-bit vector element shift counts are generally
2689 // not legal, and it is hard to see their values after they get legalized to
2690 // loads from a constant pool.
2691 case Intrinsic::arm_neon_vshifts:
2692 case Intrinsic::arm_neon_vshiftu:
2693 case Intrinsic::arm_neon_vshiftls:
2694 case Intrinsic::arm_neon_vshiftlu:
2695 case Intrinsic::arm_neon_vshiftn:
2696 case Intrinsic::arm_neon_vrshifts:
2697 case Intrinsic::arm_neon_vrshiftu:
2698 case Intrinsic::arm_neon_vrshiftn:
2699 case Intrinsic::arm_neon_vqshifts:
2700 case Intrinsic::arm_neon_vqshiftu:
2701 case Intrinsic::arm_neon_vqshiftsu:
2702 case Intrinsic::arm_neon_vqshiftns:
2703 case Intrinsic::arm_neon_vqshiftnu:
2704 case Intrinsic::arm_neon_vqshiftnsu:
2705 case Intrinsic::arm_neon_vqrshiftns:
2706 case Intrinsic::arm_neon_vqrshiftnu:
2707 case Intrinsic::arm_neon_vqrshiftnsu: {
2708 MVT VT = N->getOperand(1).getValueType();
2710 unsigned VShiftOpc = 0;
2713 case Intrinsic::arm_neon_vshifts:
2714 case Intrinsic::arm_neon_vshiftu:
2715 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
2716 VShiftOpc = ARMISD::VSHL;
2719 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
2720 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
2721 ARMISD::VSHRs : ARMISD::VSHRu);
2726 case Intrinsic::arm_neon_vshiftls:
2727 case Intrinsic::arm_neon_vshiftlu:
2728 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
2730 llvm_unreachable("invalid shift count for vshll intrinsic");
2732 case Intrinsic::arm_neon_vrshifts:
2733 case Intrinsic::arm_neon_vrshiftu:
2734 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
2738 case Intrinsic::arm_neon_vqshifts:
2739 case Intrinsic::arm_neon_vqshiftu:
2740 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
2744 case Intrinsic::arm_neon_vqshiftsu:
2745 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
2747 llvm_unreachable("invalid shift count for vqshlu intrinsic");
2749 case Intrinsic::arm_neon_vshiftn:
2750 case Intrinsic::arm_neon_vrshiftn:
2751 case Intrinsic::arm_neon_vqshiftns:
2752 case Intrinsic::arm_neon_vqshiftnu:
2753 case Intrinsic::arm_neon_vqshiftnsu:
2754 case Intrinsic::arm_neon_vqrshiftns:
2755 case Intrinsic::arm_neon_vqrshiftnu:
2756 case Intrinsic::arm_neon_vqrshiftnsu:
2757 // Narrowing shifts require an immediate right shift.
2758 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
2760 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
2763 llvm_unreachable("unhandled vector shift");
2767 case Intrinsic::arm_neon_vshifts:
2768 case Intrinsic::arm_neon_vshiftu:
2769 // Opcode already set above.
2771 case Intrinsic::arm_neon_vshiftls:
2772 case Intrinsic::arm_neon_vshiftlu:
2773 if (Cnt == VT.getVectorElementType().getSizeInBits())
2774 VShiftOpc = ARMISD::VSHLLi;
2776 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
2777 ARMISD::VSHLLs : ARMISD::VSHLLu);
2779 case Intrinsic::arm_neon_vshiftn:
2780 VShiftOpc = ARMISD::VSHRN; break;
2781 case Intrinsic::arm_neon_vrshifts:
2782 VShiftOpc = ARMISD::VRSHRs; break;
2783 case Intrinsic::arm_neon_vrshiftu:
2784 VShiftOpc = ARMISD::VRSHRu; break;
2785 case Intrinsic::arm_neon_vrshiftn:
2786 VShiftOpc = ARMISD::VRSHRN; break;
2787 case Intrinsic::arm_neon_vqshifts:
2788 VShiftOpc = ARMISD::VQSHLs; break;
2789 case Intrinsic::arm_neon_vqshiftu:
2790 VShiftOpc = ARMISD::VQSHLu; break;
2791 case Intrinsic::arm_neon_vqshiftsu:
2792 VShiftOpc = ARMISD::VQSHLsu; break;
2793 case Intrinsic::arm_neon_vqshiftns:
2794 VShiftOpc = ARMISD::VQSHRNs; break;
2795 case Intrinsic::arm_neon_vqshiftnu:
2796 VShiftOpc = ARMISD::VQSHRNu; break;
2797 case Intrinsic::arm_neon_vqshiftnsu:
2798 VShiftOpc = ARMISD::VQSHRNsu; break;
2799 case Intrinsic::arm_neon_vqrshiftns:
2800 VShiftOpc = ARMISD::VQRSHRNs; break;
2801 case Intrinsic::arm_neon_vqrshiftnu:
2802 VShiftOpc = ARMISD::VQRSHRNu; break;
2803 case Intrinsic::arm_neon_vqrshiftnsu:
2804 VShiftOpc = ARMISD::VQRSHRNsu; break;
2807 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
2808 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
2811 case Intrinsic::arm_neon_vshiftins: {
2812 MVT VT = N->getOperand(1).getValueType();
2814 unsigned VShiftOpc = 0;
2816 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
2817 VShiftOpc = ARMISD::VSLI;
2818 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
2819 VShiftOpc = ARMISD::VSRI;
2821 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
2824 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
2825 N->getOperand(1), N->getOperand(2),
2826 DAG.getConstant(Cnt, MVT::i32));
2829 case Intrinsic::arm_neon_vqrshifts:
2830 case Intrinsic::arm_neon_vqrshiftu:
2831 // No immediate versions of these to check for.
2838 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
2839 /// lowers them. As with the vector shift intrinsics, this is done during DAG
2840 /// combining instead of DAG legalizing because the build_vectors for 64-bit
2841 /// vector element shift counts are generally not legal, and it is hard to see
2842 /// their values after they get legalized to loads from a constant pool.
2843 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
2844 const ARMSubtarget *ST) {
2845 MVT VT = N->getValueType(0);
2847 // Nothing to be done for scalar shifts.
2848 if (! VT.isVector())
2851 assert(ST->hasNEON() && "unexpected vector shift");
2854 switch (N->getOpcode()) {
2855 default: llvm_unreachable("unexpected shift opcode");
2858 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
2859 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
2860 DAG.getConstant(Cnt, MVT::i32));
2865 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
2866 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
2867 ARMISD::VSHRs : ARMISD::VSHRu);
2868 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
2869 DAG.getConstant(Cnt, MVT::i32));
2875 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
2876 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
2877 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
2878 const ARMSubtarget *ST) {
2879 SDValue N0 = N->getOperand(0);
2881 // Check for sign- and zero-extensions of vector extract operations of 8-
2882 // and 16-bit vector elements. NEON supports these directly. They are
2883 // handled during DAG combining because type legalization will promote them
2884 // to 32-bit types and it is messy to recognize the operations after that.
2885 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
2886 SDValue Vec = N0.getOperand(0);
2887 SDValue Lane = N0.getOperand(1);
2888 MVT VT = N->getValueType(0);
2889 MVT EltVT = N0.getValueType();
2890 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2892 if (VT == MVT::i32 &&
2893 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
2894 TLI.isTypeLegal(Vec.getValueType())) {
2897 switch (N->getOpcode()) {
2898 default: llvm_unreachable("unexpected opcode");
2899 case ISD::SIGN_EXTEND:
2900 Opc = ARMISD::VGETLANEs;
2902 case ISD::ZERO_EXTEND:
2903 case ISD::ANY_EXTEND:
2904 Opc = ARMISD::VGETLANEu;
2907 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
2914 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
2915 DAGCombinerInfo &DCI) const {
2916 switch (N->getOpcode()) {
2918 case ISD::ADD: return PerformADDCombine(N, DCI);
2919 case ISD::SUB: return PerformSUBCombine(N, DCI);
2920 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
2921 case ISD::INTRINSIC_WO_CHAIN:
2922 return PerformIntrinsicCombine(N, DCI.DAG);
2926 return PerformShiftCombine(N, DCI.DAG, Subtarget);
2927 case ISD::SIGN_EXTEND:
2928 case ISD::ZERO_EXTEND:
2929 case ISD::ANY_EXTEND:
2930 return PerformExtendCombine(N, DCI.DAG, Subtarget);
2935 /// isLegalAddressImmediate - Return true if the integer value can be used
2936 /// as the offset of the target addressing mode for load / store of the
2938 static bool isLegalAddressImmediate(int64_t V, MVT VT,
2939 const ARMSubtarget *Subtarget) {
2946 if (Subtarget->isThumb()) { // FIXME for thumb2
2951 switch (VT.getSimpleVT()) {
2952 default: return false;
2967 if ((V & (Scale - 1)) != 0)
2970 return V == (V & ((1LL << 5) - 1));
2975 switch (VT.getSimpleVT()) {
2976 default: return false;
2981 return V == (V & ((1LL << 12) - 1));
2984 return V == (V & ((1LL << 8) - 1));
2987 if (!Subtarget->hasVFP2())
2992 return V == (V & ((1LL << 8) - 1));
2996 /// isLegalAddressingMode - Return true if the addressing mode represented
2997 /// by AM is legal for this target, for a load/store of the specified type.
2998 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
2999 const Type *Ty) const {
3000 MVT VT = getValueType(Ty, true);
3001 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3004 // Can never fold addr of global into load/store.
3009 case 0: // no scale reg, must be "r+i" or "r", or "i".
3012 if (Subtarget->isThumb()) // FIXME for thumb2
3016 // ARM doesn't support any R+R*scale+imm addr modes.
3023 int Scale = AM.Scale;
3024 switch (VT.getSimpleVT()) {
3025 default: return false;
3030 // This assumes i64 is legalized to a pair of i32. If not (i.e.
3031 // ldrd / strd are used, then its address mode is same as i16.
3033 if (Scale < 0) Scale = -Scale;
3037 return isPowerOf2_32(Scale & ~1);
3040 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3045 // Note, we allow "void" uses (basically, uses that aren't loads or
3046 // stores), because arm allows folding a scale into many arithmetic
3047 // operations. This should be made more precise and revisited later.
3049 // Allow r << imm, but the imm has to be a multiple of two.
3050 if (AM.Scale & 1) return false;
3051 return isPowerOf2_32(AM.Scale);
3058 static bool getARMIndexedAddressParts(SDNode *Ptr, MVT VT,
3059 bool isSEXTLoad, SDValue &Base,
3060 SDValue &Offset, bool &isInc,
3061 SelectionDAG &DAG) {
3062 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3065 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3067 Base = Ptr->getOperand(0);
3068 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3069 int RHSC = (int)RHS->getZExtValue();
3070 if (RHSC < 0 && RHSC > -256) {
3071 assert(Ptr->getOpcode() == ISD::ADD);
3073 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3077 isInc = (Ptr->getOpcode() == ISD::ADD);
3078 Offset = Ptr->getOperand(1);
3080 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3082 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3083 int RHSC = (int)RHS->getZExtValue();
3084 if (RHSC < 0 && RHSC > -0x1000) {
3085 assert(Ptr->getOpcode() == ISD::ADD);
3087 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3088 Base = Ptr->getOperand(0);
3093 if (Ptr->getOpcode() == ISD::ADD) {
3095 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3096 if (ShOpcVal != ARM_AM::no_shift) {
3097 Base = Ptr->getOperand(1);
3098 Offset = Ptr->getOperand(0);
3100 Base = Ptr->getOperand(0);
3101 Offset = Ptr->getOperand(1);
3106 isInc = (Ptr->getOpcode() == ISD::ADD);
3107 Base = Ptr->getOperand(0);
3108 Offset = Ptr->getOperand(1);
3112 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3116 static bool getT2IndexedAddressParts(SDNode *Ptr, MVT VT,
3117 bool isSEXTLoad, SDValue &Base,
3118 SDValue &Offset, bool &isInc,
3119 SelectionDAG &DAG) {
3120 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3123 Base = Ptr->getOperand(0);
3124 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3125 int RHSC = (int)RHS->getZExtValue();
3126 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3127 assert(Ptr->getOpcode() == ISD::ADD);
3129 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3131 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3132 isInc = Ptr->getOpcode() == ISD::ADD;
3133 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3141 /// getPreIndexedAddressParts - returns true by value, base pointer and
3142 /// offset pointer and addressing mode by reference if the node's address
3143 /// can be legally represented as pre-indexed load / store address.
3145 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3147 ISD::MemIndexedMode &AM,
3148 SelectionDAG &DAG) const {
3149 if (Subtarget->isThumb1Only())
3154 bool isSEXTLoad = false;
3155 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3156 Ptr = LD->getBasePtr();
3157 VT = LD->getMemoryVT();
3158 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3159 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3160 Ptr = ST->getBasePtr();
3161 VT = ST->getMemoryVT();
3166 bool isLegal = false;
3167 if (Subtarget->isThumb() && Subtarget->hasThumb2())
3168 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3169 Offset, isInc, DAG);
3171 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3172 Offset, isInc, DAG);
3176 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3180 /// getPostIndexedAddressParts - returns true by value, base pointer and
3181 /// offset pointer and addressing mode by reference if this node can be
3182 /// combined with a load / store to form a post-indexed load / store.
3183 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3186 ISD::MemIndexedMode &AM,
3187 SelectionDAG &DAG) const {
3188 if (Subtarget->isThumb1Only())
3193 bool isSEXTLoad = false;
3194 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3195 VT = LD->getMemoryVT();
3196 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3197 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3198 VT = ST->getMemoryVT();
3203 bool isLegal = false;
3204 if (Subtarget->isThumb() && Subtarget->hasThumb2())
3205 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3208 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3213 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3217 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3221 const SelectionDAG &DAG,
3222 unsigned Depth) const {
3223 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3224 switch (Op.getOpcode()) {
3226 case ARMISD::CMOV: {
3227 // Bits are known zero/one if known on the LHS and RHS.
3228 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3229 if (KnownZero == 0 && KnownOne == 0) return;
3231 APInt KnownZeroRHS, KnownOneRHS;
3232 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3233 KnownZeroRHS, KnownOneRHS, Depth+1);
3234 KnownZero &= KnownZeroRHS;
3235 KnownOne &= KnownOneRHS;
3241 //===----------------------------------------------------------------------===//
3242 // ARM Inline Assembly Support
3243 //===----------------------------------------------------------------------===//
3245 /// getConstraintType - Given a constraint letter, return the type of
3246 /// constraint it is for this target.
3247 ARMTargetLowering::ConstraintType
3248 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3249 if (Constraint.size() == 1) {
3250 switch (Constraint[0]) {
3252 case 'l': return C_RegisterClass;
3253 case 'w': return C_RegisterClass;
3256 return TargetLowering::getConstraintType(Constraint);
3259 std::pair<unsigned, const TargetRegisterClass*>
3260 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3262 if (Constraint.size() == 1) {
3263 // GCC RS6000 Constraint Letters
3264 switch (Constraint[0]) {
3266 if (Subtarget->isThumb1Only())
3267 return std::make_pair(0U, ARM::tGPRRegisterClass);
3269 return std::make_pair(0U, ARM::GPRRegisterClass);
3271 return std::make_pair(0U, ARM::GPRRegisterClass);
3274 return std::make_pair(0U, ARM::SPRRegisterClass);
3276 return std::make_pair(0U, ARM::DPRRegisterClass);
3280 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3283 std::vector<unsigned> ARMTargetLowering::
3284 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3286 if (Constraint.size() != 1)
3287 return std::vector<unsigned>();
3289 switch (Constraint[0]) { // GCC ARM Constraint Letters
3292 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3293 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3296 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3297 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3298 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3299 ARM::R12, ARM::LR, 0);
3302 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3303 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3304 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3305 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3306 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3307 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3308 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3309 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3311 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3312 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3313 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3314 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3318 return std::vector<unsigned>();
3321 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3322 /// vector. If it is invalid, don't add anything to Ops.
3323 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3326 std::vector<SDValue>&Ops,
3327 SelectionDAG &DAG) const {
3328 SDValue Result(0, 0);
3330 switch (Constraint) {
3332 case 'I': case 'J': case 'K': case 'L':
3333 case 'M': case 'N': case 'O':
3334 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3338 int64_t CVal64 = C->getSExtValue();
3339 int CVal = (int) CVal64;
3340 // None of these constraints allow values larger than 32 bits. Check
3341 // that the value fits in an int.
3345 switch (Constraint) {
3347 if (Subtarget->isThumb1Only()) {
3348 // This must be a constant between 0 and 255, for ADD
3350 if (CVal >= 0 && CVal <= 255)
3352 } else if (Subtarget->isThumb2()) {
3353 // A constant that can be used as an immediate value in a
3354 // data-processing instruction.
3355 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3358 // A constant that can be used as an immediate value in a
3359 // data-processing instruction.
3360 if (ARM_AM::getSOImmVal(CVal) != -1)
3366 if (Subtarget->isThumb()) { // FIXME thumb2
3367 // This must be a constant between -255 and -1, for negated ADD
3368 // immediates. This can be used in GCC with an "n" modifier that
3369 // prints the negated value, for use with SUB instructions. It is
3370 // not useful otherwise but is implemented for compatibility.
3371 if (CVal >= -255 && CVal <= -1)
3374 // This must be a constant between -4095 and 4095. It is not clear
3375 // what this constraint is intended for. Implemented for
3376 // compatibility with GCC.
3377 if (CVal >= -4095 && CVal <= 4095)
3383 if (Subtarget->isThumb1Only()) {
3384 // A 32-bit value where only one byte has a nonzero value. Exclude
3385 // zero to match GCC. This constraint is used by GCC internally for
3386 // constants that can be loaded with a move/shift combination.
3387 // It is not useful otherwise but is implemented for compatibility.
3388 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3390 } else if (Subtarget->isThumb2()) {
3391 // A constant whose bitwise inverse can be used as an immediate
3392 // value in a data-processing instruction. This can be used in GCC
3393 // with a "B" modifier that prints the inverted value, for use with
3394 // BIC and MVN instructions. It is not useful otherwise but is
3395 // implemented for compatibility.
3396 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3399 // A constant whose bitwise inverse can be used as an immediate
3400 // value in a data-processing instruction. This can be used in GCC
3401 // with a "B" modifier that prints the inverted value, for use with
3402 // BIC and MVN instructions. It is not useful otherwise but is
3403 // implemented for compatibility.
3404 if (ARM_AM::getSOImmVal(~CVal) != -1)
3410 if (Subtarget->isThumb1Only()) {
3411 // This must be a constant between -7 and 7,
3412 // for 3-operand ADD/SUB immediate instructions.
3413 if (CVal >= -7 && CVal < 7)
3415 } else if (Subtarget->isThumb2()) {
3416 // A constant whose negation can be used as an immediate value in a
3417 // data-processing instruction. This can be used in GCC with an "n"
3418 // modifier that prints the negated value, for use with SUB
3419 // instructions. It is not useful otherwise but is implemented for
3421 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3424 // A constant whose negation can be used as an immediate value in a
3425 // data-processing instruction. This can be used in GCC with an "n"
3426 // modifier that prints the negated value, for use with SUB
3427 // instructions. It is not useful otherwise but is implemented for
3429 if (ARM_AM::getSOImmVal(-CVal) != -1)
3435 if (Subtarget->isThumb()) { // FIXME thumb2
3436 // This must be a multiple of 4 between 0 and 1020, for
3437 // ADD sp + immediate.
3438 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3441 // A power of two or a constant between 0 and 32. This is used in
3442 // GCC for the shift amount on shifted register operands, but it is
3443 // useful in general for any shift amounts.
3444 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3450 if (Subtarget->isThumb()) { // FIXME thumb2
3451 // This must be a constant between 0 and 31, for shift amounts.
3452 if (CVal >= 0 && CVal <= 31)
3458 if (Subtarget->isThumb()) { // FIXME thumb2
3459 // This must be a multiple of 4 between -508 and 508, for
3460 // ADD/SUB sp = sp + immediate.
3461 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3466 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3470 if (Result.getNode()) {
3471 Ops.push_back(Result);
3474 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,